text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Testing Ising Models
Constantinos Daskalakis∗
EECS, MIT
[email protected]
Nishanth Dikkala†
EECS, MIT
[email protected]
Gautam Kamath‡
EECS, MIT
[email protected]
arXiv:1612.03147v4 [cs.DS] 30 Oct 2017
November 1, 2017
Abstract
Given samples from an unknown multivariate distribution p, is it possible to distinguish
whether p is the product of its marginals versus p being far from every product distribution?
Similarly, is it possible to distinguish whether p equals a given distribution q versus p and q
being far from each other? These problems of testing independence and goodness-of-fit have
received enormous attention in statistics, information theory, and theoretical computer science,
with sample-optimal algorithms known in several interesting regimes of parameters [BFF+ 01,
Pan08, VV17, ADK15, DK16]. Unfortunately, it has also been understood that these problems
become intractable in large dimensions, necessitating exponential sample complexity.
Motivated by the exponential lower bounds for general distributions as well as the ubiquity
of Markov Random Fields (MRFs) in the modeling of high-dimensional distributions, we initiate
the study of distribution testing on structured multivariate distributions, and in particular
the prototypical example of MRFs: the Ising Model. We demonstrate that, in this structured
setting, we can avoid the curse of dimensionality, obtaining sample and time efficient testers for
independence and goodness-of-fit. One of the key technical challenges we face along the way is
bounding the variance of functions of the Ising model.
∗
Supported by a Microsoft Research Faculty Fellowship, NSF CCF-1551875, CCF-1617730, CCF-1650733, and
ONR N00014-12-1-0999.
†
Supported by NSF CCF-1551875, CCF-1617730, CCF-1650733, and ONR N00014-12-1-0999.
‡
Supported by NSF CCF-1551875, CCF-1617730, CCF-1650733, and ONR N00014-12-1-0999.
Contents
1 Introduction
1.1 Further Technical Discussion and Highlights . . . . . . . .
1.1.1 Structural Results for Ferromagnetic Ising Models
1.1.2 A Linear Lower Bound for Testing Ising Models . .
1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
8
8
9
10
2 Preliminaries
10
2.1 Input to Goodness-of-Fit Testing Algorithms . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Symmetric KL Divergence Between Two Ising Models . . . . . . . . . . . . . . . . . 13
3 A General Purpose Localization Algorithm
13
3.1 Independence Test using Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Identity Test using Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4 Improved Tests for Forests and Ferromagnetic Ising Models
4.1 Improved Algorithms for Independence and Identity Testing on Forests . .
4.2 Ferromagnetic Ising Models: A Structural Understanding and an Improved
dence Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Random Cluster Model . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Coupling between the Random Cluster Model and the Ising model
. . . . . .
Indepen. . . . . .
. . . . . .
. . . . . .
5 An Improved Test for High-Temperature Ising Models: A Learn-then-Test
gorithm
5.1 Weak Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Weak Learning the Edges of an Ising Model . . . . . . . . . . . . . . . . .
5.2 Testing Our Learned Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Combining Learning and Testing . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Balancing Weak Learning and Testing . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Changes Required for General Independence and Identity Testing . . . . . . . . .
5.5.1 Independence Testing under an External Field . . . . . . . . . . . . . . .
5.5.2 Identity Testing under No External Field . . . . . . . . . . . . . . . . . .
5.5.3 Identity Testing under an External Field . . . . . . . . . . . . . . . . . . .
16
17
20
20
21
Al.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
28
28
31
31
32
32
32
34
35
6 Comparing Localization and Learn-then-Test Algorithms
36
7 Bounding the Variance of Functions of the Ising Model in the High-Temperature
Regime
7.1 Overview of the Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Bounding Variance of Z 0 Under No External Field . . . . . . . . . . . . . . . . . . .
0
7.3 Bounding Variance of Zcen
Under Arbitrary External Field . . . . . . . . . . . . . . .
40
40
41
42
8 Lower Bounds
8.1 Dependences on n . . .
8.2 Dependences on h, β . .
8.3 Lower Bound Proofs . .
8.3.1 Proof of Theorem
8.3.2 Proof of Theorem
43
43
44
44
44
46
. .
. .
. .
18
19
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8.3.3
Proof of Theorem 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
A Weakly Learning Rademacher Random Variables
55
B An Attempt towards Testing by Learning in KL-divergence
57
1
Introduction
The two most fundamental problems in Statistics are perhaps testing independence and goodnessof-fit. Independence testing is the problem of distinguishing, given samples from a multivariate
distribution p, whether or not it is the product of its marginals. The applications of this problem
abound: for example, a central problem in genetics is to test, given genomes of several individuals, whether certain single-nucleotide-polymorphisms (SNPs) are independent from each other. In
anthropological studies, a question that arises over and over again is testing whether the behaviors of individuals on a social network are independent; see e.g. [CF07]. The related problem of
goodness-of-fit testing is that of distinguishing, given samples from p, whether or not it equals a
specific “model” q. This problem arises whenever one has a hypothesis (model) about the random
source generating the samples and needs to verify whether the samples conform to the hypothesis.
Testing independence and goodness-of-fit have a long history in statistics, since the early days;
for some old and some more recent references see, e.g., [Pea00, Fis35, RS81, Agr12]. Traditionally,
the emphasis has been on the asymptotic analysis of tests, pinning down their error exponents as
the number of samples tends to infinity [Agr12, TAW10]. In the two decades or so, distribution
testing has also piqued the interest of theoretical computer scientists, where the emphasis has been
different [BFF+ 01, Pan08, LRR13, VV17, ADK15, CDGR16, DK16]. In contrast to much of the
statistics literature, the goal has been to minimize the number of samples required for testing. From
this vantage point, our testing problems take the following form:
Goodness-of-fit (or Identity) Testing: Given sample access to an unknown distribution p over
Σn and a parameter ε > 0, the goal is to distinguish with probability at least 2/3 between
p = q and d(p, q) > ε, for some specific distribution q, from as few samples as possible.
Independence Testing: Given sample access to an unknown distribution p over Σn and a
parameter ε > 0, the goal is to distinguish with probability at least 2/3 between p ∈ I(Σn )
and d(p, I(Σn )) > ε, where I(Σn ) is the set of product distributions over Σn , from as few
samples as possible.
In these problem definitions, Σ is some discrete alphabet, and d(·, ·) some notion of distance or
divergence between distributions, such as the total variation distance or the KL divergence. As
usual, 32 is an arbitrary choice of a constant, except that it is bounded away from 21 . It can always
be boosted to some arbitrary 1 − δ at the expense of a multiplicative factor of O(log 1/δ) in the
sample complexity.
For both testing problems, recent work has identified tight upper and lower bounds on their
sample complexity [Pan08, VV17, ADK15, DK16]: when d is taken to bethe total
variation distance,
n/2
the optimal sample complexity for both problems turns out to be Θ |Σ|ε2 , i.e. exponential in
the dimension. As modern applications commonly involve high-dimensional data, this curse of
dimensionality makes the above testing goals practically unattainable. Nevertheless, there is a
sliver of hope, and it lies with the nature of all known sample-complexity lower bounds, which
construct highly-correlated distributions that are hard to distinguish from the set of independent
distributions [ADK15, DK16], or from a particular distribution q [Pan08]. Worst-case analysis of
this sort seems overly pessimistic, as these instances are unlikely to arise in real-world data. As such,
we propose testing high-dimensional distributions which are structured, and thus could potentially
rule out such adversarial distributions.
Motivated by the above considerations and the ubiquity of Markov Random Fields (MRFs) in
the modeling of high-dimensional distributions (see [Jor10] for the basics of MRFs and the references
[STW10, KNS07] for a sample of applications), we initiate the study of distribution testing for the
1
prototypical example of MRFs: the Ising Model, which captures all binary MRFs with node and
edge potentials.1 Recall that the Ising model is a distribution over {−1, 1}n , defined in terms of
a graph G = (V, E) with n nodes. It is parameterized by a scalar parameter θu,v for every edge
(u, v) ∈ E, and a scalar parameter θv for every node v ∈ V , in terms of which it samples a vector
x ∈ {±1}V with probability:
X
X
~ ,
θu,v xu xv − Φ(θ)
(1)
p(x) = exp
θ v xv +
v∈V
(u,v)∈E
~ is the log-partition function, ensuring that the distribution
where θ~ is the parameter vector and Φ(θ)
is normalized. Intuitively, there is a random variable Xv sitting on every node of G, which may be
in one of two states, or spins: up (+1) or down (-1). The scalar parameter θv models a local (or
“external”) field at node v. The sign of θv represents whether this local field favors Xv taking the
value +1, i.e. the up spin, when θv > 0, or the value −1, i.e. the down spin, when θv < 0, and its
magnitude represents the strength of the local field. We will say a model is “without external field”
when θv = 0 for all v ∈ V . Similarly, θu,v represents the direct interaction between nodes u and v.
Its sign represents whether it favors equal spins, when θu,v > 0, or opposite spins, when θu,v < 0,
and its magnitude corresponds to the strength of the direct interaction. Of course, depending on the
structure of the Ising model and the edge parameters, there may be indirect interactions between
nodes, which may overwhelm local fields or direct interactions.
The Ising model has a rich history, starting with its introduction by statistical physicists as a
probabilistic model to study phase transitions in spin systems [Isi25]. Since then it has found a
myriad of applications in diverse research disciplines, including probability theory, Markov chain
Monte Carlo, computer vision, theoretical computer science, social network analysis, game theory,
and computational biology [LPW09, Cha05, Fel04, DMR11, GG86, Ell93, MS10]. The ubiquity
of these applications motivate the problem of inferring Ising models from samples, or inferring
statistical properties of Ising models from samples. This type of problem has enjoyed much study in
statistics, machine learning, and information theory, see, i.e., [CL68, AKN06, CT06, RWL10, JJR11,
SW12, BGS14, Bre15, VMLC16, BK16, Bha16, BM16, MdCCU16, HKM17, KM17]. Much of prior
work has focused on parameter learning, where the goal is to determine the parameters of an Ising
model to which sample access is given. In contrast to this type of work, which focuses on discerning
parametrically distant Ising models, our goal is to discern statistically distant Ising models, in the
hopes of dramatic improvements in the sample complexity. (We will come to a detailed comparison
between the two inference goals shortly, after we have stated our results.) To be precise, we study
the following problems:
Ising Model Goodness-of-fit (or Identity) Testing: Given sample access to an unknown Ising
model p (with unknown parameters over an unknown graph) and a parameter ε > 0, the
goal is to distinguish with probability at least 2/3 between p = q and dSKL (p, q) > ε, for
some specific Ising model q, from as few samples as possible.
Ising Model Independence Testing: Given sample access to an unknown Ising model p (with
unknown parameters over an unknown graph) and a parameter ε > 0, the goal is to distinguish with probability at least 2/3 between p ∈ In and dSKL (p, In ) > ε, where In are all
product distributions over {−1, 1}n , from as few samples as possible.
We note that there are several potential notions of statistical distance one could consider — classically, total variation distance and the Kullback-Leibler (KL) divergence have seen the most study.
1
This follows trivially by the definition of MRFs, and elementary Fourier analysis of Boolean functions.
2
As our focus here is on upper bounds, we consider the symmetrized KL divergence dSKL , which is
a “harder” notion of distance than both: in particular, testers for dSKL immediately imply testers
for both total variation distance and the KL divergence. Moreover, by virtue of the fact that dSKL
upper-bounds KL in both directions, our tests offer useful information-theoretic interpretations of
rejecting a model q, such as data differencing and large deviation bounds in both directions.
Sample Applications: As an instantiation of our proposed testing problems for the Ising model
one may maintain the study of strategic behavior on a social network. To offer a little bit of
background, a body of work in economics has modeled strategic behavior on a social network as
the evolution of the Glauber dynamics of an Ising model, whose graph is the social network, and
whose parameters are related to the payoffs of the nodes under different selections of actions by
them and their neighbors. For example, [Ell93, MS10] employ this model to study the adoption of
competing technologies with network effects, e.g. iPhone versus Android phones. Glauber dynamics,
as described in Section 2, define the canonical Markov chain for sampling an Ising model. Hence an
observation of the actions (e.g. technologies) used by the nodes of the social network should offer
us a sample from the corresponding Ising model (at least if the Glauber dynamics have mixed).
An analyst may not know the underlying social network or may know the social network but not
the parameters of the underlying Ising model. In either case, how many independent observations
would he need to test, e.g., whether the nodes are adopting technologies independently, or whether
their adoptions conform to some conjectured parameters? Our results offer algorithms for testing
such hypotheses in this stylized model of strategic behavior on a network.
As another application, we turn to the field of computer vision. In the Bayesian setting, it is
assumed that images are generated according to some prior distribution. Often, practitioners take
this prior to be an Ising model in the binary case, or, in general, a higher-order MRF [GG86]. As
such, a dataset of images can be pictured as random samples from this prior. A natural question to
ask is, given some distribution, does a set of images conform to this prior? This problem corresponds
to goodness-of-fit testing for Ising models.
A third application comes up in the field of medicine and computational biology. In order to
improve diagnosis, symptom prediction and classification, as well as to improve overall healthcare
outcomes, graphical models are trained on data, often using heuristic methods [FLNP00] and surgeon intuition, thereby incorporating hard-wired expert knowledge; see, i.e., the pneumonia graphical model identified in [LAFH01]. Our methods give efficient algorithms for testing the accuracy
of such models. Furthermore, when the discrepancy is large, we expect that our algorithms could
reveal the structural reasons for the discrepancy, i.e., blaming a large portion of the error on a
misspecified edge.
Main Results and Techniques: Our main result is the following:
Theorem 1. Both Ising Model Goodness-of-fit Testing and Ising Model Independence Testing can
be solved from poly n, 1ε samples in polynomial time.
There are several variants of our testing problems, resulting from different knowledge that the analyst
may have about the structure of the graph (connectivity, density), the nature of the interactions
(attracting, repulsing, or mixed), as well as the temperature (low vs high). We proceed to discuss
all these variants, instantiating the resulting polynomial sample complexity in the above theorem.
We also illuminate the techniques involved to prove these theorems. This discussion should suffice
in evaluating the merits of the results and techniques of this paper.
A. Our Baseline Result. In the least favorable regime, i.e. when the analyst is oblivious to the
structure of the Ising model
4 2 p, 2the
signs of the interactions, and their strength, the polynomial in
p
n β +n h2
Theorem 1 becomes O
. In this expression, β = max{|θu,v
|} for independence testing,
ε2
3
p
q
and β = max{max{|θu,v
|}, max{|θu,v
|}} for goodness-of-fit testing, while h = 0 for independence
p
testing, and h = max{max{|θu |}, max{|θuq |}} for goodness-of-fit testing; see Theorem 2. If the
analyst has an upper bound on the maximum
dmax (of all Ising models involved in the
2 2 degree
n dmax β 2 +ndmax h2
problem) the dependence improves to O
, while if the analyst has an upper
ε2
bound on the total number of edges m, then max{m, n} takes the role of ndmax in the previous
bound; see Theorem 2.
Technical Discussion 1.0: “Testing via Localization.” All the bounds mentioned so far
are obtained via a simple localization argument showing that, whenever two Ising models p and q
satisfy dSKL (p, q) > ε, then “we can blame it on a node or an edge;” i.e. there exists a node with
significantly different bias under p and q or a pair of nodes u, v whose covariance is significantly
different under the two models. Pairwise correlation tests are a simple screening that is often
employed in practice. For our setting, there is a straighforward and elegant way to show that
pair-wise (and not higher-order) correlation tests suffice; see Lemma 4.
For more details about our baseline localization tester see Section 3.
B. Anchoring Our Expectations. Our next results aim at improving the afore-described baseline
bound. Before stating these improvements, however, it is worth comparing the sample complexity of
our baseline results to the sample complexity of learning. Indeed, one might expect and it is often the
case that testing problems can be solved in a two-step fashion, by first learning a hypothesis p̂ that is
statistically close to the true p and then using the learned hypothesis p̂ as a proxy for p to determine
whether it is close to or far from some q, or some set of distributions. Given that the KL divergence
and its symmetrized version do not satisfy the triangle inequality, however, it is not clear how such
an approach would work. Even if it could, the only algorithm that we are aware of for proper
learning Ising models, which offers KL divergence guarantees but does not scale exponentially with
the maximum degree and β, is a straightforward
net-based algorithm. This algorithm, explained
n6 β 2 +n4 h2
in Section B, requires Ω
samples and is time inefficient. In particular, our baseline
ε2
algorithm already beats this sample complexity and is also time-efficient. Alternatively, one could
aim to parameter-learn p; see, e.g., [SW12, Bre15, VMLC16] and their references. However, these
algorithms require sample complexity that is exponential in the maximum degree [SW12], and they
typically use samples exponential in β as well [Bre15, VMLC16]. For instance, if we use [VMLC16],
which is one of the state-of-the-art algorithms, to do parameter learning prior to testing, we would
4 β·dmax
need Õ( n ·2 ε2
) samples to learn p’s parameters closely enough to be able to do the testing
afterwards. Our baseline result beats this sample complexity, dramatically so if the degrees are
unbounded.
The problem of learning the structure of Ising models (i.e., determining which edges are present
in the graph) has enjoyed much study, especially in information theory – see [Bre15, VMLC16] for
some recent results. At a first glance, one may hope that these results have implications for testing
Ising models. However, thematic similarities aside, the two problems are qualitatively very different
– our problem focuses on statistical estimation, while theirs looks at structural estimation. To
point out some qualitative differences for these two problems, the complexity of structure learning
is exponential in the maximum degree and β, while only logarithmic in n. On the other hand, for
testing Ising models, the complexity has a polynomial dependence in all three parameters, which is
both necessary and sufficient.
C. Trees and Ferromagnets. When p is a tree-structured (or forest-structured) Ising model,
then independence testing can be performed computationally efficiently without any dependence on
β, with an additional quadratic improvement with respect to the other parameters. In particular,
without external fields, i.e. max{|θup |} = 0, independence can be solved from O( nε ) samples, and
this result is tight when m = O(n); see Theorem 3 for an upper bound and Theorem 19 for a lower
4
bound. Interestingly, we show the dependence on β cannot be avoided in the presence of external
fields, or if we switch to the problem of identity testing; see Theorem 20. In the latter case, we
can at least maintain the linear dependence on n; see Theorem 4. Similar results hold when p is a
p
ferromagnet, i.e. θu,v
≥ 0, with no external fields, even if it is not a tree. In particular, the sample
complexity becomes O( max{m,n}
) (which is again tight when m = O(n)), see Theorem 5.
ε
Technical Discussion 2.0: “Testing via Strong Localization.” The improvements that we
have just discussed are obtained via the same localization approach discussed earlier, which resulted
into our baseline tester. That is, we are still going to “blame it on a node or an edge.” The removal
of the β dependence and the improved running times are due to the proof of a structural lemma,
which relates the parameter θu,v on some edge (u, v) of the Ising model to the E[Xu Xv ]. We show
that for forest-structured Ising models with no external fields, E[Xu Xv ] = tanh(θu,v ), see Lemma 8.
A similar statement holds for ferromagnets with no external field, i.e., E[Xu Xv ] ≥ tanh(θu,v ), see
Lemma 11. The proof of the structural lemma for trees/forests is straightforward. Intuitively, the
only source of correlation between the endpoints u and v of some edge (u, v) of the Ising model is
the edge itself, as besides this edge there are no other paths between u and v that would provide
alternative avenues for correlation. Significant more work is needed to prove the inequality for
ferromagnets on arbitrary graphs. Now, there may be several paths between u and v besides the
edge connecting them. Of course, because the model is a ferromagnet, these paths should intuitively
only contribute to increase E[Xu Xv ] beyond tanh(θu,v ). But making this formal is not easy, as
calculations involving the Ising model quickly become unwieldy beyond trees.2 Our argument uses
a coupling between (an appropriate generalization of) the Fortuin-Kasteleyn random cluster model
and the Ising model. The coupling provides an alternative way to sample the Ising model by first
sampling a random clustering of the nodes, and then assigning uniformly random spins to the
sampled clusters. Moreover, it turns out that the probability that two nodes u and v land in the
same cluster increases as the vector of parameters θ~ of the Ising model increases. Hence, we can
work inductively. If only edge (u, v) were present, then E[Xu Xv ] = tanh(θu,v ). As we start adding
edges, the probability that u, v land in the same cluster increases, hence the probability that they
receive the same spin increases, and therefore E[Xu Xv ] increases.
A slightly more detailed discussion of the structural result for ferromagnets is in Section 1.1.1,
and full details about our testers for trees and ferromagnets can be found in Sections 4.1 and 4.2,
respectively.
D. Dobrushin’s Uniqueness Condition and the High-Temperature Regime. Motivated by
phenomena in the physical world, the study of Ising models has identified phase transitions in the
behavior of the model as its parameters vary. A common transition occurs as the temperature of the
model changes from low to high. As the parameters θ~ correspond to inverse (individualistic) temperatures, this corresponds to a transition of these parameters from low values (high temperature)
to high values (low temperature). Often the transition to high temperature is identified with the
satisfaction of Dobrushin-type conditions [Geo11]. Under such conditions, the model enjoys a number of good properties, including rapid mixing of the Glauber dynamics, spatial mixing properties,
and uniqueness of measure. The Ising model has been studied extensively in such high-temperature
regimes [Dob56, Cha05, Hay06, DGJ08], and it is a regime that is often used in practice.
In the high-temperature regime, we show that we can improve our baseline result without making
ferromagnetic or tree-structure assumptions, using a non-localization based argument, explained
next. In particular, we show in Theorem 7 that under high temperature
with
no external
and
n10/3
fields independence testing can be done computationally efficiently from Õ ε2 d2
samples, which
max
2
We note that the partition function is #P-hard to compute[JS93].
5
improves upon our baseline result
4/3if dmax is large enough. For instance, when dmax = Ω(n), the
sample complexity becomes Õ nε2 . Other tradeoffs between β, dmax and the sample complexity
are explored in Theorem 6. Similar improvements hold when external fields are present (Theorem 9),
as well as for identity testing, without and with external fields (Theorems 10 and 11).
We offer some intuition about the improvements in Figures 1 and 2 (appearing in Section 6),
which are plotted for high temperature and no external fields. In Figure 1, we plot the number of
1
samples required for testing Ising models with no external fields when β = Θ( dmax
) as dmax varies.
The horizontal axis is logn dmax . We see that localization is the better algorithm for degrees smaller
than O(n2/3 ), above which its complexity can be improved. In particular, the sample complexity
is O(n2 /ε2 ) until degree dmax = O(n2/3 ), beyond which it drops inverse quadratically in dmax . In
Figure 2, we consider a different tradeoff. We plot the number of samples required when β = n−α
and the degree of the graph varies. In particular, we see three regimes as a function of whether the
Ising model is in high temperature (dmax = O(na )) or low temperature (dmax = ω(na )), and also
which of our techniques localization vs non-localization gives better sample complexity bounds.
Technical Discussion 3.0: “Testing via a Global Statistic, and Variance Bounds.”
One way or another all our results up to this point had been obtained via localization, namely
blaming the distance of p from independence, or from some distribution q to a node or an edge.
Our improved bounds employ non-localized statistics that look at all P
the nodes of the Ising model
simultaneously. Specifically, we employ statistics of the form Z =
e=(u,v)∈E ce Xu Xv for some
appropriately chosen signs ce .
The first challenge we encounter here involves selecting the signs ce in accordance with the sign
of each edge marginal’s expectation, E[Xu Xv ]. This is crucial to establish that the resulting statistic
will be able to discern between the two cases. While the necessary estimates of these signs could be
computed independently for each edge, this would incur an unnecessary overhead of O(n2 ) in the
number of samples. Instead we try to learn signs that have a non-trivial agreement with the correct
signs, from fewer samples. Despite the Xu Xv terms potentially having nasty correlations with each
other, a careful analysis using anti-concentration calculations allows us to sidestep this O(n2 ) cost
and generate satisfactory estimates with a non-negligible probability, from fewer samples.
The second and more significant challenge involves bounding the variance of a statistic Z of the
above form. Since Z’s magnitude is at most O(n2 ), its variance can trivially be bounded by O(n4 ).
However, applying this bound in our algorithm gives a vacuous sample complexity. As the Xu ’s
will experience a complex correlation structure, it is not clear how one might arrive at non-trivial
bounds for the variance of such statistics, leading to the following natural question:
Question 1. How can one bound the variance of statistics over high-dimensional distributions?
This meta-question is at the heart of many high-dimensional statistical tasks, and we believe it
is important to develop general-purpose frameworks for such settings. In the context of the Ising
model, in fairly general regimes, we can show the variance to be Õ(n2 ). We consider this to be
surprising – stated another way, despite the complex correlations which may be present in the Ising
model, the summands in Z behave roughly as if they were pairwise independent.
This question has been studied in-depth in two recent works [DDK17, GLP17], which prove
concentration of measure for d-linear statistics over the Ising model. We note that these results are
stronger than what we require in this work – we need only variance bounds (which are implied by
concentration of measure) for bilinear statistics. Despite these stronger bounds, for completeness,
we present a proof of the variance bounds for bilinear statistics which we require3 . This approach
3
We thank Yuval Peres for directing us towards the reference [LPW09] and the tools required to prove these
bounds.
6
uses tools from [LPW09]. It requires a bound on the spectral gap of the Markov chain, and an
expected Lipschitz property of the statistic when a step is taken at stationarity. The technique is
described in Section 7, and the variance bounds are given in Theorems 16 and 17.
E. Our Main Lower Bound. The proof of our linear lower bound applies Le Cam’s method
[LC73]. Our construction is inspired by Paninski’s lower bound for uniformity testing [Pan08],
which involves pairing up domain elements and jointly perturbing their probabilities. This style of
construction is ubiquitous in univariate testing lower bounds. A naive application of this approach
would involve choosing a fixed matching of the nodes and randomly perturbing the weight of the
√
edges, which leads to an Ω( n) lower bound. We analyze a construction of a similar nature as
a warm-up for our main lower bound, while also proving a lower bound for uniformity testing on
product distributions over a binary alphabet (which are a special case of the Ising model where
no edges are present), see Theorem 18. To achieve the linear lower bound, we instead consider a
random matching of the nodes. The analysis of this case turns out to be much more involved due
to the complex structure of the probability function which corresponds to drawing k samples from
an Ising model on a randomly chosen matching. Indeed, our proof turns out to have a significantly
combinatorial flavor, and we believe that our techniques might be helpful for proving stronger lower
bounds in combinatorial settings for multivariate distributions. Our analysis of this construction
is tight, as uniformity testing on forests can be achieved with O(n) samples. We believe that a
super-linear lower bound would be very interesting, but also quite difficult to obtain. Proving our
linear lower bound already required a very careful analysis for a relatively simple construction, and
an improved lower bound would require analyzing a distribution over dense constructions, for which
an improved structural understanding is needed. A further technical discussion of this lower bound
is in Section 1.1.2, see Section 8 and Theorem 19 for a formal statement and full analysis of our
main lower bound. As mentioned before, we also show that the sample complexity must depend on
β and h in certain cases, see Theorem 20 for a formal statement.
Table 1 summarizes our algorithmic results.
The High-Dimensional Frontier and Related Work: We emphasize that we believe the study
of high-dimensional distribution testing to be of significant importance, as real-world applications
often involve multivariate data. As univariate distribution testing is now very well understood, with
a thorough set of tools and techniques, this is the natural next frontier to attack. However, multivariate distributions pose several new technical challenges, and many of these univariate tools are
rendered obsolete – as such, we must extend these methods, or introduce new techniques entirely. It
is important to develop approaches which may be applicable in much more general high-dimensional
distribution testing settings, when there may be complex correlations between random variables.
First, it is important to get a grasp on the concentration and variance of statistics in these settings,
and we provide exposition of a technique for bounding the variance of some simple statistics. Additionally, our linear lower bound’s construction and analysis give insight into which instances cause
intractability to arise, and provide a recipe for the style of combinatorics required to analyze them.
In further works, the authors and another group have investigated more properties of multilinear
functions over the Ising model [DDK17, GLP17]. In the present work, we require and prove variance
bounds for bilinear functions of the Ising model. These other works prove concentration bounds
(which are qualitatively stronger than variance bounds) for multilinear functions of arbitrary degree
d (rather than just bilinear functions, which are of degree d = 2).
High-dimensional distribution testing has recently attracted the interest of the theoretical computer science community, with work concurrent to ours on testing Bayes networks4 [CDKS17, DP17].
4
Bayes nets are another type of graphical model, and are in general incomparable to Ising models.
7
Testing Problem
Independence
using Localization
Identity
using Localization
No External Field
2 2 2
β
Õ n dmax
ε2
Õ
n2 d2max β 2
ε2
Arbitrary External Field
2 2 2
β
Õ n dmax
ε2
Õ
n2 d2max β 2
ε2
n2 h2
ε2
+
Independence
under Dobrushin/high-temperature
Õ
n10/3 β 2
ε2
Õ
n10/3 β 2
ε2
Õ
n10/3 β 2
ε2
using Learn-Then-Test
Identity
under Dobrushin/high-temperature
Õ
n11/3 β 2
ε2
n5/3 h2
ε2
+
using Learn-Then-Test
Independence on Forests
Õ
using Improved Localization
Identity on Forests
using Improved Localization
Independence on Ferromagnets
using Improved Localization
Õ
Õ
n
ε
Õ
n·c(β)
ε
ndmax
ε
Õ
Õ
n2 β 2
ε2
n2 β 2
ε2
+
n2 h2
ε2
n2 d2max β 2
ε2
Table 1: Summary of our results in terms of the sample complexity upper bounds for the various
problems studied. n = number of nodes in the graph, dmax = maximum degree, β = maximum absolute value of edge parameters, h = maximum absolute value of node parameters (when applicable),
and c is a function discussed in Theorem 4.
It remains to be seen which other multivariate distribution classes of interest allow us to bypass the
curse of dimensionality.
We note that the paradigm of distribution testing under structural assumptions has been explored in the univariate setting, where we may assume the distribution is log-concave or k-modal.
This often allows exponential savings in the sample complexity [DDS+ 13, DKN15b, DKN15a].
1.1
Further Technical Discussion and Highlights
In this section, we give a slightly more in-depth discussion of some of the technical highlights of
our work. For full details and more discussion, the interested reader can refer to the corresponding
sections in the body of the paper.
1.1.1
Structural Results for Ferromagnetic Ising Models
Our general-purpose testing algorithm is a localization-based algorithm – in particular, it operates
based on the structural property that if two Ising models (with no external field) are far from each
other, they will have a distant edge marginal. We convert this structural property to an algorithm
by estimating each edge marginal and testing whether they match for the two models. However,
the underlying structural property is quantitatively weak, and leads to sub-optimal testing bounds.
In some cases of interest, we can derive quantitatively stronger versions of this structural result,
8
giving us more efficient algorithms.
For instance, one can consider the ferromagnetic case, where one has all edge parameters θe ≥ 0.
We would like to derive a relationship between an edge marginal (i.e., E[Xu Xv ] for an edge e =
(u, v)) and the parameter on that edge θe . For a tree-structured Ising model with no external field
(ferromagnetic or not), it is not hard to show that E[Xu Xv ] = tanh(θe ) – for small edge parameters,
this indicates a linear relationship between the edge marginal and the edge parameter. Intuitively,
if a model is ferromagnetic and contains cycles, these cycles should only increase the correlation
between adjacent nodes, i.e., we would expect that E[Xu Xv ] ≥ tanh(θe ). While this is true, it
proves surprisingly difficult to prove directly, and we must instead view the Ising model through
the Fortuin-Kastelyn random cluster model.
At a high level, the Fortuin-Kastelyn random cluster model is defined for a graph G = (V, E)
with a probability parameter 0 < re < 1 on each edge. This parameter indicates the probability of a
bond being present on edge e (i.e., the distribution gives a measure over {0, 1}E ), placing this model
into the space of bond percolation models (see Section 4.2.1 and (21) for the formal definition). It
turns out that an alternative way to draw a sample from the Ising model is through this random
cluster model. Namely, we first draw a sample from the Fortuin-Kastelyn model (defined with
appropriate parameters), and for each connected component in the resulting graph, we flip a fair
coin to determine whether all the nodes in the component should be −1 or +1.
With this correspondence in hand, we can apply results for the Fortuin-Kastelyn model – crucial
for our purposes is that the fact that the FK model’s measure is stochastically increasing. Roughly,
this means that if we increase the values of the re ’s, the probability of an edge having a 1 can only
increase. Intuitively, this leads to an increase in E[Xu Xv ] in the Ising model, since it increases the
probability that the nodes are connected in the FK model, and thus the expectation of any edge can
only increase as we increase the ferromagnetic edge parameters. Careful work is needed to carry
through the implications of this correspondence, but it allows us to conclude the nearly-optimal
sample complexity of Õ(m/ε).
Full details are provided in Section 4.2.
1.1.2
A Linear Lower Bound for Testing Ising Models
As a starting point for our lower bound, we use Le Cam’s classical two-point method. This is the
textbook method for proving lower bounds in distribution testing. It involves defining two families
of distributions P and Q, such that every distribution p ∈ P is ε-far from every distribution q ∈ Q.
We consider selecting a uniformly random pair (p, q) ∈ (P, Q) and then drawing k independent
samples from each of p and q. If we can show that the resulting two transcripts of k samples are
close in total variation distance, then k samples are insufficient to distinguish these two cases.
While this method is fairly well-understood in the univariate setting, it proves more difficult
to apply in some multivariate settings. This difficulty arises in the definition of the set Q5 . In
the univariate setting, we often decompose the domain into several disjoint sets, and define Q by
applying perturbations to each of these sets independently. This style of construction allows us
to analyze each subset locally and compose the results. In the multivariate setting, constructions
of this local nature are still possible and are not too hard to analyze – see Theorem 18. In this
construction, we consider an Ising model defined by taking a fixed perfect matching on the graph
and selecting a distribution from Q by applying a random sign vector to the edge potentials of this
√
matching. This allows us to prove an Ω( n) lower bound on the complexity of uniformity testing.
However, such local constructions prove to be limited in the multivariate setting. In order to
prove stronger lower bounds, we instead must consider an Ising model generated by taking a random
5
We note that for simplicity, P is often chosen to be a singleton.
9
perfect matching on the graph. This construction is more global in nature, since the presence of
an edge gives us information about the presence of other edges in the graph. As a result, the
calculations no longer decompose elegantly over the (known) edges in the matching. While at a
first glance, the structure of such a construction may seem too complex to analyze, we reduce it to
analyzing the structure of a random pair of matchings by exploiting combinatorial symmetries. An
important step in the proof requires us to understand the random variable representing the number
of edges shared by two random perfect matchings. This analysis allows us to prove a quadraticallybetter lower bound of Ω(n). We believe our analysis may be useful in proving lower bounds for such
global constructions in other multivariate settings.
Full details are provided in Section 8.
1.2
Organization
In Section 2, we discuss preliminaries and the notation that we use throughout the paper. In Section
3, we give a simple localization-based algorithm for independence testing and its corresponding
variant for goodness-of-fit testing. In Section 4, we present improvements to our localization-based
algorithms for forest-structured and ferromagnetic Ising models. In Section 5, we describe our main
algorithm for the high-temperature regime which uses a global statistic on the Ising model. In
Section 6, we compare our algorithms from Sections 3 and 5. In Section 7, we discuss the bounds
in [LPW09] and apply them to bounding the variance of bilinear statistics over the Ising model. In
Section 8, we describe our lower bounds.
2
Preliminaries
Recall the definition of the Ising model from Eq. (1). We will abuse notation, referring to both the
probability distribution p and the random vector X that it samples in {±1}V as the Ising model.
That is, X ∼ p. We will use Xu to denote the variable corresponding to node u in the Ising model
X. When considering multiple samples from an Ising model X, we will use X (l) to denote the lth
sample. We will use h to denote the largest node parameter in absolute value and β to denote the
largest edge parameter in absolute value. That is, |θv | ≤ h for all v ∈ V and |θe | ≤ β for all e ∈ E.
Depending on the setting, our results will depend on h and β. Furthermore, in this paper we will
use the convention that E = {(u, v) | u, v ∈ V, u 6= v} and θe may be equal to 0, indicating that edge
e is not present in the graph. We use m to denote the number of edges with non-zero parameters
in the graph, and dmax to denote the maximum degree of a node.
Throughout this paper, we will use the notation µv , E[Xv ] for the marginal expectation of a
node v ∈ V (also called node marginal), and similarly µuv , E[Xu Xv ] for the marginal expectation
of an edge e = (u, v) ∈ E (also called edge marginal). In case a context includes multiple Ising
models, we will use µpe to refer to the marginal expectation of an edge e under the model p.
We will use Un to denote the uniform distribution over {±1}n , which also corresponds to the
Ising model with θ~ = ~0. Similarly, we use In for the set of all product distributions over {±1}n .
In this paper, we will consider Rademacher random variables, where Rademacher(p) takes value
1 with probability p, and −1 otherwise.
When p~ and ~q are vectors, we will write p~ ≤ ~q to mean that pi ≤ qi for all i.
Definition 1. In the setting with no external field, θv = 0 for all v ∈ V .
Definition 2. In the ferromagnetic setting, θe ≥ 0 for all e ∈ E.
10
Definition 3 (Dobrushin’s Uniqueness Condition). Consider an Ising
P model p defined on a graph
~
G = (V, E) with |V | = n and parameter vector θ. Suppose maxv∈V u6=v tanh (|θuv |) ≤ 1 − η for
some constant η > 0. Then p is said to satisfy Dobrushin’s uniqueness condition, or be in the high
temperature regime. Note that since tanh(|x|) ≤ |x| for all x, the above condition follows from more
simplified conditions which avoid having to deal with hyperbolic functions. For instance, either of
the following two conditions:
X
max
|θuv | ≤ 1 − η or
v∈V
u6=v
βdmax ≤ 1 − η
are sufficient to imply Dobrushin’s condition (where β = maxu,v |θuv | and dmax is the maximum
degree of G).
In general, when one refers to the temperature of an Ising model, a high temperature corresponds
to small θe values, and a low temperature corresponds to large θe values. In this paper, we will only
use the precise definition as given in Definition 3.
Remark 1. We note that high-temperature is not strictly needed for our results to hold – we only need
Hamming contraction of the “greedy coupling.” This condition implies rapid mixing of the Glauber
dynamics (in O(n log n) steps) via path coupling (Theorem 15.1 of [LPW09]). See [DDK17, GLP17]
for further discussion of this weaker condition.
Lipschitz functions of the Ising model have the following variance bound, which is in Chatterjee’s
thesis [Cha05]:
Lemma 1 (Lipschitz Concentration Lemma). Suppose that f (X1 , . . . , Xn ) is a function of an Ising
model in the high-temperature regime. Suppose the Lipschitz constants of f are l1 , l2 , . . . , ln respectively. That is,
f (X1 , . . . , Xi , . . . , Xn ) − f (X1 , . . . , Xi0 , . . . , Xn ) ≤ li
for all values of X1 , . . . , Xi−1 , Xi+1 , . . . , Xn and for any Xi and Xi0 . Then for some absolute constant
c1 ,
c1 t2
Pr [|f (X) − E[f (X)]| > t] ≤ 2 exp − Pn 2 .
2 i=1 li
In particular, for some absolute constant c2 ,
Var(f (X)) ≤ c2
X
li2 .
i
We will use the symmetric KL divergence, defined as follows:
p
q
dSKL (p, q) = dKL (p, q) + dKL (q, p) = Ep log
+ Eq log
.
q
p
We will use without proof the following well-known result regarding relations between distance
measures on probability distributions.
Lemma 2 (Pinsker’s Inequality). For any two distributions p and q, we have the following relation
between their total variation distance and their KL-divergence,
2d2TV (p, q) ≤ dKL (p||q).
11
Also since dKL (p||q) ≥ 0 for any distributions P and Q, we have
dSKL (p, q) ≥ dKL (p||q) ≥ 2d2TV (p, q).
(2)
Hence the symmetric KL-divergence between two distributions upper bounds both the KL-divergence
and total variation (TV) distance between them under appropriate scaling. Therefore, our results
which hold for testing with respect to the SKL-divergence also hold for testing with respect to
KL-divergence and TV distance.
We will use the following folklore result on estimating the parameter of a Rademacher random
variable.
Lemma 3. Given iid random variables X1 , . . . , Xk ∼ Rademacher(p) for k = O(log(1/δ)/ε2 ),
there exists an algorithm which obtains an estimate p̂ such that |p̂ − p| ≤ ε with probability 1 − δ.
In Section 7 we use the Glauber dynamics on the Ising model. Glauber dynamics is the canonical
Markov chain for sampling from an Ising model. We consider the basic variant known as single-site
Glauber dynamics. The dynamics are a Markov chain defined on the set Σn where Σ = {±1}. They
proceed as follows:
1. Start at any state X (0) ∈ Σn . Let X (t) denote the state of the dynamics at time t.
2. Let N (u) denote the set of neighbors of node u. Pick a node u uniformly at random and
update X as follows:
P
(t)
exp θu + v∈N (u) θuv Xv
Xu(t+1) = 1 w.p.
P
P
(t)
(t)
exp θu + v∈N (u) θuv Xv + exp −θu − v∈N (u) θuv Xv
P
(t)
exp −θu − v∈N (u) θuv Xv
Xu(t+1) = −1 w.p.
P
P
(t)
(t)
exp θu + v∈N (u) θuv Xv + exp −θu − v∈N (u) θuv Xv
Xv(t+1) = Xv(t)
∀ v 6= u.
Glauber dynamics define a reversible, ergodic Markov chain whose stationary distribution is
identical to the corresponding Ising model. In many relevant settings, such as, for instance, the
high-temperature regime, the dynamics are fast mixing, i.e., they mix in time O(n log n) and hence
offer an efficient way to sample from Ising models.
2.1
Input to Goodness-of-Fit Testing Algorithms
To solve the goodness-of-fit testing or identity testing problem with respect to a discrete distribution
q, a description of q is given as part of the input along with sample access to the distribution p
which we are testing. In case q is an Ising model, its support has exponential size and specifying the
vector of probability values at each point in its support is inefficient. Since q is characterized by the
edge parameters between every pair of nodes and the node parameters associated with the nodes, a
succinct description would be to specify the parameters vectors {θuv }, {θu }. In many cases, we are
also interested in knowing the edge and node marginals of the model. Although these quantities
can be computed from the parameter vectors, there is no efficient method known to compute the
marginals exactly for general regimes. A common approach is to use MCMC sampling to generate
samples from the Ising model. However, for this technique to be efficient we require that the mixing
12
time of the Markov chain be small which is not true in general. Estimating and exact computation of
the marginals of an Ising model is a well-studied problem but is not the focus of this paper. Hence,
to avoid such computational complications we will assume that for the identity testing problem the
description of the Ising model q includes both the parameter vectors {θuv }, {θu } as well as the edge
and node marginal vectors {µuv = E[Xu Xv ]}, {µu = E[Xu ]}.
2.2
Symmetric KL Divergence Between Two Ising Models
We note that the symmetric KL divergence between two Ising models p and q admits a very convenient expression [SW12]:
X
X
(θep − θeq ) (µpe − µqe ) .
(3)
dSKL (p, q) =
(θvp − θvq ) (µpv − µqv ) +
v∈V
e=(u,v)∈E
This expression will form the basis for all our algorithms.
3
A General Purpose Localization Algorithm
Our first algorithm is a general purpose “localization” algorithm. While extremely simple, this
serves as a proof-of-concept that testing on Ising models can avoid the curse of dimensionality,
while simultaneously giving a very efficient algorithm for certain parameter regimes. The main
observation which enables us to do a localization based approach is stated in the following Lemma,
which allows us to “blame” a difference between models p and q on a discrepant node or edge.
Lemma 4. Given two Ising models p and q, if dSKL (p, q) ≥ ε, then either
q
p
) (µpuv − µquv ) ≥
− θuv
• There exists an edge e = (u, v) such that (θuv
• There exists a node u such that (θup − θuq ) (µpu − µqu ) ≥
ε
2m ;
or
ε
2n .
Proof of Lemma 4: We have,
dSKL (p, q) =
X
(θep − θeq ) (µpe − µqe ) +
=⇒
(θvp − θvq ) (µpv − µqv ) ≥ ε
v∈V
e=(u,v)∈E
X
X
(θep − θeq ) (µpe − µqe ) ≥ ε/2 or
X
(θvp − θvq ) (µpv − µqv ) ≥ ε/2
v∈V
e=(u,v)∈E
p
q
ε
In the first case, there has to exist an edge e = (u, v) such that (θuv
− θuv
) (µpuv − µquv ) ≥ 2m
and
p
q
p
q
ε
in the second case there has to exist a node u such that (θu − θu ) (µu − µu ) ≥ 2n thereby proving
the lemma.
Before giving a description of the localization algorithm, we state its guarantees.
2 2
Theorem 2. Given Õ mε2β samples from an Ising model p, there exists a polynomial-time algorithm which distinguishes
the
2between
cases p ∈ In and dSKL (p, In ) ≥ ε with probability at least 2/3.
m β2
n2 h2
Furthermore, given Õ ε2 + ε2
samples from an Ising model p and a description of an Ising
model q, there exists a polynomial-time algorithm which distinguishes between the cases p = q and
dSKL (p, q) ≥ ε with probability at least 2/3 where β = max{|θuv |} and h = max{|θu |}. The above
algorithms assume that m, an upper bound on the
number of edges, is known. If no upper bound is
known, we may use the trivial upper bound of n2 . If we are given as input the maximum degree of
nodes in the graph dmax , m in the above bounds is substituted by ndmax .
13
Note that the sample complexity achieved by the localization algorithm gets worse as the graph
becomes denser. This is because as the number of possible edges in the graph grows, the contribution
to the distance by any single edge grows smaller thereby making it harder to detect.
We describe the algorithm for independence testing in Section 3.1. The algorithm for testing identity
is similar, its description and correctness proofs are given in Section 3.2.
3.1
Independence Test using Localization
We start with a high-level description of the algorithm. Given sample access to Ising model X ∼ p it
will first obtain empirical estimates of the node marginals µu for each node u ∈ V and edge marginals
µuv for each pair of nodes (u, v). Denote these empirical estimates by µ̂u and µ̂uv respectively.
Using these empirical estimates, the algorithm computes the empirical estimate for the covariance
of each pair of variables in the Ising model. That is, it computes an empirical estimate of λuv =
E[Xu Xv ] − E[Xu ]E[Xv ] for all pairs (u, v). If they are all close to zero, then we can conclude that
p ∈ In . If there exists an edge for which λuv is far from 0, this indicates that p is far from In .
The reason for this follows from the expression Lemma 4 and is described in further detail in the
proof of Lemma 6. A precise description of the test is given in in Algorithm 1 and its correctness
is proven via Lemmas 5 and 6. We note that this algorithm is phrased as if an upper bound on the
number of edges m is known. If we instead know an upper bound on the maximum degree dmax ,
then we can replace m by ndmax .
Algorithm 1 Test if an Ising model p is product
1: function LocalizationTest(sample
access to Ising model p, accuracy parameter ε, β, m)
2 2
m β log n
samples from p. Denote the samples by X (1) , . . . , X (k)
2:
Draw k = O
ε2
P (i)
P (i) (i)
3:
Compute empirical estimates µ̂u = k1 i Xu for each node u ∈ V and µ̂uv = k1 i Xu Xv
for each pair of nodes (u, v)
4:
Using the above estimates compute the covariance estimates λ̂uv = µ̂uv − µ̂u µ̂v for each pair
of nodes (u, v)
If for any pair of nodes (u, v), λ̂uv ≥
6:
Otherwise, return that p ∈ In
7: end function
5:
ε
4mβ
return that dSKL (p, In ) ≥ ε
To prove correctness of Algorithm 1, we will require the following lemma, which allows us to
detect pairs u, v for which λuv is far from 0.
n
Lemma 5. Given O log
samples from an Ising model X ∼ p, there exists a polynomial-time
ε2
algorithm which, with probability at least 9/10, can identify all pairs of nodes (u, v) ∈ V 2 such that
|λuv | ≥ ε, where λuv = E[Xu Xv ] − E[Xu ]E[Xv ]. Namely, the algorithm computes the empirical
value of |λuv | for each pair of nodes and identifies pairs such that this value is sufficiently far from
0.
Proof. This lemma is a direct consequence of Lemma 3. Note that for any edge e = (u, v) ∈ E,
Xu Xv ∼ Rademacher((1+µe )/2). Also Xu ∼ Rademacher((1+µu )/2) and Xv ∼ Rademacher((1+
µv )/2). We will use Lemma 3 to show that O(log n/ε2 ) samples suffice to detect whether λe = 0
or |λe | ≥ ε with probability at least 1 − 1/10n2 . With O(log n/ε2 ) samples, Lemma 3 implies we
ε
can obtain estimates µ̂uv , µ̂u and µ̂v for µuv , µu and µv respectively such that |µ̂uv − µuv | ≤ 10
,
ε
ε
2
|µ̂u − µu | ≤ 10 and |µ̂v − µv | ≤ 10 with probability at least 1 − 1/10n . Let λ̂uv = µ̂uv − µ̂u µ̂v . Then
14
2
ε
3ε
+ 100
. It can be seen that in
from the above, it follows by triangle inequality that |λuv − λ̂uv | ≤ 10
the case when the latter term in the previous inequality dominates the first, ε is large enough that
ε2
ε
O(log n) samples suffice to distinguish the two cases. In the more interesting case, 100
≤ 10
, and
|λuv − λ̂uv | ≤
4ε
10 .
Therefore if |λuv | ≥ ε, then λ̂uv ≥
6ε
10 ,
and if |λuv | = 0, then λ̂uv ≤
4ε
10
thereby
implying that with probability at least 1 − 1/10n2 we can detect whether λuv = 0 or |λuv | ≥ ε.
Taking a union bound over all edges, the probability that we correctly identify all such edges is at
least 9/10.
With this lemma in hand, we now prove the first part of Theorem 2.
2 2
Lemma 6. Given Õ mε2β samples from an Ising model X ∼ p, Algorithm 1 distinguishes between
the cases p ∈ In and dSKL (p, In ) ≥ ε with probability at least 2/3.
Proof. We will run Algorithm 1 on all pairs Xu , Xv to identify any pair such that |λuv | is large.
This will involve using the algorithm of Lemma 5 with parameter “ε” as ε/2βm. If no such pair is
identified, output that p ∈ In , and otherwise, output that dSKL (p, In ) ≥ ε. If p ∈ In , we know that
E[Xu Xv ] = E[Xu ]E[Xv ] for all edges (u, v), and therefore, with probability 9/10, there will be no
ε
edges for which the empirical estimate of |λe | ≥ 2βm
. On the other hand, if dSKL (p, In ) ≥ ε, then
dSKL (p, q) ≥ ε for every q ∈ In . In particular, consider the product distribution q on n nodes such
that µqu = µpu for all u ∈ V . For this particular product distribution q, by (3), there must exist some
ε
e∗ such that |λe∗ | ≥ 2βm
, and the algorithm will identify this edge. This is because
X
(θvp − θvq ) (µpv − µqv ) = 0
(4)
v∈V
∴ dSKL (p, q) ≥ ε
ε
=⇒ ∃e∗ = (u, v) s.t (θep − θeq ) (µpe − µqe ) ≥
m
ε
=⇒ ∃e∗ = (u, v) s.t |(µpe − µqe )| ≥
2βm
ε
=⇒ ∃e∗ = (u, v) s.t |λe∗ | ≥
.
2βm
(5)
(6)
where (4) follows because µpv = µqv for all v ∈ V , (5) follows from Lemma 4 and (6) follows because
|θep − θeq | ≤ 2β. This completes the proof of the first part of Theorem 2.
3.2
Identity Test using Localization
If one wishes to test for identity of p to an Ising model q, the quantities whose absolute values indicate
that p is far from q are µpuv − µquv for all pairs u, v, and µpu − µqu for all u, instead of λuv . Since µquv
and µqu are given as part of the description of q, we only have to identify whether E[Xu Xv ] ≥ c and
E[Xu ] ≥ c for any constant c ∈ [−1, 1]. A variant of Lemma 5 as stated in Lemma 7 achieves this
goal. Algorithm 2 describes the localization based identity test. Its correctness proof will imply the
second part of Theorem 2 and is similar in vein to that of Algorithm 1. It is omitted here.
n
Lemma 7. Given O log
samples from an Ising model p, there exists a polynomial-time algorithm
2
ε
which, with probability at least 9/10, can identify all pairs of nodes (u, v) ∈ V 2 such that |µpuv −c|
≥ ε
n
for any constant c ∈ [−1, 1]. There exists a similar algorithm, with sample complexity O log
ε2
which instead identifies all v ∈ V such that |µpv − c| ≥ ε, where µpv = E[Xv ] for any constant
c ∈ [−1, 1].
15
Proof of Lemma 7: The proof follows along the same lines as Lemma 5. Let X ∼ p. Then, for
any pair of nodes (u, v), Xu Xv ∼ Rademacher((1 + µpe )/2). Also Xu ∼ Rademacher((1 + µpu )/2)
for any node u. For any pair of nodes u, v, with O(log n/ε2 ) samples, Lemma 3 implies we that
ε
the empirical estimate µ̂puv is such that |µ̂puv − µpuv | ≤ 10
with probability at least 1 − 1/10n2 . By
p
p
ε
ε
triangle inequality, we get |µuv − c| − 10
≤ |µ̂uv − c| ≤ |µpuv − c| + 10
. Therefore if |µpuv − c| = 0,
ε
9ε
then |µ̂puv − c| ≤ 10
w.p. ≥ 1 − 1/10n2 and if |µpuv − c| ≥ ε, then |µ̂puv − c| ≥ 10
w.p. ≥ 1 − 1/10n2 .
p
Hence by comparing whether |µ̂uv − c| to ε/2 we can distinguish between the cases |µpuv − c| = 0
and |µpuv − c| ≥ ε w.p. ≥ 1 − 1/10n2 . Taking a union bound over all edges, the probability that
we correctly identify all such edges is at least 9/10. The second statement of the Lemma about the
nodes follows similarly.
Algorithm 2 Test if an Ising model p is identical to q
1: function LocalizationTestIdentity(sample access to Ising model X ∼ p, description of
Ising model q, accuracy
parameterε,β,h,m)
(m2 β 2 +n2 h2 ) log n
2:
Draw k = O
samples from p. Denote the samples by X (1) , . . . , X (k)
ε2
P (i)
P (i) (i)
3:
Compute empirical estimates µ̂pu = k1 i Xu for each node u ∈ V and µ̂puv = k1 i Xu Xv
for each pair of nodes (u, v)
ε
4:
If for any pair of nodes (u, v), |µ̂puv − µquv | ≥ 8mβ
return that dSKL (p, q) ≥ ε
q
p
ε
5:
If for any node u, if |µ̂u − µu | ≥ 8nh return that dSKL (p, q) ≥ ε
6:
Otherwise, return that p = q
7: end function
The proof of correctness of Algorithm 2 follows along the same lines as that of Algorithm 1 and
uses Lemma 7. We omit the proof here.
4
Improved Tests for Forests and Ferromagnetic Ising Models
In this section we will describe testing algorithms for two commonly studied classes of Ising models,
namely forests and ferromagnets. In these cases, the sample complexity improves compared to the
baseline result when in the regime of no external field. The testers are still localization based (like
those of Section 3), but we can now leverage structural properties to obtain more efficient testers.
First, we consider the class of all forest structured Ising models, where the underlying graph
G = (V, E) is a forest. Such models exhibit nice structural properties which can be exploited to
obtain more efficient tests. In particular, under no external field, the edge marginals µe , which,
in general are hard to compute, have a simple closed form expression. This structural information
enables us to improve our testing algorithms from Section 3 on forest graphs. We state the improved
sample complexities here and defer a detailed description of the algorithms to Section 4.1.
Theorem 3 (Independence testing of Forest-Structured Ising Models). Algorithm 3 takes in Õ nε
samples from an Ising model X ∼ p whose underlying graph is a forest and which is under no
external field and outputs whether p ∈ In or dSKL (p, In ) ≥ ε with probability ≥ 9/10.
Remark 2. Note that Theorem 3 together with our lower bound described in Theorem 19 indicate a
tight sample complexity up to logarithmic factors for independence testing on forest-structured Ising
models under no external field.
16
Theorem 4 (Identity Testing of Forest-Structured Ising Models). Algorithm 4 takes in the edge
parameters
of an Ising model q on a forest graph and under no external field as input, and draws
Õ c(β) nε samples from an Ising model X ∼ p (where c(β) is a function of the parameter β) whose
underlying graph is a forest and under no external field, and outputs whether p = q or dSKL (p, q) ≥ ε
with probability ≥ 9/10.
Note that for identity testing, any algorithm necessarily has to have at least a β dependence
due to the lower bound we show in Theorem 20.
The second class of Ising models we consider this section are ferromagnets. For a ferromagnetic
Ising model, θuv ≥ 0 for every pair of nodes u, v. Ferromagnets may potentially contain cycles but
since all interactions are ferromagnetic, the marginal of every edge is at least what it would have
been if it was a solo edge. This intuitive property turns out to be surprisingly difficult to prove in a
direct way. We prove this structural property using an alternative view of the Ising model density
which comes from the Fortuin-Kasteleyn random cluster model. Using this structural property, we
give a quadratic improvement in the dependence on parameter m for testing independence under no
external field. We state our main result in this regime here and a full description of the algorithm
and the structural lemma are provided in Section 4.2.
Theorem 5 (Independence Testing of Ferromagnetic Ising Models). Algorithm 5 takes in Õ ndmax
ε
samples from a ferromagnetic Ising model X ∼ p which is under no external field and outputs whether
p ∈ In or dSKL (p, In ) ≥ ε with probability ≥ 9/10.
4.1
Improved Algorithms for Independence and Identity Testing on Forests
Before we present the improved algorithms, we will prove the following fact about the edge marginals
of an arbitrary Ising model with no external field where the underlying graph is a forest. This result
was known prior to this work by the community but we couldn’t find a proof of the same, hence we
provide our own proof of the lemma.
Lemma 8 (Structural Lemma for Forest-Structured Ising Models). If p is an Ising model on a
forest graph with no external field, and X ∼ p, then for any (u, v) ∈ E, we have
E [Xu Xv ] = tanh(θuv ).
Proof. Consider any edge e = (u, v) ∈ E. Consider the tree (T, ET ) which contains e. Let nT be the
number of nodes in the tree. We partition the vertex set T into U and V as follows. Remove edge
e from the graph and let U denote all the vertices which lie in the connected component of node u
except u itself. Similarly, let V denote all the vertices which lie in the connected component of node
v except node v itself. Hence, T = U ∪ V ∪ {u} ∪ {v}. Let XU be the vector random variable which
denotes the assignment of values in {±1}|U | to the nodes in U . XV is defined similarly. We will
also denote a specific value assignment to a set of nodes S by xS and −xS denotes the assignment
which corresponds to multiplying each coordinate of xS by −1. Now we state the following claim
which follows from the tree structure of the Ising model.
Claim 1. Pr [XU = xU , Xu = 1, Xv = 1, XV = xV ] = exp(2θuv ) Pr [XU = xU , Xu = 1, Xv = −1, XV = −xV ].
In particular the above claim implies the following corollary which is obtained by marginalization
of the probability to nodes u and v.
Corollary 1. If X is an Ising model on a forest graph G = (V, E) with no external field, then for
any edge e = (u, v) ∈ E, Pr [Xu = 1, Xv = 1] = exp(2θuv ) Pr [Xu = 1, Xv = −1].
17
Now,
E [Xu Xv ] = Pr [Xu Xv = 1] − Pr [Xu Xv = −1]
= 2P r [Xu = 1, Xv = 1] − 2 Pr [Xu = 1, Xv = −1]
2P r [Xu = 1, Xv = 1] − 2 Pr [Xu = 1, Xv = −1]
=
2P r [Xu = 1, Xv = 1] + 2 Pr [Xu = 1, Xv = −1]
P r [Xu = 1, Xv = 1] − Pr [Xu = 1, Xv = −1]
=
P r [Xu = 1, Xv = 1] + Pr [Xu = 1, Xv = −1]
exp(2θuv ) − 1 Pr [Xu = 1, Xv = −1]
=
exp(2θuv ) + 1 Pr [Xu = 1, Xv = −1]
= tanh(θuv )
(7)
(8)
(9)
(10)
(11)
(12)
where (8) follows because Pr [Xu = 1, Xv = 1] = Pr [Xu = −1, Xv = −1] and Pr [Xu = −1, Xv = 1] =
Pr [Xu = 1, Xv = −1] by symmetry. Line (9) divides the expression by the total probability which
is 1 and (11) follows from Corollary 1.
Given the above structural lemma, we give the following simple algorithm for testing independence on forest Ising models under no external field.
Algorithm 3 Test if a forest Ising model p under no external field is product
1: function TestForestIsing-Product(sample access to Ising model p)
p
2:
Run the algorithm of Lemma 5 to identify all edges e = (u, v) such that |E[Xu Xv ]| ≥ nε .
using Õ nε samples. If it identifies any edges, return that dSKL (p, In ) ≥ ε
3:
Otherwise, return that p is product.
4: end function
Algorithm 3, at a high level, works as follows. If there is an edge parameter whose absolute
value is larger than a certain threshold, it will be easy to detect due to the structural information
about the edge marginals. In case all edges have parameters smaller in absolute value than this
threshold, the expression for dSKL (., .) between two Ising models tells us that there still has to be at
least one edge with a significantly large value of µe in case the model is far from uniform, and hence
will still be detectable by the algorithm of Lemma 5. The proof of Theorem 3 shows this formally.
Proof of Theorem 3: Firstly, note that under no external field, the only product Ising model is the
uniform distribution Un . Therefore the problem reduces to testing whether p is uniform or not.
Consider the case when p is indeed uniform. That is, there are no edges in the underlying graph of
the Ising model. In this case with probability at least 9/10 the localization algorithm of Lemma 5
will output no edges. Hence Algorithm 3 will output that p is uniform.
In case dSKL (p, Un ) ≥ ε, we split the analysis into two cases.
pε
• Case 1: There exists an edge e = (u, v) such that |θuv | ≥
n . In this case, E[Xu Xv ] =
tanh(θuv ) and in the regime
p ε where |θ| = o(1), |tanh(θ)| ≥ |θ/2|. Hence implying that
|E[Xu Xv ]| ≥ |θuv /2| ≥
n /2 . Therefore the localization algorithm of Lemma 5 would
identify such an edge with probability at least 9/10. Note that the regime where the inequality |tanh(θ)| ≥ |θ/2| isn’t valid is easily detectable using Õ( nε ) samples, as this would imply
that |θ| ≥ 1.9 and |E[Xu Xv ]| ≥ 0.95.
18
• Case 2: All edges e = (u, v) are such that |θuv | ≤
pε
n
. In this case we have,
dSKL (p, Un ) ≥ ε
ε
=⇒ ∃ edge e = (u, v) s.t θuv E[Xu Xv ] ≥
n
r
n
ε
=⇒ ∃ edge e = (u, v) s.t |E[Xu Xv ]| ≥
×
n
ε
r
ε
=
n
(13)
(14)
(15)
(16)
Hence, the localization algorithm of Lemma 5 would identify such an edge with probability at
least 9/10.
Next, we will present an algorithm for identity testing on forest Ising models under no external
field.
Algorithm 4 Test if a forest Ising model p under no external field is identical to a given Ising
model q
1: function TestForestIsing-Identity(Ising model q,sample access to Ising model p)
2:
If the Ising model q is not a forest, or has a non-zero external field on some node, return.
dSKL (p, q) ≥ ε
3:
Run the algorithm of p
Lemma 5 to identify all edges e = (u, v) such that.
q
ε
n
|E[Xu Xv ] − tanh(θuv )| ≥
n using Õ ε samples. If it identifies any edges, return that
dSKL (p, q) ≥ ε
4:
Otherwise, return that p = q.
5: end function
Proof of Theorem 4: Consider the case when p is indeed q. In this case with probability at least
9/10 the localization algorithm of Lemma 5 will output no edges. Hence Algorithm 4 will output
that p is uniform.
In case dSKL (p, q) ≥ ε, we split the analysis into two cases.
p
p
q
• Case 1: There exists an edge e = (u, v) such that |θuv
− θuv
| ≥ nε . In this case, E[Xu Xv ] −
p
q
p
q
p
q
µquv = tanh(θuv
) − tanh(θuv
) and hence has the same sign as θuv
− θuv
. Assume that θuv
≥ θuv
.
q
p
p
q
The argument for the case θuv
> θuv
will follow similarly. If θuv
− θuv
≤ 1/2 tanh(β), then
the following inequality holds from Taylor’s theorem.
p
q
sech2 (β) (θuv
− θuv
)
2
p
2
p
q
which would imply tanh(θuv
) − tanh(θuv
) ≥ sech2 (β) nε and hence the localization algorithm
of Lemma 5 would identify edge e with probability at least 9/10 using Õ c1 (β)n
samples
ε
p
q
tanh(θuv
) − tanh(θuv
)≥
p
q
p
q
4
(where
c1 (β) = cosh
(β)). If θuv − θuv > 1/2 tanh(β), then tanh(θuv ) − tanh(θuv ) ≥ tanh(β) −
1
tanh β − 2 tanh(β)
and hence the localization algorithm of Lemma 5 would identify edge e
with probability at least 9/10 using Õ (c2 (β)) samples where c2 (β) =
19
1
.
(tanh(β)−tanh(β−1/2 tanh(β)))2
Note that as β grows small, c2 (β) gets worse. However it cannot grow unbounded as we also
p
q
have to satisfy the constraint that θuv
− θuv
≤ 2β. This implies that
1
2
c2 (β) = min β ,
(tanh(β) − tanh(β − 1/2 tanh(β)))2
samples suffice in this case. Therefore
the algorithm will give the correct output with proba
n
bility > 9/10 using Õ c(β) ε samples where c(β) = max{c1 (β), c2 (β)}.
p
q
q
• Case 2: All edges e = (u, v) are such that |θuv
− θuv
| ≤ nε . In this case we have,
dSKL (p, q) ≥ ε
ε
p
q
=⇒ ∃ edge e = (u, v) s.t (θuv
− θuv
) (E[Xu Xv ] − µquv ) ≥
n
r
n
ε
q
=⇒ ∃ edge e = (u, v) s.t |E[Xu Xv ] − µuv | ≥
×
n
ε
r
ε
=
n
(17)
(18)
(19)
(20)
Hence, the localization algorithm of Lemma 5 would identify such an edge with probability at
least 9/10.
4.2
Ferromagnetic Ising Models: A Structural Understanding and an Improved
Independence Test
In this section we will describe an algorithm for testing independence of ferromagnetic Ising models
under no external field. The tester follows the localization based recipe of Section 3 but leverages
additional structural information about ferromagnets to obtain an improved sample complexity.
At a high level, the algorithm is as follows: if there exists an edge with a large edge parameter,
then we lower bound its marginal by tanh(θuv ) where uv is the edge under consideration. This
implies that its marginal sticks out and is easy to catch via performing local tests on all edges. If
all the edge parameters were small, then Algorithm 1 is already efficient.
We first prove a structural lemma about ferromagnetic Ising models. We will use the FortuinKasteleyn random cluster model and its coupling with the Ising model (described in Chapter 10 of
[RAS15]) to argue that in any ferromagnetic Ising model µuv ≥ tanh(θuv ) for all pairs u, v.
4.2.1
Random Cluster Model
Let G = (V, E) be a finite graph. The random cluster measure is a probability distribution on the
space Ω = {0, 1}E of bond configurations denoted by η = (η(e))e∈E ∈ {0, 1}E . Each edge has an
associated bond η(e). η(e) = 1 denotes that bond e is open or present and η(e) = 0 implies that
bond e is closed or unavailable. A random cluster measure is parameterized by an edge probability
0 < r < 1 and by a second parameter 0 < s < ∞. Let k(η) denote the number of connected
components in the graph (V, η). The random cluster measure is defined by
!
Y
1
ρr,s (η) =
rη(e) (1 − r)1−η(e) sk(η)
Zr,s
e∈E
20
where Zr,s is a normalizing factor to make ρ a probability density. We consider a generalization of
the random cluster model where each edge is allowed to have its own parameter 0 < re < 1. Under
this generalization, the measure becomes
!
Y
1
η(e)
1−η(e)
re (1 − re )
sk(η) .
(21)
ρ~r,s (η) =
Z~r,s
e∈E
The random cluster measure is stochastically increasing in ~r when s ≥ 1. This property is formally
stated in Lemma 10.3 of [RAS15]. We state a generalized version of the Lemma here which holds
when each edge is allowed its own probability parameter re .
Lemma 9. [Lemma 10.3 from [RAS15]] For s ≥ 1, and r~1 ≤ r~2 coordinate-wise, ρr~1 ,s ≤ ρr~2 ,s where
given two bond configurations η1 and η2 , η1 ≥ η2 iff η1 (e) = 1 for all e such that η2 (e) = 1.
4.2.2
Coupling between the Random Cluster Model and the Ising model
We will now describe a coupling between the random cluster measure and the probability density
function for a ferromagnetic Ising model. In particular, the edge probability re under the random
cluster measure and the edge parameters θe of the Ising model are related by
re = 1 − exp(−2θe )
and the parameter s = 2 because the Ising model has two spins ±1. The coupling Q will be a
joint distribution on the spin variables X = (X1 . . . Xn ) of the Ising model and the bond variables
η = (η(e))e∈E . The measure Q is defined as
Q(X, η) =
1
Z
Y
reη(e) (1 − re )1−η(e) (1Xu =Xv + (1 − η(e))1Xu 6=Xv )
e=(u,v)∈E
where Z is a normalizing constant so as to make Q a probability measure. Under the relation stated
above between re and θe , the following properties regarding the marginal distributions of Q hold.
X
X
1
Q(X, η) = 0 exp
θuv Xu Xv
Z
u6=v
η∈{0,1}E
!
X
Y
1
Q(X, η) = 00
reη(e) (1 − re )1−η(e) 2k(η) = ρ~r,2 (η)
Z
n
X∈{±1}
e∈E
(22)
where Z 0 , Z 00 are normalizing constants to make the marginals probability densities. The above
equations imply that the measure Q is a valid coupling and more importantly they yield an alternative way to sample from the Ising model as follows:
First sample a bond configuration η according to ρ~r,2 (η). For each connected component in the
bond graph, flip a fair coin to determine if the variables in that component will be all +1 or all −1.
In addition to the above information about the marginals of Q, we will need the following simple
observations.
21
1. Q(X, η) = 0 if η(e) = 1 for any e ∈
/ E.
2. Q(X, η) = 0 if for any e = (u, v) ∈ E, η(e) = 1 and Xu 6= Xv .
Next we state another property of the coupling Q(., .) which says that if two nodes u and v
are in different connected components in the bond graph specified by η, then the probability that
Xu = Xv is the same as the probability that Xu 6= Xv .
Claim 2. Let Cη (u, v) denote the predicate that under the bond configuration η, u and v are connected with a path of open bonds. Then,
X
X
X
X
Q(X, η)
Q(X, η) =
η s.t
X s.t.
Cη (u,v)=0 Xu 6=Xv
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
The proof of the above claim is quite simple and follows by matching the appropriate terms in
the probability density Q when u and v lie in different connected components. The proof is omitted
here.
Armed with the coupling Q and its properties stated above, we are now ready to state the main
structural lemma we show for ferromagnetic Ising models.
Lemma 10. Consider two ferromagnetic Ising models p and q under no external field defined on
Gp = (V, Ep ) and Gq = (V, Eq ). Denote the parameter vector of p model by θ~p and that of q model
by θ~q . If θ~p ≥ θ~q coordinate-wise, then for any two nodes u, v ∈ V , µpuv ≥ µquv .
Proof. Since
µpuv = Pr [Xu = Xv ] − Pr [Xu 6= Xv ]
p
=⇒
µpuv
p
= 2 Pr [Xu = Xv ] − 1
p
to show that µpuv ≥ µquv it suffices to show that Prp [Xu = Xv ] ≥ Prq [Xu = Xv ]. Consider the coupling Q(X, η) described above between the random cluster measure and the Ising model probability.
Prp [Xu = Xv ] can be expressed in terms of Qp (X, η) as follows:
X X
Pr [Xu = Xv ] =
Qp (X, η)
p
X s.t.
Xu =Xv
η
Denote the sum on the right in the above equation by Sp . It suffices to show that Sp ≥ Sq .
Lemma 10.3 of [RAS15] gives that for any bond configuration η0 ,
X
X
b
b
ρE
(η)
≥
ρE
p
q (η).
η≥η0
η≥η0
This follows because the parameter vectors of p and q satisfy the condition of the lemma that
θ~p ≥ θ~q . Again, let Cη (u, v) denote the predicate that under the bond configuration η, u and v are
connected. Let H be the set of all bond configurations such that u and v are connected by a single
distinct path. Therefore Cη0 (u, v) = 1 for all η0 ∈ H. Then the set
C = {η|η ≥ η0 for some η0 ∈ H}
22
represents precisely the bond configurations in which u and v are connected. Applying Lemma 10.3
of [RAS15] on each η0 ∈ H and summing up the inequalities obtained, we get
X
X
b
b
ρE
ρE
q (η)
p (η) ≥
X
=⇒
η s.t
Cη (u,v)=1
η s.t
Cη (u,v)=1
X
X
Qp (X, η) ≥
η s.t
X
Cη (u,v)=1
X
=⇒
X
X
Qq (X, η)
η s.t
X
Cη (u,v)=1
X
Qp (X, η) ≥
X
Qq (X, η)
(23)
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
where the last inequality follows because Q(X, η) = 0 if for any pair u, v, η(uv) = 1 but Xu 6= Xv .
Also, from Claim 2, we have that for any Ising model,
X
X
X
Q(X, η) =
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
X
Q(X, η)
(24)
η s.t
X s.t.
Cη (u,v)=0 Xu 6=Xv
And since Q(., .) is a probability measure we have that for any Ising model,
X
X
X
X
X
X
Q(X, η) +
Q(X, η) +
Q(X, η) +
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
X
=⇒
X
η s.t
X s.t.
Cη (u,v)=0 Xu 6=Xv
X
Q(X, η) +
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
X
X
X
X
Q(X, η) = 1(25)
η s.t
X s.t.
Cη (u,v)=1 Xu 6=Xv
X
Q(X, η) +
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
=⇒
X
Q(X, η) = 1(26)
η s.t
X s.t.
Cη (u,v)=0 Xu 6=Xv
X
Q(X, η) + 2
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
X
X
Q(X, η) = 1(27)
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
where (26) follows because the last term in (25) is 0 and (27) follows from (24).
Equation (27) implies that
X
X
X
X
Sp =
Qp (X, η) +
Qp (X, η)
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
=
1
2
X
X
η s.t
X s.t.
Cη (u,v)=0 Xu =Xv
Qp (X, η) +
η s.t
X s.t.
Cη (u,v)=1 Xu =Xv
1
2
Therefore from (23), we get
Sp ≥ Sq
Using the above lemma, we now prove the main structural lemma for ferromagnets which will
be crucial to our algorithm for testing ferromagnetic Ising models.
Lemma 11 (Structural Lemma about Ferromagnetic Ising Models). If X ∼ p is a ferromagnetic
Ising model on a graph G = (V, E) under zero external field, then µuv ≥ tanh(θuv ) for all edges
(u, v) ∈ E.
23
Proof. Fix the edge of concern e = (u, v). If the graph doesn’t contain cycles, then from Lemma 8
µuv = tanh(θuv ) and the statement is true. To show that the statement holds for general graphs we
will use induction on the structure of the graph. Graph G can be constructed as follows. Start with
the single edge e = (u, v) and then add the remaining edges in E\{e} one by one in some order.
Denote the intermediate graphs obtained during this process as G0 , G1 , . . . , Gm = G where G0 is
the graph consisting of just a single edge. For each graph Gi we can associate the corresponding
Ising model pi to be the model which has θepi = θe for e ∈ EGi and θepi = 0 otherwise. For each
graph Gi in the sequence, we will use µpuvi to denote E [Xu Xv ] for the Ising model corresponding
to graph Gi . We will prove that µpuv ≥ tanh(θuv ) by induction on this sequence of graphs. The
statement can be easily verified to be true for G0 . In fact, µpuv0 = tanh(θuv ). Suppose the statement
p
was true for some Gi in the sequence. By Lemma 10, we have that µuvi+1 ≥ µpuvi . This implies that
Gpi+1
µuv
≥ tanh(θuv ) hence showing the statement to be true for all graphs Gi in the sequence.
Given the above structural lemma about ferromagnetic Ising models under no external field, we
present the following algorithm for testing whether a ferromagnetic Ising model is product or not.
Algorithm 5 Test if a ferromagnetic Ising model p under no external field is product
1: function TestFerroIsing-Independence(sample access to an Ising model p)
√
2:
Run the algorithm
of Lemma 5 to identify if all edges e = (u, v) such that E[Xu Xv ] ≥ ε/n.
2
using Õ nε samples. If it identifies any edges, return that dSKL (p, In ) ≥ ε
3:
Otherwise, return that p is product.
4: end function
Proof of Theorem 5: Firstly, note that under no external field, the only product Ising model is the
uniform distribution Un . To the problem reduces to testing whether p is uniform or not. Consider
the case when p is indeed uniform. That is, there are no edges in the underlying graph of the
Ising model. In this case with probability at least 9/10 the localization algorithm of Lemma 5 with
output no edges. Hence Algorithm 5 will output that p is product.
In case dSKL (p, In ) ≥ ε, we split the analysis into two cases.
p
• Case 1: There exists an edge e = (u, v) such that |θuv | ≥ nε2 . In this case, |E[Xu Xv ]| ≥
|tanh(θuv )| and in the regime where
p ε ε is a fixed constant, |tanh(θ)| ≥ |θ/2|. Hence implying that |E[Xu Xv ]| ≥ |θuv /2| ≥
/2. Therefore the localization algorithm of Lemma 5
n2
would identify such an edge with probability at least 9/10. (The regime where the inequality
2
|tanh(θ)| ≥ |θ/2| isn’t valid would be easily detectable using Õ( nε ) samples.)
p
• Case 2: All edges e = (u, v) are such that θuv ≤ nε2 . In this case we have,
dSKL (X, In ) ≥ ε
ε
=⇒ ∃ edge e = (u, v) s.t θuv E[Xu Xv ] ≥
n2 r
ε
n2
=⇒ ∃ edge e = (u, v) s.t E[Xu Xv ] ≥
×
n2
ε
r
ε
=
n2
(28)
(29)
(30)
(31)
Hence, the localization algorithm of Lemma 5 would identify such an edge with probability at
least 9/10.
24
5
An Improved Test for High-Temperature Ising Models: A Learnthen-Test Algorithm
In this section, we describe a framework for testing Ising models in the high-temperature regime
which results in algorithms which are more efficient than our baseline localization algorithm of
Section 3 for dense graphs. This is the more technically involved part of our paper and we modularize
the description and analysis into different parts. We will give a high level overview of our approach
here.
The main approach we take in this section is to consider a global test statistic over all the
variables on the Ising model in contrast to the localized statistics of Section 3. For ease of exposition,
we first describe the approach for testing independence under no external field. We then describe
the changes that need to be made to obtain tests for independence under an external field and
goodness-of-fit in Section 5.5.
Note that testing independence under no external field boils down to testing uniformity as the
only independent Ising model when there is no external field is the one corresponding to the uniform
distribution. The intuition for the core of the algorithm is as follows. Suppose we are interested in
~ Note that for the uniform Ising model,
testing uniformity of Ising model p with parameter vector θ.
θuv = θu = 0 for all u, v ∈ V . We start by obtaining an upper bound on the SKL between p and Un
~ From (3), we have that under no
which can be captured via a statistic that does not depend on θ.
external field (θu = 0 for all u ∈ V ),
X
dSKL (p, Un ) =
θuv µuv
e=(u,v)∈E
=⇒ dSKL (p, Un ) ≤
X
β |µuv |
(32)
dSKL (p, Un ) X
=⇒
≤
|µuv | .
β
(33)
u6=v
u6=v
where (32) holds because |θuv | ≤ β.
P
Given the above upper bound, we consider the statistic Z = u6=v sign(µuv ) · (Xu Xv ), where
X ∼ p and sign(µuv ) is chosen arbitrarily if µuv = 0.
X
E[Z] =
|µuv | .
u6=v
If X ∈ In , then E[Z] = 0. On the other hand, by (33), we know that if dSKL (X, In ) ≥ ε, then
E[Z] ≥ ε/β. If the sign(µe ) parameters were known, we could simply plug them into Z, and using
Chebyshev’s inequality, distinguish these two cases using Var(Z)β 2 /ε2 samples.
There are two main challenges here.
• First, the sign parameters, sign(µuv ), are not known.
• Second, it is not obvious how to get a non-trivial bound for Var(Z).
One can quickly see that learning all the sign parameters might be prohibitively expensive. For
example, if there is an edge e such that |µe | = 1/2n , there would be no hope of correctly estimating
its sign with a polynomial number of samples. Instead, we perform a process we call weak learning
– rather than trying to correctly estimate all the signs, we instead aim to obtain a ~Γ which is
correlated with the vector sign(µe ). In particular, we aim to obtain ~Γ such that, in the case where
25
P
dSKL (p, Un ) ≥ ε, E[ e=(u,v)∈E Γe (Xu Xv )] ≥ ε/ζβ, where ζ = poly(n). That is we learn a sign
vector ~Γ which is correlated enough with the true sign vector such that a sufficient portion of the
signal from the dSKL expression is still preserved. The main difficulty of analyzing this process is
due to correlations between random variables (Xu Xv ). Naively, we could get an appropriate Γe for
(Xu Xv ) by running a weak learning process independently for each edge. However, this incurs a
prohibitive cost of O(n2 ) by iterating over all edges. We manage to sidestep this cost by showing
that, despite these correlations, learning all Γe simultaneously succeeds with a probability which is
≥ 1/ poly(n), for a moderate polynomial in n. Thus, repeating this process several times, we can
obtain a ~Γ which has the appropriate guarantee with sufficient constant probability.
At this point, we are in the setting as described above – we have a statistic Z 0 of the form:
X
Z0 =
cuv Xu Xv
(34)
u6=v
V
where c ∈ {±1}( 2 ) represent the signs obtained from the weak learning procedure. E[Z 0 ] = 0
if X ∈ In , and E[Z 0 ] ≥ ε/ζβ if dSKL (X, In ) ≥ ε. These two cases can be distinguished using
Var(Z 0 )ζ 2 β 2 /ε2 samples, by Chebyshev’s inequality. At this point, we run into the second issue
mentioned above. Since the range of Z 0 is Ω(n2 ), a crude bound for Var(Z 0 ) is O(n4 ), granting us
no savings over the localization algorithm of Theorem 2. However, in the high temperature regime,
we show the following bound on the variance of Z 0 (Theorem 16).
Var(Z 0 ) = Õ(n2 ).
In other words, despite the potentially complex structure of the Ising model and potential correlations, the variables Xu Xv contribute to the variance of Z 0 roughly as if they were all independent!
We describe the result and techniques involved in the analysis of the variance bound in Section 7.
Given the tighter bound on the variance of our statistic, we run the Chebyshev-based test on all the
hypotheses obtained in the previous learning step (with appropriate failure probability) to conclude
our algorithm. Further details about the algorithm are provided in Sections 5.1-5.4.
We state the sample complexity achieved via our learn-then-test framework for independence
testing under no external field here. The corresponding statements for independence testing under
external fields and identity testing are given in Section 5.5.
Theorem 6 (Independence Testing using Learn-Then-Test, No External Field). Suppose p is an
Ising
model in the high temperature regime under no external field. Then, given
10/3
n
β2
Õ
i.i.d samples from p, the learn-then-test algorithm runs in polynomial time and distinε2
guishes between the cases p ∈ In and dSKL (p, In ) ≥ ε with probability at least 9/10.
Next, we state a corollary of Theorem 6 with sample complexities we obtain when β is close to
the high temperature threshold.
Theorem 7 (Independence Testing with β near the Threshold of High Temperature, No External
1
Field). Suppose that p is an Ising model in the high temperature regime and suppose that β = 4dmax
.
That is, β is close to the high temperature threshold. Then:
10/3
• Given Õ εn2 d2
i.i.d samples from p with no external field, the learn-then-test algorithm
max
runs in polynomial time and distinguishes between the cases p ∈ In and dSKL (p, In ) ≥ ε with
probability at least 2/3. For testing identity of p to an Ising model q in the high temperature
regime, we obtain the same sample complexity as above.
26
Figure 1 shows the dependence of sample complexity of testing as dmax is varied in the regime
of Theorem 7 for the case of no external field.
The description of our algorithm is presented in Algorithm 6. It contains a parameter τ , which
we choose to be the value achieving the minimum in the sample complexity of Theorem 8. The
algorithm follows a learn-then-test framework, which we outline here.
Algorithm 6 Test if an Ising model p under no external field is product using Learn-Then-Test
1: function Learn-Then-Test-Ising(sample access to an Ising model p, β, dmax , ε, τ )
2:
Run the localization Algorithm 1 on p with accuracy parameter nετ . If it identifies any edges,.
return that dSKL (p, In ) ≥ ε
3:
for ` = 1 to O(n2−τ ) do
4:
Run the weak learning Algorithm 7 on S = {Xu Xv }u6=v with parameters τ and ε/β to.
(`)
generate a sign vector ~Γ(`) where Γuv is weakly correlated with sign (E [Xuv ])
5:
end for
6:
Using the same set of samples for all `, run the testing algorithm of Lemma 14 on each of.
the ~Γ(`) with parameters τ2 = τ, δ = O(1/n2−τ ). If any output that dSKL (p, In ) ≥ ε, return
that dSKL (p, In ) ≥ ε. Otherwise, return that p ∈ In
7: end function
Note: The first step in Algorithm 6 is to perform a localization test to check if |µe | is not too far
away from 0 for all e. It is added to help simplify the analysis of the algorithm and is not
necessary in principle. In particular, we use the first part of Algorithm 1, which checks if
any edge looks far from uniform, to perform this first step, albeit with a smaller value of the
accuracy parameter ε than before. Similar to before, if we find a single non-uniform edge,
this is sufficient evidence to output dSKL (X, In ) ≥ ε. If we do not find any edges which are
verifiably far from uniform, we proceed onward, with the additional guarantee that |µe | is
small for all e ∈ E.
A statement of the exact sample complexity achieved by our algorithm is given in Theorem 8.
When optimized for the parameter τ , this yields Theorem 6.
2
Theorem 8. Given Õ minτ >0 n2+τ + n6−2τ βε2 i.i.d samples from an Ising model p in the hightemperature regime with no external field, there exists a polynomial-time algorithm which distinguishes between the cases p ∈ In and dSKL (p, In ) ≥ ε with probability at least 2/3.
The organization of the rest of the section is as follows. We describe and analyze our weak
learning procedure in Section 5.1. Given a vector with the appropriate weak learning guarantees,
we describe and analyze the testing procedure in Section 5.2. In Section 5.3, we describe how to
combine all these ideas – in particular, our various steps have several parameters, and we describe
how to balance the complexities to obtain the sample complexity stated in Theorem 8. Finally, in
Section 5.4, we optimize the sample complexities from Theorem 8 for the parameter τ and filter out
cleaner statement of Theorem 6. We compare the performance of our localization and learn-then-test
algorithms and describe the best sample complexity achieved in different regimes in Section 6.
27
5.1
Weak Learning
Our overall goal of this section is “weakly learn” the sign of µe = E[Xu Xv ] for all edges e = (u, v).
More specifically, we wish to output a vector ~Γ with the following guarantee:
X
cε
EX
Γe Xu Xv ≥
,
2βn2−τ2
e=(u,v)∈E
for some constant c > 0 and parameter τ2 to be specified later. Note that the “best” Γ, for which
Γe = sign(µe ), has this guarantee with τ2 = 2 – by relaxing our required learning guarantee, we
can reduce the sample complexity in this stage.
The first step will be to prove a simple but crucial lemma answering the following question:
Given k samples from a Rademacher random variable with parameter p, how well can we estimate
the sign of its expectation? This type of problem is well studied in the regime where k = Ω(1/p2 ),
in which we have a constant probability of success (see, i.e. Lemma 3), but we analyze the case
when k 1/p2 and prove how much better one can do versus randomly guessing the sign. See
Lemma 20 in Section A for more details.
With this lemma in hand, we proceed to describe the weak learning procedure. Given parameters
τ ,ε and sample access to a set S of ’Rademacher-like’
2τ random variables which may be arbitrarily
correlated with each other, the algorithm draws Õ nε2 samples from each random variable in the
set and computes their empirical expected values and outputs a signs of thus obtained empirical
expectations. The procedure is described in Algorithm 7.
Algorithm 7 Weakly Learn Signs of the Expectations of a set of Rademacher-like random variables
1: function WeakLearning(sample access to set S = {Zi }i of random variables where |S| =
O(ns ) and where Zi ∈
{−1, 0, +1} and can be arbitrarily correlated,ε, τ ,).
(1)
(k)
n2τ
2:
Draw k = Õ ε2 samples from each Zi . Denote the samples by Zi , . . . , Zi
.
P
(l)
3:
Compute the empirical expectation for each Zi : Ẑi = k1 kl=1 Zi .
4:
Output ~Γ where Γi = sign(Ẑi ).
5: end function
We now turn to the setting of the Ising model, discussed in Section 5.1.1. We invoke the weaklearning procedure of Algorithm 7 on the set S = {Xu Xv }u6=v with parameters ε/β and 0 ≤ τ ≤ 2.
By linearity of expectations and Cauchy-Schwarz, it is not hard to see that we can get a guarantee
of the form we want in expectation (see Lemma 12). However, the challenge remains to obtain
this guarantee with constant probability. Carefully analyzing the range of the random variable and
using this guarantee on the expectation allows us to output an appropriate vector ~Γ with probability
inversely polynomial in n (see Lemma 13). Repeating this process several times will allow us to
generate a collection of candidates {~Γ(`) }, at least one of which has our desired guarantees with
constant probability.
5.1.1
Weak Learning the Edges of an Ising Model
We now turn our attention to weakly learning the edge correlations in the Ising model. To recall,
our overall goal is to obtain a vector ~Γ such that
X
cε
EX∼p
Γe Xu Xv ≥
.
2βn2−τ2
e=(u,v)∈E
28
We start by proving that the weak learning algorithm 7 yields a ~Γ for which such a bound holds
in expectation. The following is fairly straightforward from Lemma 20 and linearity of expectations.
and |µe | ≤
ε
βnτ2
n2τ2 β 2
ε2
samples from an Ising model X ∼ p such that dSKL (p, In ) ≥ ε
for all e ∈ E, Algorithm 7 outputs ~Γ = {Γe } ∈ {±1}|E| such that
!2
X
X
cβ
E~Γ EX∼p
|µe | ,
Γe Xu Xv ≥ 2−τ2
εn
Lemma 12. Given k = O
e∈E
e=(u,v)∈E
for some constant c > 0.
Proof. Since for all e = (u, v) ∈ E, |µe | ≤ βnετ2 , and by our upper bound on k, all of the random
variables Xu Xv fall into the first case of Lemma 20 (the “small k” regime). Hence, we get that
√
1 c1 |µe | k
Pr [Γe = sign(µe )] ≥ +
2
2
which implies that
√ !
1 c1 |µe | k
EΓe [Γe µe ] ≥
+
|µe | +
2
2
√
= c1 |µe |2 k
√ !
1 c1 |µe | k
−
(−|µe |)
2
2
Summing up the above bound over all edges, we get
"
#
X
√ X
|µe |2
E~Γ
Γe µe ≥ c1 k
e∈E
≥
e∈E
0
τ
2
c1 n β X
ε
|µe |2 ,
e∈E
for some constant c01 > 0. Applying the Cauchy-Schwarz inequality gives us
"
#
!2
X
X
cβ
E~Γ
Γe µe ≥ 2−τ2
|µe | ,
εn
e∈E
e∈E
as desired.
Next, we prove that the desired bound holds with sufficiently high probability. The following
lemma follows by a careful analysis of the extreme points of the random variable’s range.
2τ 2
2
Lemma 13. Given k = O n ε2β i.i.d. samples from an Ising model p such that dSKL (p, In ) ≥ ε
and |µe | ≤ βnετ2 for all e ∈ E, Algorithm 7 outputs ~Γ = {Γe } ∈ {±1}|E| where: Define χτ2 to be the
event that
X
cε
,
EX∼p
Γe Xu Xv ≥
2βn2−τ2
e=(u,v)∈E
for some constant c > 0. We have that
PrΓ [χτ2 ] ≥
29
c
4n2−τ2
.
Proof. We introduce some notation which will help in the elucidation of the argument which follows.
Let r = PrΓ [χτ2 ]. Let
!2
X
cβ
T =
|µe | .
2εn2−τ2
e∈E
Let Y be the random variable defined as follows
X
Y = EX∼p
Γe Xu Xv ,
e=(u,v)∈E
U = E~Γ [Y |Y > T ] and
L = E~Γ [Y |Y ≤ T ]
Then we have
rU + (1 − r)L ≥ 2T (From Lemma 12)
2T − L
=⇒ r ≥
U −L
Since U ≤
P
e∈E
|µe |, we have
2T − L
|µ
|
−L
e
e∈E
r≥ P
Since L ≥ −
P
e∈E
|µe |,
r≥
2
2T − L
P
e∈E |µe |
2
P
Since L ≤ T , we get
r≥
T
e∈E
|µe |
Substituting in the value for T we get
2
P
cβ
e∈E |µe |
P
r≥
4εn2−τ2
e∈E |µe |
P
cβ
e∈E |µe |
=⇒ r ≥
4εn2−τ2
P
Since dSKL (p, In ) ≥ ε, this implies
e∈E |µe | ≥ ε/β and thus
r≥
c
4n2−τ2
as desired.
30
,
5.2
Testing Our Learned Hypothesis
In this section, we assume that we were successful in weakly learning a vector ~Γ which is “good”
(i.e., it satisfies χτ2 , which says that the expectation the statistic with this vector is sufficiently
large). With such a ~Γ, we show that we can distinguish between p ∈ In and dSKL (p, In ) ≥ ε.
Lemma 14. Let p be an Ising model, let X ∼ p, and let σ 2 be such that, for any ~γ = {γe } ∈ {±1}|E| ,
X
Var
γe Xu Xv ≤ σ 2 .
e=(u,v)∈E
4−2τ2 β 2 log(1/δ)
Given k = O σ 2 · n
i.i.d samples from p, which satisfies either p ∈ In or dSKL (p, In ) ≥
2
ε
ε, and ~Γ = {Γe } ∈ {±1}|E| which satisfies χτ2 (as defined in Lemma 13) in the case that dSKL (p, In ) ≥
ε, then there exists an algorithm which distinguishes these two cases with probability ≥ 1 − δ.
Proof. We prove this lemma with failure probability 1/3 – by standard boosting arguments, this
can be lowered to δ by repeating the test O(log(1/δ)) times and taking the majority result.
Denote the ith sample as X (i) . The algorithm will compute the statistic
k
1 X X
Z=
Γe Xu(i) Xv(i) .
k
i=1 e=(u,v)∈E
If Z ≤ 4βncε2−τ2 , then the algorithm will output that p ∈ In . Otherwise, it will output that
dSKL (p, In ) ≥ ε.
By our assumptions in the lemma statement, in either case,
Var (Z) ≤
σ2
.
k
If p ∈ In , then we have that
E[Z] = 0.
By Chebyshev’s inequality, this implies that
ε
16σ 2 β 2 n4−2τ2
Pr Z ≥
≤
.
2−τ
4βn 2
kc2 ε2
Substituting the value of k gives the desired bound in this case. The case where dSKL (p, In ) ≥ ε
follows similarly, but additionally using the fact that χτ2 implies that
E[Z] ≥
5.3
cε
.
2βn2−τ2
Combining Learning and Testing
In this section, we combine lemmas from the previous sections to complete the proof of Theorem 8.
Lemma 13 gives us
weak learning step gives a “good” ~Γ with probability
that a single iteration of the
1
2−τ
at least Ω n2−τ2 . We repeat this step O(n 2 ) times, generating O(n2−τ2 ) hypotheses ~Γ(`) . By
standard tail bounds on geometric random variables, this will imply that at least one hypothesis is
31
good (i.e. satisfying χτ2 ) with probability at least 9/10. We then run the algorithm of Lemma 14
on each of these hypotheses, with failure probability δ = O(1/n2−τ2 ). If p ∈ In , all the tests will
output that p ∈ In with probability at least 9/10. Similarly, if dSKL (p, In ) ≥ ε, conditioned on at
∗
least one hypothesis ~Γ(` ) being good, the test will output that dSKL (p, In ) ≥ ε for this hypothesis
with probability at least 9/10. This proves correctness of our algorithm.
To conclude our proof, we analyze its sample complexity. Combining the complexities of Lemmas
5, 13, and 14, the overall sample complexity is
2τ1 2
2+τ2 2
4−2τ2 β 2
n β log n
n
β
2n
O
+O
+O σ
log n .
ε2
ε2
ε2
Noting that the first term is always dominated by the second term we can simplify the complexity
to the following expression.
2+τ2 2
4−2τ2 β 2
n
β
2n
+O σ
log n .
(35)
O
ε2
ε2
Plugging in the variance bounds from Section 7, Theorems 16 and 17 gives Theorem 8.
5.4
Balancing Weak Learning and Testing
The sample complexities in the statement of Theorem 8 arise from a combination of two separate
algorithms and from a variance bound for our multi-linear statistic which depends on β and dmax .
To balance for the optimal value of τ in various regimes of β and dmax we use Claim 3 which can
be easily verified and arrive at Theorem 6.
2
Claim 3. Let S = Õ n2+τ + n4−2τ · σ 2 βε2 . Let σ 2 = O(ns ). The value of τ which minimizes
S is
2+s
3 .
Claim 3 together with the variance bound (Theorem 16) implies Theorem 6.
Theorem 6 (Independence Testing using Learn-Then-Test, No External Field). Suppose p is an
Ising
model in the high temperature regime under no external field. Then, given
10/3
n
β2
Õ
i.i.d samples from p, the learn-then-test algorithm runs in polynomial time and distinε2
guishes between the cases p ∈ In and dSKL (p, In ) ≥ ε with probability at least 9/10.
5.5
Changes Required for General Independence and Identity Testing
We describe the modifications that need to be done to the learn-then-test approach described in
Sections 5.1-5.4 to obtain testers for independence under an arbitrary external field (Section 5.5.1),
identity without an external field (Section 5.5.2), and identity under an external field (Section 5.5.3).
5.5.1
Independence Testing under an External Field
Under an external field, the statistic we considered in Section 5 needs to be modified.
Suppose we are interested in testing independence of an Ising model p defined on a graph G = (V, E)
with a parameter vector θ~p . Let X ∼ p. We have that dSKL (p, In ) = minq∈In dSKL (p, q). In
particular, we consider q to be the independent Ising model on graph G0 = (V, E 0 ) with parameter
32
vector θ~q such that E 0 = φ and θuq is such that µqu = µpu for all u ∈ V . Then,
dSKL (p, In ) ≤ dSKL (p, q)
X
p
θuv
(µpuv − µquv )
=
(36)
e=(u,v)∈E
=
X
p
θuv
(µpuv − µpu µpv )
e=(u,v)∈E
≤
X
β |µpuv − µpu µpv |
e=(u,v)∈E
=⇒
dSKL (p, In )
≤
β
X
|µpuv − µpu µpv | .
e=(u,v)∈E
P
The above inequality suggests a statistic Z such that E[Z] = e=(u,v)∈E |λpuv | where λpuv = µpuv −
P
(1)
(2)
(1)
(2)
µpu µpv . We consider Z = u6=v sign(λuv ) Xu − Xu
Xv − Xv
where X (1) , X (2) ∼ p are
two independent samples from p. It can be seen that Z has the desired expectation. However,
we have the same issue as before that we don’t know the sign(λuv ) parameters. Luckily, it turns
out that our weak learning procedure is general
enough
case as well. Consider
to handle this
(1)
(2)
(1)
(2)
1
the following random variable: Zuv = 4 Xu − Xu
Xv − Xv . Zuv takes on values in
0 defined as follows: Pr[Z
{−1, 0, +1}. Consider an associated Rademacher variable Zuv
uv = −1] =
0
Pr[Zuv = −1] + 1/2 Pr[Zuv = 0]. It is easy to simulate a sample from Zuv given access to a sample
0 = −1 or +1. E[Z 0 ] = E[Z ] = λuv .
from Zuv . If Zuv = 0, toss a fair coin to decide whether Zuv
uv
uv
2
λuv
1
0
Hence Zuv ∼ Rademacher 2 + 4 and by Lemma 20 with k copies of the random variable Zuv we
√
get a success probability of 1/2+c1 k |λuv | of estimating sign(λuv ) correctly. Given this guarantee,
the rest of the weak learning argument of Lemmas 12 and 13 follows analogously by replacing µe
with λe .
0
After we have weakly learnt the signs, we are left with a statistic Zcen
of the form:
X
0
Zcen
=
cuv Xu(1) − Xu(2) Xv(1) − Xv(2)
(37)
u6=v
V
where the subscript cen denotes that the statistic is a centered one and c ∈ {±1}( 2 ) . We need to
0 ). We again employ the techniques described in Section 7 to obtain
obtain a bound on Var(Zcen
0 ) in the high-temperature regime. The statement of the variance
a non-trivial bound on Var(Zcen
result is given in Theorem 17 and the details are in Section 7.3. Combining the weak learning part
and the variance bound gives us the following sample complexity for independence testing under an
external field:
2+τ
(n
+ n4−2τ σ 2 )β 2
Õ
ε2
2+τ
(n
+ n4−2τ n2 )β 2
=Õ
ε2
Balancing for the optimal value of the τ parameter gives Theorem 9.
Theorem 9 (Independence Testing using Learn-Then-Test, Arbitrary External Field). Suppose p is
an Ising model in thehigh temperature
regime under an arbitrary external field. The learn-then-test
10/3 2
algorithm takes in Õ n ε2 β i.i.d. samples from p and distinguishes between the cases p ∈ In and
dSKL (p, In ) ≥ ε with probability ≥ 9/10.
33
The tester is formally described in Algorithm 8.
Algorithm 8 Test if an Ising model p under arbitrary external field is product
1: function Learn-Then-Test-Ising(sample access to an Ising model p, β, dmax , ε, τ )
2:
Run the localization Algorithm 1 with accuracy parameter nετ . If it identifies any edges,.
return that dSKL (p, In ) ≥ ε
3:
for ` = 1 to O(n2−τ ) do
(1)
(2)
(1)
(2)
4:
Run the weak learning Algorithm 7 on S = {(Xu − Xu )(Xv − Xv )}u6=v with param-.
(`)
eters τ2 h= τ and ε/β to generate ai
sign vector ~Γ(`) where Γuv is weakly correlated with
(1)
(2)
(1)
(2)
sign E (Xu − Xu )(Xv − Xv )
5:
end for
6:
Using the same set of samples for all `, run the testing algorithm of Lemma 14 on each of.
the ~Γ(`) with parameters τ2 = τ, δ = O(1/n2−τ ). If any output that dSKL (p, In ) ≥ ε, return
that dSKL (p, In ) ≥ ε. Otherwise, return that p ∈ In
7: end function
5.5.2
Identity Testing under No External Field
We first look at the changes needed for identity testing under no external field. Similar to before,
we start by obtaining an upper bound on the SKL between the Ising models p and q. We get that,
X
p
q
dSKL (p, q) =
(θuv
− θuv
) (µpuv − µquv )
(u,v)∈E
=⇒
dSKL (p, q) X p
≤
|(µuv − µquv )|
2β
u6=v
Since we know µquv for all pairs u, v, the above upper bound suggests the statistic Z of the form
X
Z=
sign (µpuv − µquv ) (Xu Xv − µquv )
u6=v
If p = q, E[Z] = 0 and if dSKL (p, q) ≥ ε, E[Z] ≥ ε/2β. As before, there are two things we need to
do: learn a sign vector which is weakly correlated with the right sign vector and obtain a bound on
Var(Z). By separating out the part of the statistic which is just a constant, we obtain that
X
Var(Z) ≤ Var
cuv Xu Xv
u6=v
V
where c ∈ {±1}( 2 ) . Hence, the variance bound of Theorem 16 holds for Var(Z).
As for the weakly learning the signs, using Corollary 2 of Lemma
√ 20 we get that for each pair
u, v, with k samples, we can achieve a success probability 1/2 + c1 k |µpuv − µquv | of correctly estimating sign(µpuv − µquv ). Following this up with analogous proofs of Lemmas 12 and 13 where µe
is replaced by µpe − µqe , we achieve our goal of weakly learning the signs with a sufficient success
probability.
By making these changes we arrive at the following theorem for testing identity to an Ising
model under no external field.
34
Theorem 10 (Identity Testing using Learn-Then-Test, No External Field). Suppose p and q are
Ising models
in thehigh temperature regime under no external field. The learn-then-test algorithm
10/3
n
β2
i.i.d. samples from p and distinguishes between the cases p = q and dSKL (p, q) ≥
takes in Õ
ε2
ε with probability ≥ 9/10.
The tester is formally described in Algorithm 9.
Algorithm 9 Test if an Ising model p under no external field is identical to q
1: function TestIsing(sample access to an Ising model p, β, dmax , ε, τ , description of Ising model
q under no external field)
2:
Run the localization Algorithm 2 with accuracy parameter nετ . If it identifies any edges,.
return that dSKL (p, q) ≥ ε
3:
for ` = 1 to O(n2−τ ) do
4:
Run the weak learning Algorithm 7 on S = {Xu Xv − µquv }u6=v with parameters τ2 = τ and.
(`)
ε/β to generate a sign vector ~Γ(`) where Γuv is weakly correlated with sign (E [Xuv − µquv ])
5:
end for
6:
Using the same set of samples for all `, run the testing algorithm of Lemma 14 on each of.
the ~Γ(`) with parameters τ2 = τ, δ = O(1/n2−τ ). If any output that dSKL (p, q) ≥ ε, return
that dSKL (p, q) ≥ ε. Otherwise, return that p = q
7: end function
5.5.3
Identity Testing under an External Field
When an external field is present, two things change. Firstly, the terms corresponding to nodes of
the Ising model in the SKL expression no longer vanish and have to be accounted for. Secondly, it
is unclear how to define an appropriately centered statistic which has a variance of O(n2 ) in this
setting, and we consider this an interesting open question. Instead, we use the uncentered statistic
which has variance Θ(n3 ).
We now describe the first change in more detail now. Again, we start by considering an upper
bound on the SKL between Ising models p and q.
X
X
p
q
dSKL (p, q) =
(θvp − θvq ) (µpv − µqv ) +
(θuv
− θuv
) (µpuv − µquv )
v∈V
=⇒ dSKL (p, q) ≤ 2h
(u,v)∈E
X
|(µpv
v∈V
− µqv )| + 2β
X
|(µpuv − µquv )|
u6=v
Hence if dSKL (p, q) ≥ ε, then either
P
• 2h v∈V |(µpv − µqv )| ≥ ε/2 or
P
• 2β u6=v |(µpuv − µquv )| ≥ ε/2.
P
P
Moreover, if p = q, then both 2h v∈V |(µpv − µqv )| = 0 and 2β u6=v |(µpuv − µquv )| = 0. Our tester
will first test for case (i) and if that test doesn’t declare that the two Ising models are far, then
proceeds to test whether case (ii) holds.
P
We will first describe the test to detect whether v∈V |(µpv − µqv )| = 0 or is ≥ ε/2h. We observe
that the random variables Xv are Rademachers and hence we can use the weak-learning framework
35
P
we developed so far to accomplish this goal. The statistic we consider is Z = v∈V sign(µpv ) (Xv − µqv ).
Again, as before, we face two challenges: we don’t know the signs of the node expectations µpv and
we need a bound on Var(Z).
We employ the weak-learning framework described in Sections 5.1-5.4 to weakly learn a sign
vector correlated with the true sign vector. In particular, since Xv ∼ Rademacher(1/2+µv /2), from
Corollary√2, we have that with k samples we can correctly estimate sign(µpv − µqv ) with probability
1/2 + c1 k |µpv − µqv |. The rest of the argument for obtaining a sign vector which, with sufficient
probability, preserves a sufficient amount of signal from the expected value of the statistic, proceeds
in a similar way as before. However since the total number of terms we have in our expression is
only linear we get some savings in the sample complexity.
P And from Lemma 1, weV have the following bound on functions fc (.) of the form fc (X) =
v∈V cv Xv (where c ∈ {±1} ) on the Ising model:
Var(fc (X)) = O(n).
By performing
in Sections 5.3 and 5.4, we obtain that by us 5/3 2 calculations analogous to the ones
P
p
q
n h
ing Õ
samples we can test whether
v∈V |(µv − µv )| = 0 or is ≥ ε/4h with probabilε2
P
ity ≥ 19/20. If the tester outputs that v∈V |(µpv − µqv )| = 0, then we proceed to test whether
P
p
q
u6=v |(µuv − µuv )| = 0 or ≥ ε/4β.
To perform this step, we begin by looking at the statistic Z used in Section 5.5.2:
X
Z=
sign (µpuv − µquv ) (Xu Xv − µquv )
u6=v
as Z has the right expected value. We learn a sign vector which is weakly correlated with the
true sign vector. However we need to obtain a variance bound on functions of the form fc (X) =
P
q
(V2 ) . By ignoring the constant term in f (X), we get that,
c
u6=v cuv (Xu Xv − µuv ) where c ∈ {±1}
X
Var(fc (X)) = Var
cuv Xu Xv
u6=v
which can be Ω(n3 ) as it is not appropriately centered.
We employ
Lemma 1 to get a variance
11/3
n
β2
3
bound of O(n ) which yields a sample complexity of Õ
for this setting.
ε2
Theorem 11 captures the total sample complexity of our identity tester under the presence of external
fields.
Theorem 11 (Identity Testing using Learn-Then-Test, Arbitrary External Field). Suppose p and
q are Ising models in thehigh temperature
regime under arbitrary external fields. The learn-then5/3 2
11/3 β 2
test algorithm takes in Õ n h +n
i.i.d. samples from p and distinguishes between the cases
ε2
p = q and dSKL (p, q) ≥ ε with probability ≥ 9/10.
The tester is formally described in Algorithm 10.
6
Comparing Localization and Learn-then-Test Algorithms
At this point, we now have two algorithms: the localization algorithm of Section 3 and the
learn-then-test algorithm of Section 5. Both algorithms are applicable in all temperature regimes
36
Algorithm 10 Test if an Ising model p under an external field is identical to Ising model q
1: function TestIsing(sample access to an Ising model p, β, dmax , ε, τ1 , τ2 , description of Ising
model q)
2:
Run the localization Algorithm 2 on the nodes with accuracy parameter 2nετ1 . If it identifies.
any nodes, return that dSKL (p, q) ≥ ε
3:
for ` = 1 to O(n1−τ1 ) do
4:
Run the weak learning Algorithm 7 on S = {(Xu − Yu }u∈V , where Yu ∼ Rademacher(1/2 +.
(`)
µqu /2), with parameters τ1 and ε/2h to generate a sign vector ~Γ(`) where Γu is weakly
correlated with sign (E [Xu − µqu ])
5:
end for
6:
Using the same set of samples for all `, run the testing algorithm of Lemma 14 on each of.
the ~Γ(`) with parameters τ3 = τ1 , δ = O(1/n1−τ1 ). If any output that dSKL (p, q) ≥ ε, return
that dSKL (p, q) ≥ ε
7:
————————–
8:
Run the localization Algorithm 2 on the edges with accuracy parameter 2nετ2 . If it identifies.
any edges, return that dSKL (p, q) ≥ ε
9:
for ` = 1 to O(n2−τ2 ) do
10:
Run the weak learning Algorithm 7 on S = {(Xu Xv − Yuv }u6=v , where Yuv ∼.
Rademacher(1/2 + µquv /2), with parameters τ2 and ε/2β to generate a sign vector ~Γ(`)
(`)
where Γuv is weakly correlated with sign (E [Xu Xv − µquv ])
11:
end for
12:
Using the same set of samples for all `, run the testing algorithm of Lemma 14 on each of.
the ~Γ(`) with parameters τ4 = τ2 , δ = O(1/n2−τ2 ). If any output that dSKL (p, q) ≥ ε, return
that dSKL (p, q) ≥ ε. Otherwise, return that p = q
13: end function
but learn-then-test beats localization’s sample complexity in high temperature under some degree
regimes. We note that their sample complexities differ in their dependence on β and dmax . In this
section, we offer some intuition as to why the difference arises and state the best sample complexities
we achieve for our testing problems by combining these two approaches.
First, we note that if the algorithm is agnostic of the maximum degree dmax , then learn-then-test
always outperforms localization in the high temperature regime. This leads to Theorem 12.
Theorem 12. Suppose p is an Ising model in the high temperature regime. To test either independence
4 2 or identity agnostic of the maximum degree of the graph dmax , localization requires
Õ nε2β
samples from p for a success probability > 2/3. Learn-then-test, on the other hand, re 10/3 2
quires Õ n ε2 β for independence testing and identity testing under no external field. It requires
11/3 2
Õ n ε2 β for identity testing under an external field.
When knowledge of dmax is available to the tester, he can improve his sample complexities of
localization approach. Now the sample complexity of localization gets worse as dmax increases. As
noted in Section 3, the reason for this worsening is that the contribution to the distance by any
single edge grows smaller thereby making it harder to detect. However, when we are in the hightemperature regime a larger dmax implies a tighter bound on the strength of the edge interactions β
and the variance bound of Section 7 exploits this tighter bound to get savings in sample complexities
when the degree is large enough.
37
We combine the sample complexities obtained by the localization and the learn-then-test algorithms and summarize in the following theorems the best sample complexities we can achieve
for testing independence and identity by noting the parameter regimes in which of the above two
algorithms gives better sample complexity. In both of the following theorems we fix β to be n−α
for some α and present which algorithm dominates as dmax ranges from a constant to n.
Theorem 13 (Best Sample Complexity Achieved, No External Field). Suppose p is an Ising model
under no external field.
• if β = O(n−2/3 ), then for the range dmax ≤ n2/3 , localization performs better, for both inde1
pendence and identity testing. For the range n2/3 ≤ dmax ≤ 4β
, learn-then-test performs better
than localization for both independence and identity testing yielding a sample complexity which
1
is independent of dmax . If dmax ≥ 4β
, then we are no longer in the high temperature regime.
• if β = ω(n−2/3 ), then for the entire range of dmax localization performs at least as well as the
learn-then-test algorithm for both independence and identity testing.
The theorem stated above is summarized in Figure 2 for the regime when β = O(n−2/3 ).
The comparison for independence testing under the presence of an external field is similar and
is presented in Theorem 14.
Theorem 14 (Best Sample Complexity Achieved for Independence Testing, Arbitrary External
Field). Suppose p is an Ising model under an arbitrary external field.
• if β = O(n−2/3 ), then for the range dmax ≤ n2/3 , localization performs better, for independence
1
, learn-then-test performs better than localization for
testing. For the range n2/3 ≤ dmax ≤ 4β
1
independence testing yielding a sample complexity which is independent of dmax . If dmax ≥ 4β
,
then we are no longer in the high temperature regime.
• if β = ω(n−2/3 ), then for the entire range of dmax localization performs at least as well as the
learn-then-test algorithm for independence testing.
Finally, we note in Theorem 15, the parameter regimes when learn-then-test performs better for
identity testing under an external field. Here our learn-then-test approach suffers worse bounds due
to a weaker bound on the variance of our statistic.
Theorem 15 (Best Sample Complexity Achieved for Identity Testing, Arbitrary External Field).
Suppose p is an Ising model under an arbitrary external field.
1
• if β = O(n−5/6 ), then for the range n2/3 ≤ dmax ≤ 4β
, learn-then-test performs better than
localization for identity testing yielding a sample complexity which is independent of dmax . If
1
dmax ≥ 4β
, then we are no longer in the high temperature regime.
• if β = ω(n−5/6 ), then for the entire range of dmax localization performs at least as well as the
learn-then-test algorithm for identity.
38
Figure 1: Localization vs Learn-Then-Test: A plot of the sample complexity of testing identity
1
under no external field when β = 4dmax
is close to the threshold of high temperature. Note that
throughout the range of values of dmax we are in high temperature regime in this plot.
Figure 2: Localization vs Learn-Then-Test: A plot of the sample complexity of testing identity
under no external field when β ≤ n−2/3 . The regions shaded yellow denote the high temperature
regime while the region shaded blue denotes the low temperature regime. The algorithm which
achieves the better sample complexity is marked on the corresponding region.
39
7
Bounding the Variance of Functions of the Ising Model in the
High-Temperature Regime
In this section, we describe a technique for bounding the variance of our statistics on the Ising
model in high temperature. As the structure of Ising models can be quite complex, it can be
challenging to obtain non-trivial bounds on the variance of even relatively simple statistics. In
particular, to apply our learn-then-test
framework of Section 5, we must bound the variance
P
0
of statistics of the form Z 0 =
c
u6=
v uv Xu Xv (under no external field, see (34)) and Zcen =
P
(1)
(2)
(1)
(2)
Xv − Xv
(under an external field, see (37)). While the variance for
u6=v cuv Xu − Xu
both the statistics is easily seen to be O(n2 ) if the graph has no edges, to prove variance bounds
better than the trivial O(n4 ) for general graphs requires some work. We show the following two
theorems in this section.
P
The first result, Theorem 16, bounds the variance of functions of the form u6=v cuv Xu Xv under
no external field which captures the statistic used for testing independence and identity by the
learn-then-test framework of Section 5 in the absence of an external field.
V
Theorem 16 (High Temperature VariancePBound, No External Field). Let c ∈ [−1, 1]( 2 ) and
define fc : {±1}V → R as follows: fc (x) = i6=j c{i,j} xi xj . Let also X be distributed according to
an Ising model, without node potentials (i.e. θv = 0, for all v), in the high temperature regime of
Definition 3. Then
Var (fc (X)) = Õ(n2 ).
The second result of this section, Theorem 17, bounds the variance of functions of the form
(1)
(2)
(1)
(2)
u6=v cuv (Xu − Xu )(Xv − Xv ) which captures the statistic of interest for independence testing
using the learn-then-test framework of Section 5 under an external field. Intuitively, this modification is required to “recenter” the random variables. Here, we view the two samples from Ising model
p over graph G = (V, E) as coming from a single Ising model p⊗2 over a graph G(1) ∪ G(2) where
G(1) and G(2) are identical copies of G.
P
V
Theorem 17 (High Temperature Variance Bound, Arbitrary External Field). Let c ∈ [−1, 1]( 2 )
and let X be distributed according to Ising model p⊗2 over graph G(1) ∪ G(2) in P
the high temperature
0
regime of Definition 3 and define gc : {±1}V ∪V → R as follows: gc (x) =
u,v∈V cuv (xu(1) −
s.t. u6=v
xu(2) )(xv(1) − xv(2) ). Then
Var(gc (X)) = Õ n
7.1
2
.
Overview of the Technique
We will use tools from Chapter 13 of [LPW09] to obtain the variance bounds of this section. At
a high level the technique to bound the variance of a function f on a distribution µ involves first
defining a reversible Markov chain with µ as its stationary distribution. By studying the mixing
time properties (via the spectral gap) of this Markov chain along with the second moment of the
variation of f when a single step is taken under this Markov chain we obtain bounds on the second
moment of f which consequently yield the desired variance bounds.
The Markov chain in consideration here will be the Glauber dynamics chain on the Ising model
p. As stated in Section 2, the Glauber dynamics are reversible and ergodic for Ising models. Let M
be the reversible transition matrix for the Glauber dynamics on some Ising model p. Let γ∗ be the
absolute spectral gap for this Markov chain. The first step is to obtain a lower bound on γ∗ .
40
Claim 4. In the high-temperature regime/under Dobrushin conditions, γ∗ ≥ Ω
arbitrary external field.
1
n log n
under an
Proof. From Theorem 15.1 of [LPW09], we have that the mixing time of the Glauber dynamics
is O(n log n). Since the Glauber dynamics on an Ising model is ergodic and reversible, using the
relation between mixing and relaxation times (Theorem 12.4 of [LPW09]) we get that
tmix ≥ γ1∗ − 1 log(2)
(38)
=⇒
1
γ∗
=⇒ γ∗
≤
n log n
log(2)
≥Ω
+1
1
n log n
(39)
.
(40)
For a function f , define
E(f ) =
1
2
X
[f (x) − f (y)]2 π(x)M (x, y).
x,y∈{±1}n
This can be interpreted as the expected square of the difference in the function, when a step is taken
at stationarity. That is,
1
E(f ) = E (f (x) − f (y))2
(41)
2
where x is drawn from the Ising distribution and y is obtained by taking a step in the Glauber
dynamics starting from x. We now state a slight variant of Remark 13.13 which we will use as
Lemma 15.
Lemma 15. For a reversible transition matrix P on state space Ω with stationary distribution π,
let
1 X
E(f ) :=
(f (x) − f (y))2 π(x)P (x, y),
2
x,y∈Ω
where f is a function on Ω such that Varπ (f ) > 0. Also let γ∗ be the absolute spectral gap of P .
Then
E(f )
γ∗ ≤
.
Varπ (f )
Note: Remark 13.13 in [LPW09] states a bound on the spectral gap as opposed to the absolute
spectral gap bound which we use here. However, the proof of Remark 13.13 also works for
obtaining a bound on the absolute spectral gap γ∗ .
7.2
Bounding Variance of Z 0 Under No External Field
We prove Theorem 16 now. Consider the function fc (x) =
P
u,v cuv xu xv
e
Claim 5. For an Ising model under no external field, E(fc ) = O(n).
41
|V |
where c ∈ [−1, 1]( 2 ) .
Proof. Since y is obtained
by taking a single step on the Glauber dynamics from x, fc (x) − fc (y) is a
P
function of the form v cv xv where cv ∈ [−1, 1] for all v ∈ V . The coefficients {cv }v depend on which
node v0 ∈ V was updated by the Glauber dynamics. Since there are n choices for nodes to update,
and since the update might P
also leave x unchanged, i.e. y = x, fc (x) − fc (y) is one of at most n + 1
√
linear functions of the form v cv xv . Denote, by E(x, y), the event that |fc (x) − fc (y)| ≥ c n log n.
We have, from the concentration of linear functions on the Ising model around their expected value
(Lemma 1) and a union bound over the n + 1 possible linear functions, that for a sufficiently large
1
c, under no external field, Pr [E(x, y)] ≤ 10n
2 . Now,
E (fc (x) − fc (y))2
= E (fc (x) − fc (y))2 |E(x, y) Pr[E(x, y)]
+ E (fc (x) − fc (y))2 |¬E(x, y) Pr[¬E(x, y)]
2
1
1
2
≤ n2 × 10n
2 + c n log n 1 − 10n2
e
= O(n)
where we used the fact that the absolute maximum value of (fc (x) − fc (y))2 is n2 .
Claim 4 together with Claim 5 are sufficient to conclude an upper bound on the variance of fc ,
by using Lemma 15, thus giving us Theorem 16.
7.3
0
Bounding Variance of Zcen
Under Arbitrary External Field
Under the presence of an external field, we saw that we need to appropriately center our statistics
to achieve low variance. The function gc (x) of interest now is defined over the 2-sample Ising model
p⊗2 and is of the form
X
(2)
(1)
(2)
gc (x) =
cuv (x(1)
u − xu )(xv − xv )
u,v
e
where now x, y ∈ {±1}2|V | . First, note that the absolute spectral gap for p⊗2 is also at least Ω(1/n).
Now we bound E(gc ).
e
Claim 6. For an Ising model under an arbitrary external field, E(gc ) = O(n).
Proof. Since y is obtained by taking a single step on the Glauber
dynamics from x, it can be seen
P
(1)
(2)
where cv ∈ [−1, 1] for all v ∈ V . The
that gc (x) − gc (y) is a function of the form v cv xv − xv
coefficients {cv }v depend on which node v0 ∈ V was updated by the Glauber dynamics. Since there
are n choices for nodes to update, and since the update might also leavex unchanged,
i.e. y = x,
P
(1)
(2)
gc (x) − gc (y) is one of at most n + 1 linear functions of the form v cv xv − xv . Denote, by
√
E(x, y), the event that |gc (x) − gc (y)| ≥ c n log n. We have, from Lemma 1 and a union bound,
1
that for a sufficiently large c, Pr [E(x, y)] ≤ 10n
2 . Now,
E (gc (x) − gc (y))2
= E (gc (x) − gc (y))2 |E(x, y) Pr[E(x, y)]
(42)
2
c
c
+ E (gc (x) − gc (y)) |E(x, y) Pr[E(x, y) ]
(43)
2
1
1
2
2
(44)
≤ 4n × 10n2 + c n log n 1 − 10n2
e
= O(n)
(45)
where we used the fact that the absolute maximum value of (gc (x) − gc (y))2 is 4n2 .
Similar to before, Claim 4 together with Claim 6 are sufficient to conclude an upper bound on
the variance of fc , by using Lemma 15, thus giving us Theorem 17.
42
8
Lower Bounds
In this section we describe our lower bound constructions and state the main results.
8.1
Dependences on n
Our first lower bounds show dependences on n, the number of nodes, in the complexity of testing
Ising models.
To start, we prove that uniformity testing on product measures over a binary alphabet requires
√
Ω( n/ε) samples. Note that a binary product measure corresponds to the case of an Ising model
with no edges. This implies the same lower bound for identity testing, but ( not) independence
testing, as a product measure always has independent marginals, so the answer is trivial.
Theorem 18. There exists a constant c > 0 such that any algorithm, given sample access to an
Ising model p with no edges (i.e., a product measure over a binary alphabet), which distinguishes
√
between the cases p = Un and dSKL (p, Un ) ≥ ε with probability at least 99/100 requires k ≥ c n/ε
samples.
Next, we show that any algorithm which tests uniformity of an Ising model requires Ω(n/ε)
samples. In this case, it implies the same lower bounds for independence and identity testing.
Theorem 19. There exists a constant c > 0 such that any algorithm, given sample access to an
Ising model p, which distinguishes between the cases p = Un and dSKL (p, Un ) ≥ ε with probability
at least 99/100 requires k ≥ cn/ε samples. This remains the case even if p is known to have a tree
structure and only ferromagnetic edges.
The lower bounds use Le Cam’s two point method which constructs a family of distributions P
such that the distance between any P ∈ P and a particular distribution Q is large (at least ε). But
given a P ∈ P chosen uniformly at random, it is hard to distinguish between P and Q with at least
2/3 success probability unless we have sufficiently many samples.
Our construction for product measures is inspired by Paninski’s lower bound for uniformity testing p
[Pan08]. We start with the uniform Ising model and perturb each node positively or negatively
by ε/n, resulting in a model which is ε-far in dSKL from Un . The proof appears in Section 8.3.1.
Our construction for the linear lower bound builds upon this style of perturbation. In the previous construction, instead of perturbing the node potentials, we could have left the node marginals
to be uniform and perturbed the edges of some fixed, known matching to obtain the same lower
bound. To get a linear lower bound, we instead choose a random perfect matching, which turns
out to require quadratically more samples to test. Interestingly, we only need ferromagnetic edges
(i.e., positive perturbations), as the randomness in the choice of matching is sufficient to make the
problem harder. Our proof is significantly more complicated for this case, and it uses a careful combinatorial analysis involving graphs which are unions of two perfect matchings. The lower bound is
described in detail in Section 8.3.2.
√
Remark 3. Similar lower bound constructions to those of Theorems 18 and 19 also yield Ω( n/ε2 )
and Ω(n/ε2 ) for the corresponding testing problems when dSKL is replaced with dTV . In our constructions, we describe families of distributions
which are ε-far in dSKL . This is done by perturbing
p
certain parameters by a magnitude of Θ( ε/n). We can instead describe families of distributions
√
which are ε-far in dTV by performing perturbations of Θ(ε/ n), and the rest of the proofs follow
similarly.
43
8.2
Dependences on h, β
Finally, we show that dependences on the h and β parameters are, in general, necessary for independence and identity testing. Recall that h and β are upper bounds on the absolute values of the
node and edge parameters, respectively. Our constructions are fairly simple, involving just one or
two nodes, and the results are stated in Theorem 20.
Theorem 20. There is a linear lower bound on the parameters h and β for testing problems on
Ising models. More specifically,
• There exists a constant c > 0 such that, for all ε < 1 and β ≥ 0, any algorithm, given sample
access to an Ising model p, which distinguishes between the cases p ∈ In and dSKL (p, In ) ≥ ε
with probability at least 99/100 requires k ≥ cβ/ε samples.
• There exists constants c1 , c2 > 0 such that, for all ε < 1 and β ≥ c1 log(1/ε), any algorithm,
given a description of an Ising model q with no external field (i.e., h = 0) and has sample
access to an Ising model p, and which distinguishes between the cases p = q and dSKL (p, q) ≥ ε
with probability at least 99/100 requires k ≥ c2 β/ε samples.
• There exists constants c1 , c2 > 0 such that, for all ε < 1 and h ≥ c1 log(1/ε), any algorithm,
given a description of an Ising model q with no edge potentials(i.e., β = 0) and has sample
access to an Ising model p, and which distinguishes between the cases p = q and dSKL (p, q) ≥ ε
with probability at least 99/100 requires k ≥ c2 h/ε samples.
The construction and analysis appears in Section 8.3.3.
This lower bound shows that the dependence on β parameters by our algorithms cannot be
avoided in general, though it may be sidestepped in certain cases. Notably, we show that testing
independence of a forest-structured Ising model under no external field can be done using Õ nε
samples (Theorem 3).
8.3
8.3.1
Lower Bound Proofs
Proof of Theorem 18
This proof will follow via an application of Le Cam’s two-point method. More specifically, we will
consider two classes of distributions P and Q such that:
1. P consists of a single distribution p , Un ;
2. Q consists of a family of distributions such that for all distributions q ∈ Q, dSKL (p, q) ≥ ε;
3. There exists some constant c > 0 such that any algorithm which distinguishes p from a
√
uniformly random distribution q ∈ Q with probability ≥ 2/3 requires ≥ c n/ε samples.
√
The third point will be proven by showing that, with k < c n/ε samples, the following two processes
have miniscule total variation distance, and thus no algorithm can distinguish them:
• The process p⊗k , which draws k samples from p;
• The process q̄ ⊗k , which selects q from Q uniformly at random, and then draws k samples from
q.
44
We will let p⊗k
be the process p⊗k restricted to the ith coordinate of the random vectors sampled,
i
⊗k
and q̄i is defined similarly.
p
We proceed with a description of our construction. Let δ = 3ε/2n. As mentioned before, P
consists of the single distribution p , Un , the Ising model on n nodes with 0 potentials on every node
and edge. Let M be the set of all 2n vectors in the set {±δ}n . For each M = (M1 , . . . , Mn ) ∈ M,
we define a corresponding qM ∈ Q where the node potential Mi is placed on node i.
Proposition 1. For each q ∈ Q, dSKL (q, Un ) ≥ ε.
Proof. Recall that
dSKL (q, Un ) =
X
δ tanh(δ).
v∈V
Note that tanh(δ) ≥ 2δ/3 for all δ ≤ 1, which can be shown using a Taylor expansion. Therefore
dSKL (q, Un ) ≥ n · δ · 2δ/3 = 2nδ 2 /3 = ε.
The goal is to upper bound dTV (p⊗k , q̄ ⊗k ). We will use the following lemma from [AD15], which
follows from Pinsker’s and Jensen’s inequalities:
Lemma 16. For any two distributions p and q,
2d2TV (p, q)
q
≤ log Eq
.
p
Applying this lemma, the fact that Q is a family of product distributions, and that we can
picture q̄ ⊗k as the process which picks a q ∈ Q by selecting a parameter for each node in an iid
manner, we have that
"
#
⊗k
q̄
2d2TV (p⊗k , q̄ ⊗k ) ≤ n log Eq̄⊗k 1⊗k .
1
p1
We proceed to bound the right-hand side. To simplify notation, let p+ = eδ /(eδ + e−δ ) be the
probability that a node with parameter δ takes the value 1. Note that a node with parameter −δ
takes the value 1 with probability 1 − p+ . We will perform a sum over all realizations k1 for the
number of times that node 1 is observed to be 1.
"
Eq̄⊗k
1
q̄1⊗k
p⊗k
1
#
=
k
X
(q̄1⊗k (k1 ))2
p⊗k
1 (k1 )
k =0
1
2
k
k1 (1 − p )k1
(p+ )k1 (1 − p+ )k−k1 + 12 k−k
(p
)
+
+
1
=
k
k
k1 (1/2)
k1 =0
k
2k X k
=
(p+ )2k1 (1 − p+ )2(k−k1 ) + (p+ )2(k−k1 ) (1 − p+ )2k1 + 2(p+ (1 − p+ ))k
4
k1
k
X
1 k
2 k1
k1 =0
=
k
k
X
2k
k
2k X
k
(p+ (1 − p+ ))k
+2·
(p2+ )k1 ((1 − p+ )2 )k−k1
2
k1
4
k1
k1 =0
k1 =0
45
where the second equality uses the fact that q̄1⊗k chooses the Ising model with parameter on node 1
P
being δ and −δ each with probability 1/2. Using the identity kk1 =0 kk1 ak1 bk−k1 = (a + b)k gives
that
"
#
k
q̄1⊗k
4k
2k
Eq̄⊗k ⊗k = (p+ (1 − p+ ))k +
2p2+ + 1 − 2p+ .
1
2
2
p1
Substituting in the value for p+ and applying hyperbolic trigenometric identities, the above expression simplifies to
k
k
1
sech2 (δ) + 1 + tanh2 (δ)
2
k 4
≤ 1+
δ
2
2
k 9ε
= 1+
2 4n2
where the inequality follows by a Taylor expansion.
This gives us that
2
k 9ε
9k 2 ε2
2
⊗k ⊗k
2dTV (p , q̄ ) ≤ n log 1 +
≤
.
2 4n2
4n
√
If k < 0.9 · n/ε, then d2TV (p⊗k , q̄ ⊗k ) < 49/50 and thus no algorithm can distinguish between the
two with probability ≥ 99/100. This completes the proof of Theorem 18.
8.3.2
Proof of Theorem 19
This lower bound similarly applies Le Cam’s two-point method, as described in the previous section.
We proceed with a description of our construction. Assume that n is even. As before, P consists of
the single distribution p , Un , the Ising model on n nodes with 0 potentials on every node and edge.
Let M denote the set of all (n − 1)!! perfect matchings on
p the clique on n nodes. Each M ∈ M
defines a corresponding qM ∈ Q, where the potential δ = 3ε/n is placed on each edge present in
the graph.
The following proposition follows similarly to Proposition 1.
Proposition 2. For each q ∈ Q, dSKL (q, Un ) ≥ ε.
The goal is to upper bound dTV (p⊗k , q̄ ⊗k ). We again apply Lemma 16 to 2d2TV (p⊗k , q̄ ⊗k ) and
focus on the quantity inside the logarithm. Let X (i) ∈ {±1}n represent the realization of the ith
sample and Xu ∈ {±1}k represent the realization of the k samples on node u. Let H(., .) represent
the Hamming distance between two vectors, and for sets S1 and S2 , let S = S1 ] S2 be the very
commonly used multiset addition operation (i.e., combine all the elements from S1 and S2 , keeping
duplicates). Let M0 be the perfect matching with edges (2i − 1, 2i) for all i ∈ [n/2].
⊗k
X
q̄
(q̄ ⊗k (X))2
Eq̄⊗k ⊗k =
p
p⊗k (X)
X=(X (1) ,...,X (k) )
X
= 2nk
(q̄ ⊗k (X))2
X=(X (1) ,...,X (k) )
46
We can expand the inner probability as follows. Given a randomly selected perfect matching, we can
break the probability of a realization X into a product over the edges. By examining the PMF of
theIsing model,
agree,the probability is multiplied by a factor
if the two endpoints of a given edge
eδ
e−δ
of 2(eδ +e−δ ) , and if they disagree, a factor of 2(eδ +e−δ ) . Since (given a matching) the samples
are independent, we take the product of this over all k samples. We average this quantity using a
uniformly random choice of perfect matching. Writing these ideas mathematically, the expression
above is equal to
1(Xu(i) =Xv(i) )
1(Xu(i) 6=Xv(i) ) 2
k
−δ
δ
X
X
Y
Y
e
1
e
2nk
δ
−δ
δ
−δ )
(n
−
1)!!
2(e
+
e
)
2(e
+
e
(1)
(k)
X=(X
,...,X
M ∈M (u,v)∈M i=1
)
X
= 2nk
X=(X (1) ,...,X (k) )
=
eδ
=
=
+
M ∈M (u,v)∈M
X
e−δ
X=(X (1) ,...,X (k) )
eδ
eδ + e−δ
nk
eδ
eδ + e−δ
nk
Y
1
δ
2(e + e−δ )
k
2
eδ(k−H(Xu ,Xv )) e−δH(Xu ,Xv )
2
nk
eδ
X
1
(n − 1)!!
1
(n − 1)!!
X
Y
exp(−2δH(Xu , Xv ))
M ∈M (u,v)∈M
2
1
(n − 1)!!2
1
(n − 1)!!2
X
X
Y
exp(−2δH(Xu , Xv ))
M ∈M (u,v)∈M
X=(X (1) ,...,X (k) )
X
X
Y
exp(−2δH(Xu , Xv ))
X=(X (1) ,...,X (k) ) M1 ,M2 ∈M (u,v)∈M1 ]M2
At this point, we note that if we fix the matching M1 , summing over all perfect matchings M2
gives the same value irrespective of the value of M1 . Therefore, we multiply by a factor of (n − 1)!!
and fix the choice of M1 to be M0 .
eδ
eδ + e−δ
=
nk
X
1
(n − 1)!!
X
Y
exp(−2δH(Xu , Xv ))
M ∈M X=(X (1) ,...,X (k) ) (u,v)∈M0 ]M
eδ
eδ + e−δ
nk
X
X
1
(n − 1)!!
(1)
M ∈M
X
k
Y
exp −2δH Xu(1) , Xv(1)
(u,v)∈M0 ]M
We observe that multiset union of two perfect matchings will form a collection of even length
cycles (if they contain the same edge, this forms a 2-cycle), and this can be rewritten as follows.
k
eδ
eδ + e−δ
nk
X X
1
(n − 1)!!
(1)
M ∈M
X
Y
Y
cyclesC (u,v)∈C
∈M0 ]M
exp −2δH Xu(1) , Xv(1)
k
=
eδ
eδ + e−δ
nk
X Y
1
(n − 1)!!
M ∈M
X
Y
cycles C X (1) (u,v)∈C
∈M0 ]M C
47
exp −2δH Xu(1) , Xv(1)
(46)
We now simplify this using a counting argument over the possible realizations of X (1) when
restricted to edges in cycle C. Start by noting that
n/2
(1)
(1)
X Y
X
|C| − 1
|C| − 1
2δ −2H Xu ,Xv
(e )
=2
+
(e2δ )−2i .
2i
−
1
2i
(1)
i=0
XC (u,v)∈C
This follows by counting the number of possible ways to achieve a particular Hamming distance over
the cycle. The |C| − 1 (rather than |C|) and the grouping of consecutive binomial coefficients arises
as we lose one “degree of freedom” due to examining a cycle, which fixes the Hamming distance to
be even. Now, we apply Pascal’s rule and can see
n/2
n/2
X
X
|C|
|C| − 1
|C| − 1
2δ −2i
(e2δ )−2i .
+
(e )
=2
2
2i
2i − 1
2i
i=0
i=0
This is twice the sum over the even terms in the binomial expansion of (1 + e−2δ )|C| . The odd terms
may be eliminated by adding (1 − e−2δ )|C| , and thus (46) is equal to the following.
k
eδ
eδ + e−δ
nk
X Y
1
−2δ |C|
−2δ |C|
(1
+
e
)
+
(1
−
e
)
(n − 1)!!
M ∈M
cycles C
∈M0 ]M
=
eδ
eδ + e−δ
nk
X Y
1
(n − 1)!!
M ∈M
Y
= E
cycles C
∈M0 ]M
cycles C
∈M0 ]M
k
eδ + e−δ
eδ
|C|
k
!
δ
|C|
e − e−δ
1+
eδ + e−δ
1 + tanh|C| (δ)
(47)
where the expectation is from choosing a uniformly random perfect matching M ∈ M. At this
point, it remains only to bound Equation (47). Noting that for all x > 0 and t ≥ 1,
1 + tanht (δ) ≤ 1 + δ t ≤ exp δ t ,
we can bound (47) as
Y
E
cycles C
∈M0 ]M
k
k
Y
≤ E
1 + tanh|C| (δ)
cycles C
∈M0 ]M
exp δ |C|
.
For our purposes, it turns out that the 2-cycles will be the dominating factor, and we use the
following crude upper bound. Let ζ be a random variable representing the number of 2-cycles in
M0 ] M , i.e., the number of edges shared by both perfect matchings.
k
k
Y
Y
|C|
|C|
2
4
2
= E
E
exp
δ
exp
δ
exp
δ
ζk
≤
exp
δ
·
n/4
·
k
E
exp
δ
ζk
,
cycles C
cycles C
∈M0 ]M
∈M0 ]M
|C|≥4
48
where in the last inequality, we used the facts that δ |C| is maximized for |C| ≥ 4 when |C| = 4, and
that there are at most n/4 cycles of length at least 4.
We examine the distribution of ζ. Note that
E[ζ] =
n
1
n
·
=
.
2 n−1
2(n − 1)
More generally, for any positive integer z ≤ n/2,
E[ζ − (z − 1)|ζ ≥ z − 1] =
n − 2z + 2
1
n − 2z + 2
·
=
.
2
n − 2z + 1
2(n − 2z + 1)
By Markov’s inequality,
Pr[ζ ≥ z|ζ ≥ z − 1] = Pr[ζ − (z − 1) ≥ 1|ζ ≥ z − 1] ≤
Therefore,
Pr[ζ ≥ z] =
z
Y
Pr[ζ ≥ i|ζ ≥ i − 1] ≤
i=1
n − 2z + 2
.
2(n − 2z + 1)
z
Y
n − 2i + 2
.
2(n − 2i + 1)
i=1
In particular, note that for all z < n/2,
Pr[ζ ≥ z] ≤ (2/3)z .
We return to considering the expectation above:
n/2
X
Pr[ζ = z] exp δ 2 zk
E exp δ 2 ζk =
z=0
≤
n/2
X
Pr[ζ ≥ z] exp δ 2 zk
z=0
n/2
3X
≤
(2/3)z exp δ 2 zk
2
z=0
n/2
3X
exp (δ 2 k − log(3/2))z
=
2
z=0
3
1
≤ ·
,
2
2 1 − exp (δ k − log(3/2))
where the last inequality requires that exp δ 2 k − log(3/2) < 1. This is true as long as k <
· nε .
log(3/2)/δ 2 = log(3/2)
3
Combining Lemma 16 with the above derivation, we have that
3
2
⊗k ⊗k
4
2dTV (p , q̄ ) ≤ log exp(δ nk/4) ·
2(1 − exp (δ 2 k − log(3/2)))
3
= δ 4 nk/4 + log
2(1 − exp (δ 2 k − log(3/2)))
2
9ε
3
=
k + log
.
4n
2(1 − exp (3kε/n − log(3/2)))
1
If k < 25
· nε , then dTV (p⊗k , q̄ ⊗k ) < 49/50 and thus no algorithm can distinguish between the two
cases with probability ≥ 99/100. This completes the proof of Theorem 19.
49
8.3.3
Proof of Theorem 20
We provide constructions for our lower bounds of Theorem 20 which show that a dependence on β
is necessary in certain cases.
Lemma 17. There exists a constant c > 0 such that, for all ε < 1 and β ≥ 0, any algorithm, given
sample access to an Ising model p, which distinguishes between the cases p ∈ In and dSKL (p, In ) ≥ ε
with probability at least 99/100 requires k ≥ cβ/ε samples.
Proof. Consider the following two models, which share some parameter τ > 0:
1. An Ising model p on two nodes u and v, where θup = θvp = τ and θuv = 0.
2. An Ising model q on two nodes u and v, where θuq = θvq = τ and θuv = β.
(2τ +β)+exp (−2τ +β)−exp(−β)
q q
2
We note that E[Xup Xvp ] = exp
exp (2τ +β)+exp (−2τ +β)+exp(−β) and E[Xu Xv ] = tanh (τ ). By (3), these two
p p
q q
models have dSKL (p, q) = β (E[Xu Xv ] − E[Xu Xv ]). For any for any fixed β sufficiently large and
ε > 0 sufficiently small, τ can be chosen to make E[Xup Xvp ]−E[Xuq Xvq ] = βε . This is because at τ = 0,
this is equal to tanh(β) and for τ → ∞, this approaches 0, so by continuity, there must be a τ which
causes the expression to equal this value. Therefore, the SKL distance between these two models
is ε. On the other hand, it is not hard to see that dTV (p, q) = Θ (E[Xup Xvp ] − E[Xuq Xvq ]) = Θ(ε/β),
and therefore, to distinguish these models, we require Ω(β/ε) samples.
Lemma 18. There exists constants c1 , c2 > 0 such that, for all ε < 1 and β ≥ c1 log(1/ε), any
algorithm, given a description of an Ising model q with no external field (i.e., h = 0) and has sample
access to an Ising model p, and which distinguishes between the cases p = q and dSKL (p, q) ≥ ε with
probability at least 99/100 requires k ≥ c2 β/ε samples.
Proof. This construction is very similar to that of Lemma 17. Consider the following two models,
which share some parameter τ > 0:
p
= β.
1. An Ising model p on two nodes u and v, where θuv
p
2. An Ising model q on two nodes u and v, where θuv
= β − τ.
We note that E[Xup Xvp ] = tanh(β) and E[Xuq Xvq ] = tanh(β − τ ). By (3), these two models have
dSKL (p, q) = τ (E[Xup Xvp ] − E[Xuq Xvq ]). Observe that at τ = β, dSKL (p, q) = β tanh(β), and at
τ = β/2, dSKL (p, q) = β2 (tanh(β) − tanh(β/2)) = β2 (tanh(β/2) sech(β)) ≤ β exp(−β) ≤ ε, where
the last inequality is based on our condition that β is sufficiently large. By continuity, there
exists some τ ∈ [β/2, β] such that dSKL (p, q) = ε. On the other hand, it is not hard to see that
dTV (p, q) = Θ (E[Xup Xvp ] − E[Xuq Xvq ]) = Θ(ε/β), and therefore, to distinguish these models, we
require Ω(β/ε) samples.
The lower bound construction and analysis for the h lower bound follow almost identically, with
the model q consisting of a single node with parameter h.
Lemma 19. There exists constants c1 , c2 > 0 such that, for all ε < 1 and h ≥ c1 log(1/ε), any
algorithm, given a description of an Ising model q with no edge potentials(i.e., β = 0) and has sample
access to an Ising model p, and which distinguishes between the cases p = q and dSKL (p, q) ≥ ε with
probability at least 99/100 requires k ≥ c2 h/ε samples.
Together, Lemmas 17, 18, and 19 imply Theorem 20.
50
Acknowledgements
The authors would like to thank Yuval Peres for directing them towards the reference [LPW09] and
the tools used to prove the variance bounds in this paper, and Pedro Felzenszwalb for discussing
applications of Ising models in computer vision.
References
[AD15]
Jayadev Acharya and Constantinos Daskalakis. Testing Poisson binomial distributions.
In Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA ’15, pages 1829–1840, Philadelphia, PA, USA, 2015. SIAM.
[ADK15]
Jayadev Acharya, Constantinos Daskalakis, and Gautam Kamath. Optimal testing for
properties of distributions. In Advances in Neural Information Processing Systems 28,
NIPS ’15, pages 3577–3598. Curran Associates, Inc., 2015.
[Agr12]
Alan Agresti. Categorical Data Analysis. Wiley, 2012.
[AJ06]
José A. Adell and Pedro Jodrá. Exact Kolmogorov and total variation distances between some familiar discrete distributions. Journal of Inequalities and Applications,
2006(1):1–8, 2006.
[AKN06]
Pieter Abbeel, Daphne Koller, and Andrew Y. Ng. Learning factor graphs in polynomial time and sample complexity. Journal of Machine Learning Research, 7(Aug):1743–
1788, 2006.
[BFF+ 01]
Tugkan Batu, Eldar Fischer, Lance Fortnow, Ravi Kumar, Ronitt Rubinfeld, and
Patrick White. Testing random variables for independence and identity. In Proceedings
of the 42nd Annual IEEE Symposium on Foundations of Computer Science, FOCS ’01,
pages 442–451, Washington, DC, USA, 2001. IEEE Computer Society.
[BGS14]
Guy Bresler, David Gamarnik, and Devavrat Shah. Structure learning of antiferromagnetic Ising models. In Advances in Neural Information Processing Systems 27, NIPS
’14, pages 2852–2860. Curran Associates, Inc., 2014.
[Bha16]
Bhaswar B. Bhattacharya. Power of graph-based two-sample tests. arXiv preprint
arXiv:1508.07530, 2016.
[BK16]
Guy Bresler and Mina Karzand. Learning a tree-structured Ising model in order to
make predictions. arXiv preprint arXiv:1604.06749, 2016.
[BM16]
Bhaswar B. Bhattacharya and Sumit Mukherjee. Inference in Ising models. Bernoulli,
2016.
[Bre15]
Guy Bresler. Efficiently learning ising models on arbitrary graphs. In Proceedings
of the 47th Annual ACM Symposium on the Theory of Computing, STOC ’15, pages
771–782, New York, NY, USA, 2015. ACM.
[CDGR16]
Clément L. Canonne, Ilias Diakonikolas, Themis Gouleakis, and Ronitt Rubinfeld.
Testing shape restrictions of discrete distributions. In Proceedings of the 33rd Symposium on Theoretical Aspects of Computer Science, STACS ’16, pages 25:1–25:14,
Dagstuhl, Germany, 2016. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik.
51
[CDKS17]
Clément L. Canonne, Ilias Diakonikolas, Daniel M. Kane, and Alistair Stewart. Testing
Bayesian networks. In Proceedings of the 30th Annual Conference on Learning Theory,
COLT ’17, pages 370–448, 2017.
[CF07]
Nicholas A. Christakis and James H. Fowler. The spread of obesity in a large social
network over 32 years. New England Journal of Medicine, 357(4):370–379, 2007.
[Cha05]
Sourav Chatterjee. Concentration Inequalities with Exchangeable Pairs. PhD thesis,
Stanford University, June 2005.
[CL68]
C.K. Chow and C.N. Liu. Approximating discrete probability distributions with dependence trees. IEEE Transactions on Information Theory, 14(3):462–467, 1968.
[CT06]
Imre Csiszár and Zsolt Talata. Consistent estimation of the basic neighborhood of
Markov random fields. The Annals of Statistics, 34(1):123–145, 2006.
[DDK17]
Constantinos Daskalakis, Nishanth Dikkala, and Gautam Kamath. Concentration of
multilinear functions of the Ising model with applications to network data. In Advances
in Neural Information Processing Systems 30, NIPS ’17. Curran Associates, Inc., 2017.
[DDS+ 13]
Constantinos Daskalakis, Ilias Diakonikolas, Rocco A. Servedio, Gregory Valiant, and
Paul Valiant. Testing k-modal distributions: Optimal algorithms via reductions. In
Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
’13, pages 1833–1852, Philadelphia, PA, USA, 2013. SIAM.
[DGJ08]
Martin Dyer, Leslie Ann Goldberg, and Mark Jerrum. Dobrushin conditions and
systematic scan. Combinatorics, Probability and Computing, 17(6):761–779, 2008.
[DK16]
Ilias Diakonikolas and Daniel M. Kane. A new approach for testing properties of
discrete distributions. In Proceedings of the 57th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’16, pages 685–694, Washington, DC, USA, 2016.
IEEE Computer Society.
[DKN15a]
Ilias Diakonikolas, Daniel M. Kane, and Vladimir Nikishkin. Optimal algorithms and
lower bounds for testing closeness of structured distributions. In Proceedings of the
56th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’15, pages
1183–1202, Washington, DC, USA, 2015. IEEE Computer Society.
[DKN15b]
Ilias Diakonikolas, Daniel M. Kane, and Vladimir Nikishkin. Testing identity of structured distributions. In Proceedings of the 26th Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA ’15, pages 1841–1854, Philadelphia, PA, USA, 2015. SIAM.
[DMR11]
Constantinos Daskalakis, Elchanan Mossel, and Sébastien Roch. Evolutionary trees
and the Ising model on the Bethe lattice: A proof of Steel’s conjecture. Probability
Theory and Related Fields, 149(1):149–189, 2011.
[Dob56]
Roland L. Dobrushin. Central limit theorem for nonstationary Markov chains. I. Theory of Probability & Its Applications, 1(1):65–80, 1956.
[DP17]
Constantinos Daskalakis and Qinxuan Pan. Square Hellinger subadditivity for Bayesian
networks and its applications to identity testing. In Proceedings of the 30th Annual
Conference on Learning Theory, COLT ’17, pages 697–703, 2017.
52
[Ell93]
Glenn Ellison.
Learning, local interaction, and coordination.
61(5):1047–1071, 1993.
[Fel04]
Joseph Felsenstein. Inferring Phylogenies. Sinauer Associates Sunderland, 2004.
[Fis35]
Ronald A. Fisher. The Design of Experiments. Macmillan, 1935.
[FLNP00]
Nir Friedman, Michal Linial, Iftach Nachman, and Dana Pe’er. Using Bayesian networks to analyze expression data. Journal of Computational Biology, 7(3-4):601–620,
2000.
[FOS08]
Jon Feldman, Ryan O’Donnell, and Rocco A. Servedio. Learning mixtures of product
distributions over discrete domains. SIAM Journal on Computing, 37(5):1536–1564,
2008.
[Geo11]
Hans-Otto Georgii. Gibbs Measures and Phase Transitions. Walter de Gruyter, 2011.
[GG86]
Stuart Geman and Christine Graffigne. Markov random field image models and their
applications to computer vision. In Proceedings of the International Congress of Mathematicians, pages 1496–1517. American Mathematical Society, 1986.
[GLP17]
Reza Gheissari, Eyal Lubetzky, and Yuval Peres. Concentration inequalities for polynomials of contracting Ising models. arXiv preprint arXiv:1706.00121, 2017.
[Hay06]
Thomas P. Hayes. A simple condition implying rapid mixing of single-site dynamics
on spin systems. In Proceedings of the 47th Annual IEEE Symposium on Foundations
of Computer Science, FOCS ’06, pages 39–46, Washington, DC, USA, 2006. IEEE
Computer Society.
[HKM17]
Linus Hamilton, Frederic Koehler, and Ankur Moitra. Information theoretic properties
of Markov random fields, and their algorithmic applications. In Advances in Neural
Information Processing Systems 30, NIPS ’17. Curran Associates, Inc., 2017.
[Isi25]
Ernst Ising. Beitrag zur theorie des ferromagnetismus. Zeitschrift für Physik A Hadrons
and Nuclei, 31(1):253–258, 1925.
[JJR11]
Ali Jalali, Christopher C. Johnson, and Pradeep K. Ravikumar. On learning discrete
graphical models using greedy methods. In Advances in Neural Information Processing
Systems 24, NIPS ’11, pages 1935–1943. Curran Associates, Inc., 2011.
[Jor10]
Michael Jordan. Lecture notes for Bayesian modeling and inference, 2010.
[JS93]
Mark Jerrum and Alistair Sinclair. Polynomial-time approximation algorithms for the
Ising model. SIAM Journal on Computing, 22(5):1087–1116, 1993.
[KM17]
Adam Klivans and Raghu Meka. Learning graphical models using multiplicative
weights. In Proceedings of the 58th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’17, Washington, DC, USA, 2017. IEEE Computer Society.
[KNS07]
Jeongwoo Ko, Eric Nyberg, and Luo Si. A probabilistic graphical model for joint
answer ranking in question answering. In Proceedings of the 30th International ACM
SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’07,
pages 343–350, New York, NY, USA, 2007. ACM.
53
Econometrica,
[LAFH01]
Charles Lagor, Dominik Aronsky, Marcelo Fiszman, and Peter J. Haug. Automatic
identification of patients eligible for a pneumonia guideline: comparing the diagnostic
accuracy of two decision support models. Studies in Health Technology and Informatics,
84(1):493–497, 2001.
[LC73]
Lucien Le Cam. Convergence of estimates under dimensionality restrictions. The
Annals of Statistics, 1(1):38–53, 1973.
[LPW09]
David A. Levin, Yuval Peres, and Elizabeth L. Wilmer. Markov Chains and Mixing
Times. American Mathematical Society, 2009.
[LRR13]
Reut Levi, Dana Ron, and Ronitt Rubinfeld. Testing properties of collections of distributions. Theory of Computing, 9(8):295–347, 2013.
[MdCCU16] Abraham Martín del Campo, Sarah Cepeda, and Caroline Uhler. Exact goodness-of-fit
testing for the Ising model. Scandinavian Journal of Statistics, 2016.
[MS10]
Andrea Montanari and Amin Saberi. The spread of innovations in social networks.
Proceedings of the National Academy of Sciences, 107(47):20196–20201, 2010.
[Pan08]
Liam Paninski. A coincidence-based test for uniformity given very sparsely sampled
discrete data. IEEE Transactions on Information Theory, 54(10):4750–4755, 2008.
[Pea00]
Karl Pearson. On the criterion that a given system of deviations from the probable in
the case of a correlated system of variables is such that it can be reasonably supposed to
have arisen from random sampling. Philosophical Magazine Series 5, 50(302):157–175,
1900.
[RAS15]
Firas Rassoul-Agha and Timo Seppäläinen. A Course on Large Deviations with an
Introduction to Gibbs Measures. American Mathematical Society, 2015.
[RS81]
Jon N.K. Rao and Alastair J. Scott. The analysis of categorical data from complex
sample surveys: Chi-squared tests for goodness of fit and independence in two-way
tables. Journal of the Americal Statistical Association, 76(374):221–230, 1981.
[RWL10]
Pradeep Ravikumar, Martin J. Wainwright, and John D. Lafferty. High-dimensional
ising model selection using `1 -regularized logistic regression. The Annals of Statistics,
38(3):1287–1319, 2010.
[STW10]
Sujay Sanghavi, Vincent Tan, and Alan Willsky. Learning graphical models for hypothesis testing and classification. IEEE Transactions on Signal Processing, 58(11):5481–
5495, 2010.
[SW12]
Narayana P. Santhanam and Martin J. Wainwright. Information-theoretic limits of selecting binary graphical models in high dimensions. IEEE Transactions on Information
Theory, 58(7):4117–4134, 2012.
[TAW10]
Vincent Y.F. Tan, Animashree Anandkumar, and Alan S. Willsky. Error exponents
for composite hypothesis testing of Markov forest distributions. In Proceedings of the
2010 IEEE International Symposium on Information Theory, ISIT ’10, pages 1613–
1617, Washington, DC, USA, 2010. IEEE Computer Society.
54
[VMLC16]
Marc Vuffray, Sidhant Misra, Andrey Lokhov, and Michael Chertkov. Interaction
screening: Efficient and sample-optimal learning of Ising models. In Advances in Neural
Information Processing Systems 29, NIPS ’16, pages 2595–2603. Curran Associates,
Inc., 2016.
[VV17]
Gregory Valiant and Paul Valiant. An automatic inequality prover and instance optimal identity testing. SIAM Journal on Computing, 46(1):429–455, 2017.
A
Weakly Learning Rademacher Random Variables
In this section, we examine the concept of “weakly learning” Rademacher random variables. This
problem we study is classical, but our regime of study and goals are slightly different. Suppose we have k samples from a random variable, promised to either be Rademacher(1/2 + λ)
or Rademacher(1/2 − λ), for some 0 < λ ≤ 1/2. How many samples do we need to tell which case
we are in? If we wish to be correct with probability (say) ≥ 2/3, it is folklore that k = Θ(1/λ2 ) samples are both necessary and sufficient. In our weak learning setting, we focus on the regime where
we are sample limited (say, when λ is very small), and we are unable to gain a constant benefit
over randomly guessing. More precisely, we have a budget of k samples from some Rademacher(p)
random variable, and we want to guess whether p > 1/2 or p < 1/2. The “margin” λ = |p−1/2| may
not be precisely known, but we still wish to obtain the maximum possible advantage over randomly
guessing, which gives us probability of success equal √
to 1/2. We show that with any k ≤ 1/4λ2
samples, we can obtain success probability 1/2 + Ω(λ k). This smoothly interpolates within the
“low sample” regime, up to the point where k = Θ(1/λ2 ) and folklore results also guarantee a
constant probability of success. We note that in this low sample regime, standard concentration
bounds like Chebyshev and Chernoff give trivial guarantees, and our techniques require a more
careful examination of the Binomial PMF.
We go on to examine the same problem under alternate centerings – where we are trying to
determine whether p > µ or p < µ, generalizing the previous case where µ = 1/2. We provide a
simple “recentering” based reduction to the previous case, showing that the same upper bound holds
for all values of µ. We note that our reduction holds even when the centering µ is not explicitly
known, and we only have limited sample access to Rademacher(µ).
We start by proving the following lemma, where we wish to determine the direction of bias with
respect to a zero-mean Rademacher random variable.
Lemma 20. Let X1 , . . . , Xk be iid random variables, distributed as Rademacher(p) for any p ∈
[0, 1]. There exists an algorithm which takes X1 , . . . , Xk as input and outputs a value b ∈ {±1},
with the following guarantees: there exists constants c1 , c2 > 0 such that for any p 6= 12 ,
(
√
1
+
c
|λ|
k if k ≤ 4λ1 2
1
Pr (b = sign (λ)) ≥ 21
otherwise,
2 + c2
where λ = p − 12 . If p = 21 , then b ∼ Rademacher 21 .
P
Proof. The algorithm is as follows: let S = ki=1 Xi . If S 6= 0, then output b = sign(S), otherwise
output b ∼ Rademacher 12 .
The p = 1/2 case is trivial, as the sum S is symmetric about 0. We consider the case where
λ > 0 (the negative case follows by symmetry) and when k is even (odd k can be handled similarly). As the case where k > 4λ1 2 is well known (see Lemma 3), we focus on the former case,
55
where λ ≤
1
√
.
2 k
By rescaling and shifting the variables, this is equivalent to lower bounding
Pr Binomial k, + λ ≥ k2 . By a symmetry argument, this is equal to
1
1
1
+ dTV Binomial k, − λ , Binomial k, + λ
.
2
2
2
√
It remains to show this total variation distance is Ω(λ k).
1
1
dTV Binomial k, − λ , Binomial k, + λ
2
2
1
1
, Binomial k, + λ
≥
dTV Binomial k,
2
2
Z 1/2+λ
Pr (Binomial (k − 1, u) = l − 1) du
(48)
≥
k
min
1
2
`∈{dk/2e,...,dk/2+kλe} 1/2
≥
=
≥
=
≥
=
λk · Pr (Binomial (k − 1, 1/2 + λ) = k/2)
k/2
k/2−1
k−1
1
1
λk ·
+λ
−λ
k/2
2
2
r
1
1 k/2
1 k/2
Ω(λk) ·
1+ √
1− √
2k
k
k
√
1 k/2
Ω(λ k) · 1 −
k
√
1 1/2
Ω(λ k) · exp (−1/2) 1 −
k
√
Ω(λ k),
(49)
(50)
as desired.
(48) applies Proposition 2.3 of [AJ06]. (49) is by an application of Stirling’s approximation and
k
c
since λ ≤ 2√1 k . (50) is by the inequality 1 − kc ≥ 1 − kc exp(−c).
We now develop a corollary allowing us to instead consider comparisons with respect to different
centerings.
Corollary 2. Let X1 , . . . , Xk be iid random variables, distributed as Rademacher(p) for any p ∈
[0, 1]. There exists an algorithm which takes X1 , . . . , Xk and q ∈ [0, 1] as input and outputs a value
b ∈ {±1}, with the following guarantees: there exists constants c1 , c2 > 0 such that for any p 6= q,
(
√
1
+ c1 |λ| k if k ≤ 4λ1 2
2
Pr (b = sign (λ)) ≥ 1
otherwise,
2 + c2
1
where λ = p−q
2 . If p = q, then b ∼ Rademacher 2 .
This algorithm works even if only given k iid samples Y1 , . . . , Yk ∼ Rademacher(q), rather than
the value of q.
Proof. Let X ∼ Rademacher(p) and Y ∼ Rademacher(q). Consider the random variable Z defined
as follows. First, sample X and Y . If X 6= Y , output 21 (X − Y ). Otherwise, output
a random
variable sampled as Rademacher 12 . One can see that Z ∼ Rademacher 12 + p−q
.
2
56
in this method using
Our algorithm can generate k iid samples Zi ∼ Rademacher 21 + p−q
2
Xi ’s and Yi ’s, where Yi ’s are either provided as input to the algorithm or generated according to
Rademacher(q). At this point, we provide the Zi ’s as input to the algorithm of Lemma 20. By
examining the guarantees of Lemma 20, this implies the desired result.
B
An Attempt towards Testing by Learning in KL-divergence
One approach to testing problems is by learning the distribution which we wish to test. If the
distance of interest is the total variation distance, then a common approach to learning is a coverbased method. One first creates a set of hypothesis distributions H which O(ε)-covers the space.
Then by drawing k = Õ(log |H|/ε2 ) samples from p, we can output a distribution from H with the
guarantee that it is at most O(ε)-far from p. The algorithm works by computing a score based on
the samples for each of the distributions in the hypothesis class and then choosing the one with the
maximum score.
However, it is not clear if this approach would work for testing in KL-divergence (an easier problem than testing in SKL-divergence) because KL-divergence does not satisfy the triangle inequality.
In particular, if p and q are far, and we learn a distribution p̂ which is close to p, we no longer
have the guarantee that p̂ and q are still far. Even if this issue were somehow resolved, the best
known sample complexity for learning follows from the maximum likelihood algorithm. We state
the guarantees provided by Theorem 17 of [FOS08].
Theorem 21 (Theorem 17 from [FOS08]). Let b, a, ε > 0 such that a < b. Let Q be a set of
hypothesis distributions for some distribution p over the space X such that at least one q ∗ ∈ Q is
such that dKL (p||q ∗ ) ≤ ε. Suppose also that a ≤ q(x) ≤ b for all q ∈ Q and for all x such that
p(x) > 0. Then running the maximum likelihood algorithm on Q using a set S of i.i.d. samples
from p, where |S| = k, outputs a q M L ∈ Q such that dKL (p||q M L ) ≤ 4ε with probability 1 − δ where
!
−2kε2
.
δ ≤ (|Q| + 1) exp
log2 ab
To succeed with probability at least 2/3, we need that
log (3(|Q| + 1)) log2
k≥
2ε2
b
a
For the Ising model, a KL-cover Q would consist of creating a poly(n/ε) mesh for each parameter.
2
Since there are O(n2 ) parameters, the cover will have a size of poly(n/ε)n . Letting β and h denote
the maximum edge and node parameter (respectively), then the ratio b/a in the above theorem is
such that
b
≥ exp O(n2 β + nh) .
a
Therefore, the number of samples required by this approach would be
!
˙ 2
n2 log nε (n
β + nh)2
k=O
ε2
6 2
n β + n4 h2
= Õ
ε2
57
which is more expensive than our baseline, the localization algorithm of Theorem 2. Additionally,
this algorithm is computationally inefficient, as it involves iterating over all hypotheses in the
exponentially large set Q. To summarize, there are a number of issues preventing a learning-based
approach from giving an efficient tester.
58
| 10math.ST
|
arXiv:1410.0602v1 [quant-ph] 2 Oct 2014
A probabilistic evolutionary optimization approach to
compute quasiparticle braids
Roberto Santana
Ross B. McDonald
Helmut G. Katzgraber
Intelligent Systems Group
Dept. of Computer Science
and Artificial Intelligence
University of the Basque
Country (UPV/EHU)
Dept. of Physics and
Astronomy
Texas A&M University, College
Station, TX 77843-4242, USA
Dept. of Physics and
Astronomy
Texas A&M University, College
Station, TX 77843-4242, USA
Santa Fe Institute, 1399 Hyde
Park Road, Santa Fe, NM
87501, USA
[email protected]
[email protected]
ABSTRACT
Topological quantum computing is an alternative framework
for avoiding the quantum decoherence problem in quantum
computation. The problem of executing a gate in this framework can be posed as the problem of braiding quasiparticles.
Because these are not Abelian, the problem can be reduced
to finding an optimal product of braid generators where the
optimality is defined in terms of the gate approximation and
the braid’s length. In this paper we propose the use of different variants of estimation of distribution algorithms to
deal with the problem. Furthermore, we investigate how the
regularities of the braid optimization problem can be translated into statistical regularities by means of the Boltzmann
distribution. We show that our best algorithm is able to
produce many solutions that approximates the target gate
with an accuracy in the order of 10−6 , and have lengths up
to 9 times shorter than those expected from braids of the
same accuracy obtained with other methods.
Keywords
topological computing, quasiparticle braids, estimation of
distribution algorithms, probabilistic graphical models, evolutionary algorithms
1.
INTRODUCTION
The idea of using the theory of quantum mechanics to obtain computers potentially exponentially faster for certain
applications, such as the factorization of prime numbers,
arouses considerable interest and research efforts from the
scientific community nowadays. In quantum computation,
information is represented and manipulated using quantum
properties. An obstacle for the construction of large quantum computers is the problem of quantum decoherence, that
can be viewed as the loss of information of the quantum sys-
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 2014 ACM .
tem due to the interaction with the environment. It terms
of the computation goal, quantum decoherence can be also
understood as unwanted noise introduced in quantum computation [25]. One possible solution to this problem is the
design of quantum systems immune to quantum decoherence
on a hardware level.
Topological quantum computing (TQC) [3, 25] investigates quantum computing systems that, given the properties of quasiparticles they use, are not affected by quantum
decoherence. The key idea of these systems is that quantum
information can be stored in global properties of the system
and thus affected only by global operations but not by local
perturbations such as noise. In TQC, quantum gates are
carried out by adiabatically braiding quasiparticles around
each other. This braiding is used to perform the unitary
transformations of a quantum computation.
One of the essential questions to design a TQC is to find
a product of braid generators (matrices) that approximates
a quantum gate with the smallest possible error and, if possible, as short as possible to prevent loss [12]. Some approaches that treat this question as an optimization problem
have been proposed. Exhaustive search [3] has been applied
to search for braids of manageable size (up to 46 exchanges).
More recently, McDonald and Katzgraber [12] have proposed
the use of genetic algorithms to find optimal braids. They
also introduce a function that takes into account the goals of
maximizing the accuracy and minimizing the length. In this
paper, we build on their results to propose an analysis of the
braid optimization problem using probabilistic modeling of
the space of braid solutions. For braid problems of small
size, we show that the regularities that exist in the search
space can be captured by the probabilistic models. We then
extend these results to propose the application of evolutionary algorithms (EAs) able to capture statistical regularities
of the best solutions.
Estimation of distribution algorithms (EDAs) [11, 15, 10]
are EAs that apply learning and sampling of distributions
instead of classical crossover and mutation operators. Modeling the dependencies between the variables of the problem
can serve to efficiently orient the search to more promising
areas of the search space by explicitly capturing and exploiting potential relationships between the problem variables. In
addition, the information contained in the models can reveal
previously unknown information about the problem.
The paper is organized as follows: The next section briefly
describes braids and Fibonacci anyons [3]. The problem formulation, including the problem representation and fitness
functions used are presented in Section 3. The framework
for probabilistic analysis of braids is introduced in Section 4.
This section also includes a number of experiments that help
to illustrate the rationale of our approach. Section 5 presents
the different variants of EDAs proposed for the braids problem. Work related to our approach is discussed in Section 6.
Section 7 describes the experimental framework to evaluate our proposal and presents the numerical results of our
experiments. The main contributions of the paper are summarized in Section 8 where some lines for future research are
also discussed.
2.
BRAIDS AND ANYONS
Qubits play in quantum computation a role similar to that
played by bits in digital computers. A braid operation can be
represented by a matrix that acts on the qubit space. These
matrices are referred to as generators, and the quantum gate
that a braid represents is the product of the generators that
encode the individual braid operations.
Let σ1 and σ2 represent two possible generators. σ1−1 and
−1
σ2 respectively represent their inverses. Given a braid B,
len() is a function that returns the braid’s length l (e.g.
B = σ1 σ1 σ2 σ1−1 , l = len(B) = 4).
Since the product of a matrix by its inverse reduces to the
identity matrix, some braids can be simplified reducing their
length. Therefore, we also define function elen(), that has a
braid as its argument and returns the braid’s effective length
which is the length of braid after all possible simplifications
have been conducted.
σ1 σ1 σ1 σ1 σ1−1 =σ1 σ1 σ1
σ2−1 σ1 σ1 σ1−1 σ1−1 σ2 σ1−1
=σ1−1
(1)
(2)
In the braids shown in examples (1) and (2), the effective
length values are 3 and 1, respectively.
Let T represent the target matrix (gate to be emulated),
the braid error is calculated with the following metric [12]:
ǫ = |B − T |
where the matrix norm used is
sX
2
|M | =
Mij
.
(3)
for the non computational state, and therefore it can be
ignored. Overall, phases in the problem can also be ignored. Therefore the transition matrices can be projected
iπ
onto SU(2) by a multiplication with e 10 , yielding for the
generators
σ1 =
e
−i7π
10
−e
0
−iπ
σ2 =
√
0
−τ e 10
√
−i τ
−i3π
10
!
!
√
−i τ
−iπ
−τ e 10
(5)
(6)
.
where τ = 5−1
2
In this paper we address the problem of finding a product of generator matrices for Fibonacci anyon braids. Although the methodology we propose can be extended to
other braids, we focus on anyon braids since they are one
of the best known in TQC [12, 29]. As a target gate for
computing the error (3) we use
T =
i
0
0
i
.
(7)
3. PROBLEM FORMULATION
3.1 Problem representation
Let X = (X1 , . . . , Xn ) denote a vector of discrete random
variables. We use x = (x1 , . . . , xn ) to denote an assignment
to the variables. I denotes a set of indices in {1, . . . , n}, and
XI (respectively xI ) a subset of the variables of X (respectively x) determined by the indices in I.
In our representation for the quasiparticle braids problem,
X = (X1 , . . . , Xn ) represents a braid of length n, where Xi
takes values in {0, 1, . . . , 2g − 1} and g is the number of
generators. Given an order for the generators σ1 , σ2 , . . . , σg ,
Xi = j, j < g means that the matrix in position i is σj+1 . If
−1
. For
Xi = j, j ≥ g, then the matrix in position i is σ(j−g)+1
example, for generators shown in Equations (5) and (6), and
B = σ1 σ1 σ2 σ2−1 σ1−1 , the corresponding braid representation
is x = (0, 0, 1, 3, 2).
3.2 Fitness function
(4)
ij
The problem of finding braiding operations that approximate gates is then reduced to finding a product chain of the
reduced generators and their inverses that approximates the
matrix representing the quantum gate. Two elements that
describe the quality of a braid are its error ǫ and its length
l.
2.1 Fibonacci anyon braids
Anyons appear as emergent quasiparticles in fractional
quantum Hall states and as excitations in microscopic models of frustrated quantum magnets that harbor topological
quantum liquids [21]. Fibonacci anyons are the simplest
anyons with non-Abelian braiding statistics that can give
rise to universal quantum computation. Fibonacci anyon
braids [3] only encompasses one-qubit gates. In such systems, the braid transition operators result in a phase change
We are interested in the solution of an optimization problem formulated as x∗ = argmaxxf (x), where f : S → R is
called the objective or fitness function. The optimum x∗ is
not necessarily unique.
To evaluate the fitness function associated to a solution
x, firstly the product of braid matrices B is computed according to x and then the error ǫ is calculated from B as
in (3).
The fitness function [12] is defined as:
f (x) =
1−λ
λ
+
1+ǫ
l
(8)
where l is the braid’s length, and the parameter λ serves to
balance the two conflicting goals, i.e., having short braids or
low approximation error. When λ = 0, braids are optimized
only for the error and the function reaches its maximum
value when this error is minimized.
We define functions fˆ(x) and f¯(x) as two variations of
function (8). Function fˆ(x) is identical to f (x), except that
the effective length l̂ = elen(B) is used instead of the braid’s
length. Function f¯(x) outputs the maximum value of the
function for any of the braids contained in B that start from
the first position, i.e.
f¯(x) = maxy,y∈{(x1 ),(x1 ,x2 ),(x1 ,...,xi ),(x1 ,...,xn )} f (y)
4.
(9)
4.2 Statistical analysis of the braids space
PROBABILISTIC MODELING OF BRAIDS
To optimize the braid problem we use a class of evolutionary algorithms that capture and exploit statistical regularities in the best solutions. These methods assume that
such regularities exist. As a preliminary proof of concept on
the existence of such regularities, we investigate the Boltzmann distribution for braids of manageable size. A similar
approach has been successfully applied to investigate the
dependencies that arise in the configurations of simplified
protein models [24] and conductance-based neuron models
[23].
4.1 Boltzmann distribution
When the dimension of the braid problem is small, complete enumeration and evaluation of all possible solutions is
feasible. In this situation, brute force can be applied to identify the optimal solutions. We use complete enumeration to
define a probability distribution on the space of all possible
braids for n = 10. Using the fitness value as an energy function, we associate to each possible braid a probability value
p(x) according to the Boltzmann probability distribution.
The Boltzmann probability distribution pB (x) is defined as
pB (x) = P
e
x′
g(x)
T
e
g(x′ )
T
,
The reason is that f¯ considers a greater space of solutions.
Differences between functions f and fˆ can be also detected
by comparing figures 1b) and 1c). For the three functions,
the space of solutions with lower fitness is more dense that
the solutions with higher fitness.
(10)
where g(x) is a given objective function and T is the system
temperature that can be used as a parameter to smooth the
the probabilities.
The Boltzmann probability distribution is used in statistical physics to associate a probability with a system state
according to its energy [28]. In our context of application,
pB (x) assigns a higher probability to braids that gives a
more accurate approximation to the target gate. The solutions with the highest probability correspond to the braids
that maximize the objective function. We use an arbitrary
choice of the temperature, T = 1, since our idea is to compare the distributions associated to different fitness functions for the same parameter T .
Using the Boltzmann distribution we can investigate how
potential regularities of the fitness function are translated
into statistical properties of the distribution. In particular, we investigate the marginal probabilities associated to
the variables and the mutual information between pairs of
variables. Figure 1 shows the probabilities assigned by the
Boltzmann distribution to functions f , fˆ, f¯, for n = 10. The
search space comprises 410 = 1, 048, 576 braids.
It can be seen in Figure 1a) that probabilities assigned
by the Boltzmann distribution to function f and fˆ are very
similar although not identical. For both functions, only few
points have a high probability. The important difference between functions f and f¯ is evident in Figure 1b). The probability assigned by the Boltzmann distribution to a braid is
always higher or equal for function f¯ than for function f .
Figure 2 shows the univariate probabilities computed from
the Boltzmann distribution for the three functions analyzed
and the 10 variables. p1 , p2 , p3 , and p4 respectively represent the univariate probabilities for braid generators λ1 ,
−1
λ2 , λ−1
1 , and λ2 . For all the functions, higher probabilities
is more likely to be present in the
for p3 indicate that λ−1
1
best solutions. This is the type of statistical regularities that
can be detected and exploited by EAs that learn probabilistic models. If a particular configuration is more likely to
be present in the best solutions, then, these configurations
could be sampled more frequently at the time of generating
new solutions.
Analysis of Figure 2 also reveals the similarities between
functions f and fˆ since they determine similar univariate
distributions for all the variables. A remarkable fact is that
for function f¯ (Figure 2c)) the univariate probabilities notably differ for the first variables and are more similar as the
index of the variables increases. One possible explanation
for this behavior is that the first variables are present in
more solutions of those considered by function f¯. Changes
in these variables are more influential in the function. Finally, another remarkable observation is that the univariate
are always very close
probabilities associated to λ1 and λ−1
1
for the three functions.
Finally, using the Boltzmann probabilities, we compute
the bivariate marginal distributions between every pair of
variables and derive the values of the mutual information.
The mutual information is a measure of statistical dependence between the variables and can serve to identify variables that are dependent. A strong dependence between two
variables may indicate that their joint effect has a strong
influence on the function. Figure 3 shows the mutual information computed for the three functions analyzed.
It can be seen in Figure 3 that for the three functions
the strongest dependencies are between adjacent variables,
although for functions f and fˆ there is also a strong dependence between the first variable and the last variable.
Although the pattern of dependence is similar in functions
f and fˆ, the mutual information is higher for function fˆ.
It can also appreciated in Figure 3c) that the dependencies between adjacent variables decreases with the index for
function f¯.
Summarizing, the statistical analysis of the Boltzmann
distribution shows that there are at least two types of regularities of the braid problem that are translated into statistical features. Firstly, there are different frequencies associated to the generators in the space of the best solutions.
Secondly, there are strong dependencies between the variables, particularly those that are adjacent in the braid representation.
5. MODELING THE BRAID SPACE
Using the Boltzmann distribution to find the statistical
regularities is not feasible for real problems for which the
space of solutions can not be inspected exhaustively. However, statistical regularities can be detected in samples of
a)
b)
c)
Figure 1: Boltzmann distribution computed for different functions. a) f vs fˆ , b) f vs f¯, and c) fˆ vs f¯.
0.2505
0.26
0.2505
p
1
p2
0.2504
p2
0.2504
p3
0.2503
0.2503
p
p2
0.258
p3
p3
p4
0.256
p
4
4
0.2501
0.25
Probabilities
0.2502
Probabilities
0.2502
Probabilities
p1
p
1
0.2501
0.25
0.2499
0.2499
0.2498
0.2498
0.2497
0.2497
0.2496
1
0.2496
1
0.254
0.252
0.25
0.248
2
3
4
5
6
7
8
9
10
0.246
2
3
4
5
Variables
6
7
8
9
0.244
1
10
2
3
4
5
a)
6
7
8
9
10
Variables
Variables
b)
c)
Figure 2: Univariate probabilities computed from the Boltzmann distribution for : a) f , b) fˆ, and c) f¯.
−9
−9
−5
x 10
x 10
1
2.5
2
x 10
1
4.5
1
2
4
2
3
3.5
3
2
4
1.5
5
6
Variables
Variables
4
4
3
5
2.5
6
2
Variables
2
3
1.5
5
6
1
1
7
7
7
1.5
8
8
8
1
0.5
9
9
0.5
9
0.5
10
10
1
2
3
4
5
6
7
8
Variables
9
10
0
10
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
Variables
a)
b)
5
6
7
8
9
10
0
Variables
c)
Figure 3: Mutual information computed from the Boltzmann distribution for: a) f , b) fˆ, and c) f¯.
n
50
100
150
200
250
braid
σ1−1 σ23 σ1−2 σ2 σ1−1 σ2 σ1−2 σ23 σ1−1 σ2 σ1−1 σ2 σ1−2 σ2 σ1−2 σ2 σ1−1 σ23 σ1−1 σ22 σ1−1 σ2 σ1−1 σ2 σ1−2 σ22 σ1−3 σ22
σ1 σ2−2 σ14 σ2−1 σ1 σ2−4 σ1 σ2−1 σ13 σ2−1 σ12 σ2−2 σ14 σ2−1 σ1 σ2−4 σ1 σ2−1 σ16 σ22 σ1 σ2 σ1 σ22 σ13 σ25 σ1−1 σ2 σ1−3 σ2 σ13 σ25
σ2 σ1−1 σ2 σ1−1 σ2−1 σ12 σ2−1 σ1 σ2−4 σ12 σ2−2 σ1−1 σ2−1 σ1−2 σ2−4 σ12 σ2−4 σ1 σ2−1 σ12 σ2−1 σ1−1 σ2 σ1−3 σ2−1 σ1 σ2−1 σ13 σ2−1 σ1 σ2−4 σ12 σ2−4 σ12 σ2−3
σ2−2 σ14 σ2−1 σ1 σ2−4 σ1 σ2−1 σ12 σ2−1 σ13 σ2−1 σ1 σ2−4 σ1 σ2−1 σ14 σ2−1 σ1 σ2−2 σ12 σ22 σ1 σ2−1 σ1 σ2−1 σ1−5 σ2 σ1−1 σ2−6 σ1−2 σ23
σ1 σ2 σ1−1 σ2 σ1−1 σ2−1 σ1 σ2−1 σ1−3 σ24 σ1−2 σ24 σ1−1 σ2 σ12 σ2 σ1−1 σ24 σ1−2 σ24 σ1−3 σ2−1 σ1 σ2−1 σ1−1 σ22 σ1 σ2−1 σ12 σ2−1 σ1 σ2−4 σ12 σ2−4 σ1 σ2−1 σ14
σ2−1 σ1−1 σ2−1 σ1−4 σ2−1 σ1−1 σ2−1 σ14 σ2−1 σ1 σ2−4 σ12 σ2−4 σ1 σ2−2 σ1−1 σ25 σ1−1 σ2 σ1−4 σ22 σ1−4 σ22 σ1−4 σ22 σ1−1
Table 1: Best braids found by the EDAs for each value of n.
solutions selected according to their fitness. EDAs use samples of solutions to learn a model that captures some of the
regularities that may exist in the data. The pseudocode of
an EDA is shown in Algorithm 1.
EDAs that used univariate, one-order Markov, and tree
models were respectively introduced in [1, 16], [5] and [2]
and details on the methods used to learn and sample the
models can be obtained from these references.
Algorithm 1: Estimation of distribution algorithm
5.1 Enhancements to the EDAs
1 Set t ⇐ 0. Generate N solutions randomly.
2 do {
3
Evaluate the solutions using the fitness function.
We consider three enhancements to EDAs: 1) Use of a
local optimizer. 2) Partial sampling. 3) Recoding.
As is the case of other EAs, EDAs can be enhanced by
the incorporation of local optimizers [19]. We use a greedy
optimization algorithm that is applied during the evaluation
of the population by the EDA. The algorithm starts from the
solution generated by the EDA. In each iteration, the local
optimizer evaluates all the 3n solutions that are different
to the current solution in only one variable (the neighbor
solutions). The next selected solution is the neighbor that
improves the fitness of the current solution the most. The
algorithm stops when none of the neighbors improves the
fitness of the current solution.
During the sampling step of an EDA, all variables are assigned their values according to the probabilistic model and
the sampling method. For the EDA that uses the univariate
model, variables are independently sampled. For 1-order
Markov and tree, probabilistic logic sampling (PLS) [7] is
used. In both methods, all variables are assigned the new
values. However, for some problems with higher-order interactions using a base template solution can be better than
generating each new solution from scratch. For the braid
problem, careful recombination of the solutions proved to
be essential for the genetic algorithm introduced in [12].
In partial sampling, a solution of the population is selected
and only a subset of its variables are sampled according to
the model. We use two variants of partial sampling I) Partial
sampling where the number of variables to be modified is
randomly selected between 1 and n. II) Partial sampling,
where the number of variables to be modified is randomly
selected between 1 and n2 .
Recoding consists in modifying the representation of the
solution according to the evaluation. For functions fˆ and f¯
it is possible to recode the solution by eliminating redundant
generators (e.g., pairs σi σi−1 ). The rationale of using recoding is that meaningful variables will be located closer to the
beginning of the braid. Since solutions have a fixed length,
the last variables will be kept unused, i.e. garbage information. Therefore, we devised two ways to fill these gaps: I)
Leaving the unused variables as they were in the original solution. II) Replacing the unused variables by a reverse copy
of the variables used in the evaluation. The second variant
intends to replicate information that has proved to be “informative” about the problem. Equations (14) and (15) show
examples of recoding type I and II, respectively. In this hypothetical examples, the underlined variables are those that
provided the best fitness after simplifying the braid and evaluating function f¯.
4
Select a population DtS of K ≤ N solutions according to a selection method.
Calculate a probabilistic model of DtS .
Generate N new solutions sampling from the distribution represented in the model.
7
t⇐t+1
8 } until Termination criteria are met.
5
6
The model used by the EDA will determine the type of
regularities that will be captured from the data and also
the cost of the optimization algorithm since more complex
models are generally more computationally costly.
In this paper we use three types of probabilistic graphical
models: 1) Univariate model. 2) 1-order Markov model. 3)
Tree model. We work with positive distributions denoted
by p. p(xI ) denotes the marginal probability for XI = xI .
p(xi | xj ) denotes the conditional probability distribution of
Xi = xi given Xj = xj .
In the univariate model variables are considered to be independent, and the probability of a solution is the product
of the univariate probabilities for all variables:
pu (x) =
n
Y
p(xi )
(11)
i=1
In the 1-order Markov model, the configuration of variable
Xi depends on the configuration of its previous variable:
pM K (x) = p(x1 )
n
Y
i=2
p(xi | xi−1 )
(12)
A probability distribution pT (x) that is conformal with a
tree is defined as:
n
Y
p(xi |pa(xi )),
(13)
pT (x) =
i=1
where P a(Xi ) is the parent of Xi in the tree, and p(xi |pa(xi )) =
p(xi ) when pa(Xi ) = ∅, i.e. Xi is a root of the tree. We allow the existence of more than one root in the PGM (i.e.
forests) although for convenience of notation we refer to the
model as tree.
Univariate approximations are expected to work well for
functions that can be additively
decomposed into functions
P
of order one (e.g. g(x) = i xi ). However, other non additively decomposable functions can be easily
Q solvedPwith
EDAs that use univariate models (e.g. g(x) = i xi + i xi )
[14]. Therefore, it makes sense to test the univariate approximation for the braid problem. The 1-order Markov model
captures only dependencies between adjacent variables, and
the tree model can represent a maximum of n − 1 bivariate
dependencies.
(0, 3, 1, 3, 3, 3, 2, 1, 2, 2) =(0, 3, 3, 3, 2, 3, 2, 1, 2, 2)
(14)
(0, 3, 1, 3, 3, 3, 2, 1, 2, 2) =(0, 3, 3, 3, 2, 2, 3, 3, 3, 0)
(15)
6. RELATED WORK
In addition to the use of genetic algorithms [12], brute
force [3] has been proposed to search for a braids of manageable size (up to 46 exchanges). Other methods such as
the Solovay-Kitaev algorithm [4, 9, 17] provide bounds on
the accuracy and length of the braids. However, these methods do not allow the user to tune the balance between the
accuracy and the length as pioneered in [12].
The Boltzmann distribution has played an important role
in the theoretical analysis of EDAs and other authors works
have analyzed the relationship between the function structure and the dependencies in the distribution [13, 18, 22].
Other problems from physics have been previously treated
with EDAs. In particular, spin glass models with different
types of interactions and topologies have been addressed [19,
20, 27]. Two important differences between the braid problem and the spin glass models that makes it particularly
challenging is that its fitness function is multiplicative and
that the representation is non binary. In fact the cardinality
of the variables can increase with the number of generators.
n
50
100
150
200
250
l
44
70
64
62
124
ǫ
4.8435 × 10−4
8.3527 × 10−6
8.3527 × 10−6
8.3527 × 10−6
3.5038 × 10−6
log10 (ǫ)
−3.3148
−5.0782
−5.0782
−5.0782
−5.4555
3.97 1
log10
(ǫ)
116.47
633.37
633.37
633.37
841.82
Table 2: Parameters of the best braids found by the
EDAs for each value of n.
Figure 4 shows the length of all the best solutions found
for each value of n. It can be observed in Figure 4 that
EDAs are able to find several braids with different lengths
for n = 150 and n = 200.
200
EXPERIMENTS
The main objective of our experiments is to evaluate the
capacity of the EDAs to find optimal solutions to the braid
problem. We run experiments for n ∈ {50, 100, 150, 200, 250}
in order to evaluate the scalability of the algorithms. A second objective is to compare different variants of the problem
formulation and of the algorithm.
7.1 Experimental settings
Each EDA is characterized by 5 parameters:
180
160
140
Length
7.
120
100
80
60
40
• Use of local optimizer. 0: Only EDA is applied, 1:
EDA is combined with greedy search as described in
Section 5.1.
• Type of function and representation. 0: Function f ,
1: Function f¯ without recoding, 2: Function f¯ with
recoding type I, 3: Function f¯ with recoding type II.
• λ value. 0:0.0, 1:0.01, 2:0.05, 3:0.1.
• Sampling method. 0: Normal, 1: Partial sampling
type I, 2: Partial sampling type II.
• Type of probabilistic model. 0: Univariate, 1: 1-order
Markov, 2: Tree.
The total number of variants of the algorithm was 2 ×
4 × 4 × 3 × 3 = 288. All the algorithms use truncation
selection, in which the best 5% of the population is selected.
EDAs that do not incorporate the greedy local search use a
population size N = 10000. For these EDAs, the number of
generations was dependent on n as Ng = 15n.
Due to the large number of evaluations spent by the greedy
search method, the population size for all hybrid EDAs was
N = 100n and the number of generations was fixed to
Ng = 100. For each EDA variant, 100 experiments were
run.
7.2 Best solutions found by EDAs
Tables 1 and 2 respectively show the best braids achieved
by the EDAs for each value of n and the characteristics of
these solutions. In Table 2, we show an estimate of the
3.97
length of the braids (O[log10
(1/ǫ)]) that would compute
the Solovay-Kitaev algorithm to obtain the same error ǫ of
our best solutions. The lengths of our solutions compare
favorably with these estimates.
50
100
150
200
250
Maximum length
Figure 4: Length of the best solutions found for each
value of n.
7.3 Behavior of the different EDA variants
We further investigate the behavior of the different EDA
variants. Figure 5 shows the violin plots [8] with the distribution of the best values found in all the executions for: a)
All EDA variants without local optimizer (14400 runs), b)
All EDA variants that incorporate the greedy search (14400
runs), c) EDAs with local optimizer, recoding type II, and
that use partial sampling type II (300 runs). Each violin
plot shows a histogram smoothened using a kernel density
with Normal kernel. The mean and median are shown as
red crosses and green squares, respectively.
In Figure 5, the modes of the Normal distribution indicate
the existence of a local optimum with a very wide basin
of attraction. This local optimum has value log10 f (ǫ) =
−2.50785 and the majority of the EDA runs can be trapped
in this value. Differences between the EDAs due to the application of the greedy method can be appreciated for n = 200
and n = 250 (Figures 5a) and 5b)). Also, Figure 5c) reveals
how a particular combination of the EDA’s parameters can
improve the results of the search. This is shown in detail in
Table 3 that comprises all EDA variants that reached one of
the best solutions in at least one of the 100 runs.
There are a number of commonalities between the best
EDA variants included in Table 3. Except in one case, all
EDAs use recoding of type II. Similarly, except in one case,
in all the variants λ ∈ {0.01, 0.05}. Except in two cases, the
sampling method selected was partial sampling. The application of the local optimizer notably improved the results
0
0
−1
−1.5
−1
−1
−2
−2
−2
−2.5
log10 ε
ε
10
log
log
10
ε
−3
−3
−3
−3.5
−4
−4
−4
−4.5
−5
−5
−5
−5.5
−6
50
100
150
200
−6
250
50
100
Maximum length
150
200
−6
250
Maximum length
a)
50
100
150
200
250
Maximum length
b)
c)
Figure 5: Violin plots showing the distribution of the best values found in all the executions for: a) All EDAs
variants without local optimizer (14400 runs), b: All EDAs variants with local optimizer (14400 runs), c: EDAs
with local optimizer, recoding type II, and that use partial sampling type II (300 runs).
tλ
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
0
1
1
1
1
1
2
2
2
2
2
2
ts
0
1
1
1
2
2
2
0
1
1
1
2
2
2
1
1
1
1
2
2
2
1
1
1
2
2
2
pm
0
0
1
2
0
1
2
0
0
1
2
0
1
2
2
2
0
2
0
1
2
0
1
2
0
1
2
50
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
100
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
2
1
0
0
0
0
0
0
150
1
1
1
1
0
0
0
0
4
0
1
0
1
2
1
1
0
0
9
14
18
0
1
1
4
2
4
200
1
4
2
0
2
2
3
1
3
1
1
1
1
3
0
0
2
1
12
21
19
4
1
0
20
9
11
250
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
tot
2
6
3
1
2
2
4
1
7
1
2
2
3
5
1
1
2
1
21
37
38
4
2
1
24
11
16
−0.6
−0.8
−1
−1.2
log
Table 3: EDAs variants that obtained one of the
best solutions at least one in the 100 experiments.
L: Local optimizer, tf: Type of function and representation, tλ: type of λ value, ts: type of sampling,
pm: probabilistic model.
the random search, we randomly selected 10000 solutions
and selected the best solution according to function f¯, λ =
0.01. The same experiment was repeated 100 times to select
the 100 “best” solutions for n ∈ {50, 100, 150, 200, 250}.
A similar procedure was followed for the greedy local search.
The local optimizer was applied to each of the 10000 solutions until no improvement was possible. For the GA, we
used the results of 100 GA runs used for the work published
in [12]. Since these results were obtained using solutions
of different length, and with a different number of evaluations, care must be taken to interpret the differences. We
only compare the GA results with the other algorithms for
n = 50. Similarly, the results of the random search search
were very poor for n > 50 and we only include them in the
comparison for n = 50. Results are shown in Figure 6. The
results of the comparison between the EDA and the greedy
search for n > 50 are shown in Figure 7. In the boxplots,
the central mark is the median, the edges of the box are the
25th and 75th percentiles, the whiskers extend to the most
extreme data points not considered outliers, and outliers are
plotted individually.
ε
tf
3
3
3
3
3
3
3
3
3
3
3
3
3
3
2
3
3
3
3
3
3
3
3
3
3
3
3
10
L
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
1
−1.4
−1.6
−1.8
−2
for n ∈ {150, 200} but in terms of the best solution found it
did not have an important influence for the other values of
n.
As a summary, we recommend to use an EDA that incorporates the greedy search, and uses partial sampling of type
II and the 1-order Markov model since it is less complex
than the tree and results achieved by the two models are
similar.
−2.2
−2.4
−2.6
Best EDA
Random
Greedy
Algorithms
GA
Figure 6: Comparison between the Best EDA variant, a random search, the greedy local search and
the GA for n = 50.
7.4 Improvement over other search methods
As a final validation of our method, we compare it with
the results achieved using a random search, the greedy local
optimizer, and the GA introduced in [12]. We compare our
best EDA variant as described in the previous section. For
It can be seen in Figures 6 and 7 that the EDA significantly outperforms all the other methods. Furthermore, as
n increases the algorithm is able to scale and find better
solutions.
−2.5
[2]
−3
log
10
ε
−3.5
−4
[3]
−4.5
−5
100
150
200
250
Best EDA
100
150
200
Greedy search
250
Figure 7: Comparison between the Best EDA variant and the greedy local search n ∈ {100, 150, 200, 250}.
8.
CONCLUSIONS
In this paper we have proposed the use of different EDA
variants for the quasiparticle braid problem. We have shown
that the fitness function and general evolutionary optimization approach initially introduced with GAs, can be successfully extended by the application of EDAs. The best braids
obtained with our EDAs have lengths up to 9 times shorter
than those expected from braids of the same accuracy obtained with the Solovay-Kitaev algorithm and had not been
previously reported to be found by the GA approach. We
have also proposed three different methods to enhance the
behavior of EDAs. Our results show that although the local
optimizer improves the results of the EDA, it is not able to
reach solutions of similar quality when applied alone. Decoding, and particularly partial sampling, can be used as
effective methods when dealing with other real-world problems as a way to improve usability of the representation and
avoid disrupting complex solutions, respectively.
By means of analyzing the Boltzmann distribution we
have shown that some of the problem characteristics are
translated into statistical regularities of the Boltzmann distribution. In the future we plan to extend this analysis to
try to extract more problem specific information that can be
useful for designing more effective search methods. Other
evolutionary algorithms that use models able to represent
higher order dependencies, such as Bayesian networks [11],
Markov networks [26], and factor graphs [6], could be applied. We also plan to address other braid problems of higher
difficulty.
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
9.
ACKNOWLEDGMENTS
R. Santana has been partially supported by the Saiotek
and Research Groups 2013-2018 (IT-609-13) programs (Basque
Government), TIN2013-41272P (Ministry of Science and Technology of Spain), COMBIOMED network in computational
bio-medicine (Carlos III Health Institute), and by the NICaiA
Project PIRSES-GA-2009-247619 (European Commission).
H. G. Katzgraber acknowledges support from the NSF (Grant
No. DMR-1151387).
[15]
[16]
10. REFERENCES
[1] S. Baluja. Population-based incremental learning: A
method for integrating genetic search based function
optimization and competitive learning. Technical
[17]
Report CMU-CS-94-163, Carnegie Mellon University,
Pittsburgh, PA, 1994.
S. Baluja and S. Davies. Using optimal
dependency-trees for combinatorial optimization:
Learning the structure of the search space. In D. H.
Fisher, editor, Proceedings of the 14th International
Conference on Machine Learning, pages 30–38, 1997.
N. E. Bonesteel, L. Hormozi, G. Zikos, and S. H.
Simon. Braid topologies for quantum computation.
Physical review letters, 95(14):140503, 2005.
C. M. Dawson and M. A. Nielsen. The Solovay-Kitaev
algorithm. arXiv preprint quant-ph/0505030, 2005.
J. S. De Bonet, C. L. Isbell, and P. Viola. MIMIC:
Finding optima by estimating probability densities. In
Mozer et al., editor, Advances in Neural Information
Processing Systems, volume 9, pages 424–430. The
MIT Press, Cambridge, 1997.
B. H. Helmi, A. T. Rahmani, and M. Pelikan. A factor
graph based genetic algorithm. Int. J. Appl. Math.
Comput. Sci, 24(3):621–633, 2014.
M. Henrion. Propagating uncertainty in Bayesian
networks by probabilistic logic sampling. In J. F.
Lemmer and L. N. Kanal, editors, Proceedings of the
Second Annual Conference on Uncertainty in Artificial
Intelligence, pages 149–164. Elsevier, 1988.
J. L. Hintze and R. D. Nelson. Violin plots: a box
plot-density trace synergism. The American
Statistician, 52(2):181–184, 1998.
L. Hormozi, G. Zikos, N. E. Bonesteel, and S. H.
Simon. Topological quantum compiling. Physical
Review B, 75(16):165310, 2007.
P. Larrañaga, H. Karshenas, C. Bielza, and
R. Santana. A review on probabilistic graphical
models in evolutionary computation. Journal of
Heuristics, 18(5):795–819, 2012.
P. Larrañaga and J. A. Lozano, editors. Estimation of
Distribution Algorithms. A New Tool for Evolutionary
Computation. Kluwer Academic Publishers, 2002.
R. B. McDonald and H. G. Katzgraber. Genetic braid
optimization: A heuristic approach to compute
quasiparticle braids. Physical Review B, 87(5):054414,
2013.
H. Mühlenbein and T. Mahnig. Theoretical Aspects of
Evolutionary Computing, chapter Evolutionary
algorithms: from recombination to search
distributions, pages 137–176. Springer, Berlin, 2000.
H. Mühlenbein, T. Mahnig, and A. Ochoa. Schemata,
distributions and graphical models in evolutionary
optimization. Journal of Heuristics, 5(2):213–247,
1999.
H. Mühlenbein and G. Paaß. From recombination of
genes to the estimation of distributions I. Binary
parameters. In Parallel Problem Solving from Nature PPSN IV, volume 1141 of Lectures Notes in Computer
Science, pages 178–187, Berlin, 1996. Springer.
H. Mühlenbein and D. Schlierkamp-Voosen. The
science of breeding and its application to the breeder
genetic algorithm (BGA). Evolutionary Computation,
1(4):335–360, 1994.
M. A. Nielsen and I. L. Chuang. Quantum
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
computation and quantum information. Cambridge
university press, 2010.
A. Ochoa and M. R. Soto. Linking entropy to
estimation of distribution algorithms. In J. A. Lozano,
P. Larrañaga, I. Inza, and E. Bengoetxea, editors,
Towards a New Evolutionary Computation: Advances
on Estimation of Distribution Algorithms, pages 1–38.
Springer, 2006.
M. Pelikan and D. E. Goldberg. Hierarchical BOA
solves Ising spin glasses and Max-Sat. IlliGAL Report
No. 2003001, University of Illinois at
Urbana-Champaign, January 2003.
M. Pelikan and H. G. Katzgraber. Analysis of
evolutionary algorithms on the one-dimensional spin
glass with power-law interactions. In Proceedings of
the 11th Annual conference on Genetic and
evolutionary computation, pages 843–850. ACM, 2009.
N. Read and E. Rezayi. Beyond paired quantum Hall
states: parafermions and incompressible states in the
first excited Landau level. Physical Review B,
59(12):8084, 1999.
R. Santana. Estimation of distribution algorithms
with Kikuchi approximations. Evolutionary
Computation, 13(1):67–97, 2005.
R. Santana, C. Bielza, and P. Larrañaga. Conductance
interaction identification by means of boltzmann
distribution and mutual information analysis in
conductance-based neuron models. BMC
Neuroscience, 13(Suppl 1):P100, 2012.
R. Santana, P. Larrañaga, and J. A. Lozano. Protein
folding in simplified models with estimation of
distribution algorithms. IEEE Transactions on
Evolutionary Computation, 12(4):418–438, 2008.
S. D. Sarma, M. Freedman, and C. Nayak. Topological
quantum computation. Physics Today, 59(7):32–38,
2006.
S. Shakya and R. Santana, editors. Markov Networks
in Evolutionary Computation. Springer, 2012.
S. Shakya, R. Santana, and J. A. Lozano. A
Markovianity based optimisation algorithm. Genetic
Programming and Evolvable Machines, 13(2):159–195,
2012.
N. van Kampen. Stochastic Processes in Physics and
Chemistry. North Holland, 1992.
H. Xu and X. Wan. Constructing functional braids for
low-leakage topological quantum computing. Physical
Review A, 78(4):042325, 2008.
| 9cs.NE
|
1
The Gaussian Interference Channel in the Presence
of Malicious Jammers
arXiv:1712.04133v1 [cs.IT] 12 Dec 2017
Fatemeh Hosseinigoki1 and Oliver Kosut1
Abstract
This paper considers the two-user Gaussian interference channel in the presence of adversarial
jammers. We first provide a general model including an arbitrary number of jammers, and show that
its capacity region is equivalent to that of a simplified model in which the received jamming signal at
each decoder is independent. Next, existing outer and inner bounds for two-user Gaussian interference
channel are generalized for this simplified jamming model. We show that for certain problem parameters,
precisely the same bounds hold, but with the noise variance increased by the received power of the
jammer at each receiver. Thus, the jammers can do no better than to transmit Gaussian noise. For these
problem parameters, this allows us to recover the half-bit theorem. In weak and strong interference
regime, our inner bound matches the corresponding Han-Kobayashi bound with increased noise variance
by the received power of the jammer, and even in strong interference we achieve the exact capacity.
Furthermore, we determine the symmetric degrees of freedom where the signal-to-noise, interferenceto-noise and jammer-to-noise ratios are all tend to infinity. Moreover, we show that, if the jammer has
greater received power than the legitimate user, symmetrizability makes the capacity zero. The proof of
the outer bound is straightforward, while the inner bound generalizes the Han-Kobayashi rate splitting
scheme. As a novel aspect, the inner bound takes advantage of the common message acting as common
randomness for the private message; hence, the jammer cannot symmetrize only the private codeword
without being detected. This complication requires an extra condition on the signal power, so that in
general our inner bound is not identical to the Han-Kobayashi bound. We also prove a new variation
of the packing lemma that applies for multiple Gaussian codebooks in an adversarial setting.
Index Terms: Gaussian interference channel, adversarial jammer, capacity region
This material is based upon work supported by the National Science Foundation under Grant No. CCF-1453718.
This paper was presented in part at the 54th Annual Allerton Conference on Communication, Control, and Computing, Allerton
2016 [1].
1
School
of
Electrical,
Computer
{fhossei1,okosut}@asu.edu
and
Energy
Engineering,
Arizona
State
University,
Tempe,
AZ
85287.
2
I. I NTRODUCTION
The open nature of the wireless communication medium makes it inherently vulnerable to an
active attack, wherein a malicious adversary (or jammer) transmits into the medium to disrupt the
operation of the legitimate users. Therefore, developing techniques to manage the presence of an
active attacker and to characterize the effect of an attacker on the fundamental limits of wireless
communication is important. In this paper, we investigate the two-user Gaussian interference
channel (GIC) in the presence of intelligent jammers (see Fig. 1) to characterize the capacity
region under the input and state constraints.
A. Prior Work
There are a wide variety of information-theoretic studies of active attacks in the literature
such as the arbitrarily-varying channel (AVC) [2]–[4], correlated jamming [5], network coding
with adversarial errors [6], [7], multiuser channels with correlated jamming [8], and a diamond
network in the presence of a jammer [9].
In [2], Ahlswede proved that the capacity of discrete memoryless AVC for deterministic codes
is either zero or equal to the random code capacity under the average probability of error criterion.
Ericson in [10] found a sufficient condition for the capacity to be zero, and later in [3] it is
proved that the condition is also necessary. Csiszar and Narayan further studied the deterministic
code capacity of the Gaussian AVC under the average probability of error criterion in [4] while
they also considered the input and state power constraints. Here, we mainly follow the Gaussian
AVC of [4] in which an adversary does not have any knowledge about the legitimate user’s
signal (but does know the code), and it may send an arbitrary sequence across the coding block
subject to a power constraint.
On the other hand, the interference channel (without a jammer) is one of the fundamental
problems in network information theory, and the exact capacity region is still unknown in general.
However, the Han-Kobayashi inner bound [11] is optimal or near-optimal for many interference
channels. The proof of this inner bound involves rate splitting, wherein each transmitter sends a
common message, decoded by both receivers, as well as a private message, decoded by only the
intended receiver. The authors in [12] showed that for the GIC, the Han-Kobayashi comes within
half a bit of the capacity region. Furthermore, [13] obtains an outer bound on the capacity region
of the GIC, and it is shown that for sufficiently weak interference signals, treating interference
3
as noise achieves the sum capacity. The deterministic interference channel model is proposed
by Brestler and Tse [14], and they show that the capacity of this channel is within a constant
number of bits of the corresponding GIC.
B. Main Results
Our main contribution is to generalize existing inner and outer bounds for the GIC in the
presence of AVC-style jammers. We provide a generalized GIC model with G jammers (G ≥ 1),
and show that the capacity region is equivalent to the capacity region of GIC with only two
independent jammers. We show that the capacity region depends only on the received power
of the jamming signal, not on the number of jammers. Moreover, we obtain the symmetric
degrees of freedom (DoF) by taking the limit of the normalized symmetric capacity as signal-tonoise, interference-to-noise and jammer-to-noise ratios converge to infinity. This characterization
generalized the so-called “W” DoF curve in [15, p. 153]. We also recover the optimal sum-rate
for the weak interference regime, as well as the exact capacity region for the strong interference
regime. We show that our bounds are within a half-bit in some regions, but we cannot show that
this holds in general.
We show that the outer bound in [12] holds with the noise variance increased by the received
power of the corresponding jammer at each receiver. The proof, given in Section IV, follows
by applying the outer bound in [12] with the jammers choosing to transmit Gaussian noise.
Moreover, we show that if the jammer’s received power at either receiver is larger than that
of the intended transmitter, AVC symmetrizability prevents this message from being decoded,
because the receiver cannot distinguish the legitimate codeword from the jammer’s counterfeit;
thus the capacity becomes zero.
We also provide a generalization of the Han-Kobayashi inner bound. For certain problem
parameters—for example, in the symmetric case when the jammer’s received power is less than
that of the interfering user—this inner bound is precisely the Han-Kobayashi inner bound with
the noise variance again increased by the received power of the jammer. Thus, for these problem
parameters we recover the half-bit theorem of [12], although we cannot prove that it holds in
general. The proof of the inner bound, given in Section V, is somewhat more involved, as the
receivers must decode correctly no matter what the jammer transmits. One novel aspect of our
inner bound proof is that we use the common message in the rate-splitting scheme as common
4
randomness for the private message. Thus, if the jammer has more power than the private
codeword but less than both together, it cannot use symmetrization without being detected, and
thus the receiver can decode.
In both [3] and [4], Csiszar and Narayan utilized lemmas (Lemma 3 in [3] and Lemma 1 in [4])
which assert the existence of codebooks with desirable properties in order to prove achievability
results. However, they only provided these lemmas for a single codebook, and for either discrete
random vectors (in [3]) or codewords uniformly distributed on the unit ball (in [4]). On the
other hand, our proof requires a variation on the Gaussian AVC packing lemma that handles
decoding of multiple superposed Gaussian codebooks. Our main technical tools for this goal are
Lemma 4, Lemma 5 and Lemma 6, proved in Appendix B, C and D, respectively.
The remainder of this paper is organized as follows. We describe the problem and the system
model for GIC with jammers in Section II. Our main results including the outer and the
inner bound for the capacity region are given by two theorems in Section III. We also discuss
implications of our bounds for different regimes, as well as illustrate numerical results in Section
IV. The proof for the outer bound is provided in Section V, and Section VI consists of the inner
bound proof. We conclude in Section VII.
Notations: We use bold letters to indicate the n-length vectors. Notation 1(·) refers to the
indicator function. We employ h·, ·i and k · k to denote inner product and norm-2, respectively.
We indicate the positive-part function and the expectation by |·|+ and E[·], respectively. Also, for
an integer N , notation [N ] stands for the set {1, 2, 3, . . . , N }. Notation In represents the identity
matrix of size n. Each of log(·) and exp(·) functions has base 2. Moreover, C(x) = 21 log(1 + x)
and ᾱ = 1 − α.
II. P ROBLEM S TATEMENT
The Gaussian interference channel with two independent jammers is shown in Fig. 1, in which
two users send their messages to their own receivers in the presence of one or two jammers.
The jammers are assumed not to have any information about the user’s signals (but know the
code). In particular, the received signals are given by
Y1 = h11 X1 + h12 X2 + g1 W1 + V1
(1)
Y2 = h21 X1 + h22 X2 + g2 W2 + V2
5
𝑾2
𝑾1
𝑔2
𝑿1
𝑔1
𝑽1
ℎ11
𝒀1
ℎ21
𝑿2
𝑽2
ℎ12
𝒀2
ℎ22
Fig. 1: Two-user Gaussian Interference Channel with Two Independent Jammers.
where X1 and X2 are n-length vectors representing the user’s signals, W1 and W2 are the
independent adversarial jammer signals, hij and gi for i, j ∈ {1, 2} are the channel gains, and
Vi is the n-length noise vector distributed as a sequence of i.i.d. zero mean Gaussian random
variables with variance σ 2 which is independent of X1 , X2 , W1 and W2 .
The transmitter and jammer signals are constrained to satisfy power constraints kXi k2 ≤ nPi
and kWi k2 ≤ nΛ, for i = 1, 2, respectively. We define the received signal-to-noise and interferenceto-noise ratios as S1 = h211 P1 /σ 2 , S2 = h222 P2 /σ 2 , I1 = h212 P2 /σ 2 and I2 = h221 P1 /σ 2 . We
also denote the jammer-to-noise ratios as J1 = g12 Λ/σ 2 and J2 = g22 Λ/σ 2 . We assume that the
transmitters and receivers know the signal-to-noise and interference-to-noise ratios, but they need
not know the jammer-to-noise ratios. However, we require small probability of error only when
the jammer-to-noise ratios do not exceed J1 ,J2 ; thus the code is independent of the jammer’s
power up to a point, and beyond that it may fail to decode correctly.
A 2nR1 , 2nR2 , n deterministic code is given by:
•
Message sets M1 = [2nR1 ] and M2 = [2nR2 ],
•
Encoding functions xi : Mi → Rn for i = 1, 2, and
•
Decoding functions φi : Rn → Mi for i = 1, 2.
For i = 1, 2, the message Mi is chosen uniformly from the set Mi , and each transmitter
encodes its own message to Xi . At each receiver, the received signal Yi is decoded by function
(n)
φi to M̂i = φi (Yi ). The average probability of error Pe
is now given by the probability that
(M̂1 , M̂2 ) 6= (M1 , M2 ), maximized over all possible choices of jammers’ sequences W1 and
W2 . A rate pair (R1 , R2 ) is achievable if there exists a sequence of 2nR1 , 2nR2 , n codes where
(n)
lim Pe = 0. The capacity region C is the closure of the set of all achievable rate pairs (R1 , R2 ).
n→∞
6
𝑾1
𝑾2
𝑔22
𝑿1
𝑔12
𝑔21
𝑔11
𝑽1
ℎ11
𝒀1
ℎ21
𝑿2
𝑽2
ℎ12
ℎ22
𝒀2
𝑔1𝐺
𝑔2𝐺
𝑾𝐺
Fig. 2: Two-user Gaussian Interference Channel with Two Independent Cross Jammers.
A. Generalized Jamming Model
Generally speaking, if there are G jammers (G ≥ 1) with the cross matrix G
g11 g12 g13 . . . g1G
G =
g21 g22 g23 . . . g2G
(2)
as shown in Fig. 2, then the received signals at each decoder are given by
Y1 = h11 X1 + h12 X2 + g11 W1 + g12 W2 + . . . + g1G WG + V1
(3)
Y2 = h21 X1 + h22 X2 + g21 W1 + g22 W2 + . . . + g2G WG + V2
where kWi k2 ≤ nΛ for i = 1, 2, . . . , G. This includes the case where there is only one jammer
(G = 1). We refer the capacity region of this channel as CG , and state the following proposition
for the relation between CG and C . Indeed, the capacity region depends only on the received
signal at each decoder and not the number of jammers.
Proposition 1: We have CG = C as long as
|g11 | + |g12 | + . . . + |g1G | = |g1 |
(4)
|g21 | + |g22 | + . . . + |g2G | = |g2 |
where the jammer-to-noise ratios are then given by J1 = (g11 + g12 + . . . + g1G )2 Λ/σ 2 and
J2 = (g21 + g22 + . . . + g2G )2 Λ/σ 2 .
The proof is provided in Appendix A.
7
III. M AIN R ESULTS
In this section, we present inner and outer bounds on the capacity region C (two-user GIC with
two independent jammers). Before stating the main results, we define regions Ro (S1 , S2 , I1 , I2 )
and Ri (S1 , S2 , I1 , I2 ) as the previously-derived outer and inner bounds respectively for the GIC
with no jammer; namely Ro is the outer bound of [12], and Ri is the Han-Kobayashi inner bound
[11]. When we write an expression with i and j, we mean for it to hold for both (i, j) = (1, 2)
and (i, j) = (2, 1).
Define Ro (S1 , S2 , I1 , I2 ) as the set of rate pairs (R1 , R2 ) such that
Ri ≤ C (Si )
Si
+ C (Ij + Sj )
Ri + Rj ≤ C 1+I
j
S2 +I2 +I1 I2
S1 +I1 +I1 I2
+C
R1 + R2 ≤ C
1+I2
1+I1
Sj +Ij +Ii Ij
Si
.
2Ri + Rj ≤ C 1+Ij + C (Si + Ii )+ C
1+Ii
Define Ri (S1 , S2 , I1 , I2 ) as the set of rate pairs (R1 , R2 ) such that
Si
Ri < C 1+α
j Ii
Si +ᾱj Ii
αj Sj
Ri + Rj < C 1+α
+
C
1+αi Ij
j Ii
α1 S1 +ᾱ2 I1
α2 S2 +ᾱ1 I2
R1 + R2 < C
+C
1+α2 I1
1+α1 I2
Si +ᾱj Ii
αj Sj +ᾱi Ij
αi Si
2Ri + Rj < C 1+αj Ii +C 1+αj Ii +C
1+αi Ij
for some αi in [0, 1] where αi implies the portion of the private message power in the HanKobayashi inner bound proof at user i. Note that in the Han-Kobayashi inner bound proof encoder
i divides the message mi into private message mip and common message mic with power αi Pi
and ᾱi Pi respectively.
Define Si0 =
Si
1+Ji
and Ii0 =
Ii
.
1+Ji
We now state our main outer and inner bounds.
Theorem 2 (Outer Bound): C ⊆ Ro (S10 , S20 , I10 , I20 ). Moreover, if S1≤ J1 or S2≤ J2 , then C = ∅.
Theorem 3 (Inner Bound): Assume Si > Ji for i = 1, 2. Let R̃i (S10 , S20 , I10 , I20 ) be the subset
of rate pairs in Ri (S10 , S20 , I10 , I20 ) achieved by αi ∈ [0, 1] satisfying
αi Si + ᾱj Ii > Ji for (i, j) = (1, 2), (2, 1).
Then R̃i (S10 , S20 , I10 , I20 ) ⊆ C .
(5)
8
(a) S = 4, I = 3, and J between 0 and 5. For these parameters the bound Ri (S 0 , S 0 , I 0 , I 0 )
is identical to our inner bound if the jammer-to-noise ratio is less than 3.2.
(b) S = 4, J = 3.5, and I between 0 and 10. The bound
(c) S = 10, J = 3.5, and I between 0 and 40. The bound
Ri (S , S , I , I ) is identical to our inner bound for weak
Ri (S 0 , S 0 , I 0 , I 0 ) is identical to our inner bound for high
and strong interference.
signal-to-noise ratio S = 10.
0
0
0
0
Fig. 3: Bounds on the symmetric capacity Csym (S, I, J) for S1 = S2 = S, I1 = I2 = I, and
J1 = J2 = J. In addition to our inner R̃i (S 0 , S 0 , I 0 , I 0 ) and outer bounds Ro (S 0 , S 0 , I 0 , I 0 ), also
shown the bound Ri (S 0 , S 0 , I 0 , I 0 ) and shown Ri (S 0 , S 0 , I 0 , I 0 ) with sub-optimal α =
1
.
1+I 0
9
IV. D ISCUSSION AND N UMERICAL R ESULTS
Note that the inner bound differs from Ri (S10 , S20 , I10 , I20 ) only when the optimal αi parameters
do not satisfy (5). However, in several regimes of interest, this constraint is not active. For
q 0
Ij
example, if the channel has weak interference, in the sense of [15, eq (6.8)]
(1 + Ii0 ) ≤
S0
i
ρi (1 − ρj ) for some ρ1 , ρ2 ∈ [0, 1] and (i, j) = (1, 2), (2, 1) then treating interference as noise
is optimal for the sum-rate [15, Theorem 6.3]. Treating interference as noise corresponds to
α1 = α2 = 1, under which (5) holds. Therefore, in the weak interference regime, our inner
bound matches Ri (S10 , S20 , I10 , I20 ), and it also achieves the exact sum-rate capacity. On the other
hand, when the channel has strong interference in both users I20 ≥ S10 and I10 ≥ S20 , by choosing
α1 = α2 = 0 each transmitter only sends its own common message, and both messages can be
decoded at each receiver. Therefore, (5) holds if we have I1 > J1 and I2 > J2 . Thus, we obtain
the exact capacity region for the strong interference regime [15, Theorem 6.2].
In [15, Theorem 6.6], it is shown that using Han-Kobayashi inner bound with sub-optimal
choices α1 =
1
1+I20
and α2 =
1
1+I10
outer bound. Therefore, if α1 =
yields an inner bound that is always within half a bit of the
1
1+I20
and α2 =
1
1+I10
satisfy our conditions in (5) then our inner
bound is guaranteed to be within half a bit of our outer bound; that is, if J1 <
J2 <
S2
1+I10
+
I22
1+I20
S1
1+I20
+
I12
1+I10
and
, our inner and outer bounds are within half a bit.
Now, consider the symmetric case; i.e. S1 = S2 = S, I1 = I2 = I, J1 = J2 = J, and
R1 = R2 = R. Clearly in this case it is optimal to choose α1 = α2 = α for inner bound. Define
the symmetric capacity of the channel as Csym (S, I, J) = max{R : (R, R) ∈ C }. We illustrate
the bounds for Csym (S, I, J) in Fig. 3 including our outer bound Ro (S 0 , S 0 , I 0 , I 0 ), our inner
bound R̃i (S 0 , S 0 , I 0 , I 0 ) with optimal α, the Han-Kobayashi inner bound with the noise variance
increased by the received power of the jammer Ri (S 0 , S 0 , I 0 , I 0 ) and the latter with sub-optimal
α=
1
.
1+I 0
Note that we are in the strong interference regime only if I 0 ≥ S 0 .
Define the normalized symmetric capacity as dsym =
Csym (S,I,J)
.
C(S)
Then the symmetric degrees
of freedom (DoF) d∗sym is given by
d∗sym (β, δ)
Csym (S, S β , S δ )
= lim
.
S→∞
C(S)
(6)
By substituting I = S β and J = S δ in our outer and inner bounds Ro (S 0 , S 0 , I 0 , I 0 ) and
R̃i (S 0 , S 0 , I 0 , I 0 ), we find the upper bound for Csym (S, S β , S δ ) given by
10
(a) DoF for δ =
1
4
and β between 0, 2.
(b) DoF for β = 0.7 and δ between 0, 2.
Fig. 4: Symmetric degrees of freedom for the GIC.
Csym (S, S β , S δ ) ≤ max{R : (R, R) ∈ Ro (S 0 , S 0 , I 0 , I 0 )} =
2β
S+S β + S δ
S
S
S + Sβ
1
1
1+S
min C
, C
+ C
,C
,
1+S δ +S β
1 + Sδ
2
1 + Sδ + Sβ
2
1 + Sδ
2β
S+S β + S δ
S
S + Sβ
1
1
1
1+S
C
+ C
+ C
, (7)
1+S δ +S β
3
1 + Sδ + Sβ
3
1 + Sδ
3
and the lower bound for Csym (S, S β , S δ ) given by
Csym (S, S β , S δ ) ≥ max{R : (R, R) ∈ R̃i (S 0 , S 0 , I 0 , I 0 )} =
S + ᾱS β
αS
S
1
1
max
min C
, C
+ C
,
1 + S δ + αS β
2
1 + S δ + αS β
2
1 + S δ + αS β
α:αS+ᾱS β >S δ
αS + ᾱS β
1
S + ᾱS β
1
αS
1
αS + ᾱS β
, C
+ C
+ C
.
C
1 + S δ + αS β
3
1 + S δ + αS β
3
1 + S δ + αS β
3
1 + S δ + αS β
(8)
We may further lower bound the symmetric capacity by choosing α =
1
1+I 0
=
1+S δ
1+S δ +S β
as long
as this choice satisfies (5). In particular, we claim that this value of α always satisfies (5) for
sufficiently large S. We show this by substituting this value of α to find
1 + Sδ
Sβ
αS + ᾱS =
S+
Sβ
δ
β
δ
β
1+S +S
1+S +S
1+δ
2β
S+S
+S
=
.
δ
1 + S + Sβ
β
(9)
(10)
11
Since the capacity region is empty when S ≤ J = S δ , it suffices to consider (10) only for δ < 1.
The dominant power of S in (10) is given by
max{1 + δ, 2β} − max{δ, β} > max{2δ, 2β} − max{δ, β} = max{δ, β} ≥ δ.
Therefore, (10) is larger than J = S δ for sufficiently large S, thus α =
1
1+I 0
(11)
satisfies (5). Now,
we may substitute this choice of α into (8), and take the limits of (8) and (7) as S → ∞.
Therefore, we find that the symmetric DoF is given by
β δ β δ
∗
dsym (β, δ) = min max{0, 1 − δ}, max {0, 1 − β, β − δ} , max 0, 1 − − , −
2 2 2 2
(12)
which is illustrated for fixed δ = 1/4 in Fig. 4a and fixed β = 0.7 in Fig. 4b. Note that for
the interference channel with no jammer [15, p. 153], the DoF exhibits a “W” shape for a fixed
δ=
log J
.
log S
V. P ROOF OF O UTER B OUND
Consider a sequence of (2nR1 , 2nR2 , n) codes with vanishing probability of error. Since these
codes must function for arbitrary jamming signals, we may derive an outer bound by assuming
the jammers transmit Gaussian noise with variance Λ. Thus, we follow the outer bound for the
GIC with no jammer [15, Chapter 6.7.2, p. 151] and the noise power σ 2 + gi2 Λ. This yields the
outer bound Ro (S10 , S20 , I10 , I20 ).
Moreover, if J1 ≥ S1 , based on the assumption that the jammer knows the code, the jammer
can choose an arbitrary message m̃1 and transmit a scaled form of the corresponding codeword
w1 = x1 (m̃1 )h11 /g1 . Given Y1 = h11 x1 (m1 ) + h12 x2 (m2 ) + h11 x1 (m̃1 ) + V1 , decoder 1 cannot
decode the message since it does not know whether the true message is m1 or m̃1 . The same
scenario can happen for decoder 2 if J2 ≥ S2 . This attack constitutes AVC symmetrization.
VI. P ROOF OF I NNER B OUND
Our inner bound proof is a generalization of the Han-Kobayashi bound [15, Chapter 6.5.1,
p. 144]. Using rate splitting, we represent message mi from user i for i = 1, 2, by independent
common message mic at rate Ric and private message mip at rate Rip such that Ri = Ric + Rip .
Thus, each receiver will decode its own common and private messages and the common message
12
of the interfering user. Assuming Si > Ji for i = 1, 2, we show that (R1c , R1p , R2c , R2p ) is
achievable if
Rip < C
αi Si
1+Ji +αj Ii
Rip + Ric < C
Si
1+Ji +αj Ii
Rip + Rjc < C
αi Si +ᾱj Ii
1+Ji +αj Ii
Rip + Ric + Rjc < C
Si +ᾱj Ii
1+Ji +αj Ii
(13)
for some αi ∈ [0, 1] satisfying αi Si + ᾱj Ii > Ji , and again the above holds for (i, j) = (1, 2) and
(i, j) = (2, 1). This achieves the region R̃i by substituting R1 = R1c + R1p and R2 = R2c + R2p ,
and applying the Fourier-Motzkin procedure to eliminate Ric and Rip .
Before proceeding to the proof, we first define the following typical set for Gaussian random
variables X1 , . . . , Xk as:
T(n) (X1 , . . . , Xk )
1
= (x1 , . . . , xk ) : E(Xi Xj ) − ≤ hxi , xj i ≤ E(Xi Xj ) + for all i, j ∈ [1 : k] . (14)
n
Codebook generation: Fix α1 , α2 ∈ [0, 1] and γ > 0. For i = 1, 2, we generate 2nRic i.i.d
zero mean Gaussian sequences Xic (mic ) with variance (1 − γ)ᾱi Pi for each mic ∈ [2nRic ]. Also,
for each mic ∈ [2nRic ], generate 2nRip i.i.d. zero mean Gaussian sequences Xip (mic , mip ) with
variance (1 − γ)αi Pi for each mip ∈ [2nRip ] for i = 1, 2.
Encoding: For i = 1, 2, write message mi as (mic , mip ) where mic ∈ [2nRic ] and mip ∈ [2nRip ].
Transmitter i sends Xi = Xic (mic )+Xip (mic , mip ) if its power is less than Pi , otherwise it sends
zero.
Decoding: We describe the decoding procedure for receiver 1; that of receiver 2 is similar.
First, let
n
o
[
S = (m1c , m1p , m2c ) : (x1c (m1c ), x1p (m1c , m1p ), x2c (m2c ), y1 ) ∈
T(n) (X1c , X1p , X2c , Y1 )
(15)
where the union is over all joint Gaussian distributions X1c , X1p , X2c , Y1 such that (X1c , X1p , X2c ,
Y1 − h11 X1c − h11 X1p − h12 X2c ) are mutually independent.
13
Given y1 , decoder 1 finds
(m̂1c , m̂1p , m̂2c ) =
arg min
ky1 − h11 x1c (m1c ) − h11 x1p (m1c , m1p ) − h12 x2c (m2c )k .
(m1c ,m1p ,m2c )∈S
(16)
If there is more than one minimum, choose between them arbitrarily. The decoder then outputs
the message estimate m̂1 = (m̂1c , m̂1p ).
Analysis of the probability of error: Assume the two users send messages (M1c , M1p ), (M2c ,
M2p ) . We will obtain the average probability of error for decoder 1 and similarly generalize
the results for decoder 2. Define the error event
E0 = {(M1c , M1p , M2c ) ∈
/ S}.
(17)
To consider error events in which a false message set appears correct, we define the set
T = (m1c , m1p , m2c ) ∈ S : kY1 − h11 x1c (m1c ) − h11 x1p (m1c , m1p ) − h12 x2c (m2c )k2
≤ kY1 − h11 x1c (M1c ) − h11 x1p (M1c , M1p ) − h12 x2c (M2c )k2 . (18)
An error can only occur if there exists some (m1c , m1p , m2c ) ∈ T where (m1c , m1p ) 6= (M1c , M1p ).
We divide this event into the following 4 error events:
E1 = {∃ m̃1p 6= M1p : (M1c , m̃1p , M2c ) ∈ T }
(19)
E2 = {∃ m̃1c 6= M1c , m̃1p : (m̃1c , m̃1p , M2c ) ∈ T }
(20)
E3 = {∃ m̃1p 6= M1p , m̃2c 6= M2c : (M1c , m̃1p , m̃2c ) ∈ T }
(21)
E4 = {∃ m̃1c 6= M1c , m̃1p , m̃2c 6= M2c : (m̃1c , m̃1p , m̃2c ) ∈ T } .
(22)
We will prove that the probability of each one of the error events converges to zero as long as
the conditions in (13) are satisfied.
To bound these error events, we state two versions of a Gaussian AVC packing lemma. The
basic approach to these lemmas originates in [3, Lemma 3] and [4, Lemma 1], and the proof
is most similar to that of [3, Lemma 3]. These earlier lemmas showed that a single random
codebook satisfies several desirable properties with high probability. Here, we need to show that
multiple codebooks simultaneously satisfy desirable properties; thus we need a slightly more
general approach. Furthermore, we use Gaussian codewords instead of codewords uniformly
14
distributed on the unit ball. The advantage of Gaussian codewords is that superpositions of
codewords are themselves Gaussian, and we here are dealing with the summation of more than
one codeword. Lemma 5 shows that with high probability, two superposed Gaussian codebooks
yield small probability of error. While the result is stated for two codebooks for simplicity, it
applies for any number of codebooks, and so it will be used to bound events E2 , E3 , and E4 .
Note that Lemma 5 requires Λ < 1; i.e. the jammer’s power must be less than the codeword
power, which is necessary to avoid symmetrization. This requirement leads to the conditions
that Si > Ji and αi Si + ᾱj Ii > Ji . The latter, originating from event E3 , constitutes the main
difference between our inner bound and the Han-Kobayashi inner bound, but we have been
unable to eliminate it.
Lemma 4 differs from Lemma 5 in that it focuses on just one codebook, but takes into account
common randomness at the encoder and decoder. This lemma is used to bound event E1 , because
in this error event, the common message can be used as common randomness. The advantage
of AVC coding with common randomness is that it is not susceptible to symmetrization. Thus,
in Lemma 4 there is no requirement that Λ < 1.
Lemmas 4, 5 and 6 are proved in Appendix B, C and D respectively.
Lemma 4: Fix σ 2 , Λ ≥ 0, N = 2nR and K ≥ n2 . Let Xi (k) for i = 1, . . . , N , k = 1, . . . , K
be independent zero mean Gaussian random vectors with covariance matrix In . Let R satisfy
R < C(1/(Λ + σ 2 )). Define
p2 (x1 (1), . . . , x1 (K), x2 (1), . . . , x2 (K), . . . , xN (1), . . . , xN (K)|w) =
N
K
1 XX
P ∃j 6= i : kxi (k)+w+V−xj (k)k2 ≤ kw+Vk2, xi (k) ∈ T(n) (X), xj (k) ∈ T(n) (X)
N K i=1 k=1
(23)
where V is Gaussian noise distributed as V ∼ N (0, σ 2 In ). There exists ρ > 0 such that
"
#
lim P
n→∞
sup
p2 (X1 (1), . . . , X1 (K), . . . , XN (1), . . . , XN (K)|w) ≤ ρ = 0.
(24)
w:||w||2 ≤nΛ
Lemma 5: Fix θ ∈ [0, 1], σ 2 , Λ ≥ 0, N1 = 2nR1 and N2 = 2nR2 . Let X1 , . . . , XN1 ∈ Rn
and Y1 , . . . , YN2 ∈ Rn be independent zero mean Gaussian random vectors (codebooks) with
covariance matrices θ In and θ̄ In , respectively. Let Λ, R1 , R2 satisfy
15
Λ < 1,
(25)
R1 < C
θ
Λ + σ2
,
θ̄
,
R2 < C
Λ + σ2
1
R1 + R2 < C
.
Λ + σ2
(26)
(27)
(28)
(n)
Let U be the set of triples (x, y, z) ∈ T (X, Y, Z) for some Gaussian triple (X, Y, Z) where
EX 2 = θ,
EY 2 = θ̄,
(X, Y, Z) are mutually independent.
(29)
Define
p1 (x1 , . . . , xN1 , y1 , . . . , yN2 |w) =
N1 X
N2
1 X
P ∃j1 6= i1 , j2 6= i2 : kxi1 + yi2 + w + V − xj1 − yj2 k2 ≤ kw + Vk2 ,
N1 N2 i =1 i =1
1
2
(xi1 , yi2 , w + V) ∈ U, (xj1 , yj2 , xi1 + yi2 + w + V − xj1 − yj2 ) ∈ U
(30)
where V is Gaussian noise distributed as V ∼ N (0, σ 2 In ). There exists ρ > 0 such that
"
#
lim P
n→∞
sup
w:||w||2 ≤nΛ
p1 (X1 , . . . , XN1 , Y1 , . . . , YN2 |w) ≤ exp(−nρ) = 0.
(31)
Since we are dealing with more than one Gaussian codeword in this paper, we need a new
version of [3, Lemma 3] and [4, Lemma 1] not only for Gaussian vectors, but also for multiple
codebooks. It did not appear possible to use the properties derived from these lemmas on each
codebook individually; instead, we must prove a new lemma establishing joint properties among
more than one codebook. This new lemma (Lemma 6) provides the main properties that the
Gaussian codebooks need to satisfy in order to bound the probability of error event E0 by (33)
and as part of the proof of Lemma 5 by (34)-(36).
Lemma 6: Fix θ ∈ [0, 1], N1 = 2nR1 and N2 = 2nR2 . Given any random variables X, Y, W ,
define the quantity
JX;Y ;W (R1 , R2 ) = max 0, R1 −I(X; W Y ), R2 −I(Y ; W X), R1 +R2 −I(XY ; W )−I(X; Y ) .
(32)
16
Let Xi1 and Yi2 be Gaussian i.i.d. n-length random vectors (codebooks) independent from each
other with zero mean and Cov(Xi1 ) = θ In , Cov(Yi2 ) = θ̄ In where i1 ∈ {1, 2, . . . , N1 } and
i2 ∈ {1, 2, . . . , N2 }. With probability approaching 1 as n → ∞, they satisfy the following, for any
x, y, w where kwk2 ≤ nΛ and any zero mean jointly Gaussian random vector (X, Y, X 0 , Y 0 , W )
with positive definite covariance matrices with diagonals at most (θ, θ̄, θ, θ̄, Λ).
[
1
(n)
(i1 , i2 ) : (xi1 , yi2 , w) ∈
/
T (X, Y, W ) ≤ exp(−nδ()). (33)
N1 N2
(X,Y,W ) mutually independent:
EX 2 =θ,EY 2 =θ̄,EW 2 ≤Λ
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
≤ exp n JX 0 ;Y 0 ;W (R1 , R2 ) + δ()
(34)
(i1 , i2 ) : (x, y, xi1 , yi2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) ≤ exp n JX 0 ;Y 0 ;XY W (R1 , R2 ) + δ()
(35)
1
(i1 , i2 ) : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some j1 6= i1 , j2 6= i2
N1 N2
≤ 8 exp{−nδ()/4}
if JX 0 ;Y 0 ;W (R1 , R2 ) ≤ I(XY ; X 0 Y 0 W ) − 2δ().
(36)
(n,k)
We now consider each of the five error events, beginning with E0 . Define set T
S
(n)
T (X1 , . . . , Xk )
as
over all joint Gaussian distributions X1 , . . . , Xk such that (X1 , . . . , Xk )
are mutually independent. For every > 0 , we have
P(E0 ) = P{(M1c , M1p , M2c ) ∈
/ S}
n
o
[
(n)
= P (x1c (M1c ), x1p (M1c , M1p ), x2c (M2c ), Y1 ) ∈
/
T (X1c , X1p , X2c , Y1 )
= P (x1c , x1p , x2c , h12 x2p + g1 w1 + V1 ) ∈
/ T(n,4)
≤ P (x1c , x1p , x2c , x2p , w1 , V1 ) ∈
/ T(n,6)
n
o
(n,5)
≤ P (x1c , x1p , x2c , x2p , w1 ) ∈
/ T0
n
o
(n,5)
+ P (x1c , x1p , x2c , x2p , w1 , V1 ) ∈
/ T(n,6) (x1c , x1p , x2c , x2p , w1 ) ∈ T0
(37)
(38)
(39)
(40)
(41)
where the union in (38) is exactly the one in set S definition (15). (40) follows because if
(x1c , x1p , x2c , x2p , w1 , V1 ) is typical for independent distributions then (x1c , x1p , x2c , h12 x2p +
g1 w1 +V1 ) would be typical. The probabilities in (41) follows from the fact that if (x1c , x1p , x2c ,
17
(n,5)
x2p , w1 ) ∈
/ T0
(n,6)
then (x1c , x1p , x2c , x2p , w1 , V1 ) ∈
/ T
. Finally, as n → ∞ the first term in
(41) vanishes exponentially by using the general version of Lemma 6-(33), and the second term
in (41) tends to zero by using conditional typicality lemma (see (70) in appendix C). Then,
P(E0 ) tends to zero as n → ∞.
To bound the probability of event E1 , we apply Lemma 4 with the following:
•
i = m1p , j = m̃1p , k = m1c ,
•
11 x1p (m1c ,m1p )
,
xi (k) = h√
2
•
11 x1p (m1c ,m̃1p )
,
xj (k) = h√
2
•
V=
•
w=√
(1−γ)α1 σ S1
(1−γ)α1 σ S1
h12 x2p (M2c ,M2p )+V1
√
(1−γ)α1 σ 2 S1
g1 w1
(1−γ)α1 σ 2 S1
,
.
In this case, K = 2nR1c ≥ n2 for sufficiently large n as long as R1c > 0. Note that event E1
occurs if
kh11 X1p (m1c , m1p ) + g1 w1 + h12 X2p (M2c , M2p ) + V1 − h11 X1p (m1c , m̃1p )k2
Thus, by Lemma 4, if R1p
≤ kg1 w1 + h12 X2p (M2c , M2p ) + V1 k2 . (42)
(1−γ)α1 S10
(1−γ)α1 S1
< C 1+J1 +(1−γ)α2 I1 = C 1+(1−γ)α2 I 0 , then with high probability
1
the codebook X1p will be such that P(E1 ) → 0 as n → ∞.
We now bound event E2 by applying Lemma 5 with the following particularizations:
•
i1 = m1c , i2 = (m1c , m1p ), j1 = m̃1c , j2 = (m̃1c , m̃1p ),
•
h11 x1c (m1c )
1p )
x i1 = √
, yi2 = h11√x1p (m1c ,m
,
2
2
•
h11 x1c (m̃1c )
xj1 = √
, yj2 = h11√x1p (m̃1c ,2m̃1p ) ,
2
•
V=
•
w=√
(1−γ)σ S1
(1−γ)σ S1
(1−γ)σ S1
h12 x2p (M2c ,M2p )+V1
√
(1−γ)σ 2 S1
g1 w1
(1−γ)σ 2 S1
.
(1−γ)σ S1
,
18
Note that event E2 occurs if
kh11 X1c (m1c )+h11 X1p (m1c , m1p )+w+V−h11 X1c (m̃1c )−h11 X1p (m̃1c , m̃1p )k2 ≤ kw+Vk2 .
Therefore, we can conclude by Lemma 5 that with high probability as n → ∞, P(E2 ) → 0 if
J1
S1
< 1,
R1c < C
(1−γ)ᾱ1 S1
1+J1 +(1−γ)α2 I1
(43)
R1p < C
(1−γ)α1 S1
1+J1 +(1−γ)α2 I1
(44)
R1c + R1p < C
(1−γ)S1
1+J1 +(1−γ)α2 I1
.
(45)
Similarly, the probability of event E3 can be bounded using Lemma 5 as long as we have
α1 S1 + ᾱ2 I1 > J1 , (44),
ᾱ2 I1
R2c < C 1+J(1−γ)
1 +(1−γ)α2 I1
1 S1 +ᾱ2 I1 )
R1p + R2c < C (1−γ)(α
.
1+J1 +(1−γ)α2 I1
(46)
(47)
Finally, the probability of event E4 may be bounded using Lemma 5 under the conditions S1 +
ᾱ1 S1 +ᾱ2 I1 )
and R1c + R1p + R2c <
ᾱ2 I1 > J1 , (43), (44), (45), (46), (47), R1c + R2c < C (1−γ)(
1+J1 +(1−γ)α2 I1
(1−γ)(S1 +ᾱ2 I1 )
. Note that in this case we use a version of Lemma 5 for three independent
C 1+J
1 +(1−γ)α2 I1
codebooks rather than two. We finally get all equations in (13) as γ → 0.
VII. C ONCLUSION
We provided inner and outer bounds for the capacity region of the Gaussian interference
channel in the presence of malicious jammers. We showed that if input power is less than the
jammers’ power, then the capacity region is zero. Otherwise, both our inner and outer bounds
correspond to existing bounds with the noise variance increased by the jammers’ received power,
except that our inner bound has an additional constraint compared to the classical Han-Kobayashi
bound. These bounds allowed us to recover most existing results for the GIC with no jammer,
with the exception of the half-bit gap, which we have been unable to prove in general. These
results include the symmetric DoF of our channel as signal-to-noise ratio tends to infinity along
with the interference-to-noise ratio and the jammer-to-noise ratio. To prove our achievability
result, we employed three novel lemmas, including new versions of the packing lemma for
multiple Gaussian codebooks. Extensions to more than two users should be straightforward. We
19
believe that our new packing lemmas may be applicable to other network information theory
problems in the presence of a malicious intruder.
A PPENDIX A
P ROOF OF P ROPOSITION 1
First, we prove that C ⊆ CG . Consider rate pair (R10 , R20 ) ∈ C meaning that there exists a
0
0
2nR1 , 2nR2 , n code that yields an arbitrary small probability of error for any possible adversary
action of two independent jammers. Now, we must show that this same code will also work for
G jammers. In the G jammer model, let w1 , . . . , wG be any jamming signals. We may define
equivalent jamming signals for the model with two independent jammers as
w10 =
g11 w1 + g12 w2 + . . . + g1G wG
g1
w20 =
g21 w1 + g22 w2 + . . . + g2G wG
.
g2
and
Note that the received signal in the G jammer model is identical to that in the 2 jammer
model with jamming signals w10 , w20 . Moreover, in order to show that w10 and w20 satisfy power
constraints, we have
kwi0 k2 =
kgi1 w1 + gi2 w2 + . . . + giG wG k2
|gi |2
(|gi1 |kw1 k + |gi2 |kw2 k + . . . + |giG |kwG k)2
≤
|gi |2
√
√
√ 2
|gi1 | nΛ + |gi2 | nΛ + . . . + |giG | nΛ
≤
|gi |2
(48)
(49)
(50)
(|gi1 | + |gi2 | + . . . + |giG |)2
≤ nΛ
|gi |2
(51)
= nΛ
(52)
where i = 1, 2, and 52 follows from the assumption in the proposition. Therefore, the probability
of error under the G jammer model is at most that of the 2 jammer model.
Now, we prove CG ⊆ C . Let rate (R1 , R2 ) ∈ CG . Therefore, there exists a sequence of
2nR1 , 2nR2 , n code that has arbitrary small probability of error PeG for any possible adversary
20
actions with G jammers. Now, if we use this code for two-jammer scenario Fig. 1, the probability
of error at decoder i = 1, 2 is given as
Pei = max
P(M̂i 6= Mi )
0
(53)
wi
=
max
w1 ,w2 ,...,wG
P(M̂i 6= Mi )
(54)
where the last equality follows because of the same power constraints kwj k2 ≤ nΛ and kwi0 k2 ≤
nΛ (52) for j = 1, 2, . . . , G and i = 1, 2 in both models meaning that the set of received jammer
signals gi1 w1 + gi2 w2 + . . . + giG wG or gi wi0 are identical at each decoder. Note that we have
the assumption |gi1 | + |gi2 | + . . . + |giG | = |gi | for i = 1, 2. By the equivalent expression for Pei
in (54), the probability of error in the G jammer model can be lower bounded by
PeG ≥ max{Pe1 , Pe2 }.
(55)
In addition, the overall probability of error is upper bounded by
Pe(n) ≤ Pe1 + Pe2 .
(56)
Since PeG → 0, from (55) both Pe1 and Pe2 must tend to zero, too. Therefore, the sum in (56)
(n)
also tends to zero, so the overall probability of error with two jammers Pe
vanishes as n → ∞.
A PPENDIX B
P ROOF OF L EMMA 4
We prove this lemma using a random code reduction, as in [16, Lemma 12.8]. We first show
that a Gaussian codebook independent of the jammer’s signal achieves small probability of error,
and then we show that a finite number of deterministic codebooks achieve essentially the same
probability.
Let X1 , . . . , XN be Gaussian random vectors with zero mean and covariance In . We will
prove that, for any i ∈ [N ] and any w such that kwk2 ≤ nΛ
P ∃j 6= i : kXi + w + V − Xj k2 ≤ kw + Vk2 → 0
(57)
as n → ∞, where V ∼ N (0, σ 2 In ). To prove this, we adopt the basic approach of [17]. In
particular, let Z = w + V, and let U be a unitary matrix that maps Z to (kZk, 0, . . . , 0). Then
we may write
P ∃j 6= i : kXi + Z − Xj k2 ≤ kZk2 = P ∃j 6= i : kU Xi + U Z − U Xj k2 ≤ kU Zk2
(58)
21
= P ∃j 6= i : kXi + U Z − Xj k2 ≤ kU Zk2
(59)
where (59) follows from the spherical symmetry of the codebook distribution. Now if we define,
for any Σ > 0,
√
e(Σ) = P{∃j 6= i : kXi + ( nΣ, 0, . . . , 0) − Xj k2 ≤ nΣ}
(60)
then the probability in (59) may be written as Ee(kZk2 ). Note that for any δ,
lim P kZk2 − n(σ 2 + Λ) > nδ} → 0.
(61)
n→∞
Moreover, e(Σ) is non-decreasing in Σ. Thus, for any δ > 0, if we let Ṽ ∼ N (0, σ 2 + Λ + δ), for
sufficiently large n we have Ee(kZk2 ) ≤ Ee(kṼk2 ). Now, Ee(kṼk2 ) is simply the probability of
error for a Gaussian channel with noise variance σ 2 + Λ + δ. Since Gaussian codebooks achieve
capacity for Gaussian channels with minimum distance decoding, this quantity vanishes with n
as long as
R<C
1
2
σ +Λ+δ
(62)
which holds for small enough δ by the assumption that R < C( σ21+Λ ). This proves (57).
Now let Xi (k) for i = 1, . . . , N and k = 1, . . . , K be independent Gaussian vectors with zero
mean and covariance In . For any i ∈ [N ], k ∈ [K], and w such that kwk2 ≤ nΛ, let
n
E(k, i, w) = P ∃j 6= i : kXi (k) + w + V − Xj (k)k2 ≤ kw + Vk2 ,
(n)
(n)
Xi (k) ∈ T0 , Xj (k) ∈ T0
o
X1 (k), . . . , XN (k) . (63)
To prove the lemma, we need to show that, for any > 0
(
)
N
K
[
1 XX
lim P
E(k, i, w) >
→ 0.
n→∞
N
K
2
w:kwk ≤nΛ
(64)
i=1 k=1
From (57), we know that for any δ > 0 and sufficiently large n for any k, i, w, we have
EE(k, i, w) ≤ δ (The conditions on Xi and Xj only decrease the probability.) Thus, for fixed i
and w we have
(
)
K
n PK
o
1 X
E(k,i,w)
K/2
k=1
P
E(k, i, w) > /2 = P 2
>2
K k=1
−K/2
≤2
K
Y
k=1
E2E(k,i,w)
(65)
(66)
22
≤ 2−K/2 (1 + EE(1, i, w))K
(67)
= 2−K(/2−log(1+δ))
(68)
where (66) holds by Markov’s inequality, and (67) holds since 2t ≤ 1 + t if t ∈ [0, 1]. Thus, if
√
we let w1 , . . . , wL be any finite set of vectors with norm at most nΛ, we may apply the union
bound to find
(
)
K
[
1 X
P
E(k, i, wl ) >
≤ LN 2−K( 2 −log(1+δ)) .
K k=1
2
l∈[L]
(69)
i∈[N ]
In particular, let w1 , . . . , wL be a ν-dense subset of points in the sphere of radius
√
nΛ. There
exists such a set with L = 2nρ for some ρ. Since E(k, i, w) is continuous in w, for sufficiently
small ν, if the probability of error for all wl is at most /2, then the probability of error for all
w is at most . Thus we may bound the probability in (64) by
2nρ 2nR 2−K(/2−log(1+δ)) .
As long as δ is small enough so that /2 − log(1 + δ) > 0 and K/n → ∞, this probability
vanishes in n.
A PPENDIX C
P ROOF OF L EMMA 5
Throughout the appendices, we use the following conditional typicality lemma and joint
typicality lemma for Gaussian random variables; their proofs easily follow from the corresponding
proofs in [15] for discrete memory-less random variables.
(n)
Conditional Typicality Lemma: Let (X, Y ) ∼ f (x, y). Suppose that x ∈ T (X) and Y ∼
Q
f (y|x) = ni=1 fY |X (yi |xi ). Then, for every > 0 ,
lim P{(x, Y) ∈ T(n) (X, Y )} = 1.
n→∞
(70)
Joint Typicality Lemma: Let (X, Y, Z) ∼ f (x, y, z). If (x̃, ỹ) is a pair of arbitrary sequences
Q
and Z̃ ∼ ni=1 fZ|X (z̃i |x̃i ) then there exists δ() > 0 that tends to zero as → 0 such that
P{(x̃, ỹ, Z̃) ∈ T(n) (X, Y, Z)} ≤ exp(−n(I(Y ; Z|X) − δ())).
(71)
23
We apply Lemma 6 to assume the two codebooks satisfy (33)–(36). To prove (31), first note
(n)
that by (33), with high probability (xi1 , yi2 , w) ∈ T0 (X, Y, W ) where (X, Y, W ) are mutually
independent, and
EX 2 = θ,
EY 2 = θ̄,
EW 2 ≤ Λ.
(72)
Thus, by the conditional typicality lemma, for every > 0 with high probability (xi1 , yi2 , w, V) ∈
(n)
T (X, Y, W, V ) where (X, Y, W, V ) are mutually independent, and EV 2 = σ 2 . This implies
that (xi1 , yi2 , w + V) ∈ U, and also that
kw + Vk2 ≤ n(Λ + σ 2 + ).
(73)
~ = (XY X 0 Y 0 W V ). For i1 , i2 , w and any Gaussian distribution on X,
~
We use shorthand X
define
n
o
~ for some j1 6= i1 , j2 6= i2 .
eX~ (i1 , i2 , w) = P (xi1 , yi2 , xj1 , yj2 , w, V) ∈ T(n) (X)
We need to show that for some δ > 0,
1 X
N1 N2
eX~ (i1 , i2 , w) ≤ exp(−nρ)
(74)
(75)
i1 ,i2
~ where
for all X
(X, Y, W, V ) are mutually independent,
EX 2 = EX 02 = θ,
EW 2 ≤ Λ,
EY 2 = EY 02 = θ̄,
EV 2 = σ 2
E(X + Y + W + V − X 0 − Y 0 )2 ≤ Λ + σ 2
(X 0 , Y 0 , X + Y + W + V − X 0 − Y 0 ) are mutually independent.
(76)
(77)
(78)
(79)
Observe that if I(XY V ; W X 0 Y 0 ) = 0, then we would have
Λ + σ 2 ≥ E(X + Y + W + V − X 0 − Y 0 )2
(80)
= E(X + Y + V )2 + E(W − X 0 − Y 0 )2
(81)
≥ 1 + σ2.
(82)
But this cannot happen since Λ < 1 by (25). Thus, there exists η > 0 where
I(XY V ; W X 0 Y 0 ) ≥ η.
(83)
24
Recalling that I(XY V ; W ) = 0, this implies
I(XY V ; X 0 Y 0 |W ) ≥ η.
(84)
Also, by (36), we may restrict ourselves to distributions where
JX 0 ;Y 0 ;W (R1 , R2 ) ≥ I(XY ; X 0 Y 0 W ) − 2δ().
We may now write, for any (i1 , i2 ) and any w
X
eX~ (i1 , i2 , w) ≤
n
o
P (xi1 , yi2 , xj1 , yj2 , w, V) ∈ T(n)
(85)
(86)
(n)
(j1 ,j2 ):(xi1 ,yi2 ,xj1 ,yj2 ,w)∈T
n h
io
≤ exp n JX 0 ;Y 0 ;XY W (R1 , R2 ) − I(V ; X 0 Y 0 |XY W ) + δ() .
(87)
where in (87) we have applied (35), the joint typicality lemma, and the fact that I(V ; XY W ) = 0.
We consider two cases.
Case (a): JX 0 ;Y 0 ;W (R1 , R2 ) = 0. Note that JX 0 ;Y 0 ;XY W (R1 , R2 ) ≤ JX 0 ;Y 0 ;W (R1 , R2 ) so in this
case we also have JX 0 ;Y 0 ;XY W (R1 , R2 ) = 0. By (85), I(XY ; X 0 Y 0 W ) ≤ 2δ(). Thus, from (84)
η ≤ I(XY V ; X 0 Y 0 |W )
(88)
= I(XY ; X 0 Y 0 |W ) + I(V ; X 0 Y 0 |XY W )
(89)
≤ 2δ() + I(V ; X 0 Y 0 |XY W ).
(90)
From (87), we have
eX~ (i1 , i2 , w) ≤ exp{n[−I(V ; X 0 Y 0 |XY W ) + δ()]}
≤ exp{n[−η + 3δ()]}.
(91)
(92)
This vanishes exponentially fast if δ() is sufficiently small.
Case (b): JX 0 ;Y 0 ;W (R1 , R2 ) > 0. This implies that (recalling that I(X 0 ; Y 0 ) = 0)
JX 0 ;Y 0 ;W (R1 , R2 ) = max{R1 − I(X 0 ; W Y 0 ), R2 − I(Y 0 ; W X 0 ), R1 + R2 − I(X 0 Y 0 ; W )}. (93)
By (85), we have
−2δ() ≤ max{R1 −I(X 0 ; W Y 0 ), R2 −I(Y 0 ; W X 0 ), R1 +R2 −I(X 0 Y 0 ; W )}−I(XY ; X 0 Y 0 W ).
(94)
25
Note that
I(X 0 ; W Y 0 ) + I(XY ; X 0 Y 0 W ) ≥ I(X 0 ; W Y 0 ) + I(XY ; X 0 |W Y 0 )
= I(X 0 ; XY W Y 0 ).
(95)
(96)
Similarly
I(Y 0 ; W X 0 ) + I(XY ; X 0 Y 0 W ) ≥ I(Y 0 ; XY W X 0 ),
(97)
I(X 0 Y 0 ; W ) + I(XY ; X 0 Y 0 W ) ≥ I(X 0 Y 0 ; XY W ).
(98)
Thus
−2δ() ≤ max{R1 −I(X 0 ; XY W Y 0 ), R2 −I(Y 0 ; XY W X 0 ), R1 +R2 −I(X 0 Y 0 ; XY W )}. (99)
Hence
JX 0 ;Y 0 ;XY W (R1 , R2 )
≤ max{R1 − I(X 0 ; XY W Y 0 ), R2 − I(Y 0 ; XY W X 0 ), R1 + R2 − I(X 0 Y 0 ; XY W )} + 2δ().
(100)
By (87), we have
1
log eX~ (i1 , i2 , w)
n
≤ JX 0 ;Y 0 ;XY W (R1 , R2 ) − I(V ; X 0 Y 0 |XY W ) + δ()
(101)
≤ max{R1 − I(X 0 ; XY W Y 0 ), R2 − I(Y 0 ; XY W X 0 ), R1 + R2 − I(X 0 Y 0 ; XY W )}
− I(V ; X 0 Y 0 |XY W ) + 3δ()
(102)
≤ max{R1 − I(X 0 ; XY W Y 0 V ), R2 − I(Y 0 ; XY W X 0 V ), R1 + R2 − I(X 0 Y 0 ; XY W V )}
+ 3δ().
(103)
Let Z = X + Y + W + V − X 0 − Y 0 . Recalling that (X 0 , Y 0 , Z) are mutually independent and
EZ 2 ≤ Λ + σ 2 , we have
I(X 0 ; XY W Y 0 V ) ≥ I(X 0 ; X + Y + W + V − Y 0 )
(104)
= I(X 0 ; X 0 + Z)
(105)
= h(X 0 + Z) − h(X 0 + Z|X 0 )
(106)
26
1
log 2πe(θ + EZ 2 ) − h(Z|X 0 )
2
1
1
= log 2πe(θ + EZ 2 ) − log 2πeEZ 2
2
2
θ
=C
EZ 2
θ
≥C
.
Λ + σ2
=
(107)
(108)
(109)
(110)
Similarly
0
0
I(Y ; XY SX V ) ≥ C
θ̄
Λ + σ2
.
(111)
Moreover,
I(X 0 Y 0 ; XY W V ) ≥ I(X 0 Y 0 ; Z + X 0 + Y 0 )
(112)
= h(Z + X 0 + Y 0 ) − h(Z)
1
=C
EZ 2
1
≥C
.
Λ + σ2
(113)
(114)
(115)
Thus
n h
eX~ (i1 , i2 , w) ≤ exp n max R1 − C
θ
Λ+σ 2
, R2 − C
θ̄
Λ+σ 2
, R1 + R2 − C
1
Λ+σ 2
io
+ 3δ() .
(116)
Therefore, eX~ (i1 , i2 , w) is exponentially vanishing if δ() is sufficiently small and (26)–(28)
hold.
A PPENDIX D
P ROOF OF L EMMA 6
Since we frequently use [3, Lemma A1] throughout this appendix, we provide the statement
of this lemma here.
Lemma 7: Let Z1 , . . . , ZN be arbitrary random variables, and let fi (Z1 , . . . , Zi ) be arbitrary
with 0 ≤ fi ≤ 1, i = 1, . . . , N . Then the condition
E [fi (Z1 , . . . , Zi )|Z1 , . . . , Zi−1 ] ≤ a a.s., i = 1, . . . , N,
(117)
27
implies that
(
P
)
N
1 X
fi (Z1 , . . . , Zi ) > t ≤ exp{−N (t − a log e)}.
N i=1
(118)
Let Xi1 and Yi2 be Gaussian i.i.d. n-length random vectors (codebooks) independent from
(n)
(n)
each other with Var(X) = θ and Var(Y ) = θ̄. Fix x ∈ T (X), y ∈ T (Y ), w ∈ S n and
a covariance matrix Cov(X, Y, X 0 , Y 0 , W ) ∈ V 5×5 such that S n is a ν-dense subset of Rn for
w such that ||w||2 ≤ nΛ, and V 5×5 is a ν-dense subset of R5x5 for positive definite covariance
matrices with diagonals at most (θ, θ̄, θ, θ̄, Λ).
Let
[
An (X, W ) =
T(n) (X, W )
(119)
X,W independent
EX 2 =θ,EW 2 ≤Λ
and
[
An (X, Y, W ) =
T(n) (X, Y, W ).
(120)
(X,Y,W ) mutually independent
EX 2 =θ,EY 2 =θ̄,EW 2 ≤Λ
To prove (33), first define hi1 as a function of X1 , . . . , Xi1 as follows:
1, if (Xi , w) ∈
/ An (X, W )
1
hi1 (X1 , . . . , Xi1 ) =
0, otherwise .
(121)
Then the expected value of hi1 is given as
E[hi1 (X1 , . . . , Xi1 )|X1 , . . . , Xi1 −1 ] = E [1 ((Xi1 , w) ∈
/ An (X, W ))]
(122)
= P {(Xi1 , w) ∈
/ An (X, W )}
(123)
1
1
kwk2 ≥ Λ + + P
|hXi1 , wi| ≥
≤P
n
n
1
2
(124)
+ P | kXi1 k − θ| ≥
n
≤ exp(−nr1 ())
(125)
where (124) follows since the union in An (X, W ) is over independent X, W such that EX 2 = θ,
EW 2 ≤ Λ, and directly from the definition of typical set in (14) we obtain that it only suffices
to find the probability of those (Xi1 , w) that simultaneously do not satisfy the conditions of the
union and the typical set definition’s inequalities. The upper bound in (125) follows since the first
28
term in (124) is equal to zero (assumption kwk2 ≤ nΛ), and the other terms are exponentially
vanishing by using the large deviation theory for Gaussian distributions X with positive function
r1 ().
Now, using Lemma 7, we have
1
n
|{i1 : (Xi1 , w) ∈
/ A (X, W )}| > exp(−nδ1 ())
P
N1
≤ exp{−N1 [exp(−nδ1 ()) − exp(−nr1 ()) log e]} (126)
≤ exp(− exp(nρ1 ())).
(127)
where the last inequality follows as long as δ1 () < r1 () for some ρ1 () > 0. Thus, the
probability vanishes doubly exponentially as n → ∞, and with high probability we have
1
|{i1 : (xi1 , w) ∈
/ An (X, W )}| ≤ exp(−nδ1 ()).
N1
(128)
Fix xi1 , and for any i2 define h̃i2 as
h̃i2 (Y1 , . . . , Yi2 ) =
1
N1
X
1 ((xi1 , Yi2 , w) ∈/ An (X, Y, W )) .
(129)
i1 :(xi1 ,w)∈ An
(X,W )
The expected value of h̃i2 can be written as
i
h
X
1
P {(xi1 , Yi2 , w) ∈
/ An (X, Y, W )}
E h̃i2 (Y1 , . . . , Yi2 )|Y1 , . . . , Yi2 −1 =
N1
i1 :(xi1 ,w)∈ An
(X,W )
1
1
2
kwk ≥ Λ + + P
|hxi1 , wi| ≥
≤P
n
n
1
1
2
+P
kxi k − θ ≥ + P
|hYi2 , wi| ≥
n 1
n
1
1
2
+P
|hYi2 , xi1 i| ≥ + P
kYi2 k − θ̄ ≥
n
n
(130)
≤ exp(−nr2 ())
(131)
where (130) follows directly from the definition of typical set and the union’s conditions, and
(131) follows since the first three terms in (130) are equal to 0 due to the assumptions and the
other terms in (130) vanish exponentially by the large deviation theory for Gaussian distributions
Y with positive function r2 ()
29
Using Lemma 7, we have
1
X
P
N1 N2
n
i1 :(xi1 ,w)∈ A (X,W )
|{i2 : (xi1 , Yi2 , w) ∈
/ An (X, Y, W )}| > exp(−nδ2 ())
≤ exp{−N2 (exp(−nδ2 () − exp(−nr2 ()) log e))}
(132)
≤ exp(− exp(nρ2 ()))
(133)
where (133) follows if δ2 () < r2 () for some ρ2 () > 0. Therefore, with probability approaching
1, as n → ∞ we have
X
1
N1 N2
|{i2 : (xi1 , yi2 , w) ∈
/ An (X, Y, W )}| ≤ exp(−nδ2 ()).
(134)
i1 :(xi1 ,w)∈ An
(X,W )
Eventually, we easily use (128) and (134) to bound the fraction in (33) as follows:
1
|{(i1 , i2 ) :(xi1 , yi2 , w) ∈
/ An (X, Y, W )}|
N1 N2
X
1
|{i2 : (xi1 , yi2 , w) ∈
/ An (X, Y, W )}|
≤
N1 N2
n
i1 :(xi1 ,w)∈
/ A (X,W )
+
1
N1 N2
X
|{i2 : (xi1 , yi2 , w) ∈
/ An (X, Y, W )}|
(135)
i1 :(xi1 ,w)∈ An
(X,W )
≤ exp(−nδ1 ()) + exp(−nδ2 ())
(136)
≤ exp(−nδ()).
(137)
This proves (33).
Next, define function fi1 as follows:
fi1 (X1 , . . . , Xi1 ) =
1,
if (Xi1 , w) ∈ T (X 0 , W )
0,
otherwise .
(n)
(138)
Then, by joint typicality lemma we get
E[fi1 (X1 , . . . , Xi1 )|X1 , . . . , Xi1 −1 ] = E 1 (Xi1 , w) ∈ T(n) (X 0 , W )
= P (Xi1 , w) ∈ T(n) (X 0 , W )
≤ exp(−nI(X 0 , W ) + nδ()).
Thus, using Lemma 7, we have
P
i1 : (Xi1 , w) ∈ T(n) (X 0 , W )
> exp n|R1 − I(X 0 , W )|+ + n2δ()
(139)
(140)
(141)
30
≤ exp − exp n|R1 − I(X 0 , W )|+ + n2δ() +exp(−nI(X 0 , W ) + nδ() + nR1 ) log e . (142)
If R1 > I(X 0 , W ), (142) becomes less than doubly exponentially function exp(− exp(nσ()))
where σ() > 0 since for large enough n we obtain exp(nδ()) > log e. Now, if R1 ≤ I(X 0 , W )
then (142) is less than
exp{− exp(n2δ()) + exp(nR1 + nδ() − nI(X 0 , W )) log e} ≤ exp(− exp(nσ()))
where σ() > 0. In both cases, this doubly decreasing exponential function vanishes as n → ∞.
Hence, with high probability, we have
i1 : (xi1 , w) ∈ T(n) (X 0 , W )
≤ exp n |R1 − I(X 0 ; W )|+ + δ() .
(143)
For any i2 ,
P (xi1 , Yi2 , w) ∈ T(n) (X 0 , Y 0 , W ) for some i1
X
≤
P (xi1 , Yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
(144)
(n)
i1 :(xi1 ,w)∈T (X 0 ,W )
≤
i1 : (xi1 , w) ∈ T(n) (X 0 , W )
max P (x̂, Yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
x̂
≤ exp n |R1 − I(X 0 ; W )|+ − I(Y 0 ; X 0 W ) + δ() ,
(145)
(146)
where (144) follows since if (xi1 , Yi2 , w) is typical then (xi1 , w) is always typical and (146)
follows from (143) and joint typicality lemma. Thus, applying Lemma 7 in the same way that
we used it to get (143), with high probability we attain
i2 : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W ) for some i1
n h
≤ exp n R2 + |R1 − I(X 0 ; W )|+ − I(Y 0 ; X 0 W )
(n)
+
io
+ δ() . (147)
(n)
Since (xi1 , yi2 , w) ∈ T (X 0 , Y 0 , W ) implies (yi2 , w) ∈ T (Y 0 , W ), we have the simpler
bound
i2 : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W ) for some i1
(148)
i2 : (yi2 , w) ∈ T(n) (Y 0 , W )
≤ exp n |R2 − I(Y 0 ; W )|+ + δ() ,
≤
(149)
31
where (149) is similar to (143). Moreover, if we replace vector w by (yi2 , w) in (143), then for
any i2 we get
i1 : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
≤ exp n |R1 − I(X 0 ; Y 0 W )|+ + δ() .
(150)
Therefore,
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
X
≤
{i1 : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
(n)
i2 :(xi1 ,yi2 ,w)∈T
(151)
(X 0 ,Y 0 ,W ) for some i1
n h
≤ exp n R2 + |R1 − I(X 0 ; W )|+ − I(Y 0 ; X 0 W )
+
0
0
+
+ |R1 − I(X ; Y W )| + δ()
io
(152)
where (152) follows from (147) and (150). If R1 ≤ I(X 0 ; Y 0 W ), then we have
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
n h
io
+
0
+
0
0
≤ exp n R2 + |R1 − I(X ; W )| − I(Y ; X W ) + δ()
(153)
n h
io
= exp n max 0, R2 − I(Y 0 ; X 0 W ), R1 + R2 − I(X 0 Y 0 ; W ) − I(X 0 ; Y 0 ) + δ()
(154)
= exp n JX 0 ;Y 0 ;W (R1 , R2 ) + δ() .
(155)
Using (149) and (150), we alternatively bound
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
≤ exp n |R2 − I(Y 0 ; W )|+ + |R1 − I(X 0 ; Y 0 W )|+ + δ() . (156)
In particular, if R1 > I(X 0 ; Y 0 W ) then we have
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
≤ exp n |R2 − I(Y 0 ; W )|+ + R1 − I(X 0 ; W |Y 0 ) + δ()
n h
io
= exp n max R1 − I(X 0 ; W |Y 0 ), R1 + R2 − I(X 0 Y 0 ; W ) + δ()
= exp n JX 0 ;Y 0 ;W (R1 , R2 ) + δ() .
(157)
(158)
(159)
This proves (34). An identical calculation with (X, Y, W ) in place of W gives (35). We indeed
use Lemma 7 to prove that the complement of two events (34) and (35) happen with decreasingly
doubly exponential probability for sufficiently large n as follows:
32
P
(i1 , i2 ) : (xi1 , yi2 , w) ∈ T(n) (X 0 , Y 0 , W )
> exp n JX 0 ;Y 0 ;W (R1 , R2 ) + δ()
< exp(− exp(nσ())), (160)
P
(i1 , i2 ) : (x, y, xi1 , yi2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
> exp n JX 0 ;Y 0 ;XY W (R1 , R2 ) + δ()
< exp(− exp(nσ())). (161)
Now, in order to prove (36), first let
A(i1 ,i2 ) = (j1 , j2 ) : j1 < i1 , j2 6= i2 , (xj1 , Yj2 , w) ∈ T(n) (X 0 , Y 0 , W ) ,
A(i ,i ) , if |A(i ,i ) | ≤ exp {n (JX 0 ;Y 0 ;W (R1 , R2 ) + δ())}
1 2
1 2
Ã(i1 ,i2 ) =
∅,
otherwise,
(162)
(163)
where Ã(i1 ,i2 ) is defined for fixed value of x1 , . . . , xi1 −1 and random Y1 , . . . , Yj2 , i.e. Ã(i1 ,i2 ) is
a random set. Define
n
o
gi1 (x1 , . . . , xi1 ) = P (xi1 , Yi2 , xj1 , Yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some (j1 , j2 ) ∈ Ã(i1 ,i2 )
(164)
and
g̃i1 (x1 , . . . , xi1 ) =
1,
if gi1 (x1 , . . . , xi1 ) > exp(−nδ()/2)
∅,
otherwise.
(165)
It is notable that since by (34)
P
(j1 , j2 ) : (xj1 , yj2 , w) ∈ T(n) (X 0 , Y 0 , W )
> exp{n(JX 0 ;Y 0 ;W (R1 , R2 ) + δ())}
(166)
tends to zero as n grows, the probability that Ã(i1 ,i2 ) 6= A(i1 ,i2 ) for some i1 , i2 vanishes as n → ∞.
Finding the expected values of gi1 and g̃i1 , we have
E[gi1 (X1 , . . . , Xi1 )|X1 = x1 , . . . , Xi1 −1 = xi1 −1 ]
n
o
= P (Xi1 , Yi2 , xj1 , Yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some (j1 , j2 ) ∈ Ã(i1 ,i2 )
X
≤
P (Xi1 , Yi2 , xj1 , Yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
(167)
(168)
(j1 ,j2 )∈Ã(i1 ,i2 )
≤ exp(nJX 0 ;Y 0 ;W (R1 , R2 ) + δ()) max P (Xi1 , Yi2 , x̂, ŷ, w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
x̂,ŷ
(169)
33
≤ exp{−n(−JX 0 ;Y 0 ;W (R1 , R2 ) + I(XY ; X 0 Y 0 W ) − δ())}
(170)
≤ exp(−nδ())
(171)
where (169) follows since by (163) the size of Ã(i1 ,i2 ) is almost surely less than exp(nJX 0 ;Y 0 ;W (R1 ,
R2 ) + δ()), (170) follows by joint typicality lemma, and (171) follows by the condition in (36).
Moreover, by Markov’s inequality we have
E[g̃i1 (X1 , . . . , Xi1 )|X1 , . . . , Xi1 −1 ] = P{gi1 (X1 , . . . , Xi1 ) > exp(−nδ()/2)|X1 , . . . , Xi1 −1 }
≤
E[gi1 (X1 , . . . , Xi1 )|X1 , . . . , Xi1 −1 ]
exp(−nδ()/2)
(172)
≤ exp(−nδ() + nδ()/2)
(173)
= exp(−nδ()/2).
(174)
Therefore, using Lemma 7, we have
(
)
N1
1 X
P
g̃i (x1 , . . . , xi1 ) > exp(−nδ()/4)
N1 i =1 1
1
1
|{i1 : gi1 (x1 , . . . , xi1 ) > exp(−nδ()/2)}| > exp(−nδ()/4)
=P
N1
(175)
≤ exp{− exp(nR1 )(exp(−nδ()/4) − exp(−nδ()/2) log e)}
(176)
≤ exp(− exp(nσ())).
(177)
for σ() > 0 i.e. with high probability
1 X
1
g̃i1 (x1 , . . . , xi1 ) =
|{i1 : gi1 (x1 , . . . , xi1 ) > exp(−nδ()/2)}|
N1 i
N1
(178)
1
≤ exp(−nδ()/4).
(179)
Let
f(i1 ,i2 ) (y1 , . . . , yi2 ) =
1,
0,
(n)
if (xi1 , yi2 , xj1 , yj2 , w) ∈ T (X, Y, X 0 , Y 0 , W ),
for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
otherwise.
Now, fix an i1 such that gi1 (x1 , . . . , xi1 ) ≤ exp(−nδ()/2). Therefore, we have
E f(i1 ,i2 ) (Y1 , . . . , Yi2 )|Y1 , . . . , Yi2 −1
(180)
34
n
= P (xi1 , Yi2 , xj1 , Yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
o
Y1 , . . . , Yi2 −1
(181)
≤ gi1 (x1 , . . . , xi1 )
(182)
≤ exp(−nδ()/2)
(183)
where (182) and (183) follow directly from gi1 definition and our assumption for the gi1 . Thus,
using Lemma 7 we get
P
!
N2
1 X
f(i ,i ) (Y1 , . . . , Yi2 ) > exp(−nδ()/4) ≤ exp(− exp(nσ()))
N2 i =1 1 2
(184)
2
where σ() > 0. If we sum over all i1 ’s, we obtain
N1
X
N2
1 X
f(i ,i ) (Y1 , . . . , Yi2 ) > exp(−nδ()/4)
N2 i =1 1 2
P
i1 =1
!
≤ exp(nR1 − exp(nσ())),
(185)
2
that is this doubly exponential function still tends to zero as n → ∞. Therefore, with probability
approaching 1, for every i1 that gi1 (x1 , . . . , xi1 ) ≤ exp(−nδ()/2) we have
o
1 n
i2 : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
N2
≤ exp(−nδ()/4). (186)
In general, for all i1 we have
1 n
(i1 , i2 ) : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
N1 N2
for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
o
(187)
N1
1 X
1 n
≤
i2 : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
N1 i =1 N2
1
o
for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
(188)
X
1
1 n
≤ exp(−nδ()/4)+
i2 : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W )
N1
N2
i1 :gi1 ≤exp(−nδ()/2)
o
for some (j1 , j2 ) ∈ Ã(i1 ,i2 ) and j2 < i2
(189)
≤ 2 exp(−nδ()/4)
where (189) and (190) follow from (179) and (186), respectively.
(190)
35
We may use this argument to upper bound the probability in (187) for the remain three cases
(j1 < i1 , j2 > i2 ), (j1 > i1 , j2 < i2 ) and (j1 > i1 , j2 > i2 ) by defining different A(i1 ,i2 ) ’s, and
conclude the same decreasing exponential function. Finally, we obtain
1
(i1 , i2 ) : (xi1 , yi2 , xj1 , yj2 , w) ∈ T(n) (X, Y, X 0 , Y 0 , W ) for some j1 6= i1 , j2 6= i2
N1 N2
≤ 8 exp{−nδ()/4} (191)
if we have JX 0 ;Y 0 ;W (R1 , R2 ) ≤ I(XY ; X 0 Y 0 W ) − 2δ().
In order to complete the proof, since for any fixed ν the cardinality of finite set S n is
only increasingly exponentially in n, and the set V 5×5 is finite along with the doubly decreasing
exponential probabilities in (160) and (161), we derive that with probability approaching to 1, all
inequalities in (33), (34), (35) and (36) hold simultaneously for sufficiently large n. Since these
inequalities hold for every element in the finite sets S n and V 5×5 , then for any vector w, x, y and
any given covariance matrix Cov(X, Y, X 0 , Y 0 , W ) (with kxk2 = nθ, kyk2 = nθ̄, kwk2 ≤ nΛ)
which is not in its corresponding ν-dense subset, there exists a point in the corresponding ν-dense
subset that is close enough to it (in its ν distance neighborhood). Now, by using the continuity
properties of all sets, we may conclude that (33), (34), (35) and (36) hold also for any point
which is not in the ν-dense subset.
R EFERENCES
[1] F. Hosseinigoki and O. Kosut, “The gaussian interference channel in the presence of a malicious jammer,” in 2016 54th
Annual Allerton Conference on Communication, Control, and Computing (Allerton), Sept 2016, pp. 679–686.
[2] R. Ahlswede, “Elimination of correlation in random codes for arbitrarily varying channels,” Zeitschrift fr Wahrscheinlichkeitstheorie und verwandte Gebiete, vol. 44, no. 2, pp. 159–175, 1978.
[3] I. Csiszar and P. Narayan, “The capacity of the arbitrarily varying channel revisited: positivity, constraints,” Information
The, vol. 34, no. 2, pp. 181–193, 1988.
[4] ——, “Capacity of the Gaussian arbitrarily varying channel,” Information Theory, IEEE Transactions on, vol. 37, no. 1,
pp. 18–26, Jan 1991.
[5] M. Medard, “Capacity of correlated jamming channels,” in Proc. 35th Annu. Allerton Conf. Communications, Control and
Computing, 1997.
[6] R. W. Yeung and N. Cai, “Network error correction, part I: Basic concepts and upper bounds,” Comm. in Inf. and Syst.,
vol. 6, no. 1, pp. 19–36, 2006.
[7] N. Cai and R. W. Yeung, “Network error correction, part II: Lower bounds,” Comm. in Inf. and Syst., vol. 6, no. 1, pp.
37–54, 2006.
36
[8] S. Shafiee and S. Ulukus, “Mutual information games in multiuser channels with correlated jamming,” IEEE Transactions
on Information Theory, vol. 55, no. 10, pp. 4598–4607, Oct 2009.
[9] S. Mohajer and S. Diggavi, “Gaussian diamond network with adversarial jammer,” in Information Theory Workshop (ITW),
2010 IEEE, 2010, pp. 1–5.
[10] T. Ericson, “Exponential error bounds for random codes in the arbitrarily varying channel,” IEEE Transactions on
Information Theory, vol. 31, no. 1, pp. 42–48, Jan 1985.
[11] T. Han and K. Kobayashi, “A new achievable rate region for the interference channel,” Information Theory, IEEE
Transactions on, vol. 27, no. 1, pp. 49–60, Jan 1981.
[12] R. H. Etkin, D. N. C. Tse, and H. Wang, “Gaussian interference channel capacity to within one bit,” Information Theory,
IEEE Transactions on, vol. 54, no. 12, pp. 5534–5562, Dec 2008.
[13] V. S. Annapureddy and V. V. Veeravalli, “Gaussian interference networks: Sum capacity in the low-interference regime
and new outer bounds on the capacity region,” IEEE Transactions on Information Theory, vol. 55, no. 7, pp. 3032–3050,
July 2009.
[14] G. Bresler and D. Tse, “The two-user Gaussian interference channel: a deterministic view,” Eur. Trans. Telecomm., vol. 19,
no. 4, pp. 333–354, June 2008.
[15] A. E. Gamal and Y. H. Kim, Network information theory.
Cambridge University Press, 2011.
[16] I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems.
Cambri, 2011.
[17] A. Lapidoth, “Nearest neighbor decoding for additive non-Gaussian noise channels,” IEEE Transactions on Information
Theory, vol. 42, no. 5, pp. 1520–1529, Sep 1996.
| 7cs.IT
|
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
ws-ijmpc
arXiv:1402.2440v1 [cs.CE] 11 Feb 2014
International Journal of Modern Physics C
c World Scientific Publishing Company
VALIDATION EXPERIMENTS FOR LBM SIMULATIONS OF
ELECTRON BEAM MELTING
REGINA AMMER∗ and ULRICH RÜDE†
Department of Computer Science 10 System Simulation, University of Erlangen-Nuremberg,
Cauerstr. 11, 91058 Erlangen, Germany
∗ [email protected]
† [email protected]
MATTHIAS MARKL⋆ , VERA JÜCHTER† and CAROLIN KÖRNER‡
Chair of Metals Science and Technology, University of Erlangen-Nuremberg,
Martensstr. 5, 91058 Erlangen, Germany
⋆ [email protected]
† [email protected]
‡ [email protected]
Received Day Month Year
Revised Day Month Year
This paper validates 3D simulation results of electron beam melting (EBM) processes
comparing experimental and numerical data. The physical setup is presented which is
discretized by a three dimensional (3D) thermal lattice Boltzmann method (LBM). An
experimental process window is used for the validation depending on the line energy
injected into the metal powder bed and the scan velocity of the electron beam. In the
process window the EBM products are classified into the categories, porous, good and
swelling, depending on the quality of the surface. The same parameter sets are used
to generate a numerical process window. A comparison of numerical and experimental
process windows shows a good agreement. This validates the EBM model and justifies
simulations for future improvements of EBM processes. In particular numerical simulations can be used to explain future process window scenarios and find the best parameter
set for a good surface quality and dense products.
Keywords: 3D thermal lattice Boltzmann method; free surface; electron beam melting;
validation experiments.
PACS Nos.: 11.25.Hf, 123.1K
1. Introduction
Electron beam melting (EBM) is an additive manufacturing method used to produce complex metallic structures layer by layer from metal powder 1 . EBM opens
new opportunities in many industries, ranging from aircraft manufacturers to producers of medical implants. However, until now the parts cannot be manufactured
at sufficient speed to make them economically viable for any but specific very high
1
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
2
ws-ijmpc
Regina Ammer, Matthias Markl, Vera Jüchter, Carolin Körner, Ulrich Rüde
value applications. In order to accelerate the building process and improve the accuracy a better understanding of the beam-powder interaction is necessary. This
can be gained by 3D simulations of the process.
Based on the work of Körner et al.2 we use a 3D thermal LBM for the discretization for the EBM processes. This model includes hydrodynamic effects, like melt
flow, capillarity and wetting, as well as thermal effects, like absorption of the beam
energy, melting and solidification. Detailed information for parallel, optimized 3D
absorption algorithms is found in 3 and the description of the modeling aspects,
e.g., the generation of the Ti-Al6-V4 powder particles is given in 4 . The implementation of the model is embedded in the waLBerla framework (widely applicable
lattice Boltzman solver from Erlangen) which is a lattice Boltzmann based fluid
flow solver with highly parallelized kernels 5,6 . In 4 basic validation examples are
already shown, like the test of the Stefan problem as benchmark for the simulation of solid-liquid phase transition, and modest EBM process examples, e.g., the
melting of a spot regarding the interaction of powder particles and electron beam.
In this paper we extend the validation examples focusing on the EBM processes,
e.g, hatching scenarios, and compare the numerical results with experimental data.
These numerical results demonstrate the good potential of the thermal LBM approach to understand and predict complex processes like the EBM depending on
thermodydnamic as well as on fluid dynamic phenomena.
The remainder of the paper is organized as follows: Section 2 describes the 3D
thermal LBM used for the simulation of the EBM process. The next section defines
line energy, scan velocity, porosity and swelling before showing the experimental
process window. Subsequently, the generation of the numerical process window is
described, shown and compared with the experimental one. Section 4 concludes the
validation experiments and outlines future research topics.
2. Numerical methods
Based on the lattice Boltzmann equation the isothermal lattice Boltzmann method
(LBM) is introduced by 7 as an improvement of its predecessor, the lattice gas
automata (LGA) 8 . LBM is an alternative, mesoscopic approach for the numerical
solution of the Navier-Stokes equations 9 and can be parallalized easier than other
traditional CFD methods whose computation is based on the conservation of macroscopic quantities. LBM has also less computational storage and work requirements
than molecular dynamic methods which consider microscopic physics.
Thermal LB methods are classified into three categories: the multispeed
LBM 10,11 , the double-distribution or multidistribution approach 12,13,14 and hybrid approach where the scalar temperature equation is solved by finite differences
or finite volume methods 15 . The multispeed LBM can be seen as an extension of
the isothermal LBM where the equilibrium distribution function depends also on
the temperature and has higher nonlinear velocity terms. This method suffers from
numerical instabilities which have to be stabilized as it is shown in 16 . An additional
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
ws-ijmpc
Validation Experiments for LBM Simulations of Electron Beam Melting
3
disadvantage is that only one Prandtl number can be simulated by multispeed LBM
which limits the range of applications. The multidistribution LBM overcomes these
drawbacks using a separate distribution function for the temperature field.
For the simulation of the EBM process a thermal two-distribution LBM is used
which is described followed by the treatment of the free surface boundary condition.
2.1. 3D thermal lattice Boltzmann method
The model for the EBM process is based on the work of Körner et al. in 2 where a
2D thermal LBM is discussed. The idea of LBM is solving the Boltzmann equation
in the hydrodynamic limit for a particle distribution function (pdf) in the physical
momentum space. This pdf f (x, v, t) indicates the probability of finding a particle
with the macroscopic velocity v(x, t) at position x and time t. For the thermal LBM
a second pdf set h(x, v, t) is used for the numerical solution of the temperature field.
The collision operator is approximated by the BGK-collision operator 17,11 ,
Ωf = −
1
(f − f eq ) ,
τf
Ωh = −
1
(h − heq ) ,
τh
(1)
where f eq and heq denote the Maxwell equilibrium distribution 19 . τf is the relaxation time related to the viscosity ν of the fluid and given by ν = c2s ∆t(τf − 0.5)
∆x2
with the lattice speed of sound c2s = 3∆t
2 . τh belongs to the second pdf set hi and
there exists the relation to the thermal diffusivity k = c2s ∆t(τh − 0.5).
The D3Q19 stencil is used to discretize the microscopic space, i.e., a finite set of
19 discrete velocities ei and lattice weights ωi are given 19 . The Maxwell equilibrium
distribution is discretized by a Taylor series expansion 20 and is given by,
(ei · u) (ei · u)2
(ei · u)
u2
eq
eq
fi (x, t) = ωi ρ 1 +
+
− 2 , hi (x, t) = ωi E 1 +
c2s
2c4s
2cs
c2s
(2)
where u denotes the velocity of the liquid. It is sufficient to use a linearized equilibrium distribution function heq
i for the temperature field. The macroscopic quantities
density ρ, momentum ρu and energy density E are computed in the following way,
X
X
X
hi .
(3)
ei fi , E =
fi , ρu =
ρ=
i
i
i
Both pdf sets are one-way coupled by the velocity of the fluid u which is used for the
computation of heq
i . The discretization of the EBM process by a 3D coupled-thermal
4
LBGK can be written in a stream-collide algorithm,
( ′
f (x + ei ∆t, t + ∆t) = fi (x, t)
streaming i′
(4)
hi (x + ei ∆t, t + ∆t) = hi (x, t),
collide
(
fi (x, t + ∆t)
hi (x, t + ∆t)
′
= fi (x, t + ∆t) +
′
= hi (x, t + ∆t) +
eq
1
τf (fi (ρ, u) − fi (x, t + ∆t)) + Fi
eq
1
τh (hi (E, u) − hi (x, t + ∆t)) + Φi ,
(5)
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
4
ws-ijmpc
Regina Ammer, Matthias Markl, Vera Jüchter, Carolin Körner, Ulrich Rüde
where Fi denotes an external force term, e.g., gravity, and Φi is the energy injected
into the system by the electron beam,
(ei − u) (ei · u)ei
Fi = ωi ρ
· g, Φi (x, t) = ωi Eb (x, t),
(6)
+
c2s
c4s
where Eb denotes the energy of the electron beam. Detailed information how the
beam energy and absorption are computed is given in 3,4 . Next, the free surface
boundary condition is outlined.
2.2. Free surface lattice Boltzmann method
Körner et al. introduce in 21 a volume-of-fluid based free surface lattice Boltzmann
approach. This method neglects the gas phase and works completely on the liquid
phase assuming that the liquid phase covers the behavior of the flow completely. In
order to ensure mass and momentum conservation a boundary condition is imposed
at the interface which separates liquid and gas phase. More information regarding
the free surface boundary condition used for this EBM model can be found in 4 .
3. Validation experiments
The following validation experiments do not only cover particular aspects as the
solid-liquid phase transition or only concentrate on qualitative accordance but consider also the complete EBM process in a quantitative way.
3.1. Definitions and experimental setup
10mm
We use the following experimental setup for the validation experiments.
A cuboid of size (15x15x10) mm3 is generated by
hatching the Ti-Al6-V4 powder particles layer
by layer (compare the red arrows in Fig. 1). The
hatching differs by line energy and scan velocity
of the electron beam. The line energy is defined
by,
Pbeam
U ·I
=
,
EL =
vscan
vscan
(7)
m
15
m
15mm
where U denotes the acceleration voltage in [V],
Fig. 1. Experimental setup.
I the current in [A] and vscan the scan velocity
in [ ms ] of the electron beam. The parameter set
(EL , vscan ) defines the electron beam. The quality of EBM products is classified into
three categories, namely porous, good and swelling. Good samples have a smooth
surface and a relative density higher than 99.5%. If the temperature during the
process is too high swelling can occur and the dimensional accuracy cannot be
guaranteed. On the other side, if the line energy is too low and the relative density
is smaller than 99.5%, then the sample is porous (cf. Fig. 2). Next, a experimental
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
ws-ijmpc
Validation Experiments for LBM Simulations of Electron Beam Melting
porous
good
5
swelling
Increasing line energy
Fig. 2.
Categories of samples.
process window is shown in Fig. 3 which is compared with numerical simulation
results.
3.2. Experimental data
Fig. 3 shows the quality of samples with different line energies in [ kJ
m ] and scan
velocities in [ ms ] of the electron beam. The red circles stand for a porous parameter
0.8
good
porous
swelling
0.7
Line energy [kJ/m]
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
5
6
7
Scan velocity [m/s]
Fig. 3. Process window of experimental data. Built temperature 1000 K, line offset 100 µm, layer
thickness 50 µm, focused electron beam.
set, the blue squares for an optimal, good parameter set and the yellow rhombus
for a sample where swelling occurs because of too high temperatures at the surface.
It can be observed that the higher the scan velocity the lower the embedded line
energy has to be to get an optimal sample result.
3.3. Numerical results
We validate our 3D numerical EBM model against these experimental data
of Fig. 3. Because of the high computational cost of the 3D simulations we
only model the hatching of one powder layer instead of the multiple layers.
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
6
ws-ijmpc
Regina Ammer, Matthias Markl, Vera Jüchter, Carolin Körner, Ulrich Rüde
One simplified scenario for the exembeam offset
J
plary parameter set of (3.2 ms , 0.2 m
)
does already require 8 compute nodes
of LiMaa for 4 wall clock hours. LiMa
is the compute cluster where all 3D
simulations are done. However, this
simplification is in good agreement
simulated powder bed
hatching lines
with the situation expected for the full
setup as it will be shown in this Section.
We also minimize the simulated powder
Fig. 4. Sketch of simulation scenario.
particle layer, i.e., we focus only a rectangular powder layer and have seven hatching lines (cf. Fig. 4). Thus, a beam offset
per layer is defined where the electron beam is on but outside the simulated powder
particle bed.
In Fig. 5 the quality of the numerical experiments with different line energy and
scan velocity is shown. In order to determine if the numerical sample is porous the
relative density of the sample is measured and if it is less than 99.5% the sample is
porous. The definition of numerical swelling is more difficult. Our model does not
include evaporation and thus no evaporation pressure. Hence, numerical swelling
effects cannot occur. We have to find another way how we can determine regions
numerically where swelling exists experimentally. Therefore, we use the numerical
computed temperature which is higher than in real EB melting processes because
the cooling effect during the process of evaporation is missing. Furthermore, numerical artefacts cause outliers in the temperature field. In order to overcome these
numerical influences we use an averaging process for the temperature values. If these
averaged values are still higher than 7500 K we assume that swellings occur in the
experimental setup.
In comparison with the experimental data in Fig. 3 it is observed that all porous
samples are consistent in the simulation results (cf. Fig. 5). Furthermore, the numerical and experimental results for scan velocities of 3.2 ms , 4.0 ms and 6.4 ms agree
with each other. Small differences between numerical and experimental data are
only found for 0.8 ms , 1.6 ms , 2.4 ms , 4.8 ms and 5.6 ms for higher line energies; for
vscan = 0.8 ms in experimental data swelling already occur for 0.4 and 0.5 kJ
m but in
numerical data no swelling is measured. The same effect is seen for scan velocities
1.6 and 2.4 ms where swelling occurs for smaller line energies experimentally than
numerically. For higher scan velocities of 4.8 and 5.6 ms the reverse effect is observed,
i.e., more swelling occur numerically than experimentally. For vscan = 4.8 ms swelling
arises numerically already for a line energy of 0.25 kJ
m while the experimental setup
is still good. The same is seen for vscan = 5.6 ms ; line energies higher than 0.25 kJ
m
lead to swelling in numerical simulations, but in experimental data only energies
equal 0.3 kJ
m arise swelling.
a http://www.rrze.uni-erlangen.de/dienste/arbeiten-rechnen/hpc/systeme/lima-cluster.shtml
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
ws-ijmpc
Validation Experiments for LBM Simulations of Electron Beam Melting
7
0.8
swelling
porous
good
0.7
Line energy [kJ/m]
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
5
6
7
Scan velocity [m/s]
Fig. 5. Process window of numerical data. ∆x = 5 · 10−6 m, ∆t = 1.75 · 10−7 s, simulated powder
domain (1.44x0.64x0.24)·10−3 m3 , beam offset =13.56·10−3 m.
Both effects can be explained by the focus of the electron beam gun. The focus
is more precise for smaller scan velocities imposed by a smaller electron beam power
Pbeam , e.g, vscan = 1.6 ms or vscan = 2.4 ms lead to a precise, small focus and thus,
energy is brought onto a smaller area. For these velocities swelling occurs for smaller
line energies experimentally than numerically because the EB gun focus is constant
for different beam power and scan velocities, respectively. For higher scan velocities
(larger Pbeam ) like 4.8 ms the focus of the gun spreads and, subsequently, the same
amount of energy is brought into a larger area of powder particles and the maximum
temperature is smaller. As a consequence in experimental data less swelling occurs
for higher scan velocities (larger Pbeam ) than in numerical data.
(a) t = 0.20·10−3 s
(b) t = 4.27·10−3 s
Fig. 6.
Hatching of one layer, (0.15
(c) t = 12.91·10−3 s
kJ
,
m
6.4
m
).
s
m
Fig. 6 shows hatching of one line with parameter set (0.15 kJ
m , 6.4 s ) for three
different time steps. The inverse Gaussian distributed powder particles 4 are demonstrated the and the free surface is visualized by an isosurface. In Fig. 6(a) the
electron beam melted one line and powder particles in this first line are not yet
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
8
ws-ijmpc
Regina Ammer, Matthias Markl, Vera Jüchter, Carolin Körner, Ulrich Rüde
completely melted. The next Fig. 6(b) shows the second melted line by a comingback electron beam. Here, the particles of the first line are now melted completely.
In the third Fig. 6(c) the electron beam has melted seven hatching lines and the
whole powder layer is melted.
4. Conclusion
This paper describes the 3D thermal LB method for the numerical simulation of
EBM processes. Experimental data of hatching a cuboid with different line energies
and scan velocities are shown in a process window. The different samples are classified into porous, good and swelling, depending on the parameter set (line energy,
scan velocity). For the validation all parameter sets are tested numerically and a
numerical process window is generated by 3D simulations of hatching one powder
layer. Numerical and experimental process window are highly concordant especially
in the region of small line energies because parameter sets which are porous in
experimental process window are that also in the numerical one. Only in the region between good and swelling differences exist which are reducible to the focus
of the EB gun. The precision of the focus depends on the beam power but in our
simulations the focus is constant.
In order to accelerate the building process of EBM manufacturing it is important
to extend the process window, i.e., to raise up the question how the process window
will look like for a higher beam power and therefore higher scan velocities. After
the validation experiments in this paper which justify the correct behavior of our
framework and its beneficing to simulate EBM processes we will try to find better
parameter sets regarding quality of EBM products.
Acknowledgments
Our work is supported by the European Union Seventh Framework Program – Research for SME’s with full title “High Productivity Electron Beam Melting Additive
Manufacturing Development for the Part Production Systems Market” and grant
agreement number 286695. We also gratefully thank the German Research Foundation (DFG) for funding the experimental work within the Collaborative Research
Centre 814, project B2.
References
1.
2.
3.
4.
5.
6.
7.
8.
P. Heinl et al., Adv. Eng. Mater. 9 360 (2007).
C. Körner et al., J. Mater. Process. Tech. 211 978 (2011).
M. Markl et al., Procedia Comput. Sci. 18 2127 (2013).
R. Ammer et al, Comput. Math. Appl. 67 318 (2014)
C. Feichtinger et al., J. Comput. Sci. 2 105 (2011).
H. Köstler et al., IT-Information Technology 55 91 (2013).
G.R. McNamara et al., Phys. Rev. Lett. 61 2332 (1988).
U. Frisch et al., Phys. Rev. Lett. 56 1505 (1986).
February 12, 2014 3:6 WSPC/INSTRUCTION FILE
ws-ijmpc
Validation Experiments for LBM Simulations of Electron Beam Melting
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
F. Higuera et al., Europhys. Lett. 9 663 (1989)
F.J. Alexander et al., Phys. Rev. E 47 R2249 (1993)
Y. Chen et al., Phys. Rev. E 50 2776 (1994)
F. Massaioli et al., Europhys. Lett. 21 305 (1993)
X. Shan, Phys. Rev. E 55 2780 (1997)
X. He et al., J. Comput. Phys. 146 282 (1998).
P. Lallemand et al., Int. J. Mod. Phys B 17 41 (2003)
G.R. McNamara et al., J. Stat. Phys. 81 395 (1995).
P. Bhatnagar et al., Phys. Rev. 94 511 (1954).
S. Chen et al., Annu. Rev. Fluid Mech. 30 329 (1998).
X. He et al., Phys. Rev. E 56 6811 (1997)
X. He et al., Phys. Rev. E 55 R6333 (1997)
C. Körner et al., J. Stat. Physics 121 179 (2005).
9
| 5cs.CE
|
arXiv:1709.06441v1 [math.GR] 19 Sep 2017
EVERY GROUP IS THE OUTER AUTOMORPHISM GROUP
OF AN HNN-EXTENSION OF A FIXED TRIANGLE GROUP
ALAN D. LOGAN
Abstract. Fix an equilateral triangle group Ti = ha, b; ai , bi , (ab)i i with
i ≥ 6 arbitrary. Our main result is: for every presentation P of every
countable group Q there exists an HNN-extension TP of Ti such that
Out(TP ) ∼
= Q. We construct the HNN-extensions explicitly, and examples are given. The class of groups constructed have nice categorical and
residual properties. In order to prove our main result we give a method
for recognising malnormal subgroups of small cancellation groups, and we
introduce the concept of “malcharacteristic” subgroups.
1. Introduction
Every group can be realised as the outer automorphism group of some group
[Mat89]. One can ask what restrictions can be placed on the groups involved.
Several authors have achieved results in this vein [Koj88] [GP00] [DGG01]
[BG03] [FM06] [Log15] [Log16a] [Log16b]. Notably, Bumagin–Wise proved
that every countable group Q can be realised as the outer automorphism group
of a finitely generated group GQ [BW05] (here GQ is the kernel of a short
exact sequence inspired by Rips’ construction [Rip82]), and Minasyan proved
that GQ can additionally be taken to have 2-conjugacy classes and Kazdhan’s
property T [Min09] (here GQ is one of Osin’s monster groups [Osi10]).
We therefore say that a class of finitely generated groups C possesses the
Bumagin–Wise property if for every countable group Q there exists a group
GQ ∈ C such that Out(GQ ) ∼
= Q. Hence, Bumagin–Wise prove that a certain class of groups related to Rips’ construction possesses the Bumagin–Wise
property. Minasyan proved that the class of Osin’s monster groups possesses
the Bumagin–Wise property. These two classes of groups are well-known to
possess pathological properties. Therefore, the results of Bumagin–Wise and
Minasyan suggest that the Bumagin–Wise property is a pathological property.
Date: September 20, 2017.
2010 Mathematics Subject Classification. 20F28, 20F65, 20F55.
Key words and phrases. Outer automorphism groups, triangle groups, HNN-extensions,
small cancellation theory, automorphisms of free groups.
1
2
ALAN D. LOGAN
The main result of this paper is Theorem A, below, which proves that certain
classes of “nice” HNN-extensions of “nice” groups possess the Bumagin–Wise
property. Theorem B then says that this construction is functorial. These
results are in stark contrast to the results of Bumagin–Wise and Minasyan,
and suggest that the Bumagin–Wise property is not pathological. By a “nice”
HNN-extension we mean an automorphism-induced HNN-extension; we define
these HNN-extensions in Section 2. By a “nice” group we mean a triangle
group, defined below; such groups are extremely well studied [Mag74].
The proof of Theorem A has two main innovations. Firstly, we give a method
for recognising malnormal subgroups of small cancellation groups (Theorem F
says that if the presentation hx; r, si and the set s possess certain properties
then hsi is a malnormal subgroup of G = hx; ri). Secondly, we introduce and
study the notion of “malcharacteristic” subgroups (which generalise malnormal
subgroups). The key idea of this paper is that the existence of malcharacteristic subgroups in certain triangle groups implies Theorem A.
Main theorems. A triangle group is a group with a presentation of the
following form.
Ti,j,k := ha, b; ai , bj , (ab)k i
If i = j = k we shall write Ti := Ti,i,i for the corresponding equilateral triangle
group. Theorem A reveals a certain universal property possessed by equilateral
triangle groups Ti with i ≥ 6.
By a countable group presentation we mean a presentation P = hx; ri where
|x| ≤ |N| (and hence |r| ≤ |N|). All our presentations are countable so we shall
often omit the adjective “countable”. We also assume that each presentation
P contains at least one generator (that is, |x| ≥ 1); this excludes the empty
presentation of the trivial group. For a group presentation P we write π1 (P)
for the group defined by P. We often abuse notation and write Q = hx; ri for
Q a group (as we did above with Ti,j,k ). Note that a group Q is countable if
and only if there exists a countable presentation P with Q ∼
= π1 (P).
Theorem A. Fix an equilateral triangle group Ti with i ≥ 6. For every
countable group presentation P with at least one generator there exists an
automorphism-induced HNN-extension TP of Ti such that Out(TP ) ∼
= Q and
T
⋊
Q,
where
Q
:=
π
(P).
Aut(TP ) ∼
= P
1
We construct the groups TP from Theorem A in Section 4. We discuss why
we require i ≥ 6 in Theorem A after the proof of Lemma 7.7.
In Section 4 we study properties of the construction underlying Theorem A.
This analysis leads to Theorems B–E. Firstly, in Section 4 we note that countable group presentations form a category Pres, with morphisms P1 → P2
HNN-EXTENSIONS OF TRIANGLE GROUPS
3
corresponding to certain surjective homomorphisms of the groups π1 (P1 ) and
π1 (P2 ). Theorem B then says that the construction of Theorem A is functorial.
Theorem B. The map defined by P 7→ TP is a functor from the category of
countable group presentations Pres to the category of groups Grp.
Theorem B follows from Theorem 4.3. We note after the proof of Theorem 4.3 that if an appropriate subcategory of Pres is chosen then this functor
may be extended to a functor from a subcategory of Grp to Grp.
The construction of Theorem A possesses the following residual property.
Theorem C. Let TFin be the class of groups TQ where π1 (Q) is finite. If π1 (P)
is residually finite then TP is residually-TFin .
Theorem C follows from Theorem 4.4. In Section 8 we prove the following
corollary of Theorem C. Bumagin–Wise asked if every countable group Q can
be realised as the outer automorphism group of a finitely generated, residually
finite group GQ [BW05, Problem 1]. Corollary D gives a positive answer to this
question of Bumagin–Wise for all finitely generated, residually finite groups Q
by taking Q := π1 (P) and GQ := TP .
Corollary D. If the presentation P has a finite, non-empty generating set
and the group π1 (P) is residually finite then the group TP is residually finite.
The construction of Theorem A has a certain flexibility. It allows a choice
of certain subgroups Mn , and this choice may be made in such a way that we
obtain easy examples or additional properties of the groups TP . The above
theorems do not require these subgroups to be chosen in any particular way.
However, the subgroups Mn may be chosen in such a way that the following
result holds, where for presentations P = hx; ri and Q = hy; si we write P ∗ Q
for the presentation hx, y; r, si (so π1 (P ∗ Q) ∼
= π1 (P) ∗ π1 (Q)).
Theorem E. The subgroups Mn , n ∈ N, in the proof of Theorem A may be
chosen in such a way that for every presentation P with finite, non-empty
generating set and for every countable group presentation Q there exists a
surjection TP ։ TP∗Q .
Theorem E follows from Theorem 4.5. The chosen subgroups Mn are specified in Theorem 4.5.
Examples. We now give some simple examples of our construction. These examples illustrate the simplicity of the construction itself (examples of Bumagin–
Wise’s groups and of Minasyan’s groups are much harder to construct) as well
as the construction’s functorial properties. Let P∞ = hz; −i and Pk = hz; z k i
for k ∈ N (we will always assume 0 6∈ N). Fix i ≥ 6. By φ we mean the
automorphism φ : a 7→ b, b 7→ (ab)−1 of Ti = ha, b; ai , bi , (ab)i i which has order
4
ALAN D. LOGAN
three. By x and y we mean specific words over a and b which are given in
Section 3.
TP∞ = hTi , t; ty −j xy j t−1 = φ(y −j xy j ), ∀ j ∈ Zi
TPk = hTi , t; ty k t−1 = φ(y k ), ty −j xy j t−1 = φ(y −j xy j ), ∀ 0 ≤ j < ki
Note that the words y −j xy j , j ∈ Z, generate the kernel of the map F (x, y) →
π1 (P∞ ), x 7→ 1, y 7→ z, while the words y k and y −j xy j , 0 ≤ j < k, generate
the kernel of the analogous map F (x, y) → π1 (Pk ). It is clear that the map
Z/kZ 7→ TPk , k ∈ N ∪ {∞}, is a functor from a subcategory of Grp to Grp.
Triangle groups. In Theorems 7.9–7.12 we obtain results analogous to Theorems A–E where the base group H is any triangle group Ti,j,k , i, j, k ≥ 6.
The price we pay in allowing for more base groups is that the “Out(TP ) ∼
= Q”
condition of Theorem A is weakened to “Q embeds with index one or two in
Out(TP )”.
The constructions of Theorem A and Theorem 7.9 are specific cases of Theorem 4.2. This more general theorem potentially allows for other base groups
H to be taken in Theorem A, and not just triangle groups. However, it is
a highly non-trivial task to apply Theorem 4.2, and so our proofs of Theorem A and Theorem 7.9 use properties of triangle groups throughout. We
focus on triangle groups because they are well-studied groups with many nice
properties. It would be extremely surprising if the universal properties of Theorem A and Theorem 7.9 were to disappear if we replaced the groups H with
less well-behaved groups (for example Thompson’s groups F , T or V ). We
therefore record a quasi-conjecture: If a group H has a suitably rich subgroup
structure (e.g. contains a non-abelian free subgroup) then the class CH of
HNN-extensions of H possesses the Bumagin–Wise property.
Malcharacteristic subgroups. This paper introduces malcharacteristic subgroups, and most of the paper is devoted to their study. Malnormal subgroups
are central objects in geometric group theory, and malcharacteristic subgroups
generalise malnormal subgroups in the same way that characteristic subgroups
generalise normal subgroups.
We formally define malcharacteristic subgroups in Section 3 (the definition
of malnormal subgroups can be found there too). Our focus on malcharacteristic subgroups is because of Theorem 4.2, which can be interpreted as saying
“Theorem A holds if each group Ti , i ≥ 6, contains a malcharacteristic subgroup which is free of rank two”.
In Proposition 5.4 we provide concrete examples of malcharacteristic subgroups in the free group F (a, b). We use these to provide, in Proposition 7.8,
HNN-EXTENSIONS OF TRIANGLE GROUPS
5
concrete examples of malcharacteristic subgroups in the triangle groups Ti,j,k
with i, j, k ≥ 6. Theorem A then follows from Theorem 4.2.
Malnormal subgroups of small cancellation groups. In order to obtain
concrete examples of malcharacteristic subgroups we first need concrete examples of malnormal subgroups. To find these we apply Theorem F, stated
below. Theorem F gives a new method for recognising malnormal subgroups
of small cancellation groups. We refer the reader to Lyndon–Schupp’s classic
text for the definitions of the metric C ′ (1/6) and C ′ (1/4) − T (4) small cancellation conditions [LS77, Section V]. A word W in F (x) is a proper power if
there exists another word U ∈ F (x) and an integer n > 1 such that W = U n .
For a group G = hx; ri, Theorem F links the malnormality of a subgroup M
generated by a set of words s, so M = hsi, with properties of the presentation hx; r, si. Note that Theorem F allows for the explicit construction of
malnormal subgroups of small cancellation groups.
Theorem F. Suppose that hx; r, si is a C ′ (1/6) or C ′ (1/4) − T (4) small cancellation presentation, and that no element of the set s is a proper power. Then
the subgroup hsi of the small cancellation group G = hx; ri is malnormal and
free with basis s.
The assumption that no element of the set s is a proper power is necessary,
as if U n ∈ s with n > 1 then U 6∈ hsi (by small cancellation) but hU n i ≤
(U −1 hsiU ∩ hsi).
Theorem F follows from two more general theorems, Theorems 6.4 and 6.5.
These two theorems give a new method for recognising, for L and M subgroups
of a small cancellation group G, if Lg ∩ M = 1 for all g ∈ G. Also related is
Lemma 6.3, which gives a new method for recognising free subgroups of small
cancellation groups. Note that fibre products of Stalling’s graphs [Sta83, Theorem 5.5] solve for free groups the underlying decision problems of Theorem F
and of Theorems 6.4 and 6.5.
Outline of the paper. In Section 2 we prove Lemma 2.2, a result on
automorphism-induced HNN-extensions which underpins the construction of
Theorem 4.2. In Section 3 we define, and prove present basic results about,
malcharacteristic subgroups. In Section 4 we prove Theorem 4.2; this gives the
construction underlying Theorem A. This section also contains the proofs of
various properties of this construction relating to Theorems B–E. In Section 5
we give concrete examples of malcharacteristic subgroups of the free group
F (a, b). In Section 6 we give concrete examples of malnormal subgroups of
triangle groups Ti,j,k , i, j, k ≥ 6, and in particular we prove Theorem F. In
6
ALAN D. LOGAN
Section 7 we combine the results of the previous sections to give concrete examples of malcharacteristic subgroups of triangle groups Ti,j,k , i, j, k ≥ 6, and
in particular we prove Theorems A, B, C and E. In Section 8 we apply Theorem C to the question of Bumagin–Wise regarding residually finite groups,
and in particular we prove Corollary D. In Section 9 we pose certain questions
which arose in the writing of this paper.
Acknowledgements. I would like to thank Stephen J. Pride and Tara Brendle for many helpful discussions about this paper. I would also like to thank
Jim Howie for observing when the groups in Theorem A are finitely and infinitely presentable (see the discussion following the proof of Theorem 4.2), Peter
Kropholler for ideas which led to Theorem C, and Henry Wilton for pointing
out that triangle groups are LERF.
2. Automorphism-induced HNN-extensions
The main result of this section is Lemma 2.2, which gives certain isomorphisms underpinning Theorem A. We also define automorphism-induced HNNextensions and explain why they are “nice”.
Automorphism-induced HNN-extensions. Let H be a group, K H a
proper subgroup and φ ∈ Aut(H) an automorphism. Then the automorphisminduced HNN-extension of H with associated subgroup K and with associated
automorphism φ is the group with relative presentation
H∗(K,φ) = hH, t; tkt−1 = φ(k), k ∈ Ki.
The groups TP in Theorem A are automorphism-induced HNN-extensions.
That they are automorphism-induced is crucial to our paper. For example,
to prove Theorem B we apply the following fact particular to automorphisminduced HNN-extensions: if K < K0 H then H∗(K,φ) ։ H∗(K0 ,φ) .
Automorphism-induced HNN-extensions are “nice”, in the sense that they
are an easy class of groups to work with and possess nicer properties than general HNN-extensions. The simplest examples of HNN-extensions are Baumslag–
Solitar groups (HNN-extensions of the infinite cyclic group), and these also provide the standard examples of badly-behaved HNN-extensions. For example,
Baumslag–Solitar groups can have non-finitely generated outer automorphism
group [CL83]. However, if a Baumslag–Solitar group is automorphism-induced
then it has virtually cyclic outer automorphism group [GHMR00].
As another example of the tractability of automorphism-induced HNNextensions, we note that the notation H∗(K,φ) emphasises that the second
associated subgroup φ(K) can, in practice, be ignored when studying these
HNN-EXTENSIONS OF TRIANGLE GROUPS
7
groups. This in demonstrated in Theorem 2.1 and Lemma 2.2, where the
descriptions of Out(H∗(K,φ)) mention K but not φ(K).
Describing Out(G). For G = H∗(K,φ), Theorem 2.1 now describes Out(G)
subject to certain conditions. One of the conditions required by Theorem 2.1
is that the group H has Serre’s property FA, that is, every action of H on
any tree has a global fixed point. This is an extremely natural property (see,
for example, [Ser80, Theorem 15]). For δ ∈ Aut(H) we write δb to mean the
outer automorphism δ Inn(H) ∈ Out(H). For g ∈ H we write γg for the inner
automorphism of H acting as conjugation by g, so γg (h) = g −1hg for all h ∈ H.
We define the subgroup AK of Aut(H) as follows:
AK := {δ ∈ Aut(H) : δ(K) = K, ∃ g ∈ H s.t. δφ(k) = φδγg (k) ∀ k ∈ K}
Theorem 2.1 follows immediately from two results from our previous paper
[Log16b, Theorems A & C].
Theorem 2.1. Let G = H∗(K,φ) be an automorphism-induced HNN-extension.
Assume that:
(1) H has Serre’s property FA;
(2) there does not exist any b ∈ H such that φ(K) γb (K) and there does
not exist any c ∈ H such that γc (K) φ(K); and
(3) Z(H) = 1.
Then there exists a short exact sequence:
NH (K)
AK Inn(H)
1 → CH (K) ⋊
→ Out0 (G) →
→1
K
Inn(H)
where either Out0 (G) = Out(G) or there exists some δ ∈ Aut(H) and some
a ∈ H such that δ(K) = φ(K), δ 2 γa (K) = K and δφ−1 (k) = φδγa (k) for all
k ∈ K, whence Out0 (G) has index two in Out(G).
The assumption that Z(H) = 1 means that the short exact sequence here is
easily-digestible. A more extensive description of this sequence in our previous
paper requires no such assumption [Log16b, Theorem B].
Lemma 2.2 now refines the description of Out(G) given by Theorem 2.1.
We shall often use Out0 (G) to mean an index-one or-two subgroup of Out(G).
Then by Aut0 (G) we shall mean the full pre-image of Out0 (G) in Aut(G) under
the natural map (see Lemmas 2.2 and 4.1, and Theorem 4.2).
Lemma 2.2. Suppose H, K and φ are such that the following hold:
(1) H has Serre’s property FA;
(2) CH (K) is trivial;
(3) δ(K) ∩ K = 1 for all automorphisms δ 6∈ Inn(H); and
(4) φ 6∈ Inn(H).
8
ALAN D. LOGAN
Let G = H∗(K,φ) . Then there exists an index-one or-two subgroup Out0 (G)
of Out(G) such that Out0 (G) ∼
= NH (K)/K. Moreover, Aut0 (G) = Inn(G) ⋊
0
0
∼
Out (G), and so Aut (G) = G ⋊ NH (K)/K. Suppose, in addition, that the
following holds:
(5) φb has odd or infinite order in Out(H).
Then Out0 (G) = Out(G) and Aut0 (G) = Aut(G).
Proof. The three conditions of Theorem 2.1 hold here, with the second condition following from Conditions (3) and (4) here as Conditions (3) and (4)
imply that φ(K) ∩ γh (K) = 1 for all h ∈ H. Hence, the description of
Out0 (G) given by Theorem 2.1 holds here. It is therefore sufficient to prove
that AK ≤ Inn(H) that Aut0 (G) = Inn(G) ⋊ Out0 (G), and that if Condition (5) holds then Out0 (G) = Out(G).
Condition (3) implies that δ(K) 6= K for all δ 6∈ Inn(H), and so AK ≤
Inn(G) as required.
That Aut0 (G) = Inn(G) ⋊ Out0 (G) follows from the fact that AK ≤ Inn(H)
[Log16b, Lemma 5.4].
Suppose that Condition (5) holds and that Out0 (G) 6= Out(G). Then there
exists δ ∈ Aut(H) and a ∈ H such that δφ−1 (K) = K = δ 2 γa (K). Now,
b by (3). Therefore, δ 2 6∈ Inn(G), by (4) and
δφ−1 (K) = K implies that δb = φ,
(5). On the other hand, δ 2 γa (K) = K implies that δ 2 ∈ Inn(H), by (3), a
contradiction. Hence, Out0 (G) = Out(G) as required.
3. Malcharacteristic subgroups
In this section we define malcharacteristic subgroups, and present basic results about them. We also state a subgroup M of Ti,j,k which is malcharacteristic and free of rank two. Theorem 4.2 reduces the proof of Theorem A
to proving that this subgroup M is indeed malcharacteristic and free of rank
two; the purpose of Sections 5–7 is to prove these two properties of M.
Definition of malcharacteristic subgroups. A malnormal subgroup intersects its conjugates trivially apart from in the obvious place in the obvious
way. More formally, a subgroup M ≤ H is malnormal in H if the following
implication holds, where hg := g −1 hg and M g := {hg ; h ∈ M}.
M g ∩ M 6= 1 ⇒ g ∈ M
Similarly, a malcharacteristic subgroup intersects its automorphic orbit trivially apart from in the obvious place in the obvious way. More formally, a
subgroup M ≤ H is malcharacteristic in H if M is malnormal in H and for
all δ ∈ Aut(H) the following implication holds.
δ(M) ∩ M 6= 1 ⇒ δ ∈ Inn(H)
HNN-EXTENSIONS OF TRIANGLE GROUPS
9
Basic results. We now make two observations which are applied in, respectively, Lemma 3.3 and Section 4. The proofs of these observations are routine
and so omitted. Lemma 3.1 is dual to the well-known result that if N is normal
in H and C is characteristic in N then C is normal in H.
Lemma 3.1. If M is malcharacteristic in H and L is malnormal in M then
L is malcharacteristic in H.
Note that we may replace the word “malnormal” in Lemma 3.2 with the
word “malcharacteristic”.
Lemma 3.2. If M is malnormal in H and K 6= 1 is a normal subgroup of M
then NH (K) = M.
Free, malcharacteristic subgroups. Lemma 3.3 constructs free, malcharacteristic subgroups from a single “seed” subgroup M. This lemma is central
to Theorem 4.2. We use ∞ to denote the first infinite cardinal |N|.
Lemma 3.3. If a group G contains a malcharacteristic subgroup M which is
free of rank m > 1 (possibly m = ∞) then for each n ∈ N ∪ {∞} the group G
contains a malcharacteristic subgroup Mn which is free of rank n.
Proof. Suppose M = hxi is a malcharacteristic subgroup of G which is free on
the set x, where |x| = m. Let x1 , x2 ∈ x and consider L = hx1 , x2 i (clearly L
is free of rank two). Now, L is malnormal in M and so L is malcharacteristic
in G, by Lemma 3.1. Therefore, again by Lemma 3.1, it is sufficient to prove
that L ∼
= F (x1 , x2 ), the free group of rank two, contains malnormal subgroups
of arbitrary rank (possibly countably infinite).
To see that F (x1 , x2 ) contains malcharacteristic subgroups of arbitrary rank
n ∈ N∪{∞}, first note that for each such n there exists a C ′ (1/6) small cancellation set sn ⊂ F (x1 , x2 ) which has cardinality n and where no element of sn is
a proper power [BW05, Lemma 6]. If n < ∞ then Mn := hsn i is a malnormal
subgroup of F (x1 , x2 ) which is free of rank n [Wis01, Theorems 2.11 & 2.14].
If n = ∞ then write M∞ := hs∞ i. Identities of the form U0 (s∞ ) = 1 or
W (x1 , x2 )−1 U1 (s∞ )W (x1 , x2 ) = U2 (s∞ ) only apply finitely many elements of
s∞ , and so it follows from the finite-rank case that M∞ is a malnormal subgroup of F (x1 , x2 ) which is free of countably-infinite rank.
A malcharacteristic subgroup. Let i, j, k ≥ 6. We record here that the
10
ALAN D. LOGAN
subgroup M = hx, yi, x and y as below with ρ ≫ max(i, j, k), is a malcharacteristic subgroup of Ti,j,k = ha, b; ai , bj , (ab)k i which is free of rank two.
x := (ab−1 )3 (a2 b−1 )3 (ab−1 )3 (a2 b−1 )4 . . . (ab−1 )3 (a2 b−1 )ρ+2
y := (ab−1 )3 (a2 b−1 )ρ+3 (ab−1 )3 (a2 b−1 )ρ+4 . . . (ab−1 )3 (a2 b−1 )2ρ+2
This fact is proven in Lemma 7.7.
4. The construction underlying Theorem A
In this section we prove Theorem 4.2, which proves Theorem A modulo the
existence of a free, malcharacteristic subgroup M of rank two in the group
Ti for each i ≥ 6. We also prove Theorems 4.3–4.5, which are the result
underlying Theorems B–E.
The construction. In the construction of Theorem 4.2 we simply specify the
subgroups Mn and K in Lemma 4.1. Recall that by Aut0 (G) we mean the full
pre-image of Out0 (G) in Aut(G) under the natural map.
Lemma 4.1. Suppose that H and φ ∈ Aut(H) are such that Conditions (1)
and (4) of Lemma 2.2 hold. Suppose that Mn is a malcharacteristic subgroup
of H which is free of rank n > 1, and K 6= 1 is a normal subgroup of Mn .
Let G = H∗(K,φ). Then there exists an index-one or-two subgroup Out0 (G)
of Out(G) such that Out0 (G) ∼
= Mn /K. Moreover, Aut0 (G) = Inn(G) ⋊
0
0
Out (G), and so Aut (G) ∼
= G ⋊ Mn /K.
Suppose, in addition, that φ is such that Condition (5) of Lemma 2.2 holds.
Then Out0 (G) = Out(G) and Aut0 (G) = Aut(G).
Proof. By Lemma 3.2, NH (K) = Mn . Therefore, to prove the result it is sufficient to prove that H, K and φ satisfy Conditions (2) and (3) of Lemma 2.2.
The subgroup Mn is malcharacteristic in H, so (3) holds. We now prove
that CH (K) is trivial, so (2) holds. Suppose that g ∈ H is such that [k, g] = 1
for all k ∈ K. Then g ∈ Mn , by malnormality of Mn , and so CH (K) ≤ Mn .
As Mn is free we have that K is cyclic. Now, non-trivial normal subgroups of
non-cyclic free groups can never be cyclic. Thus, as Mn is free of rank n > 1
we conclude that CH (K) is trivial, as required.
If r ⊆ F (x) then we write hhrii for the normal closure of the set r in F (x),
so F (x)/hhrii = π1 (hx; ri). We say that hy; si is a quotient presentation of
hx; ri if x = y and hhrii hhsii. We now have the main result of this section.
Theorem 4.2. Suppose that the group H has
(1) Serre’s property FA;
(2) non-trivial outer automorphism group; and
(3) a malcharacteristic subgroup which is free of rank two.
HNN-EXTENSIONS OF TRIANGLE GROUPS
11
Then:
i. For every countable group presentation P with at least one generator
there exists an automorphism-induced HNN-extension HP of H such that
Out0 (HP ) ∼
= π1 (P), where Out0 (HP ) is an index-one or-two subgroup of
Out(HP ). Moreover, Aut0 (HP ) = Inn(HP )⋊Out0 (HP ), and so Aut0 (HP ) ∼
=
HP ⋊ π1 (P).
ii. For P1 and P2 countable group presentations with at least one generator, if
P2 is a quotient presentation of P1 then there is a surjection HP1 ։ HP2 .
Suppose, in addition, that the group H has
(4) a non-inner automorphism φ such that φb ∈ Out(H) has odd or infinite
order.
Then Out0 (HP ) = Out(HP ) and Aut0 (HP ) = Aut(HP ).
Proof. By Lemma 3.3, and as H contains a malcharacteristic subgroup which
is free of rank two, the group H contains malcharacteristic subgroups Mn
which are free of rank n for any given n ∈ N ∪ {∞}. Fix for each n ∈ N ∪ {∞}
such a subgroup Mn .
Let P = hx; ri be a given presentation. Let HP be the automorphisminduced HNN-extension H∗(KP ,φ) of H where φ ∈ Aut(H) is a non-inner automorphism (with φb of odd or infinite order if such an automorphism exists) and
b = hx, z0 ; r, z0 i, z0 6∈ x±1 ,
where KP is as follows: Consider the presentation P
b = x ⊔ {z0 } and b
and we shall write x
r = r ⊔ {z0 } (so Pb = hb
x; b
ri). Note that
∼
b
x| > 1, and that b
r is non-empty. Choose M|bx| to be
π1 (P) = π1 (P), that |b
malcharacteristic in H of rank |b
x| and take KP to be the normal subgroup of
M|bx| associated with hhb
rii, the normal closure in F (b
x) of b
r.
By Lemma 4.1, this construction proves Point (i) of the theorem. Also by
Lemma 4.1, if there exists φb ∈ Out(H) which has odd or infinite order then
Out0 (HP ) = Out(HP ) and Aut0 (HP ) = Aut(HP ).
To see that this construction proves Point (ii) of the theorem, note that as
P2 = hx; si is a quotient presentation of P2 = hx; ri we have that hhr, z0 ii
hhr, s, z0 ii ≤ F (x, z0 ). Therefore, adding the relators tUt−1 φ(U)−1 to HQ ,
where U ∈ M|bx| corresponds to an element of hhr, s, z0 ii \ hhr, z0ii, induces the
required group homomorphism HP1 ։ HP2 .
Note that because the subgroup KP in the proof of Theorem 4.2 is free,
the presentation of the HNN-extension HP = H∗(KP ,φ) is aspherical, and so
minimal [CCH81, Theorem 3.1]. Thus, for P = hx; ri, the group HP in the
construction is finitely presented if and only if the set x and the group π1 (P) ∼
=
Out(HP ) are both finite.
Proving Theorem A. To obtain the groups TP from Theorem A, take in
12
ALAN D. LOGAN
Theorem 4.2 the group H to be Ti , the automorphism φ to be a 7→ b, b 7→
b−1 a−1 , and the malcharacteristic subgroup in Condition (3) to be the subgroup
M := hx, yi stated in Section 3, above.
Theorem 7.9 is a related construction, and also follows from Theorem 4.2.
Indeed, Theorem 4.2 proves Theorem A and Theorem 7.9 modulo proving that
the subgroup M = hx, yi of Ti,j,k , i, j, k ≥ 6, is malcharacteristic and free of
rank two; that is, modulo proving that Condition (3) of Theorem 4.2 holds
for these triangle groups. This is because the groups Ti,j,k are well-known
to possess Conditions (1) and (2), while the groups Ti additionally possess
Condition (4) (the map φ : a 7→ b, b 7→ b−1 a−1 is non-inner of order three).
Note that non-equilateral triangle groups have outer automorphism groups of
order two or four (see Lemma 7.3), so Condition (4) does not hold for the
triangle groups of Theorem 7.9. Theorem A and Theorem 7.9 are formally
proven in Section 7.
Functorial properties. Countable group presentations form a category in
the following sense: We define the equivalence relation ∼ on presentations as
hx; ri ∼ hy; si if x = y and hhrii = hhsii. Form the category Presx whose
objects are group presentations with generating set x, under the equivalence
relation ∼ and with morphisms Pj → Pk if Pk is a quotient presentation of
Pj . Note that the normal subgroups of F (x) form a category, with morphisms
obtained from the usual subgroup partial order, and this category is naturally
isomorphic to the category Presx . Note also that the categories Presx and
Presy are isomorphic if |x| = |y|. Therefore, there exists a well-defined category
of n-generator presentations Presn , n ∈ N ∪ {∞}. The category of countable
group presentations Pres is the union of the categories Presn for n ∈ N ∪ {∞}.
Theorem 4.3 now says that the construction of Theorem 4.2 is functorial.
Theorem 4.3. The map defined by P 7→ HP is a functor from the category
of countable group presentations Pres to the category of groups Grp.
Proof. If there exists a morphism Pj → Pk then Pk is a quotient presentation
of Pj . Therefore, HPj ։ HPk by Theorem 4.2.ii.
Let Q be a fixed group and let Q be the category whose objects are quotient
groups Q/Nk and whose morphisms correspond to inclusion of kernels, so
Q/Nj → Q/Nk if Nj
Nk . Fixing a presentation of P of Q, the category
Q is isomorphic to the subcategory of Pres with initial object P. Hence, the
functor of Theorem B also encodes information about quotient groups.
Write Pk for the image of Q/Nk under the aforementioned isomorphism of
categories. Suppose that the group Q is such that Q/Nj ∼
= Q/Nk implies
Nj = Nk . Then Q is isomorphic to a subcategory of Grp and hence the map
Q/Nk 7→ TPk is a functor from a subcategory of Grp to Grp.
HNN-EXTENSIONS OF TRIANGLE GROUPS
13
On the other hand, it is impossible to chose for each group Q a presentation
P with π1 (P) = Q such that the map Q 7→ TP , factoring as Q 7→ P 7→ TP ,
is functorial; this is because there exist non-isomorphic groups Q1 and Q2
such that Q1 ։ Q2 ։ Q1 . Hence Theorem 4.3 cannot be extended in the
natural way. Indeed, Q1 and Q2 can be taken to be 2-generator, 1-relator
groups [BM06], so if we fix n > 1 and restrict to all finitely presentable ngenerator groups then the map Q 7→ TP can still not be made functorial. The
examples from the introduction show that if n = 1 then this map can be made
functorial.
Residual Properties. Let HFin be the class of groups HP where π1 (P) is
finite. Note that these presentations P are countable presentations of finite
groups. In particular, if the input presentation P = hx; ri in Theorem 4.4 has
infinite generating set, so |x| = ∞, then the presentations Pg = hx; r, si in the
following proof are presentations of finite groups but with infinite generating
set x. It follows that if |x| = ∞ then the associated subgroups KPg cannot be
finitely generated. If U and V are words then we write U ≤ V to mean that
U is a subword of V .
Theorem 4.4. If π1 (P) is residually finite then HP is residually-HFin .
Proof. Recall that HP is an HNN-extension H∗(KP ,φ) . Let g ∈ H ∗(KP ,φ)
\{1} be arbitrary. By Britton’s lemma, g is represented by a word W =
h0 tǫ1 h1 · · · tǫm hm which does not contain any subwords of the form tkt−1 or
t−1 φ(k)t for any k ∈ KP . For those hi in NH (KP ) such that thi t−1 ≤ W , write
hi := hi KP ∈ Mn /KP , and for those hi in NH (φ(KP )) such that t−1 hi t ≤ W ,
write hi := φ−1 (hi )KP ∈ Mn /KP . As Mn /KP ∼
= π1 (P) is residually finite
there exists a finite group Pg and a homomorphism σ g : Mn /KP → Pg such
that σ g (hi ) 6= 1 for each hi (if there are no hi , so for example if every hi 6∈
NH (KP ) ∪ NH (φ(KP )), then we may take Pg to be trivial). Now, there exists
a quotient presentation Pg = hx; r, si of P = hx; ri which corresponds to the
map σ g . Consider the group TPg = H∗(KPg ,φ) from Theorem A. Then, the
map σg : TP ։ TPg from Therorem 4.2.ii is obtained by adding the relators
tUt−1 φ(U)−1 , where U corresponds to an element of hhr, s, z0 ii \ hhr, z0 ii ⊆
F (x, z0 ). In particular, this map naturally corresponds to the map σ g , and so if
hi ∈ NH (KP ) and thi t−1 ≤ W then hi 6∈ KPg , and similarly if hi ∈ NH (φ(KP ))
and t−1 hi t ≤ W then hi 6∈ φ(KPg ). Note also that NH (KP ) = NH (KPg ).
We now prove that σg (g) 6= 1, which proves the result. Suppose thi t−1 ≤
W . If hi ∈ NH (KP ) then hi 6∈ KPg by the above. On the other hand, if
hi 6∈ NH (KP ) then hi 6∈ KPg ≤ NH (KP ). Hence, W does not contain any
subword of the form tkt−1 for any k ∈ KPg . Similarly, W does not contain any
14
ALAN D. LOGAN
subword of the form t−1 φ(k)t for any k ∈ KPg . That σg (g) 6= 1 now follows
from Britton’s Lemma.
Choosing the subgroups Mn . Let M0 = hx0 , y0 i denote the malcharacteristic subgroup of H which is free of rank two from Theorem 4.2. Now, the
subgroup Mn in the proof of Theorem 4.2 may be taken to be any subgroup
of M0 which is generated by a small cancellation subset sn of F (x0 , y0 ), such
that |sn | = n and sn contains no proper powers. This flexability allows for
concrete examples of this construction to be written down (for example, the
groups TP∞ and TPk from the introduction). Theorem 4.5 now demonstrates
that if we are more strict in our choice of the subgroups Mn then we can obtain
more properties of the groups HP from Theorem 4.2.
Suppose x ⊂ y, and let r ⊂ F (x). Then we write hhriiF (x) for the normal
closure of r in the free group F (x), and hhriiF (y) for the normal closure of r
in the free group F (y). Note that if we view F (x) as a subgroup of F (y) in
the natural way then hhriiF (x) ≤ hhriiF (y) .
Theorem 4.5. Let H be as in Theorem 4.2.
Then H contains a malcharacteristic subgroup hz0 , z1 , z2 , . . .i which is free
on the infinite set {z0 , z1 , z2 , . . .}. Suppose that the subgroups Mn , n ∈ N, in
the proof of Theorem 4.2 are taken to be the subgroups Mn := hz0 , z1 , . . . , zn i,
and that the subgroup M∞ is taken to be hz0 , z1 , z2 , . . .i. Then for every presentation P with finite, non-empty generating set and for every countable group
presentation Q there exists a surjection HP ։ HP∗Q .
Proof. Write zn := {z1 , z2 , . . . , zn } for n ∈ N and z∞ := {z1 , z2 , . . .}. Suppose P = hx; r(x)i and Q = hy; −i. Note that HP has associated subgroup
KP = hhr(z|x|), z0 iiF (z0 ,z|x| ) (we identify the symbol z0 here with the symbol
z0 in Theorem 4.2). Then HP∗Q is obtained from HP by adding the relators
tUt−1 φ(U)−1 to HP , for all U ∈ hhz0 , r(z|x|)iiF (z0 ,z|x⊔y| ) \ hhz0 , r(z|x|)iiF (z0 ,z|x| ) .
This induces the required group homomorphism HP ։ HP∗Q . The result then
follows for all Q from Theorem 4.2.ii.
The free product is the coproduct in the category of groups, and so Theorem 4.5 hints at the notion of a coproduct in our setting. However, it is not
clear whether or not HQ always surjects onto HP∗Q , even if Q has a finite
generating set. Consider the subcategory Pres<∞ := ∪n∈N Presn of Pres. If
the surjection HQ ։ HP∗Q always exist for Q ∈ Pres<∞ then adding the
morphisms P → P ∗ Q and Q → P ∗ Q to the category Pres<∞ to produces a new category Pres<∞ (possibly with a coproduct), and then the map
Pres<∞ → Grp given by P 7→ HP is still functorial.
HNN-EXTENSIONS OF TRIANGLE GROUPS
15
5. Malcharacteristic subgroups of free groups
In this section we find concrete examples of malcharacteristic subgroups
of the free group F (a, b). To find such subgroups we use fibre products of
Stallings’ graphs to obtain an algorithm which decides whether or not the
malcharacteristic property holds for certain subgroups of F (a, b).
The examples we obtain are used to find similar examples in triangle groups.
f{x,y} be the subgroup of F (a, b) generated by the words x and y
Indeed, let M
given in Section 3 (so generated by the same words which generated the subf{x,y} is malcharacteristic
group M of Ti,j,k ). It follows from Lemma 5.3 that M
f{x,y} is malcharacteristic in F (a, b) to
in F (a, b). We later use the fact that M
prove that M is malcharacteristic in Ti,j,k .
Maps of graphs. The principal device applied in this section is that of
maps of graphs; see Stallings’ paper for more details as well as for the relevant
definitions and terminology [Sta83].
The key tool when studying maps of graphs is Stallings’ folding algorithm
[Sta83, Section 3.2]. The algorithm begins with a map of graphs Γ1 → Γ,
and factors this map as a composition Γ1 → Γ2 → · · · → Γk → Γ such that
Γk → Γ is a locally injective map of graphs (and therefore is π1 -injective [Sta83,
Proposition 5.3]), and such that for 1 ≤ j < k the map Γj → Γj+1 is a folding
map. The graph Γk is uniquely determined, although in general the individual
foldings maps are not unique. A map of graphs Γ1 → Γ is folded if it cannot
be factored using the above algorithm.
Let F (x) be a free group with basis x. The basis x corresponds naturally to
a bouquet of circles Γx , which are directed and labeled by the letters xi ∈ x.
Any word in F (x) determines a closed path in Γx . Let s = {W1 , W2 , . . . , Wn }
be a set of words in F (x). Then this set corresponds naturally to a bouquet
Γs of n circles, with circles corresponding to the Wi ∈ s. There exists a
natural map Γs → Γx which sends the i-th loop of Γs to the closed path of
Γx corresponding to Wi . The Stallings’ graph of s is the graph Γs equipped
with the natural labeling induced by the map Γs → Γx . A Stallings’ graph Γs
bs to denote
is folded if the corresponding map Γs → Γx is folded. We use Γ
the folded Stallings’ graph of s, so the unique folded graph, equipped with the
bs → Γ. If C = hsi
natural map, such that the map Γs → Γx factors as Γs → Γ
bs )) then
bs have the same rank (that is, π1 (Γs ) ∼
and the graphs Γs and Γ
= π1 (Γ
bs ) ֒→ π1 (Γx ) corresponds to the map C ֒→ F (x).
the induced map π1 (Γ
Fibre products. The proofs in this section apply fibre products in the category of graphs [Sta83, Theorem 5.5]. If C = hsi is a subgroup of F (x) then
C is malnormal in F (x) if and only if the non-diagonal components of the
16
ALAN D. LOGAN
bs ⊗ Γ
bs form a forest. Fibre products are computable for finitely
fibre-product Γ
generated subgroups of free groups, and hence malnormality is decidable for
such subgroups of free groups. Also, if C = hsi and D = hti are subgroups of
bt ⊗ Γ
bs
F (x) then D ∩ C g = 1 for all g ∈ F (x) if and only if the fibre-product Γ
b
b
is a forest. In particular, if δ ∈ Aut(F (x)) then by computing Γδ(s) ⊗ Γs it is
decidable whether or not δ(C) ∩ C g = 1 for all g ∈ F (x).
Length preserving automorphisms. An automorphism α ∈ Aut(F (a, b))
is length-preserving if |α(a)| = 1 = |α(b)|. We begin with the following lemma,
which follows immediately from a result of Cohen–Metzler–Zimmermann [CMZ81,
Statement 3.9]. For words s ∈ F (a, b) we write s+ to mean the subsemigroup
of F (a, b) generated by the words s.1
Lemma 5.1. For all δ ∈ Aut(F (a, b)) there exists an inner automorphism
γ ∈ Inn(F (a, b)), length-preserving automorphisms α1 , α2 , and an integer m ∈
Z such that one of the following occurs:
(1) δγ is length-preserving; or
(2) α2 δα1 γ(a) = abm , α2 δα1 γ(b) = b with m 6= 0; or
(3) δα1 γ(a), δα1 γ(b) ∈ {abm , abm+1 }+ .
The length-preserving automorphism α1 corresponds to trivial changes of
notation in the images of a and b, for example replacing {abm , abm+1 }+ with
{bam , bam+1 }+ . The length-preserving automorphism α2 in Case (2) allows for
the images to be swapped, for example if δ is the map δ : a 7→ b, b 7→ abm then
we take α2 : a 7→ b, b 7→ a and take α1 and γ to be trivial.
Determining the malcharacteristic property. Lemma 5.2 now gives an
algorithm which decides if a given subgroup from a certain class C of subgroups
of F (a, b) is malcharacteristic or not. Note that it is decidable if a finite set of
words s satisfies the conditions of Lemma 5.2 (hence, membership of the class
C is decidable). Note also that the subsemigroup {a2 , a3 , b2 , b3 }+ of F (a, b)
consists precisely of the freely reduced words W in F (a, b) which do not contain
a−1 or b−1 and such that every instance of a in W is part of an a2 -term and
every instance of b in W is part of a b2 -term. Therefore, in Lemma 5.2 every
b s is part of an
instance of a in any circuit in the folded Stallings’ graph Γ
2
a -term, and similarly for b.
Lemma 5.2. Let s be a finite set of pairwise non-equal words contained in
the subsemigroup {a2 , a3 , b2 , b3 }+ of the free group F (a, b), and write C := hsi.
1
We use the “Kleene plus” s+ (subsemigroup) rather than the “Kleene star” s∗ (submonoid) as we reserve the notation s∗ for the symmetrised closure (see Section 6).
HNN-EXTENSIONS OF TRIANGLE GROUPS
17
b s contains some a3 Suppose that every circuit in the folded Stalling’s graph Γ
term and some b3 -term. Then C is malcharacteristic in F (a, b) if and only
if C is malnormal in F (a, b) and α(C) ∩ C g = 1 for all non-trivial lengthpreserving automorphism α ∈ Aut(F (a, b)) and all g ∈ F (a, b). In particular,
it is decidable whether C is malcharacteristic in F (a, b) or not.
Proof. First, suppose that C is malnormal in F (a, b) and that α(C) ∩ C g = 1
for all length-preserving automorphisms α ∈ Aut(F (a, b)) and all g ∈ F (a, b).
We show that for all non-inner δ ∈ Aut(F (a, b)) the fibre product of the folded
bδ(s) ⊗ Γ
bs is a forest; as C is malnormal this implies that C
Stallings’ graphs Γ
is malcharacteristic in F (a, b). To show that this condition holds for all such
δ ∈ Aut(F (a, b)) we consider the three cases of Lemma 5.1. The result holds
by assumption if δ ∈ Aut(F (a, b)) is such that Case (1) of Lemma 5.1 holds.
Now, suppose δ ∈ Aut(F (a, b)) is such that Case (2) of Lemma 5.1 holds,
so α2 δα1 γ(a) = abm , α2 δα1 γ(b) = b for some m ∈ Z \ {0}, γ ∈ Inn(F (a, b))
and α1 , α2 ∈ Aut(F (a, b)) length-preserving. Then
α2 δα1 γ(s) ⊆ {(abm )2 , (abm )3 , b2 , b3 }+
⊆ {abm a, (abm )2 a, b}+ =: s1 .
However, no element of s1 contains an a3 -term. Hence no circuit in the folded
bα δα γ(s) contains an a3 -term, and as γ is inner the same holds
Stallings’ graph Γ
2
1
for Γα2 δα1 (s) . As α1 , α2 are length-preserving, we further have that either no
bδ(s) contains an a3 -term or no circuit in Γ
bδ(s) contains a b3 -term.
circuit in Γ
bδ(s) ⊗ Γs is a forest, as required.
Therefore, the fibre product Γ
Finally, suppose δ ∈ Aut(F (a, b)) is such that Case (3) of Lemma 5.1 holds,
so δα1 γ(a), δα1 γ(b) ∈ {abm , abm+1 }+ for some m ∈ Z, γ ∈ Inn(F (a, b)) and
α1 ∈ Aut(F (a, b)) length-preserving. Then there exist positive words Ua and
Ub such that:
δα1 γ(s) = {Ua (abm , abm+1 )2 , Ua (abm , abm+1 )3 , Ub (abm , abm+1 )2 , Ub (abm , abm+1 )3 }
⊆ {abm , abm+1 }+ =: s2 .
If m 6= 0, −1 then no element of s2 contains an a3 -term. If m = 0 or m = −1
then s2 = {a, abǫ }+ , ǫ = ±1, and so no element of s2 contains a b3 -term.
b δα γ(s) contains an a3 -term or
Hence, no circuit in the folded Stallings’ graph Γ
1
3
b
no circuit in Γδα1 γ(s) contains a b -term. Therefore, as γ is inner and as α1 is
bδ(s) contains an a3 -term or no circuit
length-preserving, either no circuit in Γ
b δ(s) contains a b3 -term. Hence, the fibre product Γ
b δ(s) ⊗ Γs is a forest, as
in Γ
required.
18
ALAN D. LOGAN
The opposite direction is trivial. The decidability statement follows because
there are only finitely many length-preserving automorphisms α ∈ Aut(F (a, b)).
Examples of malcharacteristic subgroups. Lemma 5.3 now gives an exf of F (a, b).
ample of a malcharacteristic subgroup M
f be the subgroup of F (a, b) which is generated by the folLemma 5.3. Let M
lowing elements, with ρ ≫ 1.
x
e := a3 b3 a3 b4 . . . a3 bρ+2
ye := a3 bρ+3 a3 . . . a3 b2ρ+2
f is a non-cyclic malcharacteristic subgroup of F (a, b).
Then M
Proof. First, note that s = {e
x, ye} ⊆ {a2 , a3 , b2 , b3 }, and that every circuit
bs contains some a3 -term and some b3 -term.
in the folded Stalling’s graph Γ
Therefore, we may apply Lemma 5.2. Now, the words x
e and ye satisfy Wise’s
f is nonc(5) small cancellation condition for ρ ≫ 1, and so the subgroup M
cyclic and malnormal in F (a, b) [Wis01, Theorems 2.11 & 2.14].
Suppose that α ∈ Aut(F (a, b)) is non-trivial and length-preserving, and
f ∩M
fg 6= 1 for some g ∈ F (a, b). Then the fibre product Γα(s) ⊗ Γs
that α(M)
is not a forest. If we pick a circuit in this fibre product Γα(s) ⊗ Γs then this
circuit contains a subpath with label U = (a3 bp a3 bp+1 a3 bp+2 a3 bp+3 a3 bp+4 a3 )ǫ ,
where p > 3 and ǫ = ±1, as every circuit in Γs contains a subpath with a
label of this form. By traversing our chosen circuit in the opposite direction
if necessary, we may assume ǫ = 1.
The set α(s) is C ′ (λ) small cancellation, with λ arbitrarily small (as ρ ≫ 1).
Therefore, U is either a subword of an element of α(s)±1 (so of α(e
x)±1 or
±1
±1
of α(e
y ) ), or a subword of the product of two elements of α(s) . Hence a
subword a3 bq a3 bq+1 a3 of U is a subword of either α(e
x)±1 or of α(e
y)±1 , q > 3. It
is then clear that α is trivial (one way to see this is to note that α−1 is lengthpreserving and then to apply each non-trivial length-preserving automorphism
to a3 bq a3 bq+1 a3 and to a3 bq a3 bq+1 a3 ; the result is never a subword of x
e nor of ye).
g
f
f
This is a contradiction, so α(M ) ∩ M = 1 for all g ∈ F (a, b) as required.
f{x,y} and M
f are in the same automorphic orbit of F (a, b).
The subgroups M
f{x,y} is malcharacteristic in F (a, b). In Lemma 7.7 we
Hence, by Lemma 5.3, M
use this fact to prove that M is malcharacteristic in Ti,j,k .
Lemmas 3.3 and 5.3 combine to prove the following result.
Proposition 5.4. The free group F (a, b) contains malcharacteristic subgroups
fn of arbitrary rank n (possibly countably infinite).
M
HNN-EXTENSIONS OF TRIANGLE GROUPS
19
fn is generated by n finite subwords satisfying C ′ (1/6)
Indeed, the subgroup M
of the following infinite word:
x
eye(e
xye2 )e
xye(e
xye2)2 x
eye(e
xye2)3 · · · .
6. Malnormal subgroups from small cancellation presentations
The main results of this section are Theorems 6.4 and 6.5. We also prove
Theorem F, which is a special case of these two theorems. Theorem 6.5 is
applied in Section 7 to prove that the subgroup M of Ti,j,k , i, j, k ≥ 6, as
stated at the end of Section 3, is a free and malcharacteristic subgroup of
Ti,j,k ; Theorem A and Theorem 7.9 then follow by Theorem 4.2.
The results of this section lift conjugation in groups hx; ri to conjugation
in the ambient free group F (x). Our principal tool here is small cancellation
theory (see [LS77, Section V] for the basic definitions and results), and our key
innovation is to prove results about the subgroup hsiG of the group G = hx; ri
by studying the presentation hx; r, si. Theorems 6.4 and 6.5 give a new method
for recognising, for L and M subgroups of a small cancellation group G, if
Lg ∩ M = 1 for all g ∈ G.
Theorem 6.4 is not applied in this paper. It is included for the sake of
completeness, and because we believe that Theorem F and Theorems 6.4 and
6.5 are of general interest. These results can be viewed as a generalisation
of results of Wise, who applied small cancellation conditions on graphs to
give a method of recognising malnormal subgroups of free groups [Wis01].
We conclude this section by noting that our proofs can be adapted to Wise’s
graphical small cancellation setting.
6.1. Free subgroups of small cancellation groups. We start with Lemma
6.1, which describes how to view elements of the subgroup hsi of hx; ri as
diagrams over the presentation hx; r, si. We then apply this view to prove
Lemma 6.3, below, which gives a new method for recognising free subgroups
of small cancellation groups.
If r is a subset of F (x) then we write r∗ for the symmetrised closure of r,
that is, r∗ is the set of all cyclic shifts, inverses, and cyclic shifts of inverses
of elements of r. By an r∗ -diagram we mean a diagram over the presentation
hx; ri. Note that an r∗ -diagram is also an (r ∪ s)∗ -diagram. By an r∗ -disk
diagram we mean a diagram ∆ where R \ ∆ has exactly one component. By
an r∗ -region R we mean an r∗ -diagram which corresponds to a single relator
R ∈ r∗ . We call R a region if the set of relators is understood. A word
W ∈ F (x) is cyclically reduced if every cyclic shift of W is freely reduced.
Lemma 6.1. Let hx; si be a C(m) presentation, m ≥ 3. For every word
U = S1ǫ1 · · · Snǫn , Si ∈ s, there exists an s∗ -disk diagram DU with boundary
20
ALAN D. LOGAN
labels precisely the cyclically reduced cyclic shifts of U, and which has n regions
Si , 1 ≤ i ≤ n. For each region Si the intersection of Si with the boundary ∂DU
is connected and contains an edge, so pi := Si ∩ ∂DU is a path. Each such path
pi is the product of at least m − 2 pieces.
Proof. First form the bouquet of n petals labeled clockwise with the words
S1 , . . . , Sn respectively (different petals may have the same label). This defines,
in the obvious way, an s∗ -disk diagram ∆U with n regions Si such that Si
is a label of Si . Form the s∗ -disk diagram DU by taking ∆U and folding
small-cancellation pieces in adjacent petals together, that is, perform certain
Stallings’ folds on the edges of ∆U . Applying the C(m) condition, pi :=
Si ∩ ∂DU is connected and contains at least m − 2 pieces. Clearly some, and
hence every cyclically reduced cyclic shift of U is a label of ∂DU , and the result
follows.
Suppose that two diagrams ∆1 and ∆2 have a paths on their boundaries
with a common label, so there exist paths I1 ⊂ ∂∆1 and I2 ⊂ ∂∆2 such I1
and I2 have the same label. Then we may form a new diagram ∆ = ∆1 ∪ ∆2
where we equate the paths I1 and I2 . We call the image I ⊂ ∆ of the paths
I1 and I2 the attaching path of ∆ over ∆1 and ∆2 , or just the attaching path
if ∆, ∆1 and ∆2 are understood.
Lemma 6.2. Let hx; r, si be a C(4) presentation and let U ∈ F (x). Suppose
that ∆ is an (r ∪ s)∗ -diagram containing the subdiagram DU , formed as in
Lemma 6.1. If R is an r∗ -region of ∆ which attaches to DU over an attaching
path I then I cannot straddle three paths pi , pi+1 (mod n) and pi+2 (mod n) of DU .
Proof. Suppose I straddles three paths pi , pi+1 (mod n) and pi+2 (mod n) of DU .
Then pi+1 (mod n) contains one piece, a contradiction as pi+1 (mod n) cannot be
written as the product of fewer than two pieces by Lemma 6.1.
For a word W ∈ F (x) and a rational number c we write W > cr to mean
that there is a relator R ∈ r∗ such that W is a subword of R with |W | > c|R|.
A word W ∈ F (x) is called r∗ -reduced if W is freely reduced and there is
no subword U of W with U > 12 r. We call W cyclically r∗ -reduced if every
free reduction of every cyclic shift of W is non-empty and r∗ -reduced. Note
that a non-cyclically reduced word may be cyclically r∗ -reduced. We call W
(cyclically) Dehn reduced, rather than (cyclically) r∗ -reduced, if the underlying
set of relators r is understood.
Note that for s ⊂ F (x), if a word W ∈ hsi is freely/cyclically reduced as
a word over s±1 then it is not necessarily freely/cyclically reduced over F (x).
We therefore have to be careful what we mean by “free reduction”. If s ⊂ F (x)
we shall write Us (s), Vs (s), and so on, to mean words over the set s±1 which
HNN-EXTENSIONS OF TRIANGLE GROUPS
21
are freely reduced over s, and we shall write Ux (s), Vx (s), and so on, to mean
words over the set s±1 which are freely reduced over x.
By an r∗ -piece we mean a piece relative to the set r∗ . If G is given by the
generating set x and s ⊂ F (x) then we use hsiG to denote the subgroup of G
generated by the words s.
Lemma 6.3. Let G be given by a C ′ (1/6) or C ′ (1/4) − T (4) presentation
hx; ri. Suppose that hx; r, si is a C ′ (1/4) presentation. Then every word Us (s)
is cyclically r∗ -reduced, and hence hsiG is a free subgroup of hx; ri with basis
s.
Proof. Suppose that there exists a non-empty word U := Us (s) = S1ǫ1 . . . Snǫn ,
Si ∈ s, which is cyclically reduced as a word over s±1 but is not cyclically
r∗ -reduced. We obtain a contradiction.
Consider the (non-empty) diagram DU given by Lemma 6.1. As U is not
cyclically r∗ -reduced we may attach an r∗ -region R to the boundary ∂DU of the
diagram DU over an attaching path I such that |I| > 21 |∂R| [LS77, Theorems
V.4.5 & V.4.6].
By Lemma 6.2, the attaching path I straddles at most two paths pi and
pi+1 (mod n) of ∂DU , and hence the path I contains at most two (r ∪ s)∗ -pieces.
Therefore, |I| < 42 |∂R| = 21 |∂R|, a contradiction.
6.2. Conjugacy in small cancellation groups. Theorems 6.4 and 6.5 lift
conditions on conjugation and intersections of subgroups in small cancellation
groups hx; ri to the analogous situation in the ambient free group F (x).
An annular r∗ -diagram A is an r-diagram such that R2 \ A has exactly two
components. Such a diagram has an interior boundary ∂I A and an exterior
boundary ∂E A. Annular diagrams encode conjugation, and for C ′ (1/6) and
C ′ (1/4) − T (4) small cancellation presentations there are extremely strong
structural theorems for annular diagrams [LS77, Theorems V.5.3 & V.5.5]. In
particular, if A is an annular r∗ -diagram where hx; ri is a C ′ (1/6) or C ′ (1/4) −
T (4) small cancellation presentation then either every region of A has edges
on both the interior boundary and the exterior boundary of A, or no region
has this property. In the case that no region has this property, every region
has connected intersection with one of ∂I A of ∂E A, and each such intersection
contains at least one edge (see [LS77, Figure V.5.2]).
Note that if hx; r, si is a C ′ (1/6) presentation then the sub-presentation
hx; ri is also a C ′ (1/6) presentation. If I is a path in a graph then we use ι(I)
to denote the initial vertex of I and τ (I) to denote the terminal vertex of I.
Theorem 6.4. Suppose that hx; r, si is a C ′ (1/6) small cancellation presentation. Let G be the group given by the sub-presentation hx; ri, and let
ϕ : F (x) → hx; ri be the natural map.
22
ALAN D. LOGAN
Suppose that t ⊂ F (x) is a set of words such that every word Vt (t) is
cyclically r∗ -reduced. For all U ∈ hsi and V ∈ hti, if there exists g ∈ G such
that g −1ϕ(U)g =G ϕ(V ) then there exists W ∈ F (x) such that ϕ(W ) = g and
W −1 UW =F (x) V . In particular, there exists g ∈ G such that hsigG ∩ htiG 6=G 1
if and only if there exists W ∈ F (x) such that hsiW ∩ hti =
6 F (x) 1.
Proof. First note that every word Us (s) is cyclically r∗ -reduced by Lemma 6.3.
Therefore, suppose that A is an annular r-diagram such that its interior boundary ∂I A has a label which is a cyclic shift of some word U := U(s) = S1ǫ1 · · · Snǫn ,
Si ∈ s, and its exterior boundary ∂E A has a label which is a cyclic shift of
some word V (t). It is sufficient to prove that A contains no regions.
Consider the disk diagram DU from Lemma 6.1. Each of the paths pi on the
boundary of DU contains at least five (r ∪ s)∗ -pieces. Note that we can form
an (r ∪ s)∗ -disk diagram ∆ by attaching the boundary ∂DU to the boundary
∂A of A. Suppose that A contains at least one region; we find a contradiction.
As A contains a region, the boundary of DU connects to a region R of A
over an attaching path I such that I contains an edge.
First, suppose that R∩∂E A is empty. By the structural theorem for annular
′
C (1/6) diagrams we have that |I| > 26 |∂R|. Hence, I cannot be written as
the product of fewer than three (r ∪ s)∗ -pieces. Therefore, the path I straddles
three paths pi , pi+1 (mod n) and pi+2 (mod n) , a contradiction by Lemma 6.2.
Next, suppose that R ∩ ∂E A is non-empty. In this case every region of A
has an edge on ∂I S and an edge on ∂E S. Additionally, suppose that R does
not share an edge with any other region of A. Then we have that |I| = 21 |∂R|,
and hence I cannot be written as the product of fewer than four (r∪s)∗ -pieces.
Therefore, the path I straddles four paths pi , . . . , pi+3 (mod n) , a contradiction
by Lemma 6.2. On the other hand, suppose that R shares an edge with some
region S of A. Now, as with the path I, the path J := S ∩ ∂I A contains
an edge. We may assume that the terminal vertex of I is the initial vertex
of J, so τ (I) = ι(J). The minimal possible length of I corresponds to when
|∂E (R)| = 12 |∂R| and R connects over a piece to an additional region T of A.
Hence, |I| > (1 − 61 − 16 − 12 )|∂R| = 16 |∂R|. Hence, the attaching path I cannot
be written as the product of fewer than two pieces. Similarly, |J| > 16 |∂S| so
the attaching path J cannot be written as the product of fewer than two pieces.
We therefore have that I straddles two paths pi and pi+1 (mod n) , with ι(I) ∈ pi
and τ (I) ∈ pi+1 (mod n) , and that J straddles two paths pj and pj+1 (mod n) ,
with ι(J) ∈ pj and τ (J) ∈ pj+1 (mod n) . As τ (I) = ι(J) and as the path
pi+1 (mod n) does not consist of a single piece we have that pi+1 (mod n) = pj .
However, this means that pj decomposes into only two pieces, a contradiction
by Lemma 6.2.
HNN-EXTENSIONS OF TRIANGLE GROUPS
23
We now prove the analogous result for C ′ (1/4) − T (4) presentations. Note
that if hx; r, si is a C ′ (1/4)−T (4) presentation then the sub-presentation hx; ri
is also a C ′ (1/4) − T (4) presentation.
Theorem 6.5. Suppose that hx; r, si is a C ′ (1/4) − T (4) small cancellation
presentation. Let G be the group given by the sub-presentation hx; ri, and let
ϕ : F (x) → hx; ri be the natural map.
Suppose that t ⊂ F (x) is a set of words such that every word Vt (t) is
cyclically r∗ -reduced. For all U ∈ hsi and V ∈ hti, if there exists g ∈ G such
that g −1ϕ(U)g =G ϕ(V ) then there exists W ∈ F (x) such that ϕ(W ) = g and
W −1 UW =F (x) V . In particular, there exists g ∈ G such that hsigG ∩ htiG 6=G 1
if and only if there exists W ∈ F (x) such that hsiW ∩ hti =
6 F (x) 1.
Proof. First note that every word Us (s) is cyclically r∗ -reduced by Lemma 6.3.
Therefore, suppose that A is an annular r-diagram such that its interior boundary ∂I A has a label which is a cyclic shift of some word U := U(s) = S1ǫ1 · · · Snǫn ,
Si ∈ s, and its exterior boundary ∂E A has a label which is a cyclic shift of
some word in hti. It is sufficient to prove that A contains no regions.
Consider the disk diagram DU from Lemma 6.1. Each of the paths pi on the
boundary of DU contains at least three (r ∪ s)∗ -pieces. Note that we can form
an (r ∪ s)∗ -disk diagram ∆ by attaching the boundary ∂DU to the boundary
∂A of A. Suppose that A contains at least one region; we find a contradiction.
As A contains a region, the boundary of DU connects to a region R of A
over an attaching path I such that I contains an edge.
First, suppose that R∩∂E A is empty. By the structural theorem for annular
′
C (1/4)−T (4) diagrams we have that |I| > 14 |∂R|. Hence, I cannot be written
as the product of fewer than two (r∪s)∗ -pieces. Therefore, the path I straddles
two paths pi , pi+1 (mod n) . Then the vertex v = pi ∩ pi+1 (mod n) is an internal
vertex of a C ′ (1/4)−T (4) disc diagram which has degree three, a contradiction.
Next, suppose that R ∩ ∂E A is non-empty. In this case every region of A
has an edge on ∂I S and an edge on ∂E S. Additionally, suppose that R does
not share an edge with any other region of A. Then we have that |I| = 12 |∂R|,
and hence I cannot be written as the product of fewer than three (r ∪ s)∗ pieces. Therefore, the path I straddles three paths pi , pi+1 (mod n) , pi+2 (mod n) ,
a contradiction by Lemma 6.2. On the other hand, suppose that R shares an
edge with some region S of A. Now, as with the path I, the path J := S ∩∂I A
contains an edge. We may assume that the terminal vertex of I is the initial
vertex of J, so τ (I) = ι(J), and note that this vertex is an internal vertex of
the disc diagram. Therefore, the vertex τ (I) = ι(J) has degree at least 4, by
the T (4) condition. Hence, there is a path pi of DU such that either pi is a
subpath of I or I is a subpath of pi . If pi is a subpath of I then pi consists
of a single piece, a contradiction by Lemma 6.2. If I is a subpath of pi then
24
ALAN D. LOGAN
|I| < 41 |∂R|. Then |∂E A ∩ R| ≤ 12 |∂R|, and |R ∩ S| < 14 |∂R|, hence there
exists a second region T of A which connects over a piece to S, and where
K := T ∩ DU is a non-empty path. We may assume that τ (K) ∈ I, and so
τ (K) = ι(I). The vertex τ (K) = ι(I) is an internal vertex of the disk diagram
∆, and hence has degree four by the T (4) condition. Thus, I = pi , but then
pi consists of a single piece, a contradiction.
Proof of Theorem F. We now prove Theorem F.
Proof of Theorem F. As the set s contains no proper powers and satisfies the
C(5) small cancellation condition, the set s also satisfies Wise’s c(5) small
cancellation condition [Wis01, Definition 2.3]. Hence, hsi is a malnormal in
F (x) with minimal basis s [Wis01, Theorems 2.11 & 2.14]. The result then
follows by taking t := s in Theorem 6.4 or in Theorem 6.5, as appropriate.
C(m) − T (4) does not work. The C ′ (1/6) and C ′ (1/4) − T (4) conditions
are needed in Theorems 6.4 and 6.5 because they are metric conditions, in the
sense that pieces have bounded length. The following example demonstrates
that the conclusions of these two theorems do not hold in general if the set
r ∪ s satisfies the the non-metric condition C(m) − T (4) for any m, even if r
satisfies C ′ (λ) for arbitrarily small λ: Let x = {a, b, x1 , y1 , . . . , x2p , y2p }, and
take
R = [x1 , y1 ] · · · [x2p , y2p ], and
S = [x1 , y1 ] · · · [xp , yp ]abab2 · · · abq .
Then hx; Ri has the C ′ (1/(2p − 1)) small cancellation condition while hx; R, Si
has the C(m) − T (4) condition, where m is arbitrarily large and dependent
on p and q (note that hx; r, si has T (4) as R and S are positive words). Note
that R has a piece [x1 , y1] · · · [xp , yp ] which has length 21 |R|. Then the words
S and T = abab2 · · · abq ([xp+1 , yp+1] · · · [x2p , y2p ])−1 are both r∗ -reduced, as are
all their powers, and neither of S not T is a proper power of any element of
x. Moreover, S and T are conjugate in hx; ri but are not freely conjugate.
Hence, for all W (x) ∈ F (s) we have that hsi ∩ htiW (x) 6=F (X) 1, but there
exists g ∈ G such that hsiG ∩ htigG 6=G 1.
Wise’s graphical small cancellation theory. In the classical small cancellation theory, which underlines Theorems 6.4 and 6.5, we ensure that certain
conditions hold for all cyclic shifts of every R ∈ (r ∪ s)±1 (the symmetrised
closure (r ∪ s)∗ ). This is more powerful than is necessary for Theorems 6.4
and 6.5. In particular, we do not need all cyclic shifts of the elements of s±1
HNN-EXTENSIONS OF TRIANGLE GROUPS
25
because the diagram DU , given by Lemma 6.1, is constructed only by elements
of s±1 .
We therefore introduce the following conditions, which sit between the classical small cancellation conditions and Wise’s graphical small cancellation conditions [Wis01] (see also Gromov [Gro03]). We state below two theorems which
use these new conditions to generalise Theorems 6.4 and 6.5. We do not prove
these analogous theorems as the proofs are essentially identical to the proofs
of Theorems 6.4 and 6.5. This is because Wise’s small cancellation conditions
involve the use of a “distinguished vertex” for the boundary of each region R,
but these new conditions allow for this vertex to be placed arbitrarily on the
boundary of R. Therefore, in the proofs of Theorems 6.4 and 6.5 we can first
“place” the distinguished vertex of the region R where we want it to be in ∂R
(for example, we can chose that the distinguished vertex will be an internal
vertex of the annular diagram A of degree at least three, or will be the initial
vertex ι(I) of the attaching path I).
The new conditions are as follows; they can be roughly though of as Wise’s
graphical small cancellation conditions involving the symmetrised closure of a
subset of the relators. Let r = {R1 , . . . , Rn }. We say that the presentation
hx; r, si has the graphical c′r∗ (λ) condition if every set {R1W1 , . . . , RnWn } ∪ s,
where each RiWi is cyclically reduced, has the graphical c′ (λ) condition. The
cr∗ (m) and tr∗ (n) conditions are defined analogously.
Theorem 6.6. Suppose that G is given by the C ′ (1/6) small cancellation presentation hx; ri, and let ϕ : F (x) → hx; ri be the natural map. Suppose that
hx; r, si is a graphical c′r∗ (1/6) small cancellation presentation.
Suppose that t ⊂ F (x) is a set of words such that every word Vt (t) is
cyclically r∗ -reduced. For all U ∈ hsi and V ∈ hti, if there exists g ∈ G such
that g −1ϕ(U)g =G ϕ(V ) then there exists W ∈ F (x) such that ϕ(W ) = g and
W −1 UW =F (x) V . In particular, there exists g ∈ G such that hsigG ∩ htiG 6=G 1
if and only if there exists W ∈ F (x) such that hsiW ∩ hti =
6 F (x) 1.
Theorem 6.7. Suppose that G is given by the C ′ (1/4) − T (4) small cancellation presentation hx; ri, and let ϕ : F (x) → hx; ri be the natural map. Suppose
that hx; r, si is a graphical c′r∗ (1/4) − tr∗ (4) small cancellation presentation.
Suppose that t ⊂ F (x) is a set of words such that every word Vt (t) is
cyclically r∗ -reduced. For all U ∈ hsi and V ∈ hti, if there exists g ∈ G such
that g −1ϕ(U)g =G ϕ(V ) then there exists W ∈ F (x) such that ϕ(W ) = g and
W −1 UW =F (x) V . In particular, there exists g ∈ G such that hsigG ∩ htiG 6=G 1
if and only if there exists W ∈ F (x) such that hsiW ∩ hti =
6 F (x) 1.
26
ALAN D. LOGAN
7. Malcharacteristic free subgroups of triangle groups
In this section we obtain examples of malcharacteristic subgroups of triangle
groups. In particular, Lemma 7.7 proves that the subgroup M of Ti,j,k , i, j, k ≥
6, stated in Section 3 is malcharacteristic and free of rank two. At the end of
the section we apply this to prove Theorem A and Theorem 7.9.
M is free and malnormal. We first prove that the stated subgroup M of
Ti,j,k is free of rank two and malnormal.
Lemma 7.1. The subgroup M = hx, yi of Ti,j,k = ha, b; ai , bj , (ab)k i, i, j, k ≥ 6,
is a malnormal subgroup of Ti,j,k and is free of rank two.
Proof. The set {ai , bj , (ab)k , x, y} has the C ′ (1/4) − T (4) small cancellation
condition. The result then follows from Theorem F.
The automorphisms of Ti,j,k . After Lemma 7.1, in order to prove that M
is malcharacteristic in Ti,j,k it is sufficient to prove that if δ ∈ Aut(Ti,j,k ) such
that δ(M) ∩ M 6= 1 then δ ∈ Inn(Ti,j,k ). We therefore need to understand
the outer automorphism groups of triangle groups; we do this in Lemmas 7.2
and 7.3. The proof of Lemma 7.2 follows closely certain proofs in a paper of
Zieschang [Zie76] (but Lemma 7.2 does not follow from Zieschang’s paper).
Lemma 7.2. Suppose that Ti,j,k , with i, j, k ≥ 6, and Tt1 ,t2 ,t3 define the same
presentation, up to a permutation of the generators. Write G = ha, b; ai , bj , (ab)k i
and H = hx1 , x2 , x3 ; xt11 , xt22 , xt33 , x1 x2 x3 i. If f : G → H is an isomorphism then
there exists an inner automorphism γ ∈ Inn(H), integers p, q ∈ {1, 2, 3} with
p 6= q, and ǫ = ±1 such that f γ(a) = xǫp and f γ(b) = xǫq .
ξ
ξ
Proof. We may assume f (a) = U1 xpp U1−1 and f (b) = U2 xqq U2−1 , with p, q ∈
{1, 2, 3} not necessarily distinct, and with |ξp | ≤ 12 tp and |ξq | ≤ 12 tq [ZVC70,
Satz IV.12] (see also [FR99, Theorem 4.3.2]). Therefore, there exists an inner
ξ
ξ
automorphism γ ∈ Inn(H) such that f γ(a) = xpp and f γ(b) = Uxqq U −1 ,
where U does not begin with x±1
and does not end with x±1
p
q . We view H
′
tp
tq
using the C (1/4) − T (4) presentation PH = hxp , xq ; x , x , (xp xq )tr i. Let
s = {xtp , xtq , (xp xq )tr }. Assume that the word U is Dehn reduced (that is, U
is s∗ -reduced).
ξ
ξ
Suppose that U is not the empty word. Now, the word xpp Uxqq U −1 has finite
ξ
ξ
order in H, and so there exist some n > 1 such that (xpp Uxqq U −1 )n contains a
subword W of some R ∈ s∗ with |W | > 32 |R| [LS77, Theorem V.4.4.ii]. Then
ξ
ξ
W must contain letters from U or U −1 and at least one letter from xpp or xqq .
−ǫ ξ ǫ
Therefore, R is conjugate to (xp xq )±tr . Since tq > 2, the word xq q xpp xqq is
HNN-EXTENSIONS OF TRIANGLE GROUPS
27
ξ
not contained in W , and hence if W contains a part of xpp then W contains a
ξ
single letter from xpp , and this is either at the start or the end of W . Similarly,
−ǫ ξ ǫ
since tp > 2 the word xp p xqq xpp is not contained in W and so the analogous
ξ
statement holds for W containing a part of xqq .
ξs1
We therefore have two cases: either W contains x±1
s1 from xs1 and the other
ξs1
letters are from one of U or U −1 , or W begins with x±1
s1 from xs1 , ends with
ξs2
−1
.
x±1
s2 from xs2 , s1 6= s2 , and all other letters are from precisely one of U or U
′
tr
In the first case, U contains a part W from the defining relation (xp xq ) with
|W ′ | = |W | − 1 > 32 · 2tr − 1 > tr = 21 |R| (as tr > 3). Hence, U is not Dehn
reduced, a contradiction. In the second case, W is again part of the relation
(xp xq )±tr , therefore |U| = |W | − 2 > 23 · 2tr − 2 ≥ tr (as tr ≥ 6), and so again
U is not Dehn reduced, a contradiction. Therefore, U is the empty word, and
ξ
ξ
we have f γ(a) = xpp and f γ(b) = xqq .
We now prove that there exists some ǫ = ±1 such that ξp = ǫ = ξq , which
proves the result. Suppose otherwise. Then the longest possible subword of a
ξ ξ
ǫ ǫ ǫ
ǫ ǫ ǫ
relator contained in the word (xpp xqq )tr is either xpp xqq xpp or xqq xpp xqq , so the
ξ ξ
word (xpp xqq )tr is Dehn reduced. As this word is non-empty and Dehn reduced
it is non-trivial, a contradiction.
Write Ψ for the set consisting of the following twelve maps. Note that each
map defines an automorphism of F (a, b).
ψ(1,ǫ) : a 7→ aǫ
b 7→ bǫ
ψ(4,ǫ) : a 7→ bǫ
b 7→ aǫ
ψ(2,ǫ) : a 7→ aǫ
b 7→ (ab)−ǫ
ψ(5,ǫ) : a 7→ bǫ
b 7→ (ab)−ǫ
ψ(3,ǫ) : a 7→ (ab)ǫ
b 7→ b−ǫ
ψ(6,ǫ) : a 7→ (ab)ǫ
b 7→ a−ǫ
Lemma 7.3. Let i, j, k ≥ 6. Then:
(1) If i, j, k are pairwise non-equal then Ψi,j,k := {ψ(1,1) , ψ(1,−1) } is a transversal for Out(Ti,j,k ).
(2) If i = j 6= k then Ψi,i,k := {ψ(1,1) , ψ(1,−1) , ψ(4,1) , ψ(4,−1) } is a transversal
for Out(Ti,j,k ).
(3) If i = j = k then Ψ is a transversal for Out(Ti,i,i ).
Proof. By Lemma 7.2, if δ ∈ Aut(Ti,j,k ) then there exists some γ ∈ Inn(Ti,j,k )
such that δγ ∈ Ψ. The result follows by checking which elements of Ψ define
automorphisms of Ti,j,k for each of the three cases.
The images of x and y under automorphisms. Let s := {x, y} ⊂ F (a, b).
28
ALAN D. LOGAN
Our final step before proving that M is malcharacteristic in Ti,j,k is to prove
that for each ψ ∈ Ψ, every word Wψ(s) (ψ(s)) is cyclically Dehn reduced in
Ti,j,k . This allows us to apply Theorem 6.5 to ψ(s). Lemmas 7.4–7.6 now
analyse these words Wψ(s) (ψ(s)).
Every map ψ ∈ Ψ extends naturally to the whole group Ti,j,k , even if
ψ 6∈ Aut(Ti,j,k ). That there are maps ψ ∈ Ψ which are not necessarily automorphisms of the groups involved does not affect the proofs of Lemmas 7.4–
7.6.
Lemma 7.4. Let Ti,j,k = ha, b; ai , bj , (ab)k i, with i, j, k ≥ 6. Suppose ψ is
a map contained in the set Ψ, and we shall write A := ψ(a)ψ(b)−1 , B :=
ψ(a)2 ψ(b)−1 . Then for all p, q ≥ 0 the word B p A3 B q is freely reduced over
F (a, b) and does not contain a±4 , b±4 , ((ab)3 a)±1 or (b(ab)3 )±1 , and so is
Dehn reduced.
Proof. The proof is by inspection of the appropriate words. Indeed the following words are all freely reduced over F (a, b) for p, q ≥ 3 with the longest
subwords of relators occurring being a3 ≤ ai , b ≤ bj and babab ≤ (ab)k . Each
word represents B p A3 B q for the indicated ψ ∈ Ψ.
ψ(1,1) :
(a2 b−1 )p (ab−1 )3 (a2 b−1 )q
ψ(2,1) :
(a3 b)p (a2 b)3 (a3 b)q
ψ(3,1) :
(abab2 )p (ab2 )3 (abab2 )q
Now, if the result holds for ψ(l,1) , l ∈ {1, 2, 3} then it also holds for ψ(l+3,1)
(as ψ(l,1) and ψ(l+3,1) differ just by a switch of a and b, and possibly by an
inner automorphism). Further, if the result holds for ψ(l,1) then it also holds
for ψ(l,−1) . Hence, the proof is complete.
Lemma 7.5. Let Ti,j,k = ha, b; ai , bj , (ab)k i, with i, j, k ≥ 6. Suppose ψ is
a map contained in the set Ψ, and we shall write A := ψ(a)ψ(b)−1 , B :=
ψ(a)2 ψ(b)−1 . Then for all p, q ≥ 3, ǫ0 = ±1, after free reduction over F (a, b)
the word B ǫ0 p A3 B −ǫ0 q has the form B ǫ0 (p−1) CB −ǫ0 (q−1) for some word C ∈
F (a, b). Moreover, the resulting word B ǫ0 (p−1) CB −ǫ0 (q−1) does not contain a±4 ,
b±4 , (ab)±3 or (ba)±3 and so is Dehn reduced.
Proof. As with Lemma 7.4, the proof is by inspection of the appropriate words.
Indeed, after free reduction over F (a, b) the following words are all Dehn reduced for p, q ≥ 3, ǫ0 = ±1, with the longest subwords of relators occurring
being a±3 ≤ a±i , b±2 ≤ b±j and (babab)±1 ≤ (ab)±k . In each line the left-hand
word represents B ǫ0 p A3 B −ǫ0 q for the indicated map ψ ∈ Ψ and the right-hand
word represents the freely reduced form of the word. Each map has two cases:
HNN-EXTENSIONS OF TRIANGLE GROUPS
29
ǫ0 = 1 and ǫ0 = −1.
ψ(1,1) : (a2 b−1 )p (ab−1 )3 (a2 b−1 )−q = (a2 b−1 )p (ab−1 )2 a−1 (ba−2 )q−1
(a2 b−1 )−p (ab−1 )3 (a2 b−1 )q = (ba−2 )p−1 ba−1 b−1 (ab−1 )2 (a2 b−1 )q
ψ(2,1) :
(a3 b)p (a2 b)3 (a3 b)−q = (a3 b)p (a2 b)2 a−1 (b−1 a−3 )q−1
(a3 b)−p (a2 b)3 (a3 b)q = (b−1 a−3 )p−1 b−1 a−1 b(a2 b)2 (a3 b)q
ψ(3,1) :
(abab2 )p (ab2 )3 (abab2 )−q = (abab2 )p ab2 aba−1 (b−2 a−1 b−1 a−1 )q−1
(abab2 )−p (ab2 )3 (abab2 )q = (b−2 a−1 b−1 a−1 )p−1b−2 a−1 b(ab2 )2 (abab2 )q
Now, if the result holds for ψ(l,1) , l ∈ {1, 2, 3} then it also holds for ψ(l+3,1)
(as ψ(l,1) and ψ(l+3,1) differ just by a switch of a and b, and possibly by an
inner automorphism). Further, if the result holds for ψ(l,1) then it also holds
for ψ(l,−1) . Hence, the proof is complete.
We now combine Lemmas 7.4 and 7.5 as follows. Note that the word
W{a,b} (ψ(s)) corresponds to the image of the word Ws (s) under the map ψ
after applying free reduction in F (a, b).
Lemma 7.6. Let Ti,j,k = ha, b; ai , bj , (ab)k i, with i, j, k ≥ 6, and let s := {x, y}.
Suppose ψ is a map contained in the set Ψ and consider Ws (s). Then the word
W{a,b} (ψ(s)) is Dehn reduced.
Proof. Write A := ψ(a)ψ(b)−1 and B := ψ(a)2 ψ(b)−1 . Write Xp0 ,q0 := B p0 A3 B q0 ,
Yp1 ,q1 := B p1 A3 B −q1 and Zp2 ,q2 := B −p2 A3 B q2 , where p0 , q0 ≥ 0 and where
p1 , p2 , q1 , q2 ≥ 3, and write
s0 := {Xp0 ,q0 | p0 , q0 ≥ 0} ∪ {Yp1 ,q1 , Zp2 ,q2 | pi , qi ≥ 3}.
Now, Wψ(s) (ψ(s)) is a product of the words (A3 B 3 A3 B 4 · · · A3 B ρ+2 )±1 and
(A3 B ρ+3 A3 · · · A3 B 2ρ+2 )±1 . Therefore, there exists a word Us0 (s0 ) such that
Us0 (s0 ) ≡ Wψ(s) (ψ(s)), and where every subword of length two of Us0 (s0 ) has
the form either (Xp0 ,q0 Yp1 ,q1 )±1 , (Yp1,q1 Xp−1
)±1 , (Xp−1
Z
)±1 , (Zp2 ,q2 Xp0 ,q0 )±1 ,
0 ,q0
0 ,q0 p2 ,q2
or (Xp0 ,q0 Xp′0 ,q0′ )±1 . Note that U{a,b} (s0 ) ≡ W{a,b} (ψ(s)).
By Lemma 7.4, the words Xp0 ,q0 are freely reduced over F (a, b) and are
Dehn reduced. By Lemma 7.5, after free reduction in F (a, b) all words Yp1 ,q1
and Zp2 ,q2 are Dehn reduced and of the form respectively B p1 −1 C1 B −(q1 −1)
and B −(p2 −1) C2 B q2 −1 for some fixed words C1 , C2 ∈ F (a, b). Then, applying
Lemma 7.4 again, the following words are freely reduced over F (a, b) and Dehn
reduced:
)±1 ,
(Xp0 ,q0 B p1 −1 C1 B −(q1 −1) )±1 , (B p1−1 C1 B −(q1 −1) Xp−1
0 ,q0
B −(p2 −1) C2 B q2 −1 )±1 , (B −(p2 −1) C2 B q2 −1 Xp0 ,q0 )±1 , (Xp0 ,q0 Xp′0 ,q0′ )±1
(Xp−1
0 ,q0
30
ALAN D. LOGAN
These words represent the length-two subwords of Us0 (s0 ) after free reduction
in F (a, b), and hence the word U{a,b} (s0 ) is Dehn reduced. As U{a,b} (s0 ) ≡
W{a,b} (ψ(s)), the words W{a,b} (ψ(s)) is Dehn reduced as required.
Examples of malcharacteristic subgroups. We are finally ready to prove
that the subgroup M of Ti,j,k , i, j, k ≥ 6, is malcharacteristic.
Lemma 7.7. Let M be the subgroup of Ti,j,k = ha, b; ai , bj , (ab)k i, i, j, k ≥ 6,
which is generated by the following elements, with ρ ≫ i, j, k.
x := (ab−1 )3 (a2 b−1 )3 (ab−1 )3 (a2 b−1 )4 . . . (ab−1 )3 (a2 b−1 )ρ+2
y := (ab−1 )3 (a2 b−1 )ρ+3 (ab−1 )3 (a2 b−1 )ρ+4 . . . (ab−1 )3 (a2 b−1 )2ρ+2
Then M is a malcharacteristic subgroup of Ti,j,k and is free of rank two.
f{x,y} = hx, yi for the subgroup of F (a, b) generated by the
Proof. Write M
f{x,y} is in the same Aut(F (a, b))-orbit as the subgroup
words x and y. Then M
f defined in Lemma 5.3. Therefore, by Lemma 5.3, the subgroup M
f{x,y} is
M
malcharacteristic in F (a, b). We use this to prove that M is malcharacteristic
in Ti,j,k .
Now, by Lemma 7.1, it is sufficient to prove that if δ ∈ Aut(Ti,j,k ) is such
that δ(M) ∩ M 6= 1 then δ ∈ Inn(Ti,j,k ). So, suppose that there exists an
automorphism δ ∈ Aut(Ti,j,k ) and two words U and V such that U(x, y) =
V (δ(x), δ(y)), and we wish to prove that δ is inner. Now, by Lemma 7.3
there exist ψ ∈ Ψ and γg ∈ Inn(Ti,j,k ), g ∈ Ti,j,k , such that δ = ψγg , so
ψγg (M) ∩ M 6= 1. Now, after free reduction over F (a, b) every word in ψ(M)
is Dehn reduced, by Lemma 7.6. Hence, we may apply Theorem 6.5. Therefore,
e W (M
f{x,y} )∩M
f{x,y} 6=F (a,b) 1, where
there exists a word W ∈ F (a, b) such that ψγ
ψe is the automorphism of F (a, b) defined using the same words as ψ. Then ψe 6∈
f{x,y} is not malcharacteristic in F (a, b), a contradiction.
Inn(F (a, b)), so M
If any of i, j or k in Ti,j,k are less than 6 then it is not clear that the
subgroup M is malcharacteristic. The precise issues are found in Lemma 7.2
and in Lemmas 7.4 and 7.5, whose proofs each require i, j, k ≥ 6. Because of
the small cancellation arguments we apply, it is likely that Lemma 7.7, and
hence Theorem A and Theorem 7.9, can be extended to i, j, k > 6 or even to
i, j, k > 4. To develop our results in this direction, first Lemma 7.2 would need
to be extended appropriately. To extend Lemmas 7.4 and 7.5 it is likely that
new words x, y ∈ F (a, b) would need to be found.
Lemmas 3.3 and 7.7 combine to prove the following result.
Proposition 7.8. A triangle group Ti,j,k , i, j, k ≥ 6, contains malcharacteristic subgroups Mn of arbitrary rank n (possibly countably infinite).
HNN-EXTENSIONS OF TRIANGLE GROUPS
31
Indeed, the subgroup Mn is generated by n finite subwords satisfying C ′ (1/6)
of the following infinite word:
xy(xy 2 )xy(xy 2)2 xy(xy 2)3 · · · .
Proofs of Theorems A, B, C and E. We first prove Theorem A.
Proof of Theorem A. By Lemma 7.7, the group Ti contains a malcharacteristic
subgroup which is free of rank two. As Ti has Serre’s property FA [Ser80,
Example 6.3.5] and as φ : a 7→ b, b 7→ b−1 a−1 defines a non-inner automorphism
of Ti which has order three, the result follows from Theorem 4.2.
We now prove Theorem B.
Proof of Theorem B. Theorem B follows immediately from Theorem 4.3.
We now prove Theorem C.
Proof of Theorem C. Theorem C follows immediately from Theorem 4.4.
We now prove Theorem E.
Proof of Theorem E. Theorem E follows immediately from Theorem 4.5.
The triangle groups Ti,j,k . Theorems 7.9–7.12 prove analogous results to
Theorems A–E where the base group H is a fixed triangle group Ti,j,k rather
than an equilateral triangle group Ti .
Theorem 7.9. Fix a (non-equilateral) triangle group Ti,j,k with i, j, k ≥ 6.
Every countable group Q can be realised as an index-one or-two subgroup of
the outer automorphism group of an automorphism-induced HNN-extension
TPi,j,k of Ti,j,k . Moreover, Aut(TPi,j,k ) has an index-one or-two subgroup which
splits as TPi,j,k ⋊ Q.
Proof. By Lemma 7.7, the group Ti,j,k contains a malcharacteristic subgroup
which is free of rank two. As Ti,j,k has Serre’s property FA [Ser80, Example
6.3.5] and as φ : a 7→ a−1 , b 7→ b−1 defines a non-inner automorphism of Ti,j,k ,
the result follows from Theorem 4.2.
The proofs of Theorems 7.10–7.12 are identical to the proofs of Theorems
B, C and E, and hence are omitted.
Theorem 7.10. The map defined by P 7→ TPi,j,k is a functor from the category
of countable group presentations Pres to the category of groups Grp.
i,j,k
Theorem 7.11. Let Ti,j,k
where π1 (P) is finite.
Fin be the class of groups TP
If π1 (P) is residually finite then TPi,j,k is residually-Ti,j,k
Fin .
32
ALAN D. LOGAN
Theorem 7.12. The subgroups Mn , n ∈ N, in the proof of Theorem 7.9 may
be chosen in such a way that for every presentation P with finite, non-empty
generating set and for every countable group presentation Q there exists a
surjection TP ։ TP∗Q .
8. Residual finiteness
Bumagin–Wise asked if every countable group Q can be realised as the outer
automorphism group of a finitely generated, residually finite group GQ [BW05,
Problem 1]. We now prove Corollary D (which is a corollary of Theorem C).
Corollary D answers this question of Bumagin–Wise for all finitely generated,
residually finite groups Q by taking GQ := TP for P a presentation of Q with
finite, non-empty generating set.
Proof of Corollary D. By Theorem C it is sufficient to prove that if P = hx; ri
has finite, non-empty generating set x and π1 (P) is finite then TP is residually
finite. Under these conditions the associated subgroup KP of the group TP =
H∗(KP ,φ) has finite index in the subgroup M|x| . As M|x| is a free group of finite
rank |x|, the subgroup KP is finitely generated.
Now, triangle groups are LERF [Sco78]. Hence, KP is separable in the
triangle group H. Therefore, the automorphism-induced HNN-extension TP
is residually finite [BT78, Lemma 4.4], as required.
In order to extend Corollary D to all countable residually finite groups it
would be necessary to prove that if P = hx; ri is a presentation of a finite
group where |x| = ∞ then the group TP is residually finite. (In Corollary D
we explicitly use the fact that x is a finite set.) We therefore have the following
corollary of Theorem C.
Corollary 8.1. Let T∞
Fin be the class of groups TP where π1 (P) is finite but
where P has infinite generating set. Suppose that every element of T∞
Fin is
residually finite. Then for every countable, residually finite group Q there
exists a finitely generated, residually finite group GQ such that Out(GQ ) ∼
= Q.
Note that Corollary 8.1 represents the best application of Theorem A in
this direction, in the sense that if TP is residually finite then Q := π1 (P) is
residually finite. This is because Q embeds in Aut(TP ), so if Q is not residually
finite then TP is not residually finite [Bau63].
9. Questions
We now pose certain questions which arose in the writing of this paper.
Injectivity of the functor. The functor of Theorem B should be injective
on both objects and morphisms. Injectivity of the functor is equivalent to a
HNN-EXTENSIONS OF TRIANGLE GROUPS
33
positive answer to the following question, where ∼ is the equivalence relation
on presentations defined in Section 4.
Question 1. Is it true that TP ∼
= TQ if and only if P ∼ Q?
It is unlikely that the functor of Theorem 4.3 is injective for all choices of
groups H and automorphisms φ ∈ Aut(H).
Question 2. For which base groups H and automorphisms φ ∈ Aut(H) does it
hold that in the construction of Theorem 4.2, HP ∼
= HQ if and only if P ∼ Q?
Residual finiteness. Question 3 leads naturally on from Corollary D and
Corollary 8.1.
Question 3. Suppose P = hx; ri is a presentation of a finite group such that
the set x is infinite. Is the group TP from Theorem A residually finite?
Free groups. In Section 5 we proved that F2 , the free group of rank two,
f with M
f∼
contains a malcharacteristic subgroup M
= F2 ; it then follows from
Lemma 3.3 that F2 contains malcharacteristic subgroups of arbitrary rank.
Question 4. Do the free groups Fn , n > 2, each contain a malcharacteristic
subgroup which is free of rank two?
Malnormality is decidable in free groups. In Lemma 5.2 we gave conditions
relating to the malcharacteristic property being decidable in F2 .
Question 5. Is it decidable if a subgroup M of a free group Fn , n ≥ 2, is
malcharacteristic?
Malnormality is a generic property of subgroups of free groups: using the the
“few relator” model for random groups we see that a random finite set of words
s ⊂ Fn contains no proper powers and satisfies the classical small cancellation
conditions [Oll05, Proposition 2]. It then follows that s is malnormal in Fn
with free basis s [Wis01, Theorems 2.11 & 2.14].
Question 6. Does a generic set of words s ⊂ Fn , n ≥ 2, generate, with
overwhelming probability, a malcharacteristic subgroup of Fn ?
Other groups. As was mentioned in the introduction, we expect Theorem A to hold for many other classes of groups and not just equilateral triangle groups. Now, Thompson’s group V has Serre’s property FA [Far11] and
has outer automorphisms of any given order [BCM+ 16]. Therefore, a positive
answer to the following question would allow us to apply Theorem 4.2 to V ,
obtaining a result analogous to Theorem A.
34
ALAN D. LOGAN
Question 7. Does Thompson’s group V contain a malcharacteristic subgroup
which is free of rank two?
Thompson’s group T also has Serre’s property FA [Far11], but Out(T ) is
cyclic of order two [Bri96]. Therefore, a positive answer to the following question would allow us to apply Theorem 4.2 to T , obtaining a result analogous
to Theorem 7.9.
Question 8. Does Thompson’s group T contain a malcharacteristic subgroup
which is free of rank two?
Note that we ask about T and V and not Thompson’s group F because it
is well-known that F does not contain a non-abelian free group.
References
[Bau63]
G. Baumslag, Automorphism groups of residually finite groups, J. London Math.
Soc. 48 (1963), 117–118.
[BCM+ 16] C. Bleak, P. Cameron, Y. Maissel, A. Navas, and F. Olukoya, The further
chameleon groups of Richard Thompson and Graham Higman: Automorphisms
via dynamics for the Higman groups G{n,r} , arXiv:1605.09302 (2016).
[BG03]
G. Braun and R. Göbel, Outer automorphisms of locally finite p-groups, J. Algebra 264 (2003), no. 1, 55–67.
[BM06]
A. V. Borshchev and D. I. Moldavanskii, On the isomorphism of some groups
with one defining relation, Mathematical Notes 79 (2006), no. 1, 31–40.
[Bri96]
M.G. Brin, The chameleon groups of Richard J. Thompson: automorphisms and
dynamics, Inst. Hautes Études Sci. Publ. Math. (1996), no. 84, 5–33 (1997).
MR 1441005
[BT78]
B. Baumslag and M. Tretkoff, Residually finite HNN extensions, Comm. Algebra
6 (1978), no. 2, 179–194. MR 484178
[BW05]
I. Bumagin and D.T. Wise, Every group is an outer automorphism group of a
finitely generated group, J. Pure App. Alg. 200 (2005), no. 1, 137–147.
[CCH81] I.M. Chiswell, D.J. Collins, and J. Huebschmann, Aspherical group presentations, Math. Zeitschrift 178 (1981), no. 1, 1–36.
[CL83]
D.J. Collins and F. Levin, Automorphisms and Hopficity of certain BaumslagSolitar groups, Arch. Math. 40 (1983), no. 1, 385–400.
[CMZ81] M. Cohen, W. Metzler, and A. Zimmermann, What does a basis of F (a, b) look
like?, Math. Ann. 257 (1981), no. 4, 435–445.
[DGG01] M. Droste, M. Giraudet, and R. Göbel, All groups are outer automorphism
groups of simple groups, J. Lon. Math. Soc. 64 (2001), no. 3, 565–575.
[Far11]
D. Farley, A proof that Thompson’s groups have infinitely many relative ends,
J. Group Theory 14 (2011), no. 5, 649–656. MR 2831963
[FM06]
R. Frigerio and B. Martelli, Countable groups are mapping class groups of hyperbolic 3-manifolds, Math. Res. Lett. 13 (2006), no. 5-6, 897–910. MR 2280783
[FR99]
B. Fine and G. Rosenberger, Algebraic generalizations of discrete groups, Monographs and Textbooks in Pure and Applied Mathematics, vol. 223, Marcel
Dekker, Inc., New York, 1999, A path to combinatorial group theory through
one-relator products. MR 1712997
HNN-EXTENSIONS OF TRIANGLE GROUPS
35
[GHMR00] N.D. Gilbert, J. Howie, V. Metaftsis, and E. Raptis, Tree actions of automorphism groups, J. Group Theory 3 (2000), no. 2, 213–223.
[GP00]
R. Göbel and A.T. Paras, Outer automorphism groups of metabelian groups, J.
Pure App. Algebra 149 (2000), no. 3, 251–266.
[Gro03]
M. Gromov, Random walk in random groups, Geom. Funct. Anal. 13 (2003),
no. 1, 73–146. MR 1978492
[Koj88]
S. Kojima, Isometry transformations of hyperbolic 3-manifolds, Topology and
its Applications 29 (1988), no. 3, 297–307.
[Log15]
A.D. Logan, On the outer automorphism groups of finitely generated, residually
finite groups, J. Algebra 423 (2015), 890–901. MR 3283743
, On a question of Bumagin and Wise, New York J. Math. 22 (2016),
[Log16a]
865–873. MR 3548127
[Log16b]
, The Bass-Jiang group for automorphism-induced HNN-extensions,
arXiv:1509.01847 (2016).
[LS77]
R.C. Lyndon and P.E. Schupp, Combinatorial group theory, Classics in Mathematics, Springer, 1977.
[Mag74]
Wilhelm Magnus, Noneuclidean tesselations and their groups, Academic Press
[A subsidiary of Harcourt Brace Jovanovich, Publishers], New York-London,
1974, Pure and Applied Mathematics, Vol. 61. MR 0352287
[Mat89]
T. Matumoto, Any group is represented by an outer automorphism group, Hiroshima Math. J. 19 (1989), no. 1, 209–219.
[Min09]
Ashot Minasyan, Groups with finitely many conjugacy classes and their automorphisms, Comment. Math. Helv. 84 (2009), no. 2, 259–296. MR 2495795
[Oll05]
Yann Ollivier, A January 2005 invitation to random groups, Ensaios
Matemáticos [Mathematical Surveys], vol. 10, Sociedade Brasileira de
Matemática, Rio de Janeiro, 2005. MR 2205306
[Osi10]
D. Osin, Small cancellations over relatively hyperbolic groups and embedding
theorems, Ann. of Math. (2) 172 (2010), no. 1, 1–39. MR 2680416
[Rip82]
E. Rips, Subgroups of small cancellation groups, Bull. London Math. Soc 14
(1982), no. 1, 45–47.
[Sco78]
P. Scott, Subgroups of surface groups are almost geometric, J. London Math.
Soc. (2) 17 (1978), no. 3, 555–565. MR 0494062
[Ser80]
J.P. Serre, Trees, Springer-Verlag, Berlin-New York, 1980, Translated from the
French by John Stillwell. MR 607504
[Sta83]
J.R. Stallings, Topology of finite graphs, Invent. Math. 71 (1983), no. 3, 551–565.
[Wis01]
D.T. Wise, The residual finiteness of positive one-relator groups, Comment.
Math. Helv. 76 (2001), no. 2, 314–338.
[Zie76]
H. Zieschang, On triangle groups, Russian Math. Surveys 31 (1976), no. 5,
226–233.
[ZVC70]
H. Zieschang, E. Vogt, and H.-D. Coldewey, Flächen und ebene diskontinuierliche Gruppen, Lecture Notes in Mathematics, Vol. 122, Springer-Verlag, BerlinNew York, 1970. MR 0262378
| 4math.GR
|
1
Coordinate-Descent Diffusion Learning by
Networked Agents
arXiv:1607.01838v2 [cs.MA] 11 Oct 2017
Chengcheng Wang, Student Member, IEEE, Yonggang Zhang, Senior Member, IEEE,
Bicheng Ying, Student Member, IEEE, and Ali H. Sayed, Fellow, IEEE
Abstract—This work examines the mean-square error performance of diffusion stochastic algorithms under a generalized
coordinate-descent scheme. In this setting, the adaptation step
by each agent is limited to a random subset of the coordinates
of its stochastic gradient vector. The selection of coordinates
varies randomly from iteration to iteration and from agent to
agent across the network. Such schemes are useful in reducing
computational complexity at each iteration in power-intensive
large data applications. They are also useful in modeling situations where some partial gradient information may be missing
at random. Interestingly, the results show that the steady-state
performance of the learning strategy is not always degraded,
while the convergence rate suffers some degradation. The results
provide yet another indication of the resilience and robustness
of adaptive distributed strategies.
Index Terms—Coordinate descent, stochastic partial update,
computational complexity, diffusion strategies, stochastic gradient algorithms, strongly-convex cost.
I. I NTRODUCTION AND R ELATED W ORK
Consider a strongly-connected network of N agents, where
information can flow in either direction between any two
connected agents and, moreover, there is at least one self-loop
in the topology [2, p. 436]. We associate a strongly convex
differentiable risk, Jk (w), with each agent k and assume in
this work that all these costs share a common minimizer,
wo ∈ RM , where R denotes field of real numbers. This case
models important situations where agents work cooperatively
towards the same goal. The objective of the network is to
determine the unique minimizer wo of the following aggregate
cost, assumed to be strongly-convex:
J glob (w) ,
N
X
Jk (w)
(1)
k=1
A short version of this work appears in the conference publication [1].
This work was performed while C. Wang was a visiting student at the
UCLA Adaptive Systems Laboratory. The work of C. Wang was supported
in part by a Chinese Government Scholarship. The work of Y. Zhang was
supported in part by the National Natural Science Foundation of China
(61371173), and Fundamental Research Founds for the Central University
of Harbin Engineering University (HEUCFP201705). The work of B. Ying
and A. H. Sayed was supported in part by NSF grants CCF-1524250 and
ECCS-1407712.
C. Wang and Y. Zhang are with the College of Automation, Harbin
Engineering University, Harbin, Heilongjiang 150001 China. C. Wang is
also with the School of Electrical and Electronic Engineering, Nanyang
Technological University, 639798 Singapore (e-mail: [email protected];
[email protected]).
B. Ying is with the Department of Electrical Engineering, University of
California, Los Angeles, CA 90024 USA (e-mail: [email protected]).
A. H. Sayed is with the Ecole Polytechnique Federale de Lausanne, EPFL,
School of Engineering, CH-1015 Lausanne, Switzerland (e-mail: ali.sayed@
epfl.ch).
It is also assumed that the individual cost functions, Jk (w),
are each twice-differentiable and satisfy
0 < νd IM ≤ ∇2w Jk (w) ≤ δd IM
(2)
where ∇2w Jk (w) denotes the M ×M Hessian matrix of Jk (w)
with respect to w, νd ≤ δd are positive parameters, and IM
is the M × M identity matrix. In addition, for matrices A
and B, the notation A ≤ B denotes that B − A is positive
semi-definite. The condition in (2) is automatically satisfied
by important cases of interest, such as logistic regression or
mean-square-error designs [2], [3].
Starting from some initial conditions {wk,−1 }, the agents
work cooperatively in an adaptive manner to seek the minimizer wo of problem (1) by applying the following diffusion
strategy [2], [3]:
X
φk,i−1 =
a1,`k w`,i−1
(3a)
`∈N
k
\
ψk,i = φk,i−1 − µk ∇
(3b)
wT J k (φk,i−1 )
X
wk,i =
a2,`k ψ`,i
(3c)
`∈Nk
where the M -vector wk,i denotes the estimate by agent k
at iteration i for wo , while ψk,i and φk,i−1 are intermediate
estimates. Moreover, an approximation for the true gradient
\
vector of Jk (w), ∇
wT J k (·), is used in (3b) since it is generally
the case that the true gradient vector is not available (e.g.,
when Jk (w) is defined as the expectation of some loss function
and the probability distribution of the data is not known
beforehand to enable computation of Jk (·) or its gradient
vector). The symbol Nk in (3) refers to the neighborhood of
agent k. The N × N combination matrices A1 = [a1,`k ] and
A2 = [a2,`k ] are left-stochastic matrices consisting of convex
combination coefficients that satisfy:
aj,`k ≥ 0,
N
X
aj,`k = 1, aj,`k = 0, if ` ∈
/ Nk
(4)
`=1
for j = 1, 2. Either of these two matrices can be chosen as the
identity matrix, in which case algorithm (3) reduces to one of
two common forms for diffusion adaptation: the adapt-thencombine (ATC) form when A1 = I and the combine-thenadapt (CTA) form when A2 = I. We continue to work with the
general formulation (3) in order to treat both algorithms, and
other cases as well, in a unified manner. The parameter µk > 0
is a constant step-size factor used to drive the adaptation
process. Its value is set to a constant in order to enable
2
continuous adaptation in response to streaming data or drifting
minimizers. We could also consider a distributed implementation of the useful consensus-type [2], [4]–[11]. However, it has
been shown in [2], [12] that when constant step-sizes are used
to drive adaptation, the diffusion networks have wider stability
ranges and superior performance. This is because consensus
implementations have an inherent asymmetry in their updates,
which can cause network graphs to behave in an unstable
manner when the step-size is constant. This problem does
not occur over diffusion networks. Since adaptation is a core
element of the proposed strategies in this work, we therefore
focus on diffusion learning mechanisms.
The main distinction in this work relative to prior studies
on diffusion or consensus adaptive networks is that we now
assume that, at each iteration i, the adaptation step in (3b)
has only access to a random subset of the entries of the
approximate gradient vector. This situation may arise due to
missing data or a purposeful desire to reduce the computational
burden of the update step. We model this scenario by replacing
the approximate gradient vector by
miss
\
∇
wT J k
\
(φk,i−1 ) = Γk,i · ∇
wT J k (φk,i−1 )
(5)
where the random matrix Γk,i is diagonal and consists of
Bernoulli random variables {rk,i (m)}; each of these variables
is either zero or one with probability
Prob(rk,i (m) = 0) , rk
(6)
where 0 ≤ rk < 1 and
Γk,i = diag{rk,i (1), rk,i (2), . . . , rk,i (M )}
(7)
In the case when rk,i (m) = 0, the m-th entry of the gradient
vector is missing, and then the m-th entry of ψk,i in (3b) is
not updated. Observe that we are attaching two subscripts to
r: k and i, which means that we are allowing the randomness
in the update to vary across agents and also over time.
A. Relation to Block-Coordinate Descent Methods
Our formulation provides a nontrivial generalization of the
powerful random coordinate-descent technique studied, for
example, in the context of deterministic optimization in [13]–
[15] and the references therein. Random coordinate-descent
has been primarily applied in the literature to single-agent
convex optimization, namely, to problems of the form:
wo = arg min J(w)
w
(8)
where J(w) is assumed to be known beforehand. The traditional gradient descent algorithm for seeking the minimizer of
J(w), assumed differentiable, takes the form
wi = wi−1 − µ∇wT J(wi−1 )
(9)
where the full gradient vector is used at every iteration to
update wi−1 to wi . In a coordinate-descent implementation,
on the other hand, at every iteration i, only a subset of the
entries of the gradient vector is used to perform the update.
These subsets are usually chosen as follows. First, a collection
of K partitions of the parameter space w is defined. These
partitions are defined by diagonal matrices, {Ωk }. Each matrix
has ones and zeros on the diagonal and the matrices add up
to the identity matrix:
K
X
Ωk = I M
(10)
k=1
Multiplying w by any Ωk results in a vector of similar size,
albeit one where the only nontrivial entries are those extracted
by the unit locations in Ωk . At every iteration i, one of the
partitions is selected randomly, say, with probability
Prob(Γi = Ωk ) = ωk
(11)
where the {ωk } add up to one. Subsequently, the gradient
descent iteration is replaced by
wi = wi−1 − µΓi ∇wT J(wi−1 )
(12)
This formulation is known as the randomized block-coordinate
descent (RBCD) algorithm [13]–[15]. At each iteration, the
gradient descent step employs only a collection of coordinates
represented by the selected entries from the gradient vector.
Besides reducing complexity, this step helps alleviate the
condition on the step-size parameter for convergence.
If we reduce our formulation (3)–(5) to the single agent
case, it will become similar to (12) in that the desired cost
function is optimized only along a subset of the coordinates at
each iteration. However, our algorithm offers more randomness
in generating the coordinate blocks than the RBCD algorithm,
by allowing more random combinations of the coordinates at
each time index. In particular, we do not limit the selection
of the coordinates to a collection of K possibilities predetermined by the {Ωk }. Moreover, in our work we use a
random subset of the stochastic gradient vector instead of
the true gradient vector to update the estimate, which is
necessary for adaptation and online learning when the true risk
function itself is not known (since the statistical distribution
of the data is not known). Also, our results consider a general
multi-agent scenario involving distributed optimization where
each individual agent employs random coordinates for its own
gradient direction, and these coordinates are generally different
from the coordinates used by other agents. In other words,
the networked scenario adds significant flexibility into the
operation of the agents under model (5).
B. Relation to Partial Updating Schemes
It is also important to clarify the differences between our
formulation and other works in the literature, which rely on
other useful notions of partial information updates. To begin
with, our formulation (5) is different from the models used
in [16]–[18] where the step-size parameter was modeled as
a random Bernoulli variable, µk (i), which could assume the
values µk or zero with certain probability. In that case, when
the step-size is zero, all entries of ψk,i will not be updated
and adaptation is turned off completely. This is in contrast
to the current scenario where only a subset of the entries are
left without update and, moreover, this subset varies randomly
from one iteration to another.
3
Likewise, the useful work [19] employs a different notion
of partial sharing of information by focusing on the exchange
of partial entries of the weight estimates themselves rather
than on partial entries of the gradient directions. In other
words, the partial information used in this work relates to the
combination steps (3a) and (3c) rather than to the adaptation
step (3b). It also focuses on the special case in which the risks
{Jk (w)} are quadratic in w. In [19], it is assumed that only
a subset of the weight entries are shared (diffused) among
neighbors and that the estimate itself is still updated fully. In
comparison, the formulation we are considering diffuses all
entries of the weight estimates. Similarly, in [20] it is assumed
that some entries of the regression vector are missing, which
causes changes to the gradient vectors. In order to undo these
changes, an estimation scheme is proposed in [20] to estimate
the missing data. In our formulation, more generally, a random
subset of the entries of the gradient vector are set to zero at
each iteration, while the remaining entries remain unchanged
and do not need to be estimated.
There are also other criteria that have been used in the
literature to motivate partial updating. For example, in [21], the
periodic and sequential least-mean-squares (LMS) algorithms
are proposed, where the former scheme updates the whole
coefficient vector every N −th iteration, with N > 1, and the
latter updates only a fraction of the coefficients, which are predetermined, at each iteration. In [22], [23] the weight vectors
are partially updated by following a set-membership approach,
where updates occur only when the innovation obtained from
the data exceeds a predetermined threshold. In [23]–[25], only
entries corresponding to the largest magnitudes in the regression vector or the gradient vector at each agent are updated.
However, such scheduled updating techniques may suffer from
non-convergence in the presence of nonstationary signals [26].
Partial update schemes can also be based on dimensionality
reduction policies using Krylov subspace concepts [27]–[29].
There are also techniques that rely on energy considerations
to limit updates, e.g., [30].
The objective of the analysis that follows is to examine the
effect of random partial gradient information on the learning
performance and convergence rate of adaptive networks for
general risk functions. We clarify these questions by adapting
the framework described in [2], [3].
Notation: We use lowercase letters to denote vectors, uppercase letters for matrices, plain letters for deterministic
variables, and boldface letters for random variables. We also
use (·)T to denote transposition, (·)−1 for matrix inversion,
Tr(·) for the trace of a matrix, diag{·} for a diagonal matrix,
col{·} for a column vector, λ(·) for the eigenvalues of a matrix,
ρ(·) for the spectral radius of a matrix, k·k for the two-induced
norm of a matrix or the Euclidean norm of a vector, kxk2Σ for
the weighted square value xT Σx, ⊗ for Kronecker product,
⊗b for block Kronecker product. Besides, we use p 0
to denote that all entries of vector p are positive. Moreover,
α = O(µ) signifies that |α| ≤ c|µ| for some constant c > 0,
and α = o(µ) signifies that α/µ → 0 as µ → 0. In addition,
the notation lim supn→∞ a(n) denotes limit superior of the
sequence a(n).
II. DATA M ODEL AND A SSUMPTIONS
Let F i−1 represent the filtration (collection) of all random
events generated by the processes {wk,j } and {Γk,j } at all
agents up to time i − 1. In effect, the notation F i−1 refers to
the collection of all past {wk,j , Γk,j } for all j ≤ i − 1 and all
agents.
Assumption 1: (Conditions on indicator variables). It is
assumed that the indicator variables rk,i (m) and r`,i (n) are
independent of each other, for all `, k, m, n. In addition, the
\
variables {rk,i (m)} are independent of F i−1 and ∇
wT J k (w)
for any iterates w ∈ F i−1 and for all agents k.
Let
\
sk,i (φk,i−1 ) , ∇
wT J k (φk,i−1 ) − ∇wT J k (φk,i−1 )
(13)
denote the gradient noise at agent k at iteration i, based on
\
the complete approximate gradient vector, ∇
wT J k (w). We
introduce its conditional second-order moment
Rs,k,i (w) , E[sk,i (w)sT
k,i (w)|F i−1 ].
(14)
The following assumptions are standard and are satisfied by
important cases of interest, such as logistic regression risks or
mean-square-error risks, as already shown in [2], [3]. These
references also motivate these conditions and explain why they
are reasonable.
Assumption 2: (Conditions on gradient noise) [2, pp. 496–
497]. It is assumed that the first and fourth-order conditional
moments of the individual gradient noise processes satisfy the
following conditions for any iterates w ∈ F i−1 and for all
k, ` = 1, 2, . . . , N :
E[sk,i (w)|F i−1 ] = 0
E[sk,i (w)sT
`,i (w)|F i−1 ]
E[ksk,i (w)k4 |F i−1 ]
(15)
= 0, k 6= `
≤
βk4 kwk4
(16)
+
4
σs,k
(17)
4
almost surely, for some nonnegative scalars βk4 and σs,k
.
Assumption 3: (Smoothness conditions) [2, pp. 552,576].
It is assumed that the Hessian matrix of each individual cost
function, Jk (w), and the covariance matrix of each individual
gradient noise process are locally Lipschitz continuous in a
small neighborhood around w = wo in the following manner:
k∇2w Jk (wo + 4w) − ∇2w Jk (wo )k ≤ κc k4wk
o
o
kRs,k,i (w + 4w) − Rs,k,i (w )k ≤ κd k4wk
(18)
γ
(19)
for any small perturbations k4wk ≤ ε and for some κc ≥ 0,
κd ≥ 0, and parameter 0 < γ ≤ 4.
III. M AIN R ESULTS : S TABILITY AND P ERFORMANCE
For each agent k, we introduce the error vectors:
e k,i , wo − wk,i
w
e , wo − φk,i
φ
k,i
e
ψ , wo − ψk,i
k,i
(20)
(21)
(22)
4
We also collect all errors, along with the gradient noise
processes, from across the network into block vectors:
e i , col {w
e 1,i , w
e 2,i , . . . , w
e N,i }
w
o
n
e , col ψ
e ,ψ
e ,...,ψ
e
ψ
i
1,i
2,i
N,i
o
n
e , col φ
e ,φ
e ,...,φ
e
φ
i
1,i
2,i
N,i
si , col {s1,i , s2,i , . . . , sN,i }
(23)
(24)
(25)
(26)
For simplicity, in (26) we use the notation sk,i to replace the
gradient noise sk,i (φk,i−1 ) defined in (13), but note vector
si is dependent on the collection of {φk,i−1 } for all k. We
further introduce the extended matrices:
M , diag{µ1 , µ2 , . . . , µN } ⊗ IM
(27)
A1 , A1 ⊗ IM , A2 , A2 ⊗ IM
(28)
Γi , diag{Γ1,i , Γ2,i , . . . , ΓN,i }
(29)
Note that the main difference between the current work
and the prior work in [2] is the appearance of the random
matrices {Γk,i } defined by (5). In the special case when
the random matrices are set to the identity matrices across
the agents, i.e., {Γk,i ≡ IM }, current coordinate-descent
case will reduce to the full-gradient update studied in [2].
The inclusion of the random matrices {Γk,i } adds a nontrivial level of complication because now, agents update only
random entries of their iterates at each iteration and, importantly, these entries vary randomly across the agents. This
procedure adds a rich level of randomness into the operation
of the multi-agent system. As the presentation will reveal,
the study of the stability and limiting performance under
these conditions is more challenging than in the stochastic
full-gradient diffusion implementation due to at least two
factors: (a) First, the evolution of the error dynamics will
now involve a non-symmetric matrix (matrix D11,i defined
later in (111)); because of this asymmetry, the arguments of
[2] do not apply and need to be modified; and (b) second,
there is also randomness in the coefficient matrix for the error
dynamics (namely, randomness in the matrix B0i defined by
(39)). These two factors add nontrivial complications to the
stability, convergence, and performance analysis of distributed
coordinate-descent solutions, as illustrated by the extended
derivations in Appendices A and B. These derivations illustrate
the new arguments that are necessary to handle the networked
solution of this manuscript. For this reason, in the presentation
that follows, whenever we can appeal to a result from [2], we
will simply refer to it so that, due to space limitations, we
can focus the presentation on the new arguments and proofs
that are necessary for the current context. It is clear from the
proofs in Appendices A and B that these newer arguments are
demanding and not straightforward.
where
T
Bi , AT
2 (I − MΓi Hi−1 ) A1
Hi−1 , diag{H1,i−1 , H2,i−1 , . . . , HN,i−1 }
Z 1
ek,i−1 )dt.
Hk,i−1 ,
∇2w Jk (wo − tφ
(31)
(32)
(33)
0
Proof : Refer to [2, pp. 498–504], which is still applicable to
the current context. We only need to set in that derivation
the matrix Ao to Ao = I, and the vector b to b = 0M N .
These quantities were defined in (8.131) and (8.136) of [2].
The same derivation will lead to (30)–(33), with the main
difference being the appearance now of the random matrix
Γi in (30) and (31).
We assume that the matrix product P = A1 A2 is primitive. This condition is guaranteed automatically, for example,
for ATC and CTA scenarios when the network is stronglyconnected. This means, in view of the Perron-Frobenius Theorem [2], [3], that P has a single eigenvalue at one. We denote
the corresponding eigenvector by p, and normalize the entries
of p to add up to one. It follows from the same theorem that
the entries of p are strictly positive, written as
P p = p, 1T p = 1, p 0
(34)
with 1 being the vector of size N with all its entries equal to
one.
Theorem 1: (Network stability). Consider a stronglyconnected network of N interacting agents running the diffusion strategy (3) with the gradient vector replaced by (5).
Assume the matrix product P = A1 A2 is primitive. Assume
also that the individual cost functions, Jk (w), satisfy the
condition in (2) and that Assumptions 1–2 hold. Then, the
second and fourth-order moments of the network error vectors
are stable for sufficiently small step-sizes, namely, it holds, for
all k = 1, 2, . . . , N , that
ek,i k2 = O(µmax )
lim sup Ekw
(35)
ek,i k4 = O(µ2max )
lim sup Ekw
(36)
i→∞
i→∞
for any µmax < µo , for some small enough µo , where
µmax , max{µ1 , µ2 , . . . , µN }.
(37)
Proof : The argument requires some effort and is given in
Appendix A.
Lemma 1: (Network error dynamics). Consider a network
of N interacting agents running the diffusion strategy (3) with
the gradient vector replaced by (5). The evolution of the error
dynamics across the network relative to the reference vector
wo is described by the following recursion:
Lemma 2: (Long-term network dynamics). Consider a
strongly-connected network of N interacting agents running
the diffusion strategy (3) under (5). Assume the matrix product
P = A1 A2 is primitive. Assume also that the individual
cost functions satisfy (2), and that Assumptions 1–2 and (18)
hold. After sufficient iterations, i 1, the error dynamics
of the network relative to the reference vector wo is wellapproximated by the following model:
ei = Bi w
ei−1 + AT
w
2 MΓi si
0
ei0 = B0i w
ei−1
w
+ AT
2 MΓi si , i 1
(30)
(38)
5
where
B0i
,
AT
2
(I −
MΓi H) AT
1
H , diag{H1 , H2 , . . . , HN }
Hk , ∇2w Jk (wo )
(39)
to first-order in µmax :
MSDk , µmax
lim lim sup
µmax →0
(40)
(41)
MSDav ,
(42)
0
ek,i
lim sup Ekw
k4 = O(µ2max )
(43)
i→∞
i→∞
ei0 k2 = lim sup Ekw
ei k2 + O(µ3/2
lim sup Ekw
max ).
i→∞
Result (35) ensures that the mean-square-error (MSE) performance of the network is in the order of µmax . Using the
long-term model (38), we can be more explicit and derive
the proportionality constant that describes the value of the
network mean-square-error to first-order in µmax . To do so,
we introduce the quantity
(45)
and the gradient-noise covariance matrices:
Gk , lim Rs,k,i (wo )
(46)
G0k , E[Γk,i Gk Γk,i ].
(47)
i→∞
(49)
(50)
Likewise, we define the excess-risk (ER) for each agent k as
the average fluctuation of the normalized aggregate cost
!−1 N
N
X
X
glob
¯
J
(w) ,
qk
qk Jk (w)
(51)
k=1
k=1
with {qk } being entries of the vector q defined by (45), around
its minimum value J¯glob (wo ) at steady state to first-order in
µmax , namely [2, p. 581]:
Proof : To establish (38), we refer to the derivation in [2, pp.
553–555], and note that, in our case, kΓi k ≤ 1 and b = 0M N
(which appeared in (10.2) of [2]). Moreover, the results in (42)
and (43) can be established by following similar techniques
to the proof of Theorem 1, where the only difference is that
the random matrix Hi−1 defined in (32) is now replaced with
the deterministic matrix H defined by (40), and by noting that
the matrices {Hk } in (41) still satisfy the condition (2). With
regards to result (44), we refer to the argument in [2, pp. 557–
560] and note again that kΓi k ≤ 1.
q , diag {µ1 , µ2 , . . . , µN } A2 p
N
1 X
MSDk .
N
(44)
i→∞
µmax
ek,i k2
Ekw
k=1
More specifically, it holds for sufficiently small step-sizes that
0
ek,i
lim sup Ekw
k2 = O(µmax )
i→∞
1
Observe that Gk is the limiting covariance matrix of the
gradient noise process evaluated at wo , and is assumed to
be a constant value, while G0k is a weighted version of it.
A typical example for the existence of the limit in (46) is the
MSE network, where the covariance matrix of the gradient
noise is a constant matrix, which is independent of the time
index i [2, p. 372]. It follows by direct inspection that the
entries of G0k are given by:
(1 − rk )2 Gk (m, n), m 6= n
G0k (m, n) =
(48)
(1 − rk )Gk (m, m), m = n.
We also define the mean-square-deviation (MSD) for each
agent k, and the average MSD across the network to firstorder in µmax — see [2] for further clarifications on these
expressions where it is explained, for example, that MSDk
ek,i k2
provides the steady-state value of the error variance Ekw
ERk , µmax
×
lim lim sup
µmax →0
i→∞
1
µmax
E{J¯glob (wk,i ) − J¯glob (wo )} .
(52)
The average ER across the network is defined by
ERav ,
N
1 X
ERk .
N
(53)
k=1
By following similar arguments to [2, p. 582], it can be verified
that the excess risk can also be evaluated by computing a
weighted mean-square-error variance:
1
ek,i k21 H̄
ERk , µmax
Ekw
(54)
lim lim sup
2
µmax →0 i→∞ µmax
where H̄ denotes the Hessian matrix of the normalized aggregate cost, J¯glob (w), evaluated at the minimizer w = wo :
!−1 N
N
X
X
qk Hk
(55)
H̄ ,
qk
k=1
k=1
with Hk defined by (41). Moreover, we define the convergence
ek,i k2 ,
rate as the slowest rate at which the error variances, Ekw
converge to the steady-state region. By iterating the recursion
for the second-order moment of the error vector, we will arrive
at a relation in the following form:
ei k2 ] = E kw
e−1 k2F i+1 + c
(56)
E[kw
e−1 denotes the
for some matrix F and constant c, where w
network error vector at the initial time instant. The first-term
on the right-hand side corresponds to a transient component
that dies out with time, and the second-term denotes the
ei k2 ] converges to. Then, the
steady-state region that E[kw
2
ei k ] towards its steady-state region is
convergence rate of E[kw
dictated by ρ(F ) [2, p. 395]. The following conclusion is one
of the main results in this work. It shows how the coordinate
descent construction influences performance in comparison to
the standard diffusion strategy where all entries of the gradient
vector are used at each iteration. Following the statement of
the result, we illustrate its implications by considering several
important cases.
6
Theorem 2: (MSD and ER performance). Under the same
setting of Theorem 1, and assume also that Assumption 3
holds, it holds that, for sufficiently small step-sizes:
MSDcoor,k = MSDcoor,av
!−1 N
N
X
1 X
= Tr
qk (1 − rk )Hk
qk2 G0k
2
k=1
k=1
N
ERcoor,k = ERcoor,av
(57)
X
1
= Tr X
qk2 G0k
2
!
(58)
k=1
where the subscript “coor” denotes the stochastic coordinatedescent diffusion implementation, and matrix X is the unique
solution to the following Lyapunov equation:
!
!
N
N
X
X
X
qk (1 − rk )Hk +
qk (1 − rk )Hk X = H̄
k=1
k=1
(59)
with H̄ defined by (55). Moreover, for large enough i, the
ek,i k2 , towards
convergence rate of the error variances, Ekw
the steady-state region (57) is given by
!
N
X
+1)/N
αcoor = 1 − 2λmin
qk (1 − rk )Hk + O µ(N
max
k=1
(60)
Proof : See Appendix B.
IV. I MPLICATIONS AND U SEFUL C ASES
A. Uniform Missing Probabilities
Consider the case when the missing probabilities are identical across the agents, i.e., {rk ≡ r}.
1) Convergence time: Consider the ATC or CTA forms of
the full-gradient or coordinate-descent diffusion strategy (3a)–
(3c) and (5). From (56), we find that the error variances for
the distributed strategies evolve according to a relation of the
form:
ek,i k2 ≤ αi+1 Ekw
ek,−1 k2 + c
Ekw
(61)
for some constant c > 0, and where the parameter α determines the convergence rate. Its value is denoted by αgrad for
the full-gradient implementation and is given by [2, p. 584]:
!
N
X
αgrad = 1 − 2λmin
qk Hk + o (µmax )
(62)
k=1
Likewise, the convergence rate for the coordinate-descent variant is denoted by αcoor and is given by expression (60). It is
clear that αcoor ≥ αgrad for 0 ≤ r < 1, so that the coordinatedescent implementation converges at a slower rate as expected
(since it only employs partial gradient information). Thus, let
Tcoor and Tgrad denote the largest number of iterations that
ek,i k2 , to converge to
are needed for the error variances, Ekw
their steady-state regions. The values of Tcoor and Tgrad can be
estimated by assessing the number of iterations that it takes
ek,−1 k2 in (61) to assume a
for the transient term αi+1 E kw
higher-order value in µmax , i.e., for
Tcoor
ek,−1 k2 = dµ1+
αcoor
Ekw
max
(63)
Tgrad
ek,−1 k2
αgrad
Ekw
(64)
=
dµ1+
max
for some proportionality constant d, and small number > 0.
Then, it holds that
ln αgrad
Tcoor
=
Tgrad
ln αcoor
P
N
ln 1 − 2λmin
(a)
k=1 qk Hk
P
≈
N
ln 1 − 2λmin
q
(1
−
r)H
k
k
k=1
P
N
q
H
−2λ
k
k
min
(b)
k=1
P
≈
N
−2(1 − r)λmin
q
H
k=1 k k
1
(65)
1−r
where in step (a) we ignored the higher-order term in µmax ,
and in (b) we used ln(1 − x) ≈ −x as x → 0. Expression
(65) reveals by how much the convergence time is increased
in the coordinate-descent implementation. Note that because
of longer convergence time, the stochastic coordinate-descent
diffusion implementation may require more quantities to be
exchanged across the network compared to the full-gradient
case.
=
2) Computational complexity: Let us now compare the
computational complexity of both implementations: the
coordinate-descent and the full-gradient versions. Assume that
the computation required to calculate each entry of the gradient
\
vector ∇
wT J k (φk,i−1 ) is identical, and let cm ≥ 0 and
ca ≥ 0 denote the number of multiplications and additions,
respectively, that are needed for each entry of the gradient
vector.
Let nk , |Nk | denote the degree of agent k. Then, in the
full-gradient implementation, the adaptation step (3b) requires
cm M + M multiplications and ca M + M additions, while the
combination step (3a) or (3c) requires nk M multiplications
and (nk − 1)M additions. In the coordinate-descent implementation, the adaptation step (3b) with the gradient vector
replaced by (5) requires (1 − r) · (cm M + M ) multiplications
and (1 − r) · (ca M + M ) additions on average, while the
combination step (3a) or (3c) requires nk M multiplications
and (nk − 1)M additions. Let mcoor,k and mgrad,k denote the
combined number of multiplications required by the adaptation
and combination steps per iteration at each agent k in the
coordinate-descent and full-gradient cases. Then,
mgrad,k = (cm + nk + 1)M
(66)
mcoor,k = mgrad,k − (cm + 1)M r
(67)
If we now consider that these algorithms take Tcoor and Tgrad
iterations to reach their steady-state regime, then the total
number of multiplications at agent k, denoted by Mcoor,k and
7
!−1 N
N
X
1 X
= Tr
q k Hk
qk2 Gk +
2
k=1
k=1
!−1 N
N
X
X
r
Tr
qk Hk
qk2 diag{Gk } −
2
k=1
k=1
!
−1
N
N
X
r X
Tr
qk Hk
qk2 Gk .
2
Mgrad,k , are therefore given by
Mcoor,k = mcoor,k Tcoor
(68)
Mgrad,k = mgrad,k Tgrad
(69)
Mcoor,k
mcoor,k 1
=
Mgrad,k
mgrad,k 1 − r
(70)
so that using (65):
Now, the first term on the right hand side satisfies
mcoor,k
cm + 1
=1−
r
mgrad,k
cm + nk + 1
k=1
(71)
so that from (70) and (71):
1≤
Mcoor,k
cm + 1
= (1 − r)−1 1 −
r
Mgrad,k
cm + n k + 1
(72)
since 0 ≤ r < 1. It is clear that when it is costly to compute
the gradient entries, i.e., when cm nk , then Mcoor,k and
Mgrad,k will be essentially identical. This means that while
the coordinate-descent implementation will take longer to
converge, the savings in computation per iteration that it provides is such that the overall computational complexity until
convergence remains largely invariant (it is not increased).
This is a useful conclusion. It means that in situations where
computations at each iteration need to be minimal (e.g., when
low end sensors are used), then a coordinate-descent variant is
recommended and it will be able to deliver the same steadystate performance (to first-order in µmax , see (78) ahead)
with the total computational demand spread over a longer
number of iterations. This also means that the complexity and
convergence rate measures, when normalized by the number
of entries that are truly updated at each iteration, remain
effectively invariant. A similar analysis and conclusion holds
if we examine the total number of additions (as opposed to
multiplications) that are necessary.
3) MSD performance: The matrix G0k defined by (48) can
be written as
G0k = (1 − r)2 Gk + (1 − r) − (1 − r)2 diag{Gk }
r
= (1 − r)2 Gk +
diag{Gk }
(73)
1−r
where the term diag{Gk } is a diagonal matrix that consists
of the diagonal entries of Gk . Then, the MSD expression (57)
gives
!−1
N
X
(73) 1
MSDcoor,k = (1 − r)Tr
qk Hk
×
2
k=1
!
N
X
r
qk2 Gk +
diag{Gk }
1−r
k=1
(74)
k=1
By recognizing that the first item in (74) is exactly the MSD
expression for the stochastic full-gradient diffusion case [2, p.
594], which is denoted by “MSDgrad,k ”, we get
MSDcoor,k − MSDgrad,k
r
= Tr
2
N
X
!−1
q k Hk
k=1
N
X
!
qk2 Ǧk
(75)
k=1
where
Ǧk , diag{Gk } − Gk .
(76)
We show in Appendix C that the difference in (75) can be
positive or negative, i.e., the MSD performance can be better or
worse in the stochastic coordinate-descent case in comparison
to the stochastic full-gradient case. Recall from (49) that the
MSD performance is evaluated to first-order in µmax . Then,
the MSD gap in (75) is to first-order in the step-size parameter.
Observe that the missing probability r on the right hand side
of that equation is independent of µmax . It thus follows that
!−1 N
!
N
X
X
2
qk Ǧk = O(µmax ).
Tr
q k Hk
(77)
k=1
k=1
Corollary 1: (Small missing probabilities). Let r =
O(µεmax ) for a small number ε > 0. It holds that
MSDcoor,k − MSDgrad,k = O(µ1+ε
max ) = o(µmax ).
(78)
Proof : It follows from (75) and (77).
We proceed to provide a general upper bound for the
difference between MSDcoor,k and MSDgrad,k .
Corollary 2: (Upper bound). Under the same conditions of
Theorem 2, and when the missing probabilities are uniform,
namely, {rk ≡ r}, it holds that:
|MSDcoor,k − MSDgrad,k | ≤
!−1
N
N
1
1 X 2
r X
qk
−
qk Tr(Gk ) (79)
2
νd
δd
k=1
k=1
where the positive numbers νd ≤ δd are defined in (2), and
the matrices {Gk } are defined by (46). Furthermore, when the
matrices {Hk } or {Gk } are diagonal, it follows that
MSDcoor,k = MSDgrad,k
(80)
Proof : See Appendix D.
Corollary 3: (Uniform step-sizes). Continuing with the
setting of Corollary 2 by assuming now that the step-sizes
are uniform across all agents and A1 = I or A2 = I
(corresponding to either the ATC or CTA formulations). Let
8
{pk } be entries of the vector p defined by (34). Then, in view
of (45) and (34), qk = µpk and the {pk } add up to one. In
this case, the sum of the {qk } is equal to µ and expression
(79) simplifies to
|MSDcoor,k − MSDgrad,k | ≤
N
r
1
1 X 2
µ
−
pk Tr(Gk ). (81)
2
νd
δd
k=1
Consider now MSE networks where the risk function that
is associated with each agent k is the mean-square-error:
Jk (w) = E(dk (i) − uk,i w)2
(82)
where dk (i) denotes the desired signal, and uk,i is a (row)
regression vector. In these networks, the data {dk (i), uk,i }
are assumed to be related via the linear regression model
dk (i) = uk,i wo + vk (i)
(83)
where vk (i) is zero-mean white measurement noise with vari2
ance σv,k
and assumed to be independent of all other random
variables. The processes {dk (i), uk,i , vk (i)} are assumed to
be jointly wide-sense stationary random processes. Assume
also that the regression data {uk,i } are zero-mean, and white
over time and space with
E uT
k,i u`,j , Ru,k δk,` δi,j
(84)
where Ru,k > 0, and δk,` denotes the Kronecker delta
sequence. Consider the case when the covariance matrices of
the regressors are identical across the network, i.e., {Ru,k ≡
Ru > 0}. Then, it holds that [2, p. 598]
2
Hk ≡ 2Ru , Gk = 4σv,k
Ru .
(85)
matrices, i.e., {Ru,k ≡ Ru > 0}, it holds that
0 ≤ MSDcoor,k − MSDgrad,k ≤
!−1 N
!
N
X
X
δd
2 2
r
qk
−1 M
qk σv,k
νd
k=1
(87)
k=1
Moreover, it holds that MSDcoor,k = MSDgrad,k if, and only
if, Ru is diagonal.
Proof : It follows from Corollary 2 by using Tr (Gk ) =
2
4σv,k
Tr (Ru ) and noting that νd /2 ≤ λ (Ru ) ≤ δd /2 according to (85) and (2).
4) ER performance: Consider the scenario when the missing probabilities are identical across the agents, i.e., {rk ≡ r}.
Then, expression (59) simplifies to
!
!
N
N
X
X
(1 − r)
qk X H̄ + (1 − r)
qk H̄X = H̄ (88)
k=1
k=1
where we used the equality
follows that
PN
1
X = (1 − r)−1
2
k=1 qk Hk
N
X
=
P
N
k=1 qk
H̄, it
!−1
qk
IM .
(89)
k=1
Thus, the ER expression in (58) can be rewritten as:
!−1
!
N
N
X
X
1
−1
2 0
ERcoor,k = (1 − r)
qk
Tr
qk Gk
4
k=1
k=1
!−1 N
N
X
X
(a) 1
qk2 Tr (Gk )
(90)
=
qk
4
k=1
k=1
which is exactly the same result for the full gradient case
from [2, p. 608], and where the equality (a) holds because
Tr (G0k ) = (1 − r)Tr (Gk ) according to the definition in (48).
Substituting into (75) we have
MSDcoor,k − MSDgrad,k
!−1 N
!
N
X
X
2 2
=r
qk
qk σv,k Tr Ru−1 diag{Ru } − M
k=1
B. Uniform Individual Costs
k=1
≥0
(86)
where (86) holds because Tr Ru−1 diag{Ru } ≥ M , which
can be shown by using the property that Tr (X) Tr X −1 ≥
M 2 for any M ×M symmetric positive-definite matrix X [31,
1
1
p. 317], and choosing X = diag 2 {Ru }Ru−1 diag 2 {Ru }. In
the case of MSE networks, by exploiting the special relation
between the matrices {Hk } and {Gk } in (85), we are able
to show that the MSD in the stochastic coordinate-descent
case is always larger (i.e., worse) than or equal to that in
the stochastic full-gradient diffusion case (although by not
more than o(µmax ), as indicated by (78)). We are also able
to provide a general upper bound on the difference between
these two MSDs.
Corollary 4: (MSE networks). Under the same conditions
of Corollary 2, and for MSE networks with uniform covariance
Consider the case when the individual costs, Jk (w), are
identical across the network, namely, [2, p. 610]
Jk (w) ≡ J(w) , EQ(w; xk,i )
(91)
where Q(w; xk,i ) denotes the loss function. In this case, it
will hold that the matrices {Hk , Gk } are uniform across the
agents, i.e.,
Hk = ∇2w J(wo ) ≡ H
(92)
T
Gk = E∇wT Q(wo ; xk,i ) [∇wT Q(wo ; xk,i )] ≡ G
(93)
in view of ∇wT J(wo ) = 0. Then, (92) ensures the matrix
H̄ = H according to the definition in (55). By referring to
(59), we have
!−1
N
1 X
qk (1 − rk )
IM .
(94)
X=
2
k=1
9
Then, expressions (57) and (58) reduce to
MSDcoor,k = MSDcoor,av
1
=
2
N
X
!−1
qk (1 − rk )
k=1
N
X
qk2 Tr H −1 G0k
(95)
k=1
ERcoor,k = ERcoor,av
N
X
!−1
qk (1 − rk )
k=1
N
X
qk2 (1 − rk )Tr (G) . (96)
k=1
We proceed to compare the MSD and ER performance in the
stochastic full-gradient and coordinate-descent cases. Let
PN
PN
2
2
qk2
k=1 qk (1 − rk )
− Pk=1
(97)
α , PN
N
k=1 qk (1 − rk )
k=1 qk
PN
PN
2
qk2
k=1 qk (1 − rk )
θ , PN
− Pk=1
(98)
N
k=1 qk (1 − rk )
k=1 qk
and note that α ≤ θ, with equality if, and only if, {rk ≡ 0}.
Corollary 5: (Performance comparison). Under the same
conditions of Theorem 2, when the individual costs Jk (w) are
identical across the agents, it holds that:
a) if α ≥ 0:
0 ≤ MSDcoor,k − MSDgrad,k
1 θ
≤
Tr (G)
2 νd
Fig. 1. Network topology consisting of N = 100 agents.
10
Theoretical MSD for full update
Theoretical MSD for partial update
MSE network, full update
MSE network, partial update
0
Transient MSD (dB)
1
=
4
−10
−20
−30
−40
−50
−60
0.5
1
1.5
2
2.5
3
3.5
(99)
4
4
Iterations
x 10
b) if α < 0, and θ ≥ (1 − δd /νd ) α ≥ 0:
Fig. 2. MSD learning curves, averaged over 200 independent runs, in the
case of Corollary 1 when {rk = 0.1}. The dashed lines show the theoretical
MSD values from (57).
0 ≤ MSDcoor,k − MSDgrad,k ≤
1
1
1 θ
+
−
α Tr (G) (100)
2 νd
δd
νd
A. MSE Networks
Likewise, it holds that
θ
Tr (G) .
(102)
4
Then, in the case when either the missing probabilities or the
quantities {qk } are uniform across the agents, namely, {rk ≡
r} or {qk ≡ q}, it follows that
ERcoor,k − ERgrad,k =
ERcoor,k = ERgrad,k .
(103)
Proof : See Appendix E.
Note that for the other choices of parameter θ that are not
indicated in Corollary 5, there is no consistent conclusion on
which MSD (between MSDcoor,k and MSDgrad,k ) is lower.
10
Transient MSD (dB)
c) if α < 0, and θ ≤ (1 − νd /δd ) α ≤ 0:
1
1 θ
1
+
−
α Tr (G) ≤
2 δd
νd
δd
MSDcoor,k − MSDgrad,k ≤ 0. (101)
In the following examples, we will test performance of
the associated algorithms in the case when uniform missing
probabilities are utilized across the agents. We adopt the ATC
formulation, and set the combination matrices A1 = I, and
A2 according to the averaging rule [2, p. 664] in the first two
examples, and Metropolis rule [2, p. 664] in the third example.
In the first example, we test the case when the gradient vectors
are missing with small probabilities across the agents. Figure
1 shows a network topology with N = 100 agents. The
parameter vector wo is randomly generated with M = 10.
Theoretical MSD
MSE network, full update
MSE network, partial update
convergence rate for full update
convergence rate for partial update
0
−10
−20
−30
−40
−50
−60
0.2
V. S IMULATION R ESULTS
In this section, we illustrate the results by considering MSE
networks and logistic regression networks; both settings satisfy
condition (2) and Assumptions 1 through 3.
0.4
0.6
0.8
1
1.2
Iterations
1.4
1.6
1.8
2
4
x 10
Fig. 3. MSD learning curves, averaged over 200 independent runs, in the
case of Corollary 2 when {Hk , Gk } are diagonal. The dashed line along the
horizontal axis shows the theoretical MSD value from (57). Those along the
learning curves show the reference recursion at rates formulated by (60).
10
0
−28
−15
−28.5
−20
−29
−25
−29.5
1700
1800
1900
2000
Theoretical MSD for full update
Theoretical MSD for partial update
MSE network, full update
MSE network, partial update
0
−5
−10
−15
−20
5
Theoretical ER
MSE network, full update
MSE network, partial update
0
Transient ER (dB)
Transient MSD (dB)
−10
5
Transient MSD (dB)
Theoretical MSD for full update
Theoretical MSD for partial update
MSE network, full update
MSE network, partial update
−5
−5
−10
−15
−20
−25
−25
−30
−30
500
1000
1500
2000
−30
1000
2000
3000
4000
5000
6000
1000
2000
3000
4000
Iterations
Iterations
Iterations
(a)
(b)
(c)
5000
6000
7000
Fig. 4. Learning curves, averaged over 10000 independent runs, and theoretical results calculated from (57) and (58) respectively, for a two-agent MSE
network, with parameters {π1 = −0.34, π2 = 0.99} in (a), and {π1 = 0.34, π2 = 0.99} in (b) and (c).
The regressors are generated by the first-order autoregressive
model
q
uk,i (m) = πk uk,i (m − 1) + 1 − πk2 tk,i (m), 1 ≤ m < M
(104)
and the variances are scaled to be 1. The processes {tk,i }
are zero-mean, unit-variance, and independent and identically
distributed (i.i.d) Gaussian sequences. The parameters {πk }
are generated from a uniform distribution on the interval
(−1, 1). The noises, uncorrelated with the regression vectors,
are zero-mean white Gaussian sequences with the variances
uniformly distributed over (0.001, 0.1). The step-sizes {µk }
across the agents are generated from a uniform distribution
on the interval (0.0001, 0.0005). We choose a small missing
probability {rk = 0.1}. Figure 2 shows the simulation results, which are averaged over 200 independent runs, as well
as the theoretical MSD values calculated from (57), which
are −57.72dB and −57.61dB, respectively, for the full and
partial update case. It is clear from the figure that, when the
gradient information is missing with small probabilities, the
performance of the coordinate-descent case is close to that of
the full-gradient diffusion case.
In the second example, we test the case when the regressors
are white across the agents. We randomly generate wo of
size M = 10. The white regressors are generated from zeromean white Gaussian sequences, and the powers, which vary
from entry to entry, and from agent to agent, are uniformly
distributed over (0.05, 0.15). The noises {vk (i)}, uncorrelated
with the regressors, are zero-mean white Gaussian sequences,
2
with the variances {σv,k
} generated from uniform distribution
on the interval (0.0001, 0.01). The step-sizes are uniformly
distributed over (0.001, 0.01). The results, including the theoretical MSD value from (57) in Theorem 2, the simulated
MSD learning curves, and the convergence rates from (60),
are illustrated by Fig. 3, where the results are averaged over
200 independent runs. It is clear from the figure that, when
white regressors are utilized in MSE networks, the stochastic
coordinate-descent case converges to the same MSD level
as the full-gradient diffusion case, which verifies (80), at a
convergence rate formulated in (60).
In the third example, we revisit the two-agent MSE network
discussed in Appendix C, i.e., N = 2. We randomly generate
wo of size M = 2. The step-sizes µ1 = µ2 = 0.005 are
uniform across the agents, which gives q1 = q2 = 2.5 ×
7
14
1
4
3
13
20
11
15
18
8
16
2
6
19
17
5
10
9
12
Fig. 5. Network topology consisting of N = 20 agents.
10−3 . The missing probabilities r1 = r2 = 0.5. The noises
{v1 (i), v2 (i)} are zero-mean white Gaussian sequences with
2
2
the variances {σv,1
= 0.5, σv,2
= 5 × 10−4 }. The regressors,
uncorrelated with the noise sequences, are scaled such that the
covariance matrices are of the form
|π1 | π1
|π2 | π2
Ru,1 =
, Ru,2 =
(105)
π1
1
π2
1
with |π1 | < 1, |π2 | < 1. Now we select parameters {π1 =
−0.34, π2 = 0.99}, which satisfy condition (169a), and
{π1 = 0.34, π2 = 0.99} to illustrate the cases of MSDcoor,k <
MSDgrad,k and MSDcoor,k > MSDgrad,k respectively. Fig. 4
(a) shows the simulation results with the parameters {π1 =
−0.34, π2 = 0.99}. Figures 4 (b) and 4 (c) show the simulation results with the parameters {π1 = 0.34, π2 = 0.99}.
All results are averaged over 10000 independent runs. It is
clear from the figures that the simulation results match well
with the theoretical results from Theorem 2. In Fig. 4 (a), the
steady-state MSD of the stochastic coordinate-descent case is
slightly lower than that of the full-gradient diffusion case, by
about 0.32dB, which is close to the theoretical MSD difference
of 0.41dB from (75). The MSD performance is better in the
full-gradient diffusion case in Fig. 4 (b), and the difference
between these two MSDs at steady state is 1.71dB, which is
close to the theoretical difference of 1.49dB from (75). The
ER performance for both the stochastic coordinate-descent and
full-gradient diffusion cases are the same as illustrated in Fig.
4 (c), which verifies the theoretical result in (90).
11
−5
Theoretical ER for full update
Theoretical ER for partial update
logistic network, full update
logistic network, partial update
−20
−25
−30
−35
−15
−20
−40
−25
−42
−30
3.5
3.6
3.7
4
−35
−40
0.5
1
1.5
2
Iterations
(a)
2.5
3
−45
3.5
−15
−20
−40
−25
−42
−30
5.8
−35
x 10
5.82
5.84
5.86
4
x 10
−40
−40
−45
Theoretical ER for full update
Theoretical ER for partial update
logistic network, full update
logistic network, partial update
−10
Transient ER (dB)
−10
Transient ER (dB)
−15
Transient ER (dB)
−5
−5
Theoretical ER
logistic network, full update
logistic network, partial update
−10
0.5
1
1.5
4
x 10
2
Iterations
2.5
(b)
3
−45
3.5
4
x 10
1
2
3
Iterations
4
5
6
4
x 10
(c)
Fig. 6. ER learning curves, averaged over 1000 independent runs, and theoretical results from (96) for diffusion learning over a logistic network with full or
partial updates. Corollary 5 is tested in (a) when a uniform step-size and a doubly-stochastic combination matrix are utilized across the network. Corollary 5
is tested when the parameters {µk } and {rk } are scaled to make θ in (98) negative in (b) and positive in (c).
B. Logistic Networks
We now consider an application in the context of pattern
classification. We assign with each agent the logistic risk
n h
io
T
ρ
Jk (w) = kwk2 + E ln 1 + e−γk (i)hk,i w
(106)
2
with regularization parameter ρ > 0, and where the labels
{γk (i) = ±1} are binary random and the {hk,i } are feature
vectors. The objective is for the agents to determine a parameter vector wo to enable classification by estimating the class
o
bk (i) = hT
labels via γ
k,i w .
We proceed to test the theoretical findings in Corollary 5.
Consider the network topology in Fig. 5 with N = 20 agents.
We still adopt the ATC formulation, and set the combination
matrices A1 = I, and A2 according to the Metropolis rule in
[2, p. 664]. The feature vectors and the unknown parameter
vector are randomly generated from uncorrelated zero-mean
unit-variance i.i.d Gaussian sequences, both of size M = 10.
The parameter ρ in (106) is set to 0.01. To generate the
trajectories for the experiments, the optimal solution to (106),
wo , the Hessian matrix H, and the gradient noise covariance
matrix, G, are first estimated off-line by applying a batch
algorithm to all data points.
In the first example, we consider the case when a uniform
step-size {µk = 0.005} is utilized across the agents. All entries
of the stochastic gradient vectors are missing completely at
random with positive probabilities that are uniformly distributed over (0, 1). Figure 6 (a) shows the transient ER curves
for the diffusion strategies with complete and partial gradients,
where the results are averaged over 1000 independent runs.
The figure also shows the theoretical result calculated from
(96). It is clear from Figure 6 (a) that the same ER performance
is obtained in the stochastic coordinate-descent and fullgradient diffusion cases, by utilizing a uniform step-size and
a doubly-stochastic combination matrix across the agents (in
which case the parameters {qk } in (45) are identical across the
agents), which is in agreement with the theoretical analysis in
(103).
In the second and third examples, we randomly generate the
step-sizes {µk } and missing probabilities {rk } by following
uniform distributions on the intervals (0.001, 0.01) and (0, 1)
respectively. In Figure 6 (b), the parameters {µk } and {rk }
are scaled to get a negative value for θ in (98), and in
Fig. 6 (c), those parameters are scaled to make θ positive.
Figures 6 (b) and 6 (c) show respectively the transient ER
learning curves in these two cases for the diffusion strategies
with complete and partial gradients, where the results are
averaged over 1000 independent runs. The figures also show
the theoretical results calculated from (96). It is clear from
Figs. 6 (b) and 6 (c) that these learning curves converge to
their theoretical results at steady state. In Fig. 6 (b) where
θ < 0, the stochastic coordinate-descent case converges to
a lower ER level than the full-gradient diffusion case, and
the difference between these two ERs is 0.637dB, which is
close to the theoretical difference of 0.640dB from (102).
In Fig. 6 (c) where θ > 0, the steady-state ER in the fullgradient diffusion case is lower than that of the stochastic
coordinate-descent case, by about 0.726dB, which is close to
the theoretical difference of 0.750dB from (102).
A PPENDIX A
P ROOF OF T HEOREM 1
Let P = A1 A2 . It was argued in [2, p.510] that P admits
a Jordan canonical decomposition of the form P = V JV−1
where
T
1
1 0
−1
,J=
V , p VR , V ,
(107)
0 J
VLT
p is defined by (34), denotes an arbitrary positive scalar that
we are free to choose, and the matrix J has a Jordan structure
with appearing in the first lower diagonal rather than unit
entries. All eigenvalues of J are strictly inside the unit circle.
Then,
P , P ⊗ IM , V J V−1
(108)
where V , V ⊗ IM , J , J ⊗ IM . Using (108), we can
rewrite Bi from (31) as
T
T
Bi , V−1 (J T − D T
(109)
i )V
where
T T
T
−1
DT
i , V A2 MΓi Hi−1 A1 V
T
T
D11,i D21,i
=
T
T
D12,i D22,i
T
(110)
12
and
D11,i =
N
X
512], we have
qk Hk,i−1 Γk,i
(111)
k=1
with the vector q = {qk } defined by (45). With regards to the
norm of D11,i , we observe that contrary to the arguments
in [2, p. 511], this matrix is not symmetric anymore in
the coordinate-descent case due to the presence of Γk,i . We
therefore need to adjust the arguments, which we do next.
Let
D̄11,i , E [D11,i |F i−1 ]
N
X
=
(6)
=
qk Hk,i−1 E [Γk,i ]
k=1
N
X
qk (1 − rk )Hk,i−1
k=1
T
= E D11,i
|F i−1 .
E [Γk,i Γj,i ] =
(1 − rk )(1 − rj ), k 6= j
1 − rk ,
k=j
(113)
we introduce
i
T
D11,i − D̄11,i |F i−1
T
T
= E D11,i D11,i
|F i−1 − D̄11,i E D11,i
|F i−1 −
RD11 ,i , E
h
D11,i − D̄11,i
2
D̄11,i
E [D11,i |F i−1 ] D̄11,i +
(112)
T
2
= E D11,i D11,i
|F i−1 − D̄11,i
=
N X
N
X
(114)
qk qj Hk,i−1 E [Γk,i Γj,i ] Hj,i−1 −
k=1 j=1
N X
N
X
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1
k=1 j=1
(113)
=
N
N
X
X
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1 −
k=1 j6=k=1
N X
N
X
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1 +
k=1 j=1
N
X
=
=
k=1
N
X
Now, multiplying both sides of (30) by VT , we have
T
ei = (J T − D T
ei−1 + VT AT
VT w
i )V w
2 MΓi si
(118)
where the asymmetry of the matrix D11,i in this case leads
to a difference in the first row, compared to the arguments in
[2, pp. 514–515]. We adjust the arguments as follows. Using
Jensen’s inequality, we have [2, p. 515]:
E[kw̄i k2 |F i−1 ] ≤
1
T
E[k(IM − D11,i
)w̄i−1 k2 |F i−1 ]
1−t
1
T
+ E[kD21,i
w̌i−1 k2 |F i−1 ] + E[ks̄i k2 |F i−1 ] (122)
t
for any 0 < t < 1, where the expectation of the cross term
T
T
between s̄i and (IM − D11,i
)w̄i−1 − D21,i
w̌i−1 vanishes
conditioned on F i−1 and Γi in view of (15), and the result in
(122) follows by taking the expectations again on both sides
over Γi . Then, the first term on the right hand side of (122)
can be bounded by
T
E k(IM − D11,i
)w̄i−1 k2 |F i−1
T
T
= (w̄i−1 ) E (IM − D11,i ) IM − D11,i
|F i−1 w̄i−1
(a)
2
T
≤ λmax E (IM − D11,i ) IM − D11,i
|F i−1 kw̄i−1 k
(b)
T
= E (IM − D11,i ) IM − D11,i
|F i−1 kw̄i−1 k2
(112)
T
= IM − 2D̄11,i + E D11,i D11,i
|F i−1 kw̄i−1 k2
(114)
2
qk2 (1 − rk )Hk,i−1
2
IM − 2D̄11,i + D̄11,i
+ RD11 ,i kw̄i−1 k2
2
≤
IM − D̄11,i
+ RD11 ,i kw̄i−1 k2
(c)
2
=
IM − D̄11,i + RD11 ,i kw̄i−1 k2
2 2
≤ (1 − σ11 µmax )2 + β11
µmax kw̄i−1 k2
=
k=1
N
X
2
for some positive constants σ11 and β11
, and sufficiently small
µmax .
where (109) was used. Let
ei
pT ⊗ IM w
w̄i
ei =
VT w
,
(119)
ei
w̌i
VRT ⊗ IM w
q T ⊗ IM Γi si
s̄i
VT AT
MΓ
s
=
,
i i
2
ši
VRT ⊗ IM AT
2 MΓi si
(120)
We then rewrite (118) as
T
T
−D21,i
IM − D11,i
w̄i−1
s̄i
w̄i
+
=
T
T
w̌i−1
ši
w̌i
−D12,i
JT − D22,i
(112)
(121)
Noting that
2 2
kIM − D̄11,i k ≤ 1 − σ11 µmax , kRD11 ,i k ≤ β11
µmax (117)
2
qk2 (1 − rk )Hk,i−1
−
N
X
2
qk2 (1 − rk )2 Hk,i−1
k=1
2
qk2 (1 − rk )rk Hk,i−1
.
(115)
(123)
(116)
where in step (a) we called upon the Rayleigh-Ritz characterization of eigenvalues [32], [33], and (b), (c) hold because
kAk = λmax (A) for any symmetric positive-semidefinite
matrix A, and kA2 k = kAk2 for any symmetric matrix A.
Then, matrices D̄11,i and RD11 ,i are symmetric positivedefinite. Following similar arguments to those in [2, pp. 511–
Computing the expectations again on both sides of (123),
k=1
Recall, from (2) and (33), that
0 < νd IM ≤ Hk,i−1 ≤ δd IM .
13
Let X be a constant matrix of size M × M . Then,
(1 − rk )(1 − rj )X, k 6= j
E [Γk,i XΓj,i ] =
(134)
X 0,
k=j
we have
1
T
E E[k(IM − D11,i
)w̄i−1 k2 |F i−1 ]
1−t
1
2 2
≤
(1 − σ11 µmax )2 + β11
µmax Ekw̄i−1 k2
1− t
2 2
β11
µmax
(a)
= 1 − σ11 µmax +
Ekw̄i−1 k2
1 − σ11 µmax
where X 0 has the same form as (48), and we can further
rewrite X 0 as (73). It follows that:
T
2
E D11,i
D11,i |F i−1 − D̄11,i
(b)
0
≤ (1 − σ11
µmax ) Ekw̄i−1 k2
(124)
where in step (a) we set t = σ11 µmax , and in (b) positive
0
0
number σ11
< σ11 , and µmax is small enough such that σ11
≤
−1 2
σ11 − (1 − σ11 µmax ) β11 µmax . We can now establish (35)
by substituting (124) into (122), and completing the argument
starting from Eq. (9.69) in the proof of Theorem 9.1 in [2, pp.
516–521], where the quantity b = 0M N (appeared in (9.54)
of [2]).
=
N X
N
X
N X
N
X
for any 0 < t < 1, which appeared in (9.117) of [2]. Let
T
T
Ki , (IM − D11,i ) IM − D11,i
w̄i−1 (w̄i−1 )
T
× (IM − D11,i ) IM − D11,i
(126)
2
T
Li , (IM − D11,i ) IM − D11,i
.
(127)
Then, both matrices Ki and Li are symmetric positive semidefinite. Thus, we have
T
E k(IM − D11,i
)w̄i−1 k4 |F i−1
T
= (w̄i−1 ) E [Ki |F i−1 ] w̄i−1
(134)
=
N
N
X
X
2
qk2 (1 − rk )2 Hk,i−1
+
k=1
N X
N
X
2
qk2 (1 − rk )rk diag{Hk,i−1
}−
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1
k=1 j=1
=
N
X
2
qk2 (1 − rk )rk diag{Hk,i−1
}.
(135)
k=1
2
Recall from (116) that {Hk,i−1 > 0}. Then, {Hk,i−1
> 0}
2
and {diag{Hk,i−1 } > 0}. Computing Euclidean norms on
both sides of (135), we have
T
2
E D11,i
D11,i |F i−1 − D̄11,i
N
(a) X
≤
k=1
N
(b) X
2
qk2 (1 − rk )rk diag{Hk,i−1
}
2
qk2 (1 − rk )rk Tr Hk,i−1
k=1
≤ Tr (E [Ki |F i−1 ]) kw̄i−1 k
T
= w̄i−1
E [Li |F i−1 ] w̄i−1 kw̄i−1 k2
≤
N
X
2
qk2 (1 − rk )rk M λmax Hk,i−1
k=1
N
(116) X
4
= kE [Li |F i−1 ] kkw̄i−1 k4
(128)
where the inequality (a) holds because λmax (Σ) ≤ Tr(Σ) for
any symmetric positive semi-definite matrix Σ. We proceed to
deal with the term E [Li |F i−1 ]. Note that
Li = IM − L1,i + L2,i − L3,i + L4,i
(129)
≤
qk2 (1 − rk )rk M δd2
T
L1,i , 2D11,i + 2D11,i
2
(130)
2
T
T
T
L2,i , 3D11,i D11,i
+ D11,i
D11,i + (D11,i ) + D11,i
(131)
2
2
T
T
L3,i , (D11,i ) D11,i
+ D11,i D11,i
+
(132)
(133)
and we have E [L1,i |F i−1 ] = 4D̄11,i according to (112).
(136)
k=1
where in step (a) we used the property that kA + Bk ≤
kAk + kBk [33], and (b) holds because kXk ≤ Tr(X),
for any symmetric positive semi-definite matrix X, and that
Tr[diag{X}] = Tr[X]. Recall from (45) that [2, p. 509]
T
qk = µk (eT
k A2 p) , µmax τk (ek A2 p)
where
T
T
T
D11,i D11,i
D11,i + D11,i
D11,i D11,i
2
T
L4,i , D11,i D11,i
N
X
k=1
2
≤ λmax (E [Li |F i−1 ]) kw̄i−1 k
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1 +
k=1 j6=k=1
≤
2
≤ λmax (E [Ki |F i−1 ]) kw̄i−1 k
(a)
qk qj (1 − rk )(1 − rj )Hk,i−1 Hj,i−1
k=1 j=1
N
X
We next establish (36). Compared to the proof for Theorem
9.2 in [2], the main difference, apart from the second-order
moments evaluated in (124), is the term
1
T
E k(IM − D11,i
)w̄i−1 k4
(125)
3
(1 − t)
qk qj E [Γk,i Hk,i−1 Hj,i−1 Γj,i |F i−1 ] −
k=1 j=1
(137)
where ek denotes the k-th basis vector, which has a unit entry
at the k-th location and zeros elsewhere, and the parameter τk
satisfies µk = µmax τk . Then, we have
T
2
E D11,i
D11,i |F i−1 − D̄11,i
= O(µ2max )
(138)
14
Likewise, it follows that
h
i
2
2
E (D11,i ) |F i−1 − D̄11,i
h
i
2
2
T
E D11,i
|F i−1 − D̄11,i
Recall from (117) and (114) that
2
T
E D11,i D11,i
|F i−1 − D̄11,i
=
O(µ2max )
(139)
=
O(µ2max ).
(140)
=
O(µ2max ).
Substituting (138)–(141) into (131), we obtain
2
E [L2,i |F i−1 ] − 6 D̄11,i
= O(µ2max ).
(141)
(142)
Then, by following similar techniques shown in the proof of
Lemma 9.5 [2, pp. 542–546], we have
(I − F)−1 = [(p ⊗ p)(1 ⊗ 1)T ] ⊗ Z −1 + O(1)
(148)
where
Z,
N
X
qk (1 − rk ) [(Hk ⊗ IM ) + (IM ⊗ Hk )] .
(149)
k=1
The desired results (57) and (58) in Theorem 2 now follow
by referring to the proofs of Theorem 11.2 and Lemma 11.3
in [2, pp. 583–596], and Theorem 11.4 in [2, pp. 608–609].
Similarly, it can be verified that
3
E [L3,i |F i−1 ] − 4 D̄11,i
= O(µ3max )
4
E [L4,i |F i−1 ] − D̄11,i
= O(µ4max ).
(143)
(144)
It follows that
kE [Li |F i−1 ] k = kI − E [L1,i |F i−1 ] + E [L2,i |F i−1 ] −
E [L3,i |F i−1 ] + E [L4,i |F i−1 ] k
2
3
= I − 4D̄11,i + 6 D̄11,i − 4 D̄11,i +
4
2
D̄11,i + E [L2,i |F i−1 ] − 6 D̄11,i
3
− E [L3,i |F i−1 ] − 4 D̄11,i
4
+ E [L4,i |F i−1 ] − D̄11,i
2
3
≤ I − 4D̄11,i + 6 D̄11,i − 4 D̄11,i +
4
2
D̄11,i
+ E [L2,i |F i−1 ] − 6 D̄11,i
3
+ E [L3,i |F i−1 ] − 4 D̄11,i
4
+ E [L4,i |F i−1 ] − D̄11,i
4
= k I − D̄11,i k + O(µ2max )
Evaluating the squared Euclidean norms on both sides of
(38) and taking expectations conditioned on F i−1 , then taking
expectations again we get
n
o
0
ei0 k2bvec(I ) ] = E kw
ei−1
E[kw
k2F bvec(I ) +
NM
n NM
o
2
E ksi kE[(Γ MA )⊗ (Γ MA )]bvec(I )
(150)
i
4
≤ (1 − σ11 µmax ) + O(µ2max ).
00
≤ (1 − σ11
µmax ) Ekw̄i−1 k4
(146)
00
for some positive constant σ11
< σ11 , and for small enough
µmax , where in step (a) we set t , σ11 µmax . Then, the result
in (36) can be obtained by continuing from Eq. (9.117) (by
choosing t = σ11 µmax ) in the proof of Theorem 9.2 in [2, pp.
523–530].
T
F , E B0i ⊗b B0i
2
NM
n
E ksi k2E[(Γ
NM
o
i MA2 )⊗b (Γi MA2 )]F
n
bvec(IN M )
(151)
n=0
where the first-term corresponds to a transient component that
ek,i k2
dies out with time, and the convergence rate of Ekw
towards the steady-state regime is seen to be dictated by ρ (F)
[2, p. 592]. Now, let
Γ , EΓi = diag{(1 − rk )}N
k=1 ⊗ IM
(152)
M0 , MΓ = diag{µk (1 − rk )}N
k=1 ⊗ IM
(153)
0
T
B 0 , EB0i = AT
2 (I − M H) A1
(154)
0
We now rewrite (147) in terms of B as
(39)
T
T
T T
F = E AT
2 (I − MΓi H) A1 ⊗b A2 (I − MΓi H) A1
= (A1 ⊗b A1 ) I − I ⊗b (HΓM) − (HΓM) ⊗b I+
E [(HΓi M) ⊗b (HΓi M)] (A2 ⊗b A2 )
T
= [B 0 ⊗b B 0 ] + ∆F (µ2max )
(155)
where ∆F (µ2max ) is a matrix whose entries are in the order
of O(µ2max ). Following similar techniques to the proof of
Theorem 9.3 [2, pp. 535–540], we make the same Jordan
canonical decomposition for matrix P = A1 A2 as (108), then
substituting into (154) we get
T
B 0 = V−1 (J T − D0T )VT
(156)
where
A PPENDIX B
P ROOF OF T HEOREM 2
Define
i
NM
i
X
(145)
Substituting into (128), and taking expectations again on both
sides, we have
1
T
E k(IM − D11,i
)w̄i−1 k4
3
(1 − t)
1
4
2
≤
(1
−
σ
µ
)
+
O(µ
)
Ekw̄i−1 k4
11
max
max
(1 − t)3
(a)
= 1 − σ11 µmax + O(µ2max ) Ekw̄i−1 k4
b
where we used the weighted vector notation kxk2σ = kxk2Σ
with σ = bvec(Σ) and bvec(·) denoting the block vector
operation [2, p. 588]. Iterating the relation we get
n
o
0
ei0 k2bvec(I ) ] = E kw
e−1
E[kw
k2F i+1 bvec(I ) +
= kI − D̄11,i k4 + O(µ2max )
(117)
2
(147)
0
T
−1
D0T , VT AT
2 M HA1 V
0T
0T
D11 D21
=
0T
0T
D12 D22
T
(157)
15
and
0
D11
=
N
X
0
qk (1 − rk )Hk , D21
= O(µmax ).
(158)
k=1
0T
We now introduce the eigen-decomposition D11
, U ΛU T for
0T
the symmetric positive-definite matrix D11 , where U is unitary,
and Λ is a diagonal matrix composed of the eigenvalues of
0T
D11
. Let
2/N
(N −1)/N
T = diag{µ1/N
IM , µmax IM }
max U, µmax IM , . . . , µmax
(159)
then we have
0
0
T
B11 B12
T −1 VT B 0 V−1 T =
.
(160)
0
0
B21
B22
which means that the diagonal entries of F11 are the eigenval0
0
perturbed by a second-order term, O(µ2max ).
ues of B11
⊗ B11
Referring to Gershgorin’s Theorem, the union of the M 2
Gershgorin discs, centered at the diagonal entries of F11 with
(N +1)/N
radius O(µmax
), is disjoint from the union of the other
M 2 (N 2 −1) Gershgorin discs, centered at the diagonal entries
of F22 . Note that F has the same eigenvalues as the block
matrix on the right hand side of (163), and that eigenvalues
are invariant under a transposition operation. It follows from
(161) that
0
0
+1)/N
ρ (F) = ρ (B11
⊗ B11
) + O(µ(N
).
max
(166)
2
0
0
0
Using the fact that ρ (B11
⊗ B11
) = [ρ (B11
)] , we arrive at
the desired result (60).
(N +1)/N
0
0
It follows that B11
, IM − Λ, B12
= O(µmax
) [2,
0
p. 538]. The matrix B has the same eigenvalues as the
block matrix on the right hand side of (160). By referring to
Gershgorin’s Theorem [32], [33], it is shown in [2, pp. 539–
540] that the union of the M Gershgorin discs, each centered
(N +1)/N
0
at an eigenvalue of B11
), is disjoint
with radius O(µmax
from that of the other M (N − 1) Gershgorin discs, centered
0
at the diagonal entries of B22
, and therefore
0
+1)/N
ρ (B 0 ) = ρ (B11
) + O(µ(N
).
max
A PPENDIX C
E XAMINING THE D IFFERENCE IN (75)
We revisit the MSE networks discussed in (82). Assume
that there are only two agents in the network as shown in Fig.
2
2
7, namely, N = 2, with M = 2. Assume that σv,1
> σv,2
. For
1
(161)
Let
1
˜ F , T T V−1 ⊗b T T V−1 ∆F (µ2max )×
∆
T
T
V T −1
⊗b V T −1
1/N
o(µmax )
O(µ2max )
2/N
O(µ2max )
o(µmax )
.
=
..
o(µ2max )
.
1+1/N
O(µ2max )
o(µmax )
1
Fig. 7. A two-agent MSE network with a doubly-stochastic combination
matrix.
(162)
˜ F are
It follows from (162) that all the diagonal blocks of ∆
2
in the order of O(µmax ), the remaining block matrices in the
1/N
first row are in the order of o(µmax ), the remaining block
matrices in the first column are in the order of o(µ2max ), and
the upper and lower triangular blocks in the (2, 2)th block of
1+1/N
˜ F are in the order of o(µ2/N
∆
) respectively.
max ) and o(µmax
Then, substituting (160) and (162) into (155), we have
T
T
T T V−1 ⊗b T T V−1 F V T −1
⊗b V T −1
0
0
T
0
0
B11 B12
B11 B12
˜F
=
⊗b
+∆
0
0
0
0
B22
B22
B21
B21
T
F11 F12
=
(163)
F21 F22
where
0
0
+1)/N
F11 = B11
⊗ B11
+ O(µ2max ), F12 = O(µ(N
). (164)
max
Recall that
we have
0
B11
is a diagonal matrix, so is
0
B11
⊗
0
B11
,
0
0
diag{F11 } = diag{ λ (B11
⊗ B11
)} + O(µ2max )
2
then
(165)
simplicity, uniform parameters {qk ≡ q} are used across the
agents (which may occur, for example, in the ATC or CTA
forms when the step-sizes are uniform across the agents, i.e.,
{µk ≡ µ}, and doubly-stochastic combination matrices are
adopted. In this case, we get {qk ≡ µ/N } [2, pp. 493–494]).
Let
|π1 | π1
|π2 | π2
Ru,1 =
, Ru,2 =
(167)
π1
1
π2
1
where the numbers |π1 | < 1, |π2 | < 1 ensure that Ru,1 >
0, Ru,2 > 0. Then, expression (75) can be rewritten as
MSDcoor,k − MSDgrad,k
(85)
−1
2
= rqTr (Ru,1 + Ru,2 )
σv,1
(diag{Ru,1 } − Ru,1 ) +
2
σv,2
(diag{Ru,2 } − Ru,2 )
−1
(167)
|π1 | + |π2 | π1 + π2
= rqTr
×
π1 + π2
2
!
2
2
0
−π1 σv,1
− π2 σv,2
2
2
−π1 σv,1
− π2 σv,2
0
=
2
2
2rq(π1 + π2 )(π1 σv,1
+ π2 σv,2
)
2
2 (|π1 | + |π2 |) − (π1 + π2 )
(168)
where the denominator is positive for all |π1 | < 1, |π2 | <
1. Then, MSDcoor,k < MSDgrad,k if, and only if (π1 +
16
2
2
π2 )(π1 σv,1
+ π2 σv,2
) < 0, which implies that
(
2
2
0 < π2 < 1, −π2 < π1 < − σv,2
/σv,1
π2
2
2
−1 < π2 < 0, − σv,2 /σv,1 π2 < π1 < −π2 .
A PPENDIX E
P ROOF OF C OROLLARY 5
(169a)
(169b)
We start from the MSD expression in (95) and note first
that
rk
G0k = (1 − rk )2 G +
diag{G}
(174)
1 − rk
Otherwise, MSDcoor,k ≥ MSDgrad,k .
Substituting into (95) we have:
A PPENDIX D
P ROOF OF C OROLLARY 2
MSDcoor,k
1
=
2
k=1
k=1
=
1
2
(170)
where we substituted (76) into (75) and used the relation
Tr(Gk ) = Tr (diag{Gk }). Then, noting that
!
N
N
N
X
X
X
0<
qk λmin (Hk ) ≤ λ
qk Hk ≤
qk λmax (Hk )
k=1
k=1
k=1
(172)
we have
N
.
X
1
δd
qk
!
(a)
≤ 1
k=1
N
.X
λ
N
X
qk λmax (Hk ) ≤
!−1
≤
k=1
N
N
.X
X
(b) .
1
qk λmin (Hk ) ≤ 1
νd
qk
k=1
− rk )
×
k=1
−1
1
G +
2
N
X
!−1
qk (1 − rk )
×
k=1
!
qk2 (1 − rk )rk
Tr H −1 diag{G}
k=1
1
=
2
N
X
!−1
qk (1 − rk )
k=1
N
X
!
qk2 (1 − rk )2
×
k=1
1
Tr H −1 G + (θ − α)Tr H −1 diag{G}
2
(175)
where (175) holds because
PN
PN
2
2
qk2 (1 − rk )
k=1 qk (1 − rk )
θ − α = Pk=1
−
PN
N
k=1 qk (1 − rk )
k=1 qk (1 − rk )
PN
qk2 (1 − rk )rk
= Pk=1
(176)
N
k=1 qk (1 − rk )
with the numbers α and θ being defined in (97) and (98),
respectively. Recall that
!−1 N
N
X
1 X
MSDgrad,k =
qk
qk2 Tr H −1 G
(177)
2
k=1
qk Hk
2
k=1
is the MSD performance for the full-gradient case. Thus,
k=1
!
qk2 (1
k=1
k=1
N
X
k=1
k=1
N
X
rk
diag{G}
1 − rk
!−1 N
!
N
X
X
2
2
qk (1 − rk )
qk (1 − rk ) ×
Tr H
Then, applying the inequality [34]:
for any symmetric positive semi-definite matrices A and B,
where λmin (A) and λmax (A) represent respectively the largest
and smallest eigenvalues of A, we obtain
(
!−1
N
X
r
−
MSDcoor,k −MSDgrad,k ≤
λmax
qk Hk
2
k=1
!−1 ) N
N
X
X
qk2 Tr(Gk ) (171)
λmin
qk Hk
qk (1 − rk )
k=1
k=1
λmin (A)Tr(B) ≤ Tr(AB) ≤ λmax (A)Tr(B)
!−1
−1
Tr H
G+
In the case when the matrices {Hk } or {Gk } are diagonal, it
follows from (75) and (76) that MSDcoor,k − MSDgrad,k = 0,
which verifies (80).
More generally, according to (2) and (46), we have
!−1
N
N
N
X
X
X
q k Hk
> 0,
qk2 Gk ≥ 0,
qk2 diag{Gk } ≥ 0
N
X
!
(173)
k=1
where the inequalities (a) and (b) hold due to (2). Substituting
(173) into (171) gives the upper bound for the difference
MSDcoor,k −MSDgrad,k as shown by (79). Then, by following
a similar argument, we obtain the lower bound for the difference as the opposite number of the upper bound, which leads
to the desired result in Corollary 2.
MSDcoor,k − MSDgrad,k
!
PN
PN
2
2
2
1
k=1 qk (1 − rk )
k=1 qk
=
− PN
Tr H −1 G
PN
2
k=1 qk (1 − rk )
k=1 qk
1
+ (θ − α)Tr H −1 diag{G}
2
1
α
= Tr H −1 G + (θ − α)Tr H −1 diag{G} . (178)
2
2
Applying (170) and (2) to (178), we obtain the desired results
for the MSD performance shown in Corollary 5. The result
for the ER performance in Corollary 5 can be shown by
subtracting the ER expression, ERgrad,k , on the both sides
of (96).
17
R EFERENCES
[1] C. Wang, Y. Zhang, B. Ying, and A. H. Sayed, “Coordinate-descent
adaptation over networks,” in Proc. EUSIPCO, Kos Island, Greece,
Aug. 2017.
[2] A. H. Sayed, “Adaptation, learning, and optimization over networks,”
Foundations and Trends in Machine Learning, vol. 7, no. 4-5, pp. 311–
801, 2014.
[3] A. H. Sayed, “Adaptive networks,” Proceedings of the IEEE, vol. 102,
no. 4, pp. 460–497, Apr. 2014.
[4] S. Kar, J. M. F. Moura, and K. Ramanan, “Distributed parameter estimation in sensor networks: Nonlinear observation models and imperfect
communication,” IEEE Trans. Information Theory, vol. 58, no. 6, pp.
3575–3605, June 2012.
[5] A. Nedić and A. Ozdaglar, “Cooperative distributed multi-agent optimization,” in Convex Optimization in Signal Processing and Communications, D. P. Palomar and Y. C. Eldar, Eds., pp. 340–386. Cambridge
University Press, 2010.
[6] S. Kar and J. M. F. Moura, “Convergence rate analysis of distributed
gossip (linear parameter) estimation: Fundamental limits and tradeoffs,”
IEEE J. Sel. Top. Signal Process., vol. 5, no. 4, pp. 674–690, Aug. 2011.
[7] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione,
“Gossip algorithms for distributed signal processing,” Proc. IEEE, vol.
98, no. 11, pp. 1847–1864, Nov. 2010.
[8] S. Sardellitti, M. Giona, and S. Barbarossa, “Fast distributed average
consensus algorithms based on advection-diffusion processes,” IEEE
Trans. Signal Process., vol. 58, no. 2, pp. 826–842, Feb. 2010.
[9] P. Braca, S. Marano, and V. Matta, “Running consensus in wireless
sensor networks,” in Proc. 11th International Conference on Information
Fusion, Cologne, Germany, June 2008, pp. 1–6.
[10] U. A. Khan and J. M. F. Moura, “Distributing the Kalman filter for
large-scale systems,” IEEE Trans. Signal Process., vol. 56, no. 10, pp.
4919–4935, Oct. 2008.
[11] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,”
Syst. Control Lett., vol. 53, no. 1, pp. 65–78, Sep. 2004.
[12] S.-Y. Tu and A. H. Sayed, “Diffusion strategies outperform consensus
strategies for distributed estimation over adaptive networks,” IEEE
Trans. Signal Process., vol. 60, no. 12, pp. 6217–6234, Dec. 2012.
[13] Y. Nesterov, “Efficiency of coordinate descent methods on huge-scale
optimization problems,” SIAM Journal on Optimization, vol. 22, no. 2,
pp. 341–362, 2012.
[14] P. Richtárik and M. Takáč, “Iteration complexity of randomized blockcoordinate descent methods for minimizing a composite function,”
Mathematical Programming, vol. 144, no. 1-2, pp. 1–38, 2014.
[15] Z. Lu and L. Xiao, “On the complexity analysis of randomized blockcoordinate descent methods,” Mathematical Programming, vol. 152, pp.
615–642, 2015.
[16] X. Zhao and A. H. Sayed, “Asynchronous adaptation and learning over
networks-Part I: Modeling and stability analysis,” IEEE Trans. Signal
Process., vol. 63, no. 4, pp. 811–826, Feb. 2015.
[17] X. Zhao and A. H. Sayed, “Asynchronous adaptation and learning over
networks-Part II: Performance analysis,” IEEE Trans. Signal Process.,
vol. 63, no. 4, pp. 827–842, Feb. 2015.
[18] X. Zhao and A. H. Sayed, “Asynchronous adaptation and learning over
networks-Part III: Comparison analysis,” IEEE Trans. Signal Process.,
vol. 63, no. 4, pp. 843–858, Feb. 2015.
[19] R. Arablouei, S. Werner, Y.-F. Huang, and K. Dogancay, “Distributed
least mean-square estimation with partial diffusion,” IEEE Trans. Signal
Process., vol. 62, no. 2, pp. 472–484, Jan. 2014.
[20] M. R. Gholami, E. G. Ström, and A. H. Sayed, “Diffusion estimation
over cooperative networks with missing data,” in Proc. IEEE GlobalSIP,
Austin, TX, Dec. 2013, pp. 411–414.
[21] S. C. Douglas, “Adaptive filters employing partial updates,” IEEE Trans.
Circuits Syst. II, vol. 44, no. 3, pp. 209–216, Mar. 1997.
[22] S. Werner, M. Mohammed, Y.-F. Huang, and V. Koivunen, “Decentralized set-membership adaptive estimation for clustered sensor networks,”
in Proc. IEEE ICASSP, Las Vegas, NV, 2008, pp. 3573–3576.
[23] S. Werner and Y.-F. Huang, “Time- and coefficient- selective diffusion
strategies for distributed parameter estimation,” in Proc. Asilomar
Conference on Signals, Systems, and Computers, Pacific Grove, CA,
2010, pp. 696–700.
[24] S. C. Douglas, “A family of normalized LMS algorithms,” IEEE Signal
Process. Lett., vol. 1, no. 3, pp. 49–51, Mar. 1994.
[25] K. Doğancay, O. Tanrıkulu, “Adaptive filtering algorithms with selective
partial updates,” IEEE Trans. Circuits Syst. II, vol. 48, no. 8, pp. 762–
769, Aug. 2001.
[26] M. Godavarti and A. O. Hero, “Partial update LMS algorithms,” IEEE
Trans. Signal Process., vol. 53, no. 7, pp. 2382–2399, Jul. 2005.
[27] S. Chouvardas, K. Slavakis, and S. Theodoridis, “Trading off complexity
with communication costs in distributed adaptive learning via Krylov
subspaces for dimensionality reduction,” IEEE Journal Selected Topics
Signal Process., vol. 7, no. 2, pp. 257–273, April 2013.
[28] S. Theodoridis, K. Slavakis, and I. Yamada, “Adaptive learning in a
world of projections: A unifying framework for linear and nonlinear
classification and regression tasks,” IEEE Signal Process. Magazine,
vol. 28, no. 1, pp. 97–123, Jan. 2011.
[29] S. Chouvardas, K. Slavakis, Y. Kopsinis, and S. Theodoridis, “A sparsity
promoting adaptive algorithm for distributed learning,” IEEE Trans.
Signal Process., vol. 60, no. 10, pp. 5412–5425, Oct. 2012.
[30] O. N. Gharehshiran, V. Krishnamurthy, and G. Yin, “Distributed energyaware diffusion least mean squares: Game-theoretic learning,” IEEE J.
Sel. Top. Signal Process., vol. 7, no. 5, pp. 821–836, Oct. 2013.
[31] A. H. Sayed, Adaptive Filters, Wiley, 2008.
[32] G. H. Golub and C. F. Van Loan, Matrix Computations, The John
Hopkins University Press, Baltimore, 3 edition, 1996.
[33] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University
Press, 2003.
[34] Y. Fang, K. A. Loparo, and X. Feng, “Inequalities for the trace of
matrix product,” IEEE Trans. Automatic Control, vol. 39, no. 12, pp.
2489–2490, 1994.
Chengcheng Wang (S’17) received the B.Eng. degree in Electrical Engineering and Automation, and
Ph.D degree in Control Science and Engineering,
from the College of Automation, Harbin Engineering University, Harbin, China, in 2011 and 2017,
respectively. She is currently a Research Fellow in
the School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore. From
Sep. 2014 to Sep. 2016, she was a visiting graduate
researcher in Adaptive Systems Laboratory at the
University of California, Los Angeles, California,
USA. Her research interests include adaptive and statistical signal processing,
and distributed adaptation and learning.
Yonggang Zhang (S’06–M’07–SM’16) received the
B.S. and M.S. degrees from the College of Automation, Harbin Engineering University, Harbin,
China, in 2002 and 2004, respectively. He received
his Ph.D. degree in Electronic Engineering from
Cardiff University, UK in 2007 and worked as a
Post-Doctoral Fellow at Loughborough University,
UK from 2007 to 2008 in the area of adaptive signal
processing. Currently, he is a Professor of Control Science and Engineering in Harbin Engineering
University (HEU) in China. His current research
interests include signal processing, information fusion and their applications
in navigation technology, such as fiber optical gyroscope, inertial navigation
and integrated navigation.
Bicheng Ying (S’15) received his B.S. and M.S.
degrees from Shanghai Jiao Tong University (SJTU)
and University of California, Los Angeles (UCLA)
in 2013 and 2014, respectively. He is currently working towards the PhD degree in Electrical Engineering
at UCLA. His research interests include multi-agent
network processing, large-scale machine learning,
distributed optimization, and statistical signal processing.
18
Ali H. Sayed (S’90–M’92–SM’99–F’01) is Dean
of Engineering at EPFL, Switzerland. He has also
served as distinguished professor and former chairman of electrical engineering at UCLA. An author of over 500 scholarly publications and six
books, his research involves several areas including
adaptation and learning, statistical signal processing,
distributed processing, network and data sciences,
and biologically-inspired designs. Dr. Sayed has received several awards including the 2015 Education
Award from the IEEE Signal Processing Society, the
2014 Athanasios Papoulis Award from the European Association for Signal
Processing, the 2013 Meritorious Service Award, and the 2012 Technical
Achievement Award from the IEEE Signal Processing Society. Also, the 2005
Terman Award from the American Society for Engineering Education, the
2003 Kuwait Prize, and the 1996 IEEE Donald G. Fink Prize. He served as
Distinguished Lecturer for the IEEE Signal Processing Society in 2005 and as
Editor-in-Chief of the IEEE TRANSACTIONS ON SIGNAL PROCESSING
(2003–2005). His articles received several Best Paper Awards from the IEEE
Signal Processing Society (2002, 2005, 2012, 2014). He is a Fellow of
the American Association for the Advancement of Science (AAAS). He is
recognized as a Highly Cited Researcher by Thomson Reuters. He is serving
as President-Elect of the IEEE Signal Processing Society.
| 3cs.SY
|
Overcoming Exploration in Reinforcement Learning
with Demonstrations
arXiv:1709.10089v2 [cs.LG] 25 Feb 2018
Ashvin Nair12 , Bob McGrew1 , Marcin Andrychowicz1 , Wojciech Zaremba1 , Pieter Abbeel12
Abstract— Exploration in environments with sparse rewards
has been a persistent problem in reinforcement learning (RL).
Many tasks are natural to specify with a sparse reward, and
manually shaping a reward function can result in suboptimal
performance. However, finding a non-zero reward is exponentially more difficult with increasing task horizon or action
dimensionality. This puts many real-world tasks out of practical
reach of RL methods. In this work, we use demonstrations
to overcome the exploration problem and successfully learn to
perform long-horizon, multi-step robotics tasks with continuous
control such as stacking blocks with a robot arm. Our method,
which builds on top of Deep Deterministic Policy Gradients and
Hindsight Experience Replay, provides an order of magnitude
of speedup over RL on simulated robotics tasks. It is simple
to implement and makes only the additional assumption that
we can collect a small set of demonstrations. Furthermore, our
method is able to solve tasks not solvable by either RL or
behavior cloning alone, and often ends up outperforming the
demonstrator policy.
I. I NTRODUCTION
RL has found significant success in decision making for
solving games, so what makes it more challenging to apply
in robotics? A key difference is the difficulty of exploration,
which comes from the choice of reward function and complicated environment dynamics. In games, the reward function
is usually given and can be directly optimized. In robotics,
we often desire behavior to achieve some binary objective
(e.g., move an object to a desired location or achieve a certain
state of the system) which naturally induces a sparse reward.
Sparse reward functions are easier to specify and recent work
suggests that learning with a sparse reward results in learned
policies that perform the desired objective instead of getting
stuck in local optima [1], [2]. However, exploration in an
environment with sparse reward is difficult since with random
exploration, the agent rarely sees a reward signal.
The difficulty posed by a sparse reward is exacerbated
by the complicated environment dynamics in robotics. For
example, system dynamics around contacts are difficult to
model and induce a sensitivity in the system to small errors.
Many robotics tasks also require executing multiple steps
successfully over a long horizon, involve high dimensional
control, and require generalization to varying task instances.
These conditions further result in a situation where the agent
so rarely sees a reward initially that it is not able to learn at
all.
All of the above means that random exploration is not a
tenable solution. Instead, in this work we show that we can
use demonstrations as a guide for our exploration. To test our
1
OpenAI,
2
University of California, Berkeley.
method, we solve the problem of stacking several blocks at
a given location from a random initial state. Stacking blocks
has been studied before in the literature [3], [4] and exhibits
many of the difficulties mentioned: long horizons, contacts,
and requires generalizing to each instance of the task. We
limit ourselves to 100 human demonstrations collected via
teleoperation in virtual reality. Using these demonstrations,
we are able to solve a complex robotics task in simulation
that is beyond the capability of both reinforcement learning
and imitation learning.
The primary contribution of this paper is to show that
demonstrations can be used with reinforcement learning
to solve complex tasks where exploration is difficult. We
introduce a simple auxiliary objective on demonstrations, a
method of annealing away the effect of the demonstrations
when the learned policy is better than the demonstrations,
and a method of resetting from demonstration states that
significantly improves and speeds up training policies. By
effectively incorporating demonstrations into RL, we shortcircuit the random exploration phase of RL and reach
nonzero rewards and a reasonable policy early on in training.
Finally, we extensively evaluate our method against other
commonly used methods, such as initialization with learning
from demonstrations and fine-tuning with RL, and show that
our method significantly outperforms them.
II. R ELATED W ORK
Learning methods for decision making problems such as
robotics largely divide into two classes: imitation learning
and reinforcement learning (RL). In imitation learning (also
called learning from demonstrations) the agent receives behavior examples from an expert and attempts to solve a task
by copying the expert’s behavior. In RL, an agent attempts
to maximize expected reward through interaction with the
environment. Our work combines aspects of both to solve
complex tasks.
Imitation Learning: Perhaps the most common form of
imitation learning is behavior cloning (BC), which learns a
policy through supervised learning on demonstration stateaction pairs. BC has seen success in autonomous driving
[5], [6], quadcopter navigation [7], locomotion [8], [9]. BC
struggles outside the manifold of demonstration data. Dataset
Aggregation (DAGGER) augments the dataset by interleaving
the learned and expert policy to address this problem of
accumulating errors [10]. However, DAGGER is difficult to
use in practice as it requires access to an expert during all
of training, instead of just a set of demonstrations.
Fig. 1: We present a method using reinforcement learning to solve the task of block stacking shown above. The robot starts
with 6 blocks labelled A through F on a table in random positions and a target position for each block. The task is to move each
block to its target position. The targets are marked in the above visualization with red spheres which do not interact with the
environment. These targets are placed in order on top of block A so that the robot forms a tower of blocks. This is a complex,
multi-step task where the agent needs to learn to successfully manage multiple contacts to succeed. Frames from rollouts of
the learned policy are shown. A video of our experiments can be found at: http://ashvin.me/demoddpg-website
Fundamentally, BC approaches are limited because they
do not take into account the task or environment. Inverse
reinforcement learning (IRL) [11] is another form of imitation learning where a reward function is inferred from the
demonstrations. Among other tasks, IRL has been applied
to navigation [12], autonomous helicopter flight [13], and
manipulation [14]. Since our work assumes knowledge of a
reward function, we omit comparisons to IRL approaches.
Reinforcement Learning: Reinforcement learning methods have been harder to apply in robotics, but are heavily
investigated because of the autonomy they could enable.
Through RL, robots have learned to play table tennis [15],
swing up a cartpole, and balance a unicycle [16]. A renewal
of interest in RL cascaded from success in games [17], [18],
especially because of the ability of RL with large function
approximators (ie. deep RL) to learn control from raw pixels.
Robotics has been more challenging in general but there
has been significant progress. Deep RL has been applied to
manipulation tasks [19], grasping [20], [21], opening a door
[22], and locomotion [23], [24], [25]. However, results have
been attained predominantly in simulation per high sample
complexity, typically caused by exploration challenges.
Robotic Block Stacking: Block stacking has been studied
from the early days of AI and robotics as a task that
encapsulates many difficulties of more complicated tasks we
want to solve, including multi-step planning and complex
contacts. SHRDLU [26] was one of the pioneering works,
but studied block arrangements only in terms of logic and
natural language understanding. More recent work on task
and motion planning considers both logical and physical
aspects of the task [27], [28], [29], but requires domainspecific engineering. In this work we study how an agent
can learn this task without the need of domain-specific
engineering.
One RL method, PILCO [16] has been applied to a simple
version of stacking blocks where the task is to place a
block on a tower [3]. Methods such as PILCO based on
learning forward models naturally have trouble modelling
the sharply discontinuous dynamics of contacts; although
they can learn to place a block, it is a much harder problem
to grasp the block in the first place. One-shot Imitation [4]
learns to stack blocks in a way that generalizes to new target
configurations, but uses more than 100,000 demonstrations
to train the system. A heavily shaped reward can be used
to learn to stack a Lego block on another with RL [30]. In
contrast, our method can succeed from fully sparse rewards
and handle stacking several blocks.
Combining RL and Imitation Learning: Previous work
has combined reinforcement learning with demonstrations.
Demonstrations have been used to accelerate learning on
classical tasks such as cart-pole swing-up and balance [31].
This work initialized policies and (in model-based methods)
initialized forward models with demonstrations. Initializing
policies from demonstrations for RL has been used for
learning to hit a baseball [32] and for underactuated swingup [33]. Beyond initialization, we show how to extract more
knowledge from demonstrations by using them effectively
throughout the entire training process.
Our method is closest to two recent approaches —
Deep Q-Learning From Demonstrations (DQfD) [34] and
DDPG From Demonstrations (DDPGfD) [2] which combine
demonstrations with reinforcement learning. DQfD improves
learning speed on Atari, including a margin loss which
encourages the expert actions to have higher Q-values than
all other actions. This loss can make improving upon the
demonstrator policy impossible which is not the case for
our method. Prior work has previously explored improving
beyond the demonstrator policy in simple environments by
introducing slack variables [35], but our method uses a
learned value to actively inform the improvement. DDPGfD
solves simple robotics tasks akin to peg insertion using
DDPG with demonstrations in the replay buffer. In contrast
to this prior work, the tasks we consider exhibit additional
difficulties that are of key interest in robotics: multi-step
behaviours, and generalization to varying goal states. While
previous work focuses on speeding up already solvable tasks,
we show that we can extend the state of the art in RL with
demonstrations by introducing new methods to incorporate
demonstrations.
III. BACKGROUND
A. Reinforcement Learning
We consider the standard Markov Decision Process framework for picking optimal actions to maximize rewards over
discrete timesteps in an environment E. We assume that the
environment is fully observable. At every timestep t, an agent
is in a state xt , takes an action at , receives a reward rt ,
and E evolves to state xt+1 . In reinforcement learning, the
agent must learn a policy at = π(xt ) toPmaximize expected
T
returns. We denote the return by Rt = i=t γ (i−t) ri where
T is the horizon that the agent optimizes over and γ is
a discount factor for future rewards. The agent’s objective
is to maximize expected return from the start distribution
J = Eri ,si ∼E,ai ∼π [R0 ].
A variety of reinforcement learning algorithms have been
developed to solve this problem. Many involve constructing
an estimate of the expected return from a given state after
taking an action:
Qπ (st , at ) = Eri ,si ∼E,ai ∼π [Rt |st , at ]
= Ert ,st+1 ∼E [rt + γ Eat+1 ∼π [Qπ (st+1 , at+1 )]]
(1)
(2)
We call Qπ the action-value function. Equation 2 is a
recursive version of equation 1, and is known as the Bellman equation. The Bellman equation allows for methods to
estimate Q that resemble dynamic programming.
B. DDPG
Our method combines demonstrations with one such
method: Deep Deterministic Policy Gradients (DDPG) [23].
DDPG is an off-policy model-free reinforcement learning
algorithm for continuous control which can utilize large
function approximators such as neural networks. DDPG
is an actor-critic method, which bridges the gap between
policy gradient methods and value approximation methods
for RL. At a high level, DDPG learns an action-value
function (critic) by minimizing the Bellman error, while
simultaneously learning a policy (actor) by directly maximizing the estimated action-value function with respect to
the parameters of the policy.
Concretely, DDPG maintains an actor function π(s) with
parameters θπ , a critic function Q(s, a) with parameters θQ ,
and a replay buffer R as a set of tuples (st , at , rt , st+1 )
for each transition experienced. DDPG alternates between
running the policy to collect experience and updating the
parameters. Training rollouts are collected with extra noise
for exploration: at = π(s) + N , where N is a noise process.
During each training step, DDPG samples a minibatch
consisting of N tuples from R to update the actor and critic
networks. DDPG minimizes the following loss L w.r.t. θQ
to update the critic:
yi = ri + γQ(si+1 , π(si+1 ))
(3)
L=
1 X
(yi − Q(si , ai |θQ ))2
N i
(4)
The actor parameters θπ are updated using the policy
gradient:
1 X
∇a Q(s, a|θQ )|s=si ,a=π(s) ∇θπ π(s|θπ )|si
∇θ π J =
N i
(5)
To stabilize learning, the Q value in equation 3 is usually
computed using a separate network (called the target network) whose weights are an exponential average over time
of the critic network. This results in smoother target values.
Note that DDPG is a natural fit for using demonstrations. Since DDPG can be trained off-policy, we can use
demonstration data as off-policy training data. We also take
advantage of the action-value function Q(s, a) learned by
DDPG to better use demonstrations.
C. Multi-Goal RL
Instead of the standard RL setting, we train agents with
parametrized goals, which lead to more general policies
[36] and have recently been shown to make learning with
sparse rewards easier [1]. Goals describe the task we expect
the agent to perform in the given episode, in our case
they specify the desired positions of all objects. We sample
the goal g at he beginning of every episode. The function
approximators, here π and Q, take the current goal as an
additional input.
D. Hindsight Experience Replay (HER)
To handle varying task instances and parametrized goals,
we use Hindsight Experience Replay (HER) [1]. The key
insight of HER is that even in failed rollouts where no
reward was obtained, the agent can transform them into
successful ones by assuming that a state it saw in the rollout
was the actual goal. HER can be used with any off-policy
RL algorithm assuming that for every state we can find a
goal corresponding to this state (i.e. a goal which leads to a
positive reward in this state).
For every episode the agent experiences, we store it in
the replay buffer twice: once with the original goal pursued
in the episode and once with the goal corresponding to the
final state achieved in the episode, as if the agent intended
on reaching this state from the very beginning.
IV. M ETHOD
Our method combines DDPG and demonstrations in several ways to maximally use demonstrations to improve
learning. We describe our method below and evaluate these
ideas in our experiments.
A. Demonstration Buffer
First, we maintain a second replay buffer RD where we
store our demonstration data in the same format as R. In
each minibatch, we draw an extra ND examples from RD
to use as off-policy replay data for the update step. These
examples are included in both the actor and critic update.
This idea has been introduced in [2].
B. Behavior Cloning Loss
Second, we introduce a new loss computed only on the
demonstration examples for training the actor.
LBC =
ND
X
2
kπ(si |θπ ) − ai k
(6)
i=1
This loss is a standard loss in imitation learning, but we show
that using it as an auxiliary loss for RL improves learning
significantly. The gradient applied to the actor parameters θπ
is:
λ1 ∇θπ J − λ2 ∇θπ LBC
(7)
(Note that we maximize J and minimize LBC .) Using this
loss directly prevents the learned policy from improving
significantly beyond the demonstration policy, as the actor is
always tied back to the demonstrations. Next, we show how
to account for suboptimal demonstrations using the learned
action-value function.
C. Q-Filter
We account for the possibility that demonstrations can be
suboptimal by applying the behavior cloning loss only to
states where the critic Q(s, a) determines that the demonstrator action is better than the actor action:
LBC =
ND
X
2
kπ(si |θπ ) − ai k 1Q(si ,ai )>Q(si ,π(si ))
(8)
i=1
The gradient applied to the actor parameters is as in equation
7. We label this method using the behavior cloning loss and
Q-filter “Ours” in the following experiments.
D. Resets to demonstration states
To overcome the problem of sparse rewards in very long
horizon tasks, we reset some training episodes using states
and goals from demonstration episodes. Restarts from within
demonstrations expose the agent to higher reward states during training. This method makes the additional assumption
that we can restart episodes from a given state, as is true in
simulation.
To reset to a demonstration state, we first sample a
demonstration D = (x0 , u0 , x1 , u1 , ...xN , uN ) from the set
of demonstrations. We then uniformly sample a state xi
from D. As in HER, we use the final state achieved in the
demonstration as the goal. We roll out the trajectory with the
given initial state and goal for the usual number of timesteps.
At evaluation time, we do not use this procedure.
We label our method with the behavior cloning loss, Qfilter, and resets from demonstration states as “Ours, Resets”
in the following experiments.
V. E XPERIMENTAL S ETUP
A. Environments
We evaluate our method on several simulated MuJoCo [37]
environments. In all experiments, we use a simulated 7-DOF
Fetch Robotics arm with parallel grippers to manipulate one
or more objects placed on a table in front of the robot.
The agent receives the positions of the relevant objects
on the table as its observations. The control for the agent is
continuous and 4-dimensional: 3 dimensions that specify the
desired end-effector position1 and 1 dimension that specifies
the desired distance between the robot fingers. The agent is
controlled at 50Hz frequency.
We collect demonstrations in a virtual reality environment.
The demonstrator sees a rendering of the same observations
as the agent, and records actions through a HTC Vive
interface at the same frequency as the agent. We have
the option to accept or reject a demonstration; we only
accept demonstrations we judge to be mostly correct. The
demonstrations are not optimal. The most extreme example
is the “sliding” task, where only 7 of the 100 demonstrations
are successful, but the agent still sees rewards for these
demonstrations with HER.
B. Training Details
To train our models, we use Adam [38] as the optimizer
with learning rate 10−3 . We use N = 1024, ND = 128, λ1 =
10−3 , λ2 = 1.0/ND . The discount factor γ is 0.98. We use
100 demonstrations to initialize RD . The function approximators π and Q are deep neural networks with ReLU activations and L2 regularization with the coefficient 5×10−3 . The
final activation function for π is tanh, and the output value
is scaled to the range of each action dimension. To explore
during training, we sample random actions uniformly within
the action space with probability 0.1 at every step, and the
noise process N is uniform over ±10% of the maximum
value of each action dimension. Task-specific information,
including network architectures, are provided in the next
section.
C. Overview of Experiments
We perform three sets of experiments. In Sec. VI, we
provide a comparison to previous work. In Sec. VII we
solve block stacking, a difficult multi-step task with complex
contacts that the baselines struggle to solve. In Sec. VIII
we do ablations of our own method to show the effect of
individual components.
VI. C OMPARISON W ITH P RIOR W ORK
A. Tasks
We first show the results of our method on the simulated
tasks presented in the Hindsight Experience Replay paper
[1]. We apply our method to three tasks:
1) Pushing. A block placed randomly on the table must
be moved to a target location on the table by the robot
(fingers are blocked to avoid grasping).
2) Sliding. A puck placed randomly on the table must be
moved to a given target location. The target is outside
the robot’s reach so it must apply enough force that
the puck reaches the target and stops due to friction.
3) Pick-and-place. A block placed randomly on the table
must be moved to a target location in the air. Note
1 In
the 10cm x 10cm x 10cm cube around the current gripper position
Pushing
1.0
0.4
0.2
0.8
Success Rate
0.6
0.6
0.4
0.2
0.0
2M
4M
6M
Timesteps
8M
10M
0.6
0.4
0.2
0.0
0M
Pick and Place
1.0
Ours
HER
BC
0.8
Success Rate
0.8
Success Rate
Sliding
1.0
0.0
0M
2M
4M
6M
8M
10M
0M
Timesteps
2M
4M
6M
8M
10M
Timesteps
Fig. 2: Baseline comparisons on tasks from [1]. Frames from the learned policy are shown above each task. Our method
significantly outperforms the baselines. On the right plot, the HER baseline always fails.
that the original paper used a form of initializing from
favorable states to solve this task. We omit this for our
experiment but discuss and evaluate the initialization
idea in an ablation.
As in the prior work, we use a fully sparse reward for this
task. The agent is penalized if the object is not at its goal
position:
(
0,
if ||xi − gi || < δ
rt =
(9)
−1, otherwise
where the threshold δ is 5cm.
B. Results
Fig. 2 compares our method to HER without demonstrations and behavior cloning. Our method is significantly faster
at learning these tasks than HER, and achieves significantly
better policies than behavior cloning does. Measuring the
number of timesteps to get to convergence, we exhibit a 4x
speedup over HER in pushing, a 2x speedup over HER in
sliding, and our method solves the pick-and-place task while
HER baseline cannot solve it at all.
The pick-and-place task showcases the shortcoming of RL
in sparse reward settings, even with HER. In pick-and-place,
the key action is to grasp the block. If the robot could manage
to grasp it a small fraction of the time, HER discovers
how to achieve goals in the air and reinforces the grasping
behavior. However, grasping the block with random actions
is extremely unlikely. Our method pushes the policy towards
demonstration actions, which are more likely to succeed.
In the HER paper, HER solves the pick-and-place task
by initializing half of the rollouts with the gripper grasping
the block. With this addition, pick-and-place becomes the
easiest of the three tasks tested. This initialization is similar
in spirit to our initialization idea, but takes advantage of the
fact that pick-and-place with any goal can be solved starting
from a block grasped at a certain location. This is not always
true (for example, if there are multiple objects to be moved)
and finding such a keyframe for other tasks would be difficult, requiring some engineering and sacrificing autonomy.
Instead, our method guides the exploration towards grasping
the block through demonstrations. Providing demonstrations
does not require expert knowledge of the learning system,
which makes it a more compelling way to provide prior
information.
VII. M ULTI -S TEP E XPERIMENTS
A. Block Stacking Task
To show that our method can solve more complex tasks
with longer horizon and sparser reward, we study the task
of block stacking in a simulated environment as shown in
Fig. 1 with the same physical properties as the previous
experiments. Our experiments show that our approach can
solve the task in full and learn a policy to stack 6 blocks
with demonstrations and RL. To measure and communicate
various properties of our method, we also show experiments
on stacking fewer blocks, a subset of the full task.
We initialize the task with blocks at 6 random locations
x1 ...x6 . We also provide 6 goal locations g1 ...g6 . To form a
tower of blocks, we let g1 = x1 and gi = gi−1 + (0, 0, 5cm)
for i ∈ 2, 3, 4, 5.
By stacking N blocks, we mean N blocks reach their
target locations. Since the target locations are always on top
of x1 , we start with the first block already in position. So
stacking N blocks involves N −1 pick-and-place actions. To
solve stacking N , we allow the agent 50 ∗ (N − 1) timesteps.
This means that to stack 6 blocks, the robot executes 250
actions or 5 seconds.
We recorded 100 demonstrations to stack 6 blocks, and
use subsets of these demonstrations as demonstrations for
stacking fewer blocks. The demonstrations are not perfect;
they include occasionally dropping blocks, but our method
can handle suboptimal demonstrations. We still rejected more
than half the demonstrations and excluded them from the
demonstration data because we knocked down the tower of
blocks when releasing a block.
B. Rewards
Two different reward functions are used. To test the
performance of our method under fully sparse reward, we
Ours
Stack 2, Sparse
Stack 3, Sparse
Stack 4, Sparse
Stack 4, Step
Stack 5, Step
Stack 6, Step
99%
99%
1%
91%
49%
4%
Ours,
Resets
97%
89%
54%
73%
50%
32%
BC
HER
65%
1%
0%
-
0%
0%
0%
-
BC+
HER
65%
1%
0%
-
Fig. 3: Comparison of our method against baselines. The
value reported is the median of the best performance (success
rate) of all randomly seeded runs of each method.
reward the agent only if all blocks are at their goal positions:
rt = min 1||xi −gi ||<δ
i
(10)
The threshold δ is the size of a block, 5cm. Throughout the
paper we call this the “sparse” reward.
To enable solving the longer horizon tasks of stacking 4
or more blocks, we use the “step” reward :
X
rt = −1 +
1||xi −gi ||<δ
(11)
i
Note the step reward is still very sparse; the robot only
sees the reward change when it moves a block into its
target location. We subtract 1 only to make the reward more
interpretable, as in the initial state the first block is already
at its target.
Regardless of the reward type, an episode is considered
successful for computing success rate if all blocks are at
their goal position in their final state.
C. Network architectures
We use a 4 layer networks with 256 hidden units per layer
for π and Q for the HER tasks and stacking 3 or fewer
blocks. For stacking 4 blocks or more, we use an attention
mechanism [39] for the actor and a larger network. The
attention mechanism uses a 3 layer network with 128 hidden
units per layer to query the states and goals with one shared
head. Once a state and goal is extracted, we use a 5 layer
network with 256 hidden units per layer after the attention
mechanism. Attention speeds up training slightly but does
not change training outcomes.
D. Baselines
We include the following methods to compare our method
to baselines on stacking 2 to 6 blocks. 2
Ours: Refers to our method as described in section IV-C.
Ours, Resets: Refers to our method as described in section
IV-C with resets from demonstration states (Sec. IV-D).
BC: This method uses behavior cloning to learn a policy.
Given the set of demonstration transitions RD , we train the
2 Because of computational constraints, we were limited to 5 random seeds
per method for stacking 3 blocks, 2 random seeds per method for stacking 4
and 5 blocks, and 1 random seed per method for stacking 6 blocks. Although
we are careful to draw conclusions from few random seeds, the results are
consistent with our collective experience training these models. We report
the median of the random seeds everywhere applicable.
Stack 3, Sparse Reward
1.0
Ours
Ours, Resets
No Q-Filter
No BC
No HER
0.8
Success Rate
Task
0.6
0.4
0.2
0.0
0M
50M
100M
150M
200M
250M
300M
350M
400M
Timesteps
Fig. 4: Ablation results on stacking 3 blocks with a fully
sparse reward. We run each method 5 times with random
seeds. The bold line shows the median of the 5 runs while
each training run is plotted in a lighter color. Note “No
HER” is always at 0% success rate. Our method without
resets learns faster than the ablations. Our method with resets
initially learns faster but converges to a worse success rate.
policy π by supervised learning. Behavior cloning requires
much less computation than RL. For a fairer comparison,
we performed a large hyperparameter sweep over various
networks sizes, attention hyperparameters, and learning rates
and report the success rate achieved by the best policy found.
HER: This method is exactly the one described in Hindsight
Experience Replay [1], using HER and DDPG.
BC+HER: This method first initializes a policy (actor) with
BC, then finetunes the policy with RL as described above.
E. Results
We are able to learn much longer horizon tasks than
the other methods, as shown in Fig. 3. The stacking task
is extremely difficult using HER without demonstrations
because the chance of grasping an object using random
actions is close to 0. Initializing a policy with demonstrations
and then running RL also fails since the actor updates depend
on a reasonable critic and although the actor is pretrained,
the critic is not. The pretrained actor weights are therefore
destroyed in the very first epoch, and the result is no better
than BC alone. We attempted variants of this method where
initially the critic was trained from replay data. However,
this also fails without seeing on-policy data.
The results with sparse rewards are very encouraging. We
are able to stack 3 blocks with a fully sparse reward without
resetting to the states from demonstrations, and 4 blocks with
a fully sparse reward if we use resetting. With resets from
demonstration states and the step reward, we are able to learn
a policy to stack 6 blocks.
VIII. A BLATION E XPERIMENTS
In this section we perform a series of ablation experiments
to measure the importance of various components of our
method. We evaluate our method on stacking 3 to 6 blocks.
We perform the following ablations on the best performing
of our models on each task:
No BC Loss: This method does not apply the behavior
cloning gradient during training. It still has access to demonstrations through the demonstration replay buffer.
Stack 4, Step Reward
0.5
0.6
0.4
0.35
Stack 6, Step Reward
Ours
Ours, Resets
0.30
0.4
0.3
0.2
0.25
0.20
0.15
0.10
0.2
0.1
0.0
0.05
0.0
100M
200M
300M
400M
500M
600M
700M
800M
0M
3.0
2.5
Reward
2.0
1.5
1.0
500M
1000M
1500M
4.0
4.0
3.5
3.5
3.0
3.0
2.5
2.5
2.0
1.5
1.0
0.5
100M
200M
300M
400M
500M
Timesteps
600M
700M
800M
1000M
1500M
2000M
1000M
1500M
2000M
2.0
1.5
0.5
0.0
0M
500M
1.0
0.5
0.0
0.00
0M
Reward
0M
Reward
0.40
Ours
Ours, Resets
No Q-Filter
0.6
Success Rate
0.8
Success Rate
Stack 5, Step Reward
Ours
Ours, Resets
No Q-Filter
No BC
Success Rate
1.0
0.0
0M
500M
1000M
1500M
0M
500M
Timesteps
Timesteps
Fig. 5: Ablation results on longer horizon tasks with a step reward. The upper row shows the success rate while the lower
row shows the average reward at the final step of each episode obtained by different algorithms. For stacking 4 and 5 blocks,
we use 2 random seeds per method. The median of the runs is shown in bold and each training run is plotted in a lighter
color. Note that for stacking 4 blocks, the “No BC” method is always at 0% success rate. As the number of blocks increases,
resets from demonstrations becomes more important to learn the task.
No Q-Filter: This method uses standard behavioral cloning
loss instead of the loss from equation Eq. 8, which means
that the actor tries to mimic the demonstrator’s behaviour
regardless of the critic.
No HER: Hindsight Experience Replay is not used.
A. Behavior Cloning Loss
Without the behavior cloning loss, the method is significantly worse in every task we try. Fig. 4 shows the training
curve for learning to stack 3 blocks with a fully sparse
reward. Without the behavior cloning loss, the system is
about 2x slower to learn. On longer horizon tasks, we do
not achieve any success without this loss.
To see why, consider the training curves for stacking 4
blocks shown in Fig. 5. The “No BC” policy learns to stack
only one additional block. Without the behavior cloning loss,
the agent only has access to the demonstrations through the
demonstration replay buffer. This allows it to view highreward states and incentivizes the agent to stack more blocks,
but there is a stronger disincentive: stacking the tower higher
is risky and could result in lower reward if the agent knocks
over a block that is already correctly placed. Because of
this risk, which is fundamentally just another instance of
the agent finding a local optimum in a shaped reward, the
agent learns the safer behavior of pausing after achieving a
certain reward. Explicitly weighting behavior cloning steps
into gradient updates forces the policy to continue the task.
B. Q-Filter
The Q-Filter is effective in accelerating learning and
achieving optimal performance. Fig. 4 shows that the method
without filtering is slower to learn. One issue with the
behavior cloning loss is that if the demonstrations are suboptimal, the learned policy will also be suboptimal. Filtering
by Q-value gives a natural way to anneal the effect of the
demonstrations as it automatically disables the BC loss when
a better action is found. However, it gives mixed results
on the longer horizon tasks. One explanation is that in the
step reward case, learning relies less on the demonstrations
because the reward signal is stronger. Therefore, the training
is less affected by suboptimal demonstrations.
C. Resets From Demonstrations
We find that initializing rollouts from within demonstration states greatly helps to learn to stack 5 and 6 blocks but
hurts training with fewer blocks, as shown in Fig. 5. Note that
even where resets from demonstration states helps the final
success rate, learning takes off faster when this technique
is not used. However, since stacking the tower higher is
risky, the agent learns the safer behavior of stopping after
achieving a certain reward. Resetting from demonstration
states alleviates this problem because the agent regularly
experiences higher rewards.
This method changes the sampled state distribution, biasing it towards later states. It also inflates the Q values
unrealistically. Therefore, on tasks where the RL algorithm
does not get stuck in solving a subset of the full problem, it
could hurt performance.
IX. D ISCUSSION AND F UTURE W ORK
We present a system to utilize demonstrations along with
reinforcement learning to solve complicated multi-step tasks.
We believe this can accelerate learning of many tasks,
especially those with sparse rewards or other difficulties in
exploration. Our method is very general, and can be applied
on any continuous control task where a success criterion can
be specified and demonstrations obtained.
An exciting future direction is to train policies directly on
a physical robot. Fig. 2 shows that learning the pick-andplace task takes about 1 million timesteps, which is about
6 hours of real world interaction time. This can realistically
be trained on a physical robot, short-cutting the simulationreality gap entirely. Many automation tasks found in factories
and warehouses are similar to pick-and-place but without the
variation in initial and goal states, so the samples required
could be much lower. With our method, no expert needs
to be in the loop to train these systems: demonstrations
can be collected by users without knowledge about machine
learning or robotics and rewards could be directly obtained
from human feedback.
A major limitation of this work is sample efficiency
on solving harder tasks. While we could not solve these
tasks with other learning methods, our method requires a
large amount of experience which is impractical outside
of simulation. To run these tasks on physical robots, the
sample efficiency will have to improved considerably. We
also require demonstrations which are not easy to collect
for all tasks. If demonstrations are not available but the
environment can be reset to arbitrary states, one way to learn
goal-reaching but avoid using demonstrations is to reuse
successful rollouts as in [40].
Finally, our method of resets from demonstration states
requires the ability to reset to arbitrary states. Although we
can solve many long-horizon tasks without this ability, it is
very effective for the hardest tasks. Resetting from demonstration rollouts resembles curriculum learning: we solve a
hard task by first solving easier tasks. If the environment
does not afford setting arbitrary states, then other curriculum
methods will have to be used.
X. ACKNOWLEDGEMENTS
We thank Vikash Kumar and Aravind Rajeswaran for valuable discussions. We thank Sergey Levine, Chelsea Finn, and
Carlos Florensa for feedback on initial versions of this paper.
Finally, we thank OpenAI for providing a supportive research
environment.
R EFERENCES
[1] M. Andrychowicz et al., “Hindsight experience replay,” in Advances
in neural information processing systems, 2017.
[2] M. Večerı́k et al., “Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards,” arXiv
preprint arxiv:1707.08817, 2017.
[3] M. P. Deisenroth, C. E. Rasmussen, and D. Fox, “Learning to Control a
Low-Cost Manipulator using Data-Efficient Reinforcement Learning,”
Robotics: Science and Systems, vol. VII, pp. 57–64, 2011.
[4] Y. Duan et al., “One-shot imitation learning,” in NIPS, 2017.
[5] D. A. Pomerleau, “Alvinn: An autonomous land vehicle in a neural
network,” NIPS, pp. 305–313, 1989.
[6] M. Bojarski et al., “End to End Learning for Self-Driving Cars,” arXiv
preprint arXiv:1604.07316, 2016.
[7] A. Giusti et al., “A Machine Learning Approach to Visual Perception
of Forest Trails for Mobile Robots,” in IEEE Robotics and Automation
Letters., 2015, pp. 2377–3766.
[8] J. Nakanishi et al., “Learning from demonstration and adaptation of
biped locomotion,” in Robotics and Autonomous Systems, vol. 47, no.
2-3, 2004, pp. 79–91.
[9] M. Kalakrishnan et al., “Learning Locomotion over Rough Terrain
using Terrain Templates,” in The 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2009.
[10] S. Ross, G. J. Gordon, and J. A. Bagnell, “A Reduction of Imitation
Learning and Structured Prediction to No-Regret Online Learning,”
in Proceedings of the 14th International Conference on Artificial
Intelligence and Statistics (AISTATS), 2011.
[11] A. Ng and S. Russell, “Algorithms for Inverse Reinforcement Learning,” International Conference on Machine Learning (ICML), 2000.
[12] B. D. Ziebart et al., “Maximum Entropy Inverse Reinforcement
Learning.” in AAAI Conference on Artificial Intelligence, 2008, pp.
1433–1438.
[13] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse reinforcement learning,” in ICML, 2004, p. 1.
[14] C. Finn, S. Levine, and P. Abbeel, “Guided Cost Learning: Deep
Inverse Optimal Control via Policy Optimization,” in ICML, 2016.
[15] J. Peters, K. Mülling, and Y. Altün, “Relative Entropy Policy Search,”
Artificial Intelligence, pp. 1607–1612, 2010.
[16] M. P. Deisenroth and C. E. Rasmussen, “Pilco: A model-based and
data-efficient approach to policy search,” in ICML, 2011, pp. 465–472.
[17] V. Mnih et al., “Human-level control through deep reinforcement
learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015.
[18] D. Silver et al., “Mastering the game of Go with deep neural networks
and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, Jan 2016.
[19] S. Levine et al., “End-to-end training of deep visuomotor policies,”
CoRR, vol. abs/1504.00702, 2015.
[20] L. Pinto and A. Gupta, “Supersizing self-supervision: Learning
to grasp from 50k tries and 700 robot hours,” arXiv preprint
arXiv:1509.06825, 2015.
[21] S. Levine et al., “Learning hand-eye coordination for robotic grasping
with deep learning and large-scale data collection,” arXiv preprint
arXiv:1603.02199, 2016.
[22] S. Gu et al., “Deep Reinforcement Learning for Robotic Manipulation with Asynchronous Off-Policy Updates,” arXiv preprint
arXiv:1610.00633, 2016.
[23] T. P. Lillicrap et al., “Continuous control with deep reinforcement
learning,” arXiv preprint arXiv:1509.02971, 2015.
[24] V. Mnih et al., “Asynchronous methods for deep reinforcement learning,” in ICML, 2016.
[25] J. Schulman et al., “Trust region policy optimization,” in Proceedings
of the twenty-first international conference on Machine learning, 2015.
[26] T. Winograd, Understanding Natural Language. Academic Press,
1972.
[27] L. P. Kaelbling and T. Lozano-Perez, “Hierarchical task and motion
planning in the now,” IEEE International Conference on Robotics and
Automation, pp. 1470–1477, 2011.
[28] L. Kavraki et al., “Probabilistic roadmaps for path planning in highdimensional configuration spaces,” IEEE transactions on Robotics and
Automation, vol. 12, no. 4, pp. 566–580, 1996.
[29] S. Srivastava et al., “Combined Task and Motion Planning Through
an Extensible Planner-Independent Interface Layer,” in International
Conference on Robotics and Automation, 2014.
[30] I. Popov et al., “Data-efficient Deep Reinforcement Learning for
Dexterous Manipulation,” arXiv preprint arXiv:1704.03073, 2017.
[31] S. Schaal, “Robot learning from demonstration,” Advances in Neural
Information Processing Systems, no. 9, pp. 1040–1046, 1997.
[32] J. Peters and S. Schaal, “Reinforcement learning of motor skills with
policy gradients,” Neural Networks, vol. 21, no. 4, pp. 682–697, 2008.
[33] J. Kober and J. Peter, “Policy search for motor primitives in robotics,”
in Advances in neural information processing systems, 2008.
[34] T. Hester et al., “Learning from Demonstrations for Real World
Reinforcement Learning,” arXiv preprint arxiv:1704.03732, 2017.
[35] B. Kim et al., “Learning from Limited Demonstrations,” Neural
Information Processing Systems., 2013.
[36] T. Schaul et al., “Universal Value Function Approximators,” Proceedings of The 32nd International Conference on Machine Learning, pp.
1312–1320, 2015.
[37] E. Todorov, T. Erez, and Y. Tassa, “MuJoCo: A physics engine for
model-based control,” in The IEEE/RSJ International Conference on
Intelligent Robots and Systems, 2012.
[38] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
International Conference on Learning Representations (ICLR), 2015.
[39] D. Bahdanau, K. Cho, and Y. Bengio, “Neural Machine Translation
by Jointly Learning to Align and Translate,” in ICLR, 2015.
[40] C. Florensa et al., “Reverse Curriculum Generation for Reinforcement
Learning,” in Conference on robot learning, 2017.
| 9cs.NE
|
Random Monomial Ideals1
arXiv:1701.07130v2 [math.AC] 5 Jan 2018
Jesús A. De Loera, Sonja Petrović, Lily Silverstein, Despina Stasi, Dane Wilburne
Abstract: Inspired by the study of random graphs and simplicial complexes, and motivated
by the need to understand average behavior of ideals, we propose and study probabilistic models
of random monomial ideals. We prove theorems about the probability distributions, expectations
and thresholds for events involving monomial ideals with given Hilbert function, Krull dimension,
first graded Betti numbers, and present several experimentally-backed conjectures about regularity,
projective dimension, strong genericity, and Cohen-Macaulayness of random monomial ideals.
1. Introduction
Randomness has long been used to study polynomials. A commutative algebraist’s interest in randomness stems from the desire to understand “average” or “typical” behavior of ideals and rings. A
natural approach to the problem is to define a probability distribution on a set of ideals or varieties,
which, in turn, induces a distribution on algebraic invariants or properties of interest. In such a
formal setup, questions of expected (typical) or unlikely (rare, non-generic) behavior can be stated
formally using probability. Let us consider some examples. Already in the 1930’s, Littlewood and
Offord [32] studied the expected number of real roots of a random algebraic equation defined by random coefficients. The investigations on random varieties, defined by random coefficients on a fixed
Newton polytope support, have generated a lot of work now considered classical (see, e.g., [26, 29, 39]
and the references therein). A probabilistic analysis of algorithms called smooth analysis has been
used in algebraic geometry, see [5, 9]. Roughly speaking, smooth analysis measures the expected
performance of an algorithm under slight random perturbations of worst-case inputs. Another more
recent example of a notion of algebraic randomness appears in [15, 16], where they consider the
distribution of Betti numbers generated randomly using concepts from the Boij-Söderberg theory
[18]. Still, there are many other examples of algebraic topics in which probabilistic analysis plays a
useful role (see e.g., [13] and the references therein). Our paper introduces a new probabilistic model
on monomial ideals inside a polynomial ring.
Why work on a probabilistic model for monomial ideals? There are at least three good reasons:
First, monomial ideals are the simplest of ideals and play a fundamental role in commutative algebra,
most notably as Gröbner degenerations of general ideals, capturing the range of values shown in all
ideals for several invariants (see [12, 17]). Second, monomial ideals provide a strong link to algebraic
combinatorics (see [23, 33, 38]). Third, monomial ideals naturally generalize graphs, hypergraphs,
and simplicial complexes; and, since the seminal paper of Erdös and Rényi [19], probabilistic methods
have been successfully applied to study those objects (see, e.g., [1, 7, 27] and the references therein).
Thus, our work is an extension of both classical probabilistic combinatorics and the new trends in
stochastic topology.
Our goal is to provide a formal probabilistic setup for constructing and understanding distributions
of monomial ideals and the induced distributions on their invariants (degree, dimension, Hilbert
1
This work is supported by the NSF collaborative grants DMS-1522662 to Illinois Institute of Technology and
DMS-1522158 to the Univ. of California, Davis.
1
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
2
function, regularity, etc.). To this end, we define below a simple probability distribution on the set
of monomial ideals. Drawing a random monomial ideal from this distribution allows us to study
the average behavior of algebraic invariants. While we are interested in more general probability
distributions on ideals, some of which we describe in Section 5, we begin with the most basic model
one can consider to generate random monomial ideals; inspired by classical work, we call this family
the Erdős-Rényi-type model, or the ER-type model for random monomial ideals.
The ER-type model for random monomial ideals Let k be a field and let S = k[x1 , . . . , xn ] be
the polynomial ring in n indeterminates. We construct a random monomial ideal in S by producing
a random set of generators as follows: Given an integer D and a parameter p = p(n, D), 0 ≤ p ≤ 1,
we include independently with probability p each non-constant monomial of total degree at most D
in n variables in a generating set of the ideal. In other words, starting with B = ∅, each monomial in
S of degree at most D is added to the set B independently with equal probability p. The resulting
random monomial ideal is then simply I = (B); if B = ∅, then we let I = (0).
Henceforth, we will denote by B(n, D, p) the resulting Erdős-Rényi-type distribution on the sets
of monomials. Since sets of monomials are now random variables, a random set of monomials B
drawn from the distribution B(n, D, p) will be denoted by the standard notation for distributions
B ∼ B(n, D, p). Note that if B ⊂ S is any fixed set of monomials of degree at most D each and
B ∼ B(n, D, p), then
D+n
P (B = B) = p|B| (1 − p)( D )−|B|−1 .
In turn, the distribution B(n, D, p) induces a distribution I(n, D, p) on ideals. We will use the
notation I ∼ I(n, D, p) to indicate that B ∼ B(n, D, p) and I = (B) is a random monomial ideal
generated by the ER-type model.
Before we state our results, let us establish some necessary probabilistic notation and background.
Readers already familiar with random structures and probability background may skip this part, but
those interested in further details are directed to many excellent texts on random graphs and the
probabilistic method including [1, 7, 25].
Given a random variable X, we denote its expected value by E [X], its variance by Var [X], and its
covariance with random variable Y as Cov [X, Y ]. We will often use in our proofs four well-known
facts: the linearity property of expectation, i.e., E [X + Y ] = E [X] + E [Y ], the first moment method
(a special case of Markov’s inequality), which states that P (X > 0) ≤ E [X], and the second moment
method (a special case of Chebyshev’s inequality), which states that P (X = 0) ≤ Var [X]/(E [X])2
for X a non-negative integer-valued random variable. Finally, an indicator random variable 1A for
an event A is a random variable such that 1A = 1, if event A occurs, and 1A = 0 otherwise. Indicator
random variables behave particularly nicely with regard to taking expectations: for any event A,
E [1A ] = P (A) and Var [1A ] = P (A)(1 − P (A)). Following convention, we abbreviate by saying a
property holds a.a.s, to mean that a property holds asymptotically almost surely if, over a sequence
of sets, the probability of having the property converges to 1.
An important point is that I(n, D, p) is a parametric probability distributions on the set of monomial ideals, because it clearly depends on the value of the probability parameter p. A key concern
of our paper is to investigate how invariants evolve as p changes. To express some of our results,
we will need to use asymptotic analysis of probabilistic events: For functions f, g : N → R, we
write f (n) = o (g(n)), and g(n) = ω (f (n)) if limn→∞ f (n)/g(n) = 0. We also write f (n) ∼ g(n) if
limn→∞ f (n)/g(n) = 1, and f (n) g(n) if there exist positive constants n0 , c1 , c2 such that c1 g(n) ≤
f (n) ≤ c2 g(n) when n ≥ n0 . When a sequence of probabilistic events is given by f (n) for n ∈ N, we
say that f happens asymptotically almost surely, abbreviated a.a.s., if limn→∞ P (f (n)) = 1.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
3
In analogy to graph-theoretic properties, we define an (monomial) ideal-theoretic property Q to
be the set of all (monomial) ideals that have property Q. A property Q is monotone increasing if for
any monomial ideals I and J such that I ⊆ J, I ∈ Q implies J ∈ Q as well. We will see that several
algebraic invariants on monomial ideals are monotone.
Let I ∼ I(n, D, p). A function f : N → R is a threshold function for a monotone property Q if for
any p : N → [0, 1]:
(
0, if p = o (f (D)) ,
lim P (I ∈ Q) =
D→∞
1, if p = ω (f (D)) .
A similar definition holds for the case when f is a function of n and n → ∞. A threshold function
gives a zero/one law that specifies when a certain behavior appears.
Our results Our results, summarized in (A)-(E) below, describe the kind of monomial ideals
generated by the ER-type model, in the sense that we can get a handle on both the average and
the extreme behavior of these random ideals and how various ranges of the probability parameter p
control those properties. The following can also serve as an outline of this paper.
(A) Hilbert functions and the distribution of monomial ideals In Section 2, we begin our
investigation of random monomial ideals by showing that I(n, D, p) is not a uniform distribution on
all monomial ideals, but instead, the probability of choosing a particular monomial ideal under the
ER-type model is completely determined by its Hilbert function and the first total Betti number of
its quotient ring.
For an ideal I, denote by βi,j = βi,j (S/I) the (i, j)-th graded Betti number of S/I, that is, the
number of syzygies of degree j at step i of the minimal free resolution. We will collect the first graded
Betti numbers, counting the minimal generators of I, in a vector βˆ1 = (β1,1 , β1,2 , . . . , β1,reg0 ), and
denote by β1 = β1,1 + β1,2 + · · · + β1,reg0 the first total Betti number of S/I. We will denote by hI (·),
or simply hI , the Hilbert function of the ideal I. We can now state the following foundational result:
Theorem 1.1. Let I ⊆ S be a fixed monomial ideal generated in degree at most D. The probability
that the random monomial ideal I ∼ I(n, D, p) equals I is
PD
P (I = I) = pβ1 (S/I) (1 − p)−1+
d=1
hI (d)
.
In this way, two monomial ideals with the same Hilbert function and the same number of minimal
generators have the same probability of occurring. Then, the following natural question arises: How
many monomial ideals in n variables with generators of degree less than or equal to D and with a
given Hilbert function are there? It is well-known that one can compute the Hilbert function from the
graded Betti numbers (see e.g., [38]). Thus it is no surprise that we can state a combinatorial lemma
to count monomial ideals in terms of their graded Betti numbers. If we denote by N M on(n, D, h, βˆ1 )
the number of possible monomial ideals in n variables, generated in degree no more than D, Hilbert
function h, and first graded Betti numbers βˆ1 = (β1,1 , β1,2 , . . . , β1,reg0 ), then N M on(n, D, h, βˆ1 ) is
equal to the number of 0–1 vertices of a certain convex polytope (Lemma 2.3).
Based on Theorem 1.1 and Lemma 2.3, we can provide a closed formula for the induced distribution
on Hilbert functions under the ER-type model, that is, the probability that the ER-type model places
on any given Hilbert function. This is presented in Theorem 2.4.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
4
(B) The Krull dimension and random monomial ideals Section 3 is dedicated to a first
fundamental ring invariant: the Krull dimension dim S/I. By encoding the Krull dimension as the
transversal number of a certain hypergraph, we can show (see Theorem 3.1 for the precise statement)
that the probability that dim(S/I) is equal to t for 0P≤ t ≤ n, for a random monomial ideal
D n
I ∼ I(n, D, p), is given by a polynomial in p of degree t+1
i=1 i
i . This formula is exponentially
large but we make it explicit in some interesting values of t (Theorem 3.2); which in particular give
a complete description of the case for two- and three-variable polynomials.
Turning to asymptotic behavior, we prove that the Krull dimension of (S/I) can be controlled by
bounding the asymptotic growth of the probability parameter p = p(D) as D → ∞. The evolution
of the Krull dimension in terms of p is illustrated in Figure 1. The result is obtained by combining
the family of threshold results from Theorem 3.4 and is stated in the following corollary:
Corollary 1.2. Let I ∼ I(n, D, p), n be fixed, and 0 ≤ t < n. If the parameter p = p(D) is such
that p = ω D−(t+1) and p = o D−t as D → ∞, then dim(S/I) = t asymptotically almost surely.
It is very useful to consider the evolution – as the probability p increases from 0 to 1 – of the random
monomial ideal from the ER-type model and its random generating set B. For very small values of
p, B is all but guaranteed to be empty, and the random monomial ideal is asymptotically almost
surely the zero ideal. As p increases, the random monomial ideal evolves into a more complex ideal
generated in increasingly smaller degrees and support. Simultaneously, as the density of B continues
to increase with p, smaller-degree generators appear; these divide increasingly larger numbers of
monomials and the random ideal starts to become less complex as its minimal generators begin to
have smaller and smaller support, causing the Krull dimension to drop. Finally, the random ideal
becomes 0-dimensional and continues to evolve towards the maximal ideal.
n (zero ideal)
|
p=0
|
p = D−n
n−1
···
n−2
|
p = D−(n−1)
|
p = D−(n−2)
1
|
p = D−2
0
|
p = D−1
|
p=1
Fig 1. Evolution of the Krull dimension of S/I, where I is the random monomial ideal I ∼ I(n, D, p), as the probability
parameter p = p(D) changes.
(C) First Betti numbers and random monomial ideals The ER-type model induces a distribution on the Betti numbers of the coordinate ring of random monomial ideals. To study this
distribution, we first ask: what are the first Betti numbers generated under the model? A natural
way to ‘understand’ a distribution is to compute its expected value, i.e., the average first Betti
number. In Theorem 4.1 we establish the asymptotic behavior for the expected number of minimal
generators in a random monomial ideal.
We also establish and quantify threshold behavior of first graded Betti numbers of random monomial ideals I ∼ I(n, D, p) in two regimes, as n or D go to infinity, in Theorem 4.4. As in the Krull
dimension case, we combine the threshold results to obtain the following corollary:
Corollary 1.3. Let I ∼ I(n, D, p).
a) Let D be fixed, and d be a constant such
that 1 < d ≤ D. If the parameter p = p(n) is such that
−d
−d+1
p(n) = ω n
and p(n) = o n
then initdeg(I) = d asymptotically almost surely.
b) Let n be fixed. Suppose that di = di (D), 1 ≤ di ≤ D and limD→∞ di (D) = ∞
for i ∈ {1, 2}. If
−n
−n
the parameter p = p(D) is such that p(D) = ω d1
and p(D) = o d2
as D → ∞, then
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
5
d2 ≤ initdeg(I) ≤ d1 asymptotically almost surely. Note that this is attainable when d1 (D) =
ω (d2 (D)).
While Corollary 1.3 explains how the parameter p controls the smallest non-zero first Betti number,
it gives no information regarding the largest such number; the latter, of course, leads the complexity
of the first Betti numbers. To that end, we study the degree complexity reg0 (I), introduced by Bayer
and Mumford in [4] as the maximal degree of any reduced Gröbner basis of I; for monomial ideals,
this is simply the highest degree of a minimal generator. In Theorem 4.7 we show that p can be
specified so that, even though B contains many large degree monomials, none are minimal.
Intuitively, one can think of the above results in following terms: when the growth of p is bounded
above by that of 1/D, Corollary 1.3 provides the minimum degree of a minimal generator of I. When
the growth of p is bounded below by that of 1/D, Theorem 4.7 provides the maximum degree of a
minimal generator of I. The evolution of the minimum degree of a minimal generator for the case
when D is fixed and n tends to infinity is illustrated in Figure 2:
zero ideal
|
p = 0
|
p = n−D
D
D−1
|
p = n−(D−1)
D−2
|
p = n−(D−2)
|
p = n−(D−3)
···
···
d
|
p = n−d
|
p = n−(d−1)
|
p = 1
Fig 2. Evolution of the initial degree of S/I, where I is the random monomial ideal I ∼ I(n, D, p), as the model
parameter p = p(n) changes, for fixed D. The initial degree, that is, the smallest integer k such that β1,j = 0 for j < k
and β1,k 6= 0, is the value displayed in each box.
(D) Other probabilistic models for generating monomial ideals In Section 5 we define
more general models for random monomial ideals. This is useful to study distributions on ideals that
place varying probabilities on monomials depending on a property of interest. For example, one may
wish to change the likelihood of monomials according to their total degree. To that end, we define
a general multiparameter model for random monomials, show how the ER-type model is a special
case of it, and introduce the graded model. In Theorem 5.2 we show there is a choice of probability
parameters in the general model for random monomial ideals that recovers, as a special case, the
multiparameter model for random simplicial complexes of [11], which itself generalizes various models
for random graphs and clique complexes presented in [27, 31].
(E) Experiments & conjectures Section 6 contains a summary of computer simulations of
additional algebraic properties of random monomial ideals using the ER-type model. For each triple
of selected values of model parameters p, n, and D, we generate 1000 monomial ideals I ∼ I(n, D, p)
and compute their various properties. Our simulation results are summarized in Figures 3, 4, 5, 6,
7, and 8. The experiments suggest several trends and conjectures, which we state explicitly.
2. Hilbert functions and the Erdős-Rényi distribution on monomial ideals
The Hilbert function, and the Hilbert polynomial that it determines, are important tools in the classification of ideals via the Hilbert scheme [17]. Of course, monomial ideals are key to this enterprise:
given an arbitrary homogeneous ideal I and a monomial order on k[x1 , . . . , xn ], the monomial ideal
generated by the leading terms of the polynomials in I has the same Hilbert function as I. Therefore the study of all Hilbert functions reduces to the study of Hilbert functions of monomial ideals.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
6
Moreover, the Hilbert function hI (·) of a monomial ideal I ⊆ S has a useful combinatorial meaning:
the value of hI (d), d ≤ D, counts the number of standard monomials - monomials of degree exactly
d that are not contained in I.
In this section we explore the Erdős-Rényi distribution of monomial ideals and the induced distribution on Hilbert functions. It turns out they are intimately related.
2.1. Distribution of monomial ideals
Our first theorem precisely describes the distribution I(n, D, p) on ideals specified by the ER-type
model. Recall that Theorem 1.1 states that the Hilbert function and first total Betti number determine the probability of any given ideal in n variables generated in degree at most D.
Proof of Theorem 1.1. Fix I ⊆ S generated in degree at most D and let M be the unique minimal set
of generators of I. Then, I = I if and only if B ⊇ M and no monomial xα such that xα 6∈ I is in B. Let
A1 denote the event that each of the β1 (S/I) elements of M is in B and let A2 denote the event that
no monomial xα such that xα 6∈ I is in B. Then, the event I = I is equivalent to the event A1 ∩ A2 .
Since the events A1 and A2 are independent, P (I = I) = P (A1 ∩ A2 ) = P (A1 )P (A2 ). Observe that
P (A1 ) = pβ1 (S/I) , since each of the β1P
(S/I) elements of M is chosen to be in B independently with
PD
D
−1+
d=1 hI (d) , since there are exactly −1 +
probability p and P (A2 ) = (1 − p)
d=1 hI (d) monomials
of degree at most D not contained in I and each of them is excluded from B independently with
probability 1 − p.
Consider a special case of this theorem in the following example.
Example 2.1 (Principal random monomial ideals in 2 variables). Fix n = 2 and D > 0. In this
example, we calculate the probability of observing the principal random monomial ideal (B) =
0
0
(xα y γ ) ⊆ k[x, y], α + γ ≤ D. We observe the ideal (xα y γ ) exactly when xα y γ ∈ B and xα y γ 6∈ B
0
0
for every monomial xα y γ with α0 < α ≤ D or γ 0 < γ ≤ D. Thus, we must count the lattice points
below the “staircase” whose only corner is at (α, γ) ∈ Z2≥0 . A simple counting argument shows that
there are exactly γ2 (2D − γ + 3) + α2 (2D − α + 3) − αγ − 1 lattice points. Hence,
γ
α
P (I = (xα y γ )) = p(1 − p) 2 (2D−γ+3)+ 2 (2D−α+3)−αγ−1 .
Note that the right hand side of this expression is maximized at α = 1, γ = 0 or α = 0, γ = 1, so that
among the principal ideals, the most likely to appear under the ER-type model are (x) and (y).
A corollary of Theorem 1.1 follows directly from the well-known formulas for Stanley-Reisner rings:
Corollary 2.2. If I is a square-free monomial ideal, then the probability of I under the ER-type model
is determined by the number of minimal non-faces and faces of the associated simplicial complex.
2.2. Distribution of Hilbert functions
Since the Erdős-Rényi-type model specifies a distribution on monomial ideals, it provides a formal
probabilistic procedure to generate random Hilbert functions, in turn providing a very natural question: what is the induced probability of observing a particular Hilbert function? In other words, what
is the most likely Hilbert function under the ER-type model? If we randomly generate a monomial
ideal, are all Hilbert functions equally likely?
Not all possible non-negative vectors are first graded Betti vectors, but Lemma 2.3 below gives an
algorithmic way to determine, using polyhedral geometry, whether a particular potential set of first
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
7
graded Betti numbers can occur for a monomial ideal in n variables. We note that Onn and Sturmfels
introduced other polytopes useful in the study of initial (monomial) ideals of a zero-dimensional ideal
for n generic points in affine d-dimensional space: the staircase polytope and the simpler corner cut
polyhedron [34].
Lemma 2.3. Denote by N M on(n, D, h) the number of possible monomial ideals in n variables, with
generating monomials of degree no more than D and given Hilbert function h. Then N M on(n, D, h)
is equal to the number of vertices of the 0 − 1 convex polytope Q(n, D, h) defined by
X
|α|=L
xα =
n+L−1
− h(L), ∀L = 1, . . . D
L
xα ≤ xγ ,
∀ α ≤ γ, |α| + 1 = |γ|,
where α, γ denote exponent vectors of monomials with n variables and total degree no more than
D, thus the system has n+D
− 1 variables.
D
Proof of Lemma 2.3. The variables xα are indicator variables recording when a monomial xα , for
0 < |α| ≤ D, is chosen to be in the ideal (xα = 1) or is not chosen to be in the ideal (xα = 0).
The first type of equation forces the values of the Hilbert function at degree L to be satisfied. The
inequalities xα ≤ xγ , for all α ≤ γ, ensure the set of chosen monomials is closed under divisibility
and thus forms an ideal. A non-redundant subset of these satisfying |α| + 1 = |γ| suffices to cut out
the polyhedron.
Theorem 2.4. Let h be a Hilbert function for an ideal of the polynomial ring in n variables. As
before, Q(n, D, h) is the polytope whose vertices are in one-to-one correspondence with the distinct
monomial ideals in n variables, with generators of degree less than or equal to D, and Hilbert function
h. Then the probability of generating a random monomial ideal I from the ER-type model with Hilbert
function h is expressed by the following formula
PD
P (hI = h) = (1 − p)−1+
d=1
h(d)
X
pβ1 (v) ,
(2.1)
v∈V (Q(n,D,h))
where β1 (v) is the total first Betti number of the monomial ideal associated to the vertex v.
Proof of Theorem 2.4. To derive the above formula for the probability of randomly generating a
monomial ideal with a particular Hilbert function h using the ER model, we decompose the random
event into probability events that are combinatorially easy to count. The probability of generating
random monomial ideals with a given Hilbert function h is the sum of the probabilities of disjoint
events enumerated by the vertices v in the polytope Q(n, D, h) (see Lemma 2.3). For each choice
of v, the probability of the corresponding ideal being randomly generated by our model is given in
Theorem 1.1 and is completely determined by its Hilbert
function and first total Betti number. These
PD
ideals share a Hilbert function, h, and so (1 − p)−1+ d=1 h(d) is a common factor for all terms. The
choice of v determines the ideal and its first Betti number.
Some final remarks about Theorem 2.4. First of all, one can find examples that show that the
Hilbert function of an ideal does not uniquely determine the value of the number of minimal generators. Second, the values for the Betti numbers have explicit tight bounds from the Bigatti-HulettPardue theorem (see [6],[24], [35]). Finally, note that, from the point of view of computer science,
it is useful to know that the Hilbert function h can be specified with one polynomial (the Hilbert
polynomial) accompanied by a list of finitely many values that do not fit the polynomial.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
8
3. Krull dimension
In this section, we study the Krull dimension dim S/I of a random monomial ideal under the model
I(n, D, p). First, in Section 3.2, we give an explicit formula to calculate the probability of I having
a given Krull dimension. Second, in Section 3.3, we establish threshold results characterizing the
asymptotic behavior of the Krull dimension and we provides ranges of values of p that, asymptotically,
control the Krull dimension of the random monomial ideal.
3.1. Hypergraph Transversals and Krull dimension
Let us recall some necessary definitions. A hypergraph H is a pair H = (V, E) such that V is a finite
vertex set and E is a collection of non-empty subsets of V , called edges; we allow edges of size one.
A clutter is a hypergraph where no edge is contained in any other (see, e.g., [10, Chapter 1]). A
transversal of H, also called a hitting set or a vertex cover, is a set T ⊆ V such that no edge in E
has a nonempty intersection with T . The transversal number of H, also known as its vertex cover
number and denoted c(H), is the minimum cardinality of any transversal of H.
Given a collection B of monomials in k[x1 , . . . , xn ], the support hypergraph of B is the hypergraph
G(B) with vertex set V = {x1 , . . . , xn }, and edge set E = {supp(xα ) | xα ∈ B}, where supp(xα )
denotes the support of the monomial xα , i.e., xi ∈ supp(xα ), if and only if xi | xα . What we require
here is that for I = (B), the Krull dimension of S/I is determined by the transversal number of
G(B) through the formula
dim S/I + c (G(B)) = n.
(3.1)
This formula is known to hold for square-free monomial ideals minimally generated by B (see e.g.,
[20]). Since the support hypergraph of B is equal to that of its maximal square-free subset and
c (G(B)) is invariant under the choice of generating set B of I, the equation, obviously, also holds
for general ideals.
3.2. Probability distribution of Krull dimension
The relationship between Krull dimension and hypergraph transversals leads to a complete characterization of the probability of producing a monomial ideal with a particular fixed Krull dimension in the
ER-type model (Theorem 3.1). Explicitly computing the values given by Theorem 3.1 requires an exhaustive combinatorial enumeration, which we demonstrate for several special cases in Theorem 3.2.
Nevertheless, where the size of the problem makes enumeration prohibitive, Theorem 3.1 gives that
P (dim S/I = t) is always a polynomial in p of a specific degree, and thus can be approximated by
numerically evaluating P (dim S/I = t) for a sufficient number of p values, and interpolating.
Theorem 3.1. Let I ∼ I(n, D, p). For any integer P
t, 0 ≤ t ≤ n,
the probability that S/I has Krull
t+1 D n
dimension t is given by a polynomial in p of degree i=1 i i . More precisely,
P (dim S/I = t) =
X
Y
C∈Cn−t σ∈E(C)
D
1 − (1 − p)(|σ|)
Y
D
(1 − p)(|σ0 |) ,
σ 0 ⊂{x1 ,...,xn }
σ6⊆σ 0 ∀σ∈E(C)
where Cn−t is the set of all clutters on {x1 , . . . , xn } with transversal number n − t.
Proof. Let B be the random generating set that gave rise to I. By Equation (3.1), the probability
that S/I has Krull dimension t is equal to the probability that c(G(B)) = n − t. Let Gmin (B) be
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
9
the hypergraph obtained from G(B) by deleting all edges of G(B) that are strict supersets of other
edges in G(B): in other words, Gmin (B) contains all edges σ of G(B) for which there is no σ 0 in
G(B) such that σ 0 ( σ.
Also let Cn−t denote the set of clutters on {x1 , . . . , xn } with transversal number n − t. Then
c(G(B)) = n − t if and only if Gmin (B) = C for some C ∈ Cn−t . As these are disjoint events, we have
X
P (dim S/I = t) =
P (Gmin (B) = C) .
(3.2)
C∈Cn−t
Let σ ⊂ {x1 , . . . , xn }. There are
D
|σ|
monomials supported on σ, so
D
D
P (σ ∈ E(G(B))) = 1 − (1 − p)(|σ|) , and P (σ 6∈ E(G(B))) = (1 − p)(|σ|) .
D
Note that both expressions are polynomials in p of degree exactly |σ|
. Now Gmin (B) = C if only if:
1. every edge σ of C is an edge of G(B), and
2. every σ 0 ⊂ {x1 , . . . , xn } such that σ 0 does not contain any σ ∈ E(C) is not an edge of G(B).
Note that condition 2 is the contrapositive of: any edge of G(B) which is not an edge of C is not
minimal. Since edges are included in G(B) independently, the above are also independent events and
the probability of their intersection is
Y
Y
P (Gmin (B) = C) =
P (σ ∈ E(G(B)))
P σ 0 6∈ E(G(B))
σ∈E(C)
=
Y
σ∈E(C)
s.t. |σ|≤t+1
σ 0 ⊂{x1 ,...,xn }
σ6⊆σ 0 ∀σ∈E(C)
P (σ ∈ E(G(B)))
Y
σ 0 ⊂{x
1 ,...,xn }
s.t. |σ 0 |≤t and
0
σ6⊆σ ∀σ∈E(C)
P σ 0 6∈ E(G(B)) .
(3.3)
The last equality follows because for C ∈ Cn−t , if σ is an edge of C then |σ| ≤ t + 1, and also if σ 0 is
a subset of {x1 , . . . , xn } satisfying σ 6⊆ σ 0 for all σ ∈ E(C), then |σ 0 | ≤ t. To show the first statement,
suppose σ ∈ E(C) with |σ| > t + 1. Since C is a clutter, no proper subset of σ is an edge of C, so for
every σ 0 ∈ E(C), σ 0 6= σ, σ 0 contains at least one vertex not in σ. Hence the set T = {x1 , . . . , xn }\σ
intersects every edge of C except σ. By taking the union of T with any one vertex in σ, we create a
transversal of C of cardinality |T | + 1 = n − |σ| + 1 < n − t, contradicting c(C) = n − t.
For the second statement, suppose |σ 0 | > t. By assumption no edge of C is a subset of σ 0 , so every
edge of C contains at least one vertex in the set T = {x1 , . . . , xn }\σ 0 . Hence T is a transversal of C
with |T | = n − |σ 0 | < n − t, again a contradiction.
No subset of {x1 , . . . , xn } can appear in both
index sets of (3.3), and each subset that does appear
contributes a polynomial in p of degree Di . It follows that (3.3) is a polynomial in p of degree no
P
D n
greater than t+1
i=1 i
i , and hence so is P (dim S/I = t) as (3.2) is a sum of such polynomials.
To prove this bound is in fact the precise
of the polynomial, we show there is a particular
Pdegree
D n
clutter for which (3.2) has degree exactly t+1
i=1 i
i , and furthermore, that for every other clutter the expression in (3.3) is of strictly lower degree. Hence the sum over clutters in (3.2) has no
cancellation of this leading term.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
10
n
edges of cardinality t + 1 and no other edges.
Consider the hypergraph Knt+1 that contains all t+1
t+1
Then Kn ∈ Cn−t and
Y
Y
P Gmin (B) = Knt+1 =
P (σ ∈ E(G(B)))
P (σ 6∈ E(G(B)))
σ∈{x1 ,...,xn }
|σ|=t+1
σ∈{x1 ,...,xn }
|σ|≤t
D
n
D n
D n
= (1 − (1 − p)(t+1) )(t+1) (1 − p)Dn+( 2 )( 2 )+···+( t )( t )
which has the correct degree. On the other hand, if C ∈ Cn−t and C 6= Knt+1 , then at least one edge
σ of C is not an edge of Knt+1 ; hence |σ| ≤ t. All subsets properly containing σ are neither edges
of C, nor do they satisfy condition 2 above, hence these subsets are not indexed by either product
in 3.3. In particular there are positively many subsets of cardinality t + 1 which do not contribute
factors to P (Gmin (B) = C).
To use the main formula of Theorem 3.1, one must be able to enumerate all clutters on n vertices
with transversal number n − t. When t is very small or very close to n this is tractable, as we see in
the following theorem.
Theorem 3.2. Let I ∼ I(n, D, p). Then,
n
a) P (dim S/I = 0) = 1 − (1 − p)D .
(n−j )
D
P
2
n
D )j (1 − p)D(n−j) 1 − (1 − p)( 2 )
.
(1
−
(1
−
p)
b) P (dim S/I = 1) = n−1
j=0 j
n+D
n+D−j
n+D
P
c) P (dim S/I = n − 1) = −(1 − p)( n )−1 + nj=1 (−1)j−1 nj (1 − p)( n )−1−( n ) .
n+D
d) P (dim S/I = n) = (1 − p)( n )−1 .
For n ≤ 4, the listed formulas gives the complete probability distribution of Krull dimension induced
by I(n, D, p), for any integer D.
Proof. Proof of Part (a) For B ∼ B(n, D, p), I = (B), S/I is zero dimensional if and only if
Gmin (B) ∈ Cn . There is a single clutter on n vertices with transversal number n: the one with edge
set {{x1 }, {x2 }, . . . , {xn }}. Hence by Theorem 3.1,
P (dim S/I = 0) =
n
Y
i=1
1 − (1 − p)
(|{xDi }|)
= 1 − (1 − p)D
n
.
Proof of Part (b): S/I is one-dimensional if and only if Gmin (B) ∈ Cn−1 . We wish to describe Cn−1 .
Suppose C is a clutter on n vertices and exactly j of the vertices are contained in a 1-edge. Then
j 6= n else C would be the clutter from part a, so let 0 ≤ j ≤ n − 1, and denote by V the set of
these j vertices. Then V is a subset of any transversal of C. Let W = {x1 , . . . , xn }\V , then it can
be shown that c(C) = n − 1 if and only E(C) = {{xi } | xi ∈ V } ∪ {{xi , xk } | xi , xk ∈ W, xi 6= xk }.
Hence P (Gmin (B) = C) equals
Y
Y
Y
P ({xi } ∈ E(G(B)))
P ({xi , xk } ∈ E(G(B)))
P ({xi } 6∈ E(G(B)))
xi ∈V
xi ,xk ∈W
xi ∈W
(n−j )
D
2
= (1 − (1 − p)D )j 1 − (1 − p)( 2 )
(1 − p)D(n−j) .
The expression for P (dim S/I = 1) is obtained by summing over all
1-edges, for each 0 ≤ j ≤ n − 1.
n
j
ways of selecting the j
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
11
Proof of Part (c): For the case of (n − 1)–dimensionality, Theorem 3.1 requires us to consider clutters
with transversal number 1: clutters where some vertex appears in all the edges. However, for this
case we can give a simpler argument by looking at the monomials in B directly. Now the condition
equivalent to (n − 1)–dimensionality
is that there
is some xi that divides every monomial in B.
n+D−1
n+D
Fix an i, then there are D − 1 − D−1 monomials that xi does not divide. If Fi is the event
n+D
n+D−1
that x divides every monomial in B, then P (F ) = (1 − p)( D )−1−( D−1 ) . To get an expression for
i
i
(n − 1)-dimensionality, we need to take the union over all Fi , which we can do using an inclusionexclusion formula considering the events that two variables divide every monomial, three variables
divide every monomial, etc. Finally, we subtract the probability that B is empty.
Proof of Part (d): Since only the zero ideal has Krull dimension n, this occurs if and only if B is
n+D
empty, which has probability (1 − p)( D )−1 .
3.3. Threshold and asymptotic results
Next we study the asymptotic behavior of dim(S/I). In particular, we derive ranges of values of p
that control the Krull dimension as illustrated in Figure 1. As a special case we obtain a threshold
result for the random monomial ideal being zero-dimensional.
Lemma 3.3. Let n, t be fixed integers, with 0 ≤ t ≤ n − 1. If B ∼ B(n, D, p) and p(D) = o D−t ,
then a.a.s. B will contain no monomials of support size t or less as D tends to infinity.
Proof. By the first moment method, the probability that B contains some monomial of support at
most t is bounded above by the expected number of such monomials.
As
the number of monomials in
n D+t
n variables with support of size at most t is strictly less than t
is bounded
t , the expectation
−t and n and t are
above by the quantity p nt D+t
.
This
quantity
tends
to
zero
when
p(D)
=
o
D
t
constants, thus establishing the lemma.
Theorem 3.4. Let n, t be integers and p = p(D) with 1 ≤ t ≤ n and 0 ≤ p ≤ 1. Suppose
that B ∼ B(n, D, p) and I ∼ I(n, D, p). Then D−t is a threshold function for the property that
dim(S/I) ≤ t − 1. In other words,
(
0, if p = o D−t
.
lim P (dim(S/I) ≤ t − 1) =
D→∞
1, if p = ω D−t
Proof. Let p = o D−t . Consider the support hypergraph of B, G(B). If every monomial in B
has support of size t + 1 or more, then any (n − t)-vertex set of G(B) is a transversal, giving
c(G(B)) ≤ n−t. Thus Equation 3.1 and Lemma 3.3 give that dim(S/I) ≤ t−1 holds with probability
tending to 0.
Now, let p = ω D−t . Consider σ ⊂ {x1 , . . . , xn } with |σ| = t and let Xσ be the random variable
that records the number of monomials in B with support exactly the set σ. There are exactly Dt
such monomials of degree at most D: each monomial can be divided by all variables in σ and the
remainder is a monomial
in (some or all of) the variables of σ of degree
no greater than D − t.
D
−t
Thus, E [Xσ ] = t p tends to infinity with D, when p = ω D . Further, Var [Xσ ] ≤ E [Xσ ], as
Xσ is a sum of independent indicator random variables, one for each monomial with support exactly
σ. Hence we can apply the second moment method to conclude that as D tends to infinity,
P (Xσ = 0) ≤
Var [Xσ ]
≤
E [Xσ ]2
1
→ 0.
p
D
t
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
12
In other words, a.a.s. B will contain a monomial with support σ. Since σ was an arbitrary t-subset
of the variables {x1 , . . . , xn }, it follows that the same holds for every choice of σ.
The probability that any of the events Xσ = 0 occurs is bounded above by the sum of their
probabilities and tends to zero, as there are a finite number of sets σ. Then, with probability tending
to 1, G(B) contains all size-t edges, which implies that, for every vertex set of size n − t or less, G(B)
contains an edge of size t that is disjoint from it. Thus the transversal number of the hypergraph is
at least n − t + 1, and, by Formula 3.1, dim(S/I) = n − c(G(B)) ≤ t − 1 a.a.s.
The smallest choice of t-values in Theorem 3.4, t = 1, gives a threshold function for the event that
I ∼ I(n, D, p) has zero Krull dimension:
Corollary 3.5. Let I ∼ I(n, D, p), n be fixed and p = p(D). Then the function 1/D is a threshold
function for the 0-dimensionality of S/I.
Now we move to obtain Corollary 1.2 announced in the Introduction. Theorem
3.4 establishes
a
threshold result for each choice of constant t. But, if both p = ω D−(t+1) and p = o D−t hold,
then the theorem gives that events dim(S/I) > t and dim(S/I) < t each hold with probability
tending to 0. Therefore the probability of their union, in other words the probability that dim(S/I)
is either strictly larger than t or strictly smaller than t, also tends to 0. By combining the threshold
results applied to two consecutive constants t and t+1 in this way, we have established the Corollary.
4. Minimal generators of random monomial ideals from the ER-type model
A way to measure the complexity of a monomial ideal is to study the number of minimal generators.
As before, we will denote by β1 := β1 (S/I) the first total Betti number of the quotient ring of the
ideal I, that is, the total number of minimal generators of I. Subsection 4.1 addresses this invariant
and provides a formula for the asymptotic average value of β1 . A more refined invariant is, of course,
the set of first graded Betti numbers β1,d (S/I), recording the number of minimal generators of I of
degree d. Subsection 4.2 provides threshold results that explain how the asymptotic growth of the
parameter value p influences the appearance of minimal generators of given degree. Subsection 4.3
provides an analogous result for the degree complexity reg0 (I) of a monomial ideal I, that is, the
maximum degree of a minimal generator.
4.1. The first total Betti number
Consider the random variable β1 (S/I) for I ∼ I(n, D, p) in the regime D → ∞. Before stating
the results of this section, we introduce a concept from multiplicative number theory. The order r
divisor function, denoted by τr (k), is a function that records the number of ordered factorizations of
an integer k into exactly r parts (e.g., τ3 (4) = 6).
Theorem 4.1. Let I ∼ I(n, D, p) and let β1 (S/I) be the random variable denoting the number of
minimal generators of the random monomial ideal I. Then,
lim E [β1 (S/I)] = p
D→∞
∞
X
k=0
τn (k + 2)(1 − p)k .
(3)
Proof. For each α = (α1 , . . . , αn ) ∈ Zn≥0 , let 1α be the indicator random variable for the event
that xα is a minimal generator of I. Excluding the constant monomial 1 and xα itself, there are
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
13
Qn
Qn
monomials in S that divide xα . Therefore, E [1α ] = p(1 − p) i=1 (αi +1)−2 and letting
D → ∞ shows that
Qn
X
X
(1 − p) i=1 (αi +1)−2 .
E [1α ] = p
lim E [β1 (S/I)] =
i=1 (αi + 1) − 2
D→∞
06=α∈Zn
≥0
06=α∈Zn
≥0
Q
Q
Since ni=1 (αi + 1) − 2 is also the number of lattice points contained in the box ni=1 [0, αi ], excluding
the origin and α, it follows that every integer k ≥ 0 appears as an exponent of q in the infinite sum
above exactly as many times as k +P
2 can be factored into a product of n ordered factors, so that
k
the above sum may be rewritten as ∞
k=0 τn (k + 2)(1 − p) .
When n = 2, τ2 (k) is simply the usual divisor function which counts the number of divisors of
an integer k. The ordinary generating function of τ2 (k) has the form of a Lambert series (see [14]),
which leads to the following result in the two-variable case.
Corollary 4.2. Let I ∼ I(2, D, p). Then,
"∞
#
X (1 − p)k
p
p
lim E [β1 (S/I)] =
−
.
2
k
D→∞
(1 − p)
1−p
1 − (1 − p)
k=1
Proof. This follows from the Lambert series identity
(1−p)k
k=1 1−(1−p)k
P∞
=
P∞
k=1 τ2 (k)(1 − p)
k
in (3).
Corollary 4.2 provides an expression for the asymptotic expected number of minimal generators
of a random monomial ideal in two variables that can be easily evaluated for any fixed value of p.
For example, if I ∼ I(2, D, 10−5 ), then asymptotically (i.e., for very large values of D) one expects
I to have about 12 generators on average. On the other hand, when n ≥ 3, the right-hand side of
(3) is difficult to compute exactly, but one can obtain bounds from the fact that n ≤ τn (k) ≤ k n ,
which are valid for every k ≥ 2. Hence
n ≤ lim E [β1 (S/I)] ≤
D→∞
p
Li−n (1 − p),
(1 − p)2
(4)
P
xk
where Lis (x) denotes the polylogarithm function of order s defined by Lis (x) = ∞
k=1 ks . A polylogarithm of negative integer order can be expressed as a rational function of its argument:
n
X
1 − p j+1
Li−n (1 − p) =
j!S(n + 1, j + 1)
,
p
j=0
where S(n + 1, k + 1) denotes a Stirling number of the second kind (see [30]). The upper bound in
(4) is thus of order O (p−n ) .
4.2. The first graded Betti numbers
Here we study the behavior of the random variables β1,k (S/I) for I ∼ I(n, D, p). Specifically, we
show how the choice of p in the ER-type model controls the minimum degree of a generator of I, by
controlling the degrees of the monomials in the random set B.
Lemma 4.3. Let B ∼ B(n, D, p), I = hBi and Bd := {xα ∈ B : |α| ≤ d}.
a) If d, D is fixed, 0 < d ≤ D and p = p(n), then f (n) = n−d is a threshold function for the property
that |Bd | = 0.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
14
b) If n is fixed, d = d(D) ≤ D is such that limD→∞ d(D) → ∞ and p = p(D), then f (d) = d−n is a
threshold function for the property that |Bd | = 0.
Proof. For part a), we must show
that when p(n) = o n−d , then |Bd | = 0 with probability tending
to 1, and when p(n) = ω n−D , then |Bd | > 0 with probability tending to 1.
For d fixed, the number of monomials of degree at most d in n variables is n+d
nd (by
d
considering the usual asymptotic bounds for binomial coefficients). For any α ∈ Zn≥0 , with
P0 <
α ∈ B, so that |B | =
||α||1 ≤ d, let 1α be the indicator random variable for
the
event
that
x
d
α 1α ,
P
n+d
and therefore E [|Bd |] = α E [1α ] = p ·
− 1 . By the first moment method:
d
P (|Bd | > 0) ≤ E [|Bd |] = p ·
n+d
− 1 ≤ ped nd .
d
If p(n) = o n−d , then ped nd →
0 as n → ∞ and hence limn→∞ P (|Bd | > 0) = 0.
Now suppose p(n) = ω n−d . In this case E [|Bd |] → ∞, since:
E [|Bd |] = p ·
d
n+d
2
· pnd .
−1 ≥
d
d
The variance of |Bd | is calculated as follows:
Var [|Bd |] =
X
α
Var [1α ] +
n+d
Cov [1α , 1α0 ] =
− 1 · p · (1 − p) = E [|Bd |] · (1 − p),
d
0
X
α6=α
where we have again used that Var [1α ] = E 12α − E [1α ]2 = p − p2 and for α 6= α0 , 1α and 1α0
are independent and thus have 0 covariance. Then by the second moment method and the fact that
E [|Bd |] → ∞ when p = ω n−d :
Var [|Bd |]
1−p
1
=
≤
→ 0 as n → ∞.
2
(E [|Bd |])
E [|Bd |]
E [|Bd |]
Part b) follows in a similar fashion, once one notices that, for fixed n, n+d
=
d
tends to infinity.
P (|Bd | = 0) ≤
n+d
n
dn as d
From Lemma 4.3 we immediately obtain a threshold result for the first graded Betti numbers:
Theorem 4.4. Let B ∼ B(n, D, p) and I = (B).
a) If d ≤ D is fixed and p = p(n), then the function f (n) = n−d is a threshold function for the
property that initdeg(I) > d.
b) If n is fixed, d = d(D) ≤ D is such that limD→∞ d(D) = ∞ and p = p(D), then f (d) = d−n is a
threshold function for the property that initdeg(I) > d.
P
Proof. By definition, initdeg(I) > d if and only if dk=1 β1,k (S/I) = 0. Equivalently, initdeg(I) > d
if and only if B 6= ∅ and B contains no monomials of degree k for each 1 ≤ k ≤ d and therefore
!
d
X
P
β1,k (S/I) = 0 = P (|Bd | = 0) .
k=1
Both a) and b) now follow directly from Lemma 4.3.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
15
The case d = D of Theorem 4.4 deserves special mention, as it provides a threshold function for
the property that the random ideal generated by the ER-type model is the zero ideal.
Corollary 4.5. Let B ∼ B(n, D, p) and I = hBi.
a) Let D be fixed and p = p(n). Then n−D is a threshold function for the property that I = (0) .
b) Let n be fixed and p = p(D). Then D−n is a threshold function for the property that I = (0) .
On the other hand, in the critical region of this threshold function, both the expected number of
monomials in B and the variance of the number of monomials in B is constant.
Proposition 4.6. Let B ∼ B(n, D, p).
−
1
then E [|B|] = 1 and Var [|B|] = 1.
a) For any n, D, if p = 1/ n+D
D
b) For D fixed, if p = p(n) n−D then k1 ≤ limn→∞ E [|B|] ≤ k2 and c1 ≤ limn→∞ Var [|B|] ≤ c2 ,
for some constants k1 , k2 , c1 , c2 .
c) For n fixed, if p = p(D) D−n then k1 ≤ limD→∞ E [|B|] ≤ k2 and c1 ≤ limD→∞ Var [|B|] ≤ c2 ,
for some constants k1 , k2 , c1 , c2 .
Proof. Since B = BD , each of these claims follows directly from the calculated values of E [|BD |]
and Var [|BD |] from the proof of Lemma 4.3.
By combining threshold functions from Theorem 4.4 we can obtain Corollary 1.3 announced in the
Introduction: the result specifies ranges of p for which initdeg(I) = d asymptotically almost surely.
The argument is identical to the one used in Section 3.3 to obtain Corollary 1.2 from Theorem 3.4.
4.3. Degree complexity
In the previous subsection, we saw how the choice of p influences the smallest degree of a minimal
generator of a random monomial ideal. In this subsection, we ask the complementary question:
what can one say about the largest degree of a minimal generator of a random monomial ideal? The
following theorem establishes an asymptotic bound for the degree complexity reg0 (I) of I ∼ I(n, D, p)
for certain choices of probability parameter p.
Theorem 4.7. Let n be fixed, I ∼ I(n, D, p), p = p(D) and let r = r(D) be a function tending to
infinity as D → ∞. If p = ω 1r , then reg0 (I) ≤ nr a.a.s.
Proof. By the proof of Theorem 3.5, for each variable xi , a.a.s. B will contain a monomial of the
form xji where j ≤ r. If j1 , . . . , jn are integers such that 1 ≤ ji ≤ r for each 1 ≤ i ≤ n, then the ideal
(xj11 , . . . , xjnn ) contains all monomials of degree ≥ rn in S, since if deg(xα ) ≥ r, by the pigeonhole
principle at least one αi ≥ b nr c. Hence, with probability tending to 1, I will contain every monomial
in S of degree nr and thus every minimal generator has degree at most nr.
To illustrate this result, suppose that p = ω (1/log D) and let I ∼ I(n, D, p). Then, by Theorem 4.7, as D tends to infinity, one should expect reg0 (I) to be at most n log D. One should keep
in mind that I never contains generators of degree larger than D, and so the bound given in this
theorem is not optimal for certain choices of r.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
16
5. Other probabilistic models
As mentioned in the Introduction, the study of ‘typical’ ideals from a family of interest may require
not only the ER-type model, but also more general models of random monomial ideals. To that end,
we define the most general probabilistic model on sets of monomials which, a priori, provides no
structure, but it does provide a framework within which other models can be recovered.
A general model. Fix a degree bound D > 0. To each monomial xα ∈ k[x1 , . . . , xn ] with 0 <
deg(xα ) ≤ D, the general model for random monomials assigns an arbitrary probability 0 ≤ pα ≤ 1
of selecting the monomial xα :
P (xα ) = pα .
(5.1)
Hence the general model has many parameters, namely {pα : α ∈ Nn \ {0}, |α| ≤ D}. It is clear that
that, for a fixed n and fixed degree bound D, the ER-type model is a special case of the general
model, where pα = p(n, D), for all α ∈ Nn such that 0 < |α| ≤ D, does not depend on α.
Of course, there are many other interesting models one can consider. We define another natural
extension of ER-type model.
The graded model. Fix a degree bound D > 0. The graded model for random monomials places
the following probability on each monomial xα ∈ k[x1 , . . . , xn ] with 0 < deg(xα ) ≤ D:
pα = f (n, D, |α|),
(5.2)
where 0 ≤ pα ≤ 1. Note that, given deg(xα ) = |α|, the probability of each monomial is otherwise
constant in α. The graded model is a D-parameter family of probability distributions on random sets
of monomials that induces a distribution on monomial ideals in the same natural way as the ER-type
model. Namely, the probability model selects random monomials and they are added to a generating
set with probability according to the model, thus producing a random monomial ideal in k[x1 , . . . , xn ]
whose generators are of degree at most D. More precisely, let p(n, D) = (p1 (n, D), . . . , pD (n, D)) be
such that 0 ≤ pi (n, D) < 1 for all 1 ≤ i ≤ D. Initialize B = ∅ and for each monomial xα of degree i,
1 ≤ i ≤ D, add xα to B with probability pi (n, D). The resulting random monomial ideal is I = (B)
(again with the convention that if B = ∅, then we set I = (0)). Denote by G(n, D, p) the resulting
induced distribution on monomial ideals. As before, since ideals are now random variables, we will
write I ∼ G(n, D, p) for a random monomial ideal generated using the graded model.
The analogue of Theorem 1.1 for G(n, D, p), which has an almost identical proof, is as follows:
Theorem 5.1. Fix n, D, and the graded model parameters p(n, D) = (p1 , . . . , pD ). For any fixed
monomial ideal I ⊆ S, random monomial ideals from the graded model distribution I ∼ G(n, D, p)
satisfy the following:
D
Y
β (S/I)
P (I = I) =
pd 1,d
(1 − pd )hI (d) ,
d=1
where β1,d (S/I) is the number of degree-d minimal generators of I (that is, the first graded Betti
number of S/I), and hI (d) is its Hilbert function.
5.1. Random monomial ideals generalize random simplicial complexes
An abstract simplicial complex on the vertex set [n] is a collection of subsets (called faces) of [n] closed
under the operation of taking subsets. The dimension dim(σ) of a face σ ⊆ [n] is |σ| − 1. If Y is an
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
17
abstract simplicial complex on [n], then the dimension dim(Y ) of Y is max{dim(σ) : σ is a face of Y }
and the Stanley-Reisner ideal IY in the polynomial ring k[x1 , . . . , xn ] is the square-free monomial
ideal generated by the monomials corresponding to the non-faces τ of Y : IY := hτ : τ ⊆ [n], τ 6∈ Y i.
For example, if n = 2 and Y = {∅, {1}, {2}}, then IY = hx1 x2 i ⊂ k[x1 , x2 ]. A fundamental result in
combinatorial commutative algebra is that the Stanley-Reisner correspondence constitutes a bijection
between abstract simplicial complexes on [n] and the square-free monomial ideals in k[x1 , . . . , xn ].
See, for example, [33, Theorem 1.7].
In [11], Costa and Farber introduced a model for generating abstract simplicial complexes at
(r)
random. Their model, which we call the Costa-Farber model, is a probability distribution on ∆n ,
the set of all abstract simplicial complexes on [n] of dimension at most r. In it, one selects a vector
p̃ = (p̃0 , . . . , p̃r , 0, . . . , 0) of probabilities such that 0 ≤ p̃l ≤ 1 for all 0 ≤ l ≤ r. Then, one retains
each of the n vertices with probability p̃0 , and for each pair ij of remaining vertices, one forms the
edge ij with probability p̃1 , and for each triple of vertices ijk that are pairwise joined by an edge,
one forms the 2 dimensional face ijk with probability p̃2 , and so on.
One can study the combinatorial and topological properties of the random simplicial complexes
generated by the Costa-Farber model. In view of the Stanley-Reisner correspondence, this model can
be seen as a model for generating random square-free monomial ideals. Thus, the general model for
generating random monomial ideals can be viewed as a generalization of the Costa-Farber model,
which in turn has been proved to generalize many other models of random combinatorial objects, for
example, the Erdős-Rényi model for random graphs and Kahle’s model for random clique complexes
[27]. See [11, Section 2.3] for details.
The relationship between our models for random monomial ideals and the Costa-Farber model can
be made precise in the following way: there exists a choice of parameters pα in (5.1) such that the
resulting distribution on square-free monomial ideals in S = k[x1 , . . . , xn ] is precisely the distribution
on the abstract simplicial complexes on [n] under the Costa-Farber model.
Theorem 5.2. Let p̃ = (p̃0 , p̃1 , . . . , p̃r , 0, . . . , 0) denote the n-vector of probabilities in the Costa(r)
Farber model for random simplicial complexes. Let Y ⊂ ∆n be a simplicial complex on [n] of
dimension at most r and let IY be the Stanley-Reisner ideal corresponding to Y . Fix D = r + 1 and
specify the following probabilities pα , where α ∈ Nn and 0 < ||α||1 ≤ r + 1, for the general monomial
model (5.1):
(
1 − p̃deg(xα )−1 , if 0 6= α ∈ {0, 1}n ,
pα =
(5.3)
0,
otherwise .
Then, PCF (Y ) = P (IY ), where the former is probability under the Costa-Farber model and the latter
under the distribution on random monomial ideals induced by the general model (5.1).
In other words, the specification of probabilities in Theorem 5.2 recovers the Costa-Farber model
on random simplicial complexes as a sub-model of the model (5.1) on random monomial ideals. Note
that this specification of probabilities can be considered as an instance of the graded model with
support restricted to the vertices of the unit hypercube.
Proof. From [11, Equation (1)], the following probability holds under the Costa-Farber model:
PCF (Y ) =
r
Y
i=0
f (Y )
p̃i i
(1 − p̃i )ei (Y ) ,
where fi (Y ) denotes the number of i-dimensional faces of Y and ei (Y ) denotes the number of idimensional minimal non-faces of Y (i.e., the number of i-dimensional non-faces of Y that do not
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
18
strictly contain another non-face). The minimal non-faces of Y correspond exactly to the minimal
generators of the Stanley-Reisner ideal IY . Thus, IY has exactly ei (Y ) minimal generators of degree
i+1, that is, ei (Y ) = β1,i+1 (S/IY ). Each i-dimensional face of Y corresponds to a degree i+1 standard
square-free monomial of IY , hence we have a Hilbert function valuehIY (i + 1) = fi (Y ). Next, note
that the specification of probabilities in (5.3) depends only on the degree of each monomial, so that
0
if deg(xα ) = deg(xα ), then pα = pα0 . Hence, for each 1 ≤ j ≤ r + 1, we denote by pj the probability
assigned to the degree j monomials in (5.3), so pj = 1 − p̃j−1 . We now apply Theorem 5.1 to conclude
that
r+1
r
Y β1,j
Y
f (Y )
P (IY ) =
pj (1 − pj )hIY (j) =
(1 − p̃i )ei (Y ) p̃i i
= PCF (Y ), as desired.
j=1
i=0
Simplicial
complex Y
Ideal IY
Faces of Y
Non-faces of Y
PCF (Y )
P (IY )
Void
k[x1 , x2 ]
none
∅, x1 , x2 , x1 x2
0
0
∅
(x1 , x2 )
∅
x1 , x2 , x1 x2
(1 − p̃0 )2
p21
1
(x2 )
∅, x1
x2 , x1 x2
p̃0 (1 − p̃0 )
p1 (1 − p1 )
2
(x1 )
∅, x2
x1 , x1 x2
p̃0 (1 − p̃0 )
p1 (1 − p1 )
1
2
(x1 x2 )
∅, x1 , x2
x1 x2
p̃20 (1 − p̃1 )
(1 − p1 )(1 − p1 )p2
1
2
(0)
∅, x1 , x2 , x1 x2
none
p̃20 p̃1
(1 − p1 )(1 − p1 )(1 − p2 )
Table 1
A table illustrating the correspondence between random monomial ideals and the Costa-Farber model.
Table 5.1 illustrates Theorem 5.2 in the case n = 2 and r = 1. The Costa-Farber model generates
a random simplicial complex on {1, 2} by specifying a probability vector p̃ = (p̃0 , p̃1 ) and starting
with two vertices x1 and x2 . We retain each vertex independently with probability p̃0 . In the event
that we retain both x1 and x2 , we connect them with probability p̃1 . By Theorem 5.2, to recover this
model from Equation (5.1) we set D = 2 and p(2,0) = p(0,2) = 0, p1 = p(1,0) = p(0,1) = 1 − p̃0 , and
p2 = p(1,1) = 1 − p̃1 . The probability distribution of the set of all simplicial complexes Y on {1, 2}
and, equivalently, all square-free monomial ideals IY ⊆ k[x1 , x2 ] appear in the table.
To conclude, note that there are other models that do not sample individual monomials, but
instead sample ensembles of monomials all at the same time. Examples of these are the methods to
generate random integer partitions and Ferrers diagrams [36] and random lattice polytopes [3].
6. Experiments & conjectures
This section collects experimental work with two purposes in mind: illustrating the theorems we
proved in this paper, and also motivating further conjectures about the ER-type model. These
experiments were performed using Macaulay2 [22] and SageMath [37], including the MPFR [21] and
NumPy [2] modules. Graphics were created using SageMath.
We ran several sets of experiments using the ER-type model. All experiments considered varying
number of variables n, varying maximum degree D, and varying probability p. For each choice of
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
19
a triple (n, D, p), we generated a sample of N = 1000 monomial ideals. We then computed some
algebraic properties and tabulated the results. The figures in this section summarize the results and
support the conjectures we state.
Cohen-Macaulayness. Recall that an ideal has an (arithmetically) Cohen-Macaulay quotient
ring if its depth equals its Krull dimension. As is well-known, Cohen-Macaulayness is a very special
property from which one derives various special results; see [8] for a standard reference.
CM
CM
n=4, D=4
1
CM
n=4, D=5
1
1
CM
n=4, D=6
1
CM
n=4, D=7
1
CM
n=4, D=8
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
CM
0.1
0.15
0.2
0.25
p
CM
n=5, D=4
1
0.05
0.1
0.15
0.2
p
CM
n=5, D=5
1
1
0.05
0.1
0.15
p
CM
n=5, D=6
1
0.05
0.1
p
CM
n=5, D=7
1
0.05
0.1
p
0.2
CM
n=5, D=8
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
CM
0.05
0.1
0.15
0.2
p
0.25
n=6, D=4
1
CM
0.05
0.1
0.15
0.2
p
CM
n=6, D=5
1
1
0.05
0.1
0.15
p
CM
n=6, D=6
1
0.05
0.1
p
CM
n=6, D=7
1
0.05
0.1
p
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
CM
0.05
0.1
0.15
0.2
p
0.25
n=7, D=4
1
CM
0.05
0.1
0.15
0.2
p
CM
n=7, D=5
1
1
0.05
0.1
0.15
p
CM
n=7, D=6
1
0.05
0.1
p
CM
n=7, D=7
1
0.05
0.1
p
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
0.1
0.15
0.2
p
0.25
0.05
0.1
0.15
0.2
p
0.05
0.1
0.15
p
0.05
0.1
p
0.05
0.1
p
0.1
p
n=5, D=9
0.05
0.1
p
n=6, D=9
0.2
CM
n=7, D=8
0.05
0.2
CM
n=6, D=8
n=4, D=9
0.05
0.1
p
n=7, D=9
0.2
0.05
0.1
p
Fig 3. Frequency of random monomial ideals I ∼ I(n, D, p) for which S/I is Cohen-Macaulay for (n, D) = (4, 4)
through (n, D) = (7, 9), as the model parameter p takes values between D−n and D−1 . The sample size for each fixed
value of p is 1000.
Figure 3 shows the percent of ideals in the random samples whose quotient rings are arithmetically
Cohen-Macaulay for n = 4, . . . , 7, D = 4, . . . , 9. We see that, as n and D get larger, very few
Cohen-Macaulay ideals are generated, suggesting that Cohen-Macaulayness is a “rare” property in a
meaningful sense. We speculate that the appearances of Cohen-Macaulay ideals are largely or entirely
due to the appearance of the zero ideal and of zero-dimensional ideals, both of which are trivially
Cohen-Macaulay. For the smallest n and D values, the zero ideal appears with observable frequency
for all p values, as do zero-dimensional ideals. As p decreases toward the zero ideal threshold or
increases toward the zero-dimensional threshold, these probabilities grow, resulting in a “U” shape.
(Note that probabilities continue to increase to 1 for larger p beyond the domain of these plots.)
However, as n and D grow, the thresholds for the zero ideal and for zero-dimensionality become
pronounced, and even as p approaches either threshold the frequency of Cohen-Macaulayness stays
low. When p is so low that only the zero ideal is generated, Cohen-Macaulayness is of course observed
with probability 1, but as soon as nontrivial ideals are generated, the frequency plummets to near
zero. These experimental results suggest the following conjecture:
Conjecture 1. For large n and D, the only Cohen-Macaulay ideals generated by the ER-type model
are the trivial cases (the zero ideal or zero-dimensional ideals), with probability approaching 1. In
particular, using the threshold functions of Theorems 3.4 and 4.4, we conjecture that as n goes to
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
20
infinity, the probability
that I ∼ I(n, D, p) is Cohen-Macaulay will go to zero for p = p(D) satisfying
p(D) = o D−1 and p(D) = ω (D−n ).
Projective Dimension. Further exploring the complexity of minimal free resolutions, and how
much the ranges of Betti numbers vary, we investigate the projective dimension of S/I; i.e., the
length of the minimal free resolution of S/I.
pd
pd
n=4, D=4
5
pd
n=4, D=5
5
5
pd
n=4, D=6
5
pd
n=4, D=7
5
pd
n=4, D=8
5
4
4
4
4
4
4
3
3
3
3
3
3
2
2
2
2
2
2
1
1
1
1
1
0.05
pd
0.05
0.2
0.25
p
pd
0.1
0.15
0.2
p
0.25
n=6, D=4
7
6
5
4
3
2
1
pd
0.15
n=5, D=4
6
5
4
3
2
1
pd
0.1
0.05
0.1
0.15
0.2
p
0.25
n=7, D=4
8
7
6
5
4
3
2
1
0.05
0.1
0.15
0.2
p
0.25
0.05
0.05
0.2
p
pd
0.1
0.15
0.2
p
0.05
0.1
0.15
0.2
p
0.05
0.1
0.15
0.2
7
6
5
4
3
2
1
pd
n=7, D=5
8
7
6
5
4
3
2
1
6
5
4
3
2
1
pd
n=6, D=5
7
6
5
4
3
2
1
pd
0.15
n=5, D=5
6
5
4
3
2
1
pd
0.1
p
8
7
6
5
4
3
2
1
0.05
0.1
0.15
p
pd
n=5, D=6
0.05
0.1
0.15
p
pd
n=6, D=6
0.05
0.1
0.15
p
0.1
0.15
7
6
5
4
3
2
1
pd
n=7, D=6
0.05
6
5
4
3
2
1
p
8
7
6
5
4
3
2
1
0.05
0.1
p
pd
n=5, D=7
0.05
0.1
p
pd
n=6, D=7
0.05
0.1
p
0.1
7
6
5
4
3
2
1
pd
n=7, D=7
0.05
6
5
4
3
2
1
p
8
7
6
5
4
3
2
1
0.05
0.1
p
0.05
0.1
p
0.1
p
0.1
7
6
5
4
3
2
1
pd
n=7, D=8
0.05
6
5
4
3
2
1
pd
n=6, D=8
0.05
1
pd
n=5, D=8
p
n=4, D=9
8
7
6
5
4
3
2
1
0.05
0.1
p
n=5, D=9
0.05
0.1
p
n=6, D=9
0.05
0.1
p
n=7, D=9
0.05
0.1
p
Fig 4. Projective dimension of S/I for random monomial ideals I ∼ I(n, D, p) for (n, D) = (4, 4) through (n, D) =
(7, 9). The values of the model parameter p are 20 evenly spaced values between D−n and D−1 , and the sample size for
each value of p is 1000. For each p, circle sizes indicate the proportion of quotient rings with that projective dimension,
while the black curve indicates the mean projective dimension.
By the Hilbert syzygy theorem, the projective dimension is at most n. We see experimentally that
for large n and D, the projective dimension of S modulo any non-zero random ideal tends toward
this upper bound. (Since pd(S/I) = 0 if and only if I = (0), the projective dimension will always
concentrate at 0 below the p = D−n threshold for the zero ideal.)
Conjecture 2. As n and D increase, the projective dimension of the quotient ring of any non-zero
ideal in the ER-type model is equal to n with probability approaching 1.
Note that Conjecture 2 implies Conjecture 1. By the Auslander-Buchsbaum formula ([8], see also
[38]), pd(S/I) = n−depth(S/I) for all I. Hence Conjecture 2 implies depth(S/I) → 0, which implies
that the quotient ring of a nonzero ideal will be Cohen-Macaulay if and only if it is zero-dimensional.
Strong genericity. A monomial ideal I ⊂ k[x1 , . . . , xn ] is said to be strongly generic if no two
minimal generators agree on a non-zero exponent of the same variable. For example, x1 x2 , x21 x3
is a strongly generic monomial ideal in k[x1 , x2 , x3 ], but (x1 x2 , x1 x3 ) is not. Strong genericity is
interesting because the minimal free resolution of S/I has a combinatorial interpretation using the
Scarf polyhedral complex (see [33, Chapter 6] for details.), when I is strongly generic.
The zero ideal, with no generators, is trivially strongly generic, as is any principal ideal. At first
it might seem that increasing p, and thus increasing the expected number of monomials included in
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
SG
SG
n=4, D=4
1
SG
n=4, D=5
1
1
SG
n=4, D=6
1
SG
n=4, D=7
1
21
SG
n=4, D=8
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
SG
0.1
0.15
0.2
0.25
p
SG
n=5, D=4
1
0.05
0.1
0.15
0.2
p
SG
n=5, D=5
1
1
0.05
0.1
0.15
p
SG
n=5, D=6
1
0.05
0.1
p
SG
n=5, D=7
1
0.05
0.1
p
0.2
SG
n=5, D=8
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
SG
0.05
0.1
0.15
0.2
p
0.25
n=6, D=4
1
SG
0.05
0.1
0.15
0.2
p
SG
n=6, D=5
1
1
0.05
0.1
0.15
p
SG
n=6, D=6
1
0.05
0.1
p
SG
n=6, D=7
1
0.05
0.1
p
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
SG
0.05
0.1
0.15
0.2
p
0.25
n=7, D=4
1
SG
0.05
0.1
0.15
0.2
p
SG
n=7, D=5
1
1
0.05
0.1
0.15
p
SG
n=7, D=6
1
0.05
0.1
p
SG
n=7, D=7
1
0.05
0.1
p
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
0.1
0.15
0.2
p
0.25
0.05
0.1
0.15
0.2
p
0.05
0.1
0.15
p
0.05
0.1
p
0.05
0.1
p
0.1
p
n=5, D=9
0.05
0.1
p
n=6, D=9
0.2
SG
n=7, D=8
0.05
0.2
SG
n=6, D=8
n=4, D=9
0.05
0.1
p
n=7, D=9
0.2
0.05
0.1
p
Fig 5. Frequency of strong genericity in the ER-type model, for (n, D) = (4, 4) through (n, D) = (7, 9), as the model
parameter p takes values between D−n and D−1 . The sample size for each fixed value of p is 1000.
I, would always lower the frequency of strong genericity. However at the other extreme of p = 1 − ,
the maximal ideal (x1 , x2 , . . . , xn ) occurs frequently, and this ideal is strongly generic. Between these
extremal cases, we observe experimentally that very few random monomial ideals are strongly generic,
and suspect a connection between this behavior and our results in Section 4 about the number of
minimal generators.
Conjecture 3. As n and D increase, there will be a lower threshold function l(n, D) and an upper
threshold function u(n, D), such that the probability of being strongly generic will go to zero for
p = o (u(n, D)) and p = ω (l(n, D)), while the probability of being strongly generic will go to one for
p = o (l(n, D)) as well as for p = ω (u(n, D)).
Note that because the plots in Figure 5 display p values between D−n and D−1 , the behavior as
p approaches 1 is not visible.
We now turn to several properties of monomial ideals for which we experimentally observed noticeable patterns, but for which we do not have explicit conjectures to present.
Castelnuovo-Mumford regularity. The Castelnuovo-Mumford regularity (or simply regularity)
of an ideal is a measure of the variation and spread of the the degrees of the generators at each of
the modules in a minimal free resolution of the ideal. To compute the regularity requires computing
the resolutions and the number of rows in the Macaulay2-formatted Betti diagram. Of course, the
degree complexity studied in Section 4 is similar but less refined measure of complexity, as it counts
the number of entries in the first column of the Betti diagram only.
The simulations on the range of values realized for regularity of random monomial ideals, which
do not take into account any zero ideals generated (since their regularity is −∞), show an interesting
trend. Namely, already for the small values of n and D depicted in Figure 6, there is quite a range
of values obtained under the model. As n grows these values concentrate more tightly around their
mean, while seeming to follow a binomial distribution (discretized version of a normal distribution).
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
reg
reg
n=4, D=4
20
reg
n=4, D=5
20
20
reg
n=4, D=6
20
reg
n=4, D=7
20
22
reg
n=4, D=8
20
15
15
15
15
15
15
10
10
10
10
10
10
5
5
5
5
5
0.05
reg
0.1
0.15
0.2
0.25
p
reg
n=5, D=4
20
0.05
0.1
0.15
0.2
p
reg
n=5, D=5
20
20
0.05
0.1
0.15
p
reg
n=5, D=6
20
0.05
0.1
p
reg
n=5, D=7
20
5
0.05
0.1
p
reg
n=5, D=8
20
15
15
15
15
15
15
10
10
10
10
10
10
5
5
5
5
5
reg
0.05
0.1
0.15
0.2
p
0.25
n=6, D=4
20
reg
0.05
0.1
0.15
0.2
p
reg
n=6, D=5
20
20
0.05
0.1
0.15
p
reg
n=6, D=6
20
0.05
0.1
p
reg
n=6, D=7
20
0.1
p
reg
n=6, D=8
20
15
15
15
15
15
10
10
10
10
10
10
5
5
5
5
5
reg
0.1
0.15
0.2
p
0.25
n=7, D=4
20
reg
0.05
0.1
0.15
0.2
p
reg
n=7, D=5
20
20
0.05
0.1
0.15
p
reg
n=7, D=6
20
0.05
0.1
p
reg
n=7, D=7
20
0.1
p
reg
n=7, D=8
20
15
15
15
15
15
10
10
10
10
10
10
5
5
5
5
5
0.1
0.15
0.2
p
0.25
0.05
0.1
0.15
0.2
p
0.05
0.1
0.15
p
0.05
0.1
p
0.1
p
n=5, D=9
0.05
0.1
p
n=6, D=9
5
0.05
15
0.05
0.05
5
0.05
15
0.05
n=4, D=9
0.05
0.1
p
n=7, D=9
5
0.05
0.1
p
0.05
0.1
p
Fig 6. Castelnuovo-Mumford regularity of random monomial ideals I ∼ I(n, D, p) for (n, D) = (4, 4) through (n, D) =
(7, 9). Note that n increases across rows and D increases down columns, and that the values of the model parameter
p are 20 evenly spaced values between D−n and D−1 . The sample size for each fixed value of p is 1000. Circle sizes
indicate the relative frequencies of occurrence of random monomial ideals of corresponding regularity, and the mean
regularity for each p value is indicated by the black curves.
Simplicial homology of associated simplicial complexes. One motivation for studying random monomial ideals is that they provide a way to generate random simplicial complexes different
from earlier work, where the authors randomly generate k-dimensional faces of complexes for some
fixed integer k (see [11, 27, 31]). Instead, by Stanley-Reisner duality, the indices in a monomial we
generate are the elements of a non-face of a simplicial complex.
There are two natural ways to randomly generate sets of square-free monomials and our experiments considered both of them. First, as radicals of random monomial ideals drawn from the ER-type
model and, second, directly as random square-free monomial ideals drawn from the general model,
as described in Theorem 5.2, that places zero probability on non-square-free-monomials and probability p on square-free monomials. In our experiments, given an ideal in n variables, we computed
the Z2 -homology Hi , for i from 0 to n − 1, of the associated simplicial complex. Figure 7 displays
the homological properties of the random simplicial complexes obtained via the Stanley-Reisner correspondence from the first approach. Figure 8 displays the homological properties of the random
simplicial complexes obtained via the second method, directly generating square-free ideals, avoiding
taking the radical.
The patterns of appearance and disappearance of higher homology groups, visible in both figures,
are familiar in the world of random topology; see, for example, [28]. For each i there is a lower
threshold below which Hi always vanishes, and an upper threshold above which Hi also vanishes, while
between these two thresholds we see a somewhat normal-looking curve. Since our model parameter
p controls how many faces are removed from a complex, higher dimensional objects are associated
with lower values of p: a reversal of the typical behavior in prior random topological models. Another
interesting pattern in both sets of experiments (unfortunately not evident in the plots) is that there
was an overall tendency for each random simplicial complex to have no more than one non-trivial
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
n=4, D=4
1
n=4, D=5
1
1
n=4, D=6
1
n=4, D=7
1
23
n=4, D=8
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
0.1
0.15
0.2
0.25
p
n=5, D=4
1
0.05
0.1
0.15
0.2
p
n=5, D=5
1
0.05
1
0.1
0.15
p
n=5, D=6
0.05
1
0.1
0.2
p
n=5, D=7
0.05
1
0.1
p
n=5, D=8
0.05
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
0.1
0.15
0.2
p
0.25
n=6, D=4
1
0.05
0.1
0.15
0.2
p
n=6, D=5
1
0.05
1
0.1
0.15
p
n=6, D=6
0.05
1
0.1
0.05
1
0.1
p
n=6, D=8
0.05
1
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.05
0.1
0.15
0.2
p
0.25
n=7, D=4
1
0.05
0.1
0.15
0.2
p
n=7, D=5
1
0.05
1
0.1
0.15
p
n=7, D=6
0.05
1
0.1
0.05
1
0.1
p
n=7, D=8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.1
0.15
0.2
p
0.25
0.05
0.1
Legend:
0.15
0.2
H0
p
0.05
0.1
H1
0.15
p
0.05
H2
0.1
H3
0.1
p
n=6, D=9
0.05
1
0.8
0.05
p
0.2
p
n=7, D=7
0.1
n=5, D=9
0.2
p
n=6, D=7
n=4, D=9
0.1
p
n=7, D=9
0.2
p
0.05
H4
0.1
p
0.05
0.1
p
H5
Fig 7. Non-trivial homology groups for the random simplicial complexes whose nonfaces are radicals of random monomial ideals in the ER-type model setting. Each subplot in Figure 7 represents a fixed pair of (n, D) values, with n increasing down columns, and D increasing across rows. For each n, D and p value we generated 1000 ideals rad(I ∼ I(n, D, p)).
Each subplot registers the frequency at which the second kind of random simplicial complex exhibited Hi 6= {0} over Z2 .
The values of p in each subplot are chosen in the interval [D−n , D−1 ]. We note that I tends to the maximal ideal as
p → 1, so every curve will tend to 0.
homology group. That is, if for a particular n, D and p the experiments found that H0 was nontrivial
with frequency x, and H1 was nontrivial with frequency y, one might expect to see random simplicial
complexes with H0 and H1 simultaneously nontrivial with frequency xy. This would be the case
if these events were statistically uncorrelated. However we consistently observed frequencies much
lower than this, i.e. a negative correlation between nonzero homology at i and nonzero homology at
j, for every i 6= j.
7. Acknowledgements
We are grateful for the comments, references and suggestions from Eric Babson, Boris Bukh, Seth
Sullivant, Agnes Szanto, Ezra Miller, and Sayan Mukherjee. We are also grateful to REU student
Arina Ushakova who did some initial experiments for this project.
References
[1] Alon, N., and Spencer, J. H. The Probabilistic Method, 4th ed. Wiley, 2016.
[2] Ascher, D., Dubois, P. F., Hinsen, K., Hugunin, J., and Oliphant, T. Numerical
Python, ucrl-ma-128569 ed. Lawrence Livermore National Laboratory, Livermore, CA, 1999.
[3] Bárány, I., and Matousek, J. The randomized integer convex hull. Discrete & Computational Geometry 33, 1 (2005), 3–25.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
n=4, D=4
0.6
n=4, D=5
0.6
0.4
0.4
0.2
0.4
0.6
0.8
1
p
n=5, D=4
0.6
0.2
0.4
0.6
0.8
1
n=5, D=5
0.6
0.2
0.4
0.6
0.8
1
n=5, D=6
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1
p
n=5, D=7
0.6
0.2
0.2
0.4
0.6
0.8
1
p
n=5, D=8
0.6
0.2
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.6
0.8
1
p
n=6, D=4
0.6
0.2
0.4
0.6
0.8
1
p
n=6, D=5
0.6
0.2
0.4
0.6
0.8
1
p
n=6, D=6
0.6
0.2
0.4
0.6
0.8
1
p
n=6, D=7
0.6
0.2
0.4
0.6
0.8
1
p
n=6, D=8
0.6
0.2
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.6
0.8
1
p
n=7, D=4
0.6
0.2
0.4
0.6
0.8
1
p
n=7, D=5
0.6
0.2
0.4
0.6
0.8
1
p
n=7, D=6
0.6
0.2
0.4
0.6
0.8
1
p
n=7, D=7
0.6
0.2
0.4
0.6
0.8
1
p
n=7, D=8
0.6
0.2
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.4
0.6
0.8
1
p
0.2
Legend:
0.4
0.6
0.8
1
H0
p
0.2
0.4
0.6
H1
0.8
1
p
0.2
H2
0.4
0.6
0.8
H3
1
p
0.2
0.4
H4
0.6
0.8
1
p
0.8
1
0.8
1
0.8
1
0.8
1
p
0.4
0.6
p
0.4
0.6
p
n=7, D=9
0.6
0.4
0.6
n=6, D=9
0.6
0.4
0.4
n=5, D=9
0.6
0.4
0.2
n=4, D=9
0.6
0.4
0.2
p
n=4, D=8
0.6
0.4
0.2
p
n=4, D=7
0.6
0.4
0.2
0.2
n=4, D=6
0.6
24
0.2
0.4
0.6
p
H5
Fig 8. Non-trivial homology groups for the random simplicial complexes whose non-faces are random square-free monomial ideals. Each subplot is for a pair (n, D) values, with n increasing down columns, and D increasing across rows.
With a sample size of 1000 monomial ideals for each n, D and p, each subplot shows the frequency at which the second kind of random simplicial complex exhibited Hi 6= {0} over Z2 . The p values for this set of experiments are
{0.05n | 0 ≤ n ≤ 20}. The subplots register the frequency at which these randomly generated simplicial complexes
exhibited Hi 6= {0}.
[4] Bayer, D., and Mumford, D. What can be computed in algebraic geometry? In Computational algebraic geometry and commutative algebra (Apr. 1992), University Press, pp. 1–48.
[5] Beltrán, C., and Pardo, L. M. Smale’s 17th problem: average polynomial time to compute
affine and projective solutions. Journal of the AMS 22, 2 (2009.), 363–385.
[6] Bigatti, A. M. Upper bounds for the Betti numbers of a given Hilbert function. Commutative
Algebra 21 (1993), 2317–2334.
[7] Bollobás, B. Random graphs, 2nd ed. Cambridge University Press, 2001.
[8] Bruns, W., and Herzog, J. Cohen-Macaulay rings. Cambridge University Press, 1998.
[9] Bürgisser, P., and Cucker, F. On a problem posed by Steve Smale. Annals of Mathematics
174, 3 (2011), 1785–1836.
[10] Cornuéjols, G. Combinatorial Optimization: Packing and Covering. CBMS-NSF Regional
Conference Series in Applied Mathematics. SIAM, 2001.
[11] Costa, A., and Farber, M. Random simplicial complexes. In Configuration spaces - Geometry, Topology, and Representation Theory, INdAM Series. Springer, To appear.
[12] Cox, D., Little, J. B., and O’Shea, D. Ideals, Varieties, and Algorithms: An Introduction
to Computational Algebraic Geometry and Commutative Algebra. Springer, 2007.
[13] De Loera, J. A., Petrović, S., and Stasi, D. Random sampling in computational algebra:
Helly numbers and violator spaces. Journal of Symbolic Computation 77 (2016), 1–15.
[14] Dilcher, K. Some q-series identities related to divisor functions. Discrete mathematics 145, 1
(1995), 83–93.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
25
[15] Ein, L., Erman, D., and Lazarsfeld, R. Asymptotics of random Betti tables. Journal für
die reine und angewandte Mathematik (Crelle’s Journal). (To appear).
[16] Ein, L., and Lazarsfeld, R. Asymptotic syzygies of algebraic varieties. Inventiones mathematicae 190, 3 (2012), 603–646.
[17] Eisenbud, D. Commutative algebra with a view toward algebraic geometry, vol. 150 of Graduate
Texts in Mathematics. Springer, 1995.
[18] Eisenbud, D., and Schreyer, F.-O. Betti numbers of graded modules and cohomology of
vector bundles. Journal of the American Mathematical Society 22, 3 (2009), 859–888.
[19] Erdös, P., and Rényi, A. On random graphs, I. Publicationes Mathematicae (Debrecen) 6
(1959), 290–297.
[20] Faridi, S. Cohen-macaulay properties of square-free monomial ideals. Journal of Combinatorial
Theory, Series A 109, 2 (2005), 299–329.
[21] Fousse, L., Hanrot, G., Lefèvre, V., Pélissier, P., and Zimmermann, P. Mpfr: A
multiple-precision binary floating-point library with correct rounding. ACM Transactions on
Mathematical Software 33, 2 (June 2007).
[22] Grayson, D. R., and Stillman, M. E. Macaulay2, a software system for research in algebraic
geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[23] Herzog, J., and Hibi, T. Monomial ideals, vol. 260 of Graduate Texts in Mathematics.
Springer-Verlag London, Ltd., London, 2011.
[24] Hulett, H. Maximum Betti numbers of homogeneous ideals with a given Hilbert function.
Communications in Algebra 21 (1993), 2335–2350.
[25] Janson, S., Luczak, T., and Ruciński, A. Random graphs. Interscience Series in Discrete
Mathematics and Optimization. Wiley, New York, 2000.
[26] Kac, M. On the average number of real roots of a random algebraic equation. Bulletin of the
American Mathematical Society 49 (1943), 314–320.
[27] Kahle, M. Topology of random simplicial complexes: a survey. In Algebraic topology: applications and new directions, vol. 620 of Contemporary Mathematics. American Mathematical
Society, Providence, RI, 2014, pp. 201–221.
[28] Kahle, M., and Meckes, E. Limit theorems for Betti numbers of random simplicial complexes. Homology Homotopy Appl. 15, 1 (2013), 343–374.
[29] Kouchnirenko, A. G. Polyèdres de Newton et nombres de Milnor. Inventiones mathematicae
32 (1976), 1–32.
[30] Lewin, L. Polylogarithms and associated functions. Elsevier Science Ltd, 1981.
[31] Linial, N., and Meshulam, R. Homological connectivity of random 2-complexes. Combinatorica 26, 4 (Aug. 2006), 475–487.
[32] Littlewood, J. E., and Offord, A. C. On the Number of Real Roots of a Random Algebraic
Equation. Journal of the London Mathematical Society S1-13, 4 (1938), 288.
[33] Miller, E., and Sturmfels, B. Combinatorial commutative algebra, vol. 227 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 2005.
[34] Onn, S., and Sturmfels, B. Cutting corners. Advances in applied mathematics 23, 1 (1999),
29–48.
[35] Pardue, K. Deformation classes of graded modules and maximal Betti numbers. Illinois
Journal of Mathematics 40 (1996), 564–585.
[36] Pittel, B. On a likely shape of the random Ferrer’s diagram. Advances in Applied Mathematics
18, 4 (1997), 432 – 488.
[37] Sage Developers. SageMath, the Sage Mathematics Software System (Version 7.3), 2016.
De Loera, Petrović, Silverstein, Stasi, Wilburne/Random monomial ideals
26
http://www.sagemath.org.
[38] Stanley, R. P. Combinatorics and commutative algebra, 2nd ed., vol. 41 of Progress in Mathematics. Birkhäuser Boston, Inc., Boston, MA, 1996.
[39] Sturmfels, B. Polynomial equations and convex polytopes. The American Mathematical
Monthly 105, 10 (1998), 907–922.
| 0math.AC
|
Soft-Deep Boltzmann Machines
arXiv:1505.02462v3 [cs.NE] 21 Jun 2015
Taichi Kiwaki
Graduate School of Engineering
The University of Tokyo
[email protected]
Abstract
We present a layered Boltzmann machine (BM) that can better exploit the advantages of a distributed representation. It is widely believed that deep BMs (DBMs)
have far greater representational power than its shallow counterpart, restricted
Boltzmann machines (RBMs). However, this expectation on the supremacy of
DBMs over RBMs has not ever been validated in a theoretical fashion. In this
paper, we provide both theoretical and empirical evidences that the representational power of DBMs can be actually rather limited in taking advantages of distributed representations. We propose an approximate measure for the representational power of a BM regarding to the efficiency of a distributed representation.
With this measure, we show a surprising fact that DBMs can make inefficient
use of distributed representations. Based on these observations, we propose an
alternative BM architecture, which we dub soft-deep BMs (sDBMs). We show
that sDBMs can more efficiently exploit the distributed representations in terms
of the measure. Experiments demonstrate that sDBMs outperform several stateof-the-art models, including DBMs, in generative tasks on binarized MNIST and
Caltech-101 silhouettes.
1
Introduction
One aspect behind superior performance of deep architectures is the effective use of distributed
representations [1, 2]. A representation is said distributed if it consists of mutually non-exclusive
features [1]. Distributed representations can efficiently model complex functions with enormous
number of variations by dividing the input space to a huge number of sub-regions with a combination
features [2].
Recent analyses have proven efficient use of distributed representations in deep feed forward networks with rectified linear (ReL) activations [3, 4]. Such deep networks model complex input-output
relationships by dividing the input space to enormous number of sub-regions, that grow exponentially in the number of parameters. Multiple levels of feature representations in deep feed forward
networks successfully facilitate efficient reuse of low-level representations, and deep feed forward
networks thus can manage an exponentially greater number of sub-regions than shallow architectures.
It is interesting to ask whether deep generative models could attain such a property as deep discriminative models. To answer this question, it would be useful to compare restricted Boltzmann
machines (RBMs) and deep Boltzmann machine (DBMs). RBMs are a shallow generative model
with distributed representations [5]. Deep Boltzmann machines (DBM) are a deep extension of
RBMs [6]. DBMs are commonly expected to have a far greater representational power than RBMs
while being relatively easy to be trained compared to RBMs. However, the expectation of supremacy
of DBMs over RBMs has not ever been validated in a theoretical fashion.
In this paper, we provide both theoretical and empirical evidences that the representational power
of DBMs can actually be rather limited in exploiting the advantages of distributed representations.
1
2
2
Energy
0
0
2
2
4
4
6
8
1010
Figure 1: Illustration of
an sDBM. All the layer
pairs are connected in
different magnitudes of
strength.
F̂(v)
© ª
E(v, 0,0 )
© ª
E(v, 0,1 )
5
0
©
ª
6
E(v, 1,0 )
© ª
E(v, 1,1 )
F(v)
5
v
8
10
10
0
5
v
10
15
(a)
(b)
Figure 2: (a) Free energy F and hard-min free energy F̂ of (a)
a two-layered DBM and (b) sDBM (i.e., gBM(2)). Free energy
bounds are indicated with shaded regions. All the layers of BMs
have only one unit. The sDBM parameters are generated with Algorithm 1 and rescaling. (best view in color)
Our contributions are as follows. First, we propose an approximate measure for the efficiency of
distributed representations of BMs inspired by recent analysis on deep feedforward networks [3, 4].
Our measure is the number of linear regions of a piecewise linear function that approximates the free
energy function of a BM. This measure approximates the number of sub-regions that a BM manages
in the visible space. We show that the depth does not largely improve the representational power of a
DBM in terms of this measure. This indicates a surprising fact that DBMs can make inefficient use of
distributed representations, despite common expectations. Second, we propose a superset of DBMs,
which we dub soft-deep BMs (sDBMs). An sDBM is a layered BM where all the layer pairs are
connected with topologically defined regularization. Such relaxed connections realize soft hierarchy
as opposed to hard hierarchy of conventional deep networks where only neighboring layers are
connected. We show that the number of linear regions of the approximate free energy of an sDBM
scales exponential in the number of its layers thus can be as large as that of a general BM and can
be exponentially greater than that of an RBM or a DBM. Finally, we experimentally demonstrate
high generative performance of sDBMs. sDBMs trained without pretraining outperform state-ofthe-art generative models, including DBMs, on two benchmark datasets: MNIST and Caltech-101
silhouettes.
2
Soft-Deep BMs
We propose a soft-deep BM (sDBM): a Boltzmann Machine (BM) [7] that consists of multiple
layers where all the layer pairs are connected and connections within layers are restricted. Figure 1
illustrates an sDBM. The energy of an sDBM is defined as:
E(X ; θ sDBM ) = −
X
Nk X
Nl
X
0≤l<k≤L i=1 j=1
(k,l)
(k)
(k)
(k,l) (l)
xi Wi,j xj −
Nk
L X
X
(k) (k)
bi xi ,
(1)
k=0 i=1
(k)
where θ sDBM = {Wi,j , bi } is a set of parameters, x(k) = {xi } is the state of the k th layer,
and X = {v, H} is the set of all the units with H = {h(k) } being the set of hidden layers, and
v being the visible layer. We number layers x(k) s.t. x(0) is the visible layer, and x(k) is the k th
hidden layer. Let L be the number of layers, N be the total number of units, Nk be the number of
units in k th layer, Nvis be the number of visible units, and Nhid be the number of hidden units. An
sDBM assigns probability p(X ; θ) ∝ exp(−E(X ; θ)) to a configuration X ∈ {0, 1}N .
RBMs and DBMs are subsets of sDBMs; RBMs are sDBMs of L = 1, and DBMs are sDBMs where
(k,l)
Wi,j ≡ 0 for k − l > 1.
2
3
Quantifying the Efficiency of Distributed Representation in BMs
In the this section, we define an approximate measure for the representational power of a BM based
on its free energy function. We compare various BMs in terms of this measure and show that sDBMs
could attain richer representations than DBMs and RBMs.
The free energy of a BM is defined as the negative log probability that the network assigns to a
visible configuration without normalizing constant:
X
F (v; θ) , − log
exp(−E(v, H; θ)),
(2)
H
P
where H denotes summation over all the hidden configurations, and θ is the set of parameters. We
would be able to measure the representational power of a BM with the complexity of the free energy
function because the free energy function contains all the information on the probability distribution
that a BM models.
3.1
Hard-min Approximation of Free Energy
We here define a piecewise linear approximation of the free energy function of a BM. For RBMs,
it is widely known that the free energy function can be well approximated with a piecewise linear function [8]. This idea can be extended to general BMs that do P
not have connections between
visible units, which include sDBMs, as follows: the operation log H exp in Eq. (2) can be regarded
as a relaxed max operation; the sum is virtually dominated by the smallest energy, i.e.,
P
exp(−E(v,
H)) ≈ exp(− minH E(v, H)) where minH denotes min operation over all possiH
ble hidden configurations. The negative logarithm of the sum is thus nearly minH E(v, H). Based
on this observation, we define following approximation of the free energy:
Definition 1. Hard-min free energy F̂ of a BM with parameters θ is defined as:
F̂ (v; θ) , min E(v, H; θ).
H
(3)
Note that F̂ (v; θ) is a piecewise linear function if the BM does not have connections between visible
units because E(v, H; θ) does not have interactions involving multiple visible units.
Formally, we can show that F̂ bounds F as:
P
Theorem 2. Let Eres (v) = − log{ H exp(−E(v, H)) − exp(−F̂ (v))}. Then the free energy
F (v) is bounded as:
F̂ (v) − exp(F̂ (v) − Eres (v)) ≤ F (v) ≤ FMF (v) ≤ F̂ (v),
(4)
where FMF is the mean-filed approximation of the free energy.
The tightness of the bound is determined by the dominance of minimum energy F̂ over the free
energy. The difference between the upper and the lower bounds becomes fairly tight if F̂ is smaller
than Eres , the contribution of the non-minimum energies on the free energy.
Theorem 2 shows that F̂ is a very rough approximation for the free energy; F̂ is less accurate than
mean-field approximation FMF . Nevertheless, the bound can be tight except points where several
energy terms nearly achieve the minimum, e.g., boundaries between linear regions of F̂ . Figure 2
demonstrates this idea. Therefore, we will be able to roughly measure the complexity of the free
energy of a BM through quantifying the complexity of F̂ .
A natural way to quantify the complexity of a piecewise linear function is to count the number of
its linear regions. To quantify the representational power of a deep feedforward network with ReL
activation, this strategy was recently applied to the piecewise linear input-output function [3, 4].
Inspired by these analyses, we propose to use the number of the linear regions of F̂ to measure a
BM’s representational power. Intuitively, this measure roughly indicates the number of effective
Bernoulli mixing components of a BM; F̂ with k linear regions will be well approximated by the
negative probability function of a mixture of k Bernoulli components by assigning each component
to each region. We therefore shall call this measure the number of effective mixtures of a BM:
3
Definition 3. Suppose a BM with no connections between visible units. The number of effective
mixtures of the BM is the number of linear regions of the hard-min free energy F̂ of the BM.
Obviously from Definitions 1 and 3, the maximal number of effective mixtures of a BM is bounded
above by the number of its hidden configurations:
Proposition 4. The number of effective mixtures of a BM is upper bounded by 2Nhid .
Note that this proposition tells us nothing about whether this bound is actually achievable by a BM
with a certain parameter configuration; we provide positive results in later sections.
The number of effective mixtures of a BM approximately measures the efficiency of the distributed
representation. Each configuration of a distributed representation can give rise to a linear region
of F̂ . Therefore, an efficient distributed representation of a BM potentially manages 2Nhid subregions in the visible space. The efficiency, however, can substantially be damaged by restricted
connections.
For deep feedforward networks with ReL, Montúfar et al. [4] showed that a deeper network can
model a piecewise linear function with much more linear regions than a shallow network with the
same number of parameters. The number of the linear regions grows exponentially in the number
of the layers. Now we ask a question: is this also true for DBMs in terms of the approximate free
energy F̂ ? Surprisingly, the answer is NO. We shall provide proofs in the following sections.
3.2
The Number of Effective Mixtures of an RBM
We first analyze RBMs. The free energy function of an RBM can be approximated with a 2-layered
feedforward network with ReL [8]. The number of linear regions of the input-output function of
such a shallow network has been studied by Pascanu et al. [3] and Montúfar et al. [4]. With slight
modification on their results, we can compute the maximal number of effective mixtures of an RBM:
PN0 N1
Theorem 5. The maximal number of effective mixtures of an RBM is j=0
j .
Note that this
bound is quite smaller than the upper bound in Proposition 4 for N1 > N0 because
PN0 N1
=
Θ(N1 N0 ) 2N1 .
j=0 j
3.3
The Number of Effective Mixtures of a DBM
Next we analyze DBMs. Here we provide lower and upper bounds on the maximal number of
effective mixtures of DBMs. We have a lower bound because DBMs are a superset of RBMs:
Proposition
6. The maximal number of effective mixtures of a DBM is lower bounded by
PN0 N1
j=0 j .
A key idea of the proof on an upper bound, which we show in the appendix, is that energies associated with a same configuration in the first hidden layer h(1) have an identical gradient in the
space of v. For example, E(v, h(1) = •, h(2) = 0) and E(v, h(1) = •, h(2) = 1) have the
same gradient i.e., slope in Fig. 2 (a). This is because h(2) does not affect the statistics of v
given h(1) . The number of linear regions of F̂ is therefore bounded by 2N1 = 21 because one
of the energy terms with the same slope become globally smaller than the other energy term, e.g.,
E(v, h(1) = 0, h(2) = 0) < E(v, h(1) = 0, h(2) = 1) for any v. This generalize to any DBMs
leading to a natural but somewhat shocking result where the bound only depends on the number of
units in the first hidden layer:
Theorem 7. The number of effective mixtures of a DBM with any number of hidden layers is upper
bounded by 2N1 .
Depth does not largely help the number of effective mixtures of DBMs. This suggests that a distributed representation is inefficiently used in a DBM at least in the scope of the approximate free
energy F̂ . From Proposition 6 and Theorem 7, we can readily show a serious limitation on the
number of effective mixture of DBMs:
Proposition 8. The number of effective mixture of a DBM with N1 > N0 never achieves the bound
2Nhid .
4
-2
Energy
1
0.0
Energy
Energy
1
-4 2
0.5
Algorithm 1 A recursive construction of gBM(L)
-2 1
-8
1.0
function SOFT D EEP(L)
1
1 2
2 3
1.51.0 0.5 0.0 0.5
1
2
4
if L = 0 then
0
0
0
v
w(k,l) ← 0 for 0 ≤ l < k < ∞
(k)
b ← 0 for 0 ≤ k < ∞
(b)
(a)
return {w(k,l) }, {b(k) }
1.0
else
6.5
0.5
(k,l)
(k)
{w
}, {b } ← SOFT D EEP(L − 1)
16.0
2.0
x ← 2L−1
25.5
(L,0)
3.5
w
←x
35.0
(L)
1 0
1
2
1 1 3 5 7
b
← 0.5x(1 − x)
v
v
for l = 1 to L − 1 do
(L,l)
(l,0)
(c)
(d)
w
← −xw
end for
Figure 3: (a): gBM(L) for L ∈ {1, 2, 3} with
return {w(k,l) }, {b(k) }
unit indices and connection strengths. (b) to (d):
end if
F̂ (printed in black) and E(v, •) ∈ S(L) (in
end function
gray) with L = (b)1, (c)2, and (d)3.
3.4
The Number of Effective Mixtures of an sDBM
The key to the limited number of effective mixtures of DBMs is the independency between v and
{h(2) , h(3) , . . .} given h(1) . Conversely, if there exists dependency between the visible and the
upper hidden layers even given h(1) , the limitation over the number of effective mixtures will not
hold. Bypassing connections of sDBMs therefore might improve the number of effective mixtures.
Figure 2 (b) demonstrate this idea by showing that F̂ of an sDBM attains 2Nhid = 24 linear regions
with properly chosen parameters. In this section, we refine this idea for general sDBMs.
3.4.1
General BMs as Elemental sDBMs
We first analyze the number of effective mixtures of a general BM with only one visible unit, which
can be regarded as an elemental sDBM. Let gBM(L) be a general BM with L hidden units and one
visible unit whose energy function is defined as:
E(x(0:L) ; θ gBM(L) ) = −
X
0≤l<k≤L
x(k) w(k,l) x(l) −
L
X
b(k) x(k) ,
(5)
k=0
where we defined v = x(0) , h(k) = x(k) , and x(l:k) , {x(l) , . . . , x(k) } for 0 ≤ l < k ≤ L. Because
we regard a gBM(L) as an elemental sDBM with L layers each of which has only one unit, we index
units and parameters with superscripts. We may call the k th unit of a gBM(L) the k th layer because
of the same reason. Let S(L) be a set of one dimensional linear functions defined over the visible
unit: S(L) , {E(x(0) , x(1:L) ; θ gBM(L) )|x(1:L) ∈ {0, 1}L }.
We first analyze an arrangement of the elements of S(L) with a network construction procedure and
then analyze the number of linear regions of F̂ under this arrangement. The procedure is listed in
Algorithm 1, where a network is constructed by appending a unit in a recursive manner, starting
from the 1st unit to the Lth unit (see Fig. 3 (a) for example). With this construction, we can show
that all the elements of S(L) are arranged to be a tangent of a quadratic curve at 2L different points:
Lemma 9. Assume that {w(k,l) }, {b(k) } are computed with SOFT D EEP(M ) for a large integer M .
Then elements of S(L) for 0 < L ≤ M are tangents of a quadratic function with equally spaced
points of tangency.
From Lemma 9, we can readily show that:
Lemma 10. The number of effective mixtures of a gBM(L) reaches 2L = 2Nhid , the bound in
Proposition 4, when parameters are properly set.
5
Figures 3 from (b) to (d) demonstrate the statements of Lemma 9
and 10 for L ∈ {1, 2, 3}. Different hidden units control the slope
of the energy in different levels of magnitude (e.g., 1, 2, 4, 8, . . .).
This allows E(v, •) to have mutually different slopes and thus
leads to 2L effective mixtures.
We call connections determined with Algorithm 1 soft-deep connections because the connection strengths can be regarded to decay exponentially in the distance between layers. Let us have
units aligned in a sequential order k = 0, L, . . . , 1 as in Fig. 3 (a).
The strength of a connection from a unit to an upper unit which is
d units away from the unit under this spatial configuration is proportional to 2−d . We observe that this connection pattern is soft
Figure 4: Heatmap of F of a counterpart of the conventional deep connection pattern where
two-layered sDBM. The sDBM only adjacent layers are connected.
is constructed as a bundle of two
independent gBM(2)s. Lines
indicate the boundary of linear 3.4.2 Main Results
regions of F̂ . The parameters By applying Lemma 10 to an sDBM constructed as a bundle of
are computed with Algorithm 1 independent gBM(L)s, we can show that the maximal number of
and with rescaling.
effective mixtures of an L-layered sDBM scales exponentially in
L:
Theorem 11. Suppose an sDBM with L hidden layers each of
which contains M (≤ Nvis ) units. Then the number of effective mixtures of this sDBM reaches
2M L = 2Nhid , the bound in Proposition 4, with a certain parameter configuration.
Figure 4 demonstrates the claim of Theorem 11; the free energy function of an sDBM with four
hidden units can be well approximated with F̂ that has 24 = 16 linear regions.
Along with the analysis on DBMs and RBMs, Theorem 11 indicates that soft-deep connections that
bypass between remote layers can be vital for a deeply layered BM to have superior representational
power to shallow one in terms of the efficiency of a distributed representations. This clearly contrasts
with feedforward networks where bypassing connections do not critically affect the representational
power [9].
4
Remarks
There are two appealing properties of sDBMs other than the huge number of effective mixtures.
First, fast block Gibbs sampling can be performed. Although sampling efficiency degrades compared to DBMs due to the dependency between hidden layers introduced by soft-deep connections,
we believe that benefits from the huge representational power offset this negative effect.
Second, soft-deep connections can ease difficulties in learning deeply layered BMs. Because DBMs
do not have connections between remote layers, the effect that the visible layer exerts on remote
hidden layers decays exponentially in the depth. This phenomenon will hinder learning signals
from correctly propagating through deep layers. We believe that one of the benefits of pretraining
is to help this stochastic vanishing gradient effect. The soft-deep connections ease this problem
by bypassing between the visible layer and remote hidden layers. We believe that high generative
performance of sDBMs without pretraining shown in Section 5 is achieved not only with the huge
representational power proven in Theorem 11, but also with the less severe vanishing gradient effect.
4.1
Soft-Deep Regularization
For the number of effective mixtures of an sDBM to scale exponentially in the depth as in Theorem 11, it is essential that connection weights are in multiple levels of magnitude. Without regularization or with uniform regularization, networks do not attain such property via learning. To address
this point, we introduce soft-deep regularization where strength of L2 regularization for connections between k th and lth layers is inversely proportional to |w(k,l) |η where w(k,l) are computed
with Algorithm 1, and η is a hyper parameter. Although this technique does not strictly guarantee
6
that the weights scale as w(k,l) , we experimentally observed that this regularization improves the
performance of sDBMs.
5
Experiments
We have discussed representational power of BMs based on approximated free energy F̂ . To validate
the approximation, we experimentally demonstrated advantages of sDBMs. We performed experiments on two datasets: MNIST digits [10] and Caltech-101 silhouettes [11]. We used Theano [12]
and pylearn2 [13] to implement sDBMs. We used stochastic maximum likelihood [14] to jointly
train networks with the centering method [15] and soft-deep regularization. We did not perform
pretraining. We scheduled learning rates to linearly decay from an initial value to zero.
To evaluate networks, we used AIS [16, 17] to estimate the variational lower bound for the average
log-likelihood on test data. We evaluated the reliability of estimates by computing 3σ confidence
intervals, which we show in the supplementary material.
On both datasets, we trained 2-, 3-, and 4-layered sDBMs with various hyper parameters. The number of units in each hidden layer is fixed to 500. Hyper parameters are tuned via random sampling
[18]. See supplementary material for detail.
Figure 5: Random samples from a 4-layered sDBM (left in each cell) displayed with nearest training
(center) and test samples (right) from binarized MNIST. Generated images are probabilities that
pixels are sampled from. The nearest neighbors are computed in terms of pixelwise L2 distance.
The sDBM does not simply memorize training examples but generalize to unseen test examples.
5.1
Binarized MNIST
MNIST is a collection of gray scaled digit images that consists of 60,000 training samples and
10,000 test samples [10]. We binarized the images following the procedure by Salakhutdinov and
Murray [17] to generate training and test data.
Table 1 compares sDBMs and various models in the literature in terms of generative performance.
We can see that sDBMs greatly performed compared to other models. Even the 2-layered sDBM
outperformed the previous state-of-the-art test log-likelihood of −80.97 nat by a recent report [19].
The best-performing 4-layered sDBM achieved −66.56 nat of test log-likelihood with a 3σ confidence interval of [−67.01, −65.70] nat.
Note that sDBMs with 2- and 3-hidden layers outperformed DBMs with the same number of layers
[20]. This result would be seen to reflect the exponentially greater number of effective mixtures of
sDBMs than of DBMs with the same number of parameters.
The depth of networks largely improved the performance, though improvement of the 3-layered
model over the 2-layered model was relative small. We believe that this effect is due to insufficiency
7
Figure 6: Random samples from a 4-layered sDBM trained on Caltech-101 silhouettes displayed
with nearest training and test examples as in Fig. 5.
in parameter tuning; The 3-layered model performed worse than 2-layered model on training data
as shown in the supplementary material. The performance of models would uniformly improve as
the depth of networks with more precise parameter tuning.
Figure 5 shows random samples from the best performing sDBM. The sDBM well generalizes to
unseen test examples.
Table 1: Comparison of generative performance of various generative models on binarized MNIST.
We report average test log-likelihood measured in nat.
Model
RBM [17]
DBN 2hl [22]
DBM 2hl [20]
DBM 3hl [20]
NADE [24]
EoNADE 2hl [25]
EoNADE-5 2hl [26]
DLGM [27]
5.2
Test LL
≈ -86.34
≈ -84.55
≈ -83.43
≈ -83.02
-88.33
-85.10
-84.68
≈ -86.60
≥
Model
DLGM 8 leapfrogs[21]
DARN 1hl [23]
DARN 12hl [23]
DRAW [19]
sDBM 2hl
sDBM 3hl
sDBM 4hl
Test LL
≈ -85.51
≈ -84.13
–
–
–
–
–
≥
-88.30
-88.30
-87.72
-80.97
-76.41
-74.58
-66.56
Caltech-101 silhouettes
Caltech-101 silhouettes is a collection of binary silhouette images of various objects [11]. The
dataset contains 4,100 training samples and 2,307 test samples.
Table 2 compares sDBMs with several other models on generation of Caltech-101 silhouettes.
sDBMs outperformed the previous state-of-the-art by NADE trained with reweighted wake-sleep
(RWS) algorithm [28]. The 4-layered sDBM achieved −85.55 nat of test log-likelihood with a
3σ confidence interval of [−85.67, −85.40] nat. This result is the best average test log-likelihood
achieved on Caltech-101 silhouettes to the best of our knowledge.
The depth improves the performance of sDBMs. We believe that less precise parameter tuning
resulted in poor performance of 2-layered model as in experiments with MNIST.
Figure 6 shows samples generated from the best performing 4-layered model. The most samples
proves nice generalization by the network though some samples resemble training examples.
8
Table 2: Comparison of generative performance of various generative models on Caltech-101 silhouettes. We report average test log-likelihood as in Table 1.
Model
RBM [29]
RBM [29]
NADE-2 [26]
NADE-5[26]
6
Test LL
-109.0
-107.8
-108.8
-107.8
Model
NADE-RWS
sDBM 2hl
sDBM 3hl
sDBM 4hl
Test LL
-104.3
≥-92.4
≥-98.7
≥-85.5
Conclusion
In this paper, we proposed a BM architecture that can better exploit the distributed representation.
We proposed a measure for the efficiency of a distributed representation of a BM, the number of
effective mixtures of a BM, which is the number of linear regions of a piecewise linear function that
approximates a free energy function of a BM. We showed inefficiency of DBMs with respect to the
maximal number of effective mixtures. We proposed sDBMs, an extension of DBMs. We showed
that the maximal number of effective mixtures of an sDBM is exponentially larger than that of a
RBM or a DBM. Finally, we experimentally demonstrated high generative performance of sDBMs.
Acknowledgement
This research was supported by JSPS Grant-in-Aid for JSPS Fellows (145500000159). We thank
KyungHyun Cho, Li Yao, Takaki Makino, and Keita Tokuda for valuable discussion.
References
[1] Geoffrey E Hinton. Distributed Representations. In James L McClelland and David E Rumelhart, editors, Parallel Distributed Processing. MIT Press, 1986.
[2] Yoshua Bengio. Learning Deep Architectures for AI. Now Publishers Inc, October 2009.
[3] Razvan Pascanu, Guido Montúfar, and Yoshua Bengio. On the number of response regions of
deep feed forward networks with piece-wise linear activations. arXiv.org, December 2013.
[4] Guido Montúfar, Razvan Pascanu, KyungHyun Cho, and Yoshua Bengio. On the Number
of Linear Regions of Deep Neural Networks. In Advances in Neural Information Processing
Systems 27, 2014.
[5] Paul Smolensky. Information Processing in Dynamical Systems: Foundations of Harmony
Theory. In David E Rumelhart and James L McClelland, editors, Parallel Distributed Processing:Explorations in the Microstructure of Cognition: Foundations, pages 194–281. MIT press,
1986.
[6] Ruslan Salakhutdinov and Geoffrey Hinton. Deep Boltzmann machines. In Proceedings of the
12th International Conference on Artificial Intelligence and Statistics, pages 448–455, 2009.
[7] Geoffrey E Hinton and Terrence J Sejnowski. Learning and Relearning in Boltzmann Machines. In David E Rumelhart and James L McClelland, editors, Parallel Distributed Processing:Explorations in the Microstructure of Cognition: Foundations, pages 282–317. MIT press,
1986.
[8] James Martens, Arkadev Chattopadhyay, Toniann Pitassi, and Richard Zemel. On the Representational Efficiency of Restricted Boltzmann Machines. In Advances in Neural Information
Processing Systems 26, pages 1–21, 2013.
[9] Christopher M Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[10] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. In Proceedings of the IEEE, pages 2278–2324, 1998.
[11] Benjamin Marlin and Nando de Freitas. Asymptotic efficiency of deterministic estimators
for discrete energy-based models: Ratio matching and pseudolikelihood. arXiv preprint
arXiv:1202.3746, 2012.
9
[12] Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian Goodfellow, Arnaud
Bergeron, Nicolas Bouchard, David Warde-Farley, and Yoshua Bengio. Theano: new features
and speed improvements. arXiv.org, November 2012.
[13] Ian J Goodfellow, David Warde-Farley, Pascal Lamblin, Vincent Dumoulin, Mehdi Mirza,
Razvan Pascanu, James Bergstra, Frédéric Bastien, and Yoshua Bengio. Pylearn2: a machine
learning research library. arXiv.org, August 2013.
[14] Tijmen Tieleman. Training restricted Boltzmann machines using approximations to the likelihood gradient. In Proceedings of the 25th International Conference on Machine Learning,
pages 1064–1071. ACM, 2008.
[15] Grégoire Montavon and Klaus-Robert Muller. Deep Boltzmann machines and the centering
trick. In gregoire.montavon.name, pages 621–637. Springer, 2012.
[16] Radford M Neal. Annealed Importance Sampling. Statistics and Computing, 11:125–139,
2001.
[17] Ruslan Salakhutdinov and Iain Murray. On the quantitative analysis of deep belief networks.
In Proceedings of the 25th International Conference on Machine Learning. ACM, July 2008.
[18] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal
of Machine Learning Research, 2012.
[19] Karol Gregor, Ivo Danihelka, Alex Graves, and Daan Wierstra. DRAW: A Recurrent Neural Network For Image Generation. In Proceedings of the 32nd International Conference on
Machine Learning, 2015.
[20] Ruslan R Salakhutdinov and Geoffrey E Hinton. A better way to pretrain deep Boltzmann
machines. In Advances in Neural Information Processing Systems 25, pages 2447–2455, 2012.
[21] Tim Salimans, Diederik P Kingma, and Max Welling. Markov Chain Monte Carlo and Variational Inference: Bridging the Gap . In Proceedings of the 32nd International Conference on
Machine Learning, 2015.
[22] Iain Murray and Ruslan R Salakhutdinov. Evaluating probabilities under high-dimensional
latent variable models. In Advances in Neural Information Processing Systems 21, pages 1137–
1144, 2008.
[23] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep AutoRegressive Networks. In Proceedings of the 31st International Conference on Machine
Learning, 2014.
[24] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. Journal of
Machine Learning Research, 15:29–37, 2011.
[25] Benigno Uria, Iain Murray, and Hugo Larochelle. A Deep and Tractable Density Estimator. In
Proceedings of the 31st International Conference on Machine Learning, 2014.
[26] Tapani Raiko, Li Yao, KyungHyun Cho, and Yoshua Bengio. Iterative Neural Autoregressive
Distribution Estimator (NADE-k). In Advances in Neural Information Processing Systems 27,
2014.
[27] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic Backpropagation
and Approximate Inference in Deep Generative Models. In Proceedings of the 31st International Conference on Machine Learning, 2014.
[28] Jorg Bornschein and Yoshua Bengio. Reweighted Wake-Sleep. arXiv.org, June 2014.
[29] KyungHyun Cho, Tapani Raiko, and Alexander Ilin. Enhanced Gradient for Training Restricted Boltzmann Machines. Neural Computation, 25(3):805–831, 2013.
[30] Simon Osindero and Geoffrey E Hinton. Modeling image patches with a directed hierarchy
of Markov random fields. In Advances in Neural Information Processing Systems 21, pages
1121–1128, 2008.
[31] Nitish Srivastava and Ruslan Salakhutdinov. Multimodal Learning with Deep Boltzmann Machines. In Advances in Neural Information Processing Systems 25, 2012.
[32] Geoffrey E Hinton and Terrence J Sejnowski. Optimal perceptual inference. In Proceedings
of IEEE Conference on Computer Vision and Pattern Recognition, pages 448–453. IEEE New
York, 1983.
10
[33] Ruslan Salakhutdinov and Hugo Larochelle. Efficient learning of deep Boltzmann machines.
In AISTATs ’10, pages 693–700, 2010.
[34] Thomas Zaslavsky. Facing up to arrangements: face-count formulas for partitions of space by
hyperplanes. Mem. Amer. Math. Soc., 1(154):vii–102, 1975.
[35] Juergen Schmidhuber. Deep Learning in Neural Networks: An Overview. arXiv.org, May
2014.
[36] Honglak Lee, Chaitanya Ekanadham, and Andrew Y Ng. Sparse deep belief net model for
visual area V2. In Advances in Neural Information Processing Systems 20, 2007.
[37] Daniel J Felleman and David C Van Essen. Distributed Hierarchical Processing in the Primate
Cerebral Cortex. Cerebral Cortex, 1:1–47, 1991.
11
A
Boltzmann Machines
A Boltzmann machine (BM) is a stochastic generative model, which is typically defined over N
binary units xi ∈ {0, 1}. The probability that a BM assigns to a state X = {xi } is defined as
p(X ; θ) =
1
exp(−E(X ; θ)),
Z(θ)
(6)
where the normalization constant, or the partition function of the BM is denoted by Z(θ), and its
energy function is defined as
E(X ; θ) = −
N X
N
X
xi wi,j xj −
N
X
i=1 j=1
bi xi , s
(7)
i=1
where wi,j are symmetric (i.e., wi,j = wj,i , wi,i = 0) connection weights between units i and j, bi
are biases, and θ is the set of the parameters.
A BM with visible and hidden units can have rich representations; visible units vj ∈ v correspond
to data variables, and hidden units hi ∈ H correspond to latent features of data. All the units are
either a visible units or a hidden unit (i.e., X = {H, v}). The numbers of visible and hidden units
are denoted by Nvis and Nhid .
Various network topologies that restrict connections of BMs have attracted great research interests
[5, 6, 30, 31]. Albeit general BMs are a superset of such BMs with restricted connections, general
BMs are rarely used in practice. The main problem with general BMs is the difficulty due to the
intractability of the expectations with respect to data-dependent and model distributions. One approach is to approximate expectations via expensive MCMC [32]. The relaxation time of a Markov
chain can be quite long because general BMs have enormous number of well-separated modes to
be explored. Moreover, dense connections of BMs require generic Gibbs sampling which updates
only one unit at a time. Restriction on connections can alleviate these issues. Particularly, BMs
with layered connection patterns are widely studied because of their appealing properties such as
efficiency in sampling, less-complex energy landscapes, and simplicity of learning algorithms. We
here review two representative layered BMs: Restricted BMs (RBMs) and Deep BMs (DBMs).
A.1
RBMs
An RBM is a BM with a bipartite graph that consists of a visible layer and a hidden layer. Connections within each layer are restricted [5]. The energy function of an RBM is defined as
E({h(1) , v}; θ RBM ) = −
N1 X
N0
X
(1)
(1,0)
hi Wi,j vj −
i=1 j=1
N0
X
(0)
bj v j −
j=1
N1
X
(1) (1)
bi hi ,
(8)
i=1
(1,0)
(0)
(1)
where h(1) denotes the states of the (first) hidden layer, and θ RBM = {Wi,j , bj , bi } are model
parameters. We here use redundant notation with layer indices associated with a superscript to avoid
confusion of notations for models which we shall describe in later sections.
RBMs exhibit a nice property that conditional distributions p(h(1) |v) and p(v|h(1) ) are tractable
and factorized. This allows us to perform fast block Gibbs sampling and makes the data-dependent
expectation tractable.
However, such tractability substantially sacrifices the representation power of RBMs.
A.2
DBMs
DBMs are an extension of RBMs that have multiple hidden layers that form deep hierarchy. Connections within each layer are restricted, and units in a layer are connected to all the units in the
neighboring layers [6]. The energy function of a DBM with L layers is
E(X ; θ DBM ) = −
L−1
k+1 Nk
X NX
X
(k+1)
xi
k=0 i=1 j=1
(k+1,k) (k)
xj
Wi,j
−
Nk
L X
X
k=0 i=1
12
(k) (k)
bi xi ,
(9)
where we number layers s.t. the 0th layer is the visible layer, and the k th layer is the k th hidden
(k)
layer. The state of the k th layer is denoted by x(k) = {xi }, hence the state of the k th hidden layer
(k)
(k)
(k)
(k)
is h = x , (hi = xi ) for 0 < k ≤ L, and the state of the visible layer is v = x(0) , (vj =
PL
(0)
xj ). Let Nk be the number of units in k th layer (i.e., Nvis = N0 , Nhid = k=1 Nk ).
DBMs have several appealing properties. Fast block Gibbs sampling is also applicable to DBMs,
as to RBMs. Particularly, block sampling is highly efficient because conditional distributions of the
even layers given the odd layers and those of the odd layers given the even layers are tractable and
factorized. Moreover, DBMs possess greater representation power than RBMs because of multiple
hidden layers.
However, the improved representation power causes a serious difficulty. The data-dependent expectation needs to be approximated in learning because the conditional distribution p(H|v) is no longer
tractable; stochastic approximation procedure [14] or variational inference [6, 33] is used for approximation. At the appearance of DBMs, Salakhutdinov and Hinton [6] introduced a pre-training
algorithm to ease this problem. Recently, the centering method is proposed for joint training of
DBMs without pre-training [15].
Upon the introduction of DBMs, DBMs would have been expected to be scalable, i.e., great performance improvements can be achieved with DBMs by stacking a layer as in other deep neural
models. However, experiments suggest that this seems not true; improvements are hard to be gained
with very deep BMs with more than 3 hidden layers even with elaborated learning algorithms [6].
It is widely conceived that the poor scalability of DBMs is attributed that we cannot exploit huge
representation capacity of DBMs due to inefficient optimization methods. This will be true to some
extent. We, however, shall provide both empirical and theoretical evidences that the poor scalability
of DBMs is not only due to the optimization issues, but also because of rather limited representation
capacity of DBMs.
B
Proof of Theorems
P
Theorem 2. Let Eres (v) = − log{ H exp(−E(v, H)) − exp(−F̂ (v))}. Then the free energy
F (v) is bounded as:
F̂ (v) − exp(F̂ (v) − Eres (v)) ≤ F (v) ≤ FMF (v) ≤ F̂ (v),
(10)
where FMF is the mean-filed approximation of the free energy.
Proof. We first show the upper bound. For any approximating posterior Q(H|v), We have
F (v) =
X
H
Q(H|v)E(H, v) −
X
Q(H|v) log
H
1
− KL (Q(•|v)|P (•|v)),
Q(H|v)
(11)
where P (H|v) denotes the model’s true posterior distribution. Suppose we have a following appromximating posterior:
1 (H = Ĥ(v)) ,
Q̂(H|v) =
(12)
0 (otherwise)
where we defined Ĥ(v) = arg minH E(v, H). Note that this posterior factorizes. With this posterior, Eq. (11) becomes
F (v) = F̂ (v) − KL Q̂(•|v)|P (•|v) .
(13)
we have F (v) ≤ F̂ (v) because KL Q̂(•|v)|P (•|v) = − log P (Ĥ(v)|v) ≥ 0. The equality
holds if and only if the true posterior has its all the mass on Ĥ(v), i.e., P (H|v) = Q̂(H|v). We
have an inequality FMF (v) ≤ F̂ (v) readily from the definition of the mean-field free energy because
Q̂ is factorized.
13
Next, we prove the lower bound as:
F̂ (v) − F (v) = F̂ (v) + log
X
exp(−E({H, v}))
(14)
H
= log
exp(−F̂ (v)) + exp(−Eres (v))
!
(15)
exp(−F̂ (v))
= log 1 + exp(−Eres (v) + F̂ (v))
(16)
≤ exp(F̂ (v) − Eres (v)),
(17)
where we used log(x) ≤ x−1 in the last line. The equality holds if and only if F̂ (v) = Eres (v).
B.1
On The Number of Effective Mixtures of an RBM
Theorem 5. The maximal number of effective mixtures of an RBM is
PN0
j=0
N1
j
.
PN0 (0)
Proof. The hard-min free energy of an RBM can be written as F̂ (v; θ RBM ) = − j=1
bj vj −
PN1
PN0
(1,0)
(1)
i=1 max(0,
j=1 Wi,j vj + bi )). The number of linear regions of this function is the number
PN0
(1,0)
(1)
of regions separated by N1 hyper-planes each of them satisfies j=1
Wi,j vj + bi = 0 for
PN0 N1
0 < i ≤ N1 . The number of these regions is j=0 j [34]. This proves the claim.
B.2
On The Number of Effective Mixtures of a DBM
Here we provide lower and upper bounds for the maximal number of effective mixtures of DBMs
with respect to the parameters. Let us begin with a lower bound. Because DBMs are a superset
of RBMs, the number of effective mixtures of a DBM can be as large as the maximal number of
effective mixtures of RBMs. This observation leads us to a lower bound:
Proposition
6. The maximal number of effective mixtures of a DBM is lower bounded by
PN0 N1
.
j=0 j
We here outline the idea of the proof of Theorem 7.
A key observation is that the
marginal distribution over visible units of a DBM is written as a summation: p(v; θ DBM ) =
P
(1)
; θ DBM )p(h(1) ; θ DBM ) where p(h(1) ; θ DBM ) is the marginal distribution over h(1) .
h(1) p(v|h
This indicates that the number of the mixing components of a DBM is 2N1 ; this bounds the maximal number of effective mixtures from above . These observations lead to a natural but somewhat
shocking result where the bound only depends on the number of units in the first hidden layer:
Theorem 7. The number of effective mixtures of a DBM with any number of hidden layers is upper
bounded by 2N1 .
Proof. Suppose a set of linear functions S(h(1) ) = {E({h(L) , . . . , h(2) , h(1) , v})|h(k) ∈
N
{0, 1} k for 2 ≤ k ≤ L}. Linear functions within this set f ∈ S(h(1) ) have an identical
PN0
(1)
gradient as f (v; h(1) ) =
)vj + C where C is a constant that only depends on
j=1 αj (h
P
(1)
(1,0)
(0)
N1
{h(2) , . . . , h(L) } and αj (h(1) ) = − i=1
hi Wi,j − bj . Therefore, min S(h(1) ) is a linear
P
N0
function fmin (v; h(1) ) = j=1
αj (h(1) )vj + Cmin with Cmin = minh(2) ,...,h(L) C. The hard-min
free energy of a DBM is F̂ (v; θ DBM ) = minh(1) min S(h(1) ) = minh(1) fmin (v; h(1) ), and its
maximal number of linear regions is bounded above by the number of configurations of h(1) i.e.,
2N1 .
These results depict a serious limitation on the representation power of DBMs. There are two ways
to increase the number of effective mixtures of a DBM. The first way is to stack layers. However,
the number of effective mixtures never become greater than 2N1 , which is solely determined by N1 .
Therefore, depth does not largely help the capacity of DBMs measured in the number of effective
14
mixtures. The second way is to increase N1 . This strategy, however, at least necessitates the presence
of second layer units, which does not improve the bound 2N1 . Otherwise, the DBM is equivalent to
an RBM, and its maximal number of effective mixtures is merely Θ(N1 N0 ). Therefore, the number
of effective mixtures of a DBM is smaller than the upper bound in Proposition 4:
Proposition 8. The number of effective mixture of a DBM with N1 > N0 never achieves the bound
2Nhid .
Proof. First, suppose that the DBM has no hidden layers above the first layer. This DBM is equivalent to an RBM, thus from Theorem 5, the maximal number of effective mixtures of this DBM is
smaller than 2Nhid . Next, suppose that the DBM has more than one hidden units in its third hidden
layer with non-zero connection weights between units in the second hidden layer. From Theorem 7,
the number of effective mixtures of this DBM is bounded above by 2N1 < 2Nhid . This proves the
claim.
B.3
On The Number of Effective Mixtures of an sDBM
Lemma 9. Assume that {w(k,l) }, {b(k) } are computed with SOFT D EEP(M ) for a large integer M .
Then elements of S(L) for 0 < L ≤ M are tangents of a quadratic function with equally spaced
points of tangency.
Proof. We here show the claim with induction with a quadratic function
f (x(0) ) = −0.5(x(0) (x(0) + 1) + 0.25).
(18)
As in main text, let S(L) be a set of linear functions {E(x(0) , x(1:L) )|x(1:L) ∈ {0, 1}L }. Assume
that elements of S(L − 1) are a tangent of f (x(0) ) where the point of tangency is ξx(1:L−1) =
PL−1 (k) k−1
PL−1
2
−0.5, and the slope is − k=1 x(k) 2k−1 . We divide S(L) into two sets Sx(L) =0 (L)
k=1 x
and Sx(L) =1 (L), each of which is a set of lines that correspond to either x(L) = 0 or x(L) = 1.
We can readily show that elements of Sx(L) =0 (L) are a tangent of f (x(0) ) because S(L − 1) =
Sx(L) =0 (L).
We can show the tangency of elements of Sx(L) =1 (L) as follows. Let gx(L) =η (x(0) ; x(1:L−1) ) be an
element of Sx(L) =η (L) with hidden configuration x(1:L−1) for η ∈ {0, 1}, i.e.,
gx(L) =η (x(0) ; x(1:L−1) ) = E(x(0) , . . . , x(L) ; θ gBM(L) )|x(L) =η .
(19)
Let us consider the difference gx(L) =1 (x(0) ; x(1:L−1) ) − gx(L) =0 (x(0) ; x(1:L−1) ) = Bx(1:L−1) x(0) +
Cx(1:L−1) where Bx(1:L−1) and Cx(1:L−1) can be computed as follows:
Bx(1:L−1) = −w(L,0) = −2L−1 ,
(20)
X
(21)
and
Cx(1:L−1) = −
w(L,l) x(l) − b(L)
0<l<L
= −w(L,0)
X
x(l) 2l−1 + 0.5 (w(L,0) )2 − w(L,0)
(22)
0<l<L
= 0.5(w
(L,0) 2
) +w
(L,0)
L−1
X
!
x
(k) k−1
2
− 0.5
(23)
k=1
= 0.5(Bx(1:L−1) )2 + Bx(1:L−1) ξx(1:L−1) ,
(24)
where we used w(L,l) = w(L,0) 2l−1 .
From Eqs. 20 and 24, gx(L) =1 (x(0) ; x(1:L−1) ) is a tangent of f (x) because the difference between yintercepts cα and cα+β of two tangents of a quadratic function ax2 + bx + c with slopes α and α + β
2
is calculated as cα+β − cα = − β4a − βxα where xα is the point of tangency of the line with slope
PL
α. The point of tangency of gx(L) =1 (x(0) ; x(1:L−1) ) is ξx(1:L−1) + w(L,0) = k=1 x(k) 2k−1 − 0.5
15
PL−1
PL
and the slope is − k=1 x(k) 2k−1 − w(L,0) = − k=1 x(k) 2k−1 . Therefore, elements of S(L) are
a tangent of f (x(0) ) if elements of S(L − 1) are a tangent of f (x(0) ).
Observe that S(0) contains only one element g(x(0) ) = 0; this is a tangent of f (x(0) ) at the point of
tangency x(0) = −0.5. Therefore, elements of S(L) are a tangent of f (x(0) ) for any L ≤ M . This
proves the claim.
Lemma 10. The number of effective mixtures of a gBM(L) reaches 2L = 2Nhid , the bound in
Proposition 4, when parameters are properly set.
Proof. Assume a gBM(L) whose parameters are generated with SOFT D EEP(L). From Lemma 9,
an element of S(L) is a tangent of f (x(0) ) = −0.5(x(0) (x(0) + 1) + 0.25) at different points.
Because f is strictly concave, g(x(0) ) ≥ f (x(0) ) where g ∈ S(L) and the equality holds at the point
of tangency. Therefore, for ĝ, g ∈ S(L) (ĝ 6= g), ĝ(x̂) = f (x̂) < g(x̂) where x̂ is the point of
tangency of ĝ. Thus, at a neighbor of x̂, F̂ (x(0) ) = ĝ(x̂). Because elements of S(L) are tangents of
f (x(0) ) at 2L different points, the number of effective mixtures of the gBM(L) is 2L . This proves
the claim.
This result directly indicates that general BMs with more than one visible units can also achieve the
maximal number of effective mixtures 2Nhid with connection weights determined by our construction procedure where visible-hidden connections are replicated for all the visible units.
Figure 7: Illustration of an sDBM, which is used in the proof of Theorem 11.
Theorem 11. Suppose an sDBM with L hidden layers each of which contains M (≤ Nvis ) units.
Then the number of effective mixtures of this sDBM reaches 2M L = 2Nhid , the bound in Proposition 4, with a certain parameter configuration.
Proof. Assume an sDBM constructed as a collection of M independent gBM(L)s each of which has
PM
parameters θ gBM(L) generated with SOFT D EEP(L). Then, F̂ (v; θ sDBM ) = j=1 F̂ (vj ; θ gBM(L) ).
Because each F̂ (vj ; θ gBM(L) ) has 2L linear regions, the number of linear regions of F̂ (v; θ sDBM )
is 2M L . This proves the claim.
C
Connection to Biological Neural Nets
There has been increasingly more intense research interests on the connection between deep neural
networks and biological neural networks [35]. One prevalent aspect is that layers of deep neural
networks correspond to cortical regions that form hierarchy [36]. However, unlike conventional deep
networks, it is widely known that biological neural networks have many connections that bypass
between functionally remote cortical regions (e.g., between V1 and MT) [37]. Because bypassing
16
Energy
6.5
16.0
25.5
35.0 1
1
3
v
5
7
Figure 8: F̂ of gBM(3) displayed in a large size. F̂ is printed in black, and E(v, •) ∈ S(L) are in
gray.
connections do not largely contribute to the representation power of feedforward neural networks
[9], recent great success of deep feedforward networks do not explain the functional role of such
bypassing connections in our brain. Our results on sDBMs may help us to understand this mystery.
D
D.1
Details of Experiments
Parameters
We tuned hyper parameters via random sampling; initial learning rates were sampled from 10−[2,4]
for MNIST and from 10−[2.5,4.5] for Caltech-101 silhouettes, strengths of L2 regularization were
sampled from 10−[4,7] , η was sampled from [0.5, 3.5], and update constants for the centering parameters were sampled from 10−[5,8] . We generated 16 configurations of hyper parameters for each
experiment setting. The number of parameter updates was 106 .
Networks were trained with stochastic maximum likelihood [14]. We did not perform variational
inference [6]. The number of positive phase Markov chain updates per parameter update was 5 for
MNSIT and 1 for Caltech-101 silhouettes. The number of negative phase Markov chain updates per
parameter update was 5. The batch size was set to 100.
D.2
AIS
Throughout the training, we monitored the training and test log-likelihood of models by occasionally
performing AIS. Such monitoring AIS was executed with rather cheap settings of 100 runs and
30,000 intermediate distributions. After training, we performed more expensive AIS on several
best performing models evaluated via cheap AIS to gain thorough estimates. This expensive AIS
17
Table 3: Details of AIS estimates for sDBMs trained on MNIST. Estimated variational lower bounds
on training and test data are reported. 3σ confidence intervals are also reported in parentheses.
Model
sDBM 2hl
sDBM 3hl
sDBM 4hl
-68.80
-71.17
-61.90
Train LL
(-68.91,-68.67)
(-71.58,-70.48)
(-62.36,-61.04)
-76.41
-74.58
-66.56
Test LL
(-76.53,-76.28)
(-74.98,-73.89)
(-67.01,-65.70)
Table 4: Details of AIS estimates for sDBMs trained on Caltech-101 silhouttes as in Table 3
Model
sDBM 2hl
sDBM 3hl
sDBM 4hl
-30.16
-72.62
-38.16
Train LL
(-30.35,-29.92)
(-72.69,-72.56)
(-38.29,-38.02)
-92.37
-98.66
-85.55
Test LL
(-92.56,-92.13)
(-98.72,-98.59)
(-85.67,-85.40)
is executed with at least 1,000 runs and at least 300,000 intermediate distributions. All the figures
reported in the main text were gained with such expensive AIS.
D.3
Samples from sDBMs
Figures 9 and 10 show consecutive samples from the best-performing 4-layered sDBMs. These
figures demonstrate nice mixing of Markov chains between several classes.
Figure 9: Consecutive samples generated from a 4-layered sDBM trained on MNIST.
18
Figure 10: Consecutive samples generated from a 4-layered sDBM trained on Caltech-101 silhouettes.
19
| 9cs.NE
|
arXiv:1603.03836v1 [cs.DS] 12 Mar 2016
Near-Isometric Binary Hashing for Large-scale Datasets
Amirali Aghazadeh
Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA
AMIRALI @ RICE . EDU
Andrew Lan
Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA
MR . LAN @ SPARFA . COM
Anshumali Shrivastava
Department of Computer Science, Rice University, Houston, TX, USA
Richard Baraniuk
Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA
Abstract
We develop a scalable algorithm to learn binary hash codes for indexing large-scale datasets.
Near-isometric binary hashing (NIBH) is a datadependent hashing scheme that quantizes the
output of a learned low-dimensional embedding to obtain a binary hash code. In contrast
to conventional hashing schemes, which typically rely on an `2 -norm (i.e., average distortion) minimization, NIBH is based on a `∞ norm (i.e., worst-case distortion) minimization
that provides several benefits, including superior
distance, ranking, and near-neighbor preservation performance. We develop a practical and efficient algorithm for NIBH based on column generation that scales well to large datasets. A range
of experimental evaluations demonstrate the superiority of NIBH over ten state-of-the-art binary
hashing schemes.
1. Introduction
Hashing, one of the primitive operations in large-scale systems, seeks a low-dimensional binary embedding of a highdimensional data set. Such a binary embedding can increase the computational efficiency of a variety of tasks,
including searching, learning, near-neighbor retrieval, etc.
One of the fundamental challenges in machine learning is
the development of efficient hashing algorithms that embed
data points into compact binary codes while preserving the
geometry of the original dataset.
ANSHUMALI @ RICE . EDU
RICHB @ RICE . EDU
In this paper, we are interested in learning near-isometric
binary embeddings, i.e., hash functions that preserve the
distances between data points up to a given distortion in
the Hamming space. More rigorously, let Z and Y denote
metric spaces with metrics dZ and dY , respectively. An
embedding f : Z → Y is called near-isometric (Plan &
Vershynin, 2014, Def. 1.1) if, for every pair of data points
zi , zj ∈ Z, we have
dZ (zi , zj ) − γ ≤ dY (f (zi ), f (zj )) ≤ dZ (zi , zj ) + γ,
where γ is called the isometry constant. In words, f is nearisometric if and only if the entries of the pairwise-distortion
vector containing the distance distortion between every pair
of data points |dZ (zi , zj ) − dY (f (zi ), f (zj ))|, ∀i > j do
not exceed the isometry constant γ. A near-isometric embedding is approximately distance-preserving in that the
distance between any pairs of data points in the embedded
space Y is approximately equal to their distance in the ambient space Z (Hegde et al., 2015; Shaw & Jebara, 2007;
Weinberger & Saul, 2006).
The simplest and most popular binary hashing scheme,
random projection, simply projects the data into a lowerdimensional (lower-bit) random subspace and then quantizes to binary values. Random projections are known to be
near-isometric with high probability, due to the celebrated
Johnson-Lindenstrauss (JL) lemma (Andoni et al., 2014;
Datar et al., 2004; Plan & Vershynin, 2014). Algorithms
based on the JL lemma belong to the family of probabilistic dimensionality reduction techniques; a notable example
is locality sensitive hashing (LSH) (Andoni et al., 2014;
Datar et al., 2004). Unfortunately, theoretical results on
LSH state that the number of bits required to guarantee an
isometric embedding can be as large as the number of data
points (Datar et al., 2004; Plan & Vershynin, 2014). Even
Near-Isometric Binary Hashing for Large-scale Datasets
(a)
in practice, LSH’s requirement on the number of bits is impractically high for indexing many real-world, large-scale
datasets (Lv et al., 2007).
Consequently, several data-dependent binary hashing algorithms have been developed that leverage the structure of
the data to learn compact binary codes. These methods enable a significant reduction in the number of bits required
to index large-scale datasets compared to LSH; see (Wang
et al., 2014) for a survey. However, learning compact binary codes that preserve the local geometry of the data remains challenging.
These data-dependent hashing algorithms focus on the
choice of the distortion measure. Typically, after finding
the appropriate distortion measure, the hash functions are
learned by minimizing the average distortion, i.e., the `2 norm of the pairwise-distortion vector, which sums the distortion among all pairwise distances with equal weights.
Binary reconstructive embedding (BRE) (Kulis & Darrell,
2009), for example, uses an optimization algorithm to directly minimize the average distortion in the embedded
sapce. Spectral Hashing (SH) (Weiss et al., 2009), Anchor
Graph Hashing (AGH) (Liu et al., 2011), Multidimensional
Spectral Hashing (MDSH) (Weiss et al., 2012), and Scalable Graph Hashing (SGH) (Jiang & Li, 2015) define notions of similarity based on a function of `2 -distance between the data points kzi − zj k2 and use spectral methods
to learn hash functions that keep similar points sufficiently
close. Some other hashing algorithms first project the data
onto its principal components, e.g., PCA Hashing (PCAH)
(Jolliffe, 2002), which embeds with minimal average distortion, and then learn a rotation matrix to minimize the
quantization error (Gong & Lazebnik, 2011) or balance the
variance across the components (Isotropic Hashing (IsoHash) (Kong & Li, 2012)).
While minimizing the average distortion seems natural, this
approach can sacrifice the preservation of certain pairwise
distances in favor of others. As we demonstrate below, this
can lead to poor performance in certain applications, such
as the preservation of nearest neighbors. In response, in
this paper, we develop a new data-driven hashing algorithm
that minimizes the worst-case distortion among the pairwise distances, i.e., the `∞ -norm of the pairwise-distortion
vector.
Figure 1 illustrates the advantages of minimizing the
`∞ -norm of the pairwise-distortion vector instead of
its `2 -norm. Consider three clusters of points in a
two-dimensional space. We compute the optimal onedimensional embeddings of the data points by minimizing
the `∞ -norm and the `2 -norm of the pairwise-distortion
vector using a grid search over the angular orientation of
the line that represents the embedding. We evaluate the
near-neighbor preservation of a given query point in the
Cluster 1
Cluster 2
Cluster 3
L∞
L2
(b)
10
NN(q) = [1-2-3-4-5-6-7-8-9-10]
NN(q) = [10-4-2-1-8-6-3-7-5-9]
0
q = (0,0)
NN(q) = [1-2-3-4-5-6-7-8-9-10]
0
100
200
Figure 1. Comparison of the near-neighbor (NN) preservation
performance of hashing based on minimizing the `∞ -norm
(worst-case distortion) vs. the `2 -norm (average distortion) of
the pairwise distance distortion vector on an illustrative data set.
(a) For a dataset with three clusters (5 circles, 5 squares, and 60
stars), we found the optimal embeddings for both error metrics using grid search. (b) The projection of the data points using the `∞ optimal embedding preserves three well-separated clusters; however the projection using the `2 -optimal embedding mixes circles
and squares, projecting them into a single cluster. For the query
point q = (0, 0), all of its nearest neighbors NN(q) are preserved
with the correct ordering using the worst-based distortion embedding but not the average distortion embedding.
embedded space (shown in Fig. 1 (b)). For a query point
q, located without loss of generality at the origin, the nearest neighbor ranking from the ambient space is destroyed
by the `2 -optimal embedding, since the circle and square
clusters overlap. In contrast, the `∞ -optimal embedding
exactly preserves the rankings.
This illustration emphasizes the importance of preserving
relevant distances in data retrieval tasks. To minimize
the average distortion, the `2 -optimal embedding (dashed
red line) sacrifices the square–circle distances in favor of
the square–star and circle–star distances, which contribute
more to the `2 -norm of the pairwise distance distortion vector. In contrast, the `∞ -optimal embedding focuses on the
hardest distances to preserve (i.e., the worst-case distortion), leading to an embedding with smaller isometry constant than the `2 -optimal embedding. Preservation of these
distances is critical for near-neighbor retrieval.
1.1. Contributions
We make four distinct contributions in this paper. First,
conceptually, we advocate minimizing the worst-case distortion, which is formulated as an `∞ -norm minimization problem, and show that this approach outperforms approaches based on minimizing the average, `2 -norm distortion in a range of computer vision and learning scenarios
(Hartley & Schaffalitzky, 2004).
Second, algorithmically, since `∞ -norm minimization
problems are computationally challenging, especially for
large datasets, we develop two accelerated and scalable al-
Near-Isometric Binary Hashing for Large-scale Datasets
gorithms to find the optimal worst-case embedding. The
first, near-isometric binary hashing (NIBH), is based on
the alternating direction method of multipliers (ADMM)
framework (Boyd et al., 2011). The second, NIBH-CG, is
based on an accelerated greedy extension of the NIBH algorithm using the concept of column generation (Dantzig
& Wolfe, 1960). NIBH-CG can rapidly learn hashing functions from large-scale data sets that require the preservation
of billions of pairwise distances (e.g., MNIST).
Third, theoretically, since current data-dependent hashing
algorithms do not offer any probabilistic guarantees in
terms of preserving near-neighbors, we develop new theory
to prove that, under natural assumptions regarding the data
distribution and with a notion of hardness of near-neighbor
search, NIBH preserves the nearest neighbors with high
probability. Our analysis approach could be of independent interest for obtaining theoretical guarantees for other
data-dependent hashing schemes.
Fourth, experimentally, we demonstrate the superior performance of NIBH as compared to ten state-of-the-art binary hashing algorithms using an exhaustive set of experimental evaluations involving six diverse datasets and three
different performance metrics (near-isometry, Hamming
distance ranking, and kendall τ ranking performance). In
particular, we show that NIBH achieves the same distance
preservation and Hamming ranking performance as stateof-the-art algorithms while using up to 60% fewer bits. Our
experiments clearly show the superiority of the `∞ -norm
formulation over the more classical `2 -norm formulation
that underlies many hashing algorithms, such as BRE and
IsoHash. Our formulation also outperforms recently developed techniques that assume more structure in their hash
functions, such as Spherical Hamming Distance Hashing
(SHD) (Heo et al., 2012) and Circulant Binary Embedding
(CBE) (Yu et al., 2014).
2. Near-Isometric Binary Hashing
The standard formulation for data dependent binary hash
function embeds a data point x ∈ RN into the lowdimensional Hamming space H = {0, 1}M by first multiplying it by an embedding matrix W ∈ RM ×N and then
quantizing the entries of the product Wx to binary values:
h(Wx) =
1 + sgn(Wx)
.
2
(1)
The function sgn(·) operates element-wise on the entries of
Wx, transforming the real-valued vector Wx into a set of
binary codes depending on the sign of the entries in Wx.
2.1. Problem formulation
Consider the design of an embedding f that maps Q
high-dimensional data vectors X = {x1 , x2 , . . . , xQ } in
the ambient space RN into low-dimensional binary codes
H = {h1 , h2 , . . . , hQ } in the Hamming space with hi ∈
{0, 1}M , where hi = f (xi ), i = 1, . . . , Q, and M N .
Define the distortion of the embedding by
δ = inf sup |λdH (hi , hj ) − d(xi , xj )|,
λ>0 (i,j)∈Ω
with Ω = {(i, j) : i, j ∈ {1, 2, . . . , Q}, i > j},
where d(xi , xj ) denotes the Euclidean distance between
the data points xi , xj , dH (hi , hj ) denotes the Hamming
distance between the binary codes hi and hj , and λ is a
positive scaling variable. The distortion δ measures the
worst-case deviation from perfect isometry (i.e., optimal
distance preservation) among all pairs of data points. Define the secant set S(X ) as S(X ) = {xi −xj : (i, j) ∈ Ω},
i.e., the set of all pairwise difference vectors in X . Let
|S(X )| = |Ω| = Q(Q − 1)/2 denote the size of the secant set. Note that the common distortion measure utilized
in
average distortion, i.e.,
P other hashing algorithms is the
2
(λd
(h
,
h
)
−
d(x
,
x
))
/|Ω|.
H
i
j
i
j
i>j
We formulate the problem of minimizing the distortion parameter δ as the following optimization problem:
2
minimize sup λkh(Wxi ) − h(Wxj )k2 − kxi − xj k2 ,
W,λ>0 (i,j)∈Ω
since the squared `2 -distance between a pair of binary
codes is equivalent to their Hamming distance up to a scaling factor that can be absorbed into λ. We can rewrite the
above optimization problem as
(P∗ ) minimize kλv0 (W) − ck∞ ,
W,λ>0
where v0 ∈ NQ(Q−1)/2 is a vector containing the pairwise
Hamming distances between the embedded data vectors
2
kh(xi ) − h(xj )k2 , and c is a vector containing the pairwise `2 -distances between the original data vectors. Intuitively, the `∞ -norm objective optimizes the worst-case
distortion among all pairs of data points.
The problem (P∗ ) is a combinatorial problem with complexity O(Q2M ). To overcome the combinatorial nature of
the problem, we approximate the hash function h(·) by the
sigmoid function (also known as the inverse logit link function) σ(x) = (1 + e−x )−1 . This enables us to approximate
(P∗ ) by the following optimization problem:
(P) minimize kλv(W) − ck∞ ,
W,λ>0
Q(Q−1)/2
where v ∈ R+
is a vector containing the pairwise
`2 distances between the embedded data vectors after sig2
moid relaxation (1 + e−Wxi )−1 − (1 + e−Wxj )−1 2 .
Here the sigmoid function operates element-wise on Wxi .
In practice we use a more general definition of the sigmoid
Near-Isometric Binary Hashing for Large-scale Datasets
function, defined as σα (x) = (1 + e−αx )−1 , where α is the
rate parameter controlling how closely it approximates the
non-smooth function h(·). The following lemma characterizes the quality of such an approximation (see the Appendix
for a proof).
problem:
Lemma 1. Let x be a Gaussian random variable as x ∼
N (µ, σ 2 ). Define the distortion of the sigmoid approximation at x as |h(x) − σα (x)|. Then, the expected distortion is bounded as Ex [|h(x) − σα (x)|] ≤ σ√12πα +
where ρ is the scaling parameter in ADMM and y ∈
RQ(Q−1)/2 is the Lagrange multiplier vector. The NIBH
algorithm proceeds as follows. First, the variables W, λ, u,
and Lagrange multipliers y are initialized randomly. Then,
at each iteration, we optimize over each of the variables
W, u, and λ while holding the other variables fixed. More
specifically, in iteration `, we perform the following four
steps until convergence:
minimize kuk∞
W,u,λ>0
subject to u = λv(W) − c.
(2)
The augmented Lagrangian form of this problem can be
2
written as minimize kuk∞ + ρ2 ku − λv(W) + c + yk2 ,
W,u,λ>0
√
2
2e−( α+c/ασ ) , where c is a positive constant. As α goes
to infinity, the expected distortion goes to 0.
Remark. As has been noted in the machine vision literature (Zoran & Weiss, 2012), a natural model for an image
database is that its images are generated from a mixture of
Gaussian distributions. Lem. 3 bounds the deviation of the
sigmoid approximation from the non-smooth hash function
(1) under this model.
2.2. Near-isometry and nearest neighbor preservation
Inspired by the definition of relative contrast in (He et al.,
2012), we define a more generalized measure of data separability to preserve k-NN that we call the k-order gap
∆k := d(x0 , xk+1 ) − d(x0 , xk ), where x0 is a query point
and xk and xk+1 are its k th and k + 1th nearest neighbors,
respectively. We formally show that if the data is highly
separable (∆k is large), then the above approach preserves
all k nearest neighbors with high probability (see the Appendix for a proof and discussion).
• Optimize
over
W
via
W(`+1)
←
(`)
1
1 P
1
(`)
2
arg min 2 (i,j)∈Ω (uij −λ k 1+e−Wxi − 1+e−Wxj k2 +
W
2
(`)
kxi − xj k2 − yij )2 , where λ(`) denotes the value of λ in
(`)
(`)
the `th iteration. We also use uij and yij to denote the
entries in u(`) and y(`) that correspond to the pair xi and
xj in the dataset X . We show in our experiments below
that using the accelerated first-order gradient descent
algorithm (Nesterov, 2007) to solve this subproblem
results in good empirical convergence performance (see
the Appendix).
• Optimize over u while holding the other variables fixed;
it corresponds to solving the proximal problem of the
`∞ -norm u(`+1) ← arg min kuk∞ + ρ2 ku − λ(`)v(`+1) +
u
c + y(`) k22 . We use the low-cost algorithm described in
(Studer et al., 2014) to perform the proximal operator update.
Theorem 2. Assume that all the data points are independently generated
distribution
PP from a mixture of Gaussian N
i.e., xi ∼
denote a
p=1 πp N (µp , Σp ). Let x0 ∈ R
query data point in the ambient space, and the other data
points xi be ordered so that d(x0 , x1 ) < d(x0 , x2 ) <
. . . < d(x0 , xQ ). Let δ denote the final value of the distortion parameter computed from any binary hashing algorithm, and let»c denote a positive constant. Then, if
Ex [∆k ] ≥ 2δ + 1c log Qk
, the binary hashing algorithm
preserves all the k-nearest neighbors of a data point with
probability at least 1 − .
• Update y via y(`+1)←y(`) + η(u(`+1) − λ(`+1) v(`+1) +
c), where the parameter η controls the dual update step
size.
2.3. The NIBH algorithm
2.4. Accelerated NIBH for large-scale datasets
We now develop an algorithm to solve the optimization
problem (P). We apply the alternating direction method
of multipliers (ADMM) framework (Boyd et al., 2011) to
construct an efficient algorithm to find a (possibly local)
optimal solution of (P). Note that (P) is non-convex, and
therefore no standard optimization method is guaranteed to
converge to a globally optimal solution in general. We introduce an auxiliary variable u to arrive at the equivalent
The ADMM-based NIBH algorithm is efficient for smallscale datasets (e.g., for secant sets of size |S(X )| < 5000
or so). However, the memory requirement of NIBH is
quadratic in |X |, which would be problematic for applications involving large-scale numbers of data points and
secants. In response, we develop an algorithm that approximately solves (P) while scaling very well to largescale problems. The key idea comes from classical results
• Optimize over λ while holding the other variables fixed;
it corresponds to a positive least squares problem, where
λ is updated as λ(`+1)←arg min 21 ku(`+1) −λv(`+1) +c+
λ>0
y(`) k22 . We perform this update using the non-negative
least squares algorithm (Kim & Park, 2007).
Near-Isometric Binary Hashing for Large-scale Datasets
in optimization theory related to column generation (CG)
(Dantzig & Wolfe, 1960; Hegde et al., 2015).
The optimization problem (2) is an `∞ -norm minimization problem with an equality constraint on each secant.
The Karush-Kuhn-Tucker (KKT) condition for this problem states that, if strong duality holds, then the optimal solution is entirely specified by a (typically very small) portion of the constraints. Intuitively, the secants corresponding to these constraints are the pairwise distances that are
harder to preserve in the low-dimensional Hamming space.
We call the set of such secants the active set. In order to
solve (P), it suffices to find the active secants and solve
NIBH with a much smaller number of active constraints.
To leverage the idea of the active set, we iteratively run
NIBH on a small subset of all the secants that violate the
near-isometry condition, as detailed below:
• Solve (P) with a small random subset S0 of all the sec δ̂, and λ̂, initial
cants S(X ) using NIBH to obtain W,
estimates of the parameters. Identify the active set Sa .
Fix λ = λ̂ for the rest of the algorithm.
• Randomly select a new subset Sv ⊂ S of secants that
violate the near isometry condition using the current
c δ̂, and λ̂. Then, form an augmented
estimates of W,
secant set Saug = Sa ∪ Sv .
• Solve (P) with the secants in the set Saug using the
NIBH algorithm.
We dub this approach NIBH-CG. NIBH-CG iterates over
the above steps until no new violating secants are added
to the active set. Since the algorithm searches over all the
secants for violating secants in each iteration before terminating, NIBH-CG ensures that all of the constraints are
satisfied when it terminates. A key benefit of NIBH-CG
is that only the set of active secants (and not all secants)
needs to be stored in memory. This benefit leads to significant improvements in terms of memory complexity over
competing algorithms, since the set of all secants quickly
becomes large-scale and can exceed the system memory
capacity in large-scale applications.
3. Experiments
In this section, we validate the NIBH and NIBH-CG algorithms via experiments using a range of synthetic and realworld datasets, including three small-scale, three mediumscale, and one large-scale datasets with respect to three
metrics. We compare NIBH against ten state-of-the-art
binary hashing algorithms, including binary reconstructive embedding (BRE) (Kulis & Darrell, 2009), spectral
hashing (SH) (Weiss et al., 2009), anchor graph hashing
(AGH) (Liu et al., 2011), multidimensional spectral hashing (MDSH) (Weiss et al., 2012), scalable graph hashing
(SGH) (Jiang & Li, 2015), PCA hashing (PCAH) (Jolliffe,
2002), isotropic hashing (IsoHash) (Kong & Li, 2012),
spherical Hamming distance hashing (SHD) (Heo et al.,
2012), circulant binary embedding (CBE) (Yu et al., 2014),
and locality-sensitive hashing (LSH) (Indyk & Motwani,
1998).
3.1. Performance metrics and datasets
We compare the algorithms using the following three metrics:
Maximum distortion δ = inf ||λv̂ − c||∞ , where the vector
λ>0
v̂ contains the pairwise Hamming distances between the
learned binary codes. This metric quantifies the distance
preservation among all of the pairwise distances after projecting the training data in the ambient space into binary
codes. We also define the maximum distortion for unseen
test data δtest , which measures the distance preservation on
a hold-out test dataset using the hash function learned from
the training dataset.
Mean average precision (MAP) for near-neighbor preservation in the Hamming space. MAP is computed by first
finding the set of k-nearest neighbors for each query point
on a hold-out test data in the ambient space Lk and the
corresponding set LkH in the Hamming space and then calculating the average precision AP = |Lk ∩ LkH |/k. We then
report MAP by calculating the mean value of AP across all
data points.
Kendall τ ranking correlation coefficient. We first rank the
set of k-nearest neighbors for each data point by increasing
distance in the ambient space as T (Lk ) and in the Hamming space as T (LkH ). The Kendall τ correlation coefficient is a scalar τ ∈ [−1, 1] that measures the similarity
between the two ranked sets T (Lk ) and T (LkH ) (Kendall,
1938). The value of τ increases as the similarity between
the two rankings increases and reaches the maximum value
of τ = 1 when they are identical. We report the average
value of τ across all data points in the training dataset.
To compare the algorithms, we use the following standard
datasets from computer vision: Random consists of independently drawn random vectors in R100 from a multivariate Gaussian distribution with zero mean and identity covariance matrix. Translating squares is a synthetic
dataset consisting of 10 × 10 images that are translations
of a 3 × 3 white square on black background (Hegde et al.,
2015). MNIST is a collection of 60,000 28 × 28 greyscale
images of handwritten digits (LeCun & Cortes, 1998).
Photo-Tourism is a corpus of approximately 300,000 image patches, represented using scale-invariant feature transform (SIFT) features (Lowe, 2004) in R128 (Snavely et al.,
Near-Isometric Binary Hashing for Large-scale Datasets
CBE
NIBH
NIBH−CG
AGH
SGH
BRE
Random
3
2.5
δ
Translating squares
3
2
2
2
1.5
1.5
1
1
1
0.5
0.5
0.5
MAP
30
40
50
60
70
1
10
20
30
40
50
60
70
10
20
30
40
50
60
70
1
10
20
30
40
50
60
70
0 1
10
20
30
40
50
60
70
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0.8
1
10
20
30
40
50
60
70
0
0.8
1
10
20
30
40
50
60
70
0
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
1
10
20
30
40
50
60
70
0
MNIST
1
1
1
0
Kendall τ
20
IsoHash
LSH
2.5
1.5
10
PCAH
SH
MDSH
3
2.5
1
SHD
1
10
20 30 40 50 60
Number of bits (M)
70
Figure 2. Comparison of the NIBH and NIBH-CG algorithms against several baseline binary hashing algorithms using three small-scale
datasets with 4950 secants (Q = 100). The performance of NIBH-CG closely follows that of NIBH, and both outperform all of the other
algorithms in terms of the maximum distortion δ (superior distance preservation), mean average precision MAP of training samples
(superior nearest neighbor preservation), and Kendall τ rank correlation coefficient (superior ranking preservation).
2006). LabelMe is a collection of over 20,000 images represented using GIST descriptors in R512 (Torralba et al.,
2008). Peekaboom is a collection of 60,000 images represented using GIST descriptors in R512 (Torralba et al.,
2008). Following the experimental approaches of the hashing literature (Kulis & Darrell, 2009; Norouzi & Fleet,
2011), we pre-process the data by subtracting the mean and
then normalizing all points to lie on the unit sphere.
3.2. Small- and medium-scale experiments
We start by evaluating the performance of NIBH and
NIBH-CG using a small-scale subset of the first three
datasets. Small-scale datasets enable us to compare the
performance of NIBH vs. NIBH-CG to verify that they perform similarly. Also they help us assess the asymptotic behavior of algorithms in preserving isometry since the total
of number of secants are small compare to the bit budget in
compact binary codes.
Experimental setup. We randomly select Q = 100 data
points from the Random, Translating squares, and MNIST
datasets. We then apply the NIBH, NIBH-CG, and all the
baseline algorithms on each dataset for different target binary code word lengths M from 1 to 70 bits. We set the
NIBH and NIBH-CG algorithm parameters to the common
choice of ρ = 1 and η = 1.6. To generate hash function of length M for LSH, we draw M random vectors
from a Gaussian distribution with zero mean and an identity covariance matrix. We use the same random vectors
to initialize NIBH and other baseline algorithms. In the
near-neighbor preservation experiments, to show the direct
advantage of minimizing `∞ -norm over `2 -norm, we followed the exact procedure described in BRE (Kulis & Darrell, 2009) to select the training secants, i.e., we apply the
NIBH algorithm on only the lowest 5% of the pairwise distances (which are set to zero as in BRE) combined with the
highest 2% of the pairwise distances.
We follow the continuation approach (Wen et al., 2010) to
Near-Isometric Binary Hashing for Large-scale Datasets
2.3
MNIST
b)
3
2
2.5
1.7
2
δtest
δtest
a)
NIBH
NIBH−CG
BRE
CBE
AGH
SGH
SHD
SH
1.5
1.4
1.1
1
0.8
0.5
10
20
30 40 50
No. of bits (M)
60
70
MDSH
PCAH
IsoHash
LSH
0.5
1
1.5
δ
2
2.5
3
Figure 3. Comparison of NIBH and NIBH-CG against several state-of-the-art binary hashing algorithms in preserving isometry on
MNIST data. (a) NIBH-CG outperforms the other algorithms in minimizes the isometry constant on unseen data δtest . (b) . NIBH and
NIBH-CG provide better isometry guarantee with a small sacrifice to universality.
set the value of α. We start with a small value of α, (e.g.,
α = 1) to avoid becoming stuck in bad local minima, and
then gradually increase α as the algorithm proceeds. As
the algorithm gets closer to convergence and has obtained
a reasonably good estimate of the parameters W and λ, we
set α = 10, which enforces a good approximation of the
sign function (see Lemma 3).
Results. The plots in the top row of Figure 2 illustrate
the value of the distortion parameter δ as a function of
the number of projections (bits) M . The performance of
NIBH and NIBH-CG closely follow each other, indicating that NIBH-CG is a good approximation to NIBH. Both
NIBH and NIBH-CG outperform the other baseline algorithms in terms of the distortion parameter δ. Among these
baselines, LSH has the lowest isometry performance since
random projections are oblivious to the intrinsic geometry
of the training dataset. To achieve δ = 1, NIBH(-CG) requires 60% fewer bits M than CBE and BRE. NIBH(-CG)
also achieves better isometry performance asymptotically,
i.e., up to δ ≈ 0.5, given a sufficient number of bits (M ≥
70), while for most of the other algorithms the performance
plateaus after δ = 1. NIBH’s superior near-isometry performance extends well to unseen data. Figure 3(a) demonstrates that NIBH achieves the lowest isometry constant on
a test dataset δtest compared to other hashing algorithms.
Figure 3(b) further suggests that NIBH’s superior isometry
performance comes with smallest sacrifice to the universality of the hash functions.
The plots in the middle and bottom row of Figure 2 shows
the average precision for retrieving training data and the
Kendall τ correlation coefficient respectively, as a function of the number of bits M . We see that NIBH preserves a higher percentage of nearest neighbors compared
to other baseline algorithms as M increases with better average ranking among k = 10-nearest neighbors.
Now we showcase the performance of NIBH-CG on three
medium-scale, real-world datasets used in (Kulis & Darrell,
2009; Norouzi & Fleet, 2011), including Photo-tourism,
LabelMe, and Peekaboom for the popular machine learning
task of data retrieval. From each dataset we randomly select Q = 1000 training points, following the setup in BRE
(Kulis & Darrell, 2009), and use them to train NIBH-CG
and the other baseline algorithms. We then randomly select a separate set of Q = 1000 data points and use it to
test the performance of NIBH-CG and other baseline algorithms in terms MAP with k = 50. Figure 4 illustrates
the performance of NIBH-CG on these datasets. NIBH-CG
outperforms all the baseline algorithms with large margins
in Hamming ranking performance in term of MAP with
top-50 near-neighbors.
3.3. Large-scale experiments
We now demonstrate that NIBH-CG scales well to largescale datasets. We use the full MNIST dataset with 60,000
training images and augment it with three rotated versions
of each image (rotations of 90◦ , 180◦ , and 270◦ ) to create a larger dataset with Q = 240,000 data points. Next,
we construct 4 training sets with 1,000, 10,000, 100,000,
and 240,000 images out of this large set. We train all algorithms with M = 30 bits and compare their performance
on a test set of 10,000 images. BRE fails to execute on a
standard desktop PC with 12 GB of RAM for training sets
with more than 100,000 points due to the size of the secant
set |X |. The results for all algorithms are given in Table 1;
we tabulate their performance in terms of MAP for the top500 neighbors. The performance of NIBH-CG is significantly better than the baseline algorithms and, moreover,
improves as the size of the training set grows. This emphasizes that NIBH-CG excels at large-scaled problems thanks
to its very small memory requirement; indeed, the memory
requirement of NIBH-CG is linear in the number of active
Near-Isometric Binary Hashing for Large-scale Datasets
MAP
0.6
0.6
Photo-tourism
NIBH−CG
BRE
CBE−opt
AGH
SGH
SHD
0.6
LabelMe
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
1
10
20
30
40
50
0
1
10
20
30
40
Number of bits (M)
50
0
IsoHash
LSH
SH
MDSH
PCAH
Peekaboom
1
10
20
30
40
50
Figure 4. Hamming ranking performance comparison on three medium-scale datasets (Q = 1000). The top-50 neighbors are used to
report MAP over a test data of same size.
secants rather than the total number of secants.
4. Discussion
We have demonstrated that the worst-case, `∞ -norm-based
near-isometric binary hashing (NIBH) algorithm is superior to a wide range of algorithms based on the more traditional average-case, `2 -norm. Despite its non-convexity
and non-smoothness, NIBH admits an efficient optimization algorithm that converges to a high-performing local
minimum. Moreover, NIBH-CG, the accelerated version
of NIBH, provides significant memory advantages over existing algorithms. Our exhaustive experiments with six
datasets, three metrics, and ten algorithms have shown
that NIBH outperforms all of the state-of-the-art datadependent hashing algorithms. The results in this paper
provide a strong motivation for exploring `∞ -norm formulations in binary hashing.
References
Andoni, A., Indyk, P., Nguyen, H. L., and Razenshteyn,
I. Beyond locality-sensitive hashing. In Proc. 25th Annual ACM-SIAM Symposium on Discrete Algorithms, pp.
1018–1028, Jan 2014.
Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J.
Distributed optimization and statistical learning via the
alternating direction method of multipliers. Foundations
and Trends in Machine Learning, 3(1):1–122, Jan. 2011.
Cover, T. M. and Thomas, J. A. Elements of information
theory. John Wiley & Sons, 2012.
tributions. In Proc. 20th Annual Symposium on Computational Geometry, pp. 253–262, June 2004.
Gong, Y. and Lazebnik, S. Iterative quantization: A procrustean approach to learning binary codes. In Proc.
IEEE Conf. on Computer Vision and Pattern Recognition, pp. 817–824, June 2011.
Hartley, R. and Schaffalitzky, F. L∞ minimization in geometric reconstruction problems. In Proc. IEEE Conf. on
Computer Vision and Pattern Recognition, pp. 504–509,
June 2004.
He, J., Kumar, S., and Chang, S. On the difficulty of nearest
neighbor search. In Proc. 29th Intl. Conf. on Machine
Learning, pp. 1127–1134, June 2012.
Hegde, C., Sankaranarayanan, A. C., Yin, W., and Baraniuk, R. G. Numax: A convex approach for learning
near-isometric linear embeddings. IEEE Trans. Signal
Processing, 63(22):6109–6121, Nov. 2015.
Heo, J., Lee, Y., He, J., Chang, S., and Yoon, S. Spherical
hashing. In Proc. IEEE Conf. on Computer Vision and
Pattern Recognition, pp. 2957–2964, 2012.
Horn, R. A. and Johnson, C. R. Topics in Matrix Analysis.
Cambridge University Press, 1991.
Indyk, P. and Motwani, R. Approximate nearest neighbors:
Towards removing the curse of dimensionality. In Proc.
30th Annual ACM Symposium on Theory of Computing,
pp. 604–613, May 1998.
Dantzig, G. B. and Wolfe, P. Decomposition principle for
linear programs. Operations Research, 8(1):101–111,
Feb. 1960.
Jiang, Q. Y. and Li, W. J. Scalable graph hashing with
feature transformation. IJCAI, 2015.
Datar, M., Immorlica, N., Indyk, P., and Mirrokni, V. S.
Locality-sensitive hashing scheme based on p-stable dis-
Jolliffe, I. Principal Component Analysis. Wiley Online
Library, 2002.
Near-Isometric Binary Hashing for Large-scale Datasets
Table 1. Comparison of NIBH-CG against several baseline binary hashing algorithms on large-scale MNIST datasets with over 28 billion
secants |S(X )|. We tabulate the Hamming ranking performance in terms of mean average precision MAP for different sizes of the
dataset. All training times are in seconds.
M = 30 bits
Training size Q
Secant size |S(X )|
NIBH-CG
BRE
CBE
SPH
SH
MDSH
AGH
SGH
PCAH
IsoHash
LSH
1K
500K
52.79 (±0.15)
48.33 (±0.65)
38.70 (±1.18)
44.33 (±0.74)
40.12 (±0.00)
41.06 (±0.00)
45.81 (±0.34)
51.32 (±0.07)
39.90 (±0.00)
50.91 (±0.00)
33.69 (±0.94)
MAP / Top-500 ( MNIST + rotations)
10K
100K
240K
50M
5B
28B
54.69 (±0.18) 54.93 (±0.23) 55.52 (±0.11)
50.67(±0.33)
–
–
38.12 (±1.34) 38.50 (±2.05) 38.53 (±0.83)
44.24 (±0.61) 44.37 (±0.71) 44.32 (±0.63)
39.37 (±0.00) 38.79 (±0.00) 38.26 (±0.00)
41.23 (±0.00) 40.80 (±0.00) 40.39 (±0.00)
47.78 (± 0.38) 47.69 (±0.41) 47.38 (±0.32)
51.33 (±0.20) 51.01 (±0.23) 50.66 (±0.76)
38.53 (±0.00) 38.81 (±0.00) 37.50 (±0.00)
50.90 (±0.00) 50.72 (±0.00) 50.55 (±0.00)
33.69 (±0.94) 33.69 (±0.94) 33.69 (±0.94)
Kendall, M. G. A new measure of rank correlation.
Biometrika, 30(1–2):81–93, June 1938.
Kim, H. and Park, H. Sparse non-negative matrix factorizations via alternating non-negativity-constrained least
squares for microarray data analysis. Bioinformatics, 23
(12):1495–1502, Apr. 2007.
Kong, W. and Li, W. J. Isotropic hashing. In Advances in
Neural Information Processing Systems, pp. 1646–1654,
2012.
Kulis, B. and Darrell, T. Learning to hash with binary reconstructive embeddings. In Advances in Neural Information Processing Systems, pp. 1042–1050, Dec. 2009.
LeCun, Y. and Cortes, C. The MNIST database of handwritten digits, 1998.
Ledoux, M. and Talagrand, M. Probability in Banach
Spaces: Isoperimetry and Processes. Springer-Verlag,
1991.
Liu, W., Wang, J., Kumar, S., and Chang, S. Hashing with
graphs. In Proc. 28th intl. Conf. on Machine Learning,
pp. 1–8, July 2011.
Lowe, D. G. Distinctive image features from scaleinvariant keypoints. Intl. Journal of Computer Vision,
60(2):91–110, Nov. 2004.
Lv, Q., Josephson, W., Wang, Z., Charikar, M., and Li, K.
Multi-probe lsh: efficient indexing for high-dimensional
similarity search. In Proceedings of the 33rd international conference on Very large data bases, pp. 950–961.
VLDB Endowment, 2007.
Nesterov, Y. Gradient methods for minimizing composite objective function. Technical report, Université
Catholique de Louvain, Sep. 2007.
Training time
240K
28B
541.43
18685.51
68.94
184.46
3.05
15.00
4.49
5.89
0.08
2.82
2.29 × 10−4
Norouzi, M. and Fleet, D. J. Minimal loss hashing for compact binary codes. In Proc. 28th Intl. Conf. on Machine
Learning, pp. 353–360, June 2011.
Plan, Y. and Vershynin, R. Dimension reduction by random
hyperplane tessellations. Discrete & Computational Geometry, 51(2):438–461, Mar. 2014.
Shaw, B. and Jebara, T. Minimum volume embedding.
In Proc. 11th Intl. Conf. on Artificial Intelligence and
Statistics, pp. 460–467, Mar. 2007.
Snavely, N., Seitz, S. M., and Szeliski, R. Photo tourism:
Exploring photo collections in 3D. ACM Trans. Graphics, 25(3):835–846, July 2006.
Studer, C., Goldstein, T., Yin, W., and Baraniuk, R. G.
Democratic representations. Preprint, 2014. URL
http://www.csl.cornell.edu/~studer/
papers/14TIT-linf-submitted.pdf.
Torralba, A., Fergus, R., and Weiss, Y. Small codes and
large image databases for recognition. In Proc. IEEE
Conf. on Computer Vision and Pattern Recognition, pp.
1–8, June 2008.
Wainwright, M. J. Sharp thresholds for high-dimensional
and noisy sparsity recovery using `1 -constrained
quadratic programming (Lasso). IEEE Trans. Information Theory, 55(5):2183–2202, May 2009.
Wang, J., Shen, H. T., Song, J., and Ji, J. Hashing for similarity search: A survey. arXiv preprint arXiv:1408.2927,
2014.
Weinberger, K. Q. and Saul, L. K. Unsupervised learning
of image manifolds by semidefinite programming. Intl.
Journal of Computer Vision, 70(1):77–90, May 2006.
Near-Isometric Binary Hashing for Large-scale Datasets
Weiss, Y., Torralba, A., and Fergus, R. Spectral hashing.
In Advances in Neural Information Processing Systems,
pp. 1753–1760, Dec. 2009.
Weiss, Y., Fergus, R., and Torralba, A. Multidimensional
spectral hashing. In European Conf. on Computer Vision,
pp. 340–353. Oct. 2012.
Wen, Z., Yin, W., Goldfarb, D., and Zhang, Y. A fast algorithm for sparse reconstruction based on shrinkage, subspace optimization, and continuation. SIAM Journal on
Scientific Computing, 32(4):1832–1857, June 2010.
Yu, F. X., Kumar, S., Gong, Y., and Chang, S. Circulant
binary embedding. In Proc. 31st Intl. Conf. on Machine
Learning, pp. 946–954, June 2014.
Zoran, D. and Weiss, Y. Natural images, gaussian mixtures
and dead leaves. In Advances in Neural Information Processing Systems, pp. 1736–1744, Dec. 2012.
A. Appendix
In the appendix, we prove Lem. 3 and Thm. 4 on the performance of NIBH on k-nearest neighbor preservation. Finally, we include additional numerical simulation results
and discussions on the empirical convergence of the NIBH
algorithm.
Proof of Lem. 3
Lemma 3. Let x be a Gaussian random variable as x ∼
N (µ, σ 2 ). Define the distortion of the sigmoid approximation at x as |h(x) − σα (x)|. Then, the expected distortion
is bounded as
Ex [|h(x) − σα (x)|] ≤
√
2
1
√
+ 2e−( α+c/ασ ) ,
σ 2πα
where c is a positive constant. As α goes to infinity, the
expected distortion goes to 0.
Proof. It is easy to see that the distortion |h(x) − σα (x)|
occurs at x = 0. Therefore, among different values of µ,
µ = 0 gives the largest distortion since the density of x
peaks at x = 0. Therefore, we bound the distortion at setting µ = 0, which is an upper bound of the distortion when
µ 6= 0. By definition (1) in the main text, h(x) can be
written as
ß
1 if x ≥ 0
h(x) =
0 otherwise.
When x ∼ N (0, σ 2 ), we have
Z ∞
Ex [|h(x) − σα (x)|] =
|h(x) − σα (x)| N (x; 0, σ 2 )dx
−∞
Z ∞
=2
(h(x) − σα (x))N (x; 0, σ 2 )dx
(3)
0
Z x0
=2
(h(x) − σα (x))N (x; 0, σ 2 )dx
0
Z ∞
+2
(h(x) − σα (x))N (x; 0, σ 2 )dx
x0
Z x0
Z ∞
1
1
2
N (x; 0, σ )dx+2
N (x; 0, σ 2 )dx
≤2
2
1
+
eαx0
0
x0
(4)
2
2
x0
e−cx0 /σ
≤√
+2
1 + eαx0
2πσ
√
2
1
≤ √
+ 2e−( α+c/ασ ) ,
σ 2πα
(5)
(6)
when we set x0 = √1α and c is a positive constant. In (3),
we used the fact that σα (x) and h(x) are symmetric with
respect to the point (0, 21 ). (4) is given by the properties of
the sigmoid function, (5) is given by the Gaussian concentration inequality (Ledoux & Talagrand, 1991), and (6) is
Near-Isometric Binary Hashing for Large-scale Datasets
given by the inequality 1/(1 + eαx0 ) ≤ e−αx0 . The fact
that Ex [|h(x) − σα (x)|] → 0 as α → ∞ is obvious from
the bound above.
Proof of Thm. 4
dH (x0 , xm ) > dH (x0 , xn ) with m ∈ {1, . . . , k} and n ∈
{k + 1, . . . , Q}. Then, using the union bound (Cover &
Thomas, 2012), we have
X
P (E) ≤
P (em,n ) ≤ k(Q − k)P (ek,k+1 )
m,n
Theorem 4. Assume that all the data points are independently generated
PP from a mixture of Gaussian distribution,
N
denote a
i.e., xi ∼
p=1 πp N (µp , Σp ). Let x0 ∈ R
query data point in the ambient space, and the other data
points xi be ordered so that d(x0 , x1 ) < d(x0 , x2 ) <
. . . < d(x0 , xQ ). Let δ denote the final value of the distortion parameter computed from any binary hashing algorithm, and let»c denote a positive constant. Then, if
Ex [∆k ] ≥ 2δ + 1c log Qk
, the binary hashing algorithm
preserves the k-nearest neighbors of a point with probability at least 1 − .
= k(Q − k)P (dH (x0 , xk ) > dH (x0 , xk+1 ))
= k(Q − k)P (dH (x0 , xk+1 ) < dH (x0 , xk )),
where we have used the fact that the most possible event
among all em,n events is the one corresponding to the
order mismatch between the k th and k + 1th nearest
neighbor. Now, note that the NIBH output δ satisfies
maxi,j |dH (xi , xj ) − d(xi , xj )| ≤ δ 1 . Observe that ∆k =
d(x0 , xk+1 ) − d(x0 , xk ) ≥ 2δ is a sufficient condition for
dH (x0 , xk+1 ) ≥ dH (x0 , xk ), since
dH (x0 , xk+1 ) − dH (x0 , xk )
In order to prove this theorem, we need the following
Lemma:
Lemma 5. Let x0 , . . . , xN and ∆k be defined as in Thm. 4.
Then, there exist a constant c such that P (∆k − Ex [∆k ] <
2
t) ≤ e−ct for t > 0.
≥ d(x0 , xk+1 ) − δ − d(x0 , xk ) − δ
≥ 2δ − 2δ = 0,
by the triangular inequality. This leads to
P (dH (x0 , xk+1 ) < dH (x0 , xk ))
Proof. Since the data points x0 , xk and xk+1 are independently generated from a finite mixture of Gaussian
distributions, the random variable of their concatenation
y = [xT0 , xTk , xTk+1 ]T ∈ R3N is sub-Gaussian (Wainwright, 2009). Then, we have
∆k (y) = kx0 − xk+1 k2 − kx0 − xk k2
I
=k 0
0
0
0
0
0
I
0 yk2 − k 0
−I
0
2I 0
≤ k 0 −I
0
0
0 0
−I 0 yk2
0 0
= 1 − P (dH (x0 , xk+1 ) ≥ dH (x0 , xk ))
≤ 1 − P (∆k ≥ 2δ) = P (∆k < 2δ).
Therefore, combining all the above and Lem. 5, the probability that the k-nearest neighbor is not preserved is
bounded by
P (E) ≤ k(Q − k)P (dH (x0 , xk+1 ) < dH (x0 , xk ))
≤ k(Q − k)P (∆k < 2δ)
= k(Q − k)P (∆k − Ex [∆k ] < −(Ex [∆k ] − 2δ))
0
0 yk2
−I
≤ 2kyk2 ,
where we have used the triangular inequality in the second to last step, the Rayleigh-Ritz theorem (Horn & Johnson, 1991), and the fact that the maximum singular value
of the matrix in the step before is 2. This result means
that ∆k (y) is a Lipschitz function of y. Thus, by Talagrand’s inequality (Ledoux & Talagrand, 1991), we have
2
that P (∆k − Ex [∆k ] < −t) ≤ e−ct for some positive
constant c and t > 0, since y is sub-Gaussian.
Now we are ready to prove Thm. 4.
Proof. Let E denote the event that the set of top-k nearest
neighbors is not preserved in the Hamming space. Then,
we have E = ∪em,n , where em,n denote the event that
2
≤ k(Q − k)e−c(Ex [∆k ]−2δ)
2
≤ kQe−c(Ex [∆k ]−2δ) .
2
Now, let kQe−c(Ex [∆k ]−2δ) ≤ , we have that the requirement for the k-nearest neighbors to be exactly preserved
with probability at least 1 − is
…
1
Qk
log
.
Ex [∆k ] ≥ 2δ +
c
Remark Note that our bound on the number of nearest neighbor preserved k depends on the final outcome of
the NIBH algorithm in the value of δ. In order to relate our result to the number of binary hash functions M
required, we can make use of (Plan & Vershynin, 2014,
1
Here we assume λ = 1 without loss of generality.
Near-Isometric Binary Hashing for Large-scale Datasets
Thm. 1.10). For a bounded set K ⊂ RN with diameter 1,
let M ≥ Cδ −6 w(K)2 where w(K) = Ex [supx∈K hg, xi]
denotes the Gaussian width of K and some constant C.
Then, (Plan & Vershynin, 2014, Thm. 1.10) states that,
with high probability, h(x) as defined in (1) with a random
matrix W whose entries are independently generated from
N (0, 1) is a δ0 -isometric embedding. Therefore, if we initialize NIBH with such a random W which is likely to be
δ0 -isometric, then empirically (see Figure 5), the NIBH algorithm will learn a better embedding that is δ-isometric
with δ < δ0 . Therefore, we have that the number of hash
functions M required for k-nearest
neighbor preservation
p
is at least M ∼ (Ex [∆k ] − log(kQ/)−6 w(X)2 assuming that the training dataset X is properly normalized.
Empirical convergence of the NIBH
algorithm
Figure 5 shows the empirical loss and the actual distortion parameter δ as a function of the iteration count ` in
the NIBH algorithm as applied on 4950 secants (i.e., Q =
100) from the MNIST dataset. The behavior of the empirical loss function closely matches that of δ as they gradually converge. The curve empirically confirms that minimizing the loss function in each iteration of NIBH (using
the ADMM framework) directly penalizes the non-convex
loss function (distortion parameter δ). After initializing the
NIBH algorithm with random entries for W, the value of
the distortion parameter δ significantly drops in the first few
iterations, empirically conforming that NIBH learns an embedding with significantly lower distortion parameter after
a few iterations.
||λv−c||
Objective value
2.5
∞
δ
2
1.5
1
0.5
0
200
400
600
Iteration
800
1000
Figure 5. Empirical convergence behavior of the NIBH algorithm.
Both the maximum distortion parameter δ and the loss function
||λv − c||∞ that approximates δ gradually decrease and converge
as the number of iterations increases. We see that the loss function of NIBH closely matches the behavior of the actual distortion
parameter in each iteration.
| 8cs.DS
|
CSA++: Fast Pattern Search for Large Alphabets
Simon Gog1,2 , Alistair Moffat1 , and Matthias Petri1
arXiv:1605.05404v1 [cs.DS] 18 May 2016
1
Dept. Computing and Information Systems, The University of Melbourne, Australia
2 Inst. Theoretical Informatics, Karlsruhe Institute of Technology, Germany
Abstract. Indexed pattern search in text has been studied for many decades. For
small alphabets, the FM-Index provides unmatched performance, in terms of both
space required and search speed. For large alphabets – for example, when the
tokens are words – the situation is more complex, and FM-Index representations
are compact, but potentially slow. In this paper we apply recent innovations from
the field of inverted indexing and document retrieval to compressed pattern search,
including for alphabets into the millions. Commencing with the practical compressed suffix array structure developed by Sadakane, we show that the Elias-Fano
code-based approach to document indexing can be adapted to provide new tradeoff
options in indexed pattern search, and offers significantly faster pattern processing compared to previous implementations, as well as reduced space requirements.
We report a detailed experimental evaluation that demonstrates the relative advantages of the new approach, using the standard Pizza&Chili methodology and files,
as well as applied use-cases derived from large-scale data compression, and from
natural language processing. For large alphabets, the new structure gives rise to
space requirements that are close to those of the most highly-compressed FMIndex variants, in conjunction with unparalleled search throughput rates.
Keywords: String search, pattern matching, suffix array, Burrows-Wheeler transform, succinct data structure, experimental evaluation.
1
Introduction and Background
We study a well-known problem: given a static text T[0, n − 2] over an alphabet Σ of size
σ followed by a symbol T[n − 1] = $, with $ 6∈ Σ , preprocess T so that a sequence of
patterns P[0, m − 1], also over Σ , can be efficiently searched for, with the purpose of each
search being to identify the number of occurrences nocc of P in T. A variety of options
exist for this problem, providing different trade-offs between construction cost, memory
space required during pattern search operations, and search cost, both asymptotically and
in practical terms. Example structures include the suffix tree [3, 26] and suffix array [16].
The suffix array of T, denoted SA, requires O n log n bits of space in addition to the
O n log σ bits occupied by T, and uses that space to store the offsets SA[0, n−1] of all n
suffixes of T (denoted as T[i, ]) in lexicographic order such that T[SA[i], ] < T[SA[i+1], ]
for i ∈ [0,n − 1]. Using SA, the number of occurrences of P in T can be identified in
O m log n time, via two binary searches that determine the range (sp, ep) such that all
suffixes SA[sp, ep] are prefixed
by P. Thus, nocc = ep − sp + 1. The search cost can be
reduced to O m + log n if information about longest common prefixes is also available.
Storing this information for
all possible intervals SA[i, j] occurring in the binary search
process requires O n log n bits of additional space.
Compressed Indexes. In a compressed suffix array, or CSA, the space required is proportional to the compressed size of T. Sadakane [23] (see also Grossi and Vitter [12]) describes a CSA based on the observation that the function ψ[i] = SA−1 [(SA[i]+1) mod n]
consists of σ increasing sequences (or segments) of integers, and that each of those seg
ments is likely to be compressible, yielding a space usage of nHk (T) + O n log log σ
bits [19], where Hk denotes to the order-k empirical entropy of T. Occurrences of P are
located by performing a backward search to find the range SA[spi , epi ] matching each
suffix P[i, ], stopping if epi < spi , or if all of P has been processed.
An alternative compressed indexed is due to Ferragina and Manzini [5], and is based
on the Burrows Wheeler Transform (BWT), defined as BWT[i] = T[SA[i] − 1 mod n].
In an FM-Index the BWT
is generally encoded using a wavelet tree [11], and accessed
using Rank BWT, i, c , which returns the number of times symbol c occurs in the prefix
BWT[0, i − 1]. Again, P is processed in reverse order. Suppose SA[spi , epi ] refers to the
range in SA prefixed by P[i, ], and that P[i − 1] = c. An array C of σ log n bits stores the
number of symbols c in T smaller than c; using it, spi−1 = C[c] + Rank BWT, spi , c
and epi−1 = C[c] + Rank BWT, epi + 1, c − 1 can be computed. Overall, SA[sp, ep] is
identified using 2m Rank operations on the BWT; and when stored using a wavelet
tree, O m log σ time. For more information about these structures, and the time/space
trade-offs that they allow, see Navarro and Mäkinen [19] and Ferragina et al. [6].
In Practice. Implementations of the CSA and the FM-Index have been developed and
measured using a range of data. When σ is small – for example, σ = 4 for DNA, and
σ ≈ 100 for plain ASCII text – both provide fast pattern search based on compact
memory footprints, usually requiring half or less of the space initially occupied by T,
depending on a range of secondary structures and parameter choices [9, 13], and with
the FM-Index typically requiring less space that the CSA. But when σ is large – for
6
example, when the alphabet is words in a natural language
and σ ≈ 10 or greater –
the situation is more complex. In particular, the O log σ factor associated with the
FM-Index’s wavelet tree is a count of random accesses (as distinct from cache-friendly
accesses) and means that search costs increase with alphabet size, negating its space
advantage. In contrast, standard CSA implementations are relatively unaffected by σ ,
but each backward search step in a CSA has a dependency on log nc , where nc is the
frequency in T of the current symbol c = P[i]. Hence, if σ is fixed and does not grow
with n, CSA pattern match times will grow as T becomes longer.
Our Contribution. We introduce several improvements to the CSA index:
– We adapt and extend the uniform partitioned Elias-Fano (UEF) code of Ottaviano
and Venturini [21] to the storage of the ψ function, allowing faster backwards search
compared to previous implementations;
– We add a fourth UEF block type compared to Ottaviano and Venturini, and include
the option of coding sections of the ψ function in a runlength mode;
– We describe a way of segregating the short segments in ψ, allowing improved
compression when σ is large and many of the symbols in Σ are rare;
– We carry out detailed “at scale” experiments, including both synthetic query streams
and logs derived from use-cases, covering all of small, medium, and large alphabets.
The result is a pattern search index that we refer to as “CSA++”. It represents a significant
shift in the previous relativities between compressed index structures; and, for large
alphabets in particular, gives rise to space needs close to those of the most highlycompressed FM-Index variants, with unparalleled search throughput rates.
2
Storing Integer Lists
Operations Required. The function ψ[i] = SA−1 [(SA[i] + 1) mod n] is a critical – and
costly – component of a CSA. It can be thought of as consisting of a concatenation of
σ segments, the c th of which is a sorted list of the locations in BWT at which the c th
symbol in Σ appear. That is, each segment of ψ can be interpreted as a postings list
of occurrences of symbol
c. The key operation required to enable backwards search
is that of GEQ c, pos , which returns the smallest position pos0 such that ψ[pos0 ] is
in the c th segment, and such that ψ[pos0 ] ≥ pos. Starting with sp = 0 and ep = n − 1,
the (sp, ep) bounds are narrowed via a sequence of m pairs of sp = GEQ c, sp and
ep = GEQ c, ep + 1 − 1 operations, as c takes on values from P[m − 1] through to P[0].
The equivalence
of the CSA and FM-Index
search processes can be seen by noting that
GEQ c, pos = C[c] + Rank BWT, pos, c , and that all of the (sp, ep) pairs computed
are identical between the two. Note also that, by construction, symbol occurrences in
the BWT string are likely to appear in clusters, and hence ψ is likely to contain runs of
consecutive or near-consecutive integers, separated by large intervals, and to contain at
most σ “disruption” points at which ψ[i] > ψ[i + 1].
Integer Codes. One common way of storing postings lists is to compute gaps, or differences, and then store them using a suitable code for integers; clusters in BWT then gives
rise to runs of small or unit gaps in ψ. A range of integer codes have been developed
for this type of distribution, including Elias γ and δ codes, Rice and Golomb codes, and
the Binary Interpolative Code (see Moffat and Turpin [18, Chapter 3] for descriptions).
Several of these have been used in previous CSA implementations [23].
There has been recent interest in Elias-Fano codes (EF codes) for postings list
compression, a result of work by Vigna [25] (see also Anh and Moffat [2] for earlier
application, and Gog et al. [9] for preliminary experimentation with compressed suffix
arrays). Given a non-decreasing set of k integers in the range 0 . . . 2U − 1 for some
universe size 2U , a parameter ` is selected, and each integer is split into a high part (the
most significant U − ` bits) and a low part (the ` low-order bits). Groups are formed for
values that have the same high parts. A code for the block of k values is then constructed
by representing the size of each of the 2U−` possible groups in unary, followed by the
concatenation in order of the k low parts. For example, if U = 4 and k = 3, the sequence
[6, 7, 10] (that is, [0110, 0111, 1010] in binary) would be coded using ` = 2, and split into
high parts, [01, 01, 10], coded as group sizes in unary as 0:110:10:0; and into low parts
coded in binary, 10:11:10, where the “:”s are purely indicative, and do not appear in
the output. The EF code achieves representations close to the combinatorial minimum if
` = blog2 (2U /k)c; moreover, the length of the coded block is easily computed: k + 2U−`
bits are required for the high/unary parts, and k · ` bits for the low/binary parts.
One useful aspect of the EF code is that the unary parts can be searched via Select
operations over their “0” bits, and then the number of binary parts through until that
point computed. For example, in the unary sequence shown above, any elements from the
underlying sequence in the range 8 . . . 11 must fall in the third bucket, and Select0 (2) −
2 = 4 − 2 = 2 indicates that there are in total two binary parts contained within the first
two buckets, and hence that the binary parts associated with the third bucket (if any)
must commence from the third element of the low/binary part. On average there is O 1
item per bucket, and linear search can be used to scan them; if a worst-case bound is
required, binary search can be used if there are more than log2 n “1” bits between the
relevant pair of consecutive “0” bits, and linear search employed otherwise.
Another feature of EF codes is that in the binary part all components are of the
same bit-length `, meaning that there are no dependencies that would hinder vectorized
processing and loop-unrolling techniques and prevent them from achieving their full
potential. This is not the case with, for example Elias δ codes, which are based upon
gaps and are also of variable length, and hence must be decoded sequentially.
Partitioned Elias-Fano Codes. The term occurrences in long postings lists tend to
be clustered, a pattern that has been used as the basis for a range of improved index
compression techniques [18]. Ottaviano and Venturini [21] demonstrated that EF codes
could capture much of this effect if postings lists were broken into blocks of k values,
and then the document identifiers in each block mapped to the range 0 . . . 2U − 1 for
some suitable per-block choice of U. Ottaviano and Venturini further observed that in
some cases EF codes are less efficient than other options, and that it was helpful for
blocks to be coded in one of three distinct modes: (i) those consisting of an ascending
run of k consecutive document identifiers, in which case no further code bits are required
at all (NIL blocks); (ii) those where the document identifiers are sufficiently clustered
(but not consecutive) that a 2U -bit vector is the most economical approach (BV blocks);
and (iii) those that are best represented using EF codes, taking 2U−` + k · (1 + `) bits.
Note that the decision between these options can be made based solely on k and U.
The combination of fixed-k blocks and range-based code selection is referred to as
Uniform Elias-Fano (UEF) coding. Ottaviano and Venturini also describe a mechanism
for partitioning postings lists into approximately-homogeneous variable-length blocks
in a manner that benefits EF codes that we do not employ here.
Overall Structure of a CSA. With gaps in ψ represented by variable-length codewords,
the ability to directly identify and then search segments of ψ is lost. Instead, pseudorandom access is provided via a set of samples: ψ is broken into fixed-length blocks;
the first ψ value in each block is retained uncompressed in a sample index; and the
remaining values in that block
are coded as gaps starting from that first value [20, 23].
Computation of GEQ c, pos then involves identification of the region in the sample
index associated with the segment for symbol c, binary search in that section of the
sample index to identify the single block that contains pos or the next ψ value greater
than it; and then sequential decoding of that whole block, to reconstruct values of ψ in
order to determine the exact value. If symbol c occurs nc times in T, and if samples
are
extracted every k values, then searching the sample index requires O log(nc /k) time, a
cost that must be balanced against the
O k cost of linear search within the block. Small
values of k give faster GEQ c, pos operations, but also increase the size of the sample
index, and hence the size of the CSA.
3
Representing ψ
We store the ψ function of a CSA using the UEF approach of Ottaviano and Venturini,
using a blocksize of k as the basis for both the UEF code and the sample index [9]. A
number of further enhancements to previous implementations are now described.
Independent Structures. Rather than storing the whole of ψ as a single entity split
into blocks, we treat each segment independently, and genuinely form an inverted index
for the symbols c in BWT. The σ log n-bit array C of cumulative symbol frequencies
is retained, and hence nc = C[c + 1] − C[c]. A UEF-structured postings list of dnc /ke
blocks is then created for symbol c, with its own sample index constructed from the
first (smallest) value in each of the blocks, and also represented using an EF code, with
U 0 = dlog2 ne as the universe size for this “top level” structure, and k0 = dnc /ke the
number of values to be coded within it.
One risk with this “separate structures” approach is that symbols c for which nc is
small may incur relatively high overheads; a mechanism for addressing this concern is
presented shortly. Another potential issue is the cost of the mapping needed to provide
access to the c th of these structures, given a symbol identifier c; that process is also
described in more detail later in this section.
RL Blocks. Ottaviano and Venturini [21] employ three block types, to which we add
a fourth: run-length encoded blocks (RL blocks). The NIL blocks of Ottaviano and
Venturini account for runs of k consecutive ψ values; but there are also many instances
of shorter runs that do not span a whole block. In an RL block, the (strictly positive)
gaps between consecutive ψ values are represented using the Elias δ code. Any unit
gaps are followed by a second δ code to indicate a repeat counter, while non-unit gaps
are left as is. For example, [27, 28, 29, 45, 46, 47, 48, 70, 71, 73] would be represented as
[(+1, 2), +16, (+1, 3), +22, (+1, 1), +2], with the plus symbols and parentheses indicative only, and with the sampled value 27 held in the top-level structure.
To decide whether to apply RL mode to any given block, the space that it would
consume is found by summing the lengths of the δ codes, and comparing against the
(calculated) cost of the BV and EF alternatives. Because δ is slower to decode than
EF codes, a “relative advantage” test is applied, and blocks are coded using the RL
approach only if the RL size is less than half the size of the smaller of an equivalent BV
or EF-coded block. A flag bit at the start of each block informs the decoder which mode
is in use for that block.
Low-Frequency Symbols. When σ is large it is likely that many symbols in Σ have
relatively low frequencies and hence notably different values in ψ; and having a small
number of widely-spaced values in a block that is otherwise tightly clustered increases
the cost of every codeword in the block, because of the non-adaptive nature of the EF
code. In the “separate structures” approach we are adopting, there is also a level of persegment overhead that is relatively expensive for short segments. To address this issue,
we add a further option for storing the ψ values for low-frequency symbols, and do
not build an independent UEF structure for them. For example, consider a symbol c of
frequency nc = 2. Its segment in ψ is only two symbols long, and it is far more effective
to segregate those two values into two elements of a separate array using dlog2 ne bits
each than it is to construct a UEF structure and the associated sample index. In particular,
if those two elements are within a larger array in which all of the values for all symbols
for which nc = 2 are stored, the overhead space can be kept small.
The array C has already been mentioned, it allows nc to be computed for a symbol c.
A bitvector D of size σ with Rank support is added, with D[c] = 1 if symbol c is being
stored as a full UEF structure, and D[c] = 0 if nc ≤ L for some threshold L. We use D to
map from Σ to Σ 0 = {c | nc ≤ L}. The next component required is a wavelet tree over
the values nc , where c ∈ Σ 0 , to support Rank operations and hence determine how many
symbols c0 < c in Σ 0 have nc0 = nc . Finally, a set of L arrays are maintained, one for each
symbol frequency between 1 and L. We suppose that Ai is the i th of those arrays. With
those components available, locating the segment of ψ values corresponding to symbol
c is carried out as follows. First, D[c] is accessed and nc = C[c + 1] −C[c] is determined.
If D[c] is zero, the wavelet tree is used to compute s = |{c0 | 1 ≤ c0 < c and nc0 = nc }|,
and the nc required values of ψare at Anc [nc · s . . . nc · s + nc − 1]. On the other hand, if
D[c] = 1, then s = Rank D, c, 1 is computed, and the s th of the full UEF structures is
used to access the c th segment of ψ.
In the experiments reported in the next section we take L = k, where k is the UEF
block size and also the sample interval. That is, any symbols c for which nc ≤ k and less
than one full UEF block would be required are stored in uncompressed form as binary
values in the range 0 . . . n − 1, in contiguous sections of shared arrays. Note that as a
further small optimization the groups of nc elements that collectively comprise each of
the arrays Anc could themselves be stored using EF codes when nc ≥ 2, since the EFcompressed length of each such group is both readily calculable and identical. However,
given that naturally-occurring large-alphabet frequency distributions typically have long
tails of very low symbol frequencies, the average cost of such EF codes might be close
to dlog2 ne bits per ψ value anyway, in which case we would expect the additional gains
to be modest. We leave detailed exploration of this idea for future work.
Eliminating Double Search. As described earlier, each symbol that is processed in P
gives rise to two GEQ operations over ψ. It is thus tempting to compute these via two
calls to the same function. But much of the computation between the two calls can
be
shared, and it is more efficient to perform the first GEQ call to identify GEQ c, sp , and
then perform a finger-search from that point to compute the equivalent of GEQ c, ep .
4
Experiments
Methodology and Implementation. The baselines and CSA++ are written in C++14 on
top of the SDSL library [7] and compiled with optimizations using gcc 5.2.1.3 We also
make use of Sadakane’s source code as a further reference point [23]. The experimental
results were generated using a Intel Xeon E5640 CPU using 144 GiB RAM. All timings
reported are averaged over five runs; the variance was low and all measurements lie
3
To ensure the reproducibility of our results, our complete experimental setup, including data
files, is available at github.com/mpetri/benchmark-suffix-array/.
within approximately 10% of each reported value. All space usages reported are those
of the serialized data structures on disk.
Data Sets, Queries and Test Environment. Our experiments make use of texts T from
two different sources: four 200 MiB files drawn from the Pizza&Chili corpus4 , selected
to illustrate a range of alphabet sizes σ ; plus two 2 GiB files of natural language text,
one in German, and one in Spanish. The latter were extracted from a sentence-parsed
prefix of the German and Spanish sections of the CommonCrawl5 . The four 200 MiB
Pizza&Chili files are treated as byte streams, with σ ≤ 256 in all cases; the two larger
files are parsed in to word tokens, and then those tokens mapped to integers. There were
σ = 5,039,965 distinct words (integers) in the German-language file, and σ = 2,956,209
distinct words in the Spanish-language file.
The primary query streams applied to these files were generated by randomly selecting 50,000 locations in T and extracting m = 20-character strings for the Pizza&Chili
files, and extracting m = 4-symbol/word strings for the two natural language files. This
follows the methodology adopted by other similar experimentation carried out in the
past. As secondary query streams, we also make use of the strings generated by two
specific use-cases, described later in this section, in part as a response to the concerns
explored by Moffat and Gog [17].
Pattern Search, Small Alphabets. Figure 1 depicts the relative performance of two previous CSA implementations, and a total four of FM-Index options. The method marked
CSA reflects the description of [23], as implemented in the SDSL library; it stores the
ψ function using Elias γ codes as a single stream of gaps, with the disruptive elements
at the start of each segment represented as very large values rather than as negative
gaps, and with the samples stored uncompressed. Method CSA-SADA is Sadakane’s
implementation of the same mechanism. The CSA++ is the approach described here.
We compare against two versions of each of two FM-Index approaches. The first
pair, prefixed FM-HF, use a Huffman-shaped wavelet tree (WT) for the whole BWT [15].
The first version of this approach represents the WT by an uncompressed bitvector and
a cache-friendly rank structure (FM-HF-BVIL), and seeks to provide fast querying at
the expense of memory space; the second one uses entropy-compressed bitvectors (FMHF-RRR) to represent the WT, and is at the other extreme of the space/speed tradeoff.
The second pair of FM-Indexes are based on fixed-block compression boosting, prefixed
FM-FB. The BWT is partitioned into fixed-length blocks and a WT is created for each
block. We use a recent implementation by Gog et al. [8], and plug-in an uncompressed
bitvector and rank structure (FM-FB-BVIL), and a hybrid bitvector (FM-FB-HYB) [14].
We did not have access to an implementation of another recent CSA proposal [1].
In Figure 1, index size on the horizontal axis is expressed as a percentage relative
to the original text size, which in the case of these four files, is always 200 MiB. To
measure search times, plotted on the vertical axis, the corresponding query streams were
executed in entirety to determine an nocc count for each query, and then the overall
execution time for the stream was divided by the total number of query characters, to
4
5
See http://pizzachili.dcc.uchile.cl/texts.html.
See http://data.statmt.org/ngrams/deduped/
4k
dna.200MB
proteins.200MB
dblp.xml.200MB
english.200MB
Count time per character [ns]
3k
2k
1k
0k
4k
3k
2k
1k
0k
25%
50%
75%
100%
25%
50%
75%
100%
Index size [% of original text size]
Index
CSA
CSA++
FM-HF-RRR
CSA-SADA
FM-HF-BVIL
FM-FB-BVIL
FM-FB-HYB
Fig. 1. Cost of indexed pattern search for a set of small- and medium-alphabet Pizza&Chili files,
each 200 MiB. The preferred zone is at the lower-left.
obtain a computation time per query byte. Where there is more than one point shown
for a method, the blocksize k is the parameter being varied. As can be seen from the
four graphs in the figure, in general, the best of the FM-Indexes tested were the two
FM-FB variants, and they also outperformed the two CSA implementations. The CSA++
outperforms both implementations of the earlier CSA approach on all four files.
Pattern Search, Large Alphabets. Figure 2 shows the same experiment, applied to the
large-alphabet natural-language texts. A total of six FM-Index methods suited to large
alphabets are compared to the previous CSA (the SDSL version) and the new CSA++:
an alphabet partitioned (FM-AP) index [4] which provides O(log log σ ) rank time, and
a variant FM-AP-HYB which uses a hybrid bitvector [14]; two versions of Golynski
et al.’s [10] rank structure (GMR-RS and GMR); and again a huffman shaped WT using
either a plain bitvector (FM-HF-BVIL) or a hybrid bitvectors (FM-HF-HYB). In this
environment the CSA++ dominates all of the alternative mechanisms, requiring either
substantially less space, or offering greatly improved query rates. The careful attention
paid to the representation of infrequent terms is clearly beneficial.
Figure 3 helps explain the situation. The great majority of the symbols in Σ occur
fewer than k = 128 times; indeed, 25% of them appear only once. Reducing the per-term
overhead is thus very important. However, as is shown in the right pane, those terms
are a small percentage of the ψ array, and storing them in binary is not detrimental to
overall performance.
Count time per character [ns]
german-2048
10 k
spanish-2048
5k
2k
1k
500
40%
80%
120%
160%
40%
80%
120%
160%
Index size [% of original text size]
Index
CSA
FM-AP
FM-GMR
FM-HF-BVIL
CSA++
FM-AP-HYB
FM-GMR-RS
FM-HF-HYB
Fig. 2. Cost of indexed pattern search two 2 GiB files of natural language text, parsed in to word
tokens. The preferred zone is at the lower-left. Note that the vertical axis is logarithmic.
Ψ list lengths [% of n]
Count [% of σ ]
30%
25%
20%
15%
10%
5%
0%
0
5
10
15
20
Word bucket [b]
25
30
30%
25%
20%
15%
10%
5%
0%
0
5
10
15
20
25
30
Word bucket [b]
Fig. 3. Frequency statistics for german-2048. Bucket b contains words which occur between 2b
and 2b+1 − 1 times, with bucket b = 0 covering words that occur exactly once. The left pane shows
the percentage of σ accounted for by each bucket, and the right pane the percentage of n. The
low-frequency part of each distribution (less than k = 128 word occurrences) is marked in red.
Detailed Space Breakdown. Table 1 provide details of the space required by various
components of the improved CSA, for a small-alphabet file, a mid-alphabet file, and a
large-alphabet file. The two columns associated with each of the three files show the
space required by the named component, preceded by, where appropriate, the fraction of
the values in ψ that are handled via that option. The EF-coded samples require around
1–2% of the original space; and various other access structures, including the wavelet
tree for low-nc symbols, require a further 2–3%. The four different block types play
different roles across the three files. For the DNA data, the great majority of ψ values
are included in BV blocks; for the XML data, the emphasis is on NIL blocks; and for
the word-based large-alphabet data it is EF blocks that dominate. In the latter case, a
small but important fraction of the ψ values are coded in plain binary, as shown above
in Figure 3. The effect of this alphabet partitioning is better compression for the EFcoded values, which on this file are the dominant type; confirming that this option is
Method Component
CSA++ Samples
NIL-blocks
BV-coded blocks
RL-coded blocks
EF-coded blocks
Binary values
Other structures
Total space
DNA (200 MiB)
XML (200 MiB)
German (2 GiB)
%ψ
MiB
%ψ
MiB
%ψ
MiB
–
0.2
78.0
2.5
19.3
0.0
–
–
2.3
0.0
61.7
0.2
22.6
0.0
5.9
92.7
–
62.0
10.8
14.4
12.7
0.0
–
–
3.1
0.0
5.9
4.3
20.4
0.0
4.8
38.5
–
15.7
16.4
3.8
59.6
4.4
–
–
36.8
0.0
32.1
13.6
626.3
115.5
43.0
867.2
CSA
–
91.4
–
56.7
–
1061
FM-FB-HYB
–
51.3
–
25.6
–
–
FM-HF-HYB
–
51.8
–
32.4
–
1411
FM-AP
–
–
–
–
–
903.3
FM-AP-HYB
–
–
–
–
–
778.5
Table 1. Comparing the space costs of different pattern search indexes, using a blocksize of
k = 128 throughout. The methods listed in the lower part of the table are from the SDSL library.
Note that not all of the methods are applicable to all of the files.
an important component of the large-alphabet situations handled so well by the CSA++.
Table 1 also lists the space needed by several other compressed pattern search structures,
to provide further context for these results.
Case Study, Text Factorization. The RelaRandom
RLZ factors
tive Lempel-Ziv (RLZ) compression mecha- Blocksize
CSA CSA++ CSA CSA++
nism represents a string STR as a sequence
of factors from a dictionary D, see Petri et
k = 64 1.84 0.74 1.68 0.56
al. [22] for a description and experimental re- k = 128 2.89 0.76 2.73 0.59
sults. To greedily determine longest factors
k = 256 5.10 0.88 4.90 0.73
using a CSA, we take T = Dr , the reverse of
D, and build a compressed index. The string Table 2. Per-character time in microsecis then processed against T taking symbols onds for RLZ factorization, compared to 23from STR in left-to-right order, and perform- character random patterns.
ing a backward search in T; if a prefix of length p from STR is sufficient to ensure that
the (sp, ep) range becomes empty, then the next factor emitted is of length p − 1. That
is, the factorization process can be regarded as applying variable-length patterns to a
text T, with each pattern being as short as possible without appearing in T. To carry out
an application-driven experiment, we took the 64 GiB prefix of the GOV2 document
collection used by Petri et al., and built a set of patterns, each of which is one factor, plus
the next character from STR. The first 1,901,131,365 patterns from that set, representing
4 GiB of text, were used as queries. The average factor length was 23.6 characters, with
nocc = 0 in T in all cases. We then applied those patterns to an 256 MiB dictionary D
constructed from the whole 64 GiB, to compute the per-character cost of performing the
specified searches, and compared against the per-character cost associated with search
for randomly selected patterns. Table 2 shows the cost of backward search step in both
scenarios and confirms both that CSA++ significantly outperforms CSA, and also that
for count queries, random strings are a reasonable experimental methodology.
Case Study, Language Modeling. A comRandom
NL search
mon operation on natural language files Blocksize
CSA CSA++ CSA CSA++
is to identify informative phrases as sentences are parsed [24]. We built variablek = 64 1.86 1.05 1.67 0.63
length queries for the file german-2048, and
k = 128 2.90 0.99 2.98 0.62
measured the per-symbol processing time, k = 256 5.50 0.99 5.99 0.63
comparing actual-use queries and randomlyselected-string queries for CSA and CSA++. Table 3. Per-word time in microseconds for
In total 1,521,869 queries of average length phrase search, compared to 4-word random
3.4 words were extracted from the machine patterns.
translation process described by Shareghi et al., corresponding to 40,000 sentences randomly selected from the German part of Common Crawl. Table 3 shows the cost of
those count queries over the german-2048 file. The results again align with the performance of pattern searches for random queries extracted from the text, as was shown in
Figure 2. Note in particular that CSA++ performance is largely unaffected by k, whereas
the performance of CSA substantially decreases as k increases. As pattern search is a
major part of the cost of the machine translation process described by Shareghi et al.
[24], utilizing CSA++ leads to a significant speedup in practical performance.
5
Conclusion
We have described several enhancement’s to Sadakane’s CSA, and have demonstrated
improvements both in terms of compression effectiveness, and also in terms of query
throughput for count queries, especially for large-alphabet applications. If locate queries
are also required, all of the structures explored here must be augmented with SA samples,
to allow (sp, ep) ranges to be converted to offsets in T; as future work, we plan to
investigate space-speed tradeoffs in that regard as well.
Acknowledgment. This work was supported under Australian Research Council’s Discovery Projects funding scheme (project number DP140103256).
References
[1]
[2]
[3]
[4]
R. Agarwal, A. Khandelwal, and I. Stoica. “Succinct: Enabling Queries on Compressed
Data”. In: USENIX Symp. Networked Systems Design and Impl. 2015, pp. 337–350.
V. N. Anh and A. Moffat. “Compressed Inverted Files with Reduced Decoding Overheads”.
In: Proc. ACM SIGIR Int. Conf. Information Retrieval. 1998, pp. 290–297.
A. Apostolico, M. Crochemore, M. Farach-Colton, Z. Galil, and S. Muthukrishnan. “Forty
Years of Suffix Trees”. In: C. ACM 59.4 (Mar. 2016), pp. 66–73.
J. Barbay, T. Gagie, G. Navarro, and Y. Nekrich. “Alphabet Partitioning for Compressed
Rank/Select and Applications”. In: Proc. Int. Symp. Alg. and Comp. 2010, pp. 315–326.
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
P. Ferragina and G. Manzini. “Opportunistic Data Structures with Applications”. In: Proc.
IEEE Symp. Foundations of Comp. Science. 2000, pp. 390–398.
P. Ferragina, R. González, G. Navarro, and R. Venturini. “Compressed Text Indexes: From
Theory to Practice”. In: J. Exp. Alg. 13 (2008).
S. Gog, T. Beller, A. Moffat, and M. Petri. “From Theory to Practice: Plug and Play with
Succinct Data Structures”. In: Proc. Symp. Experimental Alg. 2014, pp. 326–337.
S. Gog, J. Kärkkäinen, D. Kempa, M. Petri, and S. J. Puglisi. “Faster, Minuter”. In: Proc.
Data Compression Conf. 2016, pp. 53–62.
S. Gog, G. Navarro, and M. Petri. “Improved and Extended Locating Functionality on
Compressed Suffix Arrays”. In: J. Discrete Alg. 32 (2015), pp. 53–63.
A. Golynski, J. I. Munro, and S. S. Rao. “Rank/Select Operations on Large Alphabets: A
Tool for Text Indexing”. In: Proc. ACM-SIAM Symp. Discrete Alg. 2006, pp. 368–373.
R. Grossi, A. Gupta, and J. S. Vitter. “High-Order Entropy-Compressed Text Indexes”. In:
Proc. ACM-SIAM Symp. Discrete Alg. 2003, pp. 841–850.
R. Grossi and J. S. Vitter. “Compressed Suffix Arrays and Suffix Trees with Applications
to Text Indexing and String Matching”. In: Proc. ACM Symp. Theory of Comput. 2000,
pp. 397–406.
H. Huo, L. Chen, J. S. Vitter, and Y. Nekrich. “A Practical Implementation of Compressed
Suffix Arrays with Applications to Self-Indexing”. In: Proc. Data Compression Conf.
2014, pp. 292–301.
J. Kärkkäinen, D. Kempa, and S. J. Puglisi. “Hybrid Compression of Bitvectors for the
FM-Index”. In: Proc. Data Compression Conf. 2014, pp. 302–311.
V. Mäkinen and G. Navarro. “Succinct Suffix Arrays Based on Run-Length Encoding”. In:
Proc. Symp. Combinatorial Pattern Matching. 2005, pp. 45–56.
U. Manber and G. W. Myers. “Suffix Arrays: A New Method For On-Line String Searches”.
In: SIAM J. Comp. 22.5 (1993), pp. 935–948.
A. Moffat and S. Gog. “String Search Experimentation Using Massive Data”. In: Philosophical Trans. Royal Society A 372.8 (2014).
A. Moffat and A. Turpin. Compression and Coding Algorithms. Boston, MA: Kluwer
Academic Publishers, 2002.
G. Navarro and V. Mäkinen. “Compressed Full-Text Indexes”. In: ACM Comp. Surveys
39.1 (2007).
D. Okanohara and K. Sadakane. “Practical Entropy-Compressed Rank/Select Dictionary”.
In: Proc. Wkshp. Alg. Engineering and Experiments. 2007.
G. Ottaviano and R. Venturini. “Partitioned Elias-Fano Indexes”. In: Proc. ACM SIGIR Int.
Conf. Information Retrieval. 2014, pp. 273–282.
M. Petri, A. Moffat, P. C. Nagesh, and A. Wirth. “Access Time Tradeoffs in Archive
Compression”. In: Proc. Asian IR Societies Conf. 2015, pp. 15–28.
K. Sadakane. “New Text Indexing Functionalities of the Compressed Suffix Arrays”. In: J.
Alg. 48.2 (2003), pp. 294–313.
E. Shareghi, M. Petri, G. Haffari, and T. Cohn. “Compact, Efficient and Unlimited Capacity:
Language Modeling with Compressed Suffix Trees”. In: Proc. Conf. Empirical Methods
in Natural Language Proc. 2015, pp. 2409–2418.
S. Vigna. “Quasi-Succinct Indices”. In: Proc. ACM Conf. Web Search & Data Min. 2013,
pp. 83–92.
P. Weiner. “Linear Pattern Matching Algorithms”. In: Proc. SWAT. 1973, pp. 1–11.
Appendix
Details of Implementations
Table 4 provides details of the methods compared in Section 4. The CSA-SADA results
were obtained by executing code authored by Kunihiko Sadakane, available from the
Pizza&Chili web site.
Abbreviation
CSA
FM-HF-BVIL
FM-HF-HYB
FM-HF-RRR
FM-AP
FM-AP-HYB
FM-GMR
FM-GMR-RS
FM-FB-BVIL
FM-FB-HYB
CSA++
Composition
csa_sada<enc_vector<coder::elias_gamma,sΨ >,
1<<20,1<<20,sa_order_sa_sampling<>,isa_sampling<>>
csa_wt<wt_huff<bit_vector_il<bs>>,1<<20,1<<20>
csa_wt<wt_huff<hyb_vector<>>,1<<20,1<<20>
csa_wt<wt_huff<rrr_vector<b>>,1<<20,1<<20>
csa_wt_int<wt_ap< wt_huff<bit_vector,rank_support_v5<1>,
select_support_scan<1>,select_support_scan<0>>,
wm_int<bit_vector,rank_support_v5<1>,
select_support_scan<1>,select_support_scan<0>>>,
1<<20,1<<20>
csa_wt_int<wt_ap< wt_huff<hyb_vector<>>,
wm_int<hyb_vector<>> >,1<<20,1<<20>
csa_wt_int<wt_gmr<>,1<<20,1<<20>
csa_wt_int<wt_gmr_rs<>,1<<20,1<<20>
csa_wt<wt_fbb<bit_vector_il<bs>>,1<<20,1<<20>
csa_wt<wt_fbb<hyb_vector<>>,1<<20,1<<20>
csa_sada2<hyb_sd_vector<s>,1<<20,1<<20,
sa_order_sa_sampling<>,isa_sampling<>>
Table 4. SDSL descriptions of methods used in experiments. Sampling parameters b ∈
{15, 31, 63, 127}, bs ∈ {128, 256, 512, 1024}, s ∈ {16, 32, 64, 128, 256, 512, 1024}, and sΨ ∈
{16, 32, 64, 128, 512, 1024} were varied in the experiments to get different time-space trade-offs.
The last three class definitions are available in the hyb_sd_vector branch of the library.
| 8cs.DS
|
The Refinement Calculus of Reactive Systems∗
Viorel Preoteasa
Iulia Dragomir
Stavros Tripakis
arXiv:1710.03979v2 [cs.LO] 8 Feb 2018
February 9, 2018
Abstract
The Refinement Calculus of Reactive Systems (RCRS) is a compositional formal framework for modeling and reasoning about reactive systems. RCRS provides a language which allows to describe atomic
components as symbolic transition systems or QLTL formulas, and composite components formed using
three primitive composition operators: serial, parallel, and feedback. The semantics of the language
is given in terms of monotonic property transformers, an extension to reactive systems of monotonic
predicate transformers, which have been used to give compositional semantics to sequential programs.
RCRS allows to specify both safety and liveness properties. It also allows to model input-output systems
which are both non-deterministic and non-input-receptive (i.e., which may reject some inputs at some
points in time), and can thus be seen as a behavioral type system. RCRS provides a set of techniques
for symbolic computer-aided reasoning, including compositional static analysis and verification. RCRS
comes with a publicly available implementation which includes a complete formalization of the RCRS
theory in the Isabelle proof assistant.
Contents
1 Introduction
1.1 Novel contributions of this paper and relation to our prior work . . . . . . . . . . . . . . . . .
3
4
2 Preliminaries
5
3 Language
3.1 An Algebra of Components . . . . . . . . . . . .
3.2 Symbolic Transition System Components . . . .
3.2.1 General STS Components . . . . . . . . .
3.2.2 Variable Name Scope . . . . . . . . . . .
3.2.3 Stateless STS Components . . . . . . . .
3.2.4 Deterministic STS Components . . . . . .
3.2.5 Stateless Deterministic STS Components
3.3 Quantified Linear Temporal Logic Components .
3.3.1 QLTL . . . . . . . . . . . . . . . . . . . .
3.3.2 QLTL Components . . . . . . . . . . . . .
3.4 Well Formed Composite Components . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
6
7
7
9
10
10
11
11
12
13
14
∗ The bulk of this work was done while Preoteasa and Dragomir were at Aalto University. Preoteasa is currently at Space
Systems Finland. Dragomir is currently at Verimag. Tripakis is affiliated with Aalto University and the University of California
at Berkeley. This work was partially supported by the Academy of Finland and the U.S. National Science Foundation (awards
#1329759 and #1139138). Dragomir was supported during the writing of this paper by the Horizon 2020 Programme Strategic
Research Cluster (grant agreement #730080 and #730086).
1
4 Semantics
4.1 Monotonic Property Transformers . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Some Commonly Used MPTs . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Operators on MPTs: Function Composition, Product, and Fusion . . .
4.1.3 Novel Operators Used in Semantical Definition of Feedback . . . . . . .
4.1.4 Refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Subclasses of MPTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Relational MPTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Guarded MPTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Other subclasses and overview . . . . . . . . . . . . . . . . . . . . . . .
4.3 Semantics of Components as MPTs . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Example: Two Alternative Derivations of the Semantics of Diagram Sum
4.3.2 Characterization of Legal Input Traces . . . . . . . . . . . . . . . . . . .
4.3.3 Semantic Equivalence and Refinement for Components . . . . . . . . . .
4.3.4 Compositionality Properties . . . . . . . . . . . . . . . . . . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
of Fig.
. . . .
. . . .
. . . .
. .
. .
. .
. .
. .
. .
. .
. .
. .
. .
2.
. .
. .
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
15
15
17
17
18
19
19
20
20
21
22
23
24
24
5 Symbolic Reasoning
5.1 Symbolic Transformation of STS Components to QLTL Components . . . . . . . . . . . . . .
5.2 Symbolic Transformations of Special Atomic Components to More General Atomic Components
5.3 Symbolic Computation of Serial Composition . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 Symbolic Serial Composition of Two QLTL Components . . . . . . . . . . . . . . . . .
5.3.2 Symbolic Serial Composition of Two General STS Components . . . . . . . . . . . . .
5.3.3 Symbolic Serial Composition of Two Stateless STS Components . . . . . . . . . . . .
5.3.4 Symbolic Serial Composition of Two Deterministic STS Components . . . . . . . . . .
5.3.5 Symbolic Serial Composition of Two Stateless Deterministic STS Components . . . .
5.3.6 Symbolic Serial Composition of Two Arbitrary Atomic Components . . . . . . . . . .
5.3.7 Correctness of Symbolic Serial Composition . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Symbolic Computation of Parallel Composition . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1 Symbolic Parallel Composition of Two QLTL Components . . . . . . . . . . . . . . .
5.4.2 Symbolic Parallel Composition of Two General STS Components . . . . . . . . . . . .
5.4.3 Symbolic Parallel Composition of Two Stateless STS Components . . . . . . . . . . .
5.4.4 Symbolic Parallel Composition of Two Deterministic STS Components . . . . . . . . .
5.4.5 Symbolic Parallel Composition of Two Stateless Deterministic STS Components . . .
5.4.6 Symbolic Parallel Composition of Two Arbitrary Atomic Components . . . . . . . . .
5.4.7 Correctness of Symbolic Parallel Composition . . . . . . . . . . . . . . . . . . . . . . .
5.5 Symbolic Computation of Feedback Composition for Decomposable Deterministic STS Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.1 Decomposable Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.2 Symbolic Feedback of a Decomposable Deterministic STS Component . . . . . . . . .
5.5.3 Symbolic Feedback of a Decomposable Stateless Deterministic STS Component . . . .
5.5.4 Correctness of Symbolic Feedback Composition . . . . . . . . . . . . . . . . . . . . . .
5.6 Closure Properties of MPT Subclasses w.r.t. Composition Operators . . . . . . . . . . . . . .
5.7 Symbolic Simplification of Arbitrary Composite Components . . . . . . . . . . . . . . . . . .
5.7.1 Deterministic and Algebraic Loop Free Composite Components . . . . . . . . . . . . .
5.7.2 Correctness of the Simplification Algorithm . . . . . . . . . . . . . . . . . . . . . . . .
5.8 Checking Validity and Compatibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.9 Checking Input-Receptiveness and Computing Legal Inputs Symbolically . . . . . . . . . . .
5.10 Checking Refinement Symbolically . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
25
26
27
27
28
28
28
28
28
28
29
30
30
30
30
30
30
30
6 Toolset and Case Studies
39
7 Related Work
41
2
31
31
31
31
31
32
32
33
34
36
37
37
8 Conclusion
1
43
Introduction
This paper presents the Refinement Calculus of Reactive Systems (RCRS), a comprehensive framework for
compositional modeling of and reasoning about reactive systems. RCRS originates from the precursor theory
of synchronous relational interfaces [78, 79], and builds upon the classic refinement calculus [10]. A number
of conference publications on RCRS exist [66, 31, 67, 65, 33]. This paper collects some of these results and
extends them in significant ways. The novel contributions of this paper and relation to our previous work
are presented in §1.1.
The motivation for RCRS stems from the need for a compositional treatment of reactive systems. Generally speaking, compositionality is a divide-and-conquer principle. As systems grow in size, they grow in
complexity. Therefore dealing with them in a monolithic manner becomes unmanageable. Compositionality
comes to the rescue, and takes many forms. Many industrial-strength systems have employed for many years
mechanisms for compositional modeling. An example is the Simulink tool from the Mathworks. Simulink is
based on the widespread notation of hierarchical block diagrams. Such diagrams are both intuitive, and naturally compositional: a block can be refined into sub-blocks, sub-sub-blocks, and so on, creating hierarchical
models of arbitrary depth. This allows the user to build large models (many thousands of blocks) while at
the same time managing their complexity (at any level of the hierarchy, only a few blocks may be visible).
But Simulink’s compositionality has limitations, despite its hierarchical modeling approach. Even relatively simple problems, such as the problem of modular code generation (generating code for a block
independently from context) require techniques not always available in standard code generators [53, 52].
Perhaps more serious, and more relevant in the context of this paper, is Simulink’s lack of formal semantics, and consequent lack of rigorous analysis techniques that can leverage the advances in the fields of
computer-aided verification and programming languages.
RCRS provides a compositional formal semantics for Simulink in particular, and hierarchical block diagram notations in general, by building on well-established principles from the formal methods and programming language domains. In particular, RCRS relies on the notion of refinement (and its counterpart,
abstraction) which are both fundamental in system design. Refinement is a binary relation between components, and ideally characterizes substitutability: the conditions under which some component can replace
another component, without compromising the behavior of the overall system. RCRS refinement is compositional in the sense that it is preserved by composition: if A0 refines A and B 0 refines B, then the composition
of A0 and B 0 refines the composition of A and B.
RCRS can be viewed as a refinement theory. It can also be viewed as a behavioral type system, similar to
type systems for programming languages, but targeted to reactive systems. By behavioral we mean a type
system that can capture not just data types of input and output ports of components (bool, int, etc.), but
also complete specifications of the behavior of those components. As discussed more extensively in [79], such
a behavioral type system has advantages over a full-blown verification system, as it is more lightweight: for
instance, it allows type checking, which does not require the user to provide a formal specification of the
properties that the model must satisfy, the model must simply type-check.
To achieve this, it is essential, as argued in [79], for the framework to be able to express non-inputreceptive (also called non-input-enabled or non-input-complete) components, i.e., components that reject
some input values. RCRS allows this. For example, a square-root component can be described√in RCRS
alternatively as: (1) a non-input-receptive component C1 with input-output contract x ≥ 0 ∧ y = x (where
x, y are the input
√ and output variables, respectively); or (2) an input-receptive component C2 with contract
x ≥ 0 → y = x. Simple examples of type-checking in the RCRS context are the following: Connecting the
non-input-receptive square-root component C1 to a component which outputs x = −1 results in a type error
(incompatibility). Connecting C1 to a non-deterministic component which outputs an arbitrary value for x
(this can be specified by the formula/contract true) also results in a type error. Yet, in both of the above
cases C2 will output a non-deterministically chosen value, even though the input constraint is not satisfied.
RCRS allows both non-input-receptive and non-deterministic components. This combination results in a
3
game-theoretic interpretation of the composition operators, like in interface theories [22, 79]. Refinement also
becomes game-theoretic, as in alternating refinement [7]. Game-theoretic composition can be used for an
interesting form of type inference. For example, if we connect the non-input-receptive square-root component
C1 above to a non-deterministic component C3 with input-output contract x ≥ u + 1 (where x is the output
of C3 , and u its input), and apply the (game-theoretic) serial composition of RCRS, we obtain the condition
u ≥ −1 on the external input of the overall composition. The constraint u ≥ −1 represents the weakest
condition on u which ensures compatibility of the connected components.
In a nutshell, RCRS consists of the following elements:
1. A modeling language (syntax), which allows to describe atomic components, and composite components
formed by a small number of primitive composition operators (serial, parallel, and feedback). The
language is described in §3.
2. A formal semantics, presented in §4. Component semantics are defined in terms of monotonic property
transformers (MPTs). MPTs are extensions of monotonic predicate transformers used in theories of
programming languages, and in particular in refinement calculus [10]. Predicate transformers transform
sets of post-states (states reached by the program after its computation) into sets of pre-states (states
where the program begins). Property transformers transform sets of a component’s output traces
(infinite sequences of output values) into sets of input traces (infinite sequences of input values). Using
this semantics we can express both safety and liveness properties.
3. A set of symbolic reasoning techniques, described in §5. In particular, RCRS offers techniques to
• compute the symbolic representation of a composite component from the symbolic representations
of its sub-components;
• simplify composite components into atomic components;
• reduce checking refinement between two components to checking satisfiability of certain logical
formulas;
• reduce input-receptiveness and compatibility checks to satisfiability;
• compute the legal inputs of a component symbolically.
We note that these techniques are for the most part logic-agnostic, in the sense that they do not depend
on the particular logic used to represent components.
4. A toolset, described briefly in §6. The toolset consists mainly of:
• a full implementation of the RCRS theory in the Isabelle proof assistant [61];
• a translator of Simulink diagrams into RCRS code.
Our implementation is open-source and publicly available from http://rcrs.cs.aalto.fi.
1.1
Novel contributions of this paper and relation to our prior work
Several of the ideas behind RCRS originated in the theory of synchronous relational interfaces [78, 79]. The
main novel contributions of RCRS w.r.t. that theory are: (1) RCRS is based on the semantic foundation
of monotonic property transformers, whereas relational interfaces are founded on relations; (2) RCRS can
handle liveness properties, whereas relational interfaces can only handle safety; (3) RCRS has been completely
formalized and most results reported in this and other RCRS papers have been proven in the Isabelle proof
assistant; (4) RCRS comes with a publicly available toolset (http://rcrs.cs.aalto.fi) which includes
the Isabelle formalization, a Translator of Simulink hierarchical block diagrams [31, 33], and a Formal
Analyzer which performs, among other functions, compatibility checking, refinement checking, and automatic
simplification of RCRS contracts.
4
RCRS was introduced in [66], which focuses on monotonic property transformers as a means to extend
relational interfaces with liveness properties. [66] covers serial composition, but not parallel nor feedback. It
also does not cover symbolic reasoning nor the RCRS implementation. Feedback is considered in [67], whose
aim is in particular to study instantaneous feedback for non-deterministic and non-input-receptive systems.
The study of instantaneous feedback is an interesting problem, but beyond the scope of the current paper.
In this paper we consider non-instantaneous feedback, i.e., feedback for systems without same-step cyclic
dependencies (no algebraic loops).
[31] presents part of the RCRS implementation, focusing on the translation of Simulink (and hierarchical
block diagrams in general) into an algebra of components with three composition primitives, serial, parallel,
and feedback, like RCRS. As it turns out, there is not a unique way to translate a graphical notation like
Simulink into an algebraic formalism like RCRS. The problem of how exactly to do it and what are the
trade-offs is an interesting one, but beyond the scope of the current paper. This problem is studied in
depth in [31] which proposes three different translation strategies and evaluates their pros and cons. [31]
leaves open the question whether the results obtained by the different translations are equivalent. [64] settles
this question, by proving that a class of translations, including the ones proposed in [31] are semantically
equivalent for any input block diagram. [65] also concerns the RCRS implementation, discussing solutions
to subtle typing problems that arise when translating Simulink diagrams into RCRS/Isabelle code.
In summary, the current paper does not cover the topics covered in [31, 64, 65, 67] and can be seen as a
significantly revised and extended version of [66]. The main novel contributions with respect to [66] are the
following: (1) a language of components (§3); (2) a revised MPT semantics (§4), including in particular novel
operators for feedback (§4.1.3) and a classification of MPT subclasses (§4.2); (3) a new section of symbolic
reasoning (§5).
2
Preliminaries
Sets, types. We use capital letters X, Y , Σ, . . . to denote types or sets, and small letters to denote
elements of these types x ∈ X, y ∈ Y , etc. We denote by B the type of Boolean values true and false. We
use ∧, ∨, ⇒, and ¬ for the Boolean operations. The type of natural numbers is denoted by N, while the
type of real numbers is denoted by R. The Unit type contains a single element denoted ().
Cartesian product. For types X and Y , X × Y is the Cartesian product of X and Y , and if x ∈ X and
y ∈ Y , then (x, y) is a tuple from X × Y . The empty Cartesian product is Unit. We assume that we have
only flat products X1 × . . . × Xn , and then we have
(X1 × . . . × Xn ) × (Y1 × . . . × Ym ) = X1 × . . . × Xn × Y1 × . . . × Ym
Functions. If X and Y are types, X → Y denotes the type of functions from X to Y . The function type
constructor associates to the right (e.g., X → Y → Z = X → (Y → Z)) and the function interpretation
associates to the left (e.g., f (x)(y) = (f (x))(y)). In order to construct functions we use lambda notation,
e.g., (λx, y : x + y + 1) : N → N → N. Similarly, we can have tuples in the definition of functions, e.g.,
(λ(x, y) : x + y + 2) : (N × N) → N. The composition of two functions f : X → Y and g : Y → Z, is a
function denoted g ◦ f : X → Z, where (g ◦ f )(x) = g(f (x)).
Predicates. A predicate is a function returning Boolean values, e.g., p : X → Y → B, p(x)(y) = (x = y).
We define the smallest predicate ⊥ : X → B where ⊥(x) = false for all x ∈ X. The greatest predicate is
> : X → B, with >(x) = true for all x ∈ X. We will often interpret predicates as sets. A predicate p : X → B
can be viewed as the set of all x ∈ X such that p(x) = true. For example, viewing two predicates p, q as sets,
we can write p ⊆ q, meaning that for all x, p(x) ⇒ q(x).
5
Relations. A relation is a predicate with at least two arguments, e.g., r : X → Y → B. For such a relation
r, we denote by in(r) : X → B the predicate in(r)(x) = (∃y : r(x)(y)). If the relation r has more than two
arguments, then we define in(r) similarly by quantifying over the last argument.
We extend point-wise all operations on Booleans to operations on predicates and relations. For example,
if r, r0 : X → Y → B are two relations, then r ∧ r0 and r ∨ r0 are the relations given by (r ∧ r0 )(x)(y) =
r(x)(y) ∧ r0 (x)(y) and (r ∨ r0 )(x)(y) = r(x)(y) ∨ r0 (x)(y). We also introduce the order on relations r ⊆ r0 =
(∀x, y : r(x)(y) ⇒ r0 (x)(y)).
The composition of two relations r : X → Y → B and r0 : Y → Z → B is a relation (r ◦ r0 ) : X → Z → B,
where (r ◦ r0 )(x)(z) = (∃y : r(x)(y) ∧ r0 (y)(z)).
Infinite sequences. If Σ is a type, then Σω = (N → Σ) is the set of all infinite sequences over Σ, also
called traces. For a trace σ ∈ Σω , let σi = σ(i) be the i-th element in the trace. Let σ i ∈ Σω denote the
suffix of σ starting from the i-th step, i.e., σ i = σi σi+1 · · · . We often view a pair of traces (σ, σ 0 ) ∈ Σω × Σ0ω
as being also a trace of pairs (λi : (σi , σi0 )) ∈ (Σ × Σ0 )ω .
Properties. A property is a predicate p over a set of infinite sequences. Formally, p ∈ (Σω → B). Just
like any other predicate, a property can also be viewed as a set. In particular, a property can be viewed as
a set of traces.
3
Language
3.1
An Algebra of Components
We model systems using a simple language of components. The grammar of the language is as follows:
component
atomic component
STS component
::=
::=
::=
composite component
::=
atomic component | composite component
STS component | QLTL component
GEN STS component | STATELESS STS component
| DET STS component | DET STATELESS STS component
component ; component | component k component | fdbk(component)
The elements of the above grammar are defined in the remaining of this section, where examples are
also given to illustrate the language. In a nutshell, the language contains atomic components of two kinds:
atomic components defined as symbolic transition systems (STS component), and atomic components defined
as QLTL formulas over input and output variables (QLTL component).
STS components are split in four categories: general STS components, stateless STS components, deterministic STS components, and deterministic stateless STS components. Semantically, the general STS
components subsume all the other more specialized STS components, but we introduce the specialized syntax
because symbolic compositions of less general components become simpler, as we shall explain in the sequel
(see §5).
Also, as it turns out, atomic components of our framework form a lattice, shown in Fig. 5, from the
more specialized ones, namely, deterministic stateless STS components, to the more general ones, namely
QLTL components. The full definition of this lattice will become apparent once we provide a symbolic
transformation of STS to QLTL components, in §5.1.
Apart from atomic components, the language also allows to form composite components, by composing
(atomic or other composite) components via three composition operators: serial ;, parallel k, and feedback
fdbk, as depicted in Fig. 1. The serial composition of two components C, C 0 is formed by connecting the
output(s) of C to the input(s) of C 0 . Their parallel composition is formed by “stacking” the two components
on top of each other without forming any new connections. The feedback of a component C is obtained by
connecting the first output of C to its first input.
Our language is inspired by graphical notations such as Simulink, and hierarchical block diagrams in
general. But our language is textual, not graphical. An interesting question is how to translate a graphical
6
x
x
y
C
y
C
0
z
u
(a) Serial composition: C ; C 0
y
x1
C
x2
..
.
v
C0
(b) Parallel composition: C k C 0
y1
C
y2
..
.
(c) Feedback composition: fdbk(C)
Figure 1: The three composition operators of RCRS.
block diagram into a term in our algebra. We will not address this question here, as the issue is quite
involved. We refer the reader to [31], which includes an extensive discussion on this topic. Suffice it to say
here that there are generally many possible translations of a graphical diagram into a term in our algebra (or
generally any algebra that contains primitive serial, parallel, and feedback composition operators). These
translations achieve different tradeoffs in terms of size, readability, computational properties, and so on.
See [31] for details.
1
1
In
z
Add
UnitDelay
1
Out
Figure 2: A Simulink diagram modeling the 1-step delayed sum of its input In. Each atomic block as well as
the entire system can be formalized as STS components (see §3.2). The entire system can also be formalized
as a composite component (see below).
As an example, consider the Simulink diagram shown in Fig. 2. This diagram can be represented in our
language as a composite component Sum defined as
Sum = fdbk(Add ; UnitDelay ; Split)
where Add, UnitDelay, and Split are atomic components (for a definition of these atomic components see
§3.2). Here Split models the “fan-out” element in the Simulink diagram (black bullet) where the output
wire of UnitDelay splits in two wires going to Out and back to Add.1
3.2
Symbolic Transition System Components
We introduce all four categories of STS components and at the same time provide syntactic mappings from
specialized STS components to general STS components.
3.2.1
General STS Components
A general symbolic transition system component (general STS component) is a transition system described
symbolically, with Boolean expressions over input, output, state, and next state variables defining the initial
states and the transition relation. When we say “Boolean expression” (here and in the definitions that
follow) we mean an expression of type B, in some arbitrary logic, not necessarily restricted to propositional
logic. For example, if x is a variable of numerical type, then x > 0 is a Boolean expression. In the definition
that follows, s0 denotes the primed, or next state variable, corresponding to the current state variable s. Both
1 Note that the Simulink input and output ports In and Out are not explicitly represented in Sum. They are represented
implicitly: In corresponds to the second input of Add, which carries over as the unique external input of Sum (thus, Sum is an
“open system” in the sense that it has open, unconnected, inputs); Out corresponds to the second output of Split, which carries
over as the unique external output of Sum.
7
can be vectors of variables. For example, if s = (s1 , s2 ) then s0 = (s01 , s02 ). We assume that s0 has the same
type as s.
Definition 1 (STS component). A (general) STS component is a tuple
sts(x : Σx , y : Σy , s : Σs , init exp : B, trs exp : B)
where x, y, s are input, output and state variables (or tuples of variables) of types Σx , Σy , Σs , respectively,
init exp is a Boolean expression on s (in some logic), and trs exp is a Boolean expression on x, y, s, s0 (in
some logic).
Intuitively, an STS component is a non-deterministic system which for an infinite input sequence σx ∈ Σω
x
produces as output an infinite sequence σy ∈ Σω
y . The system starts non-deterministically at some state
σs (0) satisfying init exp. Given first input σx (0), the system non-deterministically computes output σy (0)
and next state σs (1) such that trs exp holds (if no such values exist, then the input σx (0) is illegal, as
discussed in more detail below). Next, it uses the following input σx (1) and state σs (1) to compute σy (1)
and σs (2), and so on.
We will sometimes use the term contract to refer to the expression trs exp. Indeed, trs exp can be seen
as specifying a contract between the component and its environment, in the following sense. At each step in
the computation, the environment must provide input values that do not immediately violate the contract,
i.e., for which we can find values for the next state and output variables to satisfy trs exp. Then, it is the
responsibility of the component to find such values, otherwise it is the component’s “fault” if the contract
is violated. This game-theoretic interpretation is similar in spirit with the classic refinement calculus for
sequential programs [10].
We use Σx , Σy , Σs in the definition above to emphasize the types of the input, output and the state, and
the fact that, when composing components, the types should match. However, in practice we often omit the
types, unless they are required to unambiguously specify a component. Also note that the definition does not
fix the logic used for the expressions init exp and trs exp. Indeed, our theory and results are independent
from the choice of this logic. The choice of logic matters for algorithmic complexity and decidability. We
will return to this point in §5. Finally, for simplicity, we often view the formulas init exp and trs exp as
semantic objects, namely, as predicates. Adopting this view, init exp becomes the predicate init : Σs → B,
and trs exp the predicate trs exp : Σs → Σx → Σs → Σy → B. Equivalently, trs exp can be interpreted as
a relation trs exp : (Σs × Σx ) → (Σs × Σy ) → B.
Throughout this paper we assume that init exp is satisfiable, meaning that there is at least one valid
initial state.
Examples. In the examples provided in this paper, we often specify systems that have tuples as input,
state and output variables, in different equivalent ways. For example, we can introduce a general STS
component with two inputs as sts((n : N, x : R), s : R, y : R, s > 0, s0 > s ∧ y + s = xn ), but also as
sts((n, x) : N × R, s : R, y : R, s > 0, s0 > s ∧ y + s = xn ), or sts(z : N × R, y : R, s : R, s > 0, s0 > s ∧ y + s =
snd(z)fst(z) ), where fst and snd return the first and second elements of a pair.
Let us model a system that at every step i outputs the input received at previous step i − 1 (assume
that the initial output value is 0). This corresponds to Simulink’s commonly used UnitDelay block, which
is also modeled in the diagram of Fig. 2. This block can be represented by an STS component, where a state
variable s is needed to store the input at moment i such that it can be used at the next step. We formally
define this component as
UnitDelay = sts(x, y, s, s = 0, y = s ∧ s0 = x).
We use first-order logic to define init exp and trs exp. Here init exp is s = 0, which initializes the state
variable s with the value 0. The trs exp is y = s ∧ s0 = x, that is at moment i the current state variable s
which stores the input x received at moment i − 1 is outputed and its value is updated.
As another example, consider again the composite component Sum modeling the diagram of Fig. 2. Sum
could also be defined as an atomic STS component:
Sum = sts(x, y, s, s = 0, y = s ∧ s0 = s + x).
8
In §5 we will show how we can automatically and symbolically simplify composite component terms such
as fdbk(Add ; UnitDelay ; Split), to obtain syntactic representations of atomic components such as the one
above.
These examples illustrate systems coming from Simulink models. However, our language is more general,
and able to accommodate the description of other systems, such as state machines à la nuXmv [16], or
input/output automata [54]. In fact, both UnitDelay and Sum are deterministic, so they could also be
defined as deterministic STS components, as we will see below. Our language can capture non-deterministic
systems easily. An example of a non-deterministic STS component is the following:
C = sts(x, y, s, s = 0, x + s ≤ y).
For an input sequence σx ∈ Nω , C outputs a non-deterministically chosen sequence σy such that the transition
expression x + s ≤ y is satisfied. Since there is no formula in the transition expression tackling the next state
variable, this is updated also non-deterministically with values from N.
Our language can also capture non-input-receptive systems, that is, systems which disallow some input
values as illegal. For instance, a component performing division, but disallowing division by zero, can be
specified as follows:
x
Div = sts((x, y), z, (), true, y 6= 0 ∧ z = ).
y
Note that Div has an empty tuple of state variables, s = (). Such components are called stateless, and are
introduced in the sequel.
Even though RCRS is primarily a discrete-time framework, we have used it to model and verify continuoustime systems such as those modeled in Simulink (see §6). We do this by discretizing time using a time step
parameter ∆t > 0 and applying Euler numerical integration. Then, we can model Simulink’s Integrator
block in RCRS as an STS component parameterized by ∆t:
Integrator∆t
= sts x, y, s, s = 0, y = s ∧ s0 = s + x · ∆t
More complex dynamical system blocks can be modeled in a similar fashion. For instance, Simulink’s
Transfer Fcn block, with transfer function
s2 + 2
0.5s2 + 2s + 1
can be modeled in RCRS as the following STS component parameterized by ∆t:
TransferFcn∆t
where
= sts x, y, (s1 , s2 ), s1 = 0 ∧ s2 = 0, trs)
trs =
(y = −8 · s1 + 2 · x) ∧
(s01 = s1 + (−4 · s1 − 2 · s2 + x) · ∆t) ∧
(s02 = s2 + s1 · ∆t)
3.2.2
Variable Name Scope
We remark that variable names in the definition of atomic components are local. This holds for all atomic
components in the language of RCRS (including STS and QLTL components, defined in the sequel). This
means that if we replace a variable with another one in an atomic component, then we obtain a semantically equivalent component. For example, the two STS components below are equivalent (the semantical
equivalence symbol ≡ will be defined formally in Def. 20, once we define the semantics):
sts((x, y), z, s, s > 0, z > s + x + y) ≡ sts((u, v), w, t, t > 0, w > t + u + v)
9
3.2.3
Stateless STS Components
A special STS component is one that has no state variables:
Definition 2 (Stateless STS component). A stateless STS component is a tuple
C = stateless(x : Σx , y : Σy , io exp : B)
where x, y are the input and output variables, and io exp is a Boolean expression on x and y. Stateless STS
components are special cases of general STS components, as defined by the mapping stateless2sts:
stateless2sts(C) = sts(x, y, (), true, io exp).
Examples. A trivial stateless STS component is the one that simply transfers its input to its output (i.e.,
a “wire”). We denote such a component by Id, and we formalize it as
Id = stateless(x, y, y = x).
Another simple example is a component with no inputs and a single output, which always outputs a
constant value c (of some type). This can be formalized as the following component parameterized by c:
Constc = stateless((), y, y = c).
Component Add from Fig. 2, which outputs the sum of its two inputs, can be modeled as a stateless STS
component:
Add = stateless((x, y), z, z = x + y).
Component Split from Fig. 2 can also be modeled as a stateless STS component:
Split = stateless(x, (y, z), y = x ∧ z = x).
The Div component introduced above is stateless, and therefore can be also specified as follows:
x
.
Div = stateless (x, y), z, y 6= 0 ∧ z =
y
The above examples are not only stateless, but also deterministic. We introduce deterministic STS
components next.
3.2.4
Deterministic STS Components
Deterministic STS components are those which, for given current state and input, have at most one output
and next state. Syntactically, they are introduced as follows:
Definition 3 (Deterministic STS component). A deterministic STS component is a tuple
det(x : Σx , s : Σs , a : Σs , inpt exp : B, next exp : Σs , out exp : Σy )
where x, s are the input and state variables, a ∈ Σs is the initial value of the state variable, inpt exp is a
Boolean expression on s and x defining the legal inputs, next exp is an expression of type Σs on x and s
defining the next state, and out exp is an expression of type Σy on x and s defining the output. Deterministic
STS components are special cases of general STS components, as defined by the mapping det2sts:
det2sts(C) = (x, y, s, s = a, inpt exp ∧ s0 = next exp ∧ y = out exp)
where y is a new variable name (or tuple of new variable names) of type Σy .
Note that a deterministic STS component has a separate expression inpt exp to define legal inputs. A
separate such expression is not needed for general STS components, where the conditions for legal inputs
are part of the expression trs exp. For example, compare the definition of Div as a general STS above, and
as a stateless deterministic STS below (see §3.2.5).
10
Examples. As mentioned above, all three components, UnitDelay, Add, and Split from Fig. 2, as well as
Div and Const, are deterministic. They could therefore be specified in our language as deterministic STS
components, instead of general STS components:
UnitDelay = det x, s, 0, true, x, s
Constc = det (), (), (), true, (), c
Add = det (x, y), (), (), true, (), x + y
Split = det x, (), (), true, (), (x, x)
x
Div = det (x, y), (), (), y 6= 0, (),
y
The component Sum modeling the entire system is also deterministic, and could be defined as a deterministic STS component:
Sum = det(x, s, 0, true, s + x, s).
Note that these alternative specifications for each of those components, although syntactically distinct,
will turn out to be semantically equivalent by definition, when we introduce the semantics of our language,
in §4.
3.2.5
Stateless Deterministic STS Components
STS components which are both deterministic and stateless can be specified as follows:
Definition 4 (Stateless deterministic STS component). A stateless deterministic STS component is a tuple
C = stateless det(x : Σx , inpt exp : B, out exp : Σy )
where x is the input variable, inpt exp is a Boolean expression on x defining the legal inputs, and out exp is
an expression of type Σy on x defining the output. Stateless deterministic STS components are special cases
of both deterministic STS components, and of stateless STS components, as defined by the mappings
stateless det2det(C)
=
det(x, (), (), inpt exp, (), out exp)
stateless det2stateless(C)
=
stateless(x, y, inpt exp ∧ y = out exp)
where y is a new variable name or a tuple of new variable names.
Examples. Many of the examples introduced above are both deterministic and stateless. They could be
specified as follows:
Id
= stateless det x, true, x
Constc = stateless det (), true, c
Add
= stateless det (x, y), true, x + y
Split = stateless det x, true, (x, x)
x
Div
= stateless det (x, y), y 6= 0,
y
3.3
Quantified Linear Temporal Logic Components
Although powerful, STS components have limitations. In particular, they cannot express liveness properties [4]. To remedy this, we introduce another type of components, based on Linear Temporal Logic
(LTL) [63] and quantified propositional LTL (QPTL) [76, 48], which extends LTL with ∃ and ∀ quantifiers
over propositional variables. In this paper we use quantified first-order LTL (which we abbreviate as QLTL).
QLTL further extends QPTL with functional and relational symbols over arbitrary domains, quantification
11
of variables over these domains, and a next operator applied to variables.2 We need this expressive power
in order to be able to handle general models (e.g., Simulink) which often use complex arithmetic formulas,
and also to be able to translate STS components into semantically equivalent QLTL components (see §5.1).
3.3.1
QLTL
QLTL formulas are generated by the following grammar. We assume a set of constants and functional
symbols (0, 1, . . ., true, false, +, . . .), a set of predicate symbols (=, ≤, <, . . .), and a set of variable names
(x, y, z, . . .).
Definition 5 (Syntax of QLTL). A QLTL formula ϕ is defined by the following grammar:
term ::=
ϕ
::=
x | y | ... |
(variable names)
0 | 1 | . . . | true | . . . |
(constants)
term + term | . . . |
(functional symbol application)
# term
(next applied to a term)
term = term | (term ≤ term) | . . . | (atomic QLTL formulas)
¬ϕ |
(negation)
ϕ∨ψ |
(disjunction)
ϕUψ|
(until)
∀x : ϕ
(forall)
As in standard first order logic, the bounded variables of a formula ϕ are the variables in scope of the
universal quantifier ∀, and the free variables of ϕ are those that are not bounded. The logic connectives ∧,
⇒ and ⇔ can be expressed with ¬ and ∨. Quantification is over atomic variables. The existential quantifier
∃ can be defined via the universal quantifier usually as ¬∀¬. The primitive temporal operators are next for
terms (# ) and until ( U ). As is standard, QLTL formulas are evaluated over infinite traces, and ϕ U ψ
intuitively means that ϕ continuously holds until some point in the trace where ψ holds.
Formally, we will define the relation σ |= ϕ (σ satisfies ϕ) for a QLTL formula ϕ over free variables x, y, . . .,
and an infinite sequence σ ∈ Σω , where Σ = Σx × Σy × . . ., and Σx , Σy , . . . are the types (or domains) of
variables x, y, . . .. As before we assume that σ can be written as a tuple of sequences (σx , σy , . . .) where
ω
σ x ∈ Σω
x , σy ∈ Σy , . . .. The semantics of a term t on variables x, y, . . . is a function from infinite sequences
to infinite sequences hhtii : Σω → Σω
t , where Σ = Σx × Σy × . . ., and Σt is the type of t. When giving the
semantics of terms and formulas we assume that constants, functional symbols, and predicate symbols have
the standard semantics. For example, we assume that +, ≤, . . . on numeric values have the semantics of
standard arithmetic.
Definition 6 (Semantics of QLTL). Let x be a variable, t, t0 be terms, ϕ, ψ be QLTL formulas, P be a
predicate symbol, f be a functional symbol, c be a constant, and σ ∈ Σω be an infinite sequence.
hhxii(σ)
hhcii(σ)
hhf (t, t0 )ii(σ)
hh# tii(σ)
σ
σ
σ
σ
σ
|= P (t, t0 )
|= ¬ϕ
|= ϕ ∨ ψ
|= ϕ U ψ
|= (∀x : ϕ)
:= σx
:= (λi : c)
:= (λi : f (hhtii(σ)(i), hht0 ii(σ)(i)))
:= hhtii(σ 1 )
:=
:=
:=
:=
:=
P (hhtii(σ)(0), hht0 ii(σ)(0))
¬ σ |= ϕ
σ |= ϕ ∨ σ |= ψ
(∃n ≥ 0 : (∀ 0 ≤ i < n : σ i |= ϕ) ∧ σ n |= ψ)
(∀σx ∈ Σω
x : (σx , σ) |= ϕ)
Other temporal operators can be defined as follows. Eventually (F ϕ = true U ϕ) states that ϕ must
hold in some future step. Always (G ϕ = ¬F ¬ϕ) states that ϕ must hold at all steps. The next operator for
2 A logic similar to the one that we use here is presented in [47], however in [47] the next operator can be applied only once
to variables, and the logic from [47] uses also past temporal operators.
12
formulas X can be defined using the next operator for terms # . The formula X ϕ is obtained by replacing
all occurrences of the free variables in ϕ by their next versions (i.e., x is replaced by # x, y by # y, etc.).
For example the propositional LTL formula X (x ∧ X y ⇒ G z) can be expressed as
(# x = true ∧ (# # y = true) ⇒ G (# z = true)).
We additionally introduce the operator: ϕ L ψ := ¬(ϕ U ¬ψ). Intuitively, ϕ L ψ holds if whenever ϕ
holds continuously up to some step n − 1, ψ must hold at step n.
Two QLTL formulas ϕ and ψ are semantically equivalent, denoted ϕ ⇐⇒ ψ, if
∀σ : (σ |= ϕ) ⇐⇒ (σ |= ψ).
Lemma 1. Let ϕ be a QLTL formula. Then:
1. (∃x : G ϕ) ⇐⇒ G (∃x : ϕ) when ϕ does not contain temporal operators.
2. ϕ L ϕ ⇐⇒ G ϕ
3. true L ϕ ⇐⇒ G ϕ
4. ϕ L true ⇐⇒ true
5. ϕ L false ⇐⇒ false
6. ∀y : (ϕ L ψ) ⇐⇒ (∃y : ϕ) L ψ, when ϕ does not contain temporal operators and y is not free in ψ.
The proof of the above result, as well as of most results that follow, is omitted. All omitted proofs have
been formalized and proved in the Isabelle proof assistant, and are available as part of the public distribution
of RCRS from http://rcrs.cs.aalto.fi. In particular, all results contained in this paper can be accessed
from the theory RCRS Overview.thy – either directly in that file or via references to the other RCRS files.
Examples. Using QLTL we can express safety, as well as liveness requirements. Informally, a safety
requirement expresses that something bad never happens. An example is the formula
thermostat = G (180◦ ≤ t ∧ t ≤ 220◦ ),
which states that the thermostat-controlled temperature t stays always between 180◦ and 220◦ .
A liveness requirement informally says that something good eventually happens. An example is the
formula F (t > 200◦ ) stating that the temperature t is eventually over 200◦ .
A more complex example is a formula modeling an oven that starts increasing the temperature from an
initial value of 20◦ until it reaches 180◦ , and then keeps it between 180◦ and 220◦ .
oven = (t = 20◦ ∧ ((t < # t ∧ t < 180) U thermostat)).
In this example the formula t < # t specifies that the temperature increases from some point to the next.
3.3.2
QLTL Components
A QLTL component is an atomic component where the input-output behavior is specified by a QLTL formula:
Definition 7 (QLTL component). A QLTL component is a tuple qltl(x : Σx , y : Σy , ϕ), where x, y are
input and output variables (or tuples of variables) of types Σx , Σy , and ϕ is a QLTL formula over x and y.
Intuitively a QLTL component C = qltl(x, y, ϕ) represents a system that takes as input an infinite
ω
sequence σx ∈ Σω
x and produces as output an infinite sequence σy ∈ Σy such that (σx , σy ) |= ϕ. If there
is no σy such that (σx , σy ) |= ϕ is true, then input σx is illegal for C, i.e., C is not input-receptive. There
could be many possible σy for a single σx , in which case the system is non-deterministic.
As a simple example, we can model the oven as a QLTL component with no input variables and the
temperature as the only output variable:
qltl((), t, oven)
13
3.4
Well Formed Composite Components
Not all composite components generated by the grammar introduced in §3.1 are well formed. Two components C and C 0 can be composed in series only if the number of outputs of C matches the number of inputs
of C 0 , and in addition the input types of C 0 are the same as the corresponding output types of C. Also,
fdbk can be applied to a component C if the type of the first output of C is the same as the type of its first
input. Formally, for every component C we define below Σin (C) - the input type of C, Σout (C) - the output
type of C, and wf (C) - the well-formedness of C, by induction on the structure of C:
Σin (sts(x : Σx , y : Σy , s : Σs , init, trs))
Σin (stateless(x : Σx , y : Σy , trs))
Σin (det(x : Σx , s : Σs , a, inpt, next, out : Σy ))
Σin (stateless det(x : Σx , inpt, out : Σy ))
Σin (qltl(x : Σx , y : Σy , ϕ))
Σin (C ; C 0 )
Σin (C k C 0 )
Σin (fdbk(C))
Σout (sts(x : Σx , y : Σy , s : Σs , init, trs))
Σout (stateless(x : Σx , y : Σy , trs))
Σout (det(x : Σx , s : Σs , a, inpt, next, out : Σy ))
Σout (stateless det(x : Σx , inpt, out : Σy ))
Σout (qltl(x : Σx , y : Σy , ϕ))
Σout (C ; C 0 )
Σout (C k C 0 )
Σout (fdbk(C))
wf (sts(x, y, s, init, trs))
=
wf (stateless(x, y, trs))
=
wf (det(x, s, a, inpt, next, out))
=
wf (stateless det(x, inpt, out)) =
wf (qltl(x, y, ϕ))
=
wf (C ; C 0 )
=
wf (C k C 0 )
=
wf (fdbk(C))
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
Σx
Σx
Σx
Σx
Σx
Σin (C)
Σin (C) × Σin (C 0 )
X2 × · · · × Xn provided Σin (C) = X1 × · · · × Xn
for some n ≥ 1
Σy
Σy
Σy
Σy
Σy
Σout (C 0 )
Σout (C) × Σout (C 0 )
Y2 × · · · × Yn provided Σout (C) = Y1 × · · · × Yn
for some n ≥ 1
true
true
true
true
true
wf (C) ∧ wf (C 0 ) ∧ Σout (C) = Σin (C 0 )
wf (C) ∧ wf (C 0 )
wf (C) ∧ Σin (C) = X × X1 · · · × Xn
∧ Σout (C) = X × Y1 · · · × Ym , for some n, m ≥ 0.
In the definition above, both n and m are natural numbers. If n = 0 then X1 × · · · × Xn denotes the Unit
type.
Note that atomic components are by definition well-formed. The composite components considered in
the sequel are required to be well-formed too.
We note that the above well-formedness conditions are not restrictive. Components that do not have
matching inputs and outputs can still be composed by adding appropriate switching components which
reorder inputs, duplicate inputs, and so on. An example of such a component is the component Split,
introduced earlier. As another example, consider the diagram in Fig. 3:
This diagram can be expressed in our language as the composite component:
A ; Switch1 ; (B k Id) ; Switch2 ; (C k Id)
14
A
B
C
Figure 3: Another block diagram.
where
Switch1 = stateless det((x, y), true, (x, y, x))
Switch2 = stateless det((u, v, x), true, (u, x, v))
4
Semantics
In RCRS, the semantics of components is defined in terms of monotonic property transformers (MPTs).
This is inspired by classical refinement calculus [10], where the semantics of sequential programs is defined in
terms of monotonic predicate transformers [28]. Predicate transformers are functions that transform sets of
post-states (states reached after the program executes) into sets of pre-states (states from which the program
begins). Property transformers map sets of output traces (that a component produces) into sets of input
traces (that a component consumes).
In this section we define MPTs formally, and introduce some basic operations on them, which are necessary for giving the semantics of components. The definitions of some of these operations (e.g., product
and fusion) are simple extensions of the corresponding operations on predicate transformers [10, 9]. Other
operations, in particular those related to feedback, are new (§4.1.3). The definition of component semantics
is also new (§4.3).
4.1
Monotonic Property Transformers
ω
A property transformer is a function S : (Σω
y → B) → (Σx → B), where Σx , Σy are input and output types
of the component in question. Note that x is the input and y is the output. A property transformer has
a weakest precondition interpretation: it is applied to a set of output traces Q ⊆ Σω
y , and returns a set of
,
such
that
all
traces
in
P
are
legal
and,
when
fed
to
the
component,
are guaranteed to
input traces P ⊆ Σω
x
produce only traces in Q as output.
Interpreting properties as sets, monotonicity of property transformers simply means that these functions
are monotonic with respect to set inclusion. That is, S is monotonic if for any two properties q, q 0 , if q ⊆ q 0
then S(q) ⊆ S(q 0 ).
For an MPT S we define its set of legal input traces as legal(S) = S(>), where > is the greatest predicate
extended to traces. Note that, because of monotonicity, and the fact that q ⊆ > holds for any property q,
we have that S(q) ⊆ legal(S) for all q. This justifies the definition of legal(S) as a “maximal” set of input
traces for which a system does not fail, assuming no restrictions on the post-condition. An MPT S is said
to be input-receptive if legal(S) = >.
4.1.1
Some Commonly Used MPTs
Definition 8 (Skip). Skip is defined to be the MPT such that for all q, Skip(q) = q.
Skip models the identity function, i.e., the system that accepts all input traces and simply transfers them
unchanged to the output (this will become more clear when we express Skip in terms of assert or update
transformers, below). Note that Skip is different from Id, defined above, although the two are strongly
related: Id is a component, i.e., a syntactic object, while Skip is an MPT, i.e., a semantic object. As we
shall see in §4.3, the semantics of Id is defined as Skip.
Definition 9 (Fail). Fail is defined to be the MPT such that for all q, Fail(q) = ⊥.
15
Recall that ⊥ is the predicate that returns false for any input. Thus, viewed as a set, ⊥ is the empty set.
Consequently, Fail can be seen to model a system which rejects all inputs, i.e., a system such that for any
output property q, there are no input traces that can produce an output trace in q.
Definition 10 (Assert). Let p ∈ Σω → B be a property. The assert property transformer {p} : (Σω → B) →
(Σω → B) is defined by
{p}(q) = p ∧ q.
The assert transformer {p} can be seen as modeling a system which accepts all input traces that satisfy
p, and rejects all others. For all the traces that it accepts, the system simply transfers them, i.e., it behaves
as the identity function.
To express MPTs such as assert transformers syntactically, let us introduce some notation. First, we can
use lambda notation for predicates, as in λ(σ, σ 0 ) : (σ = σ 0 ) for some predicate p : Σω → Σω → B which
returns true whenever it receives two equal traces. Then, instead of writing {λ(σ, σ 0 ) : (σ = σ 0 )} for the
corresponding assert transformer {p}, we will use the slightly lighter notation {σ, σ 0 | σ = σ 0 }.
ω
Definition 11 (Demonic update). Let r : Σω
x → Σy → B be a relation. The demonic update property
ω
ω
transformer [r] : (Σy → B) → (Σx → B) is defined by
[r](q) = {σ | ∀σ 0 : r(σ)(σ 0 ) ⇒ σ 0 ∈ q}.
That is, [r](q) contains all input traces σ which are guaranteed to result into an output trace in q when
fed into the (generally non-deterministic) input-output relation r. The term “demonic update” comes from
the refinement calculus literature [10].
Similarly to assert, we introduce a lightweight notation for the demonic update. If r is an expression in σ
and σ 0 , then [σ ; σ 0 | r] = [λ(σ, σ 0 ) : r]. For example, [σx , σy ; σz | ∀i : σz (i) = σx (i) + σy (i)] is the system
which produces as output the sequence σz = (λi : σx (i) + σy (i)), where σx and σy are the input sequences.
If e is an expression in σ, then [σ ; e] is defined to be [σ ; σ 0 | σ 0 = e], where σ 0 is a new variable different
from σ and which does not occur free in e. For example, [σ ; (λi : σ(i) + 1)] = [σ ; σ 0 | σ 0 = (λi : σ(i) + 1)].
The following lemma states that Skip can be defined as an assert transformer, or as a demonic update
transformer.
Lemma 2. Skip = [σ ; σ 0 | σ = σ 0 ] = {>} = {σ | true}.
In general Skip, Fail, and other property transformers are polymorphic with respect to their input and
output types. In Skip the input and output types must be the same. Fail, on the other hand, may have an
input type and a different output type.
ω
Definition 12 (Angelic update). Let r : Σω
x → Σy → B be a relation. The angelic update property
ω
ω
transformer {r} : (Σy → B) → (Σx → B) is defined by
{r}(q) = {σ | ∃σ 0 : r(σ)(σ 0 ) ∧ σ 0 ∈ q}.
An input sequence σ is in {r}(q) if there exists an output sequence σ 0 such that r(σ)(σ 0 ) and σ 0 ∈ q.
Notice the duality between the angelic and demonic update transformers. Consider, for example, a relation
r = {(σ, σ 0 ), (σ, σ 00 )}. If q = {σ 0 , σ 00 }, then {r}(q) = [r](q) = {σ}. If q = {σ 0 } then {r}(q) = {σ}, while
[r](q) = ∅.
We use a lightweight notation for the angelic update transformer, similar to the one for demonic update.
If r is an expression in σ and σ 0 , then {σ ; σ 0 | r} = {λ(σ, σ 0 ) : r}.
Lemma 3. Assert is a particular case of angelic update: {p} = {σ ; σ 0 | p(σ) ∧ σ = σ 0 }.
16
4.1.2
Operators on MPTs: Function Composition, Product, and Fusion
As we shall see in §4.3, the semantics of composition operators in the language of components will be defined
by the corresponding composition operators on MPTs. We now introduce the latter operators on MPTs.
First, we begin by the operators that have been known in the literature, and are recalled here. In §4.1.3 we
introduce some novel operators explicitly designed in order to handle feedback composition.
Serial composition of MPTs (and property transformers in general) is simply function composition. Let
ω
ω
ω
S : (Σω
y → B) → (Σx → B) and T : (Σz → B) → (Σy → B) be two property transformers. Then
ω
ω
S ◦ T : (Σz → B) → (Σx → B), is the function composition of S and T , i.e., ∀q : (S ◦ T )(q) = S(T (q)). Note
that serial composition preserves monotonicity, so that if S and T are MPTs, then S ◦ T is also an MPT.
Also note that Skip is the neutral element for serial composition, i.e., S ◦ Skip = Skip ◦ S = S.
To express parallel composition of components, we need a kind of Cartesian product operation on property
transformers. We define such an operation below. Similar operations for predicate transformers have been
proposed in [9].
ω
ω
ω
Definition 13 (Product). Let S : (Σω
y → B) → (Σx → B) and T : (Σv → B) → (Σu → B). The product of
ω
ω
ω
ω
S and T , denoted S ⊗ T : (Σy × Σv → B) → (Σx × Σu → B), is given by
0
ω
0
0
0
(S ⊗ T )(q) = {(σ, σ 0 ) | ∃p : Σω
y → B, p : Σv → B : p × p ⊆ q ∧ σ ∈ S(p) ∧ σ ∈ T (p )}
where (p × p0 )(σy , σv ) = p(σy ) ∧ p0 (σv ).
Lemma 4. For arbitrary S and T , S ⊗ T is monotonic.
The neutral element for the product composition is the Skip MPT that has Unit as input and output
type.
In order to define a feedback operation on MPTs, we first define two auxiliary operations: Fusion and
IterateOmega. Fusion is an extension of a similar operator introduced previously for predicate transformers
in [9]. IterateOmega is a novel operator introduced in the sequel.
ω
Definition 14 (Fusion). If S = {Si }i∈I , Si : (Σω
y → B) → (Σx → B) is a collection of MPTs, then the
ω
ω
fusion of S is the MPT Fusioni∈I (Si ) : (Σy → B) → (Σx → B) defined by
\
\
(Fusioni∈I (Si ))(q) = {σ | ∃p : I → Σω
pi ⊆ q ∧ σ ∈
Si (pi )}
y →B:
i
i
The Fusion operator satisfies the following property.
Lemma 5. For I 6= ∅ we have
Fusioni∈I ({pi } ◦ [ri ]) = {
\
i∈I
4.1.3
pi } ◦ [
\
ri ].
i∈I
Novel Operators Used in Semantical Definition of Feedback
The IterateOmega operator is defined as follows:
Definition 15 (IterateOmega).
IterateOmega(S) = Fusionn∈N (S n ◦ [σ ; σ 0 | ∀i : i + 1 < n ⇒ σi = σi0 ])
The feedback operator consists of connecting the first output of an MPT S with its first input. Formally,
feedback is defined as follows.
ω
Definition 16 (Feedback). Let S : (Σu × Σω
y → B) → (Σu × Σx → B) be an MPT. The feedback operator
on S, denoted Feedback(S), is given by the MPT
Feedback(S)
=
{σx ; σu , σy , σx }
◦ IterateOmega([σu , σy , σx ; σu , σx , σx ] ◦ (S ⊗ Skip))
◦ [σu , σy , σx ; σy ]
17
Example. As an example we show how to derive Feedback(S) for S = [σu , σx ; 0 · σx + 0 · σu , 0 · σx + 0 · σu ],
where σ + σ 0 = (λi : σ(i) + σ 0 (i)), and 0 · σ is 0 concatenated with σ. For now, we note that S is the semantics
of the composite component Add ; UnitDelay ; Split, which corresponds to the inner part of the diagram of
Fig. 2, before applying feedback. We will complete the formal definition of the the semantics of this diagram
in §4.3. For now, we focus on deriving Feedback(S), in order to illustrate how the Feedback operator works.
Let
T = [σu , σy , σx ; σu , σx , σx ] ◦ (S ⊗ Skip) = [σu , σy , σx ; 0 · σx + 0 · σu , 0 · σx + 0 · σu , σx ]
Then, we have
T ◦T
...
Tn
=
=
[σu , σy , σx ; 0 · σx + 0 · 0 · σx + 0 · 0 · σu , 0 · σx + 0 · 0 · σx + 0 · 0 · σu , σx ]
[σu , σy , σx ; 0 · σx + . . . + 0n · σx + 0n · σu , 0 · σx + . . . + 0n · σx + 0n · σu , σx ]
where 0n is a finite sequence of n 0s. We also have
=
T n ◦ [σ ; σ 0 | ∀i : i + 1 < n ⇒ σi = σi0 ]
[σu , σy , σx ; σ | ∀i : i + 1 < n ⇒ σ(i) = (Σj<i σx (j), Σj<i σx (j), σx (i))]
Then
Feedback(S)
=
=
=
=
{σx ; σu , σy , σx } ◦ IterateOmega(T ) ◦ [σu , σy , σx ; σy ]
{σx ; σu , σy , σx } ◦ Fusionn∈N (T n ◦ [σ ; σ 0 | ∀i : i + 1 < n ⇒ σi = σi0 ]) ◦ [σu , σy , σx ; σy ]
{σx ; σu , σy , σx } ◦ [σu , σy , σx ; σ | ∀i : σ(i) = (Σj<i σx (j), Σj<i σx (j), σx (i))] ◦ [σu , σy , σx ; σy ]
[σx ; σy | ∀i : σy (i) = Σj<i σx (j)]
Finally we obtain
Feedback(S) = [σx ; σy | ∀i : σy (i) = Σj<i σx (j)]
(1)
This is the system that outputs the trace (λi : Σj<i σx (j)) for input trace σx .
4.1.4
Refinement
A key element of RCRS, as of other compositional frameworks, is the notion of refinement, which enables
substitutability and other important concepts of compositionality. Semantically, refinement is defined as
follows:
ω
Definition 17 (Refinement). Let S, T : (Σω
y → B) → (Σx → B) be two MPTs. We say that T refines S (or
that S is refined by T ), written S v T , if and only if ∀q : S(q) ⊆ T (q).
All operations introduced on MPTs preserve the refinement relation:
Theorem 1. If S, T, S 0 , T 0 are MPTs of appropriate types such that S v S 0 and T v T 0 , then
1. S ◦ T v S 0 ◦ T 0 and S ⊗ T v S 0 ⊗ T 0 and Fusion(S, T ) v Fusion(S 0 , T 0 ) ([9, 10])
2. IterateOmega(S) v IterateOmega(S 0 )
3. Feedback(S) v Feedback(S 0 )
18
4.2
Subclasses of MPTs
Monotonic property transformers are a very rich and powerful class of semantic objects. In practice, the
systems that we deal with often fall into restricted subclasses of MPTs, which are easier to represent syntactically and manipulate symbolically. We introduce these subclasses next.
4.2.1
Relational MPTs
Definition 18 (Relational property transformers). A relational property transformer (RPT) S is an MPT
of the form {p} ◦ [r]. We call p the precondition of S and r the input-output relation of S.
Relational property T
transformers
T correspond to conjunctive transformers [10]. A transformer S is conjunctive if it satisfies S( i∈I qi ) = i∈I S(qi ) for all (qi )i∈I and I 6= ∅.
Fail, Skip, any assert transformer {p}, and any demonic update transformer [r], are RPTs. Indeed, Fail
can be written as {σ | false} ◦ [σ ; σ 0 | true]. Skip can be written as {σ | true} ◦ [σ ; σ]. The assert
transformer {p} can be written as the RPT {p} ◦ [σ ; σ]. Finally, the demonic update transformer [r] can
be written as the RPT {σ | true} ◦ [r]. Angelic update transformers are generally not RPTs: the angelic
update transformer {σ ; σ 0 | true} is not an RPT, as it is not conjunctive.
Examples. Suppose we wish to specify a system that performs division. Here are two possible ways to
represent this system with RPTs:
S1 = {>} ◦ [σx , σy ; σz | ∀i : σy (i) 6= 0 ∧ σz (i) =
σx (i)
]
σy (i)
S2 = {σx , σy | ∀i : σy (i) 6= 0} ◦ [σx , σy ; σz | σz (i) =
σx (i)
]
σy (i)
Although S1 and S2 are both relational, they are not equivalent transformers. S1 is input-receptive: it accepts
all input traces. However, if at some step i the input σy (i) is 0, then the output σz (i) is arbitrary (nondeterministic). In contrast, S2 is non-input-receptive as it accepts only those traces σy that are guaranteed
to be non-zero at every step, i.e., those that satisfy the condition ∀i : σy (i) 6= 0.
Theorem 2 (RPTs are closed under serial, product and fusion compositions). Let S = {p} ◦ [r] and
S 0 = {p0 } ◦ [r0 ] be two RPTs, with p, p0 , r and r0 of appropriate types. Then
S ◦ S 0 = {σ | p(σ) ∧ (∀σ 0 : r(σ)(σ 0 ) ⇒ p0 (σ 0 ))} ◦ [r ◦ r0 ]
and
and
S ⊗ S 0 = {σx , σy | p(σx ) ∧ p0 (σy )} ◦ [σx , σy ; σx0 , σy0 | r(σx )(σx0 ) ∧ r0 (σy )(σy0 )]
Fusion(S, S 0 ) = {p ∧ p0 } ◦ [r ∧ r0 ]
RPTs are not closed under Feedback. For example, we have
Feedback([σx , σz ; σx , σx ]) = {σ ; σ 0 | true}
which is a non-relational angelic update transformer as we said above.
Next theorem shows that the refinement of RPTs can be reduced to proving a first order property.
Theorem 3. For p, p0 , r, r0 of appropriate types we have.
{p} ◦ [r] v {p0 } ◦ [r0 ] ⇐⇒ (p ⊆ p0 ∧ (∀σx : p(σx ) ⇒ r0 (σx ) ⊆ r(σx )))
19
4.2.2
Guarded MPTs
Relational property transformers correspond to systems that have natural syntactic representations, as the
composition {p} ◦ [r], where the predicate p and the relation r can be represented syntactically in some
logic. Unfortunately, RPTs are still too powerful. In particular, they allow system semantics that cannot
be implemented. For example, consider the RPT Magic = [σ ; σ 0 | false]. It can be shown that for any
output property q (including ⊥), we have Magic(q) = >. Recall that, viewed as a set, > is the set of all
traces. This means that, no matter what the post-condition q is, Magic somehow manages to produce output
traces satisfying q no matter what the input trace is (hence the name “magic”). In general, an MPT S is
said to be non-miraculous (or to satisfy the law of excluded miracle) if S(⊥) = ⊥. We note that in [28],
sequential programs are modeled using predicate transformers that are conjunctive and satisfy the low of
excluded miracle.
We want to further restrict RPTs so that miraculous behaviour does not arise. Specifically, for an RPT
S = {p} ◦ [r] and an input sequence σ, if there is no σ 0 such that r(σ)(σ 0 ) is satisfied, then we want σ to be
illegal, i.e., we want p(σ) = false. We can achieve this by taking p to be in(r). Recall that if r : X → Y → B,
then in(r)(x) = (∃y : r(x)(y)). Taking p to be in(r) effectively means that p and r are combined into a
single specification r which can also restrict the inputs. This is also the approach followed in the theory of
relational interfaces [79].
Definition 19 (Guarded property transformers). The guarded property transformer (GPT) of a relation r
is an RPT, denoted {r], defined by {r] = {in(r)} ◦ [r].
It can be shown that an MPT S is a GPT if and only if S is conjunctive and non-miraculous [10]. Fail,
Skip, and any assert property transformer are GPTs. Indeed, Fail = {⊥] and Skip = {σ ; σ | >]. The
assert transformer can be written as {p} = {σ ; σ | p(σ)]. The angelic and demonic update property
transformers are generally not GPTs. The angelic update property transformer is not always conjunctive in
order to be a GPT. The demonic update property transformer is not in general a GPT because is not always
non-miraculous (Magic(⊥) = > =
6 ⊥). The demonic update transformer [r] is a GPT if and only if in(r) = >
and in this case we have [r] = {r].
Theorem 4 (GPTs are closed under serial and parallel compositions). Let S = {r] and S 0 = {r0 ] be two
GPTs with r and r0 of appropriate types. Then
S ◦ S 0 = {σx ; σz | in(r)(σx ) ∧ ∀σy : r(σx )(σy ) ⇒ in(r0 )(σy ) ∧ (r ◦ r0 )(σx , σz )]
and
S ⊗ S 0 = {σx , σy ; σx0 , σy0 | r(σx )(σx0 ) ∧ r0 (σy )(σy0 )]
GPTs are not closed under Fusion neither Feedback. Indeed, we have already seen in the previous section
that Feedback applied to the GPT [σx , σz ; σx , σx ] is not an RPT, and therefore not a GPT either. For the
fusion operator, we have Fusion([x ; 0], [x ; 1]) = [⊥], which is not a GPT.
A corollary of Theorem 3 is that refinement of GPTs can be checked as follows:
Corollary 1.
{r] v {r0 ] ⇐⇒ (in(r) ⊆ in(r0 ) ∧ (∀σx : in(r)(σx ) ⇒ r0 (σx ) ⊆ r(σx )))
4.2.3
Other subclasses and overview
The containment relationships among the various subclasses of MPTs are illustrated in Fig. 4. In addition
to the subclasses discussed above, we introduce several more subclasses of MPTs in the sections that follow,
when we assign semantics (in terms of MPTs) to the various atomic components in our component language.
For instance, QLTL components give rise to QLTL property transformers. Similarly, STS components,
stateless STS components, etc., give rise to corresponding subclasses of MPTs. The containment relationships
between these classes will be proven in the sections that follow. For ease of reference, we provide some forward
links to these results also here. The fact that QLTL property transformers are GPTs follows by definition
20
of the semantics of QLTL components: see §4.3, equation (2). The fact that STS property transformers are
a special case of QLTL property transformers follows from the transformation of an STS component into a
semantically equivalent QLTL component: see §5.1 and Theorem 7. The inclusions for subclasses of STS
property transformers follow by definition of the corresponding components (see also Fig. 5).
Monotonic Property Transformers
Relational Property Transformers
Guarded Property Transformers
QLTL Property Transformers
STS Property Transformers
Stateless STS Property Transformers
Deterministic STS Property Transformers
Figure 4: Overview of the property transformer classes and their containment relations.
4.3
Semantics of Components as MPTs
We are now ready to define the semantics of our language of components in terms of MPTs. Let C be a well
formed component. The semantics of C, denoted [[C]], is a property transformer of the form:
[[C]] : ((Σout (C))ω → B) → ((Σin (C))ω → B).
We define [[C]] by induction on the structure of C. First we give the semantics of QLTL components and
composite components:
[[qltl(x, y, ϕ)]]
0
=
{σx ; σy | (σx , σy ) |= ϕ]
(2)
0
[[C ; C ]]
=
[[C]] ◦ [[C ]]
(3)
[[C k C 0 ]]
=
[[C]] ⊗ [[C 0 ]]
(4)
[[fdbk(C)]]
=
Feedback([[C]])
(5)
The semantics of QLTL components satisfies the following property:
Lemma 6. If ϕ is a LTL formula on variables x and y, we have:
[[qltl(x, y, (∃y : ϕ) ∧ ϕ)]] = [[qltl(x, y, ϕ)]]
To define the semantics of STS components, we first introduce some auxiliary notation.
ω
Consider an STS component C = sts(x, y, s, init exp, trs exp). We define the predicate runC : Σω
s × Σx ×
ω
Σy → B as
runC (σs , σx , σy ) = (∀i : trs exp(σs (i), σx (i))(σs (i + 1), σy (i))).
ω
ω
Intuitively, if σx ∈ Σω
x is the input sequence, σy ∈ Σy is the output sequence, and σs ∈ Σs is the sequence
of values of state variables, then runC (σs , σx , σy ) holds if at each step of the execution, the current state,
current input, next state, and current output, satisfy the trs exp predicate.
21
We also formalize the illegal input traces of STS component C as follows:
illegalC (σx ) = (∃σs , σy , k : init exp(σs (0)) ∧ (∀i < k : trs exp(σs (i), σx (i))(σs (i + 1), σy (i))) ∧
¬in(trs exp)(σs (k), σx (k)))
Essentially, illegalC (σx ) states that there exists some point in the execution where the current state and
current input violate the precondition in(trs exp) of predicate trs exp, i.e., there exist no output and next
state to satisfy trs exp for that given current state and input.
Then, the semantics of an STS component C is given by:
[[C]] = {¬illegalC } ◦ [σx ; σy | (∃σs : init exp(σs (0)) ∧ runC (σs , σx , σy ))]
(6)
We give semantics to stateless and/or deterministic STS components using the corresponding mappings
from general STS components. If C is a stateless STS, C 0 is a deterministic STS, and C 00 is a stateless
deterministic STS, then:
[[C]]
0
[[C ]]
00
[[C ]]
=
=
=
[[stateless2sts(C)]]
(7)
0
[[det2sts(C )]]
(8)
00
[[stateless det2det(C )]]
(9)
Note that the semantics of a stateless deterministic STS component C 00 is defined by converting C 00 into
a deterministic STS component, by Equation (9) above. Alternatively, we could have defined the semantics
of C 00 by converting it into a stateless STS component, using the mapping stateless det2stateless. In
order for our semantics to be well-defined, we need to show that regardless of which conversion we choose,
we obtain the same result. Indeed, this is shown by the lemma that follows:
Lemma 7. For a stateless deterministic STS C 00 we have:
[[stateless det2det(C 00 )]] = [[stateless det2stateless(C 00 )]]
(10)
Observe that, by definition, the semantics of QLTL components are GPTs. The semantics of STS
components are defined as RPTs. However, they will be shown to be GPTs in §5.1. Therefore, the semantics
of all atomic RCRS components are GPTs. This fact, and the closure of GPTs w.r.t. parallel and serial
composition (Theorem 4), ensure that we stay within the GPT realm as long as no feedback operations are
used. In addition, as we shall prove in Corollary 2, components with feedback are also GPTs, as long as
they are deterministic and do not contain algebraic loops. An example of a component whose semantics is
not a GPT is:
C = fdbk(stateless((x, z), (y1 , y2 ), y1 = x ∧ y2 = x))
Then, we have [[C]] = Feedback([σx , σz ; σx , σx ]). As stated earlier, Feedback([σx , σz ; σx , σx ]) is equal to
{σ ; σ 0 | true}, which is not a GPT neither an RPT. The problem with C is that it contains an algebraic
loop: the first output y1 of the internal stateless component where feedback is applied directly depends on
its first input x. Dealing with such components is beyond the scope of this paper, and we refer the reader
to [67].
4.3.1
Example: Two Alternative Derivations of the Semantics of Diagram Sum of Fig. 2
To illustrate our semantics, we provide two alternative derivations of the semantics of the Sum system of
Fig. 2.
First, let us consider Sum as a composite component:
Sum = fdbk(Add ; UnitDelay ; Split)
22
where
Add
=
UnitDelay =
Split
=
stateless det((u, x), true, u + x)
det(x, s, 0, true, x, s)
stateless det(x, true, (x, x))
We have
[[Sum]]
= [[fdbk(Add ; UnitDelay ; Split)]]
= Feedback([[Add]] ◦ [[UnitDelay]] ◦ [[Split]])
For Add, UnitDelay and Split, all inputs are legal,
After simplifications, we get:
[[Add]]
=
[[UnitDelay]] =
[[Split]]
=
The semantics of Sum is given by
so illegalC = ⊥ for all C ∈ {Add, UnitDelay, Split}.
[σu , σx ; σu + σx ]
[σx ; 0 · σx ]
[σx ; σx , σx ]
[[Sum]]
=
Feedback([[Add]] ◦ [[UnitDelay]] ◦ [[Split]]
=
=
=
We obtain:
Feedback([σu , σx ; σu + σx ] ◦ [σx ; 0 · σx ] ◦ [σx ; σx , σx ])
Feedback([σu , σx ; 0 · σx + 0 · σu , 0 · σx + 0 · σu ])
{Using (1)}
[σx ; σy | ∀i : σy (i) = Σj<i σx (j)]
[[Sum]] = [σx ; σy | ∀i : σy (i) = Σj<i σx (j)].
Next, let us assume that the system has been characterized already as an atomic component:
SumAtomic = sts(x, y, s, s = 0, y = s ∧ s0 = s + x).
The semantics of SumAtomic is given by
[[SumAtomic]] = {¬illegalSumAtomic } ◦ [σx ; σy | ∃σs : σs (0) = 0 ∧ runSumAtomic (σs , σx , σy )]
where illegalSumAtomic = ⊥ because there are no restrictions on the inputs of SumAtomic, and
runSumAtomic (σs , σx , σy ) = ∀i : σy (i) = σs (i) ∧ σs (i + 1) = σs (i) + σx (i)
We have
[[SumAtomic]] = [σx ; σy | ∃σs : σs (0) = 0 ∧ ∀i : σy (i) = σs (i) ∧ σs (i + 1) = σs (i) + σx (i) ]
which is equivalent to (11).
4.3.2
Characterization of Legal Input Traces
The following lemma characterizes legal input traces for various types of MPTs:
Lemma 8. The set of legal input traces of an RPT {p} ◦ [r] is p:
legal({p} ◦ [r]) = p.
23
(11)
The set of legal input traces of a GPT {r] is in(r):
legal({r]) = in(r).
The set of legal input traces of an STS component C = sts(x, y, s, init, r) is equal to ¬illegalC :
legal([[C]]) = ¬illegalC .
The set of legal input traces of a QLTL component C = qltl(x, y, ϕ) is:
legal([[C]]) = {σx | σx |= ∃y : ϕ}.
4.3.3
Semantic Equivalence and Refinement for Components
Definition 20. Two components C and C 0 are (semantically) equivalent, denoted C ≡ C 0 , if [[C]] = [[C 0 ]].
Component C is refined by component C 0 , denoted C C 0 , if [[C]] v [[C 0 ]].
The relation ≡ is an equivalence relation, and is a preorder relation (i.e., reflexive and transitive). We
also have
(C C 0 ∧ C 0 C) ⇐⇒ (C ≡ C 0 )
4.3.4
Compositionality Properties
Several desirable compositionality properties follow from our semantics:
Theorem 5. Let C1 , C2 , C3 , and C4 be four (possibly composite) components. Then:
1. (Serial composition is associative:) (C1 ; C2 ) ; C3 ≡ C1 ; (C2 ; C3 ).
2. (Parallel composition is associative:) (C1 k C2 ) k C3 ≡ C1 k (C2 k C3 ).
3. (Parallel composition distributes over serial composition:) If [[C1 ]] and [[C2 ]] are GPTs and [[C3 ]] and
[[C4 ]] are RPTs, then (C1 k C2 ) ; (C3 k C4 ) ≡ (C1 ; C3 ) k (C2 ; C4 ).
4. (Refinement is preserved by composition:) If C1 C2 and C3 C4 , then:
(a) C1 ; C3 C2 ; C4
(b) C1 k C3 C2 k C4
(c) fdbk(C1 ) fdbk(C2 )
In addition to the above, requirements that a component satisfies are preserved by refinement. Informally,
if C satisfies some requirement ϕ and C C 0 then C 0 also satisfies ϕ. Although we have not formally defined
what requirements are and what it means for a component to satisfy a requirement, these concepts are
naturally captured in the RCRS framework via the semantics of components as MPTs. In particular, since
our components are generally open systems (i.e., they have inputs), we can express requirements using Hoare
triples of the form p{C}q, where C is a component, p is an input property, and q is an output property. Then,
p{C}q holds iff the outputs of C are guaranteed to satisfy q provided the inputs of C satisfy p. Formally:
p{C}q ⇐⇒ p ⊆ [[C]](q).
Theorem 6. C C 0 iff ∀p, q : p{C}q ⇒ p{C 0 }q.
Theorem 6 shows that refinement is equivalent to substitutability. Substitutability states that a component
C 0 can replace another component C in any context, i.e., ∀p, q : p{C}q ⇒ p{C 0 }q.
24
5
Symbolic Reasoning
So far we have defined the syntax and semantics of RCRS. These already allow us to specify and reason about
systems in a compositional manner. However, such reasoning is difficult to do “by hand”. For example, if
we want to check whether a component C is refined by another component C 0 , we must resort to proving
the refinement relation [[C]] v [[C 0 ]] of their corresponding MPTs, [[C]] and [[C 0 ]]. This is not an easy task,
as MPTs are complex mathematical objects. Instead, we would like to have computer-aided, and ideally
fully automatic techniques. In the above example of checking refinement, for instance, we would like ideally
to have an algorithm that takes as input the syntactic descriptions of C and C 0 and replies yes/no based
on whether [[C]] v [[C 0 ]] holds. We say “ideally” because we know that in general such an algorithm cannot
exist. This is because we are not making a-priori any restrictions on the logics used to describe C and C 0 ,
which means that the existence of an algorithm will depend on decidability of these logics. In this section,
we describe how reasoning in RCRS can be done symbolically, by automatically manipulating the formulas
used to specify the components involved. As we shall show, most problems can be reduced to checking
satisfiability of first-order formulas formed by combinations of the formulas of the original components. This
means that the problems are decidable whenever the corresponding first-order logics are decidable.
5.1
Symbolic Transformation of STS Components to QLTL Components
Our framework allows the specification of several types of atomic components, some of which are special
cases of others, as summarized in Fig. 5. In §3, we have already shown how the different types of STS
components are related, from the most specialized deterministic stateless STS components, to the general
STS components. By definition, the semantics of the special types of STS components is defined via the
semantics of general STS components (see §4). In this subsection, we show that STS components can be
viewed as special cases of QLTL components.
qltl
sts
stateless
det
stateless det
Figure 5: Lattice of atomic components: lower types are special cases of higher types.
Specifically, we show how an STS component can be transformed into a semantically equivalent QLTL
component. This transformation also shows that STS property transformers are a special case of QLTL
property transformers, as already claimed in Fig. 4. Note that this containment is not obvious simply by
looking at the definitions of these MPT subclasses (c.f. §4.3), as QLTL property transformers are defined as
GPTs (equation 2), whereas STS property transformers are defined as RPTs (equation 6). Although RPTs
are generally a superclass, not a subclass, of GPTs, the transformation proposed below shows that the RPTs
obtained from STS components can indeed be captured as GPTs. The transformation of STS into QLTL
components also enables us to apply several algorithms which are available for QLTL components to STS
components as well.
Consider an STS component C = sts(x, y, s, init, trs). We can transform C into a QLTL component
using the mapping sts2qltl:
sts2qltl(C) = qltl(x, y, (∀s, y : init ⇒ (ϕ L ϕ0 )) ∧ (∃s : init ∧ G ϕ))
where ϕ = trs[s0 := # s] and ϕ0 = (∃s0 , y : trs).
The theorem that follows demonstrates the correctness of the above transformation:
25
(12)
Theorem 7. For any STS component C = sts(x, y, s, init, trs) s.t. init is satisfiable, C ≡ sts2qltl(C).
Example. It is instructive to see how the above transformation specializes to some special types of STS
components. In particular, we will show how it specializes to stateless STS components.
Let C = stateless(x, y, trs) and let C 0 = stateless2sts(C) = sts(x, y, (), true, trs). Applying the
sts2qltl transformation to C 0 , for which s = () and init = true, we obtain:
sts2qltl(C 0 ) = qltl x, y, (∀y : (ϕ L ϕ0 )) ∧ G ϕ)
where ϕ = trs[() := # ()] = trs and ϕ0 = (∃y : trs). Using the properties of Lemma 1, and the fact that
semantically equivalent LTL formulas result in semantically equivalent QLTL components, we can simplify
sts2qltl(C 0 ) further:
sts2qltl(C 0 )
= {Definition of sts2qltl}
qltl x, y, (∀y : (trs L (∃y : trs))) ∧ G trs
≡
{Lemma 1, trs does not contain temporal operators, and y is not free in (∃y : trs)}
qltl x, y, ((∃y : trs) L (∃y : trs)) ∧ G trs
≡
{Lemma 1}
qltl x, y, G (∃y : trs) ∧ G trs
≡
{Lemma 1 and trs does not contain temporal operators}
qltl x, y, (∃y : G trs) ∧ G trs
≡
{Lemma 6}
qltl x, y, G trs
Note that in the above derivation we use the equivalence symbol ≡, in addition to the equality symbol =.
Recall that ≡ stands for semantical equivalence of two components (c.f. §4.3.3). On the other hand, = for
components means syntactic equality. By definition of the semantics, if two QLTL formulas are equivalent,
then the corresponding QLTL components are equivalent.
Based on the above, we define the transformation stateless2qltl of a stateless component C =
stateless(x, y, trs), into a QLTL component as follows:
stateless2qltl(C) = qltl x, y, G trs
5.2
Symbolic Transformations of Special Atomic Components to More General
Atomic Components
Based on the lattice in Fig. 5, we define all remaining mappings from more special atomic components to more
general atomic components, by composing the previously defined mappings sts2qltl, stateless2qltl,
stateless2sts, det2sts, stateless det2stateless and stateless det2det, as appropriately.
For mapping stateless deterministic STS components to QLTL components, we have two possibilities:
stateless det → det → sts → qltl and stateless det → stateless → qltl. We choose the transformation stateless det → stateless → qltl because it results in a simpler formula:
stateless det2qltl(C) = stateless2qltl(stateless det2stateless(C)).
Examples.
Consider the following STS components:
C1
C2
UnitDelay
= stateless x, y, y > x
= stateless x, (), x > 0
= det x, s, 0, true, x, s
26
Then:
stateless2qltl(C1 )
= qltl x, y, G y > x
stateless2qltl(C2 )
= qltl x, (), G x > 0
det2qltl(UnitDelay)
= qltl x, y, (∀s, y : s = 0 ⇒
(true ∧ # s = x ∧ y = s) L (∃s0 , y : true ∧ s0 = x ∧ y = s)) ∧
(∃s : s = 0 ∧ G (true ∧ # s = x ∧ y = s))
Because (∃s0 , y : true ∧ s0 = x ∧ y = s)) ⇐⇒ true, using Lemma 1 and logical simplifications we obtain:
det2qltl(UnitDelay) ≡ qltl x, y, (∃s : s = 0 ∧ G (# s = x ∧ y = s))
≡ qltl x, y, (∃s : s = 0 ∧ G (# y = x ∧ y = s))
≡ qltl x, y, (∃s : s = 0 ∧ G (y = s) ∧ G (# y = x))
≡ qltl x, y, (∃s : y = 0 ∧ G (y = s) ∧ G (# y = x))
≡ qltl x, y, (∃s : G (y = s)) ∧ y = 0 ∧ G (# y = x)
≡ qltl x, y, G (∃s : y = s) ∧ y = 0 ∧ G (# y = x)
≡ qltl x, y, y = 0 ∧ G (# y = x)
5.3
Symbolic Computation of Serial Composition
Given a composite component C formed as the serial composition of two components C1 and C2 , i.e.,
C = C1 ; C2 , we would like to compute a new, atomic component Ca , such that Ca is semantically equivalent
to C. Because atomic components are by definition represented syntactically (and symbolically), being able
to reduce composite components into atomic components means that we are able to symbolically compute
composition operators.
We start by defining the symbolic serial composition of components of the same type.
5.3.1
Symbolic Serial Composition of Two QLTL Components
Let C = qltl(x, y, ϕ) and C 0 = qltl(y, z, ϕ0 ) such that C ; C 0 is well formed. Then their symbolic serial
composition, denoted serial(C, C 0 ), is the QLTL component defined by
serial(C, C 0 ) = qltl x, z, (∀y : ϕ ⇒ (∃z : ϕ0 )) ∧ (∃y : ϕ ∧ ϕ0 )
(13)
Note that in the above definition (as well as the ones that follow) we assume that the output variable
of C and the input variable of C 0 have the same name (y) and that the names x, y and z are distinct. In
general, this may not be the case. This is not a problem, as we can always rename variables such that this
condition is met. Note that variable renaming does not change the semantics of components (c.f. §3.2.2).
The intuition behind the formula in (13) is as follows. The second conjunct ∃y : ϕ ∧ ϕ0 ensures that
the both contracts ϕ and ϕ0 of the two components are enforced in the composite contract. The reason
we use ∃y : ϕ ∧ ϕ0 instead of just the conjunction ϕ ∧ ϕ0 is that we want to eliminate (“hide”) internal
variable y. (Alternatively, we could also have chosen to output y as an additional output, but would then
need an additional hiding operator to remove y.) The first conjunct ∀y : ϕ ⇒ (∃z : ϕ0 ) is a formula on
the input variable x of the composite component (since all other variables y and z are quantified). This
formula restricts the legal inputs of C to those inputs for which, no matter which output C produces, this
output is guaranteed to be a legal input for the downstream component C 0 . For an extensive discussion of
the intuition and justification behind this definition, see [79].
27
5.3.2
Symbolic Serial Composition of Two General STS Components
Let C = sts(x, y, s, init, trs) and C 0 = sts(y, z, t, init0 , trs0 ) be two general STS components such that
C ; C 0 is well formed. Then:
serial(C, C 0 ) = sts x, z, (s, t), init∧init0 , (∃s0 , y : trs)∧(∀s0 , y : trs ⇒ (∃t0 , z : trs0 ))∧(∃y : trs∧trs0 ) (14)
5.3.3
Symbolic Serial Composition of Two Stateless STS Components
Let C = stateless(x, y, trs) and C 0 = stateless(y, z, trs0 ) be two stateless STS components such that
C ; C 0 is well formed. Then
serial(C, C 0 ) = stateless x, z, (∀y : trs ⇒ (∃z : trs0 )) ∧ (∃y : trs ∧ trs0 )
(15)
5.3.4
Symbolic Serial Composition of Two Deterministic STS Components
Let C = det(x, s, a, p, next, out) and C 0 = det(y, t, b, p0 , next0 , out0 ) be two deterministic STS components
such that their serial composition C ; C 0 is well formed. Then:
serial(C, C 0 ) = det(x, (s, t), (a, b), p ∧ p0 [y := out], (next, next0 [y := out]), out0 [y := out])
(16)
where e[z := e0 ] denotes the substitution of all free occurences of variable z by expression e0 in expression e.
5.3.5
Symbolic Serial Composition of Two Stateless Deterministic STS Components
Finally, let C = stateless det(x, p, out) and C 0 = stateless det(y, p0 , out0 ) be two stateless deterministic
STS components such that their serial composition C ; C 0 is well formed. Then:
serial(C, C 0 ) = stateless det(x, p ∧ p0 [y := out], out0 [y := out])
5.3.6
(17)
Symbolic Serial Composition of Two Arbitrary Atomic Components
In general, we define the symbolic serial composition of two atomic components C and C 0 by using the
mappings of less general components to more general components (Fig. 5), as appropriate. For example, if C is a deterministic STS component and C 0 is a stateless STS component, then serial(C, C 0 ) =
serial(det2sts(C), stateless2sts(C 0 )). Similarly, if C is a QLTL component and C 0 is a deterministic STS component, then serial(C, C 0 ) = serial(C, det2qltl(C 0 )). Formally, assume that atm, atm 0 ∈
{qltl, sts, stateless, det, stateless det} are the types of the components C and C 0 , and common =
atm ∨ atm 0 is the least general component type that is more general than atm and atm 0 as defined in Fig. 5.
Then
serial(C, C 0 ) = serial(atm2common(C), atm 0 2common(C 0 )).
5.3.7
Correctness of Symbolic Serial Composition
The following theorem demonstrates that our symbolic computations of serial composition are correct, i.e.,
that the resulting atomic component serial(C, C 0 ) is semantically equivalent to the original composite component C ; C 0 :
Theorem 8. If C and C 0 are two atomic components, then
C ; C 0 ≡ serial(C, C 0 ).
28
Examples.
Consider the following STS components:
C1
= stateless(u, (x, y), true)
C2
x
= stateless det((x, y), y 6= 0, )
y
Then:
serial(C1 , C2 )
= serial(C1 , stateless det2stateless(C2 ))
= serial(stateless(u, (x, y), true), stateless((x, y), z, y 6= 0 ∧ z =
= stateless(u, z, (∀x, y : true ⇒ (∃z : y 6= 0 ∧ z =
∧ (∃x, y : true ∧ y 6= 0 ∧ z =
x
))
y
x
))
y
x
)))
y
≡ stateless(u, z, false)
As we can see, the composition results in a stateless STS component with input-output formula false. The
semantics of such a component is Fail, indicating that C1 and C2 are incompatible. Indeed, in the case of
C1 ; C2 , the issue is that C2 requires its second input, y, to be non-zero, but C1 cannot guarantee that. The
reason is that the input-output formula of C1 is true, meaning that, no matter what its input u is, C1 may
output any value for x and y, non-deterministically. This violates the input requirements of C2 , causing
an incompatibility. We will return to this point in §5.8. We also note that this type of incompatibility is
impossible to prevent, by controlling the input u. In the example that follows, we see a case where the two
components are not incompatible, because the input requirements of the downstream component can be met
by strengthening the input assumptions of the upstream component:
Consider the following QLTL components:
C3
= qltl(x, y, G (x ⇒ F y))
C4
= qltl(y, (), G F y)
Then:
serial(C3 , C4 )
= serial(qltl(x, y, G (x ⇒ F y)), qltl(y, (), G F y))
= qltl(x, (), (∀y : (G (x ⇒ F y)) ⇒ G F y) ∧ (∃y : G (x ⇒ F y) ∧ G F y))
≡ qltl(x, (), G F x)
In this example, the downstream component C4 requires its input y to be infinitely often true (G F y).
This can be achieved only if the input x of the upstream component is infinitely often true, which is the
condition derived by the serial composition of C3 and C4 (G F x). Notice that C3 does not impose any apriori requirements on its input. However, its input-output relation is the so-called request-response property
which can be expressed as: whenever the input x is true, the output y will eventually become true afterwards
(G (x ⇒ F y)). This request-response property implies that in order for y to be infinitely-often true, x must
be infinitely-often true. Moreover, this is the weakest possible condition that can be enforced on x in order
to guarantee that the condition on y holds.
5.4
Symbolic Computation of Parallel Composition
Given a composite component formed as the parallel composition of two components, C = C1 k C2 , we
would like to compute an atomic component Ca , such that Ca is semantically equivalent to C. We show
how this can be done in this subsection. The symbolic computation of parallel composition follows the same
pattern as the one for serial composition (§5.3).
29
5.4.1
Symbolic Parallel Composition of Two QLTL Components
Let C = qltl(x, y, ϕ) and C 0 = qltl(u, v, ϕ0 ).
parallel(C, C 0 ), is the QLTL component defined by
Then their symbolic parallel composition, denoted
parallel(C, C 0 ) = qltl (x, u), (y, v), ϕ ∧ ϕ0
(18)
As in the case of symbolic serial composition, we assume that variable names x, y, u, v are all distinct. If
this is not the case, then we rename variables as appropriately.
5.4.2
Symbolic Parallel Composition of Two General STS Components
Let C = sts(x, y, s, init, trs) and C 0 = sts(u, v, t, init0 , trs0 ). Then:
parallel(C, C 0 ) = sts (x, u), (y, v), (s, t), init ∧ init0 , trs ∧ trs0
5.4.3
(19)
Symbolic Parallel Composition of Two Stateless STS Components
Let C = stateless(x, y, trs) and C 0 = stateless(u, v, trs0 ). Then
parallel(C, C 0 ) = stateless (x, u), (y, v), trs ∧ trs0
5.4.4
(20)
Symbolic Parallel Composition of Two Deterministic STS Components
Let C = det(x, s, a, p, next, out) and C 0 = det(u, t, b, p0 , next0 , out0 ). Then:
parallel(C, C 0 ) = det (x, u), (s, t), (a, b), p ∧ p0 , (next, next0 ), (out, out0 )
5.4.5
(21)
Symbolic Parallel Composition of Two Stateless Deterministic STS Components
Let C = stateless det(x, p, out) and C 0 = stateless det(u, p0 , out0 ). Then:
parallel(C, C 0 ) = stateless det (x, u), p ∧ p0 , (out, out0 )
5.4.6
(22)
Symbolic Parallel Composition of Two Arbitrary Atomic Components
Similar to the symbolic serial composition, we define the symbolic parallel composition of two atomic components C and C 0 by using the mappings of less general components to more general components (Fig. 5), as
appropriate. Formally, assume that atm, atm 0 ∈ {qltl, sts, stateless, det, stateless det} are the types
of the components C and C 0 , and common = atm ∨ atm 0 is the least general component type that is more
general than atm and atm 0 as defined in Fig. 5. Then
parallel(C, C 0 ) = parallel(atm2common(C), atm 0 2common(C 0 )).
5.4.7
Correctness of Symbolic Parallel Composition
The following theorem demonstrates that our symbolic computations of parallel composition are also correct,
i.e., that the resulting atomic component parallel(C, C 0 ) is semantically equivalent to the original composite
component C k C 0 :
Theorem 9. If C and C 0 are two atomic components, then
C k C 0 ≡ parallel(C, C 0 ).
30
5.5
5.5.1
Symbolic Computation of Feedback Composition for Decomposable Deterministic STS Components
Decomposable Components
We provide a symbolic closed-form expression for the feedback composition of a deterministic STS component, provided such a component is decomposable. Intuitively, decomposability captures the fact that the
first output of the component, y1 , does not depend on its first input, x1 . This ensures that the feedback
composition (which connects y1 to x1 ) does not introduce any circular dependencies.
Definition
21
(Decomposability).
Let
C
be
a
deterministic
STS
det (x1 , . . . , xn ), s, a, p, next, (e1 , . . . , em ) or
a
stateless
deterministic
STS
stateless det (x1 , . . . , xn ), p, (e1 , . . . , em ) . C is called decomposable if x1 is not free in e1 .
component
component
e1
x1
x2 , . . . , x n
•
e2
...
em
e1
x2 , . . . , x n
(a)
•
x1
e2
...
em
(b)
Figure 6: (a) Decomposable deterministic component; (b) the same component after applying feedback,
connecting its first output to x1 .
Decomposability is illustrated in Fig. 6a. The figure shows that expression e1 depends only on inputs
x2 , . . . , x n .
5.5.2
Symbolic Feedback of a Decomposable Deterministic STS Component
For a decomposable deterministic STS component C = det((x1 , . . . , xn ), s, a, p, next, (e1 , . . . , em )), its symbolic feedback composition, denoted feedback(C), is the deterministic STS component defined by
feedback(C) = det (x2 , . . . , xn ), s, a, p[x1 := e1 ], next[x1 := e1 ], (e2 [x1 := e1 ], . . . , em [x1 := e1 ])
(23)
Thus, computing feedback symbolically consists in removing the first input of the component and replacing
the corresponding variable x1 by the expression of the first output, e1 , everywhere where x1 appears. The
feedback operator is illustrated in Fig. 6b.
5.5.3
Symbolic Feedback of a Decomposable Stateless Deterministic STS Component
For a decomposable stateless deterministic STS component C = stateless det((x1 , . . . , xn ), p, (e1 , . . . , em )),
feedback(C) is the stateless deterministic STS component defined by
feedback(C) = stateless det (x2 , . . . , xn ), p[x1 := e1 ], (e2 [x1 := e1 ], . . . , em [x1 := e1 ])
(24)
5.5.4
Correctness of Symbolic Feedback Composition
Theorem 10. If C is a decomposable deterministic STS component, then
fdbk(C) ≡ feedback(C)
31
Providing closed-form symbolic computations of feedback composition for general components, including
possibly non-deterministic STS and QLTL components, is an open problem, beyond the scope of the current
paper. We remark that the straightforward idea of adding to the contract the equality constraint x = y
where y is the output connected in feedback to input x, does not work.3
In fact, even obtaining a semantically consistent compositional definition of feedback for non-deterministic
and non-input-receptive systems is a challenging problem [67]. Nevertheless, the results that we provide here
are sufficient to cover the majority of cases in practice. In particular, the symbolic operator feedback can
be used to handle Simulink diagrams, provided these diagrams do not contain algebraic loops, i.e., circular
dependencies (see §5.7).
5.6
Closure Properties of MPT Subclasses w.r.t. Composition Operators
In addition to providing symbolic computation procedures, the results of the above subsections also prove
closure properties of the various MPT subclasses of RCRS with respect to the three composition operators.
These closure properties are summarized in Tables 1 and 2.
In a nutshell, both serial and parallel composition preserve the most general type of the composed
components, according to the lattice in Fig. 5. For instance, the serial (or parallel) composition of two
stateless STS components is a stateless STS component; the serial (or parallel) composition of a stateless
STS component and a general STS component is a general STS component; and so on. Feedback preserves
the type of its component (deterministic or stateless deterministic).
; and k
qltl
sts
stateless
det
stateless det
qltl
qltl
qltl
qltl
qltl
qltl
sts
qltl
sts
sts
sts
sts
stateless
qltl
sts
stateless
sts
stateless
det
qltl
sts
sts
det
det
stateless det
qltl
sts
stateless
det
stateless det
Table 1: Closure properties of serial and parallel compositions. The table is to be read as follows: given
atomic components C1 , C2 of types as specified in a row/column pair, the serial or parallel composition of
C1 and C2 is an atomic component of type as specified in the corresponding table entry.
fdbk
det and decomposable
det
stateless det and decomposable
stateless det
Table 2: Closure properties of feedback composition.
5.7
Symbolic Simplification of Arbitrary Composite Components
The results of the previous subsections show how to simplify into an atomic component the serial or parallel
composition of two atomic components, or the feedback composition of an atomic decomposable component. We can combine these techniques in order to provide a general symbolic simplification algorithm: the
algorithm takes as input an arbitrarily complex composite component, and returns an equivalent atomic
component. The algorithm is shown in Fig. 7.
The algorithm fails only in case it encounters the feedback of a non-decomposable component. Recall
that decomposability implies determinism (c.f. §5.5.1), which means that the test C 00 is decomposable means
3 One of several problems of this definition is that it does not preserve refinement. For example, the stateless component
with contract x 6= y refines the stateless component with contract true. Adding the constraint x = y to both contracts yields
the components with contracts x = y and false, respectively, where the latter no longer refines the former. For a more detailed
discussion, see [67].
32
atomic(C) :
if C is atomic then
return C
else if C is C 0 ; C 00 then
return serial(atomic(C 0 ), atomic(C 00 ))
else if C is C 0 k C 00 then
return parallel(atomic(C 0 ), atomic(C 00 ))
else if C is fdbk(C 0 ) then
C 00 := atomic(C 0 )
if C 00 is decomposable then
return feedback(C 00 )
else
fail
else /* impossible by definition of syntax */
fail
Figure 7: Simplification algorithm for arbitrary composite components.
that C 00 is of the form det((x1 , . . .), s, a, p, next, (e1 , . . .)) or stateless det((x1 , . . .), p, (e1 , . . .)) and x1 is
not free in e1 .
We note that in practice, our RCRS implementation on top of Isabelle performs more simplifications in
addition to those performed by the procedure atomic. For instance, our implementation may be able to
simplify a logical formula φ into an equivalent but simpler formula φ0 (e.g., by eliminating quantifiers from
φ), and consequently also simplify a component, say, qltl(x, y, φ) into an equivalent but simpler component
qltl(x, y, φ0 ). These simplifications very much depend on the logic used in the components. Describing the
simplifications that our implementation performs is outside the scope of the current paper, as it belongs in
the realm of computational logic. It suffices to say that our tool is not optimized for this purpose, and could
leverage specialized tools and relevant advances in the field of computational logic.
5.7.1
Deterministic and Algebraic Loop Free Composite Components
In order to state and prove correctness of the algorithm, we extend the notion of determinism to a composite
component. We also introduce the notion of algebraic loop free components, which capture systems with no
circular and intantaneous input-output dependencies.
A (possibly composite) component C is said to be deterministic if every atomic component of C is either
a deterministic STS component or a stateless deterministic STS component. Formally, C is deterministic iff
determ(C) is true, where determ(C) is defined inductively on the structure of C:
determ(det(x, s, a, p, n, e))
=
determ(stateless det(x, p, e)) =
determ(sts(x, y, s, init, trs))
=
determ(stateless(x, y, trs))
=
determ(C ; C 0 )
=
determ(C k C 0 )
=
determ(fdbk(C))
=
true
true
false
false
determ(C) ∧ determ(C 0 )
determ(C) ∧ determ(C 0 )
determ(C)
Notice that this notion of determinism applies to a generally composite component C, i.e., a syntactic term
in our algebra of components, involving atomic components possibly composed via the three composition operators. This notion of determinism is the generalization of the syntactically deterministic STS components,
33
which are atomic. This notion of determinism is also distinct from any semantic notion of determinism
(which we have not introduced at all in this paper, as it is not needed).
For a deterministic component we define its output input dependency relation. Let C be deterministic,
and let Σin (C) = X1 × . . . × Xn and Σout (C) = Y1 × . . . × Ym . The relation OI(C) ⊆ {1, . . . , m} × {1, . . . , n}
is defined inductively on the structure of C:
OI(det((x1 , . . . , xn ), s, a, p, next, (e1 , . . . , em )))
=
OI(stateless det((x1 , . . . , xn ), p, (e1 , . . . , em ))) =
OI(C ; C 0 )
=
OI(C k C 0 )
=
OI(fdbk(C))
=
{(i, j) | xj is free in ei }
{(i, j) | xj is free in ei }
OI(C 0 ) ◦ OI(C)
OI(C) ∪ {(i + m, j + n) | (i, j) ∈ OI(C 0 )}
where Σin (C) = X1 × . . . × Xn
and Σout (C) = Y1 × . . . × Ym
{(i, j) | i > 0 ∧ j > 0 ∧ ((i + 1, j + 1) ∈ OI(C)
∨ ((i + 1, 1) ∈ OI(C) ∧ (1, j + 1) ∈ OI(C)))}
The intuition is that (i, j) ∈ OI(C) iff the i-th output of C depends on its j-th input.
The OI relation is preserved by the symbolic operations, as shown by the following lemma:
Lemma 9. If C and C 0 are deterministic STS components, then
OI(C ; C 0 )
OI(C k C 0 )
= OI(serial(C, C 0 ))
= OI(parallel(C, C 0 )).
If C is also decomposable, then
OI(fdbk(C))
= OI(feedback(C)).
We introduce now the the notion of algebraic loop free component. Intuitively, a (possibly composite)
deterministic component C is algebraic loop free if, whenever C contains a subterm of the form fdbk(C 0 ),
the first output of C 0 does not depend on its first input. This implies that whenever a feedback connection
is formed, no circular dependency is introduced. It also ensures that the simplification algorithm will never
fail. Formally, for a component C such that determ(C) is true, loop-free(C) is defined inductively on the
structure of C:
loop-free(C)
=
loop-free(C)
=
loop-free(C ; C 0 )
=
loop-free(C k C 0 )
=
loop-free(fdbk(C)) =
5.7.2
true if C = det(x, s, a, p, next, out)
true if C = stateless det(x, p, out)
loop-free(C) ∧ loop-free(C 0 )
loop-free(C) ∧ loop-free(C 0 )
loop-free(C) ∧ (1, 1) 6∈ OI(C)
Correctness of the Simplification Algorithm
Theorem 11. Let C be a (possibly composite) component.
1. If C does not contain any fdbk operators then atomic(C) does not fail and returns an atomic component such that atomic(C) ≡ C.
2. If determ(C) ∧ loop-free(C) is true then atomic(C) does not fail and returns an atomic component
such that atomic(C) ≡ C. Moreover, atomic(C) is a deterministic STS component and
OI(C) = OI(atomic(C)).
Proof. The first part of the theorem is a consequence of the fact that the symbolic serial and parallel
compositions are defined for all atomic components and return equivalent atomic components, by Theorems 8
and 9.
34
For the second part, since we have a recursive procedure, we prove its correctness by asumming the
correctness of the recursive calls. Additionally, the termination of this procedure is ensured by the fact
that all recursive calls are made on “smaller” components. Specifically: we assume that both determ(C)
and loop-free(C) hold; and we prove that atomic(C) does not fail, atomic(C) is a deterministic STS
component, atomic(C) ≡ C, and OI(C) = OI(atomic(C)).
We only consider the case C = fdbk(C 0 ). All other cases are similar, but simpler. Because determ(C)
and loop-free(C) hold, we have that determ(C 0 ) and loop-free(C 0 ) also hold, and in addition (1, 1) 6∈
OI(C 0 ). Using the correctness assumption for the recursive call we have that atomic(C 0 ) does not fail,
C 00 = atomic(C 0 ) is a deterministic STS component, C 00 ≡ C 0 , and OI(C 00 ) = OI(C 0 ).
Because C 00 is a deterministic STS component and (1, 1) 6∈ OI(C 0 ) = OI(C 00 ), C 00 is decomposable. From
this we have that D := feedback(C 00 ) is defined. Therefore, atomic(C) returns D and does not fail. It
remains to show that D has the desired properties. By the definition of feedback(C 00 ) and the fact that C 00
is a decomposable deterministic STS component, D is also a deterministic STS component. We also have:
D = feedback(C 00 ) ≡ fdbk(C 00 ) ≡ fdbk(C 0 ) = C
where feedback(C 00 ) ≡ fdbk(C 00 ) follows from Theorem 10 and fdbk(C 00 ) ≡ fdbk(C 0 ) follows from C 00 ≡ C 0
and the semantics of fdbk.
Finally, using Lemma 9 and OI(C 00 ) = OI(C 0 ), we have
OI(D) = OI(feedback(C 00 )) = OI(fdbk(C 00 )) = OI(fdbk(C 0 )) = OI(C)
Corollary 2. If a component C does not contain any fdbk operators or if determ(C) ∧ loop-free(C) is
true, then [[C]] is a GPT.
Note that condition determ(C) ∧ loop-free(C) is sufficient, but not necessary, for [[C]] to be a GPT.
For example, consider the following components:
Constfalse = stateless det x, true, false
And
= stateless det (x, y), true, (x ∧ y, x ∧ y)
C
= Constfalse ; fdbk(And)
Constfalse outputs the constant false. And is a version of logical and with two identical outputs (we need two
copies of the output, because one will be eliminated once we apply feedback). C is a composite component,
formed by first connecting the first output of And in feedback to its first input, and then connecting the
output of Constfalse to the second input of And (in reality, to the only remaining input of fdbk(And)).
Observe that C has algebraic loops, that is, loop-free(C) does not hold. Yet it can be shown that [[C]] is a
GPT (in particular, we can show that C ≡ Constfalse ).
Example.
The simplification algorithm applied to the component from Fig. 2 results in
atomic(Sum) = atomic(fdbk(Add ; UnitDelay ; Split)) = det(y, s, 0, s + y, s).
35
To see how the above is derived, let us first calculate atomic(Add ; UnitDelay ; Split):
atomic(Add ; UnitDelay ; Split)
=
serial(atomic(Add ; UnitDelay), Split)
=
serial(serial(Add, UnitDelay), Split)
= {Expanding Add and UnitDelay and choosing suitable variable names}
serial(serial(stateless det((x, y), true, x + y), det(z, s, 0, true, z, s)), Split)
=
{Computing inner serial}
serial(det((x, y), s, 0, true, x + y, s), Split)
=
{Expanding Split and choosing suitable variable names}
serial(det((x, y), s, 0, true, x + y, s), stateless det(z, true, (z, z)))
=
{Computing remaining serial}
det((x, y), s, 0, true, x + y, (s, s))
We calculate now atomic(fdbk(det((x, y), s, 0, true, x + y, (s, s)))):
atomic(fdbk(det((x, y), s, 0, true, x + y, (s, s))))
= {x is not free in s}
feedback(det((x, y), s, 0, true, x + y, (s, s)))
= {Computing feedback}
det(y, s, 0, true, s + y, s)
5.8
Checking Validity and Compatibility
Recall the example given in §5.3.7, of the serial composition of components C1 and C2 , resulting in a
component with input-output relation false, implying that [[C1 ; C2 ]] = Fail. When this occurs, we say that
C1 and C2 are incompatible. We would like to catch such incompatibilities. This amounts to first simplifying
the serial composition C1 ; C2 into an atomic component C, and then checking whether [[C]] = Fail.
In general, we say that a component C is valid if [[C]] 6= Fail. Given a component C, we can check
whether it is valid, as follows. First, we simplify C to obtain an atomic component C 0 = atomic(C). If C 0
is a QLTL component of the form qltl(x, y, ϕ) then C 0 is valid iff ϕ is satisfiable. The same is true if C 0 is
a stateless STS component of the form stateless(x, y, ϕ). If C 0 is a general STS component then we can
first transform it into a QLTL component and check satisfiability of the resulting QLTL formula.
Theorem 12. If C is an atomic component of the form qltl(x, y, ϕ) or stateless(x, y, ϕ), then [[C]] 6= Fail
iff ϕ is satisfiable.
As mentioned in the introduction, one of the goals of RCRS is to function as a behavioral type system
for reactive system modeling frameworks such as Simulink. In the RCRS setting, type checking consists
in checking properties such as compatibility of components, as in the example C1 ; C2 of §5.3.7. When
components are compatible, by computing new (stronger) input preconditions like in the example C3 ; C4
of §5.3.7, RCRS can be seen as enabling to perform behavioral type inference. Indeed, the new derived
condition G F x in the above example can be seen as an inferred type of the composite component C3 ; C4 .
We note that the decidability of the satisfiability question for ϕ depends on the logic used and the domains
of the variables in the formula. For instance, although ϕ can be a QLTL formula, if it restricts the set of
36
constants to true, has no functional symbols, and only equality as predicate symbol, then it is equivalent to
a QPTL formula,4 for which we can use available techniques [76].
5.9
Checking Input-Receptiveness and Computing Legal Inputs Symbolically
Given a component C, we often want to check whether it is input-receptive, i.e., whether legal([[C]]) = >,
or equivalently, [[C]](>) = >. More generally, we may want to compute the legal input values for C, which
is akin to type inference as discussed above. To do this, we will provide a symbolic method to compute
legal([[C]]) as a formula legal(C). Then, checking that C is input-receptive amounts to checking that the
formula legal(C) is valid, or equivalently, checking that ¬legal(C) is unsatisfiable. We assume that C is
atomic (otherwise, we first simplify C using the algorithm of §5.7).
Definition 22. Given an atomic component C, we define legal(C), a formula characterizing the legal inputs
of C. legal(C) is defined based on the type of C:
legal qltl(x, y, ϕ) = (∃y : ϕ)
(25)
0
legal sts(x, y, s, init, trs) = (∀s, y : init ⇒ (ϕ L ϕ ))
(26)
legal stateless(x, y, trs) = G (∃y : trs)
(27)
legal det(x, s, a, p, next, out) = (∀s, y : s = a ⇒ (# s = next ∧ y = out) L p)
(28)
(29)
legal stateless det(x, p, out) = G p
where ϕ = trs[s0 := # s] and ϕ0 = (∃s0 , y : trs).
The next theorem shows that legal correctly characterizes the semantic predicate legal:
Theorem 13. If C is an atomic component , then
legal([[C]]) = {σx | σx |= legal(C)}.
It follows from Theorem 13 that a component C is input-receptive iff the formula legal(C) is valid.
5.10
Checking Refinement Symbolically
We end this section by showing how to check whether a component refines another component. Again, we
will assume that the components in question are atomic (if not, they can be simplified using the atomic
procedure).
Theorem 14. Let C1 = sts(x, y, s, init, r1 ), C10 = sts(x, y, s, init0 , r10 ), C2 = stateless(x, y, r2 ), C20 =
stateless(x, y, r20 ), C3 = qltl(x, y, ϕ), and C30 = qltl(x, y, ϕ0 ). Then:
1. C1 is refined by C10 if the formula
(init0 ⇒ init) ∧ ((∃s0 , y : r1 ) ⇒ (∃s0 , y : r10 )) ∧ ((∃s0 , y : r1 ) ∧ r10 ⇒ r)
is valid.
2. C2 is refined by C20 if and only if the formula
(∃y : r2 ) ⇒ (∃y : r20 ) ∧ (∃y : r2 ) ∧ r20 ⇒ r2
is valid.
4
For example, the atomic QLTL formula # # x = # y can be translated into the LTL formula X X x ⇔ X y, and the
formula # # # x = true into X X X x.
37
3. C3 is refined by C30 if and only if the formula
(∃y : ϕ) ⇒ (∃y : ϕ0 ) ∧ ((∃y : ϕ) ∧ ϕ0 ) ⇒ ϕ
is valid.
As the above theorem shows, checking refinement amounts to checking validity (or equivalently, satisfiability of the negation) of first-order formulas formed by the various symbolic expressions in the component
specifications. The exact logic of these formulas depends on the logics used by the components. For example,
if C3 and C30 both use quantifier-free LTL for φ and φ0 , then in order to check refinement we need to check
satisfiability of a first-order QLTL formula.
Examples.
Consider again the QLTL component C = qltl((), t, oven), introduced in §3.3, where
oven =
(t = 20 ∧ ((t < # t ∧ t < 180) U thermostat))
thermostat = G (180 ≤ t ∧ t ≤ 220)
Let us introduce a refined version C 0 of C:
C0
=
sts((), t, (s, sw), init, trs) where
init =
s = 20 ∧ sw = on
trs
(t = s) ∧
=
(if sw = on then s < s0 < s + 5 else (if s > 10 then s − 5 < s0 < s else s0 = s)) ∧
(if sw = on ∧ s > 210 then sw0 = off else
(if sw = off ∧ s < 190 then sw0 = on else sw0 = sw))
C 0 is an STS component with no input variables, output variable t, and state variables s and sw, recording
the current temperature of the oven, and the on/off status of the switch, respectively. When sw is on, the
temperature increases nondeterministically by up to 5 units, otherwise the temperature decreases nondeterministically by up to 5 units. When the temperature exceeds 210, the switch is turned off; when the
temperature is below 190, the switch is turned on; otherwise sw remains unchanged. The output t is always
equal to the current state s. Initially the temperature is 20, and sw is on.
Using Theorem 14, and the properties of sts2qltl we have:
C C0
⇐⇒
C sts2qltl(C 0 )
⇐⇒
qltl((), t, oven) qltl((), t, (∀s, sw, t : init ⇒ (ϕ L ϕ0 )) ∧ (∃s, sw : init ∧ G ϕ))
⇐⇒
where ϕ = trs[s0 , sw0 := # s, # sw] and ϕ0 = (∃s0 , sw0 , t : trs)
{Using Lemma 1, because ϕ0 ⇐⇒ true}
qltl((), t, oven) qltl((), t, (∃s, sw : init ∧ G ϕ))
⇐⇒
{Using Theorem 14}
(∃t : oven) ⇒ (∃t, s, sw : init ∧ G ϕ) ∧ ((∃t : oven) ∧ (∃s, sw : init ∧ G ϕ)) ⇒ oven is valid
⇐⇒
{Because (∃t : oven) ⇐⇒ true and (∃t, s, sw : init ∧ G ϕ) ⇐⇒ true}
(∃s, sw : init ∧ G ϕ) ⇒ oven is valid
Thus, checking whether C 0 refines C amounts to checking whether the QLTL formula (∃s, sw : init∧G ϕ) ⇒
oven is valid. This indeed holds for this example and can be shown using logical reasoning.
38
The above example is relatively simple in the sense that in the end refinement reduces to checking
implication between the corresponding contracts. Indeed, this is always the case for input-receptive systems,
as in the example above. However, refinement is not equivalent to implication in the general case of noninput-receptive systems. For example, consider the components:
C1
= stateless(x, y, x ≥ 0 ∧ y ≥ x)
C2
= stateless(x, y, x ≤ y ≤ x + 10)
Using Theorem 14, we have:
C1 C2
⇐⇒
{Theorem 14}
((∃y : x ≥ 0 ∧ y ≥ x) ⇒ (∃y : x ≤ y ≤ x + 10)) ∧
((∃y : x ≥ 0 ∧ y ≥ x) ∧ x ≤ y ≤ x + 10 ⇒ x ≥ 0 ∧ y ≥ x) is valid
⇐⇒
{Arithmetic and logical reasoning}
true
Note that the second and third parts of Theorem 14 provide necessary and sufficient conditions, while
the first part only provides a sufficient, but generally not necessary condition. Indeed, the condition is
generally not necessary in the case of STS components with state, as state space computation is ignored by
the condition. This can be remedied by transforming STS components into equivalent QLTL components
and then applying the second part of the theorem. An alternative which may be more tractable, particularly
in the case of finite-state systems, is to use techniques akin to strategy synthesis in games, such as those
proposed in [79] for finite-state relational interfaces.
Another limitation of the first part of Theorem 14 is that it requires the two STS components to have
the same state space, i.e., the same state variable s. This restriction can be lifted using the well-known idea
of data refinement [44, 8].
Theorem 15. Let C1 = sts(x, y, s, init, r), C2 = sts(x, y, t, init0 , r0 ) be two STS components, and D a
(data refinement) expression on variables s and t. Let p = (∃s0 , y : r) and p0 = (∃t0 , y : r0 ). If the formulas
(∀t : init0 ⇒ (∃s : D ∧ init))
(∀t, x, s : D ∧ p ⇒ p0 )
(∀t, x, s, t0 , y : D ∧ p ∧ r0 ⇒ (∃s0 : D[t, s := t0 , s0 ] ∧ r))
are valid, then C1 is refined by C2 .
6
Toolset and Case Studies
The RCRS framework comes with a toolset, illustrated in Fig. 8. The toolset is publicly available under
the MIT license and can be downloaded from http://rcrs.cs.aalto.fi. The toolset is described in tool
papers [32, 33] ([32] contains additional material over [33], specifically, a six-page appendix describing a
demo of the RCRS toolset). In summary, the toolset consists of:
• A full implementation of the RCRS theory in Isabelle [61]. The implementation consists of 22 theory
files and a total of 27588 lines of Isabelle code. A detailed description of the implementation can be
found in the file document.pdf available in the public distribution of RCRS.
• A formal Analyzer, which is a set of procedures implemented on top of Isabelle and the functional
programming language SML. The Analyzer performs compatibility checking, automatic contract simplification, and other functions.
39
Options
(translation strategy, etc.)
Fuel Control System Model
RCRS theory and component library
(Isabelle theory files)
This model uses only the ODEs to implement the dynamics.
theta [0 90]
1
throttle input (deg)
[0, 81.2]
1
.1s+1
1
A/F
Throttle del ay1
1
s
f(u)
8.8
ODE1
Base opening angle
~=
1
s
1
s
f(u)
pe
ODE4 Closed
double
2
~=
Fuel Cmd Open Pwr
Analyzer
f(u)
0.0
f(u)
airbyfuel_ref
14.7
lambda
f(u)
ODE3
incompatiblity detection
12.5
p
f(u)
ODE2
ODE4 Open
~=
Fuel Cmd Open
1
s
~=
i
f(u)
f(u)
Fuel Cmd Closed
InputPoly
2
engine speed (rpm)
[900,1100]
pi/30
AND
(rpm) to (rad/s)
internal variable elimination
Power Mode Guard
boolean
Starup Mode
1: Failure
0: Normal
In
Out
Startup Mode Latch
boolean
FaultInjection
In
Out
Sensor Failure Detection Latch
OR
3
controller_mode
NOT
Powertrain Control Benchmark Model
Toyota Technial Center
2014
This is a model of a hybrid automaton with polynomial dynamics, and an implementation of the 3rd model that appears in
"Powertrain Control Verification Benchmark", 2014 Hybrid Systems: Computation and Control,
X. Jin, J. V. Deshmukh, J.Kapinski, K. Ueda, and K. Butts
Simulink
diagram
Translator
RCRS model
of the diagram
(built on top of
Isabelle
theorem prover)
auto generated top-level contract
refinement checking
...
Figure 8: The RCRS toolset.
• A formalization of Simulink characterizing basic Simulink blocks as RCRS components and implementing those as a library of RCRS/Isabelle. At the time of writing this paper, 48 of Simulink’s blocks can
be handled.
• A Translator : a Python program translating Simulink hierarchical block diagrams into RCRS code.
We implemented in Isabelle a shallow embedding [14] of the language introduced in §3. The advantage of
a shallow embedding is that all datatypes of Isabelle are available for specification of components, and we
can use the existing Isabelle mechanism for renaming bound variables in compositions. The disadvantage
of this shallow embedding is that we cannot express Algorithm 7 within Isabelle (hence the “manual” proof
that we provide for Theorem 11). A deep embedding, in which the syntax of components is defined as a
datatype of Isabelle, is possible, and is left as an open future work direction.
We implemented Algorithm 7 in SML, the meta-language of Isabelle. The SML program takes as input
a component C and returns not only a simplified atomic component atomic(C), but also a proved Isabelle
theorem of the fact C ≡ atomic(C). The simplification program, as well as a number of other procedures
to perform compatibility checking, validity checking, etc., form what we call the Analyzer in Fig. 8.
The Translator takes as input a Simulink model and produces an RCRS/Isabelle theory file containing: (1)
the definition of all atomic and composite components representing the Simulink diagram; and (2) embedded
bottom-up simplification procedures and the corresponding correctness theorems. By running this theory
file in Isabelle, we obtain an atomic component corresponding to the top-level Simulink model, equivalent to
the original composite component. As a special case, if the Simulink diagram contains inconsistencies (e.g.,
division by zero), these are detected by obtaining Fail as the top-level atomic component. The error can be
localized by finding earlier points in the Simulink hierarchy (subsystems) which already resulted in Fail.
As mentioned earlier, how to obtain a composite component from a graphical block diagram is an interesting problem. This problem is studied in depth in [31], where several translation strategies are proposed.
These various strategies all yield semantically equivalent components, but with different trade-offs in terms
of size, readability, effectiveness of the simplification procedures, and so on. The Translator implements all
these translation strategies, allowing the user to explore these trade-offs. Further details on the translation
problem are provided in [31, 65]. A proof that the translation strategies yield semantically equivalent components is provided in [64]. This proof, which has been formalized in Isabelle, is a non-trivial result: the
entire formalization of the translation algorithms and the proof is 13579 lines of Isabelle code.
We have used the RCRS toolset on several case studies, including a real-life benchmark provided by
the Toyota motor company. The benchmark involves a Fuel Control System (FCS) described in [45, 46].
FCS aims at controlling the air mass and injected fuel in a car engine such that their ratio is always
optimal. This problem has important implications on lowering pollution and costs by improving the engine
performance. Toyota has made several versions of FCS publicly available as Simulink models at https:
//cps-vo.org/group/ARCH/benchmarks.
We have used the RCRS toolset to process two of the three Simulink models in the FCS benchmark suite
(the third model contains blocks that are currently not implemented in the RCRS component library). A
40
typical model in this set has a 3-layer hierarchy with a total of 104 Simulink block instances (97 basic blocks
and 7 subsystems), and 101 connections out of which 8 are feedbacks. Each basic Simulink block is modeled in
our framework by an atomic STS component (possibly stateless). These atomic STS components are created
once, and form part of the RCRS implementation, which is reused for different Simulink models. The
particular FCS diagram is translated into RCRS using the Translator, and simplified within Isabelle using
our SML simplification procedure. After simplification, we obtain an atomic deterministic STS component
with no inputs, 7 outputs, and 14 state variables. Its contract (which is 8337 characters long) includes a
condition on the state variables, in particular, that a certain state variable must always be non-negative (as
its value is fed into a square-root block). This condition makes it not immediately obvious that the whole
system is valid (i.e., not Fail). However, we can show after applying the transformation sts2qltl that the
resulting formula is satisfiable, which implies that the original model is consistent (i.e., no connections result
in incompatibilities, there are no divisions by zero, etc.). This illustrates the use of RCRS as a powerful
static analysis tool. More details on the FCS case study are provided in [31, 65, 32].
7
Related Work
Several formal compositional frameworks exist in the literature. Most closely related to RCRS are the
frameworks of FOCUS [15], input-output automata [54], reactive modules [6], interface automata [22], and
Dill’s trace theory [29]. RCRS shares with these frameworks many key compositionality principles, such as
the notion of refinement. At the same time, RCRS differs and complements these frameworks in important
ways. Specifically, FOCUS, IO-automata, and reactive modules, are limited to input-receptive systems,
while RCRS is explicitly designed to handle non-input-receptive specifications. The benefits of non-inputreceptiveness are discussed extensively in [79] and will not be repeated here. Interface automata are a
low-level formalism whereas RCRS specifications and reasoning are symbolic. For instance, in RCRS one
can naturally express systems with infinite state-spaces, input-spaces, or output-spaces. (Such systems can
even be handled automatically, provided the corresponding logic they are expressed in is decidable.) Both
interface automata and Dill’s trace theory use a single form of asynchronous parallel composition, whereas
RCRS has three primitive composition operators (serial, parallel, feedback) with synchronous semantics.
Our work adapts and extends to the reactive system setting many of the ideas developed previously in a
long line of research on correctness and compositionality for sequential programs. This line of research goes
back to the works of Floyd, Hoare, Dijkstra, and Wirth, on formal program semantics, weakest preconditions,
program development by stepwise refinement, and so on [34, 43, 27, 81]. It also goes back to game-theoretic
semantics of sequential programs as developed in the original refinement calculus [10], as well as to contractbased design [57]. Many of the concepts used in our work are in spirit similar to those used in the above
works. For instance, an input-output formula φ used in an atomic component in our language can be seen
as a contract between the environment of the component and the component itself: the environment must
satisfy the contract by providing to the component legal inputs, and the component must in turn provide
legal outputs (for those inputs). On the other hand, several of the concepts used here come from the world
of reactive systems and as such do not have a direct correspondence in the world of sequential programs.
For instance, this is the case with feedback composition.
RCRS extends refinement calculus from predicate to property transformers. Extensions of refinement
calculus to infinite behaviors have also been proposed in the frameworks of action systems [11], fair action
systems [12], and Event B [3]. These frameworks use predicate (not property) transformers as semantic
foundation; they can handle certain property patterns (e.g., fairness) by providing proof rules for these
properties, but they do not treat livenes and LTL properties in general [12, 82, 42]. The Temporal Logic of
Actions [49] can be used to specify liveness properties, but does not distinguish between inputs and outputs,
and as such cannot express non-input-receptive components.
Our specifications can be seen as “rich”, behavioral types [50, 22]. Indeed, our work is closely related to
programming languages and type theory, specifically, refinement types [35], behavioral types [60, 51, 26], and
liquid types [69].
Behavioral type frameworks have also been proposed in reactive system settings. In the SimCheck
41
framework [70], Simulink blocks are annotated with constraints on input and output variables, much like
stateless components in RCRS. RCRS is more general as it also allows to specify stateful components. RCRS
is also a more complete compositional framework, with composition operators and refinement, which are not
considered in [70]. Other behavioral type theories for reactive systems have been proposed in [23, 18, 30].
Compared to RCRS, these works are less general. In particular, [23, 30] are limited to specifications which
separate the assumptions on the inputs from the guarantees on the outputs, and as such cannot capture
input-output relations. [18] considers a synchronous model which allows to specify legal values of inputs and
outputs at the next step, given the current state. This model does not allow to capture relations between
inputs and outputs within the same step, which RCRS allows.
Our work is related to formal verification frameworks for hybrid systems [5]. Broadly speaking, these
can be classified into frameworks following a model-checking approach, which typically use automata-based
specification languages and state-space exploration techniques, and those following a theorem-proving approach, which typically use logic-based specifications. More closely related to RCRS are the latter, among
which, CircusTime [17], KeYmaera [37], and the PVS-based approach in [2]. CircusTime can handle a larger
class of Simulink diagrams than the current implementation of RCRS. In particular, CircusTime can handle
multi-rate diagrams, where different parts of the model work at different rates (periods). On the other hand,
CircusTime is based on predicate (not property) transformers, and as such cannot handle liveness properties. KeYmaera is a theorem prover based on differential dynamic logic [62], which is itself based on dynamic
logic [39]. The focus of both KeYmaera and the work in [2] is verification, and not compositionality. For
instance, these works do not distinguish between inputs and outputs and do not investigate considerations
such as input-receptiveness. The work of [68] distinguishes inputs and outputs, but provides a system model
where the output relation is separated from the transition relation, and where the output relation is assumed
to be total, meaning that there exists an output for every input and current state combination. This does
not allow to specify non-input-receptive stateless components, such as for example the Div component from
§3.
Our component algebra is similar to the algebra of flownomials [77] and to the relational model for nondeterministic dataflow [41]. In [20], graphs and graph operations which can be viewed as block diagrams
are represented by algebraic expressions and operations, and a complete equational axiomatization of the
equivalence of the graph expressions is given. This is then applied to flow-charts as investigated in [72]. The
translation of block diagrams in general and Simulink in particular has been treated in a large number of
papers, with various goals, including verification and code generation (e.g., see [80, 56, 19, 52, 73, 13, 83, 84,
85, 58]). Although we share several of the ideas of the above works, our main goal here is not to formalize
the language of block diagrams, neither their translation to other formalisms, but to provide a complete
compositional framework for reasoning about reactive systems.
RCRS is naturally related to compositional verification frameworks, such as [38, 1, 55, 74, 40, 24, 25].
In particular, compositional verification frameworks often make use of a refinement relation such as trace
inclusion or simulation [74, 40]. However, the focus of these frameworks is different than that of RCRS.
In compositional verification, the focus is to “break down” a large (and usually computationally expensive)
verification task into smaller (and hopefully easier to calculate) subtasks. For this purpose, compositional
verification frameworks employ several kinds of decomposition rules. An example of such a rule is the socalled precongruence rule (i.e., preservation of refinement by composition): if P1 refines Q1 , and P2 refines
Q2 , then the composition P1 kP2 refines Q1 kQ2 . This, together with preservation of properties by refinement,
allows us to conclude that P1 kP2 satisfies some property φ, provided we can prove that Q1 kQ2 satisfies φ.
The latter might be a simpler verification task, if Q1 and Q2 are smaller than P1 and P2 . The essence of
compositional verification is in finding such abstract versions Q1 and Q2 of the concrete processes in question,
P1 and P2 , and employing decomposition rules like the one above in the hope of making verification simpler.
RCRS can also be used for compositional verification: indeed, RCRS provides both the precongruence rule,
and preservation of properties by refinement. Note that, in traditional settings, precongruence is not always
powerful enough, and for this reason most compositional verification frameworks employ more complex
decomposition rules (e.g., see [59]). In settings which allow non-input-receptive components, such as ours,
there are indications that the precongruence rule is sufficient for compositional verification purposes [75],
42
although more work is required to establish this in the specific context of RCRS. Such work is beyond the
scope of the current paper. We also note that, beyond compositional verification with precongruence, RCRS
provides a behavioral type theory which allows to state system properties such as compatibility, which is
typically not available in compositional verification frameworks.
Refinement can be seen as the inverse of abstraction, and as such our framework is related to general
frameworks such as abstract interpretation [21]. Several abstractions have been proposed in reactive system
settings, including relational abstractions for hybrid systems, which are related to Simulink [71]. The focus
of these works is verification, and abstraction is used as a mechanism to remove details from the model
that make verification harder. In RCRS, the simplification procedure that we employ can be seen as an
abstraction process, as it eliminates internal variable information. However, RCRS simplification is an exact
abstraction, in the sense that it does not lose any information: the final system is equivalent to the original
one, and not an over- or under-approximation, as is usually the case with typical abstractions for verification
purposes.
8
Conclusion
We presented RCRS, a compositional framework for modeling and reasoning about reactive systems. This
paper focuses on the theory and methodology of RCRS, its formal semantics, and techniques for symbolic
and computer-aided reasoning.
RCRS is an ongoing project, and a number of problems remain open. Future work directions include:
• An extension of the framework to systems with algebraic loops, which necessitates handling instantaneous feedback. Here, the preliminary ideas of [67] can be helpful in defining the semantics of instantaneous feedback. However, [67] does not provide solutions on how to obtain symbolic closed-form
expression for the feedback of general components.
• Extension of the results of §5.5 to general components, possibly non-deterministic or non-decomposable.
• An extension of the framework to acausal systems, i.e., systems without a clear distinction of inputs
and outputs [36].
• The development of better symbolic reasoning techniques, such as simplification of logical formulas,
decision procedures, etc.
References
[1] M. Abadi and L. Lamport. Conjoining specifications. ACM Trans. Program. Lang. Syst., 17(3):507–535,
1995.
[2] E. Ábrahám-Mumm, M. Steffen, and U. Hannemann. Verification of hybrid systems: Formalization
and proof rules in PVS. In 7th Intl. Conf. Engineering of Complex Computer Systems (ICECCS 2001),
pages 48–57, 2001.
[3] J.-R. Abrial. Modeling in Event-B: System and Software Engineering. Cambridge University Press, New
York, NY, USA, 1st edition, 2010.
[4] B. Alpern and F. B. Schneider. Defining liveness. Information Processing Letters, 21(4):181 – 185, 1985.
[5] R. Alur, C. Courcoubetis, N. Halbwachs, T. Henzinger, P. Ho, X. Nicollin, A. Olivero, J. Sifakis, and
S. Yovine. The algorithmic analysis of hybrid systems. Theoretical Computer Science, 138:3–34, 1995.
[6] R. Alur and T. Henzinger. Reactive modules. Formal Methods in System Design, 15:7–48, 1999.
43
[7] R. Alur, T. Henzinger, O. Kupferman, and M. Vardi. Alternating refinement relations. In CONCUR’98,
volume 1466 of LNCS. Springer, 1998.
[8] R. J. Back. Correctness preserving program refinements: proof theory and applications, volume 131 of
Mathematical Centre Tracts. Mathematisch Centrum, Amsterdam, 1980.
[9] R.-J. Back and M. Butler. Exploring summation and product operators in the refinement calculus, pages
128–158. Springer Berlin Heidelberg, Berlin, Heidelberg, 1995.
[10] R.-J. Back and J. von Wright. Refinement Calculus: A Systematic Introduction. Springer, 1998.
[11] R.-J. Back and J. Wright. Trace refinement of action systems. In B. Jonsson and J. Parrow, editors,
CONCUR ’94: Concurrency Theory, volume 836 of Lecture Notes in Computer Science, pages 367–384.
Springer Berlin Heidelberg, 1994.
[12] R.-J. Back and Q. Xu. Refinement of fair action systems. Acta Informatica, 35(2):131–165, 1998.
[13] P. Boström. Contract-Based Verification of Simulink Models. In S. Qin and Z. Qiu, editors, Formal
Methods and Software Engineering, volume 6991 of Lecture Notes in Computer Science, pages 291–306.
Springer Berlin Heidelberg, 2011.
[14] R. J. Boulton, A. Gordon, M. J. C. Gordon, J. Harrison, J. Herbert, and J. V. Tassel. Experience with
embedding hardware description languages in HOL. In IFIP TC10/WG 10.2 Intl. Conf. on Theorem
Provers in Circuit Design, pages 129–156. North-Holland Publishing Co., 1992.
[15] M. Broy and K. Stølen. Specification and development of interactive systems: focus on streams, interfaces, and refinement. Springer, 2001.
[16] R. Cavada, A. Cimatti, M. Dorigatti, A. Griggio, A. Mariotti, A. Micheli, S. Mover, M. Roveri, and
S. Tonetta. The nuxmv symbolic model checker. In A. Biere and R. Bloem, editors, Computer Aided
Verification: 26th International Conference, CAV 2014,, pages 334–342, Cham, 2014. Springer.
[17] A. L. C. Cavalcanti, A. Mota, and J. C. P. Woodcock. Simulink timed models for program verification.
In Z. Liu, J. C. P. Woodcock, and H. Zhu, editors, Theories of Programming and Formal Methods Essays Dedicated to Jifeng He on the Occasion of His 70th Birthday, volume 8051 of Lecture Notes in
Computer Science, pages 82–99. Springer, 2013.
[18] A. Chakrabarti, L. de Alfaro, T. Henzinger, and F. Mang. Synchronous and bidirectional component
interfaces. In CAV, LNCS 2404, pages 414–427. Springer, 2002.
[19] C. Chen, J. S. Dong, and J. Sun. A formal framework for modeling and validating Simulink diagrams.
Formal Aspects of Computing, 21(5):451–483, 2009.
[20] B. Courcelle. A representation of graphs by algebraic expressions and its use for graph rewriting
systems. In H. Ehrig, M. Nagl, G. Rozenberg, and A. Rosenfeld, editors, Graph-Grammars and Their
Application to Computer Science, 3rd International Workshop, Warrenton, Virginia, USA, December
2-6, 1986, volume 291 of Lecture Notes in Computer Science, pages 112–132. Springer, 1986.
[21] P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs
by construction or approximation of fixpoints. In 4th ACM Symp. POPL, 1977.
[22] L. de Alfaro and T. Henzinger. Interface automata. In Foundations of Software Engineering (FSE).
ACM Press, 2001.
[23] L. de Alfaro and T. Henzinger. Interface theories for component-based design. In EMSOFT’01. Springer,
LNCS 2211, 2001.
44
[24] W. de Roever, H. Langmaack, and A. E. Pnueli. Compositionality: The Significant Difference. LNCS.
Springer, 1998.
[25] W.-P. de Roever, F. de Boer, U. Hanneman, J. Hooman, Y. Lakhnech, M. Poel, and J. Zwiers. Concurrency Verification: Introduction to Compositional and Non-compositional Methods. Cambridge University Press, 2012.
[26] K. Dhara and G. Leavens. Forcing behavioral subtyping through specification inheritance. In ICSE’96:
18th Intl. Conf. on Software Engineering, pages 258–267. IEEE Computer Society, 1996.
[27] E. Dijkstra. Notes on structured programming. In O. Dahl, E. Dijkstra, and C. Hoare, editors, Structured
programming, pages 1–82. Academic Press, London, UK, 1972.
[28] E. Dijkstra. Guarded commands, nondeterminacy and formal derivation of programs. Comm. ACM,
18(8):453–457, 1975.
[29] D. Dill. Trace theory for automatic hierarchical verification of speed-independent circuits. MIT Press,
Cambridge, MA, USA, 1987.
[30] L. Doyen, T. Henzinger, B. Jobstmann, and T. Petrov. Interface theories with component reuse. In 8th
ACM & IEEE International conference on Embedded software, EMSOFT, pages 79–88, 2008.
[31] I. Dragomir, V. Preoteasa, and S. Tripakis. Compositional Semantics and Analysis of Hierarchical Block
Diagrams. In SPIN, pages 38–56. Springer, 2016.
[32] I. Dragomir, V. Preoteasa, and S. Tripakis. The Refinement Calculus of Reactive Systems Toolset.
CoRR, abs/1710.08195, 2017.
[33] I. Dragomir, V. Preoteasa, and S. Tripakis. The Refinement Calculus of Reactive Systems Toolset. In
TACAS, 2018.
[34] R. Floyd. Assigning meanings to programs. In In. Proc. Symp. on Appl. Math. 19, pages 19–32.
American Mathematical Society, 1967.
[35] T. Freeman and F. Pfenning. Refinement Types for ML. SIGPLAN Not., 26(6):268–277, May 1991.
[36] P. Fritzson. Principles of Object-Oriented Modeling and Simulation with Modelica 3.3: A Cyber-Physical
Approach. Wiley, 2 edition, 2014.
[37] N. Fulton, S. Mitsch, J.-D. Quesel, M. Völp, and A. Platzer. KeYmaera X: An axiomatic tactical
theorem prover for hybrid systems. In A. P. Felty and A. Middeldorp, editors, CADE, volume 9195 of
LNCS, pages 527–538. Springer, 2015.
[38] O. Grumberg and D. Long. Model checking and modular verification. ACM Trans. Program. Lang.
Syst., 16(3):843–871, 1994.
[39] D. Harel, J. Tiuryn, and D. Kozen. Dynamic Logic. MIT Press, 2000.
[40] T. Henzinger, S. Qadeer, and S. Rajamani. You assume, we guarantee: Methodology and case studies.
In CAV’98, volume 1427 of LNCS. Springer-Verlag, 1998.
[41] T. T. Hildebrandt, P. Panangaden, and G. Winskel. A relational model of non-deterministic dataflow.
Mathematical Structures in Computer Science, 14(5):613–649, 10 2004.
[42] T. S. Hoang and J.-R. Abrial. Reasoning About Liveness Properties in Event-B. In Proceedings of the
13th International Conference on Formal Methods and Software Engineering, ICFEM’11, pages 456–471,
Berlin, Heidelberg, 2011. Springer-Verlag.
45
[43] C. A. R. Hoare. An axiomatic basis for computer programming. Comm. ACM, 12(10):576–580, 1969.
[44] C. A. R. Hoare. Proof of correctness of data representations. Acta Informatica, 1(4), Dec. 1972.
[45] X. Jin, J. Deshmukh, J. Kapinski, K. Ueda, and K. Butts. Benchmarks for model transformations
and conformance checking. In 1st Intl. Workshop on Applied Verification for Continuous and Hybrid
Systems (ARCH), 2014.
[46] X. Jin, J. V. Deshmukh, J. Kapinski, K. Ueda, and K. Butts. Powertrain Control Verification Benchmark. In Proceedings of the 17th International Conference on Hybrid Systems: Computation and Control, HSCC’14, pages 253–262. ACM, 2014.
[47] Y. Kesten, Z. Manna, and A. Pnueli. Temporal verification of simulation and refinement. In J. W.
de Bakker, W. P. de Roever, and G. Rozenberg, editors, A Decade of Concurrency Reflections and
Perspectives: REX School/Symposium, pages 273–346. Springer, 1994.
[48] Y. Kesten and A. Pnueli. Complete Proof System for QPTL. Journal of Logic and Computation,
12(5):701, 2002.
[49] L. Lamport. The temporal logic of actions. ACM Trans. Program. Lang. Syst., 16(3):872–923, May
1994.
[50] E. Lee and Y. Xiong. System-level types for component-based design. In EMSOFT’01: 1st Intl.
Workshop on Embedded Software, pages 237–253. Springer, 2001.
[51] B. Liskov and J. Wing. A behavioral notion of subtyping. ACM Trans. Program. Lang. Syst., 16(6):1811–
1841, 1994.
[52] R. Lublinerman, C. Szegedy, and S. Tripakis. Modular code generation from synchronous block diagrams
– modularity vs. code size. In 36th ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages (POPL’09), pages 78–89. ACM, Jan. 2009.
[53] R. Lublinerman and S. Tripakis. Modularity vs. reusability: Code generation from synchronous block
diagrams. In Design, Automation, and Test in Europe (DATE’08), pages 1504–1509. ACM, Mar. 2008.
[54] N. Lynch and M. Tuttle. An introduction to input/output automata. CWI Quarterly, 2:219–246, 1989.
[55] K. McMillan. A compositional rule for hardware design refinement. In Computer Aided Verification
(CAV’97), volume 1254 of LNCS. Springer-Verlag, 1997.
[56] B. Meenakshi, A. Bhatnagar, and S. Roy. Tool for Translating Simulink Models into Input Language of
a Model Checker. In Formal Methods and Software Engineering, volume 4260 of LNCS, pages 606–620.
Springer, 2006.
[57] B. Meyer. Applying “Design by Contract”. Computer, 25(10):40–51, 1992.
[58] S. Minopoli and G. Frehse. SL2SX Translator: From Simulink to SpaceEx Verification Tool. In 19th
ACM International Conference on Hybrid Systems: Computation and Control (HSCC), 2016.
[59] K. S. Namjoshi and R. J. Trefler. On the completeness of compositional reasoning methods. ACM
Trans. Comput. Logic, 11(3), 2010.
[60] O. Nierstrasz. Regular types for active objects. SIGPLAN Not., 28(10):1–15, 1993.
[61] T. Nipkow, L. C. Paulson, and M. Wenzel. Isabelle/HOL — A Proof Assistant for Higher-Order Logic.
LNCS 2283. Springer, 2002.
[62] A. Platzer. Differential dynamic logic for hybrid systems. J. Autom. Reas., 41(2):143–189, 2008.
46
[63] A. Pnueli. The Temporal Logic of Programs. In 18th Annual Symposium on Foundations of Computer
Science, Providence, Rhode Island, USA, 31 October - 1 November 1977, pages 46–57. IEEE Computer
Society, 1977.
[64] V. Preoteasa, I. Dragomir, and S. Tripakis. A nondeterministic and abstract algorithm for translating
hierarchical block diagrams. CoRR, abs/1611.01337, 2016.
[65] V. Preoteasa, I. Dragomir, and S. Tripakis. Type Inference of Simulink Hierarchical Block Diagrams in
Isabelle. In 37th IFIP WG 6.1 International Conference on Formal Techniques for Distributed Objects,
Components, and Systems (FORTE), 2017.
[66] V. Preoteasa and S. Tripakis. Refinement calculus of reactive systems. In Embedded Software (EMSOFT), 2014 International Conference on, pages 1–10, Oct 2014.
[67] V. Preoteasa and S. Tripakis. Towards Compositional Feedback in Non-Deterministic and Non-InputReceptive Systems. In 31st Annual ACM/IEEE Symposium on Logic in Computer Science (LICS),
2016.
[68] G. Reissig, A. Weber, and M. Rungger. Feedback refinement relations for the synthesis of symbolic
controllers. IEEE Trans. Automat. Contr., 62(4):1781–1796, 2017.
[69] P. M. Rondon, M. Kawaguci, and R. Jhala. Liquid types. In 29th ACM SIGPLAN Conference on
Programming Language Design and Implementation, PLDI ’08, pages 159–169, New York, NY, USA,
2008. ACM.
[70] P. Roy and N. Shankar. SimCheck: An expressive type system for Simulink. In C. Muñoz, editor,
2nd NASA Formal Methods Symposium (NFM 2010), NASA/CP-2010-216215, pages 149–160, Langley
Research Center, Hampton VA 23681-2199, USA, Apr. 2010. NASA.
[71] S. Sankaranarayanan and A. Tiwari. Relational abstractions for continuous and hybrid systems. In
Computer Aided Verification: 23rd International Conference, CAV 2011, pages 686–702. Springer, 2011.
[72] H. Schmeck. Algebraic characterization of reducible flowcharts. Journal of Computer and System
Sciences, 27(2):165 – 199, 1983.
[73] V. Sfyrla, G. Tsiligiannis, I. Safaka, M. Bozga, and J. Sifakis. Compositional translation of Simulink
models into synchronous BIP. In Industrial Embedded Systems (SIES), 2010 International Symposium
on, pages 217–220, July 2010.
[74] N. Shankar. Lazy compositional verification. In COMPOS’97: Revised Lectures from the International Symposium on Compositionality: The Significant Difference, pages 541–564, London, UK, 1998.
Springer-Verlag.
[75] A. Siirtola, S. Tripakis, and K. Heljanko. When do we not need complex assume-guarantee rules? ACM
Trans. Embed. Comput. Syst., 16(2):48:1–48:25, Jan. 2017.
[76] A. P. Sistla, M. Y. Vardi, and P. Wolper. The Complementation Problem for Büchi Automata with
Applications to Temporal Logic. Theoretical Computer Science, 49:217–237, 1987.
[77] G. Stefănescu. Network Algebra. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2000.
[78] S. Tripakis, B. Lickly, T. Henzinger, and E. Lee. On Relational Interfaces. In Proceedings of the 9th ACM
& IEEE International Conference on Embedded Software (EMSOFT’09), pages 67–76. ACM, 2009.
[79] S. Tripakis, B. Lickly, T. A. Henzinger, and E. A. Lee. A Theory of Synchronous Relational Interfaces.
ACM Trans. Program. Lang. Syst., 33(4):14:1–14:41, July 2011.
47
[80] S. Tripakis, C. Sofronis, P. Caspi, and A. Curic. Translating Discrete-time Simulink to Lustre. ACM
Trans. Embed. Comput. Syst., 4(4):779–818, Nov. 2005.
[81] N. Wirth. Program development by stepwise refinement. Comm. ACM, 14(4):221–227, 1971.
[82] D. Yadav and M. Butler. Verification of liveness properties in distributed systems. In Contemporary Computing, volume 40 of Communications in Computer and Information Science, pages 625–636.
Springer Berlin Heidelberg, 2009.
[83] C. Yang and V. Vyatkin. Transformation of Simulink models to IEC 61499 Function Blocks for verification of distributed control systems. Control Engineering Practice, 20(12):1259–1269, 2012.
[84] C. Zhou and R. Kumar. Semantic Translation of Simulink Diagrams to Input/Output Extended Finite
Automata. Discrete Event Dynamic Systems, 22(2):223–247, 2012.
[85] L. Zou, N. Zhany, S. Wang, M. Franzle, and S. Qin. Verifying Simulink diagrams via a Hybrid Hoare
Logic Prover. In Embedded Software (EMSOFT), Sept 2013.
48
| 6cs.PL
|
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
arXiv:1502.03434v3 [math.CV] 9 Sep 2015
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Abstract. We study a new biholomorphic invariant of holomorphic maps between domains in different dimensions based on generic initial ideals. We start with the standard
generic monomial ideals to find invariants for rational maps of spheres and hyperquadrics,
giving a readily computable invariant in this important case. For example, the generic initial
monomials distinguish all four inequivalent rational proper maps from the two to the three
dimensional ball. Next, we associate to each subspace X ⊂ O(U ) a generic initial monomial subspace, which is invariant under biholomorphic transformations and multiplication
by nonzero functions. The generic initial monomial subspace is a biholomorphic invariant for holomorphic maps if the target automorphism is linear fractional as in the case of
automorphisms of spheres or hyperquadrics.
1. Introduction
Let U ⊂ Cn and V ⊂ Cm be domains. Denote by O(U, V ) the set of holomorphic maps
f : U → V.
(1)
Write O(U) = O(U, C) as usual. A fundamental problem in several complex variables is to
understand maps in O(U, V ) up to automorphisms; that is, f : U → V and g : U → V are
equivalent if there exist biholomorphisms τ ∈ Aut(U) and χ ∈ Aut(V ) such that
f ◦ τ = χ ◦ g.
(2)
A particularly important case is when the maps are proper. A map f : U → V is proper
if for every compact K ⊂⊂ V , the set f −1 (K) is compact. If f extends continuously to
the boundary, then the proper map f takes boundary to boundary. The map f restricted
to the boundary gives a CR map, and therefore, we have a problem in CR geometry. It
is important to understand those situations where V has a large automorphism group, and
thus we focus most of our attention on the unit ball of a certain signature. For a pair of
integers (a, b), a ≥ 1, the unit ball of signature b is given by
Ba+b
b
b
a+b
o
n
X
X
2
a+b
|zj |2 < 1 .
= z = (z1 , . . . , za+b ) ∈ C
:−
|zj | +
j=1
(3)
j=1+b
is the
The ball Bn0 with signature 0 is the standard unit ball Bn . The boundary of Ba+b
b
hyperquadric
b
a+b
o
n
X
X
2
a+b
|zj |2 = 1 .
(4)
Q(a, b) = z ∈ C
:−
|zj | +
j=1
j=1+b
Date: September 9, 2015.
The second author was in part supported by NSF grant DMS-1362337 and Oklahoma State University’s
DIG and ASR grants.
1
2
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Hyperquadrics are the model hypersurfaces for Levi-nondegenerate surfaces. The hyperquadrics are also the flat models in CR geometry from the point of view of Riemannian
geometry.
The CR geometry problem of maps into hyperquadrics has a long history beginning with
Webster who in 1978 [17] proved that every algebraic hypersurface can be embedded into
a hyperquadric for large enough a and b. On the other hand Forstnerič [8] proved that in
general a CR submanifold will not map to a finite dimensional hyperquadric. Thus not every
CR submanifold can be realized as a submanifold of the flat model (i.e. the hyperquadric).
The mapping problem has been studied extensively for proper maps between balls, where
the related CR question is to classify the CR maps between spheres. That is, suppose
f : Bn → BN is a proper holomorphic map. Two such maps f and g are spherically equivalent
if there exist automorphisms τ ∈ Aut(Bn ) and χ ∈ Aut(BN ) such that f ◦ τ = χ ◦ g.
There has been considerable progress in the classification of such maps, especially for small
codimensional cases; see for example [1, 3, 4, 6, 9, 14] and the many references within.
In particular, Forstnerič [9] proved that given sufficient boundary regularity, the map
is rational and the degree bounded in terms of the dimensions only. The degree is an
invariant under spherical equivalence and the sharp bounds for the degree were conjectured
by D’Angelo [5].
In this article we introduce a new biholomorphic invariant that is far more general than
degree. When the map is rational and the domain and target are balls (possibly of nonzero
signature), we obtain invariants via generic initial ideals from commutative algebra. For
arbitrary domains and maps, we generalize the smallest degree part of the ideal.
The techniques center on the idea of generic initial monomial ideals; see e.g. Green [12].
Given a homogoeneous ideal I, the monomial ideal in(I) is generated by the initial monomials of elements of I (see section 3 for precise definitions). The generic initial ideal, denoted
gin(I), is an initial monomial ideal after precomposing I with a generic invertible linear
map. Grauert was the first to introduce generic initial ideals to several complex variables
(see [10]), and they have been used extensively for understanding singularities of varieties.
For rational maps of balls of the form fg , we homogenize f and g, and we then look at
the ideal generated by the components. We prove that the generic initial ideal generated by
the homogenizations of f and g is invariant under spherical equivalence. More precisely, the
first main result is the following.
Theorem 1.1. Suppose f1 : Bn → BN and f2 : Bn → BN are rational proper maps that are
spherically equivalent. Let F1 and F2 be the respective homogenizations. Then
gin I(F1 ) = gin I(F2 ) .
(5)
In section 3, we prove this result, and compute this new invariant for a number of wellknown examples. In particular we show how this invariant distinguishes many of these maps.
In section 4, we obtain a further invariant by considering the holomorphic decomposition of
the quotient
kf (z)k2 − |g(z)|2
.
(6)
kzk2 − 1
To generalize these results to arbitrary holomorphic maps we improve a result proved in
Grundmeier-Lebl-Vivas [13] that itself is a version of Galligo’s theorem for vector subspaces
of O(U). Given a subspace X ⊂ O(U), we precompose with a generic affine map τ , and
define the generic initial monomial subspace as the space spanned by the initial monomials of
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
3
elements of X ◦ τ , denoted gin(X). Let X be the affine span of the components of a holomorphic map F : U → CN , denoted by affine-span F . See section 5 for precise definitions. We
obtain an invariant under biholomorphic transformations of U and invertible linear fractional
transformations of CN . By using affine rather than just linear maps we obtain invariants of
the map without having a distinguished point. The second main result says this is, indeed,
an invariant.
Theorem 1.2. Let M, M ′ ⊂ Cn be connected real-analytic CR submanifolds. Suppose
F : M → Q(a, b) and G : M ′ → Q(a, b) are real-analytic CR maps equivalent in the sense
that there exists a real-analytic CR isomorphism τ : M ′ → M and a linear fractional automorphism χ of Q(a, b) such that
F ◦ τ = χ ◦ G.
(7)
Then
gin(affine-span F ) = gin(affine-span G).
(8)
In sections 5 and 6, we give precise definitions and lemmas. In particular we prove the
extended version of Galligo’s theorem for our setting. In section 7, we prove the second major
theorem of this paper, and we give examples where we compute this invariant. Finally, in
section 8 we define the generic initial subspace for the analogue of the quotient (6) for CR
maps between spheres and hyperquadrics.
The authors would like to acknowledge John D’Angelo for many conversations on the
subject.
2. The projective setting
Before we work with arbitrary maps, we consider the special case of rational maps between
hyperquadrics and spheres. In this case we directly apply the standard theory of generic
initial ideals to obtain invariants.
For z, w ∈ Ca+b define
def
hz, wib = hIb z, wi = −
b
X
zj w̄j +
j=1
a+b
X
zj w̄j
and
j=b+1
def
kzk2b = hz, zib .
(9)
By Ib we mean the (a + b) × (a + b) diagonal matrix with b (−1)’s and a 1’s on the diagonal.
We use the same definition for the homogeneous case when the index on the variables starts
with a zero. In this case the subscript still refers to the number of negatives. That is, for
Z = (Z0 , . . . , Za+b ) ∈ Ca+b+1 we write
def
hZ, W ib+1 = hIb+1 Z, W i = −
b
X
j=0
Zj W̄j +
a+b
X
j=b+1
Zj W̄j
and
def
kZk2b+1 = hZ, Zib+1. (10)
Let us homogenize Q(a, b). We add a variable Z0 and work with the homogeneous coordinates [Z0 , Z1 , . . . , Zn ] in Pn . Homogenize the equation above to obtain
def
(11)
HQ(a, b + 1) = Z ∈ Pa+b : kZk2b+1 = 0 .
If we think of Ca+b ⊂ Pa+b , then Q(a, b) is a subset of Pa+b and HQ(a, b + 1) is the closure
of Q(a, b) in Pa+b .
4
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Automorphisms of Pa+b can be represented as invertible linear maps on Ca+b+1 . The
automorphisms of HQ(a, b + 1) are those linear maps T that preserve the form in (11) up
to a real scalar λ 6= 0, that is
(12)
kT Zk2b+1 = λ kZk2b+1 .
p
If we represent T as a matrix, then the condition above is T ∗ Ib+1 T = µIb , where µ is ± |λ|.
If a 6= b + 1, then λ > 0. If λ < 0, then the automorphism swaps the sides of HQ(a,
b + 1).
a+b
Write the set of the corresponding automorphisms of P
as Aut HQ(a, b + 1) . As usual
the point Z ∈ Pa+b is an equivalence class of points of Ca+b+1 up to complex multiple. Thus
(as long as a 6= b + 1) the group Aut HQ(a, b + 1) is the group SU(a, b + 1)/K where K is
the subgroup of matrices ζI where ζ a+b+1 = 1. If a = b + 1, we need to include the matrix
that switches sides. The important fact for us is that automorphisms are represented by
matrices.
Fix (a, b) and (A, B). A rational map F : Pa+b 99K PA+B is represented by a homogeneous
polynomial map of Ca+b+1 to CA+B+1 . The equivalence we wish to consider is the following.
a+b
Two rational maps
99K PA+B and G : Pa+b 99K PA+B are equivalent if there exists
F: P
a+b+1
τ ∈ Aut P
and χ ∈ Aut PA+B+1 such that
F ◦ τ = χ ◦ G.
(13)
In the applications we will have τ ∈ Aut HQ(a, b + 1) and χ ∈ Aut HQ(A, B + 1) .
If f : Bn → BN is a rational proper map of balls, its homogenization F : Pn 99K PN takes
HQ(n, 1) to HQ(N,
1)
1). The equivalence on such maps F using the groups Aut HQ(n,
n
and Aut HQ(N, 1) is precisely the standard spherical equivalence on f ; that is, f : B → BN
and g : Bn → BN are spherically equivalent if there exist automorphisms τ and χ of Bn and
BN respectively such that f ◦ τ = χ ◦ g.
3. Generic initial ideals and rational maps
In this section we briefly introduce the relevant definitions and results from commutative
algebra. We use the setup from Green [12], and in a later section we use the techniques
developed by the authors in [13]. To be consistent with these two papers, we use the slightly
unusual monomial ordering as used by Green.
Let Z0 , Z1 , . . . , Zn denote our variables. Given a multi-index α ∈ Nn+1
we write Z α to
0
mean Z0α0 Z1α1 · · · Znαn as usual, and let |α| = α0 + · · · + αn denote the total degree. A
multiplicative monomial order is a total ordering on all monomials Z α , such that
(i) Z0 > Z1 > · · · > Zn ,
(ii) Z α > Z β ⇒ Z γ Z α > Z γ Z β ,
(iii) |α| < |β| ⇒ Z α > Z β .
Such orderings are not unique. Using the so-called graded reverse lexicographic ordering
when n = 2, we obtain
1 > Z0 > Z1 > Z2 > Z02 > Z0 Z1 > Z12 > Z0 Z2 > Z1 Z2 > Z22 > · · · .
(14)
Fix a certain monomial order. Given a set S of monomials, then the initial monomial is
the maximal monomial in S according to the ordering. For a homogeneous polynomial P ,
write
def
in(P ) = initial monomial of P .
(15)
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
5
Let I be a homogeneous ideal. Define the initial monomial ideal in(I) as the smallest ideal
such that z α ∈ in(I) whenever z α = in(P ) for some P ∈ I.
Galligo’s theorem implies that for an open dense set of linear maps T , in(I) = in(I ◦ T ).
Let T be such a generic linear map and define
def
gin(I) = in(I ◦ T ).
(16)
I(F ) = ideal generated by components of F in homogeneous coordinates.
(17)
F ◦ τ = χ ◦ G.
(18)
n
Finally, for a rational map F : P 99K P
N
define
Proposition 3.1. Suppose
F : Pn 99K PNand G : Pn 99K PN are equivalent in the sense that
n
there exist τ ∈ Aut P and χ ∈ Aut PN such that
Then
gin I(F ) = gin I(G) .
(19)
Proof. By definition of the generic initial ideal, gin I(F ◦ τ ) = gin I(F ) , as τ is an
invertible linear map. The ideal generated by linear combinations of G is the
same as the
one
generated by G, and χ is also an invertible linear map. So, gin I(χ ◦ G) = gin I(G) .
In the important special case of rational proper maps of balls we get the first main theorem
from the introduction.
Corollary 3.2. Suppose f : Bn → BN and g : Bn → BN are rational proper maps that are
spherically equivalent. Let F and G be the respective homogenizations. Then
gin I(F ) = gin I(G) .
(20)
We close this section with several examples. The following computations are done with
Macaulay2 [11] with the GenericInitialIdeal package using the standard graded reverse
lex ordering unless stated otherwise. One advantage of using gins as invariants is that they
are very simple to compute using computer algebra systems.
Example 3.3. Generic initial ideals distinguish all the maps from B2 to B3 . Faran [6] proved
that all maps sufficiently smooth up to the boundary are spherically equivalent to one of the
following 4 maps:
(i) (z1 , z2 ) 7→ (z1 , z2 , 0),
(ii) (z1 , z2 ) 7→ (z1 , z√1 z2 , z22 ),
(iii) (z1 , z2 ) 7→ (z12 , √2z1 z2 , z22 ),
(iv) (z1 , z2 ) 7→ (z13 , 3z1 z2 , z23 ).
Let us list the results in a table. The first column gives the map itself. The second column
gives the map homogenized by adding the Z0 variable. The last column gives the generators
of the generic initial ideal.
Map
Homogenized
gin
(z1 , z2 , 0)
(Z0 , Z1 , Z2 , 0)
(Z0 , Z1 , Z2 )
2
2
2
(z1 , z√1 z2 , z2 )
(Z0 , Z1 Z√
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z1 Z22 )
0 , Z1 Z2 , Z2 )
(z12 , √2z1 z2 , z22 ) (Z02 , Z12 , √2Z1 Z2 , Z22 )
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z23 )
3
3
3
3
3
(z1 , 3z1 z2 , z2 ) (Z0 , Z1 , 3Z1 Z2 Z0 , Z2 ) (Z03 , Z02 Z1 , Z0 Z12 , Z02 Z2 ,
Z14 , Z13 Z2 , Z0 Z1 Z22 , Z12 Z22 ,
Z0 Z24 , Z1 Z24 , Z25 )
6
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Notice that the generic initial ideals are all different. In particular, we distinguish the two
very similar second degree maps using the third degree part of the ideal.
Example 3.4. In [7] it was proved that the proper map of B2 to B4 given by
q
2
3
z
1
−
|a|
√
− a) 2
(z1 , z2 ) 7→ z12 , 2z1 z2 ,
,
1 − āz1
1 − āz1
z22 (z1
(21)
is spherically equivalent to a polynomial map if and only if a = 0. This map is the homogeneous second degree map from Faran’s theorem above with the last component tensored by
an automorphism of the ball B2 taking a to 0. If a = 0, then the generic initial ideal is
(Z03 , Z02 Z1 , Z0 Z12 , Z13 , Z02 Z2 , Z0 Z1 Z22 , Z12 Z22 , Z0 Z24 ).
(22)
If a = 21 , we obtain a different generic initial ideal
(Z03 , Z02 Z1 , Z0 Z12 , Z13 , Z02 Z2 , Z0 Z1 Z22 , Z12 Z22 , Z0 Z23 ).
(23)
Example 3.5. In [15] the second author classified all the maps from the two sphere, Q(2, 0),
to the hyperquadric Q(2, 1) (see also Reiter [16] for a different approach to the classification).
There are 7 equivalence classes of maps:
(i) (z1 , z2 ) 7→ (0, z1 , z2√),
(ii) (z1 , z2 ) 7→
(z22 , z12 , 2z2 ),
z2
(iii) (z1 , z2 ) 7→ zz22 , z11 , z22 ,
12 √ 1 2
√
√
z − 3 z z +z −z z 2 + 3 z z +z 2 −z z 2 +z − 3 z −1
(iv) (z1 , z2 ) 7→ 1z 2 +z +1 √23 z 2−1 1 , 1z 2 +z +1 √23 z 2−1 1 , z22 +z1 +√3 z2 −1 ,
2
2
2
1
2
2 1
√4 2 1
√
√
2(z1 z2 +iz1 ) 4 2(z1 z2 −iz1 ) z22 − 2 iz2 +1
√
√
√
(v) (z1 , z2 ) 7→ z 2 + 2 iz +1 , z 2 + 2 iz +1 , z 2 + 2 iz +1 ,
2
2
2
2
√2
2
z2 z12 −z2
2z23
z13 +3z1
(vi) (z1 , z2 ) 7→
3 3z 2 +1 , 3z 2 +1 , 3z 2 +1 ,
1
1
1
(vii) (z1 , z2 ) 7→ g(z1 , z2 ), g(z1 , z2 ), 1 for an arbitrary holomorphic function g.
Note that the maps are written slightly differently from [15] as we are ordering the negative
components first.
We list the generic initial ideal for each of the first 6 classes of maps. The final 7th class
is already distinguished as it is not transversal; that is, it maps an open neighborhood into
Q(2, 1).
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
Map
(i)
(ii)
(iii)
(iv)
(v)
Homogenized
(Z0 , 0, Z1 , Z2√
)
2
2
2
(Z0 , Z2 , Z1 , 2Z2 Z0 )
2
(Z12 , Z2 Z0 , Z1Z√
0 , Z2 )
(Z22 + Z
3Z2 Z0 − Z02 ,
1 Z0 +
√
Z12 − √3Z1 Z2 + Z22 − Z1 Z0 ,
Z12 + 3Z1 Z2√
+ Z22 − Z1 Z0 ,
2
Z22 + Z
√1 Z0 − 3Z22Z0 − Z0 )
2
(Z
2iZ2 Z0 + Z0 ,
2 +
√
4
2(Z1 Z2 + iZ1 Z0 ),
√
4
2(Z√
1 Z2 − iZ1 Z0 ),
2
2
Z2 − 2iZ2 Z√
0 + Z0 )
(3Z12 Z0 + Z03 , 3(Z2 Z12 − Z2 Z02 ),
2Z23 , Z13 + 3Z1 Z02 )
7
gin
(Z0 , Z1 , Z2 )
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z1 Z22 , Z23 )
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z1 Z22 )
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z1 Z22 )
(Z02 , Z0 Z1 , Z12 , Z0 Z2 , Z1 Z22 )
(Z03 , Z02 Z1 , Z0 Z12 , Z02 Z2 ,
Z14 , Z0 Z1 Z22 , Z12 Z22 , Z13 Z2 ,
Z0 Z24 , Z1 Z24 , Z25 )
We distinguish some maps, but not all. It is to be expected that not all maps can be
distinguished. After all, computing a generic initial ideal throws away much information
about the map. It should be noted that we only looked at small degree examples, where the
number of different possible gins is limited.
(vi)
Example 3.6. The particular monomial ordering may make a difference, even in simple
situations. It might be possible to tell some ideals (and hence maps) apart using one ordering,
but not using another. Let us give a very simple example. In P2 , the ideal generated by
(Z02 , Z1 Z2 ) has the gin
(Z02 , Z0 Z1 , Z13 )
in graded reverse lex ordering,
(Z02 , Z0 Z1 , Z0 Z22 , Z14 )
in graded lex ordering.
(24)
However, the ideal generated by (Z02 , Z12 ) has the same gin in both orderings:
(Z02 , Z0Z1 , Z13 ).
(25)
4. Invariants of the quotient for rational maps
Let F : Pa+b 99K PA+B be a rational map that takes HQ(a, b + 1) to HQ(A, B + 1), where
defined. Identify F with the homogeneous polynomial map taking Ca+b+1 to CA+B+1 . Write
kF (Z)k2B+1 = kZk2b+1 q(Z, Z̄),
(26)
q(Z, Z̄) = kh+ (Z)k2 − kh− (Z)k2 ,
(27)
H(q) = {h+ , h− }
(28)
where the quotient q is a bihomogeneous polynomial in Z. Find the holomorphic decomposition (see p. 101 of [2]) of q as
where h+ and h− are homogeneous holomorphic polynomial maps. Write
for the set of functions in the holomorphic decomposition of q. The holomorphic decomposition is not unique. However, the linear span of H(q) is unique,
and therefore the ideal
generated by H(q) is unique.
We thus study the ideal I H(q) and furthermore the generic
initial ideal gin I H(q) .
8
DUSTY GRUNDMEIER AND JIŘÍ LEBL
The reason for looking at q is that it may reveal further information about the map that
is not immediately visible from F . For example the quotient q was critically used in [5] for
the degree estimates problem. The number of functions in the decomposition of q is often
larger than the number of components of F as many of these components may cancel out
once multiplied with kZk2b+1 .
Suppose F : Pa+b 99K PA+B and G : Pa+b 99K PA+B are equivalent as before; i.e. F ◦τ = χ◦
G, where τ and χ are automorphisms preserving HQ(a, b+1) and HQ(A, B +1) respectively.
Let qF and qG be the corresponding quotients of F and G respectively. We regard τ and χ
as linear maps, which we also rescale, such that kτ Zk2b+1 = ± kZk2b+1 for Z ∈ Ca+b+1 and
kχW k2B+1 = ± kW k2B+1 for W ∈ CA+B+1 . The ± is there in case a = b + 1 or A = B + 1
and the sides are swapped, otherwise it would be a +.
As F and G are equivalent, kF (τ Z)k2B+1 = kχ ◦ G(Z)k2B+1 . Then
kZk2b+1 qF (τ Z, τ Z) = ± kτ Zk2b+1 qF (τ Z, τ Z) = ± kF (τ Z)k2B+1 =
= ± kχ ◦ Gk2B+1 = ± kG(Z)k2B+1 = ± kZk2b+1 qG (Z, Z̄). (29)
In other words qF (τ Z, τ Z) = ±qG (Z, Z̄).
a+b
Proposition 4.1. Suppose F : Pa+b 99K PA+B
99K PA+B are equivalent in the
and G : P
sense that there exist τ ∈ Aut HQ(a, b + 1) and χ ∈ Aut HQ(A, B + 1) such that
F ◦ τ = χ ◦ G.
Let qF and qG be the corresponding quotients. Then
gin I H(qF ) = gin I H(qG ) .
(30)
(31)
Proof. Above we proved qF (τ Z, τ Z) = ±qG (Z, Z̄). As we are talking about gins, the τ is not
relevant. The ± does not change the components of the holomorphic decomposition.
Again, in the important special case of rational proper maps of balls we get:
Corollary 4.2. Suppose f : Bn → BN and g : Bn → BN are rational proper maps that are
spherically equivalent. Let F and G be the respective homogenizations and let qF and qG be
the corresponding quotients. Then
gin I H(qF ) = gin I H(qG ) .
(32)
√
Example 4.3. Let us consider the Faran map (z1 , z2 ) 7→ (z13 , 3z1 z2 , z23 ) that takes B2 to
B3 . We compute the quotient
√
2
2
2
|z13 | + 3z1 z2 + |z23 | − 1
2
2
(33)
= z12 + z22 − |z1 z2 |2 + |z1 |2 + |z2 |2 + 1.
2
2
|z1 | + |z2 | − 1
Bihomogenizing the quotient, we obtain all the degree-two monomials in H(q). Therefore
the gin is generated by all the degree-two monomials:
gin I H(q) = (Z02 , Z0 Z1 , Z12 , Z0Z2 , Z1 Z2 , Z22 ).
(34)
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
9
5. Affine spans and automorphisms of the target
To deal with nonrational maps we switch to the affine setting. We work in CN and treat
it as a subset of PN . From now on z = (z1 , . . . , zN ) will be the inhomogeneous coordinates
on PN , that is, coordinates on CN . In other words, we fix a specific embedding ι : CN → PN .
For equivalence of maps F : U → CN we consider the target automorphisms to be the linear
fractional automorphisms of Aut(PN ) and consider CN ⊂ PN as above. Take χ ∈ Aut(PN ),
F : U → CN , and G : U → CN . When we write an equation such as χ ◦ G = F , we mean
χ ◦ ι ◦ G = ι ◦ F , where ι is the embedding above. That is, we consider F to be valued in PN
using the embedding ι : CN → PN . Note that after composing with automorphisms of PN
given as linear fractional maps, it may happen that the new map has poles in U. In fact,
when looking at linear fractional automorphisms of Q(a, b), there will in general be poles
that intersect Q(a, b).
The reason for not simply working in projective space is that the components in nonrational
holomorphic maps F : U → CN cannot be homogenized; the degree is unbounded. For this
same reason, we have to work with affine span instead of just linear span.
Definition 5.1. Given a collection of holomorphic functions F , the affine span of F is
affine-span F = span F ∪ {1} .
(35)
If F is a map, then affine-span F is the affine span of the components of F . Let X ⊂ O(U)
be a vector subspace and ϕ ∈ O(U). Denote by ϕX the vector subspace obtained by
multiplying every element in X by ϕ.
Lemma 5.2. Suppose F : U ⊂ Cn → CN and G : U ⊂ Cn → CN are such that there exists
χ ∈ Aut(PN ) such that F = χ ◦ G. Then there exists ϕ ∈ affine-span G such that
affine-span G = ϕ affine-span F.
(36)
P (w)
◦G(z)
Proof. Let χ(w) = Q(w)
be the linear fractional automorphism such that F (z) = PQ◦G(z)
for
N
N
N
z ∈ U. Here P : C → C and Q : C → C are affine maps. Therefore (Q ◦ G)F = P ◦ G.
Therefore the components of (Q ◦ G)F are in the affine span of G. Also Q ◦ G is in the affine
span of G. The span of the components of (Q ◦ G)F and the function (Q ◦ G) is exactly
(Q ◦ G) affine-span F , and therefore (Q ◦ G) affine-span F ⊂ affine-span G.
Clearly dim affine-span F ≤ dim affine-span G. By inverting χ and applying the argument
in reverse we get dim affine-span G ≤ dim affine-span F . Therefore, (Q ◦ G) affine-span F =
affine-span G.
The point of the lemma is to eliminate the target automorphism by looking at the vector
space generated by F1 , . . . , FN and 1, up to multiplication by a scalar-valued function. To find
invariants of the map F we need to find invariants of this new object under biholomorphisms
of the source.
6. Generic initial vector space
In this section we define the gin of a vector subspace of holomorphic maps. The idea is to
generalize the generic initial ideals to the setting of vector spaces of holomorphic functions
using affine maps instead of linear maps. We generalize the setup from Green [12], using the
techniques developed by the authors in [13].
10
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Let monomial order be defined on z1 , . . . , zn in the same way as above. That is, given a
multi-index α ∈ Nn0 we write z α to mean z1α1 z2α2 · · · znαn , and |α| = α1 + · · · + αn is the total
degree. We require the order to be multiplicative:
(i) z1 > z2 > · · · > zn ,
(ii) z α > z β ⇒ z γ z α > z γ z β ,
(iii) |α| < |β| ⇒ z α > z β .
In the following definitions we fix a point. We assume this point is 0 ∈ Cn , although any
point can be used after translation.
Definition 6.1. Fix a monomial order. By an initial monomial from a collection we mean
the largest monomial in the order. Given a Taylor series T , the initial monomial of T is
the largest monomial in T with a nonzero coefficient. Suppose U ⊂ Cn and 0 ∈ U. For
f ∈ O(U), let T0 f be the Taylor series for f at 0 and
def
in(f ) = initial monomial of T0 f .
(37)
Given X ⊂ O(U) a vector subspace, define the initial monomial subspace
def
in(X) = span{z α : z α = in(f ) for some f ∈ X}.
(38)
A subspace X ⊂ O(U) is a monomial subspace if X admits an algebraic basis consisting of
monomials.
If X is a monomial subspace then the basis of monomials defining it is unique. Therefore,
there is a one-to-one equivalence between monomial subspaces and subsets of the set of all
monomials.
A set S of monomials is affine-Borel-fixed if whenever z α ∈ S and zj |z α , then for all ℓ < j,
the monomials z α z1j and z α zzjℓ are in S. If X ⊂ O(U) is a monomial subspace, then we say
that X is affine-Borel-fixed if the basis of monomials that generates X is affine-Borel-fixed.
Take an invertible affine self map τ of Cn such that τ (0) ∈ U. For a subspace X ⊂ O(U),
define a subspace X ◦ τ ⊂ O(U ′ ) where U ′ = τ −1 (U). Notice that 0 ∈ U ′ and we may again
talk about initial monomials as above.
The initial monomial space in(X) is not always preserved under such precomposition with
affine maps. For example, let X be the span of {1, z2 }, so X = in(X). For a generic choice of
an affine map τ we have in(X ◦ τ ) = span{1, z1 }. Notice that {1, z2 } is not affine-Borel-fixed,
but in(X ◦ τ ) is.
We need to prove an analogue to Galligo’s Theorem (see Theorem 1.27 in [12]). The finite
dimensional version of the result is proved in [13].
2
The set Aff(n) of affine self maps of Cn can be parametrized as Mn (C) × Cn , or Cn × Cn .
We use the standard topology on this set.
Theorem 6.2. Suppose U ⊂ Cn and 0 ∈ U, and let X ⊂ O(U) be a vector subspace.
There is some neighbourhood N of the identity in Aff(n) and A ⊂ N of second category
(countable intersection of open dense subsets of N ), such that for τ ∈ A, the space in(X ◦ τ )
is affine-Borel-fixed. Furthermore, in(X ◦ τ ) = in(X ◦ τ ′ ) for any two affine τ and τ ′ in A.
Finally, if X is finite dimensional, A is open and dense in N .
So after a generic affine transformation the initial monomial subspace is always the same
space. It will also not be necessary to assume 0 ∈ U. We pick a generic affine τ such that
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
11
τ (0) ∈ U. Then 0 ∈ τ −1 (U), and so it makes sense to take in(X ◦ τ ) for X ⊂ O(U). Before
we prove the theorem, let us make the following definition.
Definition 6.3. Let U ⊂ Cn and let X ⊂ O(U) be a subspace. Take a generic affine τ such
that τ (0) ∈ U. Define the generic initial monomial subspace
gin(X) = in(X ◦ τ ).
(39)
Remark 6.4. While the proof of the theorem may seem formal, we are using affine transformations and precomposing formal power series with affine transformations may not make
sense.
Proof of Theorem 6.2. The finite dimensional version of the theorem is proved in [13] as
Proposition 1 (or Proposition 5.5 on arXiv).
Suppose X is not finite dimensional. The space O(U) is a separable Frechét space. For
the subspace X ⊂ O(U), pick a countable set {fk }∞
k=1 in X such that if
Xm = span{f1 , f2 , . . . , fm },
(40)
[
(41)
then
m
Xm ⊂ X ⊂ X =
[
Xm .
m
Apply the finite dimensional result to Xm for each m. The τ varies over countably many
open dense sets and the intersection of those sets is a second category set as claimed.
Given any k, the initial k monomials in {f1 , f2 , . . . , fm } must stabilize as m grows. That
is, the first k monomials are the same for all large enough m. We therefore obtain a sequence
of initial monomials. To obtain the first k monomials, simply go far enough
S until the initial
k monomials stabilize. Then z α is in this sequence if and only if z α ∈ in m Xm .
It is left to show that
[
[
Xm = in
Xm .
(42)
in X = in
m
m
α
α
Suppose this is not
true.
Find
the
first
(maximal)
monomial
z
such
that
z
∈
in
X
, but
S
α
α
z 6∈ in m Xm . Suppose z is the kth monomial in the ordering given. Let πk be the
projection of O(U) onto the space spanned by the
first k monomials. By Cauchy’s formula,
SM
πk is continous. The dimension of πk m=1 Xm stabilizes as M grows, and hence for large
enough M
[
∞
πk XM = πk
(43)
Xm = πk X .
m=1
α
There exists an f ∈ X with z = in(f ). By the above equality there must exist a g ∈ XM
such that z α = in(g). That is a contradiction.
We now generalize the gins to biholomorphic maps. Let X ⊂ O(U) be a subspace and
let f : U ′ → U be a holomorphic map. Define X ◦ f to be the vector subspace of O(U ′ )
consisting of all F ◦ f for F ∈ X.
Lemma 6.5. Let X be a vector subspace of O(U) and let f : U ′ → U be a biholomorphic
map. Then gin(X) = gin(X ◦ f ).
12
DUSTY GRUNDMEIER AND JIŘÍ LEBL
Proof. Suppose 0 is in both U ′ and U, and suppose f (0) = 0. Composing f with an invertible
linear map does not change the gin (we must change U appropriately). Hence, assume f ′ (0)
is the identity. Let
f (z) = z + E(z),
(44)
where E is of order 2 and higher. We only need to show that in(X) = in(X ◦ f ).
Suppose z α ∈ in(X), that is, there exists an element of X of the form
X
cβ z β .
(45)
g(z) = z α +
β<α
In (g ◦ f ), the terms from E create monomials of degree strictly higher than |α|. So
X
dβ z β .
(g ◦ f )(z) = z α +
(46)
β<α
α
for some dβ . Therefore z ∈ in(X ◦ f ).
Since f −1 (z) = z + F (z) for some F or higher order we obtain by symmetry that if
α
z ∈ in(X ◦ f ) then if z α ∈ in(X ◦ f ◦ f −1 ) = in(X).
Lemma 6.6. Let X be a vector subspace of O(U), 0 ∈ U, and let ϕ : U → C be a holomorphic
function that is not identically zero. Then gin(X) = gin(ϕX).
Proof. As we are talking about gins, we precompose with a generic affine map, which also
precomposes ϕ, and therefore we assume that ϕ(0) 6= 0. In fact, without loss of generality
we assume ϕ(0) = 1. Suppose z α ∈ in(X), that is, there is a g ∈ X of the form
X
g(z) = z α +
cβ z β .
(47)
β<α
Then
ϕ(z)g(z) = z α +
X
dβ z β ,
(48)
β<α
for some dβ by multiplicativity of the monomial ordering. Therefore z α ∈ in(ϕX). By
symmetry, in(X) = in(ϕX).
7. Gins as invariants of maps
If X ⊂ O(U) is a subspace and U ′ ⊂ U is an open set then the restriction X|U ′ (the space
of restrictions of maps from X to U ′ ) clearly has the same gin as X. Let us extend gins to
complex manifolds. Let X ⊂ O(U) be a subspace where U is a connected complex manifold
of dimension n. As gin is invariant under biholomorphic transformations, it is well-defined
on every chart for U. If two (connected) charts overlap, then the gin must be the same
(compute the gin on the intersection). Therefore, gin(X) is well-defined.
Theorem 7.1. Let U, W be connected complex manifolds of dimension n. Suppose F : U →
CN and G : W → CN are equivalent in the sense that there exists a biholomorphic map
τ : W → U and a linear fractional automorphism χ ∈ Aut(PN ) such that
Then
F ◦ τ = χ ◦ G.
(49)
gin(affine-span F ) = gin(affine-span G).
(50)
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
13
Proof. As noted above, we take two charts of U and W , and therefore, without loss of
generality we assume that U and W are domains in Cn .
As F ◦ τ and G are equivalent via a target automorphism, Lemma 5.2 says
affine-span(F ◦ τ ) = ϕ affine-span G.
As affine-span(F ◦ τ ) = (affine-span F ) ◦ τ , Lemma 6.5 says
gin affine-span(F ◦ τ ) = gin(affine-span F ).
(51)
(52)
Via Lemma 6.6, we get
gin(ϕ affine-span G) = gin(affine-span G).
The result follows.
(53)
The same proof is used for the following CR version. In the CR version we start with realanalytic CR maps. Since real-analytic CR maps extend to holomorphic maps, we therefore
work with the extended holomorphic maps when taking gins and affine spans.
Corollary 7.2. Let M, M ′ ⊂ Cn be connected real-analytic CR submanifolds. Suppose
F : M → Q(a, b) and G : M ′ → Q(a, b) are real-analytic CR maps equivalent in the sense
that there exists a real-analytic CR isomorphism τ : M ′ → M and a linear fractional automorphism χ of Q(a, b) such that
F ◦ τ = χ ◦ G.
(54)
Then
gin(affine-span F ) = gin(affine-span G).
(55)
Again we may take M and M ′ to be submanifolds of a complex manifold of dimension n
instead of Cn .
For examples we need only look at rational maps. If the map is rational we homogenize
with Z0 as before. The generic initial monomial vector space is simply the lowest degree part
of the generic initial ideal generated by the components of the homogenized map. That is
because the lowest degree part of the ideal is the linear span of the components of the map.
We must be careful that the ordering is compatible. In particular, the ordering must respect
the “grading” above, that is in each degree if we set Z0 = 1 we must still get a multiplicative
monomial ordering. For example the standard reverse lex ordering in each degree will not
work, although the lex order will.
√
3
As an example
take
the
Faran
map
F
(z)
=
(z
,
3z1 z2 , z23 ). We add 1 and homogenize to
1
√
3
3
3
get (Z0 , Z1 , 3Z0 Z1 Z2 , Z2 ). We compute the gin in the graded lex order. The degree 3 part
of the gin is generated by
Z03 , Z02 Z1 , Z02 Z2 , Z0 Z12 .
(56)
Therefore,
gin(affine-span F ) = span{1, z1 , z2 , z12 }.
(57)
8. Invariants of the quotient of a hyperquadric map
When the source is either the ball Ba+b
in the holomorphic case or Q(a, b) in the CR case,
b
we obtain further invariants by considering the quotient of the function composed with the
defining function of the target divided by the defining function of the source as we did in
the rational case.
14
DUSTY GRUNDMEIER AND JIŘÍ LEBL
P
P
2
Recall the definition kzk2b = − bj=1 |zj |2 + a+b
j=b+1 |zj | . The defining function for Q(a, b)
is then kzk2b − 1 = 0. Suppose F : U ⊂ Q(a, b) → Q(A, B) is a real-analytic CR map. A
real-analytic CR map is a holomorphic map of a neighborhood of U, and so we will identify
F with this holomorphic map. Define the quotient q via
kF (z)k2B − 1 = kzk2b − 1 q(z, z̄).
(58)
Near some point find the holomorphic decomposition of q:
q(z, z̄) = kh+ (z)k2 − kh− (z)k2 ,
(59)
H(q) = {h+ , h− }
(60)
where h+ and h− are possibly ℓ2 valued holomorphic maps. Write
for the set of functions in the holomorphic decomposition of q. The holomorphic decomposition is not unique and depends on the point. However we do have the following lemma.
Lemma 8.1. Suppose q : U ⊂ Cn → R is real-analytic and U is a connected open set. If
H1 (q) and H2 (q) are two holomorphic decompositions at two points of U, then
(61)
gin span H1 (q) = gin span H2 (q) .
Proof. By a standard connectedness argument using a path between the two points we only
need to consider the case where the domains of convergence of H1 and H2 overlap. Assume
that we work on this intersection. So we only need to show that if
kh+ (z)k2 − kh− (z)k2 = h′+ (z)
2
− h′− (z)
2
(62)
′
′
for maps h+ , h− , h′+ , h′− converging on a fixed open set, with {h+ , h
− } and {h+ , h− } lin′
′
early independent sets, then gin span{h+ , h− } = gin span{h+ , h− } . From (62) we have
span{h+ , h− } = span{h′+ , h′− }, and so the result follows.
Thus, we do not need to specify which point and which decomposition is used.
Theorem 8.2. Suppose F : U ⊂ Q(a, b) → Q(A, B) and G : V ⊂ Q(a, b) → Q(A, B) are
real-analytic CR maps equivalent in the sense that there exists a linear fractional automorphism τ of Q(a, b) and a linear fractional automorphism χ of Q(A, B) such that
Define the quotients qF and qG via
kF (z)k2B − 1 = kzk2b − 1 qF (z, z̄)
F ◦ τ = χ ◦ G.
kG(z)k2B − 1 = kzk2b − 1 qG (z, z̄).
and
(63)
(64)
Then taking holomorphic decompositions of qF and qG at any point in U and V respectively
gin span H(qF ) = gin span H(qG ) .
(65)
Proof. Write the linear fractional maps τ and χ as
τ′
χ′
τ = ′′
and
χ = ′′ .
τ
χ
′
′′
′
′′
for affine maps τ , τ , χ , and χ . Then
2
so
2
kτ ′ (z)kb − |τ ′′ (z)| = ±(kzk2b − 1)
2
|τ ′′ (z)| (kτ (z)k2b − 1) = ±(kzk2b − 1).
(66)
(67)
(68)
INITIAL MONOMIAL INVARIANTS OF HOLOMORPHIC MAPS
15
The ± is there again in case τ switches the sides of Q(a, b). We have the same equation for
w and χ with A, B instead of a, b.
We have kF (τ (z))k2B + 1 = kχ ◦ G(z)k2B + 1. Then
2
(kzk2b − 1)qF τ (z), τ (z) = ± |τ ′′ (z)| kτ (z)k2b − 1 qF τ (z), τ (z)
2
= ± |τ ′′ (z)| kF (τ (z))k2B − 1
2
= ± |τ ′′ (z)| kχ ◦ G(z)k2B − 1
τ ′′ (z)
= ± ′′
χ G(z)
In other words
τ ′′ (z)
= ± ′′
χ G(z)
2
kG(z)k2B − 1
(69)
2
kzk2b − 1 qG (z, z̄).
2
τ ′′ (z)
qG (z, z̄).
qF τ (z), τ (z) = ± ′′
(70)
χ G(z)
Multiplying by absolute value squared of a holomorphic function multiplies the elements
of the holomorphic decomposition by that function and so does not change the gin by
Lemma 6.6. Similarly composing with τ does not change the gin either by Lemma 6.5.
There is an equivalent theorem for holomorphic maps, when the automorhphism maps on
the source are simply the linear fractional automorphisms of Ba+b
b , that is, automorphisms of
a+b
a+b
a+b
P
preserving the closure of Bb in P . We lose no generality if we also allow swapping
sides (when a = b + 1) and hence we consider all linear fractional automorphisms of Q(a, b),
that is, self maps of Pa+b preserving HQ(a, b + 1).
Theorem 8.3. Suppose F : U ⊂ Ca+b → CA+B and G : V ⊂ Ca+b → CA+B are holomorphic
maps equivalent in the sense that there exists a linear fractional automorphism τ of Q(a, b),
where τ (V ) = U, and a linear fractional automorphism χ of Q(A, B) such that
F ◦ τ = χ ◦ G.
Define the quotients qF and qG as before:
and
kF (z)k2B − 1 = kzk2b − 1 qF (z, z̄)
kG(z)k2B − 1 = kzk2b − 1 qG (z, z̄).
(71)
(72)
Then taking holomorphic decompositions of qF and qG at any point in U and V respectively
gin span H(qF ) = gin span H(qG ) .
(73)
9. Gins of real-analytic functions
We end the article with a remark that Lemma 8.1 may be of independent interest, not
only for mapping questions. Let us consider the gin of a holomorphic decomposition of realanalytic functions. That is, for a domain U ⊂ Cn , consider two real-analytic functions of realanalytic functions r1 : U → R and r2 : U → R. Then say the functions are biholomorphically
equivalent if there exists a biholomorphism F : U → U, such that r1 = r2 ◦ F .
If we take H(rj ) to be the holomorphic decomposition of rj at any point of U, then
Lemma 8.1 says that if r1 and r2 are equivalent as above, then
gin affine-span H(r1 ) = gin affine-span H(r2 ) .
(74)
16
DUSTY GRUNDMEIER AND JIŘÍ LEBL
The gin as defined above is not a pointwise invariant. That is, a real-analytic function
defined on a connected open set has the same gin near every point. By the same argument as
before we can also define the gin of a real-analytic function on a connected complex manifold
U by noting that the gin is already well-defined if we take any connected chart.
References
[1] H. Alexander, Proper holomorphic mappings in C n , Indiana Univ. Math. J. 26 (1977), no. 1, 137–146.
MR0422699
[2] John P. D’Angelo, Several complex variables and the geometry of real hypersurfaces, Studies in Advanced
Mathematics, CRC Press, Boca Raton, FL, 1993. MR1224231
[3]
, Polynomial proper maps between balls, Duke Math. J. 57 (1988), no. 1, 211–219, DOI
10.1215/S0012-7094-88-05710-9. MR952233
, Hermitian analysis, Cornerstones, Birkhäuser/Springer, New York, 2013. From Fourier series
[4]
to Cauchy-Riemann geometry. MR3134931
[5] John P. D’Angelo, Šimon Kos, and Emily Riehl, A sharp bound for the degree of proper monomial mappings between balls, J. Geom. Anal. 13 (2003), no. 4, 581–593, DOI 10.1007/BF02921879. MR2005154
[6] James J. Faran, Maps from the two-ball to the three-ball, Invent. Math. 68 (1982), no. 3, 441–475, DOI
10.1007/BF01389412. MR669425
[7] James Faran, Xiaojun Huang, Shanyu Ji, and Yuan Zhang, Polynomial and rational maps between balls,
Pure Appl. Math. Q. 6 (2010), no. 3, 829–842, DOI 10.4310/PAMQ.2010.v6.n3.a10. MR2677315
[8] Franc Forstnerič, Embedding strictly pseudoconvex domains into balls, Trans. Amer. Math. Soc. 295
(1986), no. 1, 347–368, DOI 10.2307/2000160. MR831203
[9]
, Extending proper holomorphic mappings of positive codimension, Invent. Math. 95 (1989), no. 1,
31–61. MR969413
[10] Hans Grauert, Über die Deformation isolierter Singularitäten analytischer Mengen, Invent. Math. 15
(1972), 171–198 (German). MR0293127
[11] Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in algebraic
geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[12] Mark Green, Generic initial ideals, Six lectures on commutative algebra (Bellaterra, 1996), Progr. Math.,
vol. 166, Birkhäuser, Basel, 1998, pp. 119–186. MR1648665
[13] Dusty Grundmeier, Jiřı́ Lebl, and Liz Vivas, Bounding the rank of Hermitian forms and rigidity for CR
mappings of hyperquadrics, Math. Ann. 358 (2014), no. 3-4, 1059–1089, DOI 10.1007/s00208-013-0989-z.
MR3175150
[14] Xiaojun Huang, Shanyu Ji, and Wanke Yin, On the third gap for proper holomorphic maps between
balls, Math. Ann. 358 (2014), no. 1-2, 115–142, DOI 10.1007/s00208-013-0952-z. MR3157993
[15] Jiřı́ Lebl, Normal forms, Hermitian operators, and CR maps of spheres and hyperquadrics, Michigan
Math. J. 60 (2011), no. 3, 603–628, DOI 10.1307/mmj/1320763051. MR2861091
[16] Michael Reiter, Classification of Holomorphic Mappings of Hyperquadrics from C2 to C3 , to appear in
J. Geom. Anal. arXiv:1409.5968.
[17] S. M. Webster, Some birational invariants for algebraic real hypersurfaces, Duke Math. J. 45 (1978),
no. 1, 39–46. MR0481086
[18] Hassler Whitney, Complex analytic varieties, Addison-Wesley Publishing Co., Reading, Mass.-LondonDon Mills, Ont., 1972. MR0387634
Department of Mathematics, Ball State University, Muncie, IN 47306, USA
Current address: Department of Mathematics, Harvard University, Cambridge MA 02138, USA
E-mail address: [email protected]
Department of Mathematics, Oklahoma State University, Stillwater, OK 74078, USA
E-mail address: [email protected]
| 0math.AC
|
Quantum Recursion and Second Quantisation
Mingsheng Ying
arXiv:1405.4443v2 [quant-ph] 6 Aug 2014
University of Technology, Sydney, Australia
and
Tsinghua University, China
[email protected]; [email protected]
Abstract. This paper introduces a new notion of quantum recursion of which
the control flow of the computation is quantum rather than classical as in the notions of recursion considered in the previous studies of quantum programming.
A typical example is recursive quantum walks, which are obtained by slightly
modifying the construction of the ordinary quantum walks. The operational and
denotational semantics of quantum recursions are defined by employing the second quantisation method, and they are proved to be equivalent.
Keywords: Quantum case statement, quantum choice, quantum recursion, recursive quantum walks, second quantisation, Fock space
1 Introduction
Recursion is one of the central ideas of computer science. Most programming languages
support recursion or at least a special form of recursion such as while-loop. Recursion
has also been considered since the very beginning of the studies of quantum programming; for example, Selinger [18] introduced the notion of recursive procedure in his
functional quantum programming language QPL and defined the denotational semantics of recursive procedures in terms of complete partial orders of super-operators. Termination of quantum while-loops were analysed by Ying and Feng [20] in the case of
finite-dimensional state spaces. A quantum generalisation of Etessami and Yannakakis’s
recursive Markov chains was proposed by Feng et. al. [11]. But the control flows of
all of the quantum recursions studied in the previous literatures are classical because
branchings in them are determined by the outcomes of certain quantum measurements,
which are classical information. So, they can be appropriately called classical recursion
of quantum programs.
Quantum control flow was first introduced by Altenkirch and Grattage [3] by defining a quantum case statement in their quantum programming language QML that implements a unitary transformation by decomposing it into two orthogonal branches along
an orthonormal basis of a chosen qubit. Motivated by the construction of quantum walks
[1], [2], a different approach to quantum control flow was proposed by the author in
[21], [22] where a kind of quantum case statement was defined by employing an external quantum “coin”. Furthermore, the notion of quantum choice was defined as the
sequential composition of a “coin tossing” program and a quantum case statement. The
quantum control flow of programs is clearly manifested in these quantum case statement
and quantum choice.
2
Ying
This paper introduces a new notion of quantum recursion using quantum case statements and quantum choices. In contrast to the recursions in quantum programming
considered before, the control flows of this kind of quantum recursions are quantum
rather than classical. Interestingly, this notion of quantum recursion enables us to construct a new class of quantum walks, called recursive quantum walks, whose behaviours
seem very different from the quantum walks defined in the previous literatures.
In this paper, we define the operational and denotational semantics of quantum recursions. The equivalence between these two semantics of quantum recursions are established. Obviously, how to define the semantics of quantum recursions is a question
that can be asked within the basic framework of quantum mechanics. But surprisingly,
answering it requires the mathematical tools from second quantisation [8] because variable number of identical particles are employed to implement the quantum “coins”
involved in the computation of a quantum recursion.
This paper is organised as follows. To make the paper self-contained, in Section 2
we recall the notions of quantum case statement and quantum choice from [21], [22].
The syntax of quantum recursive programs is defined in Section 3. To give the reader
a clearer picture, we choose not to include quantum measurements in the declarations
of quantum recursions in this paper. It seems that quantum measurements can be added
by combining the ideas used in this paper and those in [21], [22], but the presentation
will be very complicated. In Section 4, recursive quantum walks are considered as an
example for further motivating the notion of quantum recursion. In particular, it is carefully explained that a formal description of the behaviour of recursive quantum walks
requires the second quantisation method - a mathematical framework in which we are
able to depict quantum systems with variable number of particles. For convenience of
the reader, the basics of second quantisation is briefly reviewed in Section 5. We define
the semantics of quantum recursions in two steps. The first step is carried out in Section
6 where we construct a domain of operators in the free Fock space and prove continuity
of semantic functionals of quantum programs with procedure identifiers. Then recursive
equations are solved in the free Fock space by introducing the creation functional and
by employing the standard fixed point theorem. The second step is completed in Section
7 where the solutions of recursive equations are symmetralised so that they can apply in
the physically meaningful framework, namely the symmetric and antisymmetric Fock
spaces of bosons and fermions. A special class of quantum recursions, namely quantum
while-loops with quantum control flow are examined in Section 8. A short conclusion
is drawn in Section 9 with several problems for further studies.
2 Quantum Case Statement and Quantum Choice
Case statement in classical programming languages is a very useful program construct
for case analysis, see [9] for example. A quantum extension of case statement was defined in terms of measurements in various quantum programming languages, for example, Sanders and Zuliani’s qGCL [17], [24] and Selinger’s QPL [18]. The author defined
another quantum case statement using external quantum “coin” and further introduced
quantum choice as a variant of quantum case statement in [21], [22]. In this section, we
recall these two program constructs from [22].
Quantum Recursion and Second Quantisation
3
2.1 Quantum “if...then...else”
Let us start from the simplest case - a quantum counterpart of “if...then...else”. Assume
that c is a qubit of which the state Hilbert space Hc has |0i, |1i as an orthonormal
basis. Furthermore, assume that U0 and U1 are two unitary transformations acting on
a quantum system q of which the state Hilbert space is Hq . The system q is called the
principal quantum system. The action of U0 on system q can be thought of as a quantum
program and is denoted U0 [q]. Similarly, we write U1 [q] for the action of U1 on q. Then
a kind of quantum “if...then...else” can be defined by employing qubit c as a “quantum
coin”, and it is written as:
qif [c] |0i → U0 [q]
fiq
|1i → U1 [q]
(1)
in a way similar to Dijktra’s guarded commands [9]. The semantics of statement (1) is
an unitary operator
△
U = (|0i → U0 )(|1i → U1 )
on the tensor product Hc ⊗ Hq (i.e. the state Hilbert space of the composed system of
“coin” c and principal system q) defined by
U |0, ψi = |0iU0 |ψi,
U |1, ψi = |1iU1 |ψi
for any |ψi in Hq . It can be represented by the following diagonal matrix
U0 0
.
U = |0ih0| ⊗ U0 + |1ih1| ⊗ U1 =
0 U1
We call U the guarded composition of U0 and U1 along with basis |0i, |1i. Moreover,
let V be a unitary operator in the state Hilbert space Hc of the “coin” c. The action of V
on c can also be thought of as a program and is denoted V [c]. Then the quantum choice
of U0 [q] and U1 [q] with “coin-tossing” V [c] is defined to be
V [c]; qif [c] |0i → U0 [q]
fiq
|1i → U1 [q]
(2)
Here and in the sequel, P ; Q denotes the sequential composition of programs P and
Q; that is, program P followed by program Q. Using a notation similar to probabilistic
choice in a probabilistic programming language like pGCL [15], program (2) can be
written as
U0 [q] ⊕V [c] U1 [q]
(3)
Obviously, the semantics of quantum choice (3) is the unitary matrix U (V ⊗ Iq ), where
Iq is the identity operator in Hq .
The idea of defining quantum “if...then...else” using “quantum coin” was actually
borrowed from quantum walks. Here, let us consider the one-dimensional quantum
walks [2] as an example.
4
Ying
Example 1 The simplest random walk is the one-dimensional walk in which a particle
moves on a lattice marked by integers Z, and at each step it moves one position left or
right, depending on the flip of a fair coin. The Hadamard walk is a quantum variant
of the one-dimensional random walk. Its state Hilbert space is Hd ⊗ Hp , where Hd =
span{|Li, |Ri}, L, R are used to indicate the direction Left and Right, respectively,
Hp = span{|ni : n ∈ Z}, and n indicates the position marked by integer n. One step
of the Hadamard walk is represented by the unitary operator W = T (H ⊗ I), where
the translation T is a unitary operator in Hd ⊗ Hp defined by
T |L, ni = |L, n − 1i,
for every n ∈ Z,
1
H= √
2
T |R, ni = |R, n + 1i
1 1
1 −1
is the Hadamard transform in the direction space Hd , and I is the identity operator in
the position space Hp . The Hadamard walk is described by repeated applications of
operator W .
Now let us see how the idea of quantum case statement and quantum choice disguises in the construction of the Hadamard walk. If we define the left and right translation operators TL and TR in the position space Hp by
TL |ni = |n − 1i,
TR |ni = |n + 1i
for each n ∈ Z, then the translation operator T can be broken into a quantum case
statement of TL and TR :
T = qif [d] |Li → TL [p]
fiq
|Ri → TR [p]
(4)
where d is a “direction coin”, and p is a variable used to denote the position. Furthermore, the single-step walk operator W can be seen as the quantum choice
TL [p] ⊕H[d] TR [p].
Recently, physicists have been very interested in implementing quantum control for
unknown subroutines [23], [7], [12], which is essentially a quantum “if...then...else”or
a quantum choice.
2.2 Quantum Case Statement and Quantum Choice with Multiple Branches
We now generalise the quantum case statement (1) and quantum choice (2) to the case
with more than two branches. Let n ≥ 2 and c denote an n−level quantum system
with state Hilbert space Hc = span{|0i, |1i, ..., |n − 1i}. For each 0 ≤ i < n, let Ui
Quantum Recursion and Second Quantisation
5
be a unitary operator or the zero operator in the state Hilbert space Hq of the principal
system q. Using system c as a “quantum coin”, we can define a quantum case statement:
qif [c] (i · |ii → Ui [q]) qif = qif [c] |0i → U0 [q]
|1i → U1 [q]
..........
fiq
(5)
|n − 1i → Un−1 [q]
The reason for allowing some of Ui ’s being the zero operator is that if Ui [q] is a program
containing recursion then it may not terminate. In the case that Ui is the zero operator,
we usually drop of the ith branch of the statement (5). Furthermore, let V be a unitary
operator in the “coin” space Hc . Then we can define a quantum choice:
M
|ii → Ui [q]) = V [c]; qif [c] (i · |ii → Ui [q]) qif
(6)
V [c] (
i
The semantics of quantum case statement (5) is the unitary operator
△
U = (c, |ii → Ui )
in Hc ⊗ Hq defined by U |i, ψi = |iiUi |ψi for any 0 ≤ i < n and |ψi in Hq . The operator U is called the guarded composition of Ui ’s along with basis {|ii}. It is represented
by the diagonal matrix
U0
0
n−1
X
U1
(7)
(|iic hi| ⊗ Ui ) = diag(U0 , U1 , ..., Un−1 ) =
U=
...
i=0
0
Un−1
The semantics of quantum choice (6) is then the operator U (V ⊗ Iq ), where Iq is the
identity operator in Hq .
Quantum walks on a graph [1] can be conveniently expressed in terms of the above
generalised quantum case statement and choice, as shown in the following:
Example 2 A random walk on a directed graph G = (V, E) is described by repeated
applications of stochastic matrix P = (Puv )u,v∈V , where
(
1
if (u, v) ∈ E,
Puv = du
0
otherwise
where du is the outgoing degree of u, i.e. the number of edges outgoing from u. In
particular, if G is d−regular, i.e. all nodes have the same degree d, then Puv = d1 for
all u, v ∈ V . A quantum walk on graph G is a quantum counterpart of the random walk.
Let HV = span{|vi : v ∈ V } be the Hilbert space spanned by states corresponding
to the vertices in G. We now assume that G is d−regular. Then each edge in G can be
labelled by a number among 1, 2, ..., d so that for any 1 ≤ a ≤ d, the edges labelled a
6
Ying
form a permutation. Let HA = span{|1i, |2i, ..., |di} be an auxiliary Hilbert space of
dimension d, called the “coin space”. The shift operator S is defined in HA ⊗ HV by
S|a, vi = |a, va i
for 1 ≤ a ≤ d and v ∈ V , where va is the a−th neighbour of v, i.e. the vertex
reached from v through the outgoing edge labelled a. Furthermore, let C be a unitary
operator in HA , called the “coin-tossing operator”. Then one step of the quantum walk
is modelled by the operator W = S(C ⊗ I), where I is the identity operator in HV .
The quantum walk is described by repeated applications of W .
If for each 1 ≤ a ≤ d, we define the a−th shift operator Sa in HV by
Sa |vi = |va i
for any v ∈ V , then the shift operator S can be seen as a quantum case statement:
S = qif [c] (a · |ai → Sa [q]) qif
= qif [c] |1i → S1 [q]
|2i → S2 [q]
fiq
..........
|di → Sd [q]
where c and q are two variables denoting quantum systems with state spaces HA
and HV , respectively. Consequently, the single-step walk operator W is the quantum
choice:
M
W = C[c](
|ai → Sa [q])
a
The quantum case statement (5) and quantum choice (6) can be further generalised
to the case where unitary transformations U0 [q], U1 [q], ..., Un−1 [q] are replaced by general quantum programs that may contain quantum measurements. It is quite involved to
define the semantics of such general quantum case statement and choice; for details we
refer to [21], [22].
3 Syntax of Quantum Recursive Programs
A new notion of quantum recursion with quantum control flow can be defined based on
quantum case statement and quantum choice discussed in the last section. In this short
section, we formally define the syntax of quantum recursive programs.
We assume two sets of quantum variables: principal system variables, ranged over
by p, q, ..., and “coin” variables, ranged over by c, d, .... These two sets are required to
be disjoint. We also assume a set of procedure identifiers, ranged over by X, X1 , X2 , ....
Then program schemes are defined by the following syntax:
P ::= X | abort | skip | P1 ; P2 | U [c, q] | qif [c](i · |ii → Pi ) fiq
where:
Quantum Recursion and Second Quantisation
7
– X is a procedure identifier; programs abort and skip are the same as in a classical
programming language; P1 ; P2 is the sequential composition of P1 and P2 .
– In unitary transformation U [c, q], c a sequence of “coin” variables, q is a sequence
of principal system variables, and U is a unitary operator in the state Hilbert space
of the system consisting of c and q. We will always put “coin” variables before principal system variables. Both of c and q are allowed to be empty. When c is empty,
we simply write U [q] for U [c, q] and it describes the evolution of the principal system q; when q is empty, we simply U [c] for U [c, q] and it describes the evolution of
the “coins” c. If both c and q are not empty, then U [c, q] describes the interaction
between “coins” c the principal system q.
– In quantum case statement qif [c](i · |ii → Pi ) fiq, c is a “coin” variable, and
{|ii} is an orthonormal basis of the state Hilbert space of c. It is required not to
occur in all the subprograms Pi ’s because according to its physical interpretation,
a “coin” is always external to the principal system. This program construct is a
generalisation of equation (5).
As a generalisation of equation (6), the program construct of quantum choice can be
defined in terms of quantum case statement and sequential composition. Let P be a program contains only “coin” c, let {|ii} be an orthonormal basis of the state Hilbert space
of c, and let Pi be a program for each i . Then the quantum choice of Pi ’s according to
P along the basis {|ii} is defined as
[P (c)]
M
i
△
(|ii → Pi ) = P ; qif [c] (i · |ii → Pi ) fiq.
(8)
If the “coin” is a qubit, then quantum choice (8) can be abbreviated as
P0 ⊕P P1 .
Intuitively, quantum choice (8) runs a “coin-tossing” subprogram P followed by an
alternation of a family of subprograms P0 , P1 , .... The “coin-tossing” subprogram P
creates a superposition of the execution paths of P0 , P1 , ..., and during the execution of
the alternation, each Pi is running along its own path, but the whole program is executed
in a superposition of execution paths of P0 , P1 , .... This picture can be imaginatively
termed as the superposition-of-programs paradigm.
The semantics of quantum programs without procedure identifiers (and thus without
recursion) can be easily defined. The principal system of a quantum program P is the
composition of the systems denoted by principal variables appearing in P . We write H
for the state Hilbert space of the principal system.
Definition 1 The semantics JP K of a program P without procedure identifiers is inductively defined as follows:
1. If P = abort, then JP K = 0 (the zero operator in H), and if P = skip, then
JP K = I (the identity operator in H);
2. If P is an unitary transformation U [c, q], then JP K is the unitary operator U (in the
state Hilbert space of the system consisting of c and q);
3. If P = P1 ; P2 , then JP K = JP2 K · JP1 K;
8
Ying
4. If P = qif [c](i · |ii → Pi ) fiq, then
△
JP K = (c, |ii → JPi K) =
X
i
(|iic hi| ⊗ JPi K)
(see equation (7) for the special case of unitary operators).
Finally, we can define the syntax of quantum recursive programs. If a program
scheme contains at most the procedure identifiers X1 , ..., Xm , then we write P =
P [X1 , ..., Xm ].
Definition 2 1. Let X1 , ..., Xm be different procedure identifiers. A declaration for
X1 , ..., Xm is a system of equations:
X1 ⇐ P1 ,
D:
......
Xm ⇐ Pm ,
where for every 1 ≤ i ≤ m, Pi = Pi [X1 , ..., Xm ] is a program scheme containing
at most procedure identifiers X1 , ..., Xm .
2. A recursive program consists of a program scheme P = P [X1 , ..., Xm ], called
the main statement, and a declaration D for X1 , ..., Xm such that all “coin” variables in P do not appear in D; that is, they do not appear in the procedure bodies
P1 , ..., Pm .
The requirement in the above definition that the “coins” in the main statement P
and those in the declaration D are distinct is obviously necessary because a “coin” used
to define a quantum case statement is always considered to be external to its principal
system.
Now the question is: how to define the semantics of quantum recursive programs?
4 Motivating Example: Recursive Quantum Walks
As a motivating example of quantum recursive program, let us consider a variant of
quantum walks, called recursive quantum walks. For simplicity, we focus on the recursive Hadamard walk - a modification of Example 1. Recursive quantum walks on a
graph can be defined by modifying Example 2 in a similar way.
4.1 Specification of Recursive Quantum Walks
Recall that the single-step operator W of the Hadamard walk is a quantum choice,
which is the sequential composition of a “coin-tossing” Hadamard operator H on the
“direction coin” d and translation operator T on the position variable p. The translation
T [p] is a quantum case statement that selects left or right translations according to the
basis states |Li, |Ri of the “coin” d. If d is in state |Li then the walker moves one
position left, and if d is in state |Ri then it moves one position right. An essential
Quantum Recursion and Second Quantisation
9
difference between a random walk and a quantum walk is that the “coin” of the latter
can be in a superposition of the basis states |Li, |Ri, and thus a superposition of left and
right translations TL [p] and TR [p] is created. The Hadamard walk is then defined in a
simple way of recursion with the single-step operator W , namely repeated applications
of W . Now we modify slightly the Hadamard walk using a little bit more complicated
recursion.
Example 3 1. The unidirectionally recursive Hadamard walk first runs the “cointossing” Hadamard operator H[d] and then a quantum case statement: if the “direction coin” d is in state |Li then the walker moves one position left, and if d is
in state |Ri then it moves one position right, followed by a procedure behaving as
the recursive walk itself. In the terminology of programming languages, the recursive Hadamard walk is defined to a program X declared by the following recursive
equation:
X ⇐ TL [p] ⊕H[d] (TR [p]; X)
(9)
where d, p are the direction and position variables, respectively.
2. The bidirectionally recursive Hadamard walk first runs the “coin-tossing” Hadamard
operator H[d] and then a quantum case statement: if the “direction coin” d is in
state |Li then the walker moves one position left, followed by a procedure behaving as the recursive walk itself, and if d is in state |Ri then it moves one position
right, also followed by a procedure behaving as the recursive walk itself. More
precisely, the walk can be defined to be the program X declared by the following
two recursive equations:
X ⇐ (TL [p]; X) ⊕H[d] (TR [p]; X).
(10)
3. A variant of the bidirectionally recursive Hadamard walk is the program X (or Y )
declared by the following system of recursive equations:
(
X ⇐ TL [p] ⊕H[d] (TR [p]; Y ),
(11)
Y ⇐ (TL [p]; X) ⊕H[d] TR [p].
4. Note that we used the same “coin” d in the two equations of (11). If two different “coins” d and e are used, then we have another variant of the bidirectionally
recursive Hadamard walk specified by
(
X ⇐ TL [p] ⊕H[d] (TR [p]; Y ),
Y ⇐ (TL [p]; X) ⊕H[e] TR [p].
5. We can define a recursive quantum walk in another way if quantum case statement
with three branches is employed:
X ⇐ U [d]; qif [d] |Li → TL [p]
fiq
|Ri → TR [p]
|Ii → X
10
Ying
where d is not a qubit but a qutrit with state space Hd = span{|Li, |Ri, |Ii}, L, R
stand for the directions Left and Right, respectively, and I for Iteration, and U is a
3 × 3 unitary matrix, e.g. the 3−dimensional Fourier transform:
1 1
1
4
2
F3 = 1 e 3 πi e 3 πi .
4
2
1 e 3 πi e 3 πi
Now let us have a glimpse of the behaviours of recursive quantum walks. We use
E to denote the empty program or termination. A configuration is defined to be a pair
(P, |ψi) with P being a program or the empty program E, and |ψi a pure state of the
quantum system. Then the behaviour of a program can be visualised by a sequence of
transitions between superpositions of configurations. Here, we only consider the unidirectionally recursive quantum walk X declared by equation (9). Assume that it is
initialised in state |Lid |0ip ; that is, the “coin” is in direction L and the walker is at
position 0. Then we have:
1
1
(X, |Lid |0ip ) → √ (E, |Lid | − 1ip ) + √ (X, |Rid |1ip )
2
2
1
1
1
→ √ (E, |Lid | − 1ip ) + (E, |Rid |Lid1 |0ip ) + (X, |Rid |Rid1 |2ip )
2
2
2
→ ......
n
X
1
√
(E, |Rid0 ...|Ridi−1 |Lidi |i − 1ip )
→
2i+1
i=0
1
(X, |Rid0 ...|Ridn−1 |Ridn |n + 1ip )
+√
2n+1
(12)
Here, d0 = d, and new quantum “coins” d1 , d2 , ... that are identical to the original
“coin” d are introduced in order to avoid the conflict of variables for “coins”.
The above recursive quantum walks are good examples of quantum recursion, but
their behaviours are not very interesting. It has been well-understood that the major
difference between the behaviours of classical random walks and quantum walks is
caused by quantum interference - two separate paths leading to the same point may be
out of phase and cancel one another [2]. It is clear from equation (12) that quantum
interference does not happen in the unidirectionally recursive quantum walk. Similarly,
no quantum interference occurs in the bidirectionally recursive quantum walks defined
in the above example. The following is a much more interesting recursive quantum walk
that shows a new phenomenon of quantum interference: the paths that are cancelled in
a quantum walk are finite. However, it is possible that infinite paths are cancelled in a
recursive quantum walk.
Example 4 Let n ≥ 2. A variant of unidirectionally recursive quantum walk can be
defined as the program X declared by the following recursive equation:
X ⇐ (TL [p] ⊕H[d] TR [p])n ; ((TL [p]; X) ⊕H[d] (TR [p]; X))
(13)
Quantum Recursion and Second Quantisation
11
Here, we use P n to denote the sequential composition of n P ’s. Now let us look at the
behaviour of this walk. We assume that the walk is initialised in state |Lid |0ip . Then
the first three steps of the walk are given as follows:
1
(X, |Lid |0ip ) → √ [(X1 , |Lid | − 1ip + (X1 , |Rid |1ip )]
2
1
→ [(X2 , |Lid | − 2ip ) + (X2 , |Rid |0ip ) + (X2 , |Lid |0ip ) − (X2 , |Rid |2ip )]
2
1
→ √ [(X3 , |Lid | − 3ip ) + (X3 , |Rid | − 1ip ) + (X3 , |Lid | − 1ip ) − (X3 , |Rid |1ip )
2 2
+ (X3 , |Lid | − 1ip ) + (X3 , |Rid |1ip ) − (X3 , |Lid |1ip ) + (X3 , |Rid |3ip )]
1
= √ [(X3 , |Lid | − 3ip ) + (X3 , |Rid | − 1ip ) + 2(X3 , |Lid | − 1ip )
2 2
− (X3 , |Lid |1ip ) + (X3 , |Rid |3ip )]
(14)
where
Xi = (TL [p] ⊕H[d] TR [p])n−i ; ((TL [p]; X) ⊕H[d] (TR [p]; X))
for i = 1, 2, 3. We observe that in the last step of equation (14) two configurations
−(X, |Rid |1ip ) and (X, |Rid |1ip ) cancel one another in the last part of the above
equation. It is clear that both of them can generate infinite paths because they contain
the recursive walk X itself.
The behaviour of the recursive program specified by the following equation:
X ⇐ (TL [p] ⊕H[d] TR [p])n ; ((TL [p]; X) ⊕H[d] (TR [p]; X))
(15)
is even more puzzling. Note that equation (15) is obtained from equation (13) by changing the order of the two subprograms in its right-hand side.
4.2 How to solve recursive quantum equations?
We have already seen the first steps of the recursive quantum walks. But a precise
description of their behaviours amounts to solving recursive equations (9), (10), (11)
and (13). In mathematics, a standard method for finding the least solution to an equation
x = f (x) with f being a function from a lattice into itself is as follows: let x0 be the
least element of the lattice. We take the iterations of f starting from x0 :
(
x(0) = x0 ,
x(n+1) = f (x(n) ) for n ≥ 0.
If f is monotone and the lattice is complete, then the limit limn→∞ x(n) of iterations
exists; and furthermore if f is continuous, then this limit is the least solution of the
equation. In the theory of programming languages [4], a syntactic variant of this method
12
Ying
is employed to define the semantics of a recursive program declared by, say, equation
X ⇐ F (X), where F (·) is presented in a syntactic rather than semantic way: let
(
X (0) = Abort,
X (n+1) = F [X (n) /X] for n ≥ 0.
where F [X (n) /X] is the result of substitution of X in F (X) by X (n) . The program
X (n) is called the nth syntactic approximation of X. Roughly speaking, the syntactic
approximations X (n) (n = 0, 1, 2, ...) describe the initial fragments of the behaviour of
the recursive program X. Then the semantics JXK of X is defined to be the limit of the
semantics JX (n) K of its syntactic approximations X (n) :
JXK = lim JX (n) K.
n→∞
Now we apply this method to the unidirectionally recursive Hadamard walk and construct its syntactic approximations as follows:
X (0) = abort,
X (1) = TL [p] ⊕H[d] (TR [p]; abort),
X (2) = TL [p] ⊕H[d] (TR [p]; TL [p] ⊕H[d1 ] (TR [p]; abort)),
X (3) = TL [p] ⊕H[d] (TR [p]; TL [p] ⊕H[d1 ] (TR [p]; TL [p] ⊕H[d2 ] (TR [p]; abort))),
............
(16)
However, a problem arises in constructing these approximations: we have to continuously introduce new “coin” variables in order to avoid variable conflict; that is, for every
n = 1, 2, ..., we introduce a new “coin” variable dn in the (n + 1)th syntactic approximation. Obviously, variables d, d1 , d2 , ... must denote identical particles. Moreover, the
number of the “coin” particles that are needed in running the recursive Hadamard walk
is usually unknown beforehand because we do not know when the walk terminates. It
is clear that this problem appears only in the quantum case but not in the theory of
classical programming languages because it is caused by employing an external “coin”
system in defining a quantum case statement. Therefore, a solution to this problem requires a mathematical framework in which we can deal with quantum systems where
the number of particles of the same type - the “coins” - may vary.
5 Second Quantisation
Fortunately, physicists had developed a formalism for describing quantum systems with
variable particle number, namely second quantisation, more than eighty years ago. For
convenience of the reader, we recall basics of the second quantum method in this section.
Quantum Recursion and Second Quantisation
13
5.1 Fock Spaces
Let H be the state Hilbert space of one particle. For any n ≥ 1, we write H⊗n for
the n−fold tensor product of H. If we introduce the vacuum state |0i, then the 0−fold
tensor product of H can be defined as the one-dimensional space H⊗0 = span{|0i}.
Furthermore, the free Fock space over H is defined to be the direct sum [5]:
F (H) =
∞
M
n=0
H⊗n .
The principle of symmetrisation in quantum physics [8] indicates that the states
of n identical particles are either completely symmetric or completely antisymmetric
with respect to the permutations of the particles. These particles are called bosons in
the symmetric case and fermions in the antisymmetric case. For each permutation π of
1, ..., n, we define the permutation operator Pπ in H⊗n by
Pπ |ψ1 ⊗ ... ⊗ ψn i = |ψπ(1) ⊗ ... ⊗ ψπ(n) i
for all |ψ1 i, ..., |ψn i in H. Furthermore, we define the symmetrisation and antisymmetrisation operators in H⊗n as follows:
S+ =
1 X
Pπ ,
n! π
S− =
1 X
(−1)π Pπ
n! π
where π ranges over all permutations of 1, ..., n, and (−1)π is the signature of the
permutation π. For v = +, − and any |ψ1 i, ..., |ψn i in H, we write
|ψ1 , ..., ψn iv = Sv |ψ1 ⊗ ... ⊗ ψn i.
Then the state space of n bosons and that of fermions are
Hv⊗n = Sv H⊗n = span{|ψ1 , ..., ψn iv : |ψ1 i, ..., |ψn i are in H}
for v = +, −, respectively. If we set Hv⊗0 = H⊗0 , then the space of the states of
variable particle number is the symmetric or antisymmetric Fock space:
Fv (H) =
∞
M
n=0
Hv⊗n
where v = + for bosons and v = − for fermions. The elements of the Fock space
Fv (H) (resp. the free Fock space F (H)) are of the form
|Ψ i =
∞
X
n=0
|Ψ (n)i
with |Ψ (n)i ∈ Hv⊗n (resp. |Ψ (n)i ∈ H⊗n ) for n = 0, 1, 2, ... and
∞.
P∞
n=0 hΨ (n)|Ψ (n)i
<
14
Ying
5.2 Operators in the Fock Spaces
For each n ≥ 1, let A(n) be an operator in H⊗n . Then operator
A=
∞
X
A(n)
(17)
n=0
is defined in the free Fock space F (H) as follows:
A
∞
X
n=0
|Ψ (n)i =
∞
X
A(n)|Ψ (n)i
n=0
P∞
for any |Ψ i = n=0 |Ψ (n)i in F (H), where A(0) = 0; that is, the vacuum state is
considered to be an eigenvector of operator A with eigenvalue 0.
If for each n ≥ 0 and for each permutation π of 1, ..., n, Pπ and A(n) commute;
that is,
Pπ A(n) = A(n)Pπ ,
P∞
then operator A is said to be symmetric. A symmetric operator A = n=0 A(n) is an
operator both in the symmetric Fock space F+ (H) and in the antisymmetric Fock space
F− (H): A(Fv (H)) ⊆ Fv (H) for v = +,P−. We can introduce the symmetrisation
∞
functional S that maps every operator A = n=0 A(n) to a symmetric operator:
S(A) =
∞
X
S(A(n))
(18)
n=0
where for each n ≥ 0,
S(A(n)) =
1 X
Pπ A(n)Pπ−1
n! π
(19)
with π traversing over all permutations of 1, ..., n.
Observables in the Fock Spaces If for each n ≥ 1, the operator A(n) in equation
(17) is an observable of n particles, then A is an extensive observable in the free Fock
space F(H). In particular, let us consider one-body observables. Assume that A is a
single-particle observable. Then we can define one-body observable
X (n)
Aj
A(n) =
j
(n)
for the system of n particles, where Aj = I ⊗(n−1) ⊗ A ⊗ I ⊗(n−j) (with I being the
identity operator in H) is the action of A on the jth factor of the tensor product H⊗n ;
that is,
(n)
Aj |ψ1 ⊗ ... ⊗ ψn i = |ψ1 ⊗ ... ⊗ ψj−1 ⊗ Aψj ⊗ ψj+1 ⊗ ... ⊗ ψn i
Quantum Recursion and Second Quantisation
15
for all |ψ1 i, ...|ψn i in H. It is easy to see that A(n) commutes with the permutations:
A(n)|ψ1 , ..., ψn iv =
n
X
j=1
|ψ1 , ..., ψj−1 , Aψj , ψj+1 , ..., ψn iv .
P
Therefore, A = ∞
n=0 A(n) is symmetric. It is called a one-body observable in the
Fock space Fv (H) for v = +, −. Similarly, we can define a k−body observable with
k ≥ 2.
Evolutions in the Fock Spaces Let the (discrete-time) evolution of one particle is
represented by unitary operator U . Then the evolution of n particles without mutual
interactions can be described by operator U(n) = U ⊗n in H⊗n :
U(n)|ψ1 ⊗ ... ⊗ ψn i = |U ψ1 ⊗ ... ⊗ U ψn i
(20)
for all |ψ1 i, ..., |ψn i in H. It is easy to verify that U(n) commutes with the permutations:
U(n)|ψ1 , ..., ψn iv = |U ψ1 , ..., U ψn iv .
P
So, the symmetric operator U = ∞
n=0 U(n) depicts the evolution of particles without
mutual interactions in the Fock space Fv (H) for v = +, −.
Creation and Annihilation of Particles The operator U defined by equation (17)
maps states of n particles to states of particles of the same number. The transitions between states of different particle numbers are described by the creation and annihilation
operators. To each one-particle state |ψi in H, we associate the creation operator a† (ψ)
in Fv (H) defined by
√
a† (ψ)|ψ1 , ..., ψn iv = n + 1|ψ, ψ1 , ..., ψn iv
for any n ≥ 0 and all |ψ1 i, ..., |ψn i in H. This operator adds a particle in the individual
state |ψi to the system of n particles without modifying their respective states. The
annihilation operator a(ψ) is defined to be the Hermitian conjugate of a† (ψ), and it is
not difficult to show that
a(ψ)|0i = 0,
n
1 X i−1
(v) hψ|ψi i|ψ1 , ..., ψi−1 , ψi+1 , ..., ψn iv
a(ψ)|ψ1 , ..., ψn iv = √
n i=1
Intuitively, operator a(ψ) decreases the number of particles by one unit, while preserving the symmetry of the state.
6 Solving Recursive Equations in the Free Fock Space
Second quantisation provides us with the necessary tool for defining the semantics of
quantum recursions. We first show how to solve recursive equations in the free Fock
spaces without considering symmetry or antisymmetry of the particles that are used to
implement the quantum “coins”.
16
Ying
6.1 A Domain of Operators in the Free Fock Space
Let C be a set of quantum “coins”. For each c ∈ C, let Hc be the state Hilbert space of
“coin” c and F (Hc ) the free Fock space over Hc . We write
O
G(HC ) =
F (Hc ).
c∈C
We also assume that H is the state Hilbert space of the principal system. Let ω be the
set of nonnegative integers. Then ω C is the set of C−indexed tuples of nonnegative
integers: n = {nc }c∈C with nc ∈ ω for all c ∈ C, and we have:
!
M O
⊗nc
G(HC ) ⊗ H ≡
Hc ⊗ H .
n∈ω C
c∈C
We write O(G(HC ) ⊗ H) for the set of all operators of the form
X
A=
A(n),
n∈ω C
N
where A(n) is an operator in c∈C Hc⊗nc ⊗ H for each n ∈ ω C . Recall that a binary
relation ⊑ on a nonempty set S if it is reflexive, transitive and antisymmetric. In this
case, we often call (S, ⊑) or even simply S a partial order. We define a partial order ≤
on ω C as follows: n ≤ m if and only if nc ≤ mc for all c ∈ C. A subset Ω ⊆ ω C is
said to be below-closed if n ∈ Ω and m ≤ n imply
the flat order ⊑
P∞ m ∈ Ω. We defineP
∞
on O(G(HC ) ⊗ H) as follows: for any A = n∈ωC A(n) and B = n∈ωC B(n) in
O(G(HC ) ⊗ H),
– A ⊑ B if and only if there exists a below-closed subset Ω ⊆ ω C such that A(n) =
B(n) for all n ∈ Ω and A(n) = 0 for all n ∈ ω C \ Ω.
Let (S, ⊑) be a partial order. A nonempty subset T ⊆ S is called a chain if for any
x, y ∈ T , it holds that x ⊑ y or y ⊑ x. A partial order is said to be
F complete if it has
the least element and every chain T in it has the least upper bound T .
Lemma 1 (O(G(HC ) ⊗ H), ⊑) is a complete partial order (CPO).
Proof. First, ⊑ is reflexive because ω C itself is below-closed. To show that ⊑ is transitive, we assume that A ⊑ B and B ⊑ C. Then there exist below-closed Ω, Γ ⊆ ω C
such that
1. A(n) = B(n) for all n ∈ Ω and A(n) = 0 for all n ∈ ω C \ Ω;
2. B(n) = C(n) for all n ∈ Γ and B(n) = 0 for all n ∈ ω C \ Γ .
Clearly, Ω ∩ Γ is below-closed, and A(n) = B(n) = C(n) for all n ∈ Ω ∩ Γ . On the
other hand, if n ∈ ω C \ (Ω ∩ Γ ) = (ω C \ Ω) ∪ [Ω ∩ (ω C \ Γ )], then either n ∈ ω C \ Ω
and it follows from clause 1 that A(n) = 0, or n ∈ Ω ∩ (ω C \ Γ ) and by combining
clauses 1 and 2 we obtain A(n) = B(n) = 0. Therefore, A ⊑ C. Similarly, we can
prove that ⊑ is antisymmetric. So, (O(G(HC ) ⊗ H), ⊑) is a partial order.
Quantum Recursion and Second Quantisation
17
P
the operator A = n∈ωC A(n) with A(n) = 0 (the zero operator in
N Obviously,
⊗nc
⊗ H) for all n ∈ ω C is the least element of (O(G(HC ) ⊗ H), ⊑). Now it
c∈C Hc
suffices to show that any chain {Ai } in (O(G(HC ) ⊗ H), ⊑) has the least upper bound.
For each i, we put
∆i = {n ∈ ω C : Ai (n) = 0},
∆i ↓ = {m ∈ ω C : m ≤ n for some n ∈ ∆i }.
Here
P ∆i ↓ is the below-completion of ∆i . Furthermore, we define operator A =
m∈ω C A(n) as follows:
(
Ai (n) if n ∈ ∆i ↓ for some i,
S
A(n) =
0
if n ∈
/ i (∆i ↓).
Claim 1: A is well-defined; that is, if n ∈ ∆i ↓ and n ∈ ∆j ↓, then Ai (n) =
Aj (n). In fact, since {Ai } is a chain, we have Ai ⊑ Aj or Aj ⊑ Ai . We only
consider the case of Ai ⊑ Aj (the case of Aj ⊑ Ai is proved by duality). Then there
exists below-closed Ω ⊆ ω C such that Ai (n) = Aj (n) for all n ∈ Ω and A(n) = 0
for all n ∈ ω C \ Ω. It follows from n ∈ ∆i ↓ that n ⊑ m for some m with Ai (m) 6= 0.
/ ω C \ Ω, i.e. m ∈ Ω, we have n ∈ Ω because Ω is below-closed. So,
Since m ∈
Ai (n) = Aj (n). F
Claim 2: A = i Ai . In fact, for each i, ∆i ↓ is below-closed, and Ai (n) = A(n)
for all n ∈ ∆i ↓ and Ai (n) = 0 for all n ∈ ω C \(∆i ↓). So, Ai ⊑ A, and A is an upper
bound of {Ai }. Now assume that B is an upper bound of {Ai }: for all i, Ai ⊑ B; that
is, there exists below-closed Ωi ⊆ ω C such that Ai (n) = B(n) for all n ∈ Ωi and
Ai (n) = 0 for all n ∈ ω C \ Ωi . By theSdefinition of ∆i and below-closeness of Ωi ,
we know that ∆i ↓⊆ Ωi . We take Ω = i (∆i ↓). Clearly, Ω is below-closed, and if
n ∈ ω C \ Ω, then A(n) = 0. On the other hand, if n ∈ Ω, then for some i, we have
n ∈ ∆i ↓, and it follows that n ∈ Ωi and A(n) = Ai (n) = B(n). Therefore, A ⊑ B.
P
P
For any operators A = n∈ωC A(n) and B = n∈ωC B(n) in O(G(HC ) ⊗ H),
their product is naturally defined as
X
(A(n) · B(n)) ,
A·B=
(21)
n∈ω C
which is also in O(G(HC ) ⊗ H). We can define guarded composition of operators in
Fock spaces by extending
equation (7). Let c ∈ C and {|ii} be an orthonormal basis of
P
Hc , and let Ai = n∈ωC Ai (n) be an operator in O(G(HC ) ⊗ H) for each i. Then
the guarded composition of Ai ’s along with the basis {|ii} is
!
X X
(c, |ii → Ai ) =
(|iic hi| ⊗ Ai (n)) .
(22)
n∈ω C
Note that for each n ∈ ω C ,
P
i
(|iic hi| ⊗ Ai (n)) is an operator in
O
Hdnd ⊗ H,
Hc⊗(nc +1) ⊗
i
d∈C\{c}
18
Ying
and thus (c, |ii → Ai ) ∈ O(G(HC ) ⊗ H). Recall that a mapping f from a CPO
(S1 , ⊑) into another CPO (S2 , ⊑) is said to be continuous if for any chainFT in S1 ,
its F
image f (T ) = {f (x) : x ∈ T } under f has the least upper bound and f (T ) =
f ( T ). The following lemma shows that both product and guarded composition of
operators in the free Fock space are continuous.
Lemma 2 Let {Aj }, {Bj } and {Aij } for each i be chains in (O(G(HC ) ⊗ H), ⊑).
Then
F
F
F
.
·
B
A
1. j (Aj · Bj ) =
j
j
j
j
F
F
.
2. j (c, |ii → Aij ) = c, |ii →
j Aij
Proof. We only prove part 2. The proof of part 1 is similar. For each i, we assume that
X
G
Aij = Ai =
Ai (n).
j
n∈ω C
By the construction of least upper
P bound in (O(G(HC ) ⊗ H), ⊑) given in
Sthe proof of
Lemma 1, we can write Aij = n∈Ωij Ai (n) for some Ωij ⊆ ω C with j Ωij = ω C
for every i. By appending zero operators to the end of shorter summations, we may
further ensure that Ωij ’s for all i are the same, say Ωj . Then by the defining equation
(22) we obtain:
!
G X X
G
(|iic hi| ⊗ Ai (n))
(c, |ii → Aij ) =
j n∈Ωj
j
=
X
n∈ω C
i
X
i
!
(|iic hi| ⊗ Ai (n))
= (c, |ii → Ai ) .
6.2 Semantic Functionals of Program Schemes
Let P = P [X1 , ..., Xm ] be a program scheme. We write C for the set of “coins” occuring in P . For each c ∈ C, let Hc be the state Hilbert space of quantum “coin” c. As
said in Section 3, the principal system of P is the composition of the systems denoted
by principal variables appearing in P . Let H be the state Hilbert space of the principal
system.
Definition 3 The semantic functional of program scheme P is a mapping
JP K : O(G(HC ) ⊗ H)m → O(G(HC ) ⊗ H).
For any operators A1 , ..., Am ∈ O(G(HC ) ⊗ H), JP K(A1 , ..., Am ) is inductively defined as follows:
P
1. If P = abort, then JP K(A1 , ..., AmN
) is the zero operator in A = n∈ωC A(n)
with A(n) = 0 (the zero operator in c∈C Hc⊗nc ⊗ H) for all n ∈ ω C ;
Quantum Recursion and Second Quantisation
19
P
2. If P = skip, then JP K(A1 , ..., Am ) is the
A = n∈ωC A(n)
Nidentity⊗noperator
with A(n) = I (the identity operator in c∈C Hc c ⊗ H) for all n ∈ ω C with
nc 6= 0 for every c ∈ C;
3. P
If P = U [c, q], then JP K(A1 , ..., Am ) is the cylindrical extension of U : A =
n∈ω C A(n) with A(n) = I1 ⊗ I2 (n) ⊗ U ⊗ I3 , where:
(a) I1 is the identity operator in the state Hilbert space of those “coins” that are
not in c;
N
⊗(n −1)
(b) I2 (n) is the identity operator in c∈c Hc c ; and
(c) I3 is the identity operator in the state Hilbert space of those principal variables
that are not in q for all n ≥ 1 ;
4. If P = Xj (1 ≤ j ≤ m), then JP K(A1 , ..., Am ) = Aj ;
5. If P = P1 ; P2 , then
JP K(A1 , ..., Am ) = JP2 K(A1 , ..., Am ) · JP1 K(A1 , ..., Am )
(see the defining equation (21) of product of operators in the free Fock space);
6. If P = qif [c](i · |ii → Pi ) fiq, then
JP K(A1 , ..., Am ) = (c, |ii → JPi K(A1 , ..., Am ))
(see the defining equation (22) of guarded composition of operators in the free Fock
space).
Whenever m = 0; that is, P contains no procedure identifiers, then the above definition degenerates to Definition 1.
The cartesian power O(G(HC ) ⊗ H)m is naturally equipped with the order ⊑ defined componently from the order in O(G(HC )⊗H): for any A1 , ..., Am , B1 , ..., Bm ∈
O(G(HC ) ⊗ H),
– (A1 , ..., Am ) ⊑ (B1 , ..., Bm ) if and only if for every 1 ≤ i ≤ m, Ai ⊑ Bi .
Then (O(G(HC ) ⊗ Hq )m , ⊑) is a CPO too. Furthermore, we have:
Theorem 1 (Continuity of Semantic Functionals) The semantic functional JP K : (O(G(HC )⊗
H)m , ⊑) → (O(G(HC ) ⊗ H), ⊑) is continuous.
Proof. It can be easily proved by induction on the structure of P using Lemma 2.
For each “coin” c ∈ C, we introduce the creation functional:
Kc : O(G(HC ) ⊗
P
H) → O(G(HC ) ⊗ H) defined as follows: for any A = n∈ωC A(n) ∈ O(G(HC ) ⊗
H),
X
(Ic ⊗ A(n))
Kc (A) =
n∈ω C
N
where Ic is the identity operator in Hc . We observe that A(n) is an operator in d∈C
N
⊗(n +1)
Hd⊗nd ⊗ H, whereas Ic ⊗ A(n) is an operator in Hc c
⊗ d∈C\{d} Hd⊗nd ⊗ H.
Intuitively, the creation functional Kc moves all copies of Hc one position to the right
so that ith copy becomes (i + 1)th copy for all i = 0, 1, 2, .... Thus, a new position is
created at the left end for a new copy of Hc . For other “coins” d, Kc does not move
20
Ying
any copy of Hd . It is clear that for any two “coins” c, d, Kc and Kd commute; that
is, Ka ◦ Kd = Kd ◦ Kc . Note that the set C of “coins” in P is finite. Suppose that
C = {c1 , c2 , ..., ck }. Then we can define the creation functional
KC = Kc1 ◦ Kc2 ◦ ... ◦ Kck .
For the special case where the set C of “coins” is empty, C is the identity functional;
that is, C(A) = A for all A.
Lemma 3 For each c ∈ C, the creation functionals Kc and KC : (O(G(HC ) ⊗ H),
⊑) → (O(G(HC ) ⊗ H), ⊑) are continuous.
Proof. Straightforward by definition.
Combining continuity of semantic functional and the creation functional (Theorem
1 and Lemma 3), we obtain:
Corollary 1 Let P = P [X1 , ..., Xm ] be a program scheme and C the set of “coins”
occurring in P . Then the functional:
m
Km
C ◦ JP K : (O(G(HC ) ⊗ H) , ⊑) → (O(G(HC ) ⊗ H), ⊑),
m
(KC ◦ JP K)(A1 , ..., Am ) = JP K(KC (A1 ), ..., KC (Am ))
for any A1 , ..., Am ∈ O(G(HC ) ⊗ H), is continuous.
6.3 Fixed Point Semantics
Now we are ready to define the denotational semantics of recursive programs using the
standard fixed point technique. Let us consider a recursive program P declared by the
system of recursive equations:
X1 ⇐ P1 ,
D:
(23)
......
Xm ⇐ Pm ,
where Pi = Pi [X1 , ..., Xm ] is a program scheme containing at most procedure identifiers X1 , ..., Xm for every 1 ≤ i ≤ m. The system D of recursive equations naturally
induces a semantic functional:
JDK : O(G(HC ) ⊗ H)m → O(G(HC ) ⊗ H)m ,
JDK(A1 , ..., Am ) = ((Km
C ◦ JP1 K)(A1 , ..., Am ), ...,
(Km
C ◦ JPm K)(A1 , ..., Am ))
(24)
for all A1 , ..., Am ∈ O(G(HC ) ⊗ H), where C is the set of “coins” appearing in D;
that is, in one of P1 , ..., Pm . It follows from Theorem 4.20 in [14] and Corollary 1
that JDK : (O(G(HC ) ⊗ H)m , ⊑) → (O(G(HC ) ⊗ H)m , ⊑) is continuous. Then the
Knaster-Tarski Fixed Point Theorem asserts that JDK has the least fixed point µJDK.
Quantum Recursion and Second Quantisation
21
Definition 4 The fixed point (denotational) semantics of the recursive program P declared by D is
JP Kf ix = JP K(µJDK);
that is, if µJDK = (A∗1 , ..., A∗m ) ∈ O(G(HC )⊗ H)m , then JP Kf ix = JP K(A∗1 , ..., A∗m )
(see Definition 3).
6.4 Syntactic Approximation
We now turn to consider the syntactic approximation technique for defining the semantics of recursive programs. As discussed at the end of Section 4 and further clarified
in Example 5, a problem that was not present in the classical case is that we have to
carefully avoid the conflict of quantum “coin” variables when defining the notion of
substitution. To overcome it, we assume that each “coin” variable c ∈ C has infinitely
many copies c0 , c1 , c2 , ... with c0 = c. The variables c1 , c2 , ... are used to represent a
sequence of particles that are all identical to the particle c0 = c. Then the notion of
program scheme defined in Section 3 will be used in a slightly broader way: a program
scheme may contain not only a “coin” c but also some of its copies c1 , c2 , .... If such a
generalised program scheme contains no procedure identifiers, then it is called a generalised program. With these assumptions, we can introduce the notion of substitution.
Definition 5 Let P = P [X1 , ..., Xm ] be a generalised program scheme that contains
at most procedure identifiers X1 , ..., Xm , and let Q1 , ..., Qm be generalised programs
without any procedure identifier. Then the simultaneous substitution P [Q1 /X1 , ..., Qm /
Xm ] of X1 , ..., Xm by Q1 , ..., Qm in P is inductively defined as follows:
1. If P = abort, skip or an unitary transformation, then P [Q1 /X1 , ..., Qm /Xm ] =
P;
2. If P = Xi (1 ≤ i ≤ m), then P [Q1 /X1 , ..., Qm /Xm ] = Qi ;
3. If P = P1 ; P2 , then
P [Q1 /X1 , ..., Qm /Xm ] = P1 [Q1 /X1 , ..., Qm /Xm ]; P2 [Q1 /X1 , ..., Qm /Xm ].
4. If P = qif [c](i · |ii → Pi ) fiq, then
P [Q1 /X1 , ..., Qm /Xm ] = qif [c](i · |ii → Pi′ ) fiq
where for every i, Pi′ is obtained through replacing the jth copy cj of c in Pi [Q1 /X1 ,
..., Qm /Xm ] by the (j + 1)th copy cj+1 of c for all j.
Note that in Clause 4 of the above definition, since P is a generalised program
scheme, the “coin” c may not be an original “coin” but some copy dk of an original
“coin” d ∈ C. In this case, the jth copy of c is actually the (k + j)th copy of d:
cj = (dk )j = dk+j for j ≥ −d.
The semantics of a generalised program P can be given using Definition 1 in the
way where a “coin” c and its copies c1 , c2 , ... are treated as distinct variables to each
other. For each “coin” c, let nc be the greatest index
N n such that the copy cn appears in
P . Then the semantics JP K of P is an operator in c∈C Hc⊗nc ⊗ H. Furthermore, it can
22
Ying
P
be identified with its cylindrical extension in O(G(HC ) ⊗ H): N m∈ωC (I(m) ⊗ JP K) ,
where for each m ∈ ω C , I(m) is the identity operator in c∈C Hc⊗mc . Based on
this observation, the semantics of substitution defined above is characterised by the
following:
Lemma 4 For any (generalised) program scheme P = P [X1 , ..., Xm ] and (generalised) programs Q1 , ..., Qm , we have:
JP [Q1 /X1 , ..., Qm /Xm ]K = (Km
C ◦ JP K)(JQ1 K, ..., JQm K)
= JP K(KC (JQ1 K), ..., KC (JQm K)),
where KC is the creation functional with C being the set of “coins” in P .
Proof. We prove the lemma by induction on the structure of P .
Case 1. P = abort, skip or an unitary transformation. Obvious.
Case 2. P = Xj (1 ≤ j ≤ m). Then P [Q1 /X1 , ..., Qm /Xm ] = Qm . On the other
hand, since the set of “coins” in P is empty, KC (JQi K) = JQi K for all 1 ≤ i ≤ m.
Thus, by clause 4 of Definition 3 we obtain:
JP [Q1 /X1 ,..., Qm /Xm ]K = JQm K
= JP K(JQ1 K, ..., JQm K) = JP K(KC (JQ1 K), ..., KC (JQm K)).
Case 3. P = P1 ; P2 . Then by clause 3 of Definition 1, clause 5 of Definition 3 and
the induction hypothesis, we have:
JP [Q1 /X1 ,..., Qm /Xm ]K = JP1 [Q1 /X1 , ..., Qm /Xm ]; P2 [Q1 /X1 , ..., Qm /Xm ]K
= JP2 [Q1 /X1 , ..., Qm /Xm ]K · JP1 [Q1 /X1 , ..., Qm /Xm ]K
= JP2 K(KC (JQ1 K), ..., KC (JQm K)) · JP1 K(KC (JQ1 K), ..., KC (JQm K))
= JP1 ; P2 K(KC (JQ1 K), ..., KC (JQm K))
= JP K(KC (JQ1 K), ..., KC (JQm K)).
Case 4. P = qif [c](i · |ii → Pi ) fiq. Then
P [Q1 /X1 , ..., Qm /Xm ] = qif [c](i · |ii → Pi′ ) fiq,
where Pi′ is obtained according to clause 4 of Definition 5. For each i, by the induction
hypothesis we obtain:
JPi [Q1 /X1 , ..., Qm /Xm ]K = JPi K(KC\{c} (JQ1 K), ..., KC\{c} (JQm K))
because the “coin” c does not appear in Pi′ . Furthermore, it follows that
JPi′ K = Kc (JPi [Q1 /X1 , ..., Qm /Xm ]K)
= Kc (JPi K(KC\{c} (JQ1 K), ..., KC\{c} (JQm K)))
= JPi K((Kc ◦ KC\{c} )(JQ1 K), ..., (Kc ◦ KC\{c} )(JQm K))
= JPi K(KC (JQ1 K), ..., KC (JQm K)).
Quantum Recursion and Second Quantisation
23
Therefore, by clause 4 of Definition 1, clause 6 of Definition 3 and equation (22), we
have:
X
JP [Q1 /X1 , ..., Qm /Xm ]K =
(|iihi| ⊗ JPi′ K)
i
= (c, |ii → JPi K(KC (JQ1 K), ..., KC (JQm K))
= JP K(KC (JQ1 K), ..., KC (JQm K)).
The notion of syntactic approximation can be defined based on Definition 5.
Definition 6 1. Let X1 , ..., Xm be procedure identifiers declared by the system D of
recursive equations (23). Then for each 1 ≤ k ≤ m, the nth syntactic approxima(n)
tion Xk of Xk is inductively defined as follows:
(
(0)
Xk = abort,
(n+1)
(n)
(n)
Xk
= Pk [X1 /X1 , ..., Xm /Xm ] for n ≥ 0.
2. Let P = P [X1 , ..., Xm ] be a recursive program declared by the system D of equations (23). Then for each n ≥ 0, its nth syntactic approximation P (n) is inductively
defined as follows:
(
P (0) = abort,
(n)
(n)
P (n+1) = P [X1 /X1 , ..., Xm /Xm ] for n ≥ 0.
Syntactic approximation actually gives an operational semantics of quantum recursive programs. As in the theory of classical programming, substitution represents an
application of the so-called copy rule - at runtime a procedure call is treated like the
procedure body inserted at the place of call (see, for example, [13]). Of course, sim(n)
plification may happen within Xk by operations of linear operators; for example,
C[q1 , q2 ]; X[q2 ]; C[q1 , q2 ] can be replaced by X[q2 ], where q1 , q2 are principal system
variables, C is the CNOT gate and X is the NOT gate. To simplify the presentation, we
choose not to explicitly describe simplification.
The major difference between the classical case and the quantum case is that in
the latter we need to continuously introduce new “coin” variables to avoid variable
conflict when we unfold a quantum recursive program using its syntactic approximations: for each n ≥ 0, a new copy of each “coin” in Pk is created in the substitu(n+1)
(n)
(n)
tion Xk
= P [X1 /X1 , ..., Xm /Xm ] (see Clause 4 of Definition 5). Thus, a
quantum recursive program should be understood as a quantum system with variable
particle number and described in the second quantisation formalism. Note that for all
(n)
1 ≤ k ≤ m and n ≥ 0, the syntactic approximation Xk is a generalised program con(n)
taining no procedure identifiers. Thus, its semantics JXk K can be given by a slightly
extended version of Definition 1: a “coin”c and its copies c1 , c2 , ... are allowed to appear in the same (generalised) program and they are considered as distinct variables.
As before, the principal system is the composite system of the subsystems denoted by
principal variables appearing in P1 , ..., Pm and its state Hilbert space is denoted by H.
24
Ying
Assume that C is the set of “coin” variables appearing in P1 , ..., Pm . For each c ∈ C,
we write Hc for the state Hilbert space of quantum “coin” c. Then it is easy to see
Ln
N
(n)
⊗n
that JXk K is an operator in j=0 HC
⊗ H , where HC = c∈C Hc . So, we can
(n)
imagine that JXk K ∈ O(G(HC ) ⊗ H). Furthermore, we have:
(n)
Lemma 5 For each 1 ≤ k ≤ m, {JXk K}∞
n=0 is an increasing chain and thus
(∞)
JXk
(n)
△
K = lim JXk K =
n→∞
∞
G
(n)
JXk K
(25)
n=0
exists in (O(G(HC ) ⊗ H), ⊑).
(n)
(n+1)
Proof. We show that JXk K ⊑ JXk
K by induction on n. The case of n = 0 is
(0)
trivial because JXk K = JabortK = 0. In general, by the induction hypothesis on
n − 1 and Corollary 1, we have:
(n)
(n−1)
JXk K = JPk K(KC (JX1
(n−1)
K))
K), ..., KC (JXm
(n)
(n+1)
(n)
⊑ JPk K(KC (JX1 K), ..., KC (JXm
K)) = JXk
K,
where C is the set of “coins” in D. Then existence of the least upper bound (25) follows
immediately from Lemma 1.
Definition 7 Let P be a recursive program declared by the system D of equations (23).
Then its operational semantics is
(∞)
JP Kop = JP K(JX1
(∞)
K, ..., JXm
K).
The operational semantics of recursive program P can be characterised by the limit
of its syntactic approximations (with respect to its declaration D).
Proposition 1 JP Kop =
F∞
n=0 JP
(n)
K.
Proof. It follows from Lemma 4 that
∞
G
JP (n) K =
n=0
=
=
∞
G
JP (n) K
n=0
∞
G
(n)
(n)
JP [X1 /X1 , ..., Xm
/Xm ]K
n=0
∞
G
(n)
(n)
JP K(KC (JX1 K), ..., KC (JXm
K))
n=0
where KC is the creation functional with respect to the “coins” C in P . However, all
(n)
(n)
the “coins” C in P do not appear in X1 , ..., Xm (see the condition in Definition 2.2).
Quantum Recursion and Second Quantisation
(n)
25
(n)
So, KC (JXk K) = JXk K for every 1 ≤ k ≤ m, and by Theorem 1 we obtain:
∞
G
JP (n) K =
n=0
∞
G
(n)
(n)
JP K(JX1 K, ..., JXm
K)
n=0
∞
G
= JP K
(n)
JX1 K, ...,
n=0
∞
G
!
(n)
JXm
K
n=0
∞
= JP K(JX1∞ K, ..., JXm
K) = JP Kop .
The equivalence between denotational and operational semantics of recursive programs is established in the following:
Theorem 2 (Equivalence of Denotational Semantics and Operational Semantics) For
any recursive program P , we have JP Kf ix = JP Kop .
(∞)
(∞)
Proof. By Definitions 4 and 7, it suffices to show that (JX1 K, ..., JXm K) is the
least fixed point of semantic functional JDK, where D is the declaration of procedure
identifiers in P . With Theorem 1 and Lemmas 3 and 4, we obtain:
(∞)
JXk K
=
∞
G
(n)
JXk K
n=0
=
=
∞
G
n=0
∞
G
(n)
(n)
/Xm ]K
JPk [X1 /X1 , ..., Xm
(n)
(n)
JPk K(KC (JX1 K), ..., KC (JXm
K))
n=0
=
=
∞
G
∞
G
(n)
(n)
JPk K KC
JX1 K , ..., KC
JXm
K
n=0
n=0
(∞)
(∞)
JPk K(KC (JX1 K), ..., KC (JXm
K))
!
!!
(∞)
(∞)
for every 1 ≤ k ≤ m, where C is the set of “coins” in D. So, (JX1 K, ..., JXm K)
is a fixed point of JDK. On the other hand, if (A1 , ..., Am ) ∈ O(G(HC ) ⊗ H)m is a
(n)
(n)
fixed point of JDK, then we can prove that for every n ≥ 0, (JX1 K, ..., JXm K) ⊑
(A1 , ..., Am ) by induction on n. Indeed, the case of n = 0 is obvious. In general, using
the induction hypothesis on n − 1, Corollary 1 and Lemma 4 we obtain:
(A1 , ..., Am ) = JDK(A1 , ..., Am )
m
= ((Km
C ◦ JP1 K)(A1 , ..., Am ), ..., (KC ◦ JPm K)(A1 , ..., Am ))
(n−1)
⊒ ((Km
C ◦ JP1 K)(JX1
(n)
(n)
= (JX1 K, ..., JXm
K).
(n−1)
(n−1)
K, ..., JXm
K), ..., (Km
C ◦ JPm K)(JX1
(n−1)
K, ..., JXm
K))
Therefore, it holds that
(∞)
(∞)
(JX1 K, ..., JXm
K)
=
∞
G
(n)
(n)
(JX1 K, ..., JXm
K) ⊑ (A1 , ..., Am ),
n=0
(∞)
and (JX1
(∞)
K, ..., JXm K) is the least fixed point of JDK.
26
Ying
In light of this theorem, we will simply write JP K for both the denotational (fixed
point) and operational semantics of a recursive program P . But we should carefully distinguish the semantics JP K ∈ O(G(HC )⊗H) of a recursive program P = P [X1 , ..., Xm ]
declared by a system of recursive equations about X1 , ..., Xm from the semantic functional JP K : O(G(HC )⊗H)m → O(G(HC )⊗H) of program scheme P = P [X1 , ..., Xm ].
Usually, such a difference can be recognised from the context.
6.5 Examples
Now let us reconsider the recursive quantum walks defined in Section 4.
Example 5 (Unidirectionally recursive Hadamard walk) The semantics of the nth approximation of the unidirectionally recursive Hadamard walk specified by equation (9)
is
i−1
n−1
O
X
(26)
|Ridj hR| ⊗ |Lidi hL| H(i) ⊗ TL TRi
JX (n) K =
i=0
j=0
where d0 = d, H(i) is the operator in Hd⊗i defined from the Hadamard operator H
by equation (20). This can be easily shown by induction on n, starting from the first
three approximations displayed in equation (16). Therefore, the semantics of the unidirectionally recursive Hadamard walk is the operator:
JXK = lim JX (n) K
n→∞
∞
i−1
X
O
|Ridj hR| ⊗ |Lidi hL| H(i) ⊗ TL TRi
=
i=0
j=0
(27)
∞
i−1
X
O
=
|Ridj hR| ⊗ |Lidi hL| ⊗ TL TRi (H ⊗ I)
i=0
j=0
in F (Hd ) ⊗ Hp , where Hd = span{L, R}, Hp = span{|ni : n ∈ Z}, I is the
identity
P operator in the position Hilbert space Hp , H(i) is as in equation (26), and
H= ∞
i=0 H(i) is the extension of H in the free Fock space F (Hd ) over the direction
Hilbert space Hd .
Example 6 (Bidirectionally recursive Hadamard walk) Let us consider the semantics
of the bidirectionally recursive Hadamard walk declared by equation (11). For any
string Σ = σ0 σ1 ...σn−1 of L and R, its dual is defined to be Σ = σ0 σ1 ...σn−1 , where
L = R and R = L. Moreover, we write |Σi = |σ0 id0 ⊗ |σ1 id1 ⊗ ... ⊗ |σn−1 idn−1 ,
TΣ = Tσn−1 ...Tσ1 Tσ0 and
ρΣ = |ΣihΣ| =
n−1
O
j=0
|σj idj hσj |.
Quantum Recursion and Second Quantisation
Then the semantics of procedures X and Y are
"∞
#
X
JXK =
(ρΣn ⊗ Tn ) (H ⊗ Ip ) ,
JY K =
n=0
∞
X
"
n=0
#
′
ρΣ n ⊗ T n
where H is as in Example 5, and
(
27
(28)
(H ⊗ Ip ) ,
(RL)k L
if n = 2k + 1,
k
(RL) RR if n = 2k + 2,
(
TL if n is odd,
Tn = TΣn =
TR2 if n is even,
(
TR if n is odd,
′
Tn = TΣn =
TL2 if n is even.
Σn =
It is clear from equations (27) and (28) that the behaviours of unidirectionally and
bidirectionally recursive Hadamard walks are very different: the unidirectionally one
can go to any one of the positions −1, 0, 1, 2, ..., but the bidirectionally walk X can
only go to the positions −1 and 2, and Y can only go to the positions 1 and −2.
7 Recovering Symmetry and Antisymmetry
The solutions of recursive equations found in the free Fock space using the techniques
developed in the last section cannot directly apply to the symmetric Fock space for
bosons or the antisymmetric Fock space for fermions because they may not preserve
symmetry. In this section, we consider symmetrisation of these solutions of recursive
equations.
7.1 Symmetrisation Functional
We first examine the domain of symmetric operators in the Fock spaces used in defining
semantics of recursive programs. As in Subsection 6.1, let H be the state Hilbert space
of the principal system and C the set of “coins”, and
!
M O
F (Hc ) ⊗ H ,
G(HC ) ⊗ H =
n∈ω C
c∈C
where ω is the set of nonnegative integers, and for each c ∈ C, F (Hc )P
is the free Fock
space over the state Hilbert space Hc of “coin” c. For any operator A = n∈ωC A(n) ∈
O(G(HC ) ⊗ H), we say that A is symmetric if for each n ∈ ω c , for each c ∈ C and
for each permutation π of 0, 1, ..., nc − 1, Pπ and A(n) commute; that is,
Pπ A(n) = A(n)Pπ .
28
Ying
(Note that in the above
Pπ actually stands for its cylindrical extension Pπ ⊗
N equation
N
⊗nd
H
⊗ H, where Id is the identity operator in Hd for every
I
⊗
I
in
d∈C
d
d∈C\{c} d
d ∈ C \ {c}, and I is the identity operator in H.) We write SO(G(HC ) ⊗ H) for the
set of symmetric operators A ∈ O(G(HC ) ⊗ H).
Lemma 6 (SO(G(HC ) ⊗ H), ⊑) as a sub-partial order of (O(G(HC ) ⊗ H), ⊑) is
complete.
Proof. It suffices to observe that symmetry of operators is preserved
F by the least upper
bound in (O(G(HC ) ⊗ H), ⊑) ; that is, if Ai is symmetric, so is i Ai , as constructed
in the proof of Lemma 1.
Now we generalise the symmetrisation functional defined by equations (18) and
(19) into the space G(HC ) N
⊗ H). For each n ∈ ω C , the symmetrisation functional S
over operators in the space c∈C Hc⊗nc ⊗ H is defined by
Y 1 X
·
S(A) =
nc !
c∈C
{πc }
"
Y
c∈C
Pπc
!
A
Y
Pπ−1
c
c∈C
!#
N
for every operator A in c∈C Hc⊗nc ⊗ H, where {πc } traverses over all C−indexed
families with πc being a permutation of 0, 1, ..., nc − 1 for every c ∈ C. This symmetrisation functional can be extended to O(G(HC ) ⊗ H) in a natural way:
S(A) =
X
S(A(n))
n∈ω C
P
for any A = n∈ωC A(n) ∈ O(G(HC ) ⊗ H). Obviously, S(A) ∈ SO(G(HC ) ⊗ H).
Furthermore, we have:
Lemma 7 The symmetrisation functional S : (O(G(HC ) ⊗ H), ⊑) → (SO(G(HC ) ⊗
H), ⊑) is continuous.
F
F
Proof. What we need to prove is that SF
( i Ai ) = i S(Ai ) for any chain {Ai } in
(O(G(HC )P
⊗ H), ⊑). Assume that A
P= i Ai . Then by the proof of Lemma 1, we can
write A = n∈ω A(n) and Ai = n∈Ωi A(n) for some Ωi with supi Ωi = ω C . So,
it holds that
G X
G
X
S(Ai ) =
S(A(n)) =
S(A(n)) = S(A).
i
i n∈Ωi
n∈ω C
7.2 Symmetrisation of the Semantics of Recursive Programs
Now we are ready to present the semantics of recursive programs in the symmetric or
antisymmetric Fock space.
Quantum Recursion and Second Quantisation
29
Definition 8 Let P = P [X1 , ..., Xm ] be a recursive program declared by the system D
of recursive equations (23). Then its symmetric semantics JP Ksym is the symmetrisation
of its semantics JP K in the free Fock space:
JP Ksym = S(JP K)
where JP K = JP Kf ix = JP Kop ∈ O(G(HC ) ⊗ K) (see Theorem 2), C is the set of
“coins” in D, and H is the state Hilbert space of the principal system of D.
As a symmetrisation of Proposition 1, we have:
F∞
Proposition 2 JP Ksym = n=0 S(JP (n) K).
Proof. It follows from Proposition 1 and Lemma 7 (continuity of the symmetrisation
functional) that
!
∞
∞
G
G
(n)
JP Ksym = S(JP K) = S
JP K =
S(JP (n) K).
n=0
n=0
Again, let us consider the examples of recursive Hadamard walks.
Example 7 (Continuation of Example 5) For each i ≥ 0, we have:
i−1
i−1
O
X
O
1
S
|Ridj hR| ⊗ |Lidi hL| =
|Ridj hR| ⊗ |Lidi hL| Pπ−1
Pπ
(i
+
1)!
π
j=0
j=0
(where π traverses over all permutations of 0, 1, ..., i)
=
△
1
i+1
i
X
j=0
|Rid0 hR| ⊗ ... ⊗ |Ridj−1 hR| ⊗ |Lidj hL| ⊗ |Ridj+1 hR| ⊗ ... ⊗ |Ridi hR|
= Gi .
Therefore, the symmetric semantics of the unidirectionally recursive Hadamard walk
defined by equation (9) is
!
∞
X
Gi ⊗ TL TRi (H ⊗ I).
S(JXK) =
i=0
Example 8 (Continuation of Example 6) The symmetric semantics of the bidirectionally recursive Hadamard walk specified by equaltion (11) is:
"∞
#
X
JXK =
(γn ⊗ Tn ) (H ⊗ Ip ) ,
JY K =
"
n=0
∞
X
#
(δn ⊗ Tn ) (H ⊗ Ip )
n=0
30
Ying
where:
γ2k+1 =
X
1
ρΓ ,
k
Γ
2k + 1
δ2k+1 =
X
1
ρ∆
k
∆
2k + 1
with Γ ranging over all strings of (k + 1) L’s and k R’s and ∆ ranging over all strings
of k L’s and (k + 1) R’s, and
γ2k+2 =
X
1
ρΓ ,
k
Γ
2k + 2
σ2k+2 =
X
1
ρ∆
k
∆
2k + 2
with Γ ranging over all strings of k L’s and (k + 2) R’s and ∆ ranging over all strings
of (k + 2) L’s and k R’s.
7.3 Principal System Semantics of Quantum Recursion
Let P be a recursive program with H being the state Hilbert space of its principal
variables and C being the set of its “coins”. We consder the computation
of P with input
N
|ψi ∈ H. Assume that the “coins” are initialised in state |Ψ i ∈ c∈C Fvc (Hc ), where
for each c ∈ C, Hc is the state Hilbert space of “coin” c, Fvc (Hc ) is the symmetric or
antisymmetric Fock space over Hc , and vc = + or − if “coin” c is implemented by a
boson or a fermion, respectively. Then the computation of the program starts in state
|Ψ i|ψi. What actually concerns us is the output in the principal system. This observation
leads to the following:
N
Definition 9 Given a state |Ψ i ∈ c∈C Fvc (Hc ). The principal system semantics of
program P with respect to “coin” initialisation |Ψ i is the mapping JP, Ψ K from pure
states in H to partial density operators [18], i.e. positive operators with trace ≤ 1, in
H:
JP, Ψ K(|ψi) = trNc∈C Fvc (Hc ) (|ΦihΦ|)
for each pure state |ψi in H, where
|Φi = JP Ksym (|Ψ i ⊗ |ψi),
JP K
is the symmetric semantics of P , and trNc∈C Fvc (Hc ) is the partial trace over
N sym
c∈C Fvc (Hc ) (see [16], Section 2.4.3).
Example 9 (Continuation of Example 6) We consider the bidirectionally recursive Hadamard
walk declared by equation (11) once again and suppose that it starts from the position
0.
1. If the “coins” are bosons initialised in state
|Ψ i = |L, L, ..., Li+ = |Lid0 ⊗ |Lid1 ⊗ ... ⊗ |Lidn−1 ,
Quantum Recursion and Second Quantisation
31
then we have
P
1
Γ |Γ i ⊗ | − 1i if n = 2k + 1,
k
√
n
2 2k + 1
P
JXKsym (|Ψ i ⊗ |0i) =
1
if n = 2k + 2,
∆ |∆i ⊗ |2i
√
k
2n
2k + 2
where Γ traverses over all strings of (k + 1) L’s and k R’s, and ∆ traverses over
all strings of k L’s and (k + 2) R’s. Therefore, the principal system semantics with
the “coin” initialisation |Ψ i is:
JX, Ψ K(|0i) =
(
1
2n | − 1ih−1|
1
2n |2ih2|
if n is odd,
if n is even.
2. Recall from [8] that for each single-particle state |ψi in Hd , the corresponding
coherent state of bosons in the symmetric Fock space F+ (Hd ) over Hd is defined
as
X
∞
1
[a† (ψ)]n
|ψicoh = exp − hψ|ψi
|0i
2
n!
n=0
where |0i is the vacuum state and a† (·) the creation operator. If the “coins” are
initialised in the coherent state |Licoh of bosons corresponding to |Li, then we
have:
∞
X
1
1 X
|Γk i
JXKsym (|Licoh ⊗ |0i) = √
⊗ | − 1i
√
e
k
Γk
k=0
22k+1
2k + 1
∞
1 X
1
+√
√
e
k=0
22k+2
k
2k + 2
X
∆k
|∆k i
⊗ |2i,
where Γk ranges over all strings of (k + 1) L’s and k R’s, and ∆k ranges over all
strings of k L’s and (k + 2) R’s. So, the principal system semantics with “coin”
initialisation |Licoh is:
1
JX, Lcoh K(|0i) = √
e
1
= √
e
∞
X
k=0
1
22k+1
| − 1ih−1| +
∞
X
k=0
2
1
| − 1ih−1| + |2ih2| .
3
3
1
22k+2
|2ih2|
!
32
Ying
8 Quantum Loop
In this section, we consider a special class of quantum recursions. Arguably, while-loop
is the simplest and most popular form of recursion used in programming languages. In
classical programming, the while-loop
while b do S od
can be seen as the program X declared by the recursive equation:
X ⇐ if b then X else skip fi
(29)
We can define a kind of quantum while-loop by using quantum case statement and
quantum choice in the place of classical case statement if ...then...else fi in equation
(29).
Example 10 (Quantum while-loop)
1. The first form of quantum while-loop:
qwhile [c] = |1i do U [q] od
(30)
is defined to be the recursive program X declared by
X ⇐ qif [c] |0i → skip
|1i → U [q]; X
(31)
fiq
where c is a quantum “coin” variable denoting a qubit, q is a principal quantum
variable, and U is a unitary operator in the state Hilbert space Hq of system q.
2. The second form of quantum while-loop
qwhile V [c] = |1i do U [q] od
(32)
is defined to be the recursive program X declared by
X ⇐ skip ⊕V [c] (U [q]; X)
≡ V [c]; qif [c] |0i → skip
fiq
|1i → U [q]; X
(33)
Note that the recursive equation (33) is obtained by replacing the quantum case
statement qif ...fiq in equation (31) by the quantum choice ⊕V [c] .
3. Actually, quantum loops (30) and (32) are not very interesting because there is not
any interaction between the quantum “coin” and the principal quantum system q
in them. This situation is corresponding to the trivial case of classical loop (29)
where the loop guard b is irrelevant to the loop body S. The classical loop (29)
becomes truly interesting only when the loop guard b and the loop body S share
Quantum Recursion and Second Quantisation
33
some program variables. Likewise, a much more interesting form of quantum whileloop is
qwhile W [c; q] = |1i do U [q] od
(34)
which is defined to be the program X declared by the recursive equation
X ⇐ W [c, q]; qif [c] |0i → skip
|1i → U [q]; X
fiq
where W is a unitary operator in the state Hilbert space Hc ⊗ Hq of the composed system of the quantum “coin” c and the principal system q. The operator
W describes the interaction between the “coin” c and the principal system q. It
is obvious that the loop (34) degenerates to the loop (32) whenever W = V ⊗ I,
where I is the identity operator in Hq . The semantics of the loop (34) in the free
Fock space is the operator:
JXK =
∞
X
k=1
(|1ic0 h1| ⊗ (|1ic1 h1| ⊗ ...(|1ick−2 h1| ⊗ (|0ick−1 h0| ⊗ U k−1 [q])
W [ck−1 , q])W [ck−2 , q]...)W [c1 , q])W [c0 , q]
∞
k−1
k−2
X
Y
O
W [cj , q] .
=
|1icj h1| ⊗ |0ick−1 h0| ⊗ U k−1 [q]
k=1
j=0
j=0
Furthermore, the symmetric semantics of the loop is:
∞
k−1
X
Y
A(k) ⊗ U k−1 [q]
W [cj , q] ,
JXKsym =
k=1
j=0
where:
A(k) =
k−1
1X
|1ic0 h1| ⊗ ... ⊗ |1icj−1 h1| ⊗ |0icj h0| ⊗ |1icj+1 h1| ⊗ ... ⊗ |1ick−1 h1|.
k j=0
9 Conclusion
In this paper, we introduced the notion of quantum recursion based on quantum case
statement and quantum choice defined in [21], [22]. Recursive quantum walks and
quantum while-loops were presented as examples of quantum recursion. The denotational and operational semantics of quantum recursion were defined by using second
quantisation, and they were proved to be equivalent. But we are still at the very beginning of the studies of quantum recursion, and a series of problems are left unsolved:
– First of all, it is not well understood what kind of computational problems can be
solved more conveniently by using quantum recursion.
34
Ying
– Second, how to build a Floyd-Hoare logic for quantum while-loops defined in Example 10? Blute, Panangaden and Seely [6] observed that Fock space can serve as a
model of linear logic with exponential types. Perhaps, such a program logic can be
established through combining linear logic with the techniques developed in [19].
– Another important open question is: what kind of physical systems can be used to
implement quantum recursion where new “coins” must be continuously created?
– Finally, we even do not fully understand how does a quantum recursion use its
“coins” in its computational process. In the definition of the principal system semantics of a recursive program (Definition 9), a state |Ψ i in the Fock space of
“coins” is given a priori. This means that the states of a “coin” and its copies are
given once for all. Another possibility is that the states of the copies of a “coin” are
created step by step, as shown in the following:
Example 11 Consider the recursive program X declared by
X ⇐ a†c (|0i); Ry [c, p]; qif [c] |0i → skip
|1i → TR [p]; X
fiq
where c is a “coin” variable with state space Hc = span{|0i, |1i}, the variable p
and operator TR are as in the Hadamard walk,
Ry [c, p] =
∞ h
i
π
X
Ry n+1 ⊗ |nip hn|
2
n=0
and Ry (θ) is the rotation of a qubit about the y−axis in the Bloch sphere. Intuitively, Ry [c, p] is a controlled rotation where position of p is used to determine the
rotated angle. It is worth noting that this program X is a quantum loop defined
in equation (34) but modified by adding a creation operator at the beginning. Its
initial behaviour starting at position 0 with the “coin” c being in the vacuum state
|0i is visualised by the following transitions:
a†d (|0i)
Rx [d,p] 1
|0i|0ip −→
|0i|0ip −→ √ (|0i + |1i) |0ip
2
qif ...fiq 1
−→ √ [(E, |0i|0ip ) + (X, |1i|1ip )] .
2
The first configuration at the end of the above equation terminates, but the second
continues the computation as follows:
a† (|0i)
Rx [d,p]
d
|1i|1ip −→
|0, 1iv |0ip −→ · · · .
It is clear from the above example that the computation of a recursive program
with the creation operator is very different from that without it. A careful study of
quantum recursions that allows the creation operator appear in their syntax will be
carried out in another paper.
Quantum Recursion and Second Quantisation
35
Acknowledgement
I’m very grateful to Professor Prakash Panangaden for teaching me the second quantisation method during his visit at the University of Technology, Sydney in 2013. The
first version of this paper is the text of the third part of my talk “Quantum programming:
from superposition of data to superposition of programs” at the Tsinghua Software Day,
April 21-22, 2014 (see: http://sts.thss.tsinghua.edu.cn/tsd2014/home.html. The first part
of the talk is based on [19], and the second part is based on [22]). I’m also grateful to
Professors Jean-Pierre Jouannaud and Ming Gu for inviting me.
References
1. D. Aharonov, A. Ambainis, J. Kempe and U. Vazirani, U, Quantum walks on graphs, In:
Proceedings of the 33rd ACM Symposium on Theory of Computing (STOC), 2001, pp. 5059.
2. A. Ambainis, E. Bach, A. Nayak, A. Vishwanath and J. Watrous, One-dimensional quantum
walks, In: Proceedings of the 33rd ACM Symposium on Theory of Computing (STOC), 2001,
pp. 37-49.
3. T. Altenkirch and J. Grattage, A functional quantum programming language, In: Proceedings
of the 20th Annual IEEE Symposium on Logic in Computer Science (LICS), 2005, pp.249258.
4. K. R. Apt, F. S. de Boer and E. -R. Olderog, Verification of Sequential and Concurrent
Programs, Springer, London, 2009.
5. S. Attal, Fock spaces, http://math.univ-lyon1.fr/∼attal/Mescours/fock.pdf
6. R. F. Blute, P. Panangaden and R. A. G. Seely, Holomorphic models of exponential types
in linear logic, In: Proceedings of the 9th Conference on Mathematical Foundations of Programming Semantics (MFPS), Springer LNCS 802, 1994, pp. 474-512.
7. G. Chiribella, G. M. D’Ariano, P. Perinotti and B. Valiron, Quantum computations without
definite causal structure, Physical Review A 88 (2013), art. no. 022318.
8. Ph. A. Martin and F. Rothen, Many-Body Problems and Quantum Field Theory: An Introduction, Springer, Berlin, 2004.
9. E. W. Dijkstra, Guarded commands, nondeterminacy and formal derivation of programs,
Communications of the ACM 18 (1975), 453-457.
10. K. Etessami and M. Yannakakis, Recursive Markov chains, stochastic grammars, and monotone systems of nonlinear equations, Journal of the ACM 56 (2009), art. no. 1.
11. Y. Feng, N. K. Yu and M. S. Ying, Reachability analysis of recursive quantum Markov chains,
In: Proceedings of the 38th International Symposium on Mathematical Foundations of Computer Science (MFCS), Springer LNCS 8087, 2013, pp. 385-396.
12. N. Friis, V. Dunjko, W. Dür and H. J. Briegel, Implementing quantum control for unknown
subroutines, Physical Review A 89 (2014), art. no. 030303.
13. C. A. R. Hoare, Procedures and parameters: an axiomatic approach, In: Symposium on Semantics of Algorithmic Languages, Springer Lecture Notes in Mathematics 188, 1971, pp
102-116.
14. J. Loeckx and K. Sieber, The Foundations of Program Verification (second edition), John
Wiley & Sons, Chichester, 1987.
15. A. McIver and C. Morgan, Abstraction, Refinement and Proof for Probabilistic Systems,
Springer, New York, 2005.
16. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, Cambridge University Press, 2000.
36
Ying
17. J. W. Sanders and P. Zuliani, Quantum programming, In: Proceedings of Mathematics of
Program Construction 2000, Springer LNCS 1837, 2000, pp. 88-99.
18. P. Selinger, Towards a quantum programming language, Mathematical Structures in Computer Science 14 (2004), 527-586.
19. M. S. Ying, Floyd-Hoare logic for quantum programs, ACM Transactions on Programming
Languages and Systems 39 (2011), art. no. 19.
20. M. S. Ying and Y. Feng, Quantum loop programs, Acta Informatica 47 (2010), 221-250.
21. M. S. Ying, N. K. Yu and Y. Feng, Defining quantum control flow, arXiv:1209.4379,
http://xxx.lanl.gov/abs/1209.4379.
22. M. S. Ying, N. K. Yu and Y. Feng, Quantum alternation: from superposition of data to superposition of programs, arXiv:1402.5172, http://xxx.lanl.gov/abs/1402.5172.
23. X. -Q. Zhou, T. C. Ralph, P. Kalasuwan, M. Zhang, A. Peruzzo, B. P. Lanyon and J. L.
O’Brien, Adding control to arbitrary unknown quantum operations, Nature Communications 2 (2011), art. no. 413.
24. P. Zuliani, Quantum Programming, D.Phil. Thesis, University of Oxford, 2001.
| 6cs.PL
|
Resumptions, Weak Bisimilarity and Big-Step Semantics for
While with Interactive I/O:
An Exercise in Mixed Induction-Coinduction
Keiko Nakata and Tarmo Uustalu
Institute of Cybernetics at Tallinn University of Technology, Akadeemia tee 21, EE-12618 Tallinn, Estonia
{keiko|tarmo}@cs.ioc.ee
We look at the operational semantics of languages with interactive I/O through the glasses of constructive type theory. Following on from our earlier work on coinductive trace-based semantics
for While [17], we define several big-step semantics for While with interactive I/O, based on resumptions and termination-sensitive weak bisimilarity. These require nesting inductive definitions
in coinductive definitions, which is interesting both mathematically and from the point-of-view of
implementation in a proof assistant.
After first defining a basic semantics of statements in terms of resumptions with explicit internal
actions (delays), we introduce a semantics in terms of delay-free resumptions that essentially removes
finite sequences of delays on the fly from those resumptions that are responsive. Finally, we also look
at a semantics in terms of delay-free resumptions supplemented with a silent divergence option. This
semantics hinges on decisions between convergence and divergence and is only equivalent to the
basic one classically. We have fully formalized our development in Coq.
1
Introduction
Interactive programs are those programs that take inputs, do some computation, output results, and iterate
this cycle possibly infinitely. Operating systems and data base systems are typical examples. They are
important programs and have attracted formal study to guarantee their correctness/safety. For instance,
a web application should protect confidentiality of the data it processes in interaction with possibly
untrusted agents, and a certified compiler should preserve input/output behavior of the source program
in the compiled code. These works call for formal semantics of interactive programs.
Continuing our previous work [17] on a trace-based coinductive big-step semantics for potentially
nonterminating programs, we present a constructive account of interactive input-output resumptions1 ,
their important properties, such as weak bisimilarity and responsiveness (a program always eventually
performs input or output unless it terminates) and big-step semantics of reactive programs. We devise
both constructive-style and classical-style concepts and identify their relationships. Classical-style concepts rely on upfront decisions of whether a computation is going to terminate, make an observable
action, i.e., perform input or output, or silently diverge. The problem is generally undecidable. Hence,
classical-style concepts tend to be too strong for constructive reasoning.
Our operational semantics are resumption-based. A resumption is roughly a tree representing possible runs of a program. The tree branches on inputs, each edge corresponding to each possible input,
and has infinitely deep paths if the program may diverge. We begin the paper by formalizing important
properties of resumptions, among which (termination-sensitive) weak bisimilarity is the most interesting
1 The word ‘resumption’ is sometimes reserved for denotations of parallel threads. We apply it more liberally to datastructures recording evolution in small steps. This usage dates back to Plotkin [20] and was reinforced by Cenciarelli and Moggi [5].
L. Aceto, P. Sobociński (Eds.): Seventh Workshop on
Structural Operational Semantics 2010 (SOS ’10)
EPTCS 32, 2010, pp. 57–75, doi:10.4204/EPTCS.32.5
58
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
one technically, requiring nesting of induction into coinduction. We give a constructive-style formulation of weak bisimilarity and relate it to the classical-style version adapted from previous work [13, 3].
We then present three big-step semantics for interactive While, i.e., While extended with input/output
statements: a basic semantics which explicitly deals with internal actions (delay steps) and assigns a
resumption for all configurations (statement-state pairs); a delay-free semantics for responsive configurations; and a classical-style semantics, which is total classically for all configurations. The two latter
semantics collapse finite sequences of delay steps on the fly. The classical-style semantics can in addition recognize silent divergence; classical-style resumptions include a distinguished element to represent
divergence. Moreover, all three semantics are equivalent under suitable assumptions. Our approach with
big-step semantics in terms of resumptions allow for reasoning about operational behaviors of programs
in a syntax-independent way. We therefore argue that it is more abstract than approaches by means of
small-step semantics, or labelled transition systems (in terms of configurations involving a residual program or a control point). To compare our big-step semantics to more traditional approaches, we also
define an uncontroversial small-step semantics with an associated notion of weak bisimilarity of configurations and show that it agrees with our basic big-step semantics. These technical results form the main
contributions of the paper.
Why do we want to be constructive? First, let us state that our choice is neither motivated nor
depends on any argument of truth: we are not claiming in this paper that classical logic is less true
than intuitionistic logic and none of the points we make hinge on this being the case. Nevertheless, we
do think that working in a constructive logic is very useful also if one has no philosophical problem
in accepting non-constructive arguments. Our reasons are these. For us, using constructive logic is
primarily a technical way to be conscious about the principles we depend on in our arguments. We are
by no means limiting ourselves: when we really need some non-constructive principle in a constructive
argument, we can always explicitly assume this principle (or the specific instance that we need). But
it so happens that a need for unexpectedly strong principles is often a sign of some imperfect design
choice in the setup of a theory. Another reason to be constructive as a semanticist is that programming
is about computable functions only. In constructive logic, we do not have to specifically worry about
computability: only computable functions are there and can be spoken about. For example, the formula
∀x. p x ∨ ¬(p x) is not a tautology, it states that p is a decidable: there is a computable function mapping
any x to a proof of p x or a proof of ¬(p x) (so also to yes or no, should one not care about the proofs).
In big-step semantics, although we specify evaluation as a relation in this paper, it is important for us
that it can be turned into a function, or else we do not capture the intuitive idea that programs represent
computable functions from initial configurations into behaviors.
We have formalized the development in Coq version 8.2pl1. This gives us greater confidence in the
correctness of our reasoning, in particular regarding the productivity of coinductive proofs, since the
type checker of Coq helps us avoid mistakes by ruling out improductivity. We rely on Mendler-style
coinduction to circumvent the limitations imposed by syntactic guardedness approach [9] of Coq. The
Coq development is available at http://cs.ioc.ee/~keiko/sophie.tgz.
The language we consider is the While language extended with input and output primitives, with
statements s : stmt defined inductively by
s ::= skip | s0 ; s1 | x := e | if e then st else s f | while e do st | input x | output e
We assume given the sets of variables and (pure) expressions, whose elements are ranged over by
metavariables x and e respectively. We assume the set of values to be the integers, non-zero integers
counting as truth and zero as falsity. The metavariable v ranges over values. A state, ranged over by σ ,
maps variables to values. The notation σ [x 7→ v] denotes the update of a state σ with v at x. We assume
K. Nakata & T. Uustalu
59
given an evaluation function JeKσ , which evaluates e in the state σ . We write σ |= e and σ 6|= e to denote
that e is true, resp. false in σ .
2
Resumptions
We will define operational semantics of interactive While in terms of states and (interactive input-output)
resumptions. Informally, a resumption is a datastructure that captures all possible evolutions of a configuration (a statement-state pair), a computation tree branching according to the external non-determinism
resulting from interactive input.2
Basic (delayful) resumptions r : res are defined coinductively by the rules3
σ : state
ret σ : res
f : Int → res
in f : res
v : Int r : res r : res
out v r : res
δ r : res
so a resumption either has terminated with some final state, ret σ , takes an integer input v and evolves
into a new resumption f v, in f , outputs an integer v and evolves into r, out v r, or performs an internal
action (observable at best as a delay) and becomes r, δ r. For simplicity, we assume input totality; i.e.,
input resumptions, represented by total functions, accept any integers. But we could instead have had
them partial, e.g., by letting the constructor in take the intended domain of definedness as an additional
argument. We also define (strong) bisimilarity of two resumptions, r ≈ r∗ , coinductively by
ret σ ≈ ret σ
∀v. f v ≈ f∗ v
in f ≈ in f∗
r ≈ r∗
out v r ≈ out v r∗
r ≈ r∗
δ r ≈ δ r∗
Bisimilarity is straightforwardly seen to be an equivalence. We think of bisimilar resumptions as equal,
i.e., type-theoretically we treat resumptions as a setoid with bisimilarity as the equivalence relation4 .
Accordingly, we have to make sure that all functions and predicates we define on resumptions are setoid
functions and predicates, i.e., insensitive to bisimilarity.
Here are some examples of resumptions, defined by corecursion:
⊥
rep n
rep′ n
echo σ
echo′
=
=
=
=
=
δ⊥
δ (δ (out n (rep n)))
δ (out n (rep′ n))
in (λ n. δ (if n 6= 0 then out n (echo σ ) else ret σ ))
in (λ n. δ (if n 6= 0 then out n echo′ else ⊥))
⊥ represents a resumption that silently diverges. rep outputs an integer n forever. rep′ is similar but has
shorter latency. Both echo and echo′ echo input interactively; the former terminates when the input is 0,
whereas the latter diverges in this situation.
Convergence, r ↓ r′ , states that r converges in a finite number of steps to a resumption r′ , which
has terminated or makes an observable action (performs input/output) as its first move. It is defined
inductively by
ret σ ↓ ret σ
∀v. f v ≈ f∗ v
in f ↓ in f∗
r ≈ r∗
out v r ↓ out v r∗
r ↓ r′
δ r ↓ r′
2 There are alternatives. We could have chosen to work, e.g., with functions from streams of input values into traces, i.e.,
computation paths.
3 We mark inductive definitions by single horizontal rules and coinductive definitions by double horizontal rules.
4 Classically, strong bisimilarity is equality. But we work in an intensional type theory where strong bisimilarity of colists is
weaker than equality (just as equality of two functions on all arguments is weaker than equality of these two functions). E.g.,
⊥ and δ ⊥ are only strongly bisimilar.
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
60
In contrast, divergence, r ↑, states that r diverges silently. It is defined coinductively by
r↑
δ r↑
For instance, we have δ (δ (ret σ )) ↓ ret σ , rep n ↓ out n (rep n) and ⊥ ↑.
Both convergence and divergence are setoid predicates. Constructively, it is not the case that ∀r. (∃r′ . r ↓
′
r ) ∨ r ↑, which amounts to decidability of convergence. But classically, this dichotomy is true. In particular, ∀r. ¬ (∃r′ . r ↓ r′ ) → r ↑ is constructively provable, but ∀r. ¬ r ↑ → ∃r′ . r ↓ r′ holds only classically.
We can now introduce a useful notion of responsiveness. A resumption r is responsive, if it keeps
converging. It is defined coinductively with the help of the convergence predicate by
r ↓ in f ∀v. ( f v) ⇓
r ↓ out v r′
r ↓ ret σ
[resp-ret]
[resp-in]
r⇓
r⇓
r⇓
r′ ⇓
[resp-out]
For instance, rep n, rep′ n and echo σ are responsive, but ⊥ and echo′ are not.
Classically, a resumption is responsive, if it can never evolve into a diverging resumption. Indeed, by
augmenting the definition of responsiveness with a divergence option we obtain a classically tautological
predicate, r m, that we call commitedness.
r ↓ ret σ
rm
[comm-ret]
r ↓ in f
∀v. ( f v) m
rm
[comm-in]
r ↓ out v r′
rm
r′ m
[comm-out]
r↑
rm
[comm-div]
For a resumption r to be committed, it must be the case that it always either converges or diverges. So,
classically, any resumption is committed.
Lemma 2.1 Classically, for all r, r m.
Proof Specifically, we use an instance of excluded middle, ∀r. (∃r′ . r ↓ r′ ) ∨ ¬(∃r′ . r ↓ r′ ), which amounts
to assuming that convergence is decidable.
Lemma 2.2 Convergence, divergence, responsiveness and committedness are setoid predicates.
3
Weak Bisimilarity
Two resumptions are weakly bisimilar, if they are bisimilar modulo collapsing finite sequences of delay
steps between observable actions. It is conceivable that, in practice, weak bisimilarity is what is needed:
one may well be interested only in observable behavior, disregarding finite delays. For instance, to
guarantee correctness of a compiler optimization, we would want to prove that the optimization does not
change the observable behavior of the source program, including termination and divergence behaviors,
but the optimized code may perform fewer internal steps and thus be faster. We therefore formalize
termination-sensitive weak bisimilarity, which distinguishes termination and silent divergence.
Technically, getting the definition of weak bisimilarity right is not straightforward, especially not in
a constructive setting. It requires both induction and coinduction: we need to collapse a finite number of
delay steps between observable actions possibly infinitely. Here we present two equivalent formulations
(actually, we will also give a third one for classical reasoning, which is only equivalent to the first two
classically). The first is closer to the formulations typically found in process calculi literature (except that,
in process calculi, one usually works with termination-insensitive weak bisimilarity). The second nests
induction into coinduction, exhibiting a useful technique for implementation in Coq. In our development,
K. Nakata & T. Uustalu
61
we use both formulations and their equivalence result, freely choosing the one of the two that facilitates
the proof.
The first one, noted r ∼
= r∗ , uses coinduction atop the inductive definition of convergence and is
defined by the rules
r ↓ ret σ r∗ ↓ ret σ
r∼
= r∗
r ↓ in f
r∗ ↓ in f∗ ∀v. f v ∼
= f∗ v r ↓ out v r′
r∼
= r∗
r∗ ↓ out v r∗′
r∼
= r∗
r′ ∼
= r∗′
r∼
= r∗
δ r∼
= δ r∗
so two resumptions are weakly bisimilar if they converge at the same action or can both perform an
internal action, with weakly bisimilar residual resumptions. In particular, two terminating resumptions
are derived to be weakly bisimilar by a single application of the first rule, whereas two silently diverging
resumptions are weakly bisimilar by corecursive application of the fourth rule. For instance, we have
rep n ∼
= echo′ does not hold.
= rep′ n but echo σ ∼
Lemma 3.1 For any r, r′ and r∗ , if r ↓ r′ and r∗ ↑ then ¬ r ∼
= r∗ .
As a corollary, we obtain that the silently diverging resumption ⊥ and resumptions that have terminated,
ret σ , are not weakly bisimilar.
The second formulation, denoted r ∼
=◦ r∗ nests induction into coinduction. We first define ↓X ↓ inductively in terms of X , for any relation (read: setoid relation) X , and then define ∼
=◦ coinductively in terms
◦
of ↓∼
= ↓. For binary relations X , Y , X ⊆ Y denotes ∀x, x∗ . x X x∗ → xY x∗ .
ret σ ↓X↓ ret σ
∀v. f v X f∗ v r ↓X↓ r∗
r X r∗
out v r ↓X↓ out v r∗ in f ↓X↓ in f∗ δ r ↓X↓ r∗
r∼
X ⊆∼
=◦ r∗
=◦ r ↓X↓ r∗
r∼
δ r∼
=◦ r∗
=◦ δ r∗
r ↓X↓ r∗
r ↓X↓ δ r∗
Intuitively, r ↓X ↓ r∗ means that r and r∗ converge to resumptions related by X .
In the first rule of ∼
=◦ , we have used Mendler-style coinduction in order to enable Coq’s syntactic
guarded corecursion for ∼
=◦ . The natural (Park-style) rule to stipulate would have been:
r ↓∼
=◦ ↓ r∗
r∼
=◦ r∗
Coq’s guardedness condition for induction nested into coinduction is too weak to work with the Parkstyle rule: we cannot construct the corecursive functions (coinductive proofs) that we need. With our
definition, the Park-style rule is derivable. We can also prove that ↓X ↓ is monotone in X , which allows
us to recover the natural inversion principle for ∼
=◦ .
Induction and coinduction can be mixed in several ways. An inductive definition can be mutual with
a coinductive definition, if the occurrence of one predicate in the definition of the other is contravariant.5
But this is not our situation. Instead, in our case, we have an inductive and a coinductive definition that
use each other covariantly, but one is nested in the other. Specifically, we have the inductive definition
nested in the coinductive definition6 , since we want finite chunks of ↓∼
=◦ ↓ derivations to be weaved
into an infinite ∼
=◦ derivation. The Agda developer community is currently exploring a novel approach to
coinductive types (based on suspension types) [6, 7] where this form of mixing induction and coinduction
is easily encoded while nesting the other way is problematic.
The two definitions of weak bisimilarity are equivalent.
5 This
means looking for a least X and greatest Y solving a system of equations X = F(Y, X), Y = G(X,Y ), where F and G
are contravariant in their first arguments and covariant in the second arguments.
6 i.e., we have a definition of the form ν X. G( µ Y. F(X,Y ), X) with both F and G covariant in both arguments
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
62
Lemma 3.2 For any r and r∗ , r ∼
= r∗ iff r ∼
=◦ r∗ .
Weak bisimilarity is a setoid predicate and an equivalence relation.
Lemma 3.3 Weak bisimilarity is a setoid predicate: For any r, r′ , r∗ , r∗′ , if r ≈ r′ , r ∼
= r∗ and r∗ ≈ r∗′ ,
′
′
then r ∼
= r∗ . Weak bisimilarity is an equivalence.
Proof Reflexivity and symmetry are straightforward to prove by coinduction. Below we sketch the proof
for transitivity with the second formulation, r ∼
=◦ r∗ , to show Mendler-style coinduction working in our
favour. For binary relations X ,Y , let X ◦Y denote their composition; namely, x(X ◦Y )x′ if there is x′′ such
that x X x′′ and x′′ Y x′ . We first prove, by induction, the transitivity for ↓X ↓, i.e., that, for any resumptions
r0 , r1 , r2 and setoid relations X ,Y , if r0 ↓X ↓ r1 and r1 ↓Y ↓ r2 , then r0 ↓(X ◦Y )↓ r2 . The transitivity of ∼
=◦
◦
◦
◦
states that, for any resumptions r0 , r1 and r2 , if r0 ∼
= r1 and r1 ∼
= r2 , then r0 ∼
= r2 . The proof of this is
◦
◦
∼
∼
by coinduction and inversion on r0 = r1 and r1 = r2 . We show the main case. Suppose we have r0 ∼
=◦ r1
◦
◦
◦
∼
∼
∼
and r1 = r2 , because r0 ↓X ↓ r1 and r1 ↓Y ↓ r2 for some X and Y such that X ⊆ = and Y ⊆ = . By the
transitivity of ↓X ↓ (which was proved by induction separately above), we obtain r0 ↓X ◦Y ↓ r2 . Using the
coinduction hypothesis, we have X ◦Y ⊆ ∼
=◦ , which closes the case. Notably, the invocation
=◦ ⊆ ∼
=◦ ◦ ∼
of the coinduction hypothesis here is properly guarded thanks to our use of Mendler’s trick.
As one should expect, strongly bisimilar resumptions are weakly bisimilar.
Corollary 3.1 For any r, r∗ , r ≈ r∗ , then r ∼
= r∗ .
Proof Immediate from ∼
= being a reflexive setoid predicate.
Termination-sensitive bisimilarity has previously been considered by Kučera and Mayr [13] and
Bohannon et al. [3] (but see also Bergstra et al. [1]). Their version is best suited for classical reasoning
in the sense that terminating and silently diverging resumptions are distinguished by an upfront choice
between convergence and divergence. This version of weak bisimilarity, denoted r ∼
=c r∗ , is defined
coinductively by
r ↓ ret σ r∗ ↓ ret σ
r∼
=c r∗
r ↓ out v r′
r∗ ↓ out v r∗′
r∼
=c r∗
r′ ∼
=c r∗′
r ↓ in f
r∗ ↓ in f∗ ∀v. f v ∼
=c f ∗ v
∼
r =c r∗
r ↑ r∗ ↑
r∼
=c r∗
Only the fourth rule is different from the rules of ∼
= and refers directly to divergence.
The classical-style version of weak bisimilarity, ∼
=c , is stronger than the constructive-style version,
∼
=. The converse is only true classically.
Lemma 3.4 For any r and r∗ , if r ∼
=c r∗ , then r ∼
= r∗ . Classically, for any r and r∗ , if r ∼
= r∗ , then r ∼
=c r∗ .
We insist on the use constructive-style weak bisimilarity, ∼
=, in the constructive setting, because the
classical-style notion fails to enjoy some fundamental properties constructively.
Lemma 3.5 Classical-style weak bisimilarity is a setoid predicate. Classically, it is also an equivalence
weaker than strong bisimilarity.
Proof We only prove that ∼
=c r by coin=c is an equivalence. Reflexivity: We prove that for any r, r ∼
duction. Classically, we have ∀r0 . (∃r0′ . r0 ↓ r0′ ) ∨ r0 ↑. Should r ↑ hold, we immediately conclude r ∼
=c r.
Suppose there exists r′ such that r ↓ r′ . Moreover suppose r′ = in f for some f . The coinduction hypothesis gives us that for any v, f v ∼
=c f v, from which r ∼
=c r follows. The other cases, i.e., when
′
′′
′′
′
r = out v r for some v and r or r = ret σ for some σ , are similar. Symmetry: We prove constructively
that for any r and r′ , if r ∼
=c r by coinduction and inversion on r ∼
=c r′ . Transitivity: We prove
=c r′ then r′ ∼
′′
′′
′
′′
′
′
∼
∼
∼
constructively that for any r, r and r , if r =c r and r =c r then r =c r by coinduction and inversion
on r ∼
=c r′′ .
=c r′ and r′ ∼
K. Nakata & T. Uustalu
63
Constructively, it is not possible to show classical-style weak bisimilarity reflexive and hence we
cannot show any two strong bisimilar resumptions classical-style weakly bisimilar.
A simple example of a resumption r not classical-style weakly bisimilar to itself constructively is
given by any search process that is classically total, but cannot be proved terminating constructively,
since no bound on the search can be given. By definition, a resumption can only be classical-style weakly
bisimilar to another if it terminates or diverges. Constructively, the resumption r is only nondiverging,
we cannot show it terminating.
4
Big-Step Semantics
We now proceed to a first, basic (delayful) big-step operational semantics for our reactive While in terms
of delayful resumptions. Evaluation (s, σ ) ⇒ r, expressing that running a statement s from a state σ
produces a resumption r, is defined coinductively by the rules in Figure 1. The rules for sequence and
∗
while implement the necessary sequencing with the help of extended evaluation (s, r) ⇒ r′ , expressing
that running a statement s from the last state (if it exists) of an already accumulated resumption r results
in a total resumption r′ . Extended evaluation is also defined coinductively, as the coinductive prefix
closure of evaluation.
Input and output statements produce corresponding resumptions that perform input or output actions
and terminate thereafter. We consider assignments and testing of guards of if- and while-statements to
constitute internal actions, observable as delays. This way we avoid introducing semantic anomalies,
by making sure that any while-loop always progresses. But this choice also ensures that evaluation is
total—as we should expect. Given that it is deterministic as well7 , we can equivalently turn our relational
big-step semantics into a functional one: the unique resumption for a given configuration (statementstate pair) is definable by corecursion.8 This semantics is a straightforward adaptation of the trace-based
coinductive big-step semantics of non-interactive While from our previous work [17], where the details
can be found and where we motivate all our design choices (e.g., why skip takes no time whereas the
boolean guards do; we argue that our design is canonical).
Lemma 4.1 Evaluation is a setoid predicate. It is total and deterministic up to bisimilarity.
Let us look at some examples. We have (while true do skip, σ ) ⇒ ⊥ for any σ . I.e., while true do skip
silently diverges. We also have (input x; while true do (output x; x := x + 1), σ ) ⇒ in (λ n. up n) where
up is defined corecursively by up n = δ (out n (δ (up (n + 1)))). I.e., the statement counts up from the
given input n. The two delays around every output action account for the internal actions of the assignment and testing of the boolean guard. An interactive adder takes two inputs and outputs their sum, and
repeats this process, that is, we have (while true do (input x; input y; output (x + y)), σ ) ⇒ sum where
sum is defined cocursively by sum = δ (in (λ m. in (λ n. out (m + n) sum))).
Weak bisimilarity is useful for reasoning about soundness of program transformations, where we accept that transformations may change the timing of a resumption. For instance, we have (while true do (z :=
x; output z), σ ) ⇒ rep (σ x), where rep is defined corecursively by rep n = δ δ (out n (rep n)), and
(z := x; while true do output z, σ ) ⇒ δ (rep′ (σ x)), where rep′ is defined corecursively by rep′ n =
δ (out n (rep′ n)), with rep n ∼
= rep′ n. The latter resumption is faster than the former, but they are weakly
7 Note
that the external nondeterminism resulting from input actions is encapsulated in resumptions.
aspect makes our big-step operational semantics very close in spirit to denotational semantics, specifically, denotational semantics in terms of Kleisli categories, here, the Kleisli category of a resumptions monad.
8 This
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
64
(s0 , σ ) ⇒ r
(x := e, σ ) ⇒ δ (ret σ [x 7→ JeKσ ])
e |= σ
(skip, σ ) ⇒ ret σ
∗
(st , δ (ret σ )) ⇒ r
(if e then st else s f , σ ) ⇒ r
e |= σ
∗
(st , δ (ret σ )) ⇒ r
(input x, σ ) ⇒ in (λ v.ret σ [x 7→ v])
∗
∗
∀v. (s, f v) ⇒ f ′ v
(s, ret σ ) ⇒ r
∗
(if e then st else s f , σ ) ⇒ r
∗
⇒ r′
∗
(s, in f ) ⇒ in f ′
(s0 ; s1 , σ ) ⇒ r′
(s f , δ (ret σ )) ⇒ r
e 6|= σ
(while e do st , r) ⇒ r′
(while e do st , σ )
(s, σ ) ⇒ r
e 6|= σ
∗
(s1 , r) ⇒ r′
(while e do st , σ ) ⇒ δ (ret σ )
(output e, σ ) ⇒ out (JeKσ ) (ret σ )
∗
(s, r) ⇒ r′
∗
∗
(s, δ r) ⇒ δ r′
(s, r) ⇒ r′
∗
(s, out v r) ⇒ out v r′
Figure 1: Basic (delayful) big-step semantics
bisimilar. In fact, we can prove while e do (z := x; s) and z := x; while e do s to be weakly bisimilar whenever e is true of the initial state and s does not change x. Here, the latter statement is obtained from the
former by loop-invariant code motion, a well-known compiler optimization; the optimization preserves
the observable behaviour of the source statement, irrespective of its termination behaviour, which it must
respect as well. We note that output 1 is not observationally equivalent to (while true do skip); output 1.
More importantly, output 1 is not observationally equivalent to output 1; (while true do skip), since our
weak bisimilarity is termination-sensitive. Of course, we can deal with more interesting program equivalences, such as the equivalence of mult = while true do (input x; input y; z := 0; while x 6= 0 do (z :=
z + y; x := x − 1); output z) and mult opt = while true do (input x; input y; if x ≥ 0 then output x ∗
y else (while true do skip)), slow and fast interactive multipliers, which silently diverge when given a
negative first operand.
5
Small-Step Semantics
In this section, we introduce an equivalent small-step semantics and define weak bisimilarity of configurations (statement-state pairs) in terms of it. We then prove two configurations to be weakly bisimilar if
and only if their evaluations produce weakly bisimilar resumptions.
A configuration (s, σ ) is a pair of a statement and state. Labelled configurations c : lconf are defined
by the sum9 :
s : state
ret σ : lconf
s : stmt g : Int → state
in s g : lconf
v : Int s : stmt σ : state
out v s σ : lconf
s : stmt σ : state
δ s σ : lconf
A terminality predicate/one-step reduction relation → is defined in Figure 2 (top half). If c = ret σ ,
then the proposition (s, σ ) → c means that the configuration (s, σ ) has terminated at state σ . In other
cases, it corresponds to a labelled transition: if c = in s′ g, we take an input v and evolve to a configuration
(s′ , g v); if c = out v s′ σ ′ , we output v and evolve to (s′ , σ ′ ); if c = δ s′ σ ′ , the configuration (s, σ ) evolves
to a configuration (s′ , σ ′ ) in a delay step. We have chosen to label configurations rather than transitions
so that labelled configurations become “trunks” of resumptions.
9 The
definition is non-recursive, but we pretend that it is inductive, as we also do in Coq.
K. Nakata & T. Uustalu
65
(x := e, σ ) → δ skip (σ [x 7→ JeKσ ])
(s0 , σ ) → in s′0 f
(s0 ; s1 , σ ) → in (s′0 ; s1 )
(s0 , σ ) → ret σ (s1 , σ ) → c
(s0 ; s1 , σ ) → c
′
′
f
σ |= e
(if e then st else s f , σ ) → δ st σ
(s0 , σ ) → out v s′0 σ ′
(s0 , σ ) → δ s′0 σ ′
(s0 ; s1 , σ ) → out v (s′0 ; s1 ) σ ′
(s0 ; s1 , σ ) → δ (s′0 ; s1 ) σ ′
σ 6|= e
(if e then st else s f , σ ) → δ s f σ
σ |= e
(while e do st , σ ) → δ (st ; while e do st ) σ
(input x, σ ) → in skip (λ v.σ [x 7→ v])
(s, σ ) → ret σ ′
(s, σ )
(s, σ ) → in
s′
g
ret
(s, σ )
σ 6|= e
(while e do st , σ ) → δ skip σ
(output e, σ ) → out (JeKσ ) skip σ
(s, σ ) → δ s′ σ ′
σ′
∀v. (s′ , g
(skip, σ ) → ret σ
(s′ , σ ′ )
(s, σ )
v)
fv
r
δr
(s, σ ) → out v s′ σ ′
(s, σ )
in f
(s′ , σ ′ )
r
out v r
Figure 2: Small-step semantics
Weak bisimilarity of two configurations is defined in terms of terminality/one-step reduction. Again,
convergence, (s, σ ) ↓ c, states that either (s, σ ) terminates or performs an observable action in a finite
number of steps. It is defined inductively by
(s, σ ) → ret σ ′
(s, σ ) ↓ ret σ ′
(s, σ ) → out v s′ σ ′
(s, σ ) ↓ out v s′ σ ′
(s, σ ) → in s′ g (s, σ ) → δ s′ σ ′ (s′ , σ ′ ) ↓ c
(s, σ ) ↓ in s′ g
(s, σ ) ↓ c
(We overload the same notations for resumptions and configurations without ambiguity.) Weak bisimilarity on configurations is defined coinductively by
(s, σ ) ↓ ret σ ′
(s∗ , σ∗ ) ↓ ret σ ′
(s, σ ) ∼
= (s∗ , σ∗ )
(s, σ ) ↓ in
s′
g
(s∗ , σ∗ ) ↓ in s′∗ g∗ ∀v. (s′ , g v) ∼
= (s′∗ , g∗ v)
∼ (s∗ , σ∗ )
(s, σ ) =
(s, σ ) ↓ out v (s′ , σ ′ )
(s, σ ) → δ s′ σ ′
(s∗ , σ∗ ) ↓ out v (s′∗ , σ∗′ )
∼ (s∗ , σ∗ )
(s, σ ) =
(s∗ , σ∗ ) → δ s′∗ σ∗′
(s, σ ) ∼
= (s∗ , σ∗ )
(s′ , σ ′ ) ∼
= (s′∗ , σ∗′ )
(s′ , σ ′ ) ∼
= (s′∗ , σ∗′ )
Two configurations are weakly bisimilar if and only if their evaluations yield weakly bisimilar resumptions.
Lemma 5.1 For any s, s∗ , σ and σ∗ , (s, σ ) ∼
= (s∗ , σ∗ ) iff there exist r and r∗ such that (s, σ ) ⇒ r and
(s∗ , σ∗ ) ⇒ r∗ and r ∼
= r∗ .
The evaluation relation of the small-step semantics is defined in Figure 2 (bottom half). It is the
terminal many-step reduction relation, defined coinductively. The proposition (s, σ )
r means that
running s from the state σ produces the resumption r.
The big-step and small-step semantics are equivalent.
Proposition 5.1 For any s, σ and r, (s, σ ) ⇒ r iff (s, σ ) r.
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
66
6
Delay-Free Big-Step Semantics
So far we explicitly dealt with delay steps in a fully general and constructive manner. However, it is
also possible to define big-step semantics in terms of resumptions without delay steps, by collapsing
them on the fly, if they come in finite sequences. In this section, we define a delay-free semantics for
configurations that lead to responsive resumptions.
We define delay-free resumptions, r : resr , and their (strong) bisimilarity coinductively by
σ : state
retr σ : resr
f : Int → resr
inr f : resr
v : Int r : resr
outr v r : resr
retr σ ≈ retr σ
∀v. f v ≈ f∗ v
inr f ≈ inr f∗
r ≈ r∗
outr v r ≈ outr v r∗
A responsive delayful resumption r : res can be normalized into a delay-free resumption by collapsing
the finite sequences of delay steps it has between observable actions. We define normalization, norm : (r :
res) → r ⇓ → resr , and embedding of delay-free resumptions into delayful resumptions, emb : resr → res
by corecursion. In the definition of norm, we examine the proof of r ⇓, i.e., r’s responsiveness.
emb (retr σ ) =
norm r (resp-ret σ ) = retr σ
norm r (resp-in f k) = inr (λ v. norm ( f v) (k v)) emb (inr f ) =
emb (outr v r) =
norm r (resp-out v r′ h) = outr v (norm r′ h)
ret σ
in (λ v. emb ( f v))
out v (emb r)
A delayful resumption is weakly bisimilar to a delay-free one if and only if it is responsive and its
normal form is strongly bisimilar to the same.
Lemma 6.1 For any r : res and r∗ : resr , r ∼
= emb r∗ iff norm r h ≈ r∗ for some h : r ⇓.
(The convergence proofs of a resumption are strong bisimilar, so h is unique up to that extent.)
Corollary 6.1 (i) For any r : res, h : r ⇓, r ∼
= emb (norm r h). (ii) For any r, h : r ⇓ and r∗ , h∗ : r∗ ⇓, r ∼
= r∗
iff norm r h ≈ norm r∗ h∗ .
In Figure 3, we define the delay-free big-step semantics for responsive programs. Here we have
an inductive definition of a parameterized evaluation relation ⇒↓(X ) defined in terms of X , for any
∗
relation X , nested into a coinductive definition of an extended evaluation relation ⇒, defined in terms of
∗
∗
⇒↓(⇒). Finally, the actual evaluation relation ⇒r of interest is obtained by instantiating ⇒↓ at ⇒. Since
we collapse delay-steps on the fly, an assignment immediately terminates at the updated state. Likewise,
testing the guard of a condition or a while-loop takes no time. The crucial rules are those for sequence
and while-loop. If the first statement of a sequence or the body of a while-loop terminate silently, the
second statement or the new iteration of the loop are run using the inductive evaluation. The coinductive
extended evaluation is used only if the first statement or the body perform at least one input or output
action.
This way, we make sure that only a finite number of delay steps may be collapsed between two
observable actions, while allowing for diverging runs which perform input and output every now and
then. Indeed, if we replaced the while-ret rule with
e |= σ
∗
(st , σ ) ⇒↓(X) retr σ ′ (while e do st , retr σ ′ ) ⇒ r′
(while e do st , σ ) ⇒↓(X) r′
we would obtain semantic anomalies. E.g., (while true do skip, σ ) ⇒r r would be derived for any r : resr .
K. Nakata & T. Uustalu
67
(x := e, σ ) ⇒↓(X) retr (σ [x 7→ JeKσ ])
(skip, σ ) ⇒↓(X) retr σ
(s0 , σ ) ⇒↓(X) inr f ∀v. (s1 , f v) X f ′ v
(s0 ; s1 , σ ) ⇒↓(X) inr f ′
e |= σ (st , σ ) ⇒↓(X) r
(if e then st else s f , σ ) ⇒↓(X) r
e |= σ
e |= σ
(s0 , σ ) ⇒↓(X) retr σ ′ (s1 , σ ′ ) ⇒↓(X) r
(s0 ; s1 , σ ) ⇒↓(X) r
(s0 , σ ) ⇒↓(X) outr v r (s1 , r) X r′
(s0 ; s1 , σ ) ⇒↓(X) outr v r′
e 6|= σ (s f , σ ) ⇒↓(X) r
(if e then st else s f , σ ) ⇒↓(X) r
(st , σ ) ⇒↓(X) retr σ ′ (while e do st , σ ′ ) ⇒↓(X) r
(while e do st , σ ) ⇒↓(X) r
(st , σ ) ⇒↓(X) inr f ∀v. (while e do st , f v) X f ′ v
(while e do st , σ ) ⇒↓(X) inr f ′
e |= σ
(st , σ ) ⇒↓(X) outr v r (while e do st , r) X r′
(while e do st , σ ) ⇒↓(X) outr v r′
e 6|= σ
(while e do st , σ ) ⇒↓(X) retr σ
(input x, σ ) ⇒↓(X) inr (λ v.retr σ [x 7→ v])
∗
X ⊆ ⇒ (s, σ ) ⇒↓(X) r
∗
(s, ret r σ ) ⇒ r
(output e, σ ) ⇒↓(X) outr (JeKσ ) (retr σ )
∗
∗
∀v. (s, f v) ⇒ f ′ v
∗
(s, inr f ) ⇒ inr f ′
(s, r) ⇒ r′
∗
(s, outr v r) ⇒ outr v r′
∗
(s, σ ) ⇒↓(⇒) r
(s, σ ) ⇒r r
Figure 3: Delay-free big-step semantics
Coming back to the examples of the previous section, we have (input x; while true do (output x; x :=
x + 1), σ ) ⇒r in (λ n. upr n) where upr is defined corecursively by upr n = out r n (upr (n + 1)). We also
have (while true do z := x; output z, σ )⇒r repr (σ x) and (z := x; while true do output z, σ ) ⇒ repr (σ x)
where repr is defined corecursively by repr n = outr n (repr n). Since the delay steps are collapsed on the
fly in the delay-free semantics, the two statements produce the same, i.e., strongly bisimilar, (delay-free)
resumptions. The delay-free semantics does not account for (i.e., does not assign a resumption to) nonresponsive configurations, such as while true do skip and the interactive multipliers from the previous
section (since they diverge given a negative input for the first operand), with any initial state.
We state adequacy of the delay-free semantics by relating it to the delayful semantics of Section 4.
Namely, for configurations leading to responsive resumptions they agree.
Proposition 6.1 (Soundness) For any s, σ , r : resr , if (s, σ ) ⇒r r then there exists r′ : res such that
(s, σ ) ⇒ r′ and emb r ∼
= r′ .
Proposition 6.2 (Completeness) For any s, σ , r : res and h : r ⇓, if (s, σ ) ⇒ r, then (s, σ ) ⇒r norm r h.
The proofs are omitted due to the space limitation. They are nontrivial and the details can be found
in the accompanying Coq development. Below we demonstrate the key proof technique on an example.
Consider the statement count = while true do (if i > 0 then i := i − 1 else (output x; x := x + 1; i :=
x)). It counts up from 0, so we should have (count, σ ) ⇒r upr 0 for an initial state σ that maps x and i
to 0. We need coinduction since count performs outputs infinitely often; we also need induction, nested
68
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
into coinduction, since the loop silently iterates n times each time before outputting n. Note that the
latency is finite but unbounded.
We cannot perform induction inside coinduction naı̈vely. That would be rejected by Coq’s syntactic
guardedness checker, which is there to ensure productivity of coinduction. Mendler-style coinduction
comes to rescue. Let (s, r) R r′ be a relation on pairs (s, r) of a statement and a resumption and resumptions r′ , defined inductively by
(count, retr σ [x 7→ n, i 7→ n]) R upr n
(s, r) R r′
(s, outr v r) R outr v r′
(s, σ ) ⇒↓(R) r
(s, retr σ ) R r
∗
The key fact is that R is stronger than ⇒ (Lemma 6.6 below).
We first prove that ⇒↓ is monotone by induction.
Lemma 6.2 For any X , Y , s, σ and r such that X ⊆ Y , if (s, σ ) ⇒↓(X ) r, then (s, σ ) ⇒↓(Y ) r.
The following two lemmata are proved by straightforward application of the rules in Figure 3.
Lemma 6.3 For any n, (if i > 0 then i := i − 1 else (output x; x := x + 1; i := x), σ [x 7→ n, i 7→ 0]) ⇒↓(R)
out r n (ret r σ [x 7→ n + 1, i 7→ n + 1]).
Lemma 6.4 For any n and m, (if i > 0 then i := i − 1 else (output x; x := x + 1; i := x), σ [x 7→ n, i 7→
m + 1]) ⇒↓(R) ret r σ [x 7→ n, i 7→ m].
The next lemma is proved by induction on m, using the two lemmata just proved.
Lemma 6.5 For any n and m, (count, σ [x 7→ n, i 7→ m]) ⇒↓(R) outr n (upr (n + 1)).
Corollary 6.2 For any n, (count, σ [x 7→ n, i 7→ n]) ⇒↓(R) upr n.
∗
We can now prove that R is stronger than ⇒ by coinduction and inversion on (s, r) R r′ . Here is the
∗
crux: corollary 6.2 together with the coinduction hypothesis gives (count, ret r σ [x 7→ n, i 7→ n]) ⇒ upr n,
and the use of the coinduction hypothesis is properly guarded.
∗
Lemma 6.6 For any s, r and r′ , if (s, r) R r′ then (s, r) ⇒ r′
The main proposition follows from corollary 6.2, lemma 6.6 and the monotonicity of ⇒↓ (lemma 6.2).
Proposition 6.3 For any n, (count, σ [x 7→ n, i 7→ n]) ⇒r upr n.
7
Classical-Style Big-Step Semantics
In Section 2, we augmented the definition of responsiveness with a divergence option to obtain a concept
of committedness, which is a classically tautological predicate. Similarly, we can obtain a delay-free
semantics for committed configurations from the delay-free semantics for responsive configurations of
the previous section. To do so, we extend the definition of delay-free resumptions with a “black hole”
constructor, •, representing silent divergence, arriving at classical-style resumptions, and adjust the definition of (strong) bisimilarity:
σ : state
retc σ : resc
retc σ ≈ retc σ
f : Int → resc
inc f : resc
∀v. f v ≈ f∗ v
inc f ≈ inc f∗
r : resc
outc v r : resc
• : resc
r ≈ r∗
outc v r ≈ outc v r∗
•≈•
K. Nakata & T. Uustalu
69
Given a proof h : r m of committedness of a delayful resumption r : res, we can normalize r into
a classical resumption by collapsing the finite delays between observable actions and sending silent
divergence into the black hole.
norm r (comm-ret σ )
norm r (comm-in f k)
norm r (comm-out v r′ h)
norm r (comm-div )
=
=
=
=
retc σ
emb (retc σ )
inc (λ v. norm ( f v) (k v)) emb (inc f )
outc v (norm r′ h)
emb (outc v r)
•
emb •
=
=
=
=
ret σ
in (λ v. emb ( f v))
out v (emb r)
δ (emb •)
Again, a delayful resumption is weakly bisimilar to a classical-style one if and only if it is committed
and its normal form is strongly bisimilar.
Lemma 7.1 For any r : res and r∗ : resc , r ∼
= emb r∗ iff norm r h ≈ r∗ for some h : r m.
In Figure 4, we define the classical-style semantics in terms of classical-style resumptions. We have
an inductive parameterized evaluation relation ⇒↓(X ), defined in terms of X , for any relation X , for
convergent runs; its inference rules are the same as those in the previous section. But we also have a
coinductive parameterized evaluation ⇒↑(X ), again defined in terms of X , for any relation X , for silently
∗
diverging runs, so that (s, σ ) ⇒↑(⇒) expresses that running a statement s from a state σ diverges without
performing input or output. It uses the inductive evaluation in case the first statement of a sequence or
the first iteration of the body of a while-loop silently terminates, but the whole sequence or while-loop
∗
silently diverges. Then we define coinductively an extended evaluation relation ⇒, in terms of these two
∗
evaluation relations, nesting the latter into the former. Finally, we instantiate both ⇒↓ and ⇒↑ at ⇒ to
obtain the “real” evaluation relation ⇒c . Note that, to derive an evaluation proposition in this semantics,
one has to decide upfront whether inductive or coinductive evaluation should be used—a decision that
can be made classically, but not constructively.
The classical-style semantics is adequate wrt. the basic semantics of Section 4.
Proposition 7.1 (Soundness) For any s, σ and r : resc , if (s, σ ) ⇒c r, then there exists r′ : res such that
(s, σ ) ⇒ r′ and emb r ∼
= r′ .
Proposition 7.2 (Completeness) For any s, σ and r : res and h : r m, if (s, σ ) ⇒ r, then (s, σ ) ⇒c
norm r h.
Corollary 7.1 Classically, for any s, σ and r : res, if (s, σ ) ⇒ r, then there exists r′ : resc such that
(s, σ ) ⇒c r′ and r ∼
= emb r′ .
The classical-style semantics is more expressive than responsive semantics, since it offers the option
of “detected” divergence. In particular we have (while true do skip, σ ) ⇒c • and our interactive multipliers are assigned a classical-style resumption mult defined corecursively by mult = inc (λ m. inc (λ n. if m ≥
0 then outc (m ∗ n) mult else •)); i.e., we have (mult, σ ) ⇒c mult and (mult opt, σ ) ⇒c mult.
8
Related Work
Formalized semantics are an important ingredient in the trusted computing base of certified compilers.
Proof assistants, like Coq, are a good tool for such formalization projects, as both the object semantics
of interest and its metatheory can be developed in the same framework. For introductions, see [2].
To account for nontermination or silent divergence properly in a big-step semantics is nontrivial
already for languages without interactive I/O. Leroy and Grall [14] introduced two big-step semantics
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
70
(x := e, σ ) ⇒↓(X) retc (σ [x 7→ JeKσ ])
(skip, σ ) ⇒↓(X) retc σ
(s0 , σ ) ⇒↓(X) inc f ∀v. (s1 , f v) X f ′ v
(s0 ; s1 , σ ) ⇒↓(X) inc f ′
e |= σ (st , σ ) ⇒↓(X) r
(if e then st else s f , σ ) ⇒↓(X) r
e |= σ
e |= σ
(s0 , σ ) ⇒↓(X) retc σ ′ (s1 , σ ′ ) ⇒↓(X) r
(s0 ; s1 , σ ) ⇒↓(X) r
(s0 , σ ) ⇒↓(X) outc v r (s1 , r) X r′
(s0 ; s1 , σ ) ⇒↓(X) outc v r′
e 6|= σ (s f , σ ) ⇒↓(X) r
(if e then st else s f , σ ) ⇒↓(X) r
(st , σ ) ⇒↓(X) retc σ ′ (while e do st , σ ′ ) ⇒↓(X) r
(while e do st , σ ) ⇒↓(X) r
(st , σ ) ⇒↓(X) inc f ∀v. (while e do st , f v) X f ′ v
(while e do st , σ ) ⇒↓(X) inc f ′
e |= σ
(st , σ ) ⇒↓(X) outc v r (while e do st , r) X r′
(while e do st , σ ) ⇒↓(X) outc v r′
e 6|= σ
(while e do st , σ ) ⇒↓(X) retc σ
(input x, σ ) ⇒↓(X) inc (λ v.retc σ [x 7→ v])
(s0 , σ ) ⇒↑(X)
(output e, σ ) ⇒↓(X) outc (JeKσ ) (retc σ )
(s0 , σ ) ⇒↓(X) retc σ ′
(s0 ; s1 , σ ) ⇒↑(X)
e |= σ
(s1 , σ ′ ) ⇒↑(X)
(s0 ; s1 , σ ) ⇒↑(X)
e 6|= σ
(st , σ ) ⇒↑(X)
(s f , σ ) ⇒↑(X)
(if e then st else s f , σ ) ⇒↑(X) (if e then st else s f , σ ) ⇒↑(X)
e |= σ
(st , σ ) ⇒↑(X)
e |= σ
(st , σ ) ⇒↓(X) retc σ ′
(while e do st , σ ) ⇒↑(X)
∗
X ⊆ ⇒ (s, σ ) ⇒↓(X) r
∗
(s, ret c σ ) ⇒ r
(while e do st , σ ′ ) ⇒↑(X)
(while e do st , σ ) ⇒↑(X)
∗
X ⊆ ⇒ (s, σ ) ⇒↑(X)
∗
∗
∗
(s, retc σ ) ⇒ •
(s, inc f ) ⇒ inc f ′
∗
∗
∀v. (s, f v) ⇒ f ′ v
(s, σ ) ⇒↓(⇒) r
(s, σ ) ⇒c r
(s, r) ⇒ r′
∗
(s, outc v r) ⇒ outc v r′
∗
(s, •) ⇒ •
∗
(s, σ ) ⇒↑(⇒)
(s, σ ) ⇒c •
Figure 4: Classical-style big-step semantics
for lambda-calculus. One is classical in spirit, with two evaluation relations, inductive and coinductive,
for terminating and diverging runs, and relies on decidability between termination and divergence. The
other, with a single coinductive evaluation relation, is essentially suited for constructive reasoning, but
contains a semantic anomaly (a function can continue reducing after the argument diverges), which
results from its ability to collapse an infinite sequence of internal actions (contraction steps).
In our work [17] on While with nontermination, we developed a trace-based coinductive big-step
semantics where traces were non-empty colists of intermediate states, agreeing with the very standard
coinductive small-step trace-based semantics. This semantics relied on traces being a monad; a central
component in the definition was an extended evaluation relation, corresponding to the Kleisli extension
of evaluation. Capretta [4] studied constructive denotational semantics of nontermination as the Kleisli
semantics for the delayed state monad, corresponding to hiding the intermediate states in the trace monad
as internal actions and quotienting by termination-sensitive weak bisimilarity. Rutten [21] carried out a
similar project in classical set theory where the quotient is the state space extended with an extra element
K. Nakata & T. Uustalu
71
for nontermination.
Operational semantics of interactive programs is most often described in the small-step style where
it amounts to a labelled transition system. Especially, this is the dominating approach in process calculi.
Big-step semantics is closer to denotational semantics. In this field, resumption-based descriptions go
back to Plotkin [20], Gunter et al. [10] and Cenciarelli and Moggi [5]. Resumptions are a monad and
resumptions-based denotational semantics is a Kleisli semantics. Our big-step semantics are directly
inspired by this approach, except that we work in a constructive setting and must take extra care to avoid
the need to invoke classical principles where they are dispensable.
We are not aware of many other works on constructive semantics of interactive I/O. But similar
in its spirit to ours is the work of Hancock et al. [11] on stream processors and the stream functions
that these induce by “eating”. Stream processors are like our delay-free resumptions, except that the
authors emphasize parallel composition of stream processors (one processor’s output becomes another
processor’s input) and, for this to be well-defined, a stream processor must not terminate and may only
do a finite number of input actions consecutively. Hancock et al. [8] also characterize realizable stream
functions. In a precursor work, Hancock and Setzer [12] studied a model of interaction where a client
sends a server commands and expects responses.
Weak bisimilarity tends to be defined termination-insensitively, identifying termination and divergence. In particular, this is also the approach of CCS [16]. Termination-sensitive weak bisimilarity has
been considered by Bergstra, Klop and Olderog [1], Kučera and Mayr [13] and Bohannon et al. [3], but
only in what we call the classical-style version, relying on decisions between convergence and divergence. (The weak bisimilarity of Capretta [4] is termination-sensitive and tailored for constructive reasoning, but restricted to behaviours without I/O. Weak bisimilarity also motivated the study of Danielsson
and Altenkirch [6] on mixed induction-coinduction.)
Mixed inductive-coinductive definitions in the form of induction nested into coinduction
(ν X . µY. F (X ,Y ) or, more generally, ν X . G(µY. F (X ,Y ), X )) seem to be quite fundamental in applications (e.g., the stream processors of Hancock et al., our delay-free semantics). Danielsson and Altenkirch [6, 7] argue for making this mix the basic form of inductive-coinductive definitions in the
dependently-typed programming language Agda. Unfortunately, nestings the other way around (definitions µ X . νY. F (X ,Y )) seem to become difficult or impossible to code. With our approach, coinduction
nested into induction is handled symmetrically to induction nested into coinduction [19].
Mendler-style (co)recursion originates from Mendler [15]. It uses that a monotone (co)inductive
definition is equivalent to a positive one, via a syntactic left (right) Kan extension along identity (instead
of µ X . F X one works with µ X . ∃Y. (Y → X ) → F Y ). We exploited this fact to enable Coq’s guarded
corecursion for a coinductive definition with a nested inductive definition, at the price of impredicativity.
We have previously developed and formalized a Hoare logic for the trace-based semantics of While
with nontermination [18]. A similar enterprise should be possible for resumptions, weak bisimilarity and
While with interactive I/O.
9
Conclusion
We have developed a constructive treatment of resumption-based big-step semantics of While with interactive I/O. We have devised constructive-style definitions of important concepts on resumptions such
as termination-sensitive weak bisimilarity and responsiveness, and devised two variations of delay-free
big-step semantics for programs that produce responsive and committed resumptions, respectively. Responsiveness is for interactive computation what termination is for noninteractive computation. And
72
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
likewise, committedness compares to a decided domain of definedness. Indeed, all three variations of
big-step semantics for While with interactive I/O have counterparts in big-step semantics for noninteractive While (see Appendix). Mathematically, we find it reassuring that observations made for a more
simpler noninteractive While naturally scale to a more involved language with interactive I/O. The central ideas are a concept of termination-sensitive weak bisimilarity tailored for constructive reasoning and
the organization of evaluation in the delay-free semantics as an induction nested into coinduction.
Technically, we have carried out an advanced exercise in programming and reasoning with mixed
induction and coinduction, which we have also formalized in Coq. The challenges in this exercise were
both mathematical and tool-related (Coq-specific). We deem that the mathematical part was more interesting and important. The main new aspect in comparison to our earlier development of coinductive
trace-based big-step semantics for noninteractive While was the need to deal with definitions of predicates that nest induction into coinduction—a relatively unexplored area in type theory. In Coq, we
formalized them by parameterizing the inductive definition and converting the coinductive definition
into Mendler-like format. Apparently, this technique is novel for the Coq community.
As future work, we would like to scale our development to concurrency.
Acknowledgments We thank Andreas Lochbihler, Nils Anders Danielsson and Thorsten Altenkirch
for discussions.
This research was supported by the Estonian Centre of Excellence in Computer Science, EXCS,
funded by the European Regional Development Fund, and the Estonian Science Foundation grant no. 6940.
References
[1] J. Bergstra, J. Klop & E.-R. Olderog (1987): Failures without chaos: A new process semantics for fair
abstraction. In: M. Wirsing (ed.) Proc. of 3rd IFIP TC2/WG2.2 Working Conf. on Formal Description of
Programming Concepts (Ebberup, Aug. 1986). North Holland, Amsterdam, pp. 77–101.
[2] Y. Bertot (2007):
A survey of programming language semantics styles. Coq development,
http://www-sop.inria.fr/marelle/Yves.Bertot/proofs.html
[3] A. Bohannon, B. C. Pierce, V. Sjöberg, S. Weirich & S. Zdancewic (2009): Reactive noninterference. In:
Proc. of 2009 ACM Conf. on Computer and Communications Security, CCS 2009 (Chicago, IL, Nov. 2009),
ACM Press, New York, pp. 79–90.
[4] V. Capretta (2005): General recursion via coinductive types. Logical Methods in Computer Science 1(2),
article 1.
[5] P. Cenciarelli & E. Moggi (1993): A syntactic approach to modularity in denotational semantics. In: Proc. of
5th Biennial Meeting on Category Theory and Computer Science, CTCS ’93 (Amsterdam, Sept. 1993) Tech.
report, CWI, Amsterdam.
[6] N. A. Danielsson & T. Altenkirch (2009):
Mixing
http://www.cs.nott.ac.uk/~nad/publications/.
induction
and
coinduction.
Draft,
[7] N. A. Danielsson & T. Altenkirch (2010): Subtyping, declaratively: an exercise in mixed induction and
coinduction. In C. Bolduc, J. Desharnais & B. Ktari (Eds.): Proc. of 10th Int. Conf. on Mathematics of
Program Construction, MPC 2010 (Québec City, July 2010), Lect. Notes in Comput. Sci. 6120. Springer,
Berlin, pp. 100–118.
[8] M. Ghani, P. Hancock & D. Pattinson (2009): Continuous functions on final coalgebras. In: S. Abramsky,
M. Mislove & C. Palamidessi (eds.) Proc. of 25th Conf. on Mathematical Foundations of Programming
Semantics, MFPS-25 (Oxford, Apr. 2009). Electr. Notes in Theor. Comput. Sci. 249. Elsevier, Amsterdam,
pp. 3–18.
K. Nakata & T. Uustalu
73
[9] E. Giménez (1995): Codifying guarded definitions with recursive schemes. In P. Dybjer, B. Nordström & J.
M. Smith (Eds.): Selected Papers from Int. Wksh. on Types for Proofs and Programs, TYPES ’94 (Båstad,
June 1994), Lect. Notes in Comput. Sci. 996. Springer, Berlin, pp. 39–59
[10] C. A. Gunter, P. D. Mosses & D. S. Scott (1989): Semantic Domains and Denotational Semantics.
[11] P. Hancock, D. Pattinson & N. Ghani (2009): Representations of stream processors using nested fixed points.
Logical Methods in Computer Science 5(3), article 9.
[12] P. Hancock & A. Setzer, A. (2000): Interactive programs in dependent type theory. In: P. Clote &
H. Schwichtenberg (eds.): Proc. of 14th Int. Wksh. on Computer Science Logic, CSL 2000 (Fischbachau,
Aug. 2000). Lect. Notes in Comput. Sci. 1862. Springer, Berlin, pp. 317–331.
[13] A. Kučera & R. Mayr (2002): Weak bisimilarity between finite-state systems and BPA or normed BPP is
decidable in polynomial time. Theor. Comput. Sci. 270(1–2), pp. 677–700.
[14] X. Leroy & H. Grall (2009): Coinductive big-step operational semantics. Inform. and Comput. 207(2),
pp. 285-305.
[15] N. P. Mendler (1991): Inductive types and type constraints in the second-order lambda calculus. Ann. of
Pure and Appl. Logic 51(1–2), pp. 159–172.
[16] R. Milner (1989): Communication and Concurrency. Prentice Hall, New York.
[17] K. Nakata & T. Uustalu (2009): Trace-based coinductive operational semantics for While: big-step and
small-step, relational and functional styles. In: S. Berghofer, T. Nipkow, C. Urban & M. Wenzel (eds.) Proc.
of 22nd Int. Conf. on Theorem Proving in Higher-Order Logics, TPHOLs 2009 (Munich, Aug. 2009). Lect.
Notes in Comput. Sci. 5674. Springer, Berlin, pp. 375-390.
[18] K. Nakata & T. Uustalu (2010): A Hoare logic for the coinductive trace-based big-step semantics of While.
In: A. D. Gordon (ed.) Proc. of 19th Europ. Symp. on Programming, ESOP 2010 (Paphos, March 2010).
Lect. Notes in Comput. Sci. 6012. Springer, Berlin, pp. 488–506.
[19] K. Nakata & T. Uustalu (2010): Mixed induction-coinduction at work for Coq (abstract). Abstract of talk
presented at 2nd Coq Workshop (Edinburgh, July 2010), with accompanying slides and Coq development.
[20] G. D. Plotkin (1983): Domains (“Pisa Notes”). Unpublished notes.
[21] J. Rutten (1999): A note on coinduction and weak bisimilarity for While programs. Theor. Inform. and Appl.
33(4-5), pp. 393-400.
A
Resumptions, Weak Bisimilarity, Delayful, Delay-Free and ClassicalStyle Big-step Semantics for While
The notions of resumptions and weak bisimilarity and the evaluation relations in the three big-step semantics shown of the main text are fairly involved, because of the amount of detail. Therefore, we also
spell out what they specialize (or degenerate) to in the case of ordinary non-interactive While, to better
highlight the phenomena that arise even in the absence of interaction.
A.1
Resumptions, Bisimilarity, Weak Bisimilarity
Delayful resumptions, with their strong bisimilarity, specialize to delayed states r : res à la Capretta [4]
defined coinductively.
σ : state
ret σ : res
r : res
δ r : res
ret σ ≈ ret σ
r ≈ r∗
δ r ≈ δ r∗
Resumptions, Weak Bisimilarity and Big-Step Semantics for While
74
Convergence and (silent) divergence are defined inductively resp. coinductively; convergence reduces to
termination at a final state.
r ↓ r′
δ r ↓ r′
ret σ ↓ ret σ
r↑
δ r↑
Responsiveness reduces to termination. Commitedness becomes decidability between and termination
and divergence. Commitedness is tautologically true only classically.
Weak bisimilarity is defined in terms of convergence coinductively exactly as Capretta [4] did.
r∼
= r∗
δ r∼
= δ r∗
r ↓ ret σ r∗ ↓ ret σ
r∼
= r∗
Any terminating delayed state can be normalized into a state. Any decided delayed state can be normalized into a choice between a state or a special divergence token.
A.2
Delayful Semantics
In the delayful big-step semantics, evaluation and extended evaluation are defined mutually coinductively
as follows.
(s0 , σ ) ⇒ r
(x := e, σ ) ⇒ δ (ret σ [x 7→ JeKσ ])
e |= σ
(skip, σ ) ⇒ ret σ
∗
(st , δ (ret σ )) ⇒ r
(if e then st else s f , σ ) ⇒ r
e |= σ
∗
e 6|= σ
(s0 ; s1 , σ ) ⇒ r′
∗
(s f , δ (ret σ )) ⇒ r
(if e then st else s f , σ ) ⇒ r
∗
(st , δ (ret σ )) ⇒ r
e 6|= σ
(while e do st , r) ⇒ r′
(while e do st , σ )
(while e do st , σ ) ⇒ δ (ret σ )
⇒ r′
(s, σ ) ⇒ r
∗
(s, ret σ ) ⇒ r
∗
(s1 , r) ⇒ r′
∗
(s, r) ⇒ r′
∗
(s, δ r) ⇒ δ r′
We have previously [17] conducted a thorough study of a variation of this semantics (with intermediate
states instead of delays), explaining the design considerations in great detail. We have also [18] developed
a Hoare logic for this semantics.
A.3
Delay-Free Semantics
Delay-free resumptions are the same as states.
In the delay-free semantics, there is one inductive evaluation relation for terminating configurations.
There is no need for a separate extended evaluation relation (which would coincide with evaluation
anyhow, since resumptions and states are the same thing) and no need to parameterize the evaluation
relation.
(x := e, σ ) ⇒↓ σ [x 7→ JeKσ ]
(skip, σ ) ⇒↓ σ
e |= σ (st , σ ) ⇒↓ σ ′
(if e then st else s f , σ ) ⇒↓ σ ′
e |= σ
e 6|= σ
(s0 , σ ) ⇒↓ σ ′ (s1 , σ ′ ) ⇒↓ σ ′′
(s0 ; s1 , σ ) ⇒↓ σ ′′
(s f , σ ) ⇒↓ σ ′
(if e then st else s f , σ ) ⇒↓ σ ′
(st , σ ) ⇒↓ σ ′ (while e do st , σ ′ ) ⇒↓ σ ′′
(while e do st , σ ) ⇒↓ σ ′′
e 6|= σ
(while e do st , σ ) ⇒↓ σ
K. Nakata & T. Uustalu
75
The delay-free semantics agrees with the delayful semantics for terminating delayed states.
It is the textbook big-step semantics of While, which accounts for terminating configurations and
assigns no evaluation result to diverging configurations.
A.4
Classical-Style Semantics
A classical-style resumption is a state or the special token • for divergence.
σ : state
retc σ : resc
• : resc
The classical-style semantics has an inductively defined terminating evaluation relation (defined exactly as that of the delay-free semantics) and a coinductively defined diverging evaluation relation. The
latter depends on the former, but not the other way around. There is no need for an extended evaluation
relation.
(x := e, σ ) ⇒↓ σ [x 7→ JeKσ ]
(skip, σ ) ⇒↓ σ
e 6|= σ
e |= σ (st , σ ) ⇒↓ σ ′
(if e then st else s f , σ ) ⇒↓ σ ′
e |= σ
(s0 , σ ) ⇒↑
(s0 , σ ) ⇒↓ σ ′ (s1 , σ ′ ) ⇒↓ σ ′′
(s0 ; s1 , σ ) ⇒↓ σ ′′
(s f , σ ) ⇒↓ σ ′
(if e then st else s f , σ ) ⇒↓ σ ′
(st , σ ) ⇒↓ σ ′ (while e do st , σ ′ ) ⇒↓ σ ′′
(while e do st , σ ) ⇒↓ σ ′′
(s0 , σ ) ⇒↓ σ ′
(s0 ; s1 , σ ) ⇒↑
(s1 , σ ′ ) ⇒↑
(s0 ; s1 , σ ) ⇒↑
e |= σ
(st , σ ) ⇒↑
e |= σ
e 6|= σ
(while e do st , σ ) ⇒↓ σ
(st , σ ) ⇒↑
(if e then st else s f , σ ) ⇒↑
e |= σ
(st
, σ ) ⇒↓ σ ′
(while e do st , σ ) ⇒↑
(s, σ ) ⇒↓ retc σ ′
(s, σ ) ⇒c σ ′
e 6|= σ
(s f , σ ) ⇒↑
(if e then st else s f , σ ) ⇒↑
(while e do st , σ ′ ) ⇒↑
(while e do st , σ ) ⇒↑
(s, σ ) ⇒↑
(s, σ ) ⇒c •
The classical-style semantics agrees with the delayful semantics for decided delayed states (classically, any delayed state is decided).
A semantics in this spirit (with separate convergent and divergent evaluation relations) was proposed
for untyped lambda calculus by Leroy and Grall [14].
The delayful semantics (together with the identification of weakly bisimilar delayed states) and the
classical-style semantics have the same purposes, but the delayful semantics is better behaved from the
constructive point-of-view. As a practical consequence, it has the advantage that the evaluation relation
can be turned into a function (highly desirable, if one wants to be able to directly execute the big-step
semantics). This is not possible with the classical-style semantics, as one would have to be able to decide
whether a configuration terminates before actually running it.
| 6cs.PL
|
Two-stage algorithms for covering array construction
arXiv:1606.06730v1 [cs.DS] 21 Jun 2016
Kaushik Sarkar and Charles J. Colbourn
School of Computing, Informatics, and Decision Systems Engineering
Arizona State University, PO Box 878809
Tempe, Arizona, 85287-8809, U.S.A.
June 23, 2016
Abstract
Modern software systems often consist of many different components, each with a number of options.
Although unit tests may reveal faulty options for individual components, functionally correct components
may interact in unforeseen ways to cause a fault. Covering arrays are used to test for interactions
among components systematically. A two-stage framework, providing a number of concrete algorithms,
is developed for the efficient construction of covering arrays. In the first stage, a time and memory
efficient randomized algorithm covers most of the interactions. In the second stage, a more sophisticated
search covers the remainder in relatively few tests. In this way, the storage limitations of the sophisticated
search algorithms are avoided; hence the range of the number of components for which the algorithm can
be applied is extended, without increasing the number of tests. Many of the framework instantiations
can be tuned to optimize a memory-quality trade-off, so that fewer tests can be achieved using more
memory. The algorithms developed outperform the currently best known methods when the number
of components ranges from 20 to 60, the number of options for each ranges from 3 to 6, and t-way
interactions are covered for t ∈ {5, 6}. In some cases a reduction in the number of tests by more than
50% is achieved.
Keywords: Covering array, Software interaction testing, Combinatorial construction algorithm
1
Introduction
Real world software and engineered systems are composed of many different components, each with a number
of options, that are required to work together in a variety of circumstances. Components are factors, and
options for a component form the levels of its factor. Although each level for an individual factor can be tested
in isolation, faults in deployed software can arise from interactions among levels of different factors. When
an interaction involves levels of t different factors, it is a t-way interaction. Testing for faults caused by t-way
interactions for every t is generally infeasible, as a result of a combinatorial explosion. However, empirical
research on real world software systems indicates that testing all possible 2-way (or 3-way) interactions
would detect 70% (or 90%) of all faults [25]. Moreover, testing all possible 6-way interactions is sufficient
for detection of 100% of all faults in the systems examined in [25]. Testing all possible t-way interactions for
some 2 ≤ t ≤ 6 is pseudo-exhaustive testing [24], and is accomplished with a combinatorial array known as
a covering array.
Formally, let N, t, k, and v be integers with k ≥ t ≥ 2 and v ≥ 2. A covering array CA(N ; t, k, v) is an
N × k array A in which each entry is from a v-ary alphabet Σ, and for every N × t sub-array B of A and
every x ∈ Σt , there is a row of B that equals x. Then t is the strength of the covering array, k is the number
of factors, and v is the number of levels.
When k is a positive integer, [k] denotes the set {1, . . . , k}. A t-way interaction is {(ci , ai ) : 1 ≤ i ≤
t, ci ∈ [k], ci 6= cj for i 6= j, and ai ∈ Σ}. So an interaction is an assignment of levels from Σ to t of the k
factors. It,k,v denotes the set of all kt v t interactions for given t, k and v. An N × k array A covers the
1
interaction ι = {(ci , ai ) : 1 ≤ i ≤ t, ci ∈ [k], ci 6= cj for i 6= j, and ai ∈ Σ} if there is a row r in A such that
A(r, ci ) = ai for 1 ≤ i ≤ t. When there is no such row in A, ι is not covered in A. Hence a CA(N ; t, k, v)
covers all interactions in It,k,v .
Covering arrays are used extensively for interaction testing in complex engineered systems. To ensure
that all possible combinations of options of t components function together correctly, one needs examine
all possible t-way interactions. When the number of components is k, and the number of different options
available for each component is v, each row of CA(N ; t, k, v) represents a test case. The N test cases
collectively test all t-way interactions. For this reason, covering arrays have been used in combinatorial
interaction testing in varied fields like software and hardware engineering, design of composite materials,
and biological networks [8, 24, 26, 32, 34].
The cost of testing is directly related to the number of test cases. Therefore, one is interested in covering
arrays with the fewest rows. The smallest value of N for which CA(N ; t, k, v) exists is denoted by CAN(t, k, v).
Efforts to determine or bound CAN(t, k, v) have been extensive; see [12, 14, 24, 31] for example. Naturally one
would prefer to determine CAN(t, k, v) exactly. Katona [22] and Kleitman and Spencer [23] independently
showed that
for t = v = 2, the minimum number of rows N in a CA(N ; 2, k, 2) is the smallest N for which
−1
. Exact determination of CAN(t, k, v) for other values of t and v has remained open. However,
k≤ N
N
d2e
some progress has been made in determining upper bounds for CAN(t, k, v) in the general case; for recent
results, see [33].
For practical applications such bounds are often unhelpful, because one needs explicit covering arrays to
use as test suites. Explicit constructions can be recursive, producing larger covering arrays using smaller ones
as ingredients (see [14] for a survey), or direct. Direct methods for some specific cases arise from algebraic,
geometric, or number-theoretic techniques; general direct methods are computational in nature. Indeed
when k is relatively small, the best known results arise from computational techniques [13], and these are
in turn essential for the successes of recursive methods. Unfortunately, the existing computational methods
encounter difficulties as k increases, but is still within the range needed for practical applications. Typically
such difficulties arise either as a result of storage or time limitations or by producing covering arrays that
are too big to compete with those arising from simpler recursive methods.
Cohen [11] discusses commercial software where the number of factors often exceeds 50. Aldaco et al. [1]
analyze a complex engineered system having 75 factors, using a variant of covering arrays. Android [3] uses
a Configuration class to describe the device configuration; there are 17 different configuration parameters
with 3 − 20 different levels. In each of these cases, while existing techniques are effective when the strength
is small, these moderately large values of k pose concerns for larger strengths.
In this paper, we focus on situations in which every factor has the same number of levels. These cases
have been most extensively studied, and hence provide a basis for making comparisons. In practice, however,
often different components have different number of levels, which is captured by extending the notion of a
covering array. A mixed covering array MCA(N ; t, k, (v1 , v2 , . . . , vk )) is an N × k array in which the ith
column contains vi symbols for 1 ≤ i ≤ k. When {i1 , . . . , it } ⊆ {1, . . . , k} is
Qat set of t columns, in the N × t
subarray obtained by selecting columns i1 , . . . , it of the MCA, each of the j=1 vij distinct t-tuples appears
as a row at least once. Although we examine the uniform case in which v1 = · · · = vk , the methods developed
here can all be directly applied to mixed covering arrays as well.
Inevitably, when k > max(t + 1, v + 2), a covering array must cover some interactions more than once,
for if not they are orthogonal arrays [20]. Treating the rows of a covering array in a fixed order, each row
covers some number of interactions not covered by any earlier row. For a variety of known constructions,
the initial rows cover many new interactions, while the later ones cover very few [7]. Comparing this rate
of coverage for a purely random method and for one of the sophisticated search techniques, one finds little
difference in the initial rows, but very substantial differences in the final ones. This suggests strategies to
build the covering array in stages, investing more effort as the number of remaining uncovered interactions
declines.
In this paper we propose a new algorithmic framework for covering array construction, the two-stage
framework. In the first stage, a randomized row construction method builds a specified number of rows to
cover all but at most a specified, small number of interactions. As we see later, by dint of being randomized
2
this uses very little memory. The second stage is a more sophisticated search that adds few rows to cover the
remaining uncovered interactions. We choose search algorithms whose requirements depend on the number
of interactions to be covered, to profit from the fact that few interactions remain. By mixing randomized and
deterministic methods, we hope to retain the fast execution and small storage of the randomized methods,
along with the accuracy of the deterministic search techniques.
We introduce a number of algorithms within the two-stage framework. Some improve upon best known
bounds on CAN(t, k, v) (see [33]) in principle. But our focus is on the practical consequences: The twostage algorithms are indeed quite efficient for higher strength (t ∈ {5, 6}) and moderate number of levels
(v ∈ {3, 4, 5, 6}), when the number of factors k is moderately high (approximately in the range of 20 − 80
depending on value of t and v). In fact, for many combination of t, k and v values the two-stage algorithms
beat the previously best known bounds.
Torres-Jimenez et al. [36] explore a related two-stage strategy. In their first stage, an in-parameter-order
greedy strategy (as used in ACTS [24]) adds a column to an existing array; in their second stage, simulated
annealing is applied to cover the remaining interactions. They apply their methods when t = v = 3, when
the storage and time requirements for both stages remain acceptable. In addition to the issues in handling
larger strengths, their methods provide no a priori bound on the size of the resulting array. In contrast with
their methods, ours provide a guarantee prior to execution with much more modest storage and time.
The rest of the paper is organized as follows. Section 2 reviews algorithmic methods of covering array
construction, specifically the randomized algorithm and the density algorithm. This section contrasts these
two methods and points out their limitations. Then it gives an intuitive answer to the question of why a two
stage based strategy might work and introduces the general two-stage framework. Section 3 introduces some
specific two-stage algorithms. Section 3.1 analyzes and evaluates the naı̈ve strategy. Section 3.2 describes a
two-stage algorithm that combines the randomized and the density algorithm. Section 3.3 introduces graph
coloring based techniques in the second stage. Section 3.4 examines the effect of group action on the size
of the constructed covering arrays. Section 4 compares the results of various two-stage algorithms with the
presently best known sizes. In Section 5 we discuss the Lovász local lemma (LLL) bounds on CAN(t, k, v)
and provide a Moser-Tardos type randomized algorithm for covering array construction that matches the
bound. Although the bound was known [18], the proof was non-constructive, and a constructive algorithm to
match this bound seems to be absent in the literature. We explore potentially better randomized algorithms
for the first stage using LLL based techniques, We also obtain a two-stage bound that improves the LLL
bound for CAN(t, k, v). We conclude the paper in Section 6.
2
Algorithmic construction of covering arrays
Available algorithms for the construction of covering arrays are primarily heuristic in nature; indeed exact
algorithms have succeeded for very few cases. Computationally intensive metaheuristic search methods such
as simulated annealing, tabu search, constraint programming, and genetic algorithms have been employed
when the strength is relatively small or the number of factors and levels is small. These methods have
established many of the best known bounds on sizes of covering arrays [13], but for many problems of
practical size their time and storage requirements are prohibitive. For larger problems, the best available
methods are greedy. The IPO family of algorithms [24] repeatedly adds one column at a time, and
then adds
new rows to ensure complete coverage. In this way, at any point in time, the status of v t k−1
t−1 interactions
may be stored. AETG [10] pioneered a different method, which greedily selects one row at a time to cover
a large number of as-yet-uncovered interactions. They establish that if a row can be chosen that covers the
maximum number, a good a priori bound on the size of the covering array can be computed. Unfortunately
selecting the maximum is NP-hard, and even if one selects the maximum there is no guarantee that the
covering array is the smallest possible
[7], so AETG resorts to a good heuristic selection of the next row by
examining the stored status of v t kt interactions. None of the methods so far mentioned therefore guarantee
to reach an a priori bound. An
extension of the AETG strategy, the density algorithm [5, 6, 15], stores
additional statistics for all v t kt interactions in order to ensure the selection of a good next row, and hence
guarantees to produce an array with at most the precomputed number of rows. Variants of the density
3
Algorithm 1: A randomized algorithm for covering array construction.
Input: t : strength of the covering array, k : number of factors, v : number of levels for each factor
Output: A : a CA(N ;
t, k, v)
log (kt)+t log v
1 Set N :=
;
vt
log
2
3
4
5
6
7
8
9
10
11
12
v t −1
repeat
Construct an N × k array A where each entry is chosen independently and uniformly at random
from a v-ary alphabet;
Set covered := true;
for each interaction ι ∈ It,k,v do
if ι is not covered in A then
Set covered := false;
break;
end
end
until covered = true;
Output A;
algorithm have proved to be most effective for problems of moderately large size. For even larger problems,
pure random approaches have been applied.
To produce methods that provide a guarantee on size, it is natural to focus on the density algorithm
in order to understand its strengths and weaknesses. To do this, we contrast it with a basic randomized
algorithm. Algorithm 1 shows a simple randomized algorithm for covering array construction. The algorithm
constructs an array of a particular size randomly and checks whether all the interactions are covered. It
repeats until it finds an array that covers all the interactions.
log (kt)+t log v
is guaranteed to exist:
A CA(N ; t, k, v) with N =
vt
log
v t −1
Theorem 1. [21, 27, 35] (Stein-Lovász-Johnson (SLJ) bound): Let t, k, v be integers with k ≥ t ≥ 2, and
v ≥ 2. Then as k → ∞,
log kt + t log v
CAN(t, k, v) ≤
t
log vtv−1
In fact, the probability that the N × k array constructed in line 3 of Algorithm 1 is a valid covering array
is high enough that the expected number of times the loop in line 2 is repeated is a small constant.
An alternative strategy is to add rows one by one instead of constructing the full array at the outset. We
start with an empty array, and whenever we add a new row we ensure that it covers at least the expected
number of previously uncovered interactions for a randomly chosen row. The probability that an uncovered
interaction is covered by a random row is 1/v t . If the number of uncovered interactions is u, then by linearity
of expectation, the expected number of newly covered interactions in a randomly chosen row is uv −t . If each
row added covers exactly this expected number, we obtain the same number of rows as the SLJ bound,
realized in Algorithm 1. But because the actual number of newly covered interactions is always an integer,
each added row covers at least duv −t e interactions. This is especially helpful towards the end when the
expected number is a small fraction.
Algorithm 2 follows this strategy. Again the probability that a randomly chosen row covers at least the
expected number of previously uncovered interactions is high enough that the expected number of times the
row selection loop in line 6 of Algorithm 2 is repeated is bounded by a small constant.
We can obtain an upper bound on the size produced by Algorithm 2 by assuming that each new row
added covers exactly duv −t e previously uncovered interactions. This bound is the discrete Stein-Lovász4
Algorithm 2: A randomized algorithm for covering array construction using the discrete SLJ strategy.
Input: t : strength of the covering array, k : number of factors, v : number of levels for each factor
Output: A : a CA(N ; t, k, v)
1 Let A be an empty array;
k t
2 Initialize a table T indexed by all t v interactions, marking every interaction “uncovered”;
3 while there is an interaction marked “uncovered” in T do
4
Let u be the number of interactions marked “uncovered” in T ;
5
Set expectedCoverage := d vut e;
6
repeat
7
Let r be a row of length k where each entry is chosen independently and uniformly at random
from a v-ary alphabet;
8
Let coverage be the number of “uncovered” interactions in T that are covered in row r;
9
until coverage > expectedCoverage;
10
Add r to A;
11
Mark all interactions covered by r as “covered” in T ;
12 end
13 Output A;
Johnson (discrete SLJ) bound. Figure 1 compares the sizes of covering arrays obtained from the SLJ and
the discrete SLJ bounds for different values of k when t = 6 and v = 3. Consider a concrete example, when
t = 5, k = 20, and v = 3. The SLJ bound guarantees the existence of a covering array with 12499 rows,
whereas the discrete SLJ bound guarantees the existence of a covering array with only 8117 rows.
The density algorithm replaces the loop at line 6 of Algorithm 2 by a conditional expectation derandomized method. For fixed v and t the density algorithm selects a row efficiently (time polynomial in k) and
deterministically that is guaranteed to cover at least duv −t e previously uncovered interactions. In practice,
for small values of k the density algorithm works quite well, often covering many more interactions than
the minimum. Many of the currently best known CAN(t, k, v) upper bounds are obtained by the density
algorithm in combination with various post-optimization techniques [13].
However, the practical applicability of Algorithm 2 and the density algorithm is limited by the storage
of the table T , representing each of the kt v t interactions. Even when t = 6, v = 3, and k = 54, there are
18,828,003,285 6-way interactions. This huge memory requirement renders the density algorithm impractical
for rather small values of k when t ∈ {5, 6} and v ≥ 3. We present an idea to circumvent this large
requirement for memory, and develop it in full in Section 3.
2.1
Why does a two stage based strategy make sense?
Compare the two extremes, the density algorithm and Algorithm 1. On one hand, Algorithm 1 does not suffer
from any substantial storage restriction, but appears to generate many more rows than the density algorithm.
On the other hand, the density algorithm constructs fewer rows for small values of k, but becomes impractical
when k is moderately large. One wants algorithms that behave like Algorithm 1 in terms of memory, but
yield a number of rows competitive with the density algorithm.
For t = 6, k = 16, and v = 3, Figure 2 compares the coverage profile for the density algorithm and
Algorithm 1. We plot the number of newly covered interactions for each row in the density algorithm, and
the expected number of newly covered interactions for each row for Algorithm 1. The qualitative features
exhibited by this plot are representative of the rates of coverage for other parameters.
Two key observations are suggested by Figure 2. First, the expected coverage in the initial random rows
is similar to the rows chosen by the density algorithm. In this example, the partial arrays consisting of the
first 1000 rows exhibit similar coverage, yet the randomized algorithm needed no extensive bookkeeping.
Secondly, as later rows are added, the judicious selections of the density algorithm produce much larger
coverage per row than Algorithm 1. Consequently it appears sensible to invest few computational resources
5
4
3.5
x 10
SLJ bound
Discrete SLJ bound
N − number of rows
3
2.5
2
1.5
1
0.5
0
100
200
300
400
500
k
600
700
800
900
1000
Figure 1: Comparison of covering array sizes obtained from SLJ bound and discrete SLJ bound for different
values of k, when t = 6 and v = 3.
9000
Density
Basic Random
Number of newly covered interactions
8000
7000
6000
5000
4000
3000
2000
1000
0
0
2000
4000
6000
Row number
8000
10000
12000
Figure 2: For t = 6, k = 16 and v = 3, the actual number of newly covered interactions of the density
algorithm and the expected number of newly covered interactions in a random array.
6
on the initial rows, while making more careful selections in the later ones. This forms the blueprint of our
general two-stage algorithmic framework shown in Algorithm 3.
Algorithm 3: The general two-stage framework for covering array construction.
Input: t : strength of the required covering array, k : number of factors, v : number of levels for each
factor
Output: A : a CA(N ; t, k, v)
1 Choose a number n of rows and a number ρ of interactions;
// First Stage
0
2 Use a randomized algorithm to construct an n × k array A ;
0
3 Ensure that A covers all but at most ρ interactions;
0
4 Make a list L of interactions that are not covered in A (L contains at most ρ interactions);
// Second Stage
0
5 Use a deterministic procedure to add N − n rows to A to cover all the interactions in L;
6 Output A;
A specific covering array construction algorithm results by specifying the randomized method in the first
stage, the deterministic method in the second stage, and the computation of n and ρ. Any such algorithm
produces a covering array, but we wish to make selections so that the resulting algorithms are practical while
still providing a guarantee on the size of the array. In Section 3 we describe different algorithms from the
two-stage family, determine the size of the partial array to be constructed in the first stage, and establish
upper bound guarantees. In Section 4 we explore how good the algorithms are in practice.
3
Two-stage framework
For the first stage we consider two methods:
Rand
MT
the basic randomized algorithm
the Moser-Tardos type algorithm
We defer the development of method MT until Section 5. Method Rand uses a simple variant of Algorithm
1, choosing a random n × k array.
For the second stage we consider four methods:
Naive
Greedy
Den
Col
the
the
the
the
naı̈ve strategy, one row per uncovered interaction
online greedy coloring strategy
density algorithm
graph coloring algorithm
Using these abbreviations, we adopt a uniform naming convention for the algorithms: TS hA, Bi is the algorithm in which A is used in the first stage, and B is used in the second stage. For example, TS hMT, Greedyi
denotes a two-stage algorithm where the first stage is a Moser-Tardos type randomized algorithm and the
second stage is a greedy coloring algorithm. Later when the need arises we refine these algorithm names.
3.1
One row per uncovered interaction in the second stage (TS hRand, Naivei)
In the second stage each of the uncovered interactions after the first stage is covered using a new row.
Algorithm 4 describes the method in more detail.
This simple strategy improves on the basic randomized strategy when n is chosen judiciously. For
example, when t = 6, k = 54 and v = 3, Algorithm 1 constructs a covering array with 17, 236 rows. Figure
3 plots an upper bound on the size of the covering array against the number n of rows in the partial array.
7
Algorithm 4: Naı̈ve two-stage algorithm (TS hRand, Naivei).
Input: t : strength of the covering array, k : number of factors, v : number of levels for each factor
Output: A : a CA(N ; t, k, v)
t
log (kt)+t log v+log log vtv−1
;
1 Let n :=
vt
log
1
v t −1
2
Let ρ =
3
repeat
Let A be an n × k array where each entry is chosen independently and uniformly at random from
a v-ary alphabet;
Let uncovNum := 0 and unCovList be an empty list of interactions;
Set covered := true;
for each interaction ι ∈ It,k,v do
if ι is not covered in A then
Set uncovNum :=uncovNum+1;
Add ι to unCovList;
if uncovNum > ρ then
Set covered := false;
break;
end
end
end
until covered= true;
for each interaction ι ∈uncovList do
Add a row to A that covers ι;
end
Output A;
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
log
vt
v t −1
;
8
4
Total number of rows in the Covering array
1.75
x 10
1.7
1.65
1.6
1.55
1.5
1.45
1.4
1.35
1.3
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
n −− number of rows in the partial array of the first stage
1.8
4
x 10
Figure 3: An upper bound on the size of the covering array against n, the size of the partial array constructed
in the first stage when t = 6, k = 54, and v = 3, with one new row added per uncovered interaction in the
second stage. The minimum size of 13, 162 is obtained when n = 12, 402. Algorithm 1 requires 17, 236 rows,
and the currently best known covering array has 17, 197 rows.
9
The smallest covering array is obtained when n = 12, 402 which, when completed, yields a covering array
with at most 13, 162 rows—a big improvement over Algorithm 1.
A theorem from [33] tells us the optimal value of n in general:
Theorem 2. [33] Let t, k, v be integers with k ≥ t ≥ 2, and v ≥ 2. Then
t
log kt + t log v + log log vtv−1 + 1
.
CAN(t, k, v) ≤
t
log vtv−1
log (kt)+t log v+log log
t
log vtv−1
vt
v t −1
. The expected number of uncovered
The bound is obtained by setting n =
t
interactions is exactly ρ = 1/ log vtv−1 .
Figure 4 compares SLJ, discrete SLJ and two-stage bounds for k ≤ 100, when t = 6 and v = 3. The
two-stage bound does not deteriorate in comparison to discrete SLJ bound as k increases; it consistently
takes only 307-309 more rows. Thus when k = 12 the two-stage bound requires only 6% more rows and
when k = 100 only 2% more rows than the discrete SLJ bound.
4
2.2
x 10
SLJ bound
Discrete SLJ bound
Two−stage bound
2
N − number of rows
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
10
20
30
40
50
60
70
80
90
100
k
Figure 4: Comparison of covering array sizes obtained from SLJ bound, discrete SLJ bound and two-stage
bound for k ≤ 100, when t = 6 and v = 3. In this range of k values the two-stage bound requires 307-309
more rows than the discrete SLJ bound, that is, 2-6% more rows.
To ensure that the loop in line 7 of Algorithm 4 does not repeat too many times we need to know the
probability with which a random n × k array leaves at most ρ interactions uncovered. Using Chebyshev’s
inequality and the second moment method developed in [2, Chapter 4], we next show that in a random n ×
nk
array the number of uncovered interactions is almost always close to its expectation, i.e. kt v t 1 − v1t .
Substituting the value of n from line 1, this expected value is equal to µ, as in line 2. Therefore, the probability
that a random n × k array covers the desired number of interactions is constant, and the expected number
of times the loop in line 7 is repeated is also a constant (around 2 in practice).
10
Because the theory of the second moment method is developed
considerable detail in [2], here we briefly
Pin
m
mention the relevant concepts and results. Suppose that X = i=1 Xi , where Xi is the indicator random
variable for event Ai for 1 ≤ i ≤ m. For indices i, j, we write i ∼ j if i 6= j and the events Ai , Aj are not
independent. Also suppose that X1 , . . . , Xm are symmetric, i.e. for every i 6= j there is a measure
preserving
P
mapping of the underlying probability space that sends event Ai to event Aj . Define ∆∗ = j∼i Pr [Aj |Ai ].
Then by [2, Corollary 4.3.4]:
Lemma 3. [2] If E[X] → ∞ and ∆∗ = o(E[X]) then X ∼ E[X] almost always.
In our case, Ai denotes the event that the ith interaction is not covered in a n × k array where each entry
n
is chosen independently and uniformly at random from a v-ary alphabet. Then Pr[Xi ] = 1 − v1t . Because
n
there are kt v t interactions in total, by linearity of expectation, E[X] = kt v t 1 − v1t , and E[X] → ∞ as
k → ∞.
Distinct events Ai and Aj are independent if the ith and jth interactions share no column. Therefore,
P
P
k
the event Ai is not independent of at most t t−1
other events Aj . So ∆∗ = j∼i Pr [Aj |Ai ] ≤ j∼i 1 ≤
k
t t−1
= o(E[X]) when v and t are constants. By Lemma 3, the number of uncovered interactions in a random
n × k array is close to the expected number of uncovered interactions. This guarantees that Algorithm 4 is
an efficient randomized algorithm for constructing covering arrays with a number of rows upper bounded by
Theorem 2.
In keeping with the general two-stage framework, Algorithm 4 does not store the coverage status of
each interaction. We only need store the interactions that are uncovered in A, of which there are at most
1
≈ v t . This quantity depends only on v and t and is independent of k, so is effectively a
ρ =
t
log vtv−1
constant that is much smaller than kt v t , the storage requirement for the density algorithm. Hence the
algorithm can be applied to a higher range of k values.
Although Theorem 5 provides asymptotically tighter bounds than Theorem 2, in a range of k values that
are relevant for practical application, Theorem 2 provides better results. Figure 5 compares the bounds on
CAN(t, k, v) with the currently best known results.
4
3
6
x 10
2.5
Best known
Two−stage (simple)
GSS bound
2
2
N − number of rows
N − number of rows
2.5
1.5
1
1.5
1
0.5
0.5
0
0
x 10
Best known
Two−stage (simple)
GSS bound
10
20
30
40
50
60
70
80
90
k
0
5
10
15
20
25
30
35
40
45
50
k
(a) t = 6, v = 3
(b) t = 6, v = 6
Figure 5: Comparison of GSS bound and two-stage bound with the currently best known results
3.2
The density algorithm in the second stage (TS hRand, Deni)
Next we apply the density algorithm in the second stage. Figure 6 plots an upper bound on the size of the
covering array against the size of the partial array constructed in the first stage when the density algorithm
is used in the second stage, and compares it with TS hRand, Naivei. The size of the covering array decreases
11
as n decreases. This is expected because with smaller partial arrays, more interactions remain for the second
stage to be covered by the density algorithm. In fact if we cover all the interactions using the density
algorithm (as when n = 0) we would get an even smaller covering array. However, our motivation was
precisely to avoid doing that. Therefore, we need a ”cut-off” for the first stage.
4
Total number of rows in the Covering array
1.9
x 10
Basic Two−stage
Two−stage with density in second stage
1.8
1.7
1.6
1.5
1.4
1.3
1.2
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
n −− number of rows in the partial array of the first stage
1.8
4
x 10
Figure 6: Comparison of covering array sizes from two-stage algorithms with Den and Naive in the second
stage. With Den there is no minimum point in the curve; the size of the covering array keeps decreasing as
we leave more uncovered interactions for the second stage.
We are presented with a trade-off. If we construct a smaller partial array in the first stage, we obtain a
smaller covering array overall. But we then pay for more storage and computation time for the second stage.
To appreciate the nature of this trade-off, look at Figure 7, which plots an upper bound on the covering
array size and the number of uncovered interactions in the first stage against n. The improvement in the
covering array size plateaus after a certain point. The three horizontal lines indicate ρ (≈ v t ), 2ρ and 3ρ
uncovered interactions in the first stage. (In the naı̈ve method of Section 3.1, the partial array after the first
stage leaves at most ρ uncovered interactions.) In Figure 7 the final covering array size appears to plateau
when the number of uncovered interactions left by the first stage is around 2ρ. After that we see diminishing
returns — the density algorithm needs to cover more interactions in return for a smaller improvement in the
covering array size.
Let r be the maximum number of interactions allowed to remain uncovered after the first stage. Then r
can be specified in the two-stage algorithm. To accommodate this, we denote by TS hA, B; ri the two-stage
algorithm where A is the first stage strategy, B is the second stage strategy, and r is the maximum number
of uncovered interactions after the first stage. For example, TS hRand, Den; 2ρi applies the basic randomized
algorithm in the first stage to cover all but at most 2ρ interactions, and the density algorithm to cover the
remaining interactions in the second stage.
3.3
Coloring in the second stage (TS hRand, Coli and TS hRand, Greedyi)
Now we describe strategies using graph coloring in the second stage. Construct a graph G = (V, E), the
incompatibility graph, in which V is the set of uncovered interactions, and there is an edge between two
12
Number of rows / Number of uncovered interactions
18000
16000
Num. of rows in the completed CA
Num. of uncovered interaction in first stage
14000
12000
10000
8000
6000
4000
2000
0
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
4
n −− number of rows in the partial array of the first stage
x 10
Figure 7: Final covering array size against the number of uncovered interactions after the first stage. As the
size n of the partial array decreases, the number of uncovered interactions in the first stage increases. Den is
used in the second stage. From bottom to top, the green lines denote ρ, 2ρ, and 3ρ uncovered interactions.
interactions exactly when they share a column in which they have different symbols. A single row can cover
a set of interactions if and only if it forms an independent set in G. Hence the minimum number of rows
required to cover all interactions of G is exactly its chromatic number χ(G), the minimum number of colors
in a proper coloring of G. Graph coloring is an NP-hard problem, so we employ heuristics to bound the
chromatic number. Moreover, G only has vertices for the uncovered interactions after the first stage, so is
size is small relative to the total number of interactions.
The expected number of edges in the incompatibility
graph after
n choosing n rows uniformly at random
k t Pt
t
t k−t
1
1 n
1
t−i
is γ = 2 t v
) 1 − vt
1 − (vt −vt−i ) . Using the elementary upper bound on
i=1 i t−i (v − v
q
the chromatic number χ ≤ 12 + 2m + 14 , where m is the number of edges [16, Chapter 5.2], we can surely
q
cover the remaining interactions with at most 12 + 2m + 14 rows.
The actual number of edges m that remain after the first stage is a random variable with mean γ. In
principle, the first stage could be repeatedly applied until m ≤ γ, so we call m = γ the optimistic estimate.
To ensure that the first stage is expected to be run a small constant number of times, we increase the
estimate. With probability more than 1/2 the incompatibility graph has m ≤ 2γ edges, so m = 2γ is the
conservative estimate.
For t = 6, k = 56, and v = 3, Figure 8 shows the effect on the minimum number of rows when the bound
on the chromatic number in the second stage is used, for the conservative or optimistic estimates. The Naı̈ve
method is plotted for comparison. Better coloring bounds shift the minima leftward, reducing the number
of rows produced in both stages.
Thus far we have considered bounds on the chromatic number. Better estimation of χ(G) is complicated
by the fact that we do not have much information about the structure of G until the first stage is run. In
practice, however, G is known after the first stage and hence an algorithmic method to bound its chromatic
number can be applied. Because the number of vertices in G equals the number of uncovered interactions
13
4
2.2
x 10
Conservative estimate
Optimistic estimate
Simple
Number of rows required
2
1.8
1.6
1.4
1.2
1
1
1.1
1.2
1.3
1.4
N
1.5
1.6
1.7
1.8
4
x 10
Figure 8: Size of the partial array vs. size of the complete CA. t = 6, k = 56, v = 3. Stein-Lovász-Johnson
bound requires 17, 403 rows, discrete Stein-Lovász-Johnson bound requires 13, 021 rows. Simple estimate
for the two stage algorithm is 13, 328 rows, conservative estimate assuming m = 2γ is 12, 159 rows, and
optimistic estimate assuming m = γ is 11, 919 rows. Even the conservative estimate beats the discrete
Stein-Lovász-Johnson bound.
14
after the first stage, we encounter the same trade-off between time and storage, and final array size, as seen
earlier for density. Hence we again parametrize by the expected number of uncovered interactions in the
first stage.
We employ two different greedy algorithms to color the incompatibility graph. In method Col we first
construct the incompatibility graph G after the first stage. Then we apply the commonly used smallest last
order heuristic to order the vertices for greedy coloring: At each stage, find a vertex vi of minimum degree
in Gi , order the vertices of Gi − vi , and then place vi at the end. More precisely, we order the vertices of G
as v1 , v2 , . . . , vn , such that vi is a vertex of minimum degree in Gi , where Gi = G − {vi+1 , . . . , vn }. A graph
is d-degenerate if all of its subgraphs have a vertex with degree at most d. When G is d-degenerate but not
(d − 1)-degenerate, the Coloring number col(G) is d + 1. If we then greedily color the vertices with the first
available color, at most col(G) colors are used.
In method Greedy we employ an on-line, greedy approach that colors the interactions as they are discovered in the first stage. In this way, the incompatibility graph is never constructed. We instead maintain a set
of rows. Some entries in rows are fixed to a specific value; others are flexible to take on any value. Whenever
a new interaction is found to be uncovered in the first stage, we check if any of the rows is compatible with
this interaction. If such a row is found then entries in the row are fixed so that the row now covers the
interaction. If no such row exists, a new row with exactly t fixed entries corresponding to the interaction is
added to the set of rows. This method is much faster than method Col in practice.
3.4
Using group action
Covering arrays that are invariant under the action of a permutation group on their symbols can be easier
to construct and are often smaller [15]. Direct and computational constructions using group actions are
explored in [9, 28]. Sarkar et al. [33] establish the asymptotically tightest known bounds on CAN(t, k, v)
using group actions. In this section we explore the implications of group actions on two-stage algorithms.
Let Γ be a permutation group on the set of symbols. The action of this group partitions the set of
t-way interactions into orbits. We construct an array A such that for every orbit, at least one row covers
an interaction from that orbit. Then we develop the rows of A over Γ to obtain a covering array that is
invariant under the action of Γ. Effort then focuses on covering all the orbits of t-way interactions, instead
of the individual interactions.
If Γ acts sharply transitively
on the set of symbols
(for example, if Γ is a cyclic group of order v) then
the action of Γ partitions kt v t interactions into kt v t−1 orbits of length v each. Following
the lines of the
v t−1
+1
log (kt)+(t−1) log v+log log t−1
−1
v
that covers at
proof of Theorem 2, there exists an n × k array with n =
v t−1
log
v t−1 −1
least one interaction from each orbit. Therefore,
log
CAN(t, k, v) ≤ v
k
t
+ (t − 1) log v + log log
v t−1
log vt−1
−1
v t−1
v t−1 −1
+1
.
(1)
Similarly, we can employ a Frobenius group. When v is a prime power, the Frobenius group is the group
of permutations of Fv of the form {x 7→ ax + b : a, b ∈ Fv , a 6= 0}. The action of the Frobenius group
t−1
partitions the set of t-tuples on v symbols into v v−1−1 orbits of length v(v − 1) (full orbits) each and 1 orbit
of length v (a short orbit). The short orbit consists of tuples of the form (x1 , . . . , xt ) where x1 = . . . = xt .
Therefore, we can obtain a covering array by first constructing an array that covers all the full orbits, and
then developing all the rows over the Frobenius group and adding v constant rows. Using the two stage
strategy in conjunction with the Frobenius group action we obtain:
t−1
v t−1
log kt + log v v−1−1 + log log vt−1
−v+1 + 1
CAN(t, k, v) ≤ v(v − 1)
+ v.
(2)
t−1
v
log vt−1
−v+1
15
4
1.5
x 10
N − number of rows
1.45
Two−stage (simple)
Two−stage (cyclic group action)
Two−stage (Frobenius group action)
1.4
1.35
1.3
1.25
50
55
60
65
70
75
k
Figure 9: Comparison of the simple two-stage bound with the cyclic and the Frobenius two-stage bounds.
t = 6, v = 3 and 50 ≤ k ≤ 75. Group action reduces the required number of rows slightly.
Figure 9 compares the simple two-stage bound with the cyclic and Frobenius two-stage bounds. For
t = 6, v = 3 and 12 ≤ k ≤ 100, the cyclic bound requires 7-21 (on average 16) fewer rows than the simple
bound. In the same range the Frobenius bound requires 17 − 51 (on average 40) fewer rows.
Group action can be applied in other methods for the second stage as well. Colbourn [15] incorporates
group action into the density algorithm, allowing us to apply method Den in the second stage.
Greedy extends easily to use group action, as we do not construct an explicit incompatibility graph.
Whenever we fix entries in a row to cover an uncovered orbit, we commit to a specific orbit representative.
However, applying group action to the incompatibility graph coloring for Col is more complicated. We
need to modify the definition of the incompatibility graph for two reasons. First the vertices no longer
represent uncovered interactions, but rather uncovered orbits of interaction. Secondly, and perhaps more
importantly, pairwise compatibility between every two orbits in a set no longer implies mutual compatibility
among all orbits in the set.
One approach is to form a vertex for each uncovered orbit, placing an edge between two when they share
a column. Rather than the usual coloring, however, one asks for a partition of the vertex set into classes so
that every class induces an acyclic subgraph. Problems of this type are generalized graph coloring problems
[4]. Within each class of such a vertex partition, consistent representatives of each orbit can be selected to
form a row; when a cycle is present, this may not be possible. Unfortunately, heuristics for solving these
types of problems appear to be weak, so we adopt a second approach. As we build the incompatibility
graph, we commit to specific orbit representatives. When a vertex for an uncovered orbit is added, we check
its compatibility with the orbit representatives chosen for the orbits already handled with which it shares
columns; we commit to an orbit representative and add edges to those with which it is now incompatible.
Once completed, we have a (standard) coloring problem for the resulting graph.
Because group action can be applied using each of the methods for the two stages, we extend our naming
to TS hA, B; r, Γi, where Γ can be Trivial (i.e. no group action), Cyclic, or Frobenius.
16
4
Computational results
Figure 5 indicates that even a simple two-stage bound can improve on best known covering array numbers.
Therefore we investigate the actual performance of our two-stage algorithms for covering arrays of strength
5 and 6.
First we present results for t = 6, when v ∈ {3, 4, 5, 6} and no group action is assumed. Table 1 shows the
results for different v values. In each case we select the range of k values where the two-stage bound predicts
smaller covering arrays than
previously known best ones, setting the maximum number of uncovered
tthe
v
interactions as ρ = 1/ log vt −1 ≈ v t . For each value of k we construct a single partial array and then run
the different second stage algorithms on it consecutively. In this way all the second stage algorithms cover
the same set of uncovered interactions.
The column tab lists the best known CAN(t, k, v) upper bounds from [13]. The column bound shows the
upper bounds obtained from the two-stage bound (2). The columns naı̈ve, greedy, col and den show results
obtained from running the TS hRand, Naive; ρ, Triviali, TS hRand, Greedy; ρ, Triviali, TS hRand, Col; ρ, Triviali
and TS hRand, Den; ρ, Triviali algorithms, respectively.
The naı̈ve method always finds a covering array that is smaller than the two-stage bound. This happens
because we repeat the first stage of Algorithm 4 until the array has fewer than v t uncovered interactions.
(If the first stage were not repeated, the algorithm still produce covering arrays that are not too far from
the bound.) For v = 3 Greedy and Den have comparable performance. Method Col produces covering arrays
that are smaller. However, for v ∈ {4, 5, 6} Den and Col are competitive.
Table 2 shows the results obtained by the different second stage algorithms when the maximum number
of uncovered interactions in the first stage is set to 2ρ and 3ρ respectively. When more interactions are
covered in the second stage, we obtain smaller arrays as expected. However, the improvement in size does
not approach 50%. There is no clear winner.
Next we investigate the covering arrays that are invariant under the action of a cyclic group. In Table 3 the
column bound shows the upper bounds from Equation (1). The columns naı̈ve, greedy, col and den show
results obtained from running TS hRand, Naive; ρ, Cyclici, TS hRand, Greedy; ρ, Cyclici, TS hRand, Col; ρ, Cyclici
and TS hRand, Den; ρ, Cyclici, respectively.
Table 4 presents results for cyclic group action based algorithms when the number of maximum uncovered
interactions in the first stage is set to 2ρ and 3ρ respectively.
For the Frobenius group action, we show results only for v ∈ {3, 5} in Table 5. The column bound shows
the upper bounds obtained from Equation (2).
Table 6 presents results for Frobenius group action algorithms when the number of maximum uncovered
interactions in the first stage is 2ρ or 3ρ.
Next we present a handful of results when t = 5. In the cases examined, using the trivial group action
is too time consuming to be practical. However, the cyclic or Frobenius cases are feasible. Tables 7 and 8
compare two stage algorithms when the number of uncovered interactions in the first stage is at most 2ρ.
In almost all cases there is no clear winner among the three second stage methods. Methods Den and
Greedy are, however, substantially faster and use less memory than method Col; for practical purposes they
would be preferred.
All code used in this experimentation is available from the github repository
https://github.com/ksarkar/CoveringArray
under an open source GPLv3 license.
5
Limited dependence and the Moser-Tardos algorithm
Here we explore a different randomized algorithm that produces smaller covering arrays than Algorithm 1.
When k > 2t, there are interactions that share no column. The events of coverage of such interactions are
independent. Moser et al. [29, 30] provide an efficient randomized construction method that exploits this
17
k
tab
bound
53
54
55
56
57
13021
14155
17161
19033
20185
13076
13162
13246
13329
13410
39
40
41
42
43
44
68314
71386
86554
94042
99994
104794
65520
66186
66834
67465
68081
68681
31
32
33
34
35
36
233945
258845
281345
293845
306345
356045
226700
229950
233080
236120
239050
241900
17
18
19
20
21
22
23
24
25
26
27
506713
583823
653756
694048
783784
844834
985702
1035310
1112436
1146173
1184697
486310
505230
522940
539580
555280
570130
584240
597660
610460
622700
634430
naı̈ve
greedy
t = 6, v = 3
13056
12421
13160
12510
13192
12590
13304
12671
13395
12752
t = 6, v = 4
65452
61913
66125
62573
66740
63209
67408
63819
68064
64438
68556
65021
t = 6, v = 5
226503 213244
229829 216444
232929 219514
235933 222516
238981 225410
241831 228205
t = 6, v = 6
486302 449950
505197 468449
522596 485694
539532 502023
555254 517346
569934 531910
584194 545763
597152 558898
610389 571389
622589 583473
634139 594933
col
den
12415
12503
12581
12665
12748
12423
12512
12591
12674
12757
61862
62826
63160
64077
64935
65739
61886
62835
63186
64082
64907
65703
212942
217479
219215
222242
226379
230202
212940
217326
219241
222244
226270
229942
448922
467206
484434
500788
516083
530728
544547
557917
570316
582333
593857
447864
466438
483820
500194
515584
530242
548307
557316
569911
582028
593546
Table 1: Comparison of different TS hRand, −; ρ, Triviali algorithms.
18
k
greedy
2ρ
col
53
54
55
56
57
11968
12135
12286
12429
12562
11958
12126
12129
12204
12290
39
40
41
42
43
44
59433
60090
60715
61330
61936
62530
59323
60479
61527
62488
61839
62899
31
32
33
34
35
36
204105
207243
210308
213267
216082
218884
203500
206659
209716
212675
215521
218314
17
18
19
20
21
22
23
24
25
26
27
425053
443236
460315
476456
491570
505966
519611
532612
544967
556821
568135
-
den
t = 6, v
11968
12050
12131
12218
12296
t = 6, v
59326
59976
60615
61242
61836
62428
t = 6, v
203302
206440
209554
212508
215389
218172
t = 6, v
420333
438754
455941
472198
487501
502009
515774
528868
541353
553377
564827
greedy
=3
11716
11804
11877
11961
12044
=4
58095
58742
59369
59974
60575
61158
=5
199230
202342
205386
208285
211118
213872
=6
412275
430402
447198
463071
478269
492425
505980
518746
531042
542788
554052
3ρ
col
den
11705
11787
11875
12055
12211
11708
11790
11872
11950
12034
57951
58583
59867
61000
60407
61004
57888
58544
59187
59796
60393
60978
198361
201490
204548
-
197889
201068
204107
207060
209936
212707
-
405093
423493
440532
456725
471946
486306
500038
513047
525536
537418
548781
Table 2: Comparison of TS hRand, −; 2ρ, Triviali and TS hRand, −; 3ρ, Triviali algorithms.
19
k
tab
bound
53
54
55
56
57
13021
14155
17161
19033
20185
13059
13145
13229
13312
13393
k
39
40
41
42
43
44
tab
68314
71386
86554
94042
99994
104794
bound
65498
66163
66811
67442
68057
68658
31
32
33
34
35
36
226000
244715
263145
235835
238705
256935
226680
229920
233050
236090
239020
241870
17
18
19
20
21
22
23
24
25
26
27
506713
583823
653756
694048
783784
844834
985702
1035310
1112436
1146173
1184697
486290
505210
522910
539550
555250
570110
584210
597630
610430
622670
624400
naı̈ve
greedy
t = 6, v = 3
13053
12405
13119
12489
13209
12573
13284
12660
13368
12744
t = 6, v = 4
naı̈ve
greedy
65452
61896
66080
62516
66740
63184
67408
63800
68032
64408
68556
64988
t = 6, v = 5
226000 213165
229695 216440
233015 219450
235835 222450
238705 225330
241470 228140
t = 6, v = 6
485616 449778
504546 468156
522258 485586
539280 501972
554082 517236
569706 531852
583716 545562
597378 558888
610026 571380
622290 583320
633294 594786
col
den
12405
12543
12663
12651
12744
12411
12546
12663
12663
12750
col
61860
62820
63144
63780
64692
64964
den
61864
62784
63152
63784
64680
64976
212945
217585
221770
222300
225130
229235
212890
217270
221290
222210
225120
229020
448530
467232
490488
500880
521730
530832
549660
557790
575010
582546
598620
447732
466326
488454
500172
519966
530178
548196
557280
573882
582030
597246
Table 3: Comparison of TS hRand, −; ρ, Cyclicialgorithms.
20
k
greedy
2ρ
col
53
54
55
56
57
11958
12039
12120
12204
12276
11955
12027
12183
12342
12474
39
40
41
42
43
44
59412
60040
60700
61320
61908
62512
59336
59996
61156
62196
63192
64096
31
32
33
34
35
36
204060
207165
207165
213225
216050
218835
203650
209110
209865
212830
217795
218480
17
18
19
20
21
22
23
24
25
26
27
424842
443118
460014
476328
491514
505884
519498
532368
544842
543684
568050
422736
440922
457944
474252
489270
503580
517458
530340
542688
543684
566244
den
t = 6, v
11958
12036
12195
12324
12450
t = 6, v
59304
59964
61032
61976
62852
63672
t = 6, v
203265
208225
209540
212510
217070
218155
t = 6, v
420252
438762
455994
472158
487500
501852
515718
528828
541332
543684
564756
greedy
=3
11700
11790
11862
11949
12027
=4
58076
58716
59356
59932
60568
61152
=5
199180
202255
205380
208225
211080
213770
=6
411954
430506
447186
463062
478038
492372
505824
518700
530754
542664
553704
3ρ
col
den
11691
11874
12057
11937
12021
11694
11868
12027
11943
12024
57976
58616
59252
59840
61124
61048
57864
58520
59160
59760
60904
60988
198455
204495
204720
207790
213425
213185
197870
203250
204080
207025
212040
212695
409158
427638
456468
460164
486180
489336
502806
515754
538056
539922
560820
405018
423468
449148
456630
479970
486264
500040
512940
532662
537396
555756
Table 4: Comparison of TS hRand, −; 2ρ, Cyclici and TS hRand, −; 3ρ, Cyclici algorithms.
21
k
tab
bound
53
54
55
56
57
13021
14155
17161
19033
20185
13034
13120
13203
13286
13366
31
32
33
34
35
36
233945
258845
281345
293845
306345
356045
226570
229820
232950
235980
238920
241760
naı̈ve
greedy
t = 6, v = 3
13029
12393
13071
12465
13179
12561
13245
12633
13365
12723
t = 6, v = 5
226425 213025
229585 216225
232725 219285
234905 222265
238185 225205
241525 227925
col
den
12387
12513
12549
12627
12717
12393
12531
12567
12639
12735
212865
216085
219205
223445
227445
231145
212865
216065
219145
223265
227065
230645
Table 5: Comparison of TS hRand, −; ρ, Frobeniusi algorithms.
k
greedy
2ρ
col
53
54
55
56
57
70
75
80
85
90
11931
12021
12105
12171
12255
13167
13473
13773
14031
14289
11919
12087
12237
12171
12249
13155
13473
13767
14025
14283
31
32
33
34
35
36
50
55
60
65
203785
206965
209985
213005
215765
218605
250625
259785
268185
275785
203485
208965
209645
214825
215545
218285
250365
259625
268025
275665
den
greedy
t = 6, v = 3
11931
11700
12087
11790
12231
11862
12183
11949
12255
12027
13179
13479
13779
14037
14301
t = 6, v = 5
203225 198945
208065 201845
209405 205045
214145 208065
215265 210705
218025 213525
250325
259565
267945
275665
-
3ρ
col
den
11691
11874
12057
11937
12021
-
11694
11868
12027
11943
12024
-
198445
204505
209845
207545
210365
213105
-
197825
203105
207865
206985
209885
212645
-
Table 6: Comparison of TS hRand, −; 2ρ, Frobeniusi and TS hRand, −; 3ρ, Frobeniusi algorithms.
22
k
67
68
69
70
71
tab
59110
60991
60991
60991
60991
greedy
48325
48565
48765
49005
49245
col
48285
48565
49005
48985
49205
den
48305
48585
48985
49025
49245
Table 7: Comparison of TS hRand, −; 2ρ, Frobeniusi algorithms. t = 5, v = 5
k
49
50
51
52
53
tab
122718
125520
128637
135745
137713
greedy
108210
109014
109734
110556
111306
col
108072
108894
110394
110436
111180
den
107988
108822
110166
110364
111120
Table 8: Comparison of TS hRand, −; 2ρ, Cyclici algorithms. t = 5, v = 6
limited dependence. Specializing their method to covering arrays, we obtain Algorithm 5. For the specified
value of N in the algorithm it is guaranteed that the expected number of times the loop in line 3 of Algorithm
5 is repeated is linearly bounded in k (See Theorem 1.2 of [30]).
The upper bound on CAN(t, k, v) guaranteed by Algorithm 5 is obtained by applying the Lovász local
lemma (LLL).
Lemma 4. (Lovász local lemma; symmetric case) (see [2]) Let A1 , A2 , . . . , An events in an arbitrary probability space. Suppose that each event Ai is mutually independent of a set of all other events Aj except for at
most d, and that Pr[Ai ] ≤ p for all 1 ≤ i ≤ n. If ep(d + 1) ≤ 1, then Pr[∩ni=1 Āi ] > 0.
The symmetric version of Lovász local lemma provides an upper bound on the probability of a “bad”
event in terms of the maximum degree of a bad event in a dependence graph, so that the probability that
all the bad events are avoided is non zero. Godbole, Skipper, and Sunley [18] apply Lemma 4 essentially to
obtain the bound on CAN(t, k, v) in line 1 of Algorithm 5.
Theorem 5. [18] Let t, v and k ≥ 2t be integers with t, v ≥ 2. Then
n
o
+ t log v + 1
log kt − k−t
t
CAN (t, k, v) ≤
t
log vtv−1
The bound on the size of covering arrays obtained from Theorem 5 is asymptotically tighter than the
one obtained from Theorem 1. Figure 10 compares the bounds for t = 6 and v = 3.
The original proof of LLL is essentially non-constructive and does not immediately lead to a polynomial
time construction algorithm for covering arrays satisfying the bound of Theorem 5. Indeed no previous
construction algorithms appear to be based on it. However the Moser-Tardos method of Algorithm 5 does
provide a construction algorithm running in expected polynomial time. For sufficiently large values of k
Algorithm 5 produces smaller covering arrays than the Algorithm 1.
But the question remains: Does Algorithm 5 produce smaller covering arrays than the currently best
known results within the range that it can be effectively computed? Perhaps surprisingly, we show that
the answer is affirmative. In Algorithm 5 we do not need to store the coverage information of individual
interactions in memory because each time an uncovered interaction is encountered we re-sample the columns
involved in that interaction and start the check afresh (checking the coverage in interactions in the same
order each time). Consequently, Algorithm 5 can be applied for larger values of k than the density algorithm.
23
Algorithm 5: Moser-Tardos type algorithm for covering array construction.
Input: t : strength of the covering array, k : number of factors, v : number of levels for each factor
Output: A : a CA(N ; t, k, v)
log{(kt)−(k−t
log v+1
)}+t.
t
;
1 Let N :=
vt
log
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
v t −1
Construct an N × k array A where each entry is chosen independently and uniformly at random from
a v-ary alphabet;
repeat
Set covered := true;
for each interaction ι ∈ It,k,v do
if ι is not covered in A then
Set covered := false;
Set missing-interaction := ι;
break;
end
end
if covered = false then
Choose all the entries in the t columns involved in missing-interaction independently and
uniformly at random from the v-ary alphabet;
end
until covered = true;
Output A;
Smaller covering arrays can be obtained by exploiting a group action using LLL, as shown in [33]. Table
9 shows the sizes of the covering arrays constructed by a variant of Algorithm 5 that employs cyclic and
Frobenius group actions. While this single stage algorithm produces smaller arrays than the currently best
known [13], these are already superseded by the two-stage based algorithms.
k
56
57
58
59
60
tab
19033
20185
23299
23563
23563
MT
16281
16353
16425
16491
16557
(a) Frobenius. t = 6, v = 3
k
44
45
46
47
48
tab
411373
417581
417581
423523
423523
MT
358125
360125
362065
363965
365805
(b) Frobenius. t = 6, v = 5
k
25
26
27
28
29
tab
1006326
1040063
1082766
1105985
1149037
MT
1020630
1032030
1042902
1053306
1063272
(c) Cyclic. t = 6, v = 6
Table 9: Comparison of covering array size from Algorithm 5 (MT) with the best known results [13] (tab).
5.1
Moser-Tardos type algorithm for the first stage
The linearity of expectation arguments used in the SLJ bounds permit one to consider situations in which
a few of the “bad” events are allowed to occur, a fact that we exploited in the first stage of the algorithms
thus far. However, the Lovász local lemma does not address this situation directly. The conditional Lovász
local lemma (LLL) distribution, introduced in [19], is a very useful tool.
Lemma 6. (Conditional LLL distribution; symmetric case) (see [2, 33]) Let A = {A1 , A2 , . . . , Al } be a set
of l events in an arbitrary probability space. Suppose that each event Ai is mutually independent of a set of
all other events Aj except for at most d, and that Pr[Ai ] ≤ p for all 1 ≤ i ≤ l. Also suppose that ep(d+1) ≤ 1
(Therefore, by LLL (Lemma 4) Pr[∩li=1 Āi ] > 0). Let B ∈
/ A be another event in the same probability space
24
5
10
N − number of rows
SLJ bound
GSS bound
4
10
3
10
1
10
2
3
10
10
k
4
10
5
10
Figure 10: Comparison of SLJ (Theorem 1) and GSS (Theorem 5) bounds for t = 6 and v = 3. The graph
is plotted in log-log scale to highlight the asymptotic difference between the two bounds.
with Pr[B] ≤ q, such that B is also mutually independent of a set of all other events Aj ∈ A except for at
most d. Then Pr[B| ∩li=1 Āi ] ≤ eq.
We apply the conditional
LLL distribution to obtain an upper bound on the size of partial array that
t
v
leaves at most log vt −1 ≈ v t interactions uncovered. For a positive integer k, let I = {j1 , . . . , jρ } ⊆ [k]
where j1 < . . . < jρ . Let A be an n × k array where each entry is from the set [v]. Let AI denote the n × ρ
array in which AI (i, `) = A(i, j` ) for 1 ≤ i ≤ N and 1 ≤ ` ≤ ρ; AI is the projection of A onto the columns
in I.
Let M ⊆ [v]t be a set of m t-tuples of symbols, and C ∈ [k]
be a set of t columns. Suppose the
t
entries in the array A are chosen independently from [v] with uniform probability.
Let BC denote the event
that at least one of the tuples in M is not covered in AC . There are η = kt such events, and for all of
n
them Pr[BC ] ≤ m 1 − v1t . Moreover, when k ≥ 2t, each of the events is mutually independent of all
k
other events except for at most ρ = kt − k−t
− 1 < t t−1
. Therefore, by the Lovász local lemma, when
t
1 n
eρm 1 − vt ≤ 1, none of the events BC occur. Solving for n, when
n≥
log(eρm)
t
log vtv−1
(3)
there exists an n × k array A over [v] such that for all C ∈ [k]
t , AC covers all the m tuples in M . In fact
we can use a Moser-Tardos type algorithm to construct such an array.
Let ι be an interaction whose t-tuple
n of symbols is not in M . Then the probability that ι is not covered
in an n × k array is at most 1 − v1t
when each entry of the array is chosen independently from [v] with
uniform probability. Therefore, by the
conditional LLL distribution the probability that ι1 is
n not covered
in the array A where for all C ∈ [k]
. Moreover,
t , AC covers all the m tuples in M is at most e 1 − v t
there are η(v t − m) such interactions ι. By the linearity of expectation, the expected number of uncovered
25
n
interactions in A is less than v t when η(v t − m)e 1 − v1t ≤ v t . Solving for n, we obtain
log ηe 1 − vmt
.
n≥
t
log vtv−1
Therefore, there exists an n × k array with n = max
log{ηe(1− vmt )}
log(eρm)
,
t
t
log vtv−1
log vtv−1
(4)
that has at most v t
uncovered interactions. To compute n explicitly, we must choose m. We can select a value of m to minimize
n graphically for given values of t, k and v. For example, Figure 11 plots Equations 3 and 4 against m for
t = 3, k = 350, v = 3, and finds the minimum value of n.
460
445
max(Equation (3), Equation (4))
440
n − number of rows in the partial array
n − number of rows in the partial array
Equation (3)
Equation (4)
420
400
380
360
340
0
5
10
15
m
20
25
30
(a) Equations 3 and 4 against m.
440
435
430
425
420
0
5
10
15
m
20
25
30
(b) Maximum of the two sizes against m.
Figure 11: The minimum is at n = 422, when m = 16. t = 3, k = 350, v = 3
We compare the size of the partial array from the naı̈ve two-stage method (Algorithm 4) with the size
obtained by the graphical methods in Figure 12. The Lovász local lemma based method is asymptotically
better than the simple randomized method. However, except for the small values of t and v, in the range
of k values relevant for practical applications the simple randomized algorithm requires fewer rows than the
Lovász local lemma based method.
5.2
Lovász local lemma based two-stage bound
We can apply the techniques from Section 5.1 to obtain a two-stage bound similar to Theorem 2 using the
Lovász local lemma and conditional LLL distribution. First we extend a result from [33].
Theorem
7. Let t, k, v be integers with k ≥ t ≥ 2, v ≥ 2 and let η = kt , and ρ = kt − k−t
t . If
vt
v t −1
ηv t log
ρ
≤ v t Then
log
CAN(t, k, v) ≤
k
t
+ t log v + log log
t
log vtv−1
vt
v t −1
+2
η
− .
ρ
Proof. Let M ⊆ [v]t be a set of m t-tuples of symbols. Following the arguments of Section 5.1, when
there exists an n × k array A over [v] such that for all C ∈ [k] , AC covers all m tuples in M .
n ≥ log(eρm)
t
vt
log
v t −1
At most η(v t − m) interactions are uncovered in such an array. Using the conditional
n LLL distribution,
the probability that one such interaction is not covered in A is at most e 1 − v1t . Therefore, by the
26
n − number of rows in partial array with vt missing interactions
n − number of rows in partial array with vt missing interactions
550
500
450
400
350
300
250
200
150
Randomized (Algorithm 4)
LLL based
100
50
0
200
400
600
k
800
1000
1200
2500
2000
1500
1000
500
0
0
Randomized (Algorithm 4)
LLL based
500
1000
1500
2000
2500
3000
3500
4000
4500
k
(a) t = 3, v = 3.
(b) t = 4, v = 3.
Figure 12: Comparison of the size of the partial array constructed in the first stage. Compares the size of
the partial array specified in Algorithm 4 in Section 3.1, and the size derived in Section 5.1.
t
n
linearity of expectation, we can find one such array A that leaves at most eη(v t − m) 1 − v1t = ηρ vm − 1
interactions uncovered. Adding one row per uncovered interactions to A, we obtain a covering array with at
most N rows, where
log(eρm)
η vt
+
N=
−
1
t
ρ m
log tv
The value of N is minimized when m =
v −1
t
ηv t log vtv−1
ρ
. Because m ≤ v t , we obtain the desired bound.
When m = v t this recaptures the bound of Theorem 5.
Figure 13 compares the LLL based two-stage bound from Theorem 7 to the standard two-stage bound
from Theorem 2, the Godbole et al. bound in Theorem 5, and the SLJ bound in Theorem 1. Although the
LLL based two-stage bound is tighter than the LLL based Godbole et al. bound, even for quite large values
of k the standard two-stage bound is tighter than the LLL based two-stage bound. In practical terms, this
specific LLL based two-stage method does not look very promising, unless the parameters are quite large.
6
Conclusion and open problems
Many concrete algorithms within a two-stage framework for covering array construction have been introduced
and evaluated. The two-stage approach extends the range of parameters for which competitive covering arrays
can be constructed, each meeting an a priori guarantee on its size. Indeed as a consequence a number of
best known covering arrays have been improved upon. Although each of the methods proposed has useful
features, our experimental evaluation suggests that TS hRand, Greedy; 2ρ, Γi and TS hRand, Den; 2ρ.Γi with
Γ ∈ {Cyclic, Frobenius} realize a good trade-off between running time and size of the constructed covering
array.
Improvements in the bounds, or in the algorithms that realize them, are certainly of interest. We
mention some specific directions. Establishing tighter bounds on the coloring based methods of Section 3.3
is a challenging problem. Either better estimates of the chromatic number of the incompatibility graph after
a random first stage, or a first stage designed to limit the chromatic number, could lead to improvements in
the bounds.
In Section 5.1 and 5.2 we explored using a Moser-Tardos type algorithm for the first stage. Although this
is useful for asymptotic bounds, practical improvements appear to be limited. Perhaps a different approach of
27
10000
500
9000
450
8000
400
7000
N − number of rows
N − number of rows
550
350
300
250
SLJ bound
Godbole
LLL−2−stage
2−stage
200
150
100
0
50
100
150
200
250
300
350
400
6000
5000
4000
SLJ bound
Godbole
LLL−2−stage
2−stage
3000
2000
450
k
(a) t = 3, v = 3.
1000
0
1000
2000
3000
4000
k
5000
6000
7000
8000
(b) t = 4, v = 4.
Figure 13: Comparison among the LLL based two-stage bound from Theorem 7, the standard two-stage
bound from Theorem 2, the Godbole et al. bound in Theorem 5, and the SLJ bound in Theorem 1.
reducing the number of bad events to be avoided explicitly by the algorithm may lead to a better algorithm.
A potential approach may look like following: “Bad” events would denote non-coverage of an interaction
over a t-set of columns. We would select a set of column t-sets such that the dependency graph of the
corresponding bad events have a bounded maximum degree (less than the original dependency graph). We
would devise a Moser-Tardos type algorithm for covering all the interactions on our chosen column t-sets,
and then apply the conditional LLL distribution to obtain an upper bound on the number of uncovered
interactions. However, the difficulty lies in the fact that “all vertices have degree ≤ ρ” is a non-trivial,
“hereditary” property for induced subgraphs, and for such properties finding a maximum induced subgraph
with the property is an NP-hard optimization problem [17]. There is still hope for a randomized or “nibble”
like strategy that may find a reasonably good induced subgraph with a bounded maximum degree. Further
exploration of this idea seems to be a promising research avenue.
In general, one could consider more than two stages. Establishing the benefit (or not) of having more
than two stages is also an interesting open problem. Finally, the application of the methods developed to
mixed covering arrays appears to provide useful techniques for higher strengths; this merits further study as
well.
Acknowledgments
The research was supported in part by the National Science Foundation under Grant No. 1421058.
References
[1] A. N. Aldaco, C. J. Colbourn, and V. R. Syrotiuk. Locating arrays: A new experimental design for
screening complex engineered systems. SIGOPS Oper. Syst. Rev., 49(1):31–40, Jan. 2015.
[2] N. Alon and J. H. Spencer. The probabilistic method. Wiley-Interscience Series in Discrete Mathematics
and Optimization. John Wiley & Sons, Inc., Hoboken, NJ, third edition, 2008. With an appendix on
the life and work of Paul Erdős.
[3] Android. Android configuration class, 2016. http://developer.android.com/reference/android/content/
res/Configuration.html.
[4] J. I. Brown. The complexity of generalized graph colorings. Discrete Appl. Math., 69(3):257–270, 1996.
28
[5] R. C. Bryce and C. J. Colbourn. The density algorithm for pairwise interaction testing. Software
Testing, Verification, and Reliability, 17:159–182, 2007.
[6] R. C. Bryce and C. J. Colbourn. A density-based greedy algorithm for higher strength covering arrays.
Software Testing, Verification, and Reliability, 19:37–53, 2009.
[7] R. C. Bryce and C. J. Colbourn. Expected time to detection of interaction faults. Journal of Combinatorial Mathematics and Combinatorial Computing, 86:87–110, 2013.
[8] J. N. Cawse. Experimental design for combinatorial and high throughput materials development. GE
Global Research Technical Report, 29:769–781, 2002.
[9] M. A. Chateauneuf, C. J. Colbourn, and D. L. Kreher. Covering arrays of strength 3. Des. Codes
Crypt., 16:235–242, 1999.
[10] D. M. Cohen, S. R. Dalal, M. L. Fredman, and G. C. Patton. The AETG system: An approach to
testing based on combinatorial design. IEEE Transactions on Software Engineering, 23:437–44, 1997.
[11] M. B. Cohen. Designing test suites for software interaction testing. PhD thesis, The University of
Auckland, Department of Computer Science, 2004.
[12] C. J. Colbourn. Combinatorial aspects of covering arrays. Le Matematiche (Catania), 58:121–167, 2004.
[13] C. J. Colbourn. Covering array tables, 2005-2015. http://www.public.asu.edu/∼ccolbou/src/tabby.
[14] C. J. Colbourn. Covering arrays and hash families. In Information Security and Related Combinatorics,
NATO Peace and Information Security, pages 99–136. IOS Press, 2011.
[15] C. J. Colbourn. Conditional expectation algorithms for covering arrays. Journal of Combinatorial
Mathematics and Combinatorial Computing, 90:97–115, 2014.
[16] R. Diestel. Graph Theory. Graduate Texts in Mathematics. Springer, fourth edition, 2010.
[17] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NPCompleteness. W. H. Freeman & Co., New York, NY, USA, 1979.
[18] A. P. Godbole, D. E. Skipper, and R. A. Sunley. t-covering arrays: upper bounds and Poisson approximations. Combinatorics, Probability and Computing, 5:105–118, 1996.
[19] B. Haeupler, B. Saha, and A. Srinivasan. New constructive aspects of the Lovász local lemma. J. ACM,
58(6):Art. 28, 28, 2011.
[20] A. S. Hedayat, N. J. A. Sloane, and J. Stufken. Orthogonal Arrays. Springer-Verlag, New York, 1999.
[21] D. S. Johnson. Approximation algorithms for combinatorial problems. J. Comput. System Sci., 9:256–
278, 1974.
[22] G. O. H. Katona. Two applications (for search theory and truth functions) of Sperner type theorems.
Periodica Math., 3:19–26, 1973.
[23] D. Kleitman and J. Spencer. Families of k-independent sets. Discrete Math., 6:255–262, 1973.
[24] D. R. Kuhn, R. Kacker, and Y. Lei. Introduction to Combinatorial Testing. CRC Press, 2013.
[25] D. R. Kuhn and M. Reilly. An investigation of the applicability of design of experiments to software
testing. In Proc. 27th Annual NASA Goddard/IEEE Software Engineering Workshop, pages 91–95, Los
Alamitos, CA, 2002. IEEE.
29
[26] D. R. Kuhn, D. R. Wallace, and A. M. Gallo. Software fault interactions and implications for software
testing. IEEE Trans. Software Engineering, 30:418–421, 2004.
[27] L. Lovász. On the ratio of optimal integral and fractional covers. Discrete Math., 13(4):383–390, 1975.
[28] K. Meagher and B. Stevens. Group construction of covering arrays. J. Combin. Des., 13:70–77, 2005.
[29] R. A. Moser. A constructive proof of the Lovász local lemma. In STOC’09—Proceedings of the 2009
ACM International Symposium on Theory of Computing, pages 343–350. ACM, New York, 2009.
[30] R. A. Moser and G. Tardos. A constructive proof of the general Lovász local lemma. J. ACM, 57(2):Art.
11, 15, 2010.
[31] C. Nie and H. Leung. A survey of combinatorial testing. ACM Computing Surveys, 43(2):#11, 2011.
[32] A. H. Ronneseth and C. J. Colbourn. Merging covering arrays and compressing multiple sequence
alignments. Discrete Appl. Math., 157:2177–2190, 2009.
[33] K. Sarkar and C. J. Colbourn. Upper bounds on the size of covering arrays. ArXiv e-prints.
[34] G. Seroussi and N. H. Bshouty. Vector sets for exhaustive testing of logic circuits. IEEE Trans. Inform.
Theory, 34:513–522, 1988.
[35] S. K. Stein. Two combinatorial covering theorems. J. Combinatorial Theory Ser. A, 16:391–397, 1974.
[36] J. Torres-Jimenez, H. Avila-George, and I. Izquierdo-Marquez. A two-stage algorithm for combinatorial
testing. Optimization Letters, pages 1–13, 2016.
30
| 8cs.DS
|
DynASP2.5: Dynamic Programming on
Tree Decompositions in Action∗
Johannes K. Fichte, Markus Hecher, Michael Morak, Stefan Woltran
TU Wien, Vienna, Austria
[email protected]
arXiv:1706.09370v1 [cs.LO] 28 Jun 2017
April 13, 2018
Abstract
A vibrant theoretical research area are efficient exact parameterized algorithms. Very recent
solving competitions such as the PACE challenge show that there is also increasing practical
interest in the parameterized algorithms community. An important research question is whether
dedicated parameterized exact algorithms exhibit certain practical relevance and one can even beat
well-established problem solvers. We consider the logic-based declarative modeling language and
problem solving framework Answer Set Programming (ASP). State-of-the-art ASP solvers rely
considerably on Sat-based algorithms. An ASP solver (DynASP2), which is based on a classical
dynamic programming on tree decompositions, has been published very recently. Unfortunately,
DynASP2 can outperform modern ASP solvers on programs of small treewidth only if the question
of interest is to count the number of solutions. In this paper, we describe underlying concepts of our
new implementation (DynASP2.5) that shows competitive behavior to state-of-the-art ASP solvers
even for finding just one solution when solving problems as the Steiner tree problem that have been
modeled in ASP on graphs with low treewidth. Our implementation is based on a novel approach
that we call multi-pass dynamic programming (M-DPSINC ).
1
Introduction
Answer set programming (ASP) is a logic-based declarative modelling language and problem solving
framework [18], where a program consists of sets of rules over propositional atoms and is interpreted
under an extended stable model semantics [23]. Problems are usually modelled in ASP in such a way that
the stable models (answer sets) of a program directly form a solution to the considered problem instance.
Computational problems for disjunctive, propositional ASP such as deciding whether a program has an
answer set are complete for the second level of the Polynomial Hierarchy [9]. In consequence, finding
answer sets usually involves a Sat part (finding a model of the program) and an Unsat part (minimality
check). A variety of CDCL-based ASP solvers have been implemented [19, 4] and proven to be very
successful in solving competitions [14]. Very recently, a dynamic programming based solver (DynASP2 )
that builds upon ideas from parameterized algorithmics was proposed [11]. For disjunctive input
programs, the runtime of the underlying algorithms is double exponential in the incidence treewidth
and linear in the input size (so-called fixed-parameter linear algorithms). DynASP2 (i) takes a tree
decomposition of a certain graph representation (incidence graph) of a given input program and (ii) solves
the program via dynamic programming (DP) on the tree decomposition by traversing the tree exactly
once. Both finding a model and checking minimality are considered at the same time. Once the root
node has been reached, complete solutions (if exist) for the input program can be constructed. This
approach pays off for counting answer sets, but is not competitive for outputting just one answer set.
∗ This is the authors self-archived copy including detailed proofs. Research was supported by the Austrian Science Fund
(FWF), Grant Y698.
1
The reason for that lies in the exhaustive nature of dynamic programming as all potential values are
computed locally for each node of the tree decomposition. In consequence, space requirements can
be quite extensive resulting in long running times. Moreover, dynamic programming algorithms on
tree decompositions may yield extremely diverging run-times on tree decompositions of the exact same
width [1]. In this paper, we propose a multi-pass approach (M-DPSINC ) for dynamic programming on
tree decompositions as well as a new implementation (DynASP2.5). In contrast to classical dynamic
programming algorithms for problems on the second level of the Polynomial Hierarchy, M-DPSINC traverses
the given tree decomposition multiple times. Starting from the leaves, we compute and store (i) sets of
atoms that are relevant for the Sat part (finding a model of the program) up to the root. Then we
go back again to the leaves and compute and store (ii) sets of atoms that are relevant for the Unsat
part (checking for minimality). Finally, we go once again back to the leaves and (iii) link sets from past
Passes (i) and (ii) that might lead to an answer set in the future. As a result, we allow for early cleanup
of candidates that do not lead to answer sets.
Further, we present technical improvements (including working on non-normalized tree decompositions) and employ dedicated customization techniques for selecting tree decompositions. Our
improvements are main ingredients to speedup the solving process for DP algorithms. Experiments
indicate that DynASP2.5 is competitive even for finding one answer set using the Steiner tree problem
on graphs with low treewidth. In particular, we are able to solve instances that have an upper bound on
the incidence treewidth of 14 (whereas DynASP2 solved instances of treewidth at most 9).
Our main contributions can be summarized as follows:
1. We establish a novel fixed-parameter linear algorithm (M-DPSINC ), which works in multiple passes
and computes Sat and Unsat parts separately.
2. We present an implementation (DynASP2.5)1 and an experimental evaluation.
Related Work. Jakl, Pichler, and Woltran [17] have considered ASP solving when parameterized by the
treewidth of a graph representation and suggested fixed-parameter linear algorithms. Fichte et al. [11]
have established additional algorithms and presented empirical results on an implementation that is
dedicated to counting answer sets for the full ground ASP language. The present paper extends their
work by a multi-pass dynamic programming algorithm. Bliem et al. [5] have introduced a general multipass approach and an implementation (D-FLATˆ2) for dynamic programming on tree decompositions
solving subset minimization tasks. Their approach allows to specify dynamic programming algorithms
by means of ASP. In a way, one can see ASP in their approach as a meta-language to describe table
algorithms2 , whereas our work presents a dedicated algorithm to find an answer set of a program. In
fact, our implementation extends their general ideas for subset minimization (disjunctive rules) to also
support weight rules. However, due to space constraints we do not report on weight rules in this paper.
Beyond that, we require specialized adaptions to the ASP problem semantics, including three valued
evaluation of atoms, handling of non-normalized tree decompositions, and optimizations in join nodes
to be competitive. Abseher, Musliu, and Woltran [2] have presented a framework that computes tree
decompositions via heuristics, which is also used in our solver. Other tree decomposition systems can be
found on the PACE challenge website [7]. Note that improved heuristics for finding a tree decomposition
of smaller width (if possible) directly yields faster results for our solver.
2
Formal Background
2.1
Tree Decompositions
Let G = (V, E) be a graph, T = (N, F, n) a rooted tree, and χ : N → 2V a function that maps
each node t ∈ N to a set of vertices. We call the sets χ(·) bags and N the set of nodes. Then,
the pair T = (T, χ) is a tree decomposition (TD) of G if the following conditions hold: (i) for every
1 The
2 See
source code of our solver is available at https://github.com/daajoe/dynasp/releases/tag/v2.5.0.
Algorithm 1 for the concept of table algorithms.
2
vertex v ∈ V there is a node t ∈ N with v ∈ χ(t); (ii) for every edge e ∈ E there is a node t ∈ N with
e ⊆ χ(t); and (iii) for any three nodes t1 , t2 , t3 ∈ N , if t2 lies on the unique path from t1 to t3 , then
χ(t1 ) ∩ χ(t3 ) ⊆ χ(t2 ). We call max{|χ(t)| − 1 | t ∈ N } the width of the TD. The treewidth tw (G) of a
graph G is the minimum width over all possible TDs of G.
Note that each graph has a trivial TD (T, χ) consisting of the tree ({n}, ∅, n) and the mapping
χ(n) = V . It is well known that the treewidth of a tree is 1, and a graph containing a clique of size k
has at least treewidth k − 1. For some arbitrary but fixed integer k and a graph of treewidth at most k,
3
we can compute a TD of width 6 k in time 2O(k ) · |V | [6]. Given a TD (T, χ) with T = (N, ·, ·), for a
node t ∈ N we say that type(t) is leaf if t has no children; join if t has children t0 and t00 with t0 =
6 t00
0
00
0
0
0
and χ(t) = χ(t ) = χ(t ); int (“introduce”) if t has a single child t , χ(t ) ⊆ χ(t) and |χ(t)| = |χ(t )| + 1;
rem (“removal”) if t has a single child t0 , χ(t0 ) ⊇ χ(t) and |χ(t0 )| = |χ(t)| + 1. If every node t ∈ N has
at most two children, type(t) ∈ {leaf, join, int, rem}, and bags of leaf nodes and the root are empty, then
the TD is called nice. For every TD, we can compute a nice TD in linear time without increasing the
width [6]. Later, we traverse a TD bottom up, therefore, let post-order(T, t) be the sequence of nodes in
post-order of the induced subtree T 0 = (N 0 , ·, t) of T rooted at t.
2.2
Answer Set programming (ASP)
ASP is a declarative modelling and problem solving framework that combines techniques of knowledge
representation and database theory. A main advantage of ASP is its expressiveness and when using
non-ground programs the advanced declarative problem modelling capability. Prior to solving, nonground programs are usually compiled into grounded by a grounder. In this paper, we restrict ourselves
to ground ASP programs. For a comprehensive introduction, see, e.g., [18]. Let `, m, n be non-negative
integers such that ` ≤ m ≤ n, a1 , . . ., an distinct propositional atoms and l ∈ {a1 , ¬a1 }. A choice
rule is an expression of the form {a1 ; . . . ; a` } ← a`+1 , . . . , am , ¬am+1 , . . . , ¬an , intuitively some subset
of {a1 , . . . , a` } is true if all atoms a`+1 , . . . , am are true and there is no evidence that any atom of
am+1 , . . . , an is true. A disjunctive rule is of the form a1 ∨ · · · ∨ a` ← a`+1 , . . . , am , ¬ am+1 , . . ., ¬an ,
intuitively at least one atom of a1 , . . . , a` must be true if all atoms a`+1 , . . . , am are true and there is no
evidence that any atom of am+1 , . . . , an is true. An optimization rule is an expression of the form l.
with the intuitive meaning that when literal l is true, this incurs a penalty of weight w. A rule is either
a disjunctive, a choice, or an optimization rule. For a choice or disjunctive rule r, let Hr := {a1 , . . . , a` },
Br+ := {a`+1 , . . . , am }, and Br− := {am+1 , . . . , an }. Usually, if Br− ∪ Br+ = ∅ we write for a rule r
simply Hr instead of Hr ← . For an optimization rule r, if l = a1 , let Br+ := {a1 } and Br− := ∅; and
if l = ¬a1 , let Br− := {a1 } and Br+ := ∅. For a rule r, let at(r) := Hr ∪ Br+ ∪ Br− denoteSits atoms
and Br := Br+ ∪ {¬b | b ∈ Br− } its body. Let a program P be a set of rules, and at(P ) := r∈P at(r)
denote its atoms. and let CH(P ), DISJ(P ), and OPT(P ) denote the set of all choice, disjunctive, and
optimization rules in P , respectively. A set M ⊆ at(P ) satisfies a rule r if (i) (Hr ∪ Br− ) ∩ M 6= ∅
or Br+ 6⊆ M for r ∈ DISJ(P ) or (ii) r is a choice or optimization rule. M is a model of P , denoted by
M P , if M satisfies every rule r ∈ P .
M = ∅} of rules, and
The reduct rM (i) of a choice rule r is the set {a ← Br+ | a ∈ Hr ∩ M, Br− ∩S
(ii) of a disjunctive rule r is the singleton {Hr ← Br+ | Br− ∩ M = ∅}. P M := r∈P rM is called GL
reduct of P with respect to M . A set M ⊆ at(P ) is an answer set of P if (i) M P and (ii) there is no
M 0 ( M such that M 0 P M , that is, M is subset minimal with respect to P M .
We call cst(P, M ) := |{r | r ∈ P, r is an optimization rule, (Br+ ∩ M ) ∪ (Br− \ M ) 6= ∅}| the cost of
answer set M for P . An answer set M of P is optimal if its cost is minimal over all answer sets.
Example 1. Consider program
rab
rbc
rcd
rad
rb
rd
rc1
rc2
r¬c
z }| { z }| { z }| { z }| { z }|
{ z }| { z
}|
{ z
}|
{ z }| {
P = {{eab }; {ebc }; {ecd }; {ead }; ab ← eab ; ad ← ead ; ac ← ab , ebc ; ac ← ad , ecd ; ← ¬ac }.
The set A = {eab , ebc , ab , ac } is an answer set of P , since {eab , ebc , ab , ac } is the only minimal model
of P A = {eab ←; ebc ←; ab ← eab ; ad ← ead ; ac ← ab , ebc ; ac ← ad , ecd }. Then, consider program
R = {a ∨ c ← b; b ← c, ¬g; c ← a; b ∨ c ← e; h ∨ i ← g, ¬c; a ∨ b; g ← ¬i; c; {d} ← g}. The
set B = {b, c, d, g} is an answer set of R since {b, c, d, g} and {a, c, d, g} are the minimal models of
3
Solve local
probl. A(t, . . . )
Visit next node no
t in post-order
3. DPA (T )
Store results
in A-Tabs[t]
Done?
yes
1. Construct graph G
Store witnesses
in W-Tabs[t]
←−DynASP2 DynASP2.5−→
2. Comp. TD T of G
3.I done?
no
Visit next node
t in post-order
3.I. DPW (T )
yes
Purge non-witnesses
Store counter-witnesses in C-Tabs[t]
3.II done?
yes
Purge noncounter-witnesses
4. Print solution
Compute witnesses of W(t, . . . )
no
Visit next node
t in post-order
3.II. DPC (T )
Store result in
W, C-Tabs[t]
yes
Compute counterwits. of C(t, . . . )
3.III done?
Link counter-wits.
to witnesses
no
Visit next node
t in post-order
3.III. DPLW,C (T , W-Tabs, C-Tabs)
Figure 1
Control flow for DP-based ASP solver (DynASP2, left) and for DynASP2.5 (right).
RB = {a ∨ c ← b; c ← a; b ∨ c ← e; a ∨ b; g; c; d ← g}.
Given a program P , we consider the problems of computing an answer set (called AS), outputting
the number of optimal answer sets (called #AspO), and listing all optimal answer sets of P (called
EnumAsp). Further, given a propositional formula F and an atom sol, we use the entailment problem
of listing every subset-minimal model M of F with sol ∈ M (called EnumMinSAT1).
2.3
Graph Representations of Programs
In order to use TDs for ASP solving, we need dedicated graph representations of programs. The incidence
graph I(P ) of P is the bipartite graph that has the atoms and rules of P as vertices and an edge a r if
a ∈ at(r) for some rule r ∈ P [11]. The semi-incidence graph S(P ) of P is a graph that has the atoms
and rules of P as vertices and (i) an edge a r if a ∈ at(r) for some rule r ∈ P as well as (ii) an edge a b
for disjoint atoms a, b ∈ Hr where r ∈ P is a choice rule. Since for every program P the incidence
graph I(P ) is a subgraph of the semi-incidence graph, we have that tw (I(P )) ≤ tw (S(P )). Further,
by definition of a TD and the construction of a semi-incidence graph that head atoms of choice rules,
respectively, occur in at least one common bag of the TD.
2.4
Sub-Programs
Let T = (T, χ) be a nice TD of graph representation S(P ) of a program P . Further, let T = (N, ·, n) and
t ∈ N . The bag-program is defined as Pt := P ∩ χ(t). Further, the set at≤t := {a | a ∈ at(P ) ∩ χ(t0 ), t0 ∈
post-order(T, t)} is called atoms below t, the program below t is defined as P≤t := {r | r ∈ Pt0 , t0 ∈
post-order(T, t)}, and the program strictly below t is P<t := P≤t \ Pt . It holds that P≤n = P<n = P and
at≤n = at(P ).
3
A Single Pass DP Algorithm
A dynamic programming based ASP solver, such as DynASP2 [11], splits the input program P into
“bag-programs” based on the structure of a given nice tree decomposition for P and evaluates P in parts,
thereby storing the results in tables for each TD node. More precisely, the algorithm works as outlined
on the left and middle of Figure 1 and encompasses the following steps:
1. Construct a graph representation G(P ) of the given input program P .
2. Compute a TD T of the graph G(P ) by means of some heuristic, thereby decomposing G(P ) into
several smaller parts and fixing an ordering in which P will be evaluated.
4
Algorithm 1: Algorithm DPA (T ) for Dynamic Programming on TD T for ASP [11].
1
2
3
In: Table algorithm A, nice TD T = (T, χ) with T = (N, ·, n) of G(P ) according to A.
Out: A-Tabs: maps each TD node t ∈ T to some computed table τt .
Child-Tabst := {A-Tabs[t0 ] | t0 is a child of t in T }
for iterate t in post-order(T,n) do
A-Tabs[t] ← A(t, χ(t), Pt , at≤t , Child-Tabst )
3. Algorithm 1 – DPA (T ) – sketches the general scheme for this step, assuming that an algorithm A,
which highly depends on the graph representation, is given. We usually call A the table algorithm 3 .
For every node t ∈ T in the tree decomposition T = ((T, E, n), χ) (in a bottom-up traversal),
run A and compute A-Tabs[t], which are sets of tuples (or rows for short). Intuitively, algorithm A
transforms tables of child nodes of t to the current node, and solves a “local problem” using
bag-program Pt . The algorithm thereby computes (i) sets of atoms called (local) witness sets and
(ii) for each local witness set M subsets of M called counter-witness sets [11], and directly follows
the definition of answer sets being (i) models of P and (ii) subset minimal with respect to P M .
4. For root n interpret the table A-Tabs[n] (and tables of children, if necessary) and print the solution
to the considered ASP problem.
Next, we propose a new table algorithm (SINC) for programs without optimization rules. Since our
algorithm trivially extends to counting and optimization rules by earlier work [11], we omit such rules.
The table algorithm SINC employs the semi-incidence graph and is depicted in Algorithm 2. DPSINC
merges two earlier algorithms for the primal and incidence graph [11] resulting in slightly different worst
case runtime bounds (c.f., Theorem 1).
Our table algorithm SINC computes and stores (i) sets of atoms (witnesses) that are relevant for the
Sat part (finding a model of the program) and (ii) sets of atoms (counter-witnesses) that are relevant
for the Unsat part (checking for minimality). In addition, we need to store for each set of witnesses as
well as its set of counter-witnesses satisfiability states (sat-states for short). For the following reason:
By Definition of TDs and the semi-incidence graph, it is true for every atom a and every rule r of a
program that if atom a occurs in rule r, then a and r occur together in at least one bag of the TD.
In consequence, the table algorithm encounters every occurrence of an atom in any rule. In the end,
on removal of r, we have to ensure that r is among the rules that are already satisfied. However, we
need to keep track whether a witness satisfies a rule, because not all atoms that occur in a rule occur
together in exactly one bag. Hence, when our algorithm traverses the TD and an atom is forgotten we
still need to store this sat-state, as setting the forgotten atom to a certain truth value influences the
satisfiability of the rule. Since the semi-incidence graph contains a clique on every set A of atoms that
occur together in choice rule head, those atoms A occur together in a common bag of any TD of the
semi-incidence graph. For that reason, we do not need to incorporate choice rules into the satisfiability
state, in contrast to the algorithm for the incidence graph [11]. We can see witness sets together with
its sat-state as witness. Then, in Algorithm 2 (SINC) a row in the table τt is a triple hM, σ, Ci. The
set M ⊆ at(P ) ∩ χ(t) represents a witness set. The family C of sets concerns counter-witnesses, which we
will discuss in more detail below. The sat-state σ for M represents rules of χ(t) satisfied by a superset
of M . Hence, M witnesses a model M 0 ⊇ M where M 0 P<t ∪ σ. We use binary operator ∪ to combine
sat-states, which ensures that rules satisfied in at least one operand remain satisfied. We compute a new
sat-state σ from a sat-state and satisfied rules, formally, SatPr(Ṙ, M ) := {r | (r, R) ∈ Ṙ, M R} for
M ⊆ χ(t) \ Pt and program Ṙ(r) constructed by Ṙ, mapping rules to local-programs (Definition 1).
Definition 1. Let P be a program, T = (·, χ) be a TD of S(P ), t be a node of T and R ⊆ Pt . The
local-program R(t) is obtained from R ∪ {← Br | r ∈ R is a choice rule, Hr ( at≤t }4 by removing from
(t)
every rule all literals a, ¬a with a 6∈ χ(t). We define Ṙ(t) : R → 2R by Ṙ(t) (r) := {r}(t) for r ∈ R.
3 The
table algorithm SINC for example is given in Algorithm 2.
require to add {← Br | r ∈ R is a choice rule, Hr ( at≤t } in order to decide satisfiability for corner cases of choice
rules involving counter-witnesses of Line 3 in Algorithm 2.
5 For set S and element s, we denote S + := S ∪ {s} and S − := S \ {s}.
s
s
4 We
5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Algorithm 2: Table algorithm SINC(t, χt , Pt , at≤t , Child-Tabst ).
In: Bag χt , bag-program Pt , atoms-below at≤t , child tables Child-Tabst of t. Out: Tab. τt .
if type(t) = leaf then τt ← {h∅, ∅, ∅i}
/* Abbreviations see Footnote 5. */
else if type(t) = int, a ∈ χt \ Pt is introduced and τ 0 ∈ Child-Tabs
t then
(t,M + )
(t)
τt ← {hMa+ , σ ∪ SatPr(Ṗt , Ma+ ), {hCa+ , ρ ∪ SatPr(Ṗt a , Ca+ )i | hC, ρi ∈ C} ∪
(t,M + )
(t,M + )
{hC, ρ ∪ SatPr(Ṗt a , C)i | hC, ρi ∈ C} ∪ {hM, σ ∪ SatPr(Ṗt a , M )i}i | hM , σ, Ci ∈ τ 0 }
(t)
(t,M )
∪ {hM , σ ∪ SatPr(Ṗt , M ), {hC, ρ ∪ SatPr(Ṗt
, C)i | hC, ρi ∈ C}i
| hM , σ, Ci ∈ τ 0 }
0
else if type(t) = int, r ∈ χt ∩ Pt is introduced and τ ∈ Child-Tabst then
τt ← {hM , σ ∪ SatPr({ṙ}(t) , M ), {hC, ρ ∪ SatPr({ṙ}(t,M ) , C)i | hC, ρi ∈ C}i
| hM , σ, Ci ∈ τ 0 }
0
else if type(t) = rem, a 6∈ χt is removed atom and τ ∈ Child-Tabst then
τt ← {hMa− , σ, {hCa− , ρi | hC, ρi ∈ C}i
| hM , σ, Ci ∈ τ 0 }
0
else if type(t) = rem, r 6∈ χt is removed rule and τ ∈ Child-Tabst then
τt ← {hM , σr− , hC, ρ−
| hM , σ, Ci ∈ τ 0 , r ∈ σ}
r i | hC, ρi ∈ C, r ∈ ρ i
else if type(t) = join and τ 0 , τ 00 ∈ Child-Tabst with τ 0 6= τ 00 then
τt ← {hM , σ 0 ∪ σ 00 , {hC, ρ0 ∪ ρ00 i | hC, ρ0 i ∈ C 0 , hC, ρ00 i ∈ C 00 } ∪ {hM, ρ ∪ σ 00 i | hM, ρi ∈ C 0 } ∪
{hM, σ 0 ∪ ρi | hM, ρi ∈ C 00 }i
| hM , σ 0 , C 0 i ∈ τ 0 , hM , σ 00 , C 00 i ∈ τ 00 }
(t )
(t )
Example 2. Observe Pt4 4 = {← ebc , rb } and Pt5 5 = {c ←} for Pt4 and Pt5 of Figure 2.
In Example 3 we give an idea how we compute models of a given program using the semi-incidence
graph. The resulting algorithm MOD is obtained from SINC, by taking only the first two row positions
(red and green parts). The remaining position (blue part), can be seen as an algorithm (CMOD) that
computes counter-witnesses (see Example 5). Note that we discuss selected cases, and we assume row
numbers in each table τt , i.e., the i th -row corresponds to ut.i = hMt.i , σt.i i.
Example 3. Consider program P from Example 1, TD T = (·, χ) in Figure 2, and the tables τ1 ,. . .,
τ34 , which illustrate computation results obtained during post-order traversal of T by DPMOD . Note that
Figure 2 (left) does not show every intermediate node of TD T . Table τ1 = {h∅, ∅i} as type(t1 ) =
leaf (see Algorithm 2 L1). Table τ3 is obtained via introducing rule rab , after introducing atom eab
(type(t2 ) = type(t3 ) = int). It contains two rows due to two possible truth assignments using atom eab
(L3–5). Observe that rule rab is satisfied in both rows M3.1 and M3.2 , since the head of choice rule rab is
in at≤t3 (see L7 and Definition 1). Intuitively, whenever a rule r is proven to be satisfiable, sat-state σt.i
marks r satisfiable since an atom of a rule of S(P ) might only occur in one TD bag. Consider table τ4
with type(t4 ) = rem and rab ∈ χ(t3 ) \ χ(t4 ). By definition (TDs and semi-incidence graph), we have
encountered every occurrence of any atom in rab . In consequence, MOD enforces that only rows where
rab is marked satisfiable in τ3 , are considered for table τ4 . The resulting table τ4 consists of rows of τ3
with σ4.i = ∅, where rule rab is proven satisfied (rab ∈ σ3.1 , σ3.2 , see L 11). Note that between nodes t6
and t10 , an atom and rule remove as well as an atom and rule introduce node is placed. Observe that the
second row u6.2 = hM6.2 , σ6.2 i ∈ τ6 does not have a “successor row” in τ10 , since rb 6∈ σ6.2 . Intuitively,
join node t34 joins only common witness sets in τ17 and τ33 with χ(t17 ) = χ(t33 ) = χ(t34 ). In general, a
join node marks rules satisfied, which are marked satisfied in at least one child (see L13–14).
Since we already explained how to obtain models, we only briefly describe how to compute counterwitnesses. Family C consists of rows (C, ρ) where C ⊆ at(P ) ∩ χ(t) is a counter-witness set in t to M .
Similar to the sat-state σ, the sat-state ρ for C under M represents whether rules of the GL reduct PtM
are satisfied by a superset of C. We can see counter-witness sets together with its sat-state as counter0
witnesses. Thus, C witnesses the existence of C 0 ( M 0 satisfying C 0 (P<t ∪ ρ)M since M witnesses
a model M 0 ⊇ M where M 0 P<t . In consequence, there exists an answer set of P if the root table
contains h∅, ∅, ∅i. We require local-reducts for deciding satisfiability of counter-witness sets.
Definition 2. Let P be a program, T = (·, χ) be a TD of S(P ), t be a node of T , R ⊆ Pt and M ⊆ at(P ).
(t,M )
M
We define local-reduct R(t,M ) by [R(t) ] and Ṙ(t,M ) : R → 2R
by Ṙ(t,M ) (r) :={r}(t,M ) , r ∈ R.
Proposition 1 (c.f. [11]). Let P be a program and k :=tw (S(P )). Then, the algorithm DPSINC is correct
k+2
and runs in time O(22
· kS(P )k).
6
τ12
hM12.i , σ12.i i
h{ebc , ab }, ∅i
h{ebc }, {rc1 }i
h{ab }, {rc1 }i
h∅, {rc1 }i
hM12.i , σ12.i , C12.i i
τ12
T: ∅ t36
t15 rc1 , ac
h{ebc , ab }, ∅, {h{ebc }, {rc1 }i}i
hM32.i , σ32.i i
t34
h{ebc , ab }, ∅, ∅i
t12
r¬c , ac h{ac }, {rc2 }i h{ebc }, {rc1 }, ∅i
rc1 , ebc , ab
h∅, {rc2 }i
h{ab }, {rc1 }, {h∅, {rc1 }i}i
h∅, ∅i
t10
h{ab }, {rc1 }, ∅i
r ,a
τ32 h∅, {rc1 }, ∅i
rbc , ebc , ab c2 c t32
hM6.i , σ6.i i τ6
rc2 , ecd , ad t29
r ,e ,a
h{eab , ab }, {rb }i b ab b t6
h{eab }, ∅i
rb , eab t5 rcd , ecd , ad t27
h{ab }, {rb }i
h∅, {rb }i
eab t4
rd , ead , ad t23
hM3.i , σ3.i i τ3
r
,
e
t
ab
ab 3
h{eab }, {rab }i
rad , ead t20
h∅, {rab }i
eab t2
hM1.i , σ1.i i τ1
∅ t18
∅ t1
h∅, ∅i
hM6.i , σ6.i , C6.i i τ6
τ32
h{eab , ab }, {rb }, {
hM32.i , σ32.i , C32.i i
h{eab }, ∅i}i
h{ac }, {rc2 }, {
h{eab }, ∅, ∅i
h{ab }, {rb }, {h∅, {rb }i}i h{ac }, {rc2 }i,
h∅, {rc2 }i}i
h∅, {rb }, ∅i
h{ac }, {rc2 }, {
hM3.i , σ3.i , C3.i i τ3
h∅, {rc2 }i}i
h{eab }, {rab }, {h∅, ∅i}i h∅, {rc2 }, {
h∅, {rc2 }i}i
h∅, {rab }, ∅i
h∅, {rc2 }, ∅i
hM1.i , σ1.i , C1.i i
τ1 h∅, ∅, {h∅, {rc2 }i}i
h∅, ∅, ∅i
h∅, ∅, ∅i
Figure 2 A TD T of the semi-incidence graph S(P ) for program P from Example 1 (center). Selected
DP tables after DPMOD (left) and after DPSINC (right) for nice TD T .
4
DynASP2.5: Implementing a III Pass DP Algorithm
The classical DP algorithm DPSINC (Step 3 of Figure 1) follows a single pass approach. It computes
both witnesses and counter-witnesses by traversing the given TD exactly once. In particular, it stores
exhaustively all potential counter-witnesses, even those counter-witnesses where the witnesses in the
table of a node cannot be extended in the parent node. In addition, there can be a high number of
duplicates among the counter-witnesses, which are stored repeatedly. In this section, we propose a
multi-pass approach (M-DPSINC ) for DP on TDs and a new implementation (DynASP2.5), which fruitfully
adapts and extends ideas from a different domain [5]. Our novel algorithm allows for an early cleanup
(purging) of witnesses that do not lead to answer sets, which in consequence (i) avoids to construct
expendable counter-witnesses. Moreover, multiple passes enable us to store witnesses and counterwitnesses separately, which in turn (ii) avoids storing counter-witnesses duplicately and (iii) allows for
highly space efficient data structures (pointers) in practice when linking witnesses and counter-witnesses
together. Figure 1 (right, middle) presents the control flow of the new multi-pass approach DynASP2.5,
where M-DPSINC introduces a much more elaborate computation in Step 3.
4.1
The Algorithm
Our algorithm (M-DPSINC ) executed as Step 3 runs DPMOD , DPCMOD and DPLMOD,CMOD in three passes (3.I,
3.II, and 3.III) as follows:
3.I. First, we run the algorithm DPMOD , which computes in a bottom-up traversal for every node t
in the tree decomposition a table MOD-Tabs[t] of witnesses for t. Then, in a top-down traversal
for every node t in the TD remove from tables MOD-Tabs[t] witnesses, which do not extend to a
witness in the table for the parent node (“Purge non-witnesses”); these witnesses can never be
used to construct a model (nor answer set) of the program.
3.II. For this step, let CMOD be a table algorithm computing only counter-witnesses of SINC (blue
parts of Algorithm 2). We execute DPCMOD , compute for all witnesses counter-witnesses at once
and store the resulting tables in CMOD-Tabs[·]. For every node t, table CMOD-Tabs[t] contains
counter-witnesses to witness being ⊂-minimal. Again, irrelevant rows are removed (“Purge
non-counter-witnesses”).
7
Algorithm 3: Algorithm DPLW,C (T , W-Tabs, C-Tabs) for linking counter-witnesses to witnesses.
1
2
3
4
5
6
7
8
9
In: Nice TD T = (T, χ) with T = (N, ·, n) of a graph S(P ), and mappings W-Tabs[·], C-Tabs[·].
Out: W, C-Tabs: maps node t ∈ T to some pair (τtW , τtC ) with τtW ∈ W-Tabs[t], τtC ∈ C-Tabs[t].
Child-Tabst :={W, C-Tabs[t0 ] | t0 is a child of t in T }
/* Get for a node t tables of (preceeding) combined child rows (CCR)
*/
CCRt :=Π̂τ 0 ∈Child-Tabst τ 0
/* For Abbreviations see Footnote 6. */
/* Get for a row ~
u its combined child rows (origins)
*/
origt (~
u) :={S | S ∈ CCRt , ~
u ∈ τ, τ = W(t, χ(t), Pt , at≤t , fw (S))}
/* Get for a table S of combined child rows its successors (evolution)
*/
evolt (S) :={~
u|~
u ∈ τ, τ = C(t, χ(t), Pt , at≤t , τ 0 ), τ 0 ∈ S}
for iterate t in post-order(T,n) do
/* Compute counter-witnesses (≺-smaller rows) for a witness set M
*/
subs≺ (f, M, S) :={~
u|~
u ∈ C-Tabs[t], ~
u ∈ evolt (f (S)), ~
u = hC, · · · i, C ≺ M }
/* Link each witness ~
u to its counter-witnesses and store the results
*/
W, C-Tabs[t] ← {(~
u, subs( (fw , M, S) ∪ subs⊆ (fcw , M, S)) | ~
u ∈ W-Tabs[t], ~
u = hM, · · · i, S ∈ origt (u)}
hM3.i , σ3.i , C3.i i
h{a1 , a2 , f }, ∅, ∅i
h{a2 , f }, ∅, ∅i
τ3
hM1.i , σ1.i , C1.i i
τ1
h{a1 , a2 , f }, {rc , rcf }, {
h{a1 , f }, {rcf }i, h{a2 , f }, {rcf }i,
h{f }, {rcf }i, h{a1 , a2 }, {rc }i,
hM2.i , σ2.i , C2.i i τ2
h{a1 }, {rc }i, h{a2 }, {rc }i, h∅, {rc }i}i
h{a1 , a2 , f }, {rf , r2 }, ∅i h{a1 , a2 }, {rc , rcf }, {h{a1 }, {rc , rcf }i,
h{a1 , a2 }, {r2 }, {
h{a2 }, {rc , rcf }i, h∅, {rc , rcf }i}i
h{a1 }, ∅i,
h{a1 , f }, {rc , rcf }, {h{f }, {rcf }i,
h{a2 }, {r2 }i, h∅, ∅i}i
h{a1 }, {rc }i, h∅, {rc }i}i
h{a1 , f }, {rf }, ∅i
h{a1 }, {rc , rcf }, {h∅, {rc , rf }i}i
h{a1 }, ∅, {h∅, ∅i}i
h{a2 , f }, {rc , rcf }, {h{a2 }, {rc }i,
h{a2 , f }, {rf , r2 }, ∅i
h{f }, {rcf }i}i
h{a2 }, {r2 }, {h∅, ∅i}i
h{a2 }, {rc , rcf }, {h∅, {rc , rcf }i}i
h{f }, {rf }, ∅i
h{f }, {rc , rcf }, {h∅, {rc }i}i
h∅, ∅, ∅i
h∅, {rc , rcf }, ∅i
Figure 3
hM3.i , σ3.i , C3.i i
h{a1 , a2 , f }, ∅, ∅i
h{a2 , f }, ∅, ∅i
τ3
hM2.i , σ2.i , C2.i i
τ2
h{a1 , a2 , f }, {rf , r2 }, ∅i
h{a2 , f }, {rf , r2 }, ∅i
hM1.i , σ1.i , C1.i i
τ1
h{a1 , a2 , f }, {rc , rcf }, ∅
h{a2 , f }, {rc , rcf }, ∅i
hC1.i , ρ1.i i
τ1CMOD
h{a1 , f }, {rcf }i, h{a2 , f }, {rcf }i,
h{f }, {rcf }i, h{a1 , a2 }, {rc }i, h{
a1 }, {rc }i, h{a2 }, {rc }i, h∅, {rc }i
Selected DP tables after DPSINC (left) and after M-DPSINC (right) for TD T .
3.III. Finally, in a bottom-up traversal for every node t in the TD, witnesses and counter-witnesses are
linked using algorithm DPLMOD,CMOD (see Algorithm 3). DPLMOD,CMOD takes previous results and
maps rows in MOD-Tabs[t] to a table (set) of rows in CMOD-Tabs[t].
We already explained the table algorithms DPMOD and DPCMOD in the previous section. The main
part of our multi-pass algorithm is the algorithm DPLMOD,CMOD based on the general algorithm DPLW,C
(Algorithm 3) with W = MOD, C = CMOD, which links those separate tables together. Before we
quickly discuss the core of DPLW,C in Lines 5–9, note that Lines 2–4 introduce auxiliary definitions. Line 2
combines rows of the child nodes of given node t, which is achieved by a product over sets6 , where we
drop the order and keep sets only. Line 3 concerns determining for a row ~u its origins (finding preceding
combined rows that lead to ~u using table algorithm W). Line 4 covers deriving succeeding rows for a
certain child row combination its evolution rows via algorithm C. In an implementation, origin as well
as evolution are not computed, but represented via pointer data structures directly linking to W-Tabs[·]
or C-Tabs[·], respectively. Then, the table algorithm DPLW,C applies a post-order traversal and links
witnesses to counter-witnesses in Line 9. DPLW,C searches for origins (orig) of a certain witness ~u, uses
the counter-witnesses (fcw ) linked to these origins, and then determines the evolution (evol) in order to
derive counter-witnesses (using subs) of ~u.
6 For
Q
Q
set I = {1, . . . , n} and sets Si , we define i∈I Si :=S1 × · · · × Sn = {(s1 , . . . , sn ) : si ∈ Si }. Moreover, for
i∈I Si ,
Q
Q
Q
let ˆ i∈I Si :={{{s1 }, . . . , {sn }} | (s1 , . . . ,S
sn ) ∈ i∈I Si }. If for each S ∈ ˆ i∈IS
Si and {si } ∈ S, si is a pair with a witness
and a counter-witness part, let fw (S) := {(Wi ,Ci )}∈S {{Wi }} and fcw (S) := {(Wi ,Ci )}∈S {{Ci }} restrict S to the resp.
(counter-)witness parts.
8
Theorem 1. For a program P of semi-incidence treewidth k := tw (S(P )), the algorithm M-DPSINC is
k+2
correct and runs in time O(22
· kP k).
Proof (Sketch). Due to space constraints, we only sketch the proof idea for enumerating answer sets of
disjunctive ASP programs by means of M-DPSINC . Let P be a disjunctive program and k :=tw (S(P )).
We establish a reduction R(P, k) of EnumAsp to EnumMinSAT1, such that there is a one-to-one
correspondence between answer sets and models of the formula, more precisely, for every answer set M
of P and for the resulting instance (F, k 0 ) = R(P, k) the set M ∪ {sol} is a subset-minimal model of F
03
and k 0 = tw(I(F )) with k 0 ≤ 7k + 2. We compute in time 2O(k ) · kF k a TD of width at most k 0 [6]
and add sol to every bag. Using a table algorithm designed for SAT [22] we compute witnesses and
counter-witnesses. Conceptually, one could also modify MOD for this task. To finally show correctness
of linking counter-witnesses to witnesses as presented in DPLMOD,MOD , we have to extend earlier work [5,
Theorem 3.25 and 3.26]. Therefore, we enumerate subset-minimal models of F by following each
witness set containing sol at the root having counter-witnesses ∅ back to the leaves. This runs in
(7k+2)+2
time O(22
· kP k), c.f., [5, 11]. A more involved (direct) proof, allows to decrease the runtime
k+2
to O(22
· kP k) (even for choice rules).
Example 4. Let k be some integer and Pk be some program that contains the following rules rc :={a1 , · · · , ak } ←
f , r2 := ← ¬a2 , . . ., rk := ← ¬ak , and rf := ← ¬f and rcf :={f } ← . The rules r1 , . . ., rk simulate
that only certain subsets of {a1 , · · · , ak } are allowed. Rules rf and rcf enforce that f is set to true. Let
T = (T, χ, t3 ) be a TD of the semi-incidence graph S(Pk ) of program Pk where T = (V, E) with V =
{t1 , t2 , t3 }, E = {(t1 , t2 ), (t2 , t3 )}, χ(t1 ) = {a1 , · · · , ak , f, rc , rcf }, χ(t2 ) = {a1 , · · · , ak , r2 , · · · , rk , rf },
and χ(t3 ) = ∅. Figure 3 (left) illustrates the tables for program P2 after DPSINC , whereas Figure 3 (right)
presents tables using M-DPSINC , which are exponentially smaller in k, mainly due to cleanup. Observe that
Pass 3.II M-DPSINC , “temporarily” materializes counter-witnesses only for τ1 , presented in table τ1CMOD .
Hence, using multi-pass algorithm M-DPSINC results in an exponential speedup. Note that we can trivially
extend the program such that we have the same effect for a TD of minimum width and even if we take
the incidence graph. In practice, programs containing the rules above frequently occur when encoding
by means of saturation [9]. The program Pk and the TD T also reveal that a different TD of the same
width, where f occurs already very early in the bottom-up traversal, would result in a smaller table τ1
even when running DPSINC .
4.2
Implementation Details
Efficient implementations of dynamic programming algorithms on TDs are not a by-product of computational complexity theory and involve tuning and sophisticated algorithm engineering. Therefore, we
present additional implementation details of algorithm M-DPSINC into our prototypical multi-pass solver
DynASP2.5, including two variations (depgraph, joinsize TDs).
Even though normalizing a TD can be achieved without increasing its width, a normalization may
artificially introduce additional atoms. Resulting in several additional intermediate join nodes among
such artificially introduced atoms requiring a significant amount of total unnecessary computation in
practice. On that account, we use non-normalized tree decompositions. In order to still obtain a fixedparameter linear algorithm, we limit the number of children per node to a constant. Moreover, linking
counter-witnesses to witnesses efficiently is crucial. The main challenge is to deal with situations where
a row (witness) might be linked to different set of counter-witnesses depending on different predecessors
of the row (hidden in set notation of the last line in Algorithm 3). In these cases, DynASP2.5 eagerly
creates a “clone” in form of a very light-weighted proxy to the original row and ensures that only the
original row (if at all required) serves as counter-witness during pass three. Together with efficient caches
of counter-witnesses, DynASP2.5 reduces overhead due to clones in practice.
Dedicated data structures are vital. Sets of Witnesses and satisfied rules are represented in the
DynASP2.5 system via constant-size bit vectors. 32-bit integers are used to represent by value 1 whether
an atom is set to true or a rule is satisfied in the respective bit positions according to the bag. A
restriction to 32-bit integers seems reasonable as we assume for now (practical memory limitations)
9
750
500
CPU time [s]
1000
St (best TD)
0
1000
750
500
250
# instances
20
40
Clasp 3.3.0
DynASP2.5
DynASP2.5 depgraph
DynASP2.5 joinsize
60
CPU time [s]
1250
1250
250
0
St (avg TD)
# instances
20
40
Clasp 3.3.0
DynASP2.5
DynASP2.5 depgraph
DynASP2.5 joinsize
60
solver\time
#
Avg solved runtime
TO TD 3.I 3.II 3.III
Clasp 3.3.0
34
-
-
-
-
DynASP2
68
-
-
-
-
DynASP2.5 28
Σ
334.5
-
0.5 3.5 16.6 264.4 285.0
...depgraph
24 10.9 3.4 15.3 251.3 280.9
...joinsize
32
0.7 3.5 16.5 278.3 299.0
Figure 4 Cactus plots showing best and average runtime among five TDs (left). Number of Timeouts
(TO) and average runtime among solved instances (right).
that our approach works well on TDs of width ≤ 20. Since state-of-the-art computers handle such
constant-sized integers extremely efficient, DynASP2.5 allows for efficient projections and joins of rows,
and subset checks in general. In order to not recompute counter-witnesses (in Pass 3.II) for different
witnesses, we use a three-valued notation of counter-witness sets consisting of atoms set to true (T) or
false (F) or false but true in the witness set (TW) used to build the reduct. Note that the algorithm
enforces that only (TW)-atoms are relevant, i.e., an atom has to occur in a default negation or choice
rule.
Minimum width is not the only optimization goal when computing TDs by means of heuristics.
Instead, using TDs where a certain feature value has been maximized in addition (customized TDs)
works seemingly well in practice [1, 21]. While DynASP2.5 (M-DPSINC ) does not take additional TD
features into account, we also implemented a variant (DynASP2.5 depgraph), which prefers one out of
ten TDs that intuitively speaking avoids to introduce head atoms of some rule r in node t, without
having encountered every body atom of r below t, similar to atom dependencies in the program [15].
The variant DynASP2.5 joinsize minimizes bag sizes of child nodes of join nodes, c.f. [2].
4.3
Experimental Evaluation
We performed experiments to investigate the runtime behavior of DynASP2.5 and its variants, in order
to evaluate whether our multi-pass approach can be beneficial and has practical advantages over the
classical single pass approach (DynASP2). Further, we considered the dedicated ASP solver Clasp 3.3.07 .
Clearly, we cannot hope to solve programs with graph representations of high treewidth. However,
programs involving real-world graphs such as graph problems on transit graphs admit TDs of acceptable
width to perform DP on TDs. To get a first intuition, we focused on the Steiner tree problem (St) for
our benchmarks. Note that we support the most frequently used SModels input format [24] for our
implementation.
We mainly inspected the CPU time using the average over five runs per instance (five fixed
seeds allow certain variance for heuristic TD computation). For each run, we limited the environment to 16 GB RAM and 1200 seconds CPU time. We used Clasp with options “--stats=2 --optstrategy=usc,pmres,disjoint,stratify --opt-usc-shrink=min -q”, which enable very recent improvements
for unsatisfiable cores [3], and disabled solution printing/recording. We also benchmarked Clasp with
branch-and-bound, which was, however, outperformed by the unsat. core options on all our instances.
Note that without the very recent unsatisfiable core advances Clasp timed out on almost every instance.
We refer to an extended version [12] for more details on the benchmark instances, encodings, and
benchmark environment. The left plot in Figure 4 shows the result of always selecting the best among
five TDs, whereas the right plot concerns average runtime. The table in Figure 4 reports on average
running times (TD computation and Passes 3.I, 3.II, 3.III) among the solved instances and the total
number of timeouts (TO). We consider an instance to time out, when all five TDs exceeded the limit. For
the variants depgraph and joinsize, runtimes for computing and selecting among ten TDs are included.
Our empirical benchmark results confirm that DynASP2.5 exhibits competitive runtime behavior even
for TDs of treewidth around 14. Compared to state-of-the-art ASP solver Clasp, DynASP2.5 is capable
7 Clasp
is available at https://github.com/potassco/clasp/releases/tag/v3.3.0.
10
of additionally delivering the number of optimal solutions. In particular, variant “depgraph” shows
promising runtimes.
5
Conclusion
In this paper, we presented a novel approach for ASP solving based on ideas from parameterized
complexity. Our algorithms runs in linear time assuming bounded treewidth of the input program. Our
solver applies DP in three passes, thereby avoiding redundancies. Experimental results indicate that our
ASP solver is competitive for certain classes of instances with small treewidth, where the latest version
of the well-known solver Clasp hardly keeps up. An interesting question for future research is whether a
linear amount of passes (incremental DP) can improve the runtime behavior.
References
[1] M. Abseher, F. Dusberger, N. Musliu, and S. Woltran. Improving the efficiency of dynamic
programming on tree decompositions via machine learning. JAIR, 58:829–858, 2017.
[2] M. Abseher, N. Musliu, and S. Woltran. htd – a free, open-source framework for (customized) tree
decompositions and beyond. In CPAIOR’17, 2017. To appear.
[3] M. Alviano and C. Dodaro. Anytime answer set optimization via unsatisfiable core shrinking.
TPLP, 16(5-6):533—551, 2016.
[4] M. Alviano, C. Dodaro, W. Faber, N. Leone, and F. Ricca. WASP: A native ASP solver based on
constraint learning. In LPNMR’13, volume 8148 of LNCS, pages 54–66. Springer, 2013.
[5] B. Bliem, G. Charwat, M. Hecher, and S. Woltran. D-FLATˆ2: Subset minimization in dynamic
programming on tree decompositions made easy. Fund. Inform., 147:27–34, 2016.
[6] H. Bodlaender and A. M. C. A. Koster. Combinatorial optimization on graphs of bounded treewidth.
The Computer J., 51(3):255–269, 2008.
[7] H. Dell. The 2st parameterized algorithms and computational experiments challenge – Track A:
Treewidth. Technical report, 2017.
[8] A. Durand, M. Hermann, and P. G. Kolaitis. Subtractive reductions and complete problems for
counting complexity classes. Th. Comput. Sc., 340(3), 2005.
[9] T. Eiter and G. Gottlob. On the computational cost of disjunctive logic programming: Propositional
case. Ann. Math. Artif. Intell., 15(3–4):289–323, 1995.
[10] J. K. Fichte, M. Hecher, M. Morak, and S. Woltran. Answer Set Solving using Tree Decompositions
and Dynamic Programming - The DynASP2 System -. Technical Report DBAI-TR-2016-101, TU
Wien, 2016.
[11] J. K. Fichte, M. Hecher, M. Morak, and S. Woltran. Answer set solving with bounded treewidth
revisited. In LPNMR’17, 2017. To appear.
[12] J. K. Fichte, M. Hecher, and S. Woltran. DynASP2.5: Dynamic Programming on Tree Decompositions in Action. CoRR, abs/cs/arXiv:1702.02890, 2017.
[13] M. Gebser, J. Bomanson, and T. Janhunen. Rewriting optimization statements in answer-set
programs. Technical Communications of ICLP 2016, 2016.
[14] M. Gebser, M. Maratea, and F. Ricca. What’s hot in the answer set programming competition. In
AAAI’16, pages 4327–4329. The AAAI Press, 2016.
11
[15] Georg Gottlob, Francesco Scarcello, and Martha Sideri. Fixed-parameter complexity in AI and
nonmonotonic reasoning. AIJ, 138(1-2):55–86, 2002.
[16] M. Hermann and R. Pichler. Complexity of counting the optimal solutions. Th. Comput. Sc.,
410(38–40), 2009. URL: http://dx.doi.org/10.1016/j.tcs.2009.05.025.
[17] M. Jakl, R. Pichler, and S. Woltran. Answer-set programming with bounded treewidth. In IJCAI’09,
2009.
[18] T. Janhunen and I. Niemelä. The answer set programming paradigm. 2016.
[19] B. Kaufmann, M. Gebser, R. Kaminski, and T. Schaub. clasp – a conflict-driven nogood learning
answer set solver, 2015.
[20] C. Koch and N. Leone. Stable model checking made easy. In IJCAI’99, 1999.
[21] M. Morak, N. Musliu, R. Pichler, S. Rümmele, and S. Woltran. Evaluating tree-decomposition
based algorithms for answer set programming. In LION’12, LNCS, pages 130–144. Springer, 2012.
[22] M. Samer and S. Szeider. Algorithms for propositional model counting. J. Desc. Alg., 8(1), 2010.
[23] P. Simons, I. Niemelä, and T. Soininen. Extending and implementing the stable model semantics.
AIJ, 138(1-2):181–234, 2002.
[24] T. Syrjänen. Lparse 1.0 user’s manual. tcs.hut.fi/Software/smodels/lparse.ps, 2002.
12
A
Additional Example
We assume again row numbers per table τt , i.e., ut.i = hMt.i , σt.i , Ct.i i is the ith -row. Further, for each
counter-witness hCt.i.j , ρt.i.j i ∈ Ct.i , j marks its “order” (as depicted in Figure 2 (right)) in set Ct.i .
Example 5. Again, we consider P of Example 1 and T = (·, χ) of Figure 2 as well as tables τ1 , . . ., τ34
of Figure 2 (right) using DPSINC . We only discuss certain tables. Table τ1 = {h∅, ∅, ∅i} as type(t1 ) = leaf.
Node t2 introduces atom eab , resulting in table {h{eab }, ∅, {(∅, ∅)}i, h∅, ∅, ∅i} (compare to Algorithm 2
L3–5). Then, node t3 introduces rule rab , which is removed in node t4 . Note that C3.1.1 = h∅, ∅i ∈ C3.1.1
does not have a “successor row” in table τ4 since rab is not satisfied (see L11 and Definition 2). Table τ6
is then the result of a chain of introduce nodes, and contains for each witness set M6.i every possible
counter-witness set C6.i.j with C6.i.j ( M6.i . We now discuss table τ12 , intuitively containing (a
projection of ) (counter-)witnesses of τ10 , which satisfy rule rbc after introducing rule rc1 . Observe that
there is no succeeding witness set for M6.2 = {eab } in τ10 (nor τ12 ), since eab ∈ M6.2 , but ab 6∈ M6.2
(required to satisfy rb ). Rows u12.1 , u12.4 form successors of u6.3 , while rows u12.2 , u12.5 succeed u6.1 ,
since counter-witness set C6.1.1 has no succeeding row in τ10 because it does not satisfy rb . Remaining
rows u12.3 , u12.6 have “origin” u6.4 in τ6 .
B
B.1
Omitted Proofs
Correctness of DynASP2.5
Bliem et al. [5] have shown that augmentable W-Tabs can be transformed into W, W-Tabs, which easily
allows reading off subset-minimal solutions starting at the table W, W-Tabs[n] for TD root n. We follow
their concepts and define a slightly extended variant of augmentable tables. Therefore, we reduce the
problem of enumerating disjunctive programs to EnumMinSAT1 and show that the resulting tables
of algorithm MOD (see Algorithm 2) are augmentable. In the end, we apply an earlier theorem [5]
transforming MOD-Tabs obtained by DPMOD into MOD, MOD-Tabs via the augmenting function aug(·)
proposed in their work. To this extent, we use auxiliary definitions Child-Tabst , origt (·) and evolt (·)
specified in Algorithm 3.
Definition 3. Let T = (T, χ) be a TD where T = (N, ·, ·), W be a table algorithm, t ∈ N , and
τ ∈ W-Tabs[t] be the table for node t. For tuple ~u = hM, σ, · · · i ∈ τ, we define α(~u) :=M , β(~u) :=σ. We
inductively define
[
[
α∗ (τ) :=
α(~u) ∪
α∗ (τ 0 ), and
τ 0 ∈Child-Tabst
~
u∈τ
β ∗ (τ) :=
[
[
β(~u) ∪
β ∗ (τ 0 ).
τ 0 ∈Child-Tabst
~
u∈τ
Moreover, we inductively define the extensions of a row ~u ∈ τ as
[
E(~u) := {~u} ∪ U | U ∈
{τ1 ∪ · · · ∪ τk | τi ∈ E(~u0i ) for all 1 ≤ i ≤ k} .
{{~
u01 },...,{~
u0k }}∈origt (~
u)
Remark 1. Any extension U ∈ E(~u) contains ~u and exactly one row from each table
Q that is a descendant
of τ. If ~u is a row of a leaf table, E(~u) = {{~u}} since origt (~u) = {∅} assuming i∈∅ Si = {()}.
Definition 4. Let τn be the table in W-Tabs for TD root n. We define the set sol(W-Tabs) of solutions
of W-Tabs as sol(W-Tabs) :={α∗ (U ) | ~u ∈ τn , U ∈ E(~u)}
Definition 5. Let τ be a table in W-Tabs such that τ10 , . . . , τk0 are the child tables Child-Tabst and let
~u, ~v ∈ τ. We say that x ∈ X(~u) has been X−illegally introduced at ~u if there are {{~u01 }, . . . , {~u0k }} ∈
origt (~u) such that for some 1 ≤ i ≤ k it holds that x ∈
/ X(~u0i ) while x ∈ X ∗ (τi0 ). Moreover, we say that
x ∈ X(~v ) \ X(~u) has been X−illegally removed at ~u if there is some U ∈ E(~u) such that x ∈ X(U ).
13
Definition 6. We call a table τ augmentable if the following conditions hold:
1. For all rows of the form hM, · · · , Ci, we have C = ∅.
2. For all ~u, ~v ∈ τ with ~u 6= ~v it holds that α(~u) ∪ β(~u) 6= α(~v ) ∪ β(v).
3. For all ~u = hM, σ, · · · i ∈ τ, {{~u01 }, . . . , {~u0k }} ∈ origt (~u), 1 ≤ i < j ≤ k, I ∈ E(~u0i ) and J ∈ E(~u0j )
it holds that α∗ (I) ∩ α∗ (J) ⊆ M and β ∗ (I) ∩ β ∗ (J) ⊆ σ.
4. No element of α∗ (τ) has been α-illegally introduced and no element of β ∗ (τ) has been β-illegally
introduced.
5. No element of α∗ (τ) has been α-illegally removed and no element of β ∗ (τ) has been β-illegally
removed.
We call W-Tabs augmentable if all its tables are augmentable.
It is easy to see that MOD-Tabs are augmentable, that is, Algorithm 2 (DPMOD (·)) computes only
augmentable tables.
Observation 1. MOD-Tabs are augmentable, since DPMOD (·) computes augmentable tables. CMOD-Tabs
are augmentable, since DPCMOD (·) computes augmentable tables.
The following theorem establishes that we can reduce an instance of EnumAsp (restricted to disjunctive input programs) when parameterized by semi-incidence treewidth to an instance of EnumMinSAT1
when parameterized by the treewidth of its incidence graph.
Lemma 1. Given a disjunctive program P of semi-incidence treewidth k = tw (S(P )). We can produce
in time O(kP k) a propositional formula F such that the treewidth k 0 of the incidence graph I(F )8 is
k 0 ≤ 7k + 2 and the answer sets of P and subset-minimal models of F * are in a particular one-toone correspondence. More precisely, M is an answer set of P if and only if M ∪ Maux ∪ {sol} is a
subset-minimal model of F where Maux is a set of additional variables occurring in F , but not in P and
variables introduced by Tseitin normalization.
Proof. Let P be a disjunctive program of semi-incidence treewidth k = tw (S(P )). First, we construct
a formula F consisting of a conjunction over formulas F r , F impl , F sol , F min followed by Tseitin
normalization of F to obtain F * . Among the atoms9 of our formulas will the atoms at(P ) of the program.
Further, for each atom a such that a ∈ B − (r) for some rule r ∈ P , we introduce
atom a0 . In
S a fresh
−
− 0
0
0
−
:=
the following, we denote by Z the set {z : z ∈ Z} for any set Z and by BP
r∈P Br . Hence, (BP )
denotes a set of fresh atoms for atoms occurring in any negative body. Then, we construct the following
formulas:
F r (r) :=Hr ∨ ¬Br+ ∨ (Br− )0
F
impl
F
for r ∈ P
0
(a) :=a → a
for a ∈
0
sol
(a) :=sol → (a → a)
_
F min :=¬sol →
(a0 ∧ ¬a)
for a ∈
BP−
BP−
(1)
(2)
(3)
(4)
− 0
a0 ∈(BP
)
F :=
^
r∈P
F r (r) ∧
^
−
a∈BP
F impl (a) ∧
^
F sol (a) ∧ F min
(5)
−
a∈BP
Next, we show that M is an answer set of P if and only if M ∪ ([M ∩ BP− ])0 ∪ {sol} is a subset-minimal
model of F .
8 The incidence graph I(F ) of a propositional formula F in CNF is the bipartite graph that has the variables and
clauses of F as vertices and an edge v c if v is a variable that occurs in c for some clause c ∈ F [22].
9 Note that we do not distinguish between atoms and propositional variables in terminology here.
14
(⇒): Let M be an answer set of P . We transform M into Y :=M ∪ ([M ∩ BP− ])0 ∪ {sol}. Observe that Y
satisfies all subformulas of F and therefore Y F . It remains to show that Y is a minimal model of F .
Assume towards a contradiction that Y is not a minimal model. Hence, there exists X with X ( Y . We
distinguish the following cases:
1. sol ∈ X: By construction of F we have X a0 ↔ a for any a0 ∈ (BP− )0 , which implies that
X ∩ at(P ) P M . However, this contradicts our assumption that M is an answer set of P .
2. sol 6∈ X: By construction of F there is at least one atom a ∈ BP− with a0 ∈ X, but a 6∈ X.
Consequently, X ∩ at(P ) P M . This contradicts again that M is an answer set of P .
(⇐): Given a formula F that has been constructed from a program P as given above. Then, let Y be a
subset-minimal model of F such that sol ∈ Y . By construction we have for every a0 ∈ Y ∩ (BP− )0 that
a ∈ Y . Hence, we let M = at(P ) ∩ Y . Observe that M satisfies every rule r ∈ P according to (A1) and
is in consequence a model of P . It remains to show that M is indeed an answer set. Assume towards a
contradiction that M is not an answer set. Then there exists a model N ( M of the reduct P M . We
distinguish the following cases:
1. N is not a model of P : We construct X :=N ∪ [Y ∩ (BP− )0 ] and show that X is indeed a model
of F . For this, for every r ∈ P where B − (r) ∩ M 6= ∅ we have X F r (r), since (Y ∩ (BP− )0 ) ⊆ X
by definition of X. For formulas (A1) constructed by F r (r) using remaining rules r, we also have
X F r (r), since N {r}M . In conclusion, X F and X ( Y , and therefore X contradicts Y is
a subset-minimal model of F .
2. N is also a model of P : Observe that then X :=N ∪ [N ∩ BP− ]0 ∪ {sol} is also a model of F , which
contradicts optimality of Y since X ( Y .
By Tseitin normalization, we obtain F ∗ , thereby introducing fresh atoms la0 for each a0 ∈ (BP− )0 :
F r* (r) :=Hr ∨ ¬Br+ ∨ (Br− )0
impl*
F
F
(a) :=¬a ∨ a
0
for r ∈ P
for a ∈
0
sol*
(a) :=¬sol ∨ (¬a ∨ a)
_
F1min :=sol ∨
(la0 )
for a ∈
(A1*)
BP−
BP−
(A2*)
(A3*)
(A4.1*)
−
a0 ∈BP
0
F2min (a) :=¬a0 ∨ a ∨ la0
F3min (a) :=¬la0
F4min (a) :=¬la0
0
for a ∈ BP−
(A4.2*)
BP−
BP−
(A4.3*)
∨a
for a ∈
∨ ¬a
for a ∈
(A4.4*)
Observe that the Tseitin normalization is correct and that there is a bijection between models of F ∗
and F .
Observe that our transformations runs in linear time and that the size of F * is linear in kP k. It
remains to argue that tw(I(F * )) ≤ 7k + 2. For this, assume that T = (T, χ, n) is an arbitrary but fixed
TD of S(P ) of width w. We construct a new TD T 0 :=(T, χ0 , n) where χ0 is defined as follows. For each
TD node t,
[
χ0 (t) :=
{a0 , la0 } ∪ [χ(t) ∩ at(P )] ∪ {sol} ∪ cl(t)
−
a∈BP
∩χ(t)
where
cl(t) :=
[
[F impl* (a), F sol* (a), F2min (a), F3min (a), F4min (a)] ∪ {F1min } ∪
−
a∈BP
∩χ(t)
[
{F r* (r)}.
r∈P ∩χ(t)
It is easy to see that T 0 is indeed a TD for I(F ∗ ) and that the width of T 0 is at most 7w + 2.
15
Definition 7. We inductively define an augmenting function aug(W-Tabs) that maps each table τ ∈
W-Tabs[t] for node t from an augmentable table to a table in W, W-Tabs[t]. Let the child tables of τ be
called τ10 , . . . , τk0 . For any 1 ≤ i ≤ k and ~u ∈ τi , we write res(~u) to denote {~v ∈ aug(τi0 ) | α(~u) = α(~v )}.
We define aug(τ) as the smallest table that satisfies the following conditions:
Q
1. For any ~u ∈ τ, {{~u01 }, . . . , {~u0k }} ∈ origt (~u) and {{~v10 }, . . . , {~vk0 }} ∈ ˆ 1≤i≤k res(~u0i ), there is a row
~v ∈ aug(τ) with α(~u) = α(~v ) and {{~v10 }, . . . , {~vk0 }} ∈ origt (~v ).
2. For any ~u, ~v ∈ aug(τ) with ~u = h· · · , Ci such that α(~v ) ⊆ α(~u), {{~u01 }, . . . , {~u0k }} ∈ origt (~u) and
{{~v10 }, . . . , {~vk0 }} ∈ origt (~v ) the following holds: Let 1 ≤ i ≤ k with ~u0i = h· · · , Ci i, ~ci = hCi , · · · i ∈
(Ci ∪ {~u0i }) with Ci ⊆ α(~vi0 ), and 1 ≤ j ≤ k with ~cj 6= ~u0j or α(~v ) ( α(~u). Then, there is a row
~c ∈ C with α(~c) ⊆ α(~v ) if and only if ~c ∈ τ and {{~c1 }, . . . , {~ck }} ∈ origt (~c).
For W-Tabs, we write aug(W-Tabs) to denote the result isomorphic to W, W-Tabs where each table τ in
W-Tabs corresponds to aug(τ).
Proposition 2. Let W-Tabs be augmentable. Then,
sol(aug(W-Tabs)) = {M ∈ sol(W-Tabs) | @M 0 ∈ sol(W-Tabs) : M 0 ( M }.
Proof (Sketch). The proof follows previous work [5]. We sketch only differences from their work. Any
row ~u ∈ τ of any table τ not only consists of set α(~u) being subject to subset-minimization and relevant
to solving EnumAsp. In addition, our definitions presented above also allow “auxiliary” sets β(~u) per
row ~u, which are not subject to the minimization. Moreover, by the correctness of the table algorithm
INC by Fichte and Szeider [11], we only require to store a set C of counter-witnesses hC, · · · i ∈ C per
witness set M , where each C forms a strictly ⊂-smaller model of M . As a consequence, there is no need
to differ between sets of counter-witnesses, which are strictly included or not, see [5]. Finally, we do not
need to care about duplicate rows (solved via compression function compr(·) in [5]) in τ, since τ is a set.
(7k+4)
Theorem 2. EnumAsp when the input is restricted to disjunctive programs can be solved in time 22
kP k computing aug(DPMOD (·)), where k refers to the treewidth of S(P ).
·
Proof (Sketch). First, we use reduction R(P, k) = (F ∗ , k 0 ) defined in Lemma 1 to construct an instance
of SAT given our disjunctive ASP program P . Note that k 0 = tw(I(F ∗ )) ≤ 7k + 2. Then, we can
03
compute in time 2O(k ) · |I(F ∗ )| a tree decomposition of width at most k 0 [6]. Note that since we require
to look for solutions containing sol at the root, we modify each bag of T such that it contains sol.
We call the resulting tree decomposition T 0 . We compute aug(DPMOD (T 0 )) using formula F ∗ (see
Algorithm 2). Finally, by Theorem 2 and Lemma 1, we conclude that answer sets of P correspond to
{M ∈ sol(aug(DPMOD (T 0 ))) | sol ∈ M, @M 0 ∈ sol(DPMOD (T 0 )) : M 0 ( M }.
The complexity proof sketched in [5] only cares about the runtime being polynomial. In fact, the
algorithm can be carried out in linear time, following complexity proofs presented by Fichte et al. [11],
(7k+4)
which leads to a worst-case runtime of 22
· kP k.
We can now even provide a “constructive definition” of the augmenting function aug(·).
Proposition 3. The resulting table aug(W-Tabs) obtained via DPW (T ) for any TD T is equivalent to
the table DPLW,W (T ) as given in Algorithm 3.
Proof (Idea). Intuitively, (1.) of Definition 7 concerns completeness, i.e., ensures that no row is left out
during the augmentation, and is ensured by row 7 of Algorithm 3 since each ~u ∈ W-Tabs is preserved.
The second condition (2.) enforces that there is no missing counter-witness for any witness, and the idea
is that whenever two witnesses ~u, ~v ∈ τ are in a subset relation (α(~v ) ⊆ α(~u)) and their corresponding
linked counter-witnesses (fcw ) of the corresponding origins (orig) are in a strict subset relation, then
there is some counter-witness c for u if and only if ~c ∈ τ is the successor (evol) of these corresponding
linked counter-witnesses. Intuitively, we can not miss any counter-witnesses in DPLW,W (T ) required by
16
(2.), since this required that there are two rows ~u0 , ~v 0 ∈ τ 0 with α(~v ) = α(~u) for one table τ 0 . Now let
the corresponding succeeding rows ~u, ~v ∈ τ (i.e., ~u ∈ evolt ({{~u0 }}), ~v ∈ evolt ({{~v 0 }}), respectively) with
α(~v ) ( α(~u), β(~v ) 6⊆ β(~u) and β(~v ) 6⊇ β(~u), mark the first encounter of a missing counter-witness. Since
β(~v ) is incomparable to β(~u), we conclude that the first encounter has to be in a table preceeding τ. To
conclude, one can show that DPLW,W (T ) does not contain “too many” rows, which do not fall under
conditions (1.) and (2.).
Theorem 2 works not only for disjunctive ASP via reduction to EnumMinSAT1, where witnesses
and counter-witnesses are derived with the same table algorithm MOD. In fact, one can also link
counter-witnesses to witnesses by means of DPLW,C (·), thereby using table algorithms W, C for computing
witnesses and counter-witnesses, respectively. In order to show correctness of algorithm DPLMOD,CMOD (·)
(Theorem 1) working for any ASP program, it is required to extend the definition of the augmenting
function aug(·) such that it is capable of using two different tables.
Proposition 4. Problem EnumAsp can be solved in time f (k)·kP k computing DPLMOD,CMOD (·), where k
refers to the treewidth of S(P ) and f is a computable function.
C
C.1
Additional Information on the Benchmarks
Benchmark Sets
In this paper, we mainly presented10 the Steiner tree problem using public transport networks. We
also considered benchmarks for counting answer sets as carried out in earlier work [11]. DynASP2.5
performs slightly better than DynASP2 on those instances. Hence, we do not report them here.
C.1.1
Transit Graphs
The instance graphs have been extracted from publicly available mass transit data feeds and splited by
transportation type, e.g., train, metro, tram, combinations. We heuristically computed tree decompositions [2] and obtained relatively fast decompositions of small width unless detailed bus networks were
present. Among the graphs considered were public transit networks of the cities London, Bangladesh,
Timisoara, and Paris.
C.1.2
Steiner Tree Problem
Picture yourself as the head of a famous internet service provider, which is about to provide high-speed
internet for their most-prestigious customers in public administration in order to increase productivity
levels beyond usual standards. However, these well-paying customers have to be connected via expensive
fibre cables. The good news is that the city council already confirmed that you are allowed to use
existing cable ducts, which basically adhere to the city’s transit network. We assumed for simplicity,
that edges have unit costs, and randomly generated a set of terminal stations — which are compliant
with the customers — among our transit stations (vertices). The goal is to search for a set of transit
connections of minimal cardinality such that the customers can be connected for example when putting
fibre cables along the transit network.
An encoding for this problem is depicted in Listing 1 and assumes a specification of the graph
(via edge) and the facilities (terminalVertex ) of our customers in public administration as well as the
number (numVertices) of vertices. The encoding is based on the saturation technique [9] and in fact
outperformed a different encoding presented in Listing 2 on all our instances using both solvers, Clasp
and DynASP2.5. At first sight, this observation seems quite surprising, however, we benchmarked on
more than 60 graphs with 10 varying decompositions for each solver variant and additional configurations
and different encodings for Clasp.
10 Benchmarks,
encodings, and results are available at https://github.com/daajoe/dynasp experiments/tree/ipec2017.
17
%
%
%
vertex ( X ) ← edge (X , _ ) .
vertex ( Y ) ← edge (_ , Y ) .
edge (X , Y ) ← edge (Y , X ) .
%
0 { selectedEdge (X , Y ) } 1 ← edge (X , Y ) , X < Y .
%
s1 ( X ) ∨ s2 ( X ) ← vertex ( X ) .
%
saturate
saturate
%
%
saturate
saturate
← selectedEdge (X , Y ) , s1 ( X ) , s2 ( Y ) , X < Y .
← selectedEdge (X , Y ) , s2 ( X ) , s1 ( Y ) , X < Y .
← N #count { X : s1 ( X ) , term inalVert ex ( X ) } , numVertices ( N ) .
← N #count { X : s2 ( X ) , term inalVert ex ( X ) } , numVertices ( N ) .
s1 ( X ) ← saturate , vertex ( X ) .
s2 ( X ) ← saturate , vertex ( X ) .
← not saturate .
%
#minimize { 1 ,X , Y : selectedEdge (X , Y ) }.
Listing 1
Encoding for St
%
%
edge (X , Y ) ← edge (Y , X ) .
{ selectedEdge (X , Y ) : edge (X , Y ) , X < Y }.
%
reached ( Y ) ← Y = #min { X : terminal Vertex ( X ) }.
reached ( Y ) ← reached ( X ) , selectedEdge (X , Y ) .
reached ( Y ) ← reached ( X ) , selectedEdge (Y , X ) .
← term inalVert ex ( X ) , not reached ( X ) .
%
#minimize { 1 ,X , Y : selectedEdge (X , Y ) }.
Listing 2
Alternative encoding for St
Example 6 (Steiner Tree). Let G = (V, E) be a graph and VT ⊆ V . A uniform Steiner tree on G is a
subgraph SG = (VS , ES ) of G such that VT ⊆ VS and for each distinct pair v, w ∈ VT there is a path
from v to w. The Steiner tree problem Enum-St asks to output all uniform Steiner trees. We encode
Enum-St into an ASP program as follows: Among the atoms of our program will be an atom av for each
vertex v ∈ VT , and an atom evw for each edge vw ∈ E assuming v < w for an arbitrary, but fixed total
ordering < among V . Let s be an arbitrary vertex s ∈ VT . We generate program P (G, VT ) :={{evw } ←
;
evw | vw ∈ E} ∪ {av ← aw , evw ; aw ← av , evw | vw ∈ E, v < w} ∪ {← ¬av | v ∈ VT } ∪ {as ← }.
It is easy to see that the answer sets of the program and the uniform Steiner trees are in a one-to-one
correspondence.
C.1.3
Other Graph Problems
We refer to the technical report [10] for comprehensive benchmark results using common graph problems
and comparing different table algorithms of DynASP2 and DynASP2.5. There we investigated various
18
graph problems including, but not limited to, variants of graph coloring, dominating set and vertex
cover. Note that, however, the cluster setup was slightly different.
C.2
Benchmark Environment
The experiments presented ran on an Ubuntu 16.04.1 LTS Linux cluster of 3 nodes with two Intel Xeon
E5-2650 CPUs of 12 physical cores each at 2.2 Ghz clock speed and 256 GB RAM. All solvers have been
compiled with gcc version 4.9.3 and executed in single core mode.
19
| 2cs.AI
|
Ref: International Conference on Artificial Neural Networks (ICANN), Springer LNCS,
Vol. 9887, pp. 170–178, Barcelona, Spain, September 2016.
DNN-Buddies: A Deep Neural Network-Based
Estimation Metric for the Jigsaw Puzzle Problem
Dror Sholomon1 , Eli (Omid) David1 , and Nathan S. Netanyahu1,2
1
arXiv:1711.08762v1 [cs.CV] 23 Nov 2017
2
Department of Computer Science, Bar-Ilan University, Ramat-Gan 52900, Israel
[email protected], [email protected], [email protected]
Center for Automation Research, University of Maryland, College Park, MD 20742
[email protected]
Abstract. This paper introduces the first deep neural network-based estimation metric for the jigsaw puzzle problem. Given two puzzle piece edges,
the neural network predicts whether or not they should be adjacent in the
correct assembly of the puzzle, using nothing but the pixels of each piece.
The proposed metric exhibits an extremely high precision even though no
manual feature extraction is performed. When incorporated into an existing puzzle solver, the solution’s accuracy increases significantly, achieving
thereby a new state-of-the-art standard.
(a)
(b)
Fig. 1: Jigsaw puzzle before and after reassembly using our DNN-Buddies scheme in
an enhanced solver.
1
Introduction
Jigsaw puzzles are a popular form of entertainment, available in different variation
of difficulty to challenge children, adults and even professional players. Given n × m
2
D. Sholomon, E.O. David, N.S. Netanyahu
different non-overlapping tiles of an image, the objective is to reconstruct the original image, taking advantage of both the shape and chromatic information of each
piece. Despite the popularity and vast distribution of jigsaw puzzles, their assembly
is not trivial computationally, as this problem was proven to be NP-hard [1] [8].
Nevertheless, a computational jigsaw solver may have applications in many realworld applications, such as biology [16], chemistry [25], literature [18], speech descrambling [27], archeology [2] [15], image editing [5], and the recovery of shredded
documents or photographs [3,7,14,17]. Regardless, as noted in [11], research of the
topic may be justified solely due to its intriguing nature.
Recent years have witnessed a vast improvement in the research and development
of automatic jigsaw puzzle solvers, manifested in both puzzle size, solution accuracy,
and amount of manual human intervention required. In its most basic form, every
puzzle solver requires some function to evaluate the compatibility of adjacent pieces
and a strategy for placing the pieces as accurately as possible. Most strategies are
greedy and rely heavily on some “trick” to estimate whether two pieces are truly
adjacent (e.g. two pieces that are each the most compatible piece from all pieces to
one another, four pieces that form a loop where each pair’s compatibility is above a
threshold, etc). Such heuristics were dubbed an “estimation metric” in [20], as they
allow estimating the adjacency correctness of two pieces without knowing the correct
solution. The majority of recent works focused on devising elaborate, hand-crafted
compatibility functions and high-precision estimation metrics.
Despite the proven effectiveness of neural networks in the field of computer vision,
no attempt has been made to automatically devise a high-precision estimation metric
for the jigsaw puzzle problem. This might be due to the highly imbalanced nature
of the puzzle problem, as in each n × m puzzle, there are O(n × m) matching
piece-pairs and O(n2 × m2 ) possible mismatching ones. In this paper we propose a
novel estimation metric relying on neural networks. The proposed metric achieves
extremely high precision despite the lack of any manually extracted features.
The proposed metric proves to be highly effective in real-world scenarios. We
incorporated the metric in our GA-based solver, using no hand-crafted sophisticated compatibility measure and experimented with the currently known challenging
benchmarks of the hardest variant of the jigsaw puzzle problem: non-overlapping,
(28 × 28) square pieces (i.e. only chromatic information is available to the solver)
where both piece orientation and puzzle dimensions are unknown. The enhanced
solver proposed sets a new state-of-the-art in terms of the accuracy of the solutions
obtained and the number of perfectly reconstructed puzzles.
2
Previous Work
Jigsaw puzzles were first introduced around 1760 by John Spilsbury, a Londonian
engraver and mapmaker. Nevertheless, the first attempt by the scientific community to computationally solve the problem is attributed to Freeman and Garder [9]
who in 1964 presented a solver which could handle up to nine-piece problems. Ever
since then, the research focus regarding the problem has shifted from shape-based
to merely color-based solvers of square-tile puzzles. In 2010 Cho et al. [4] presented
Deep Neural Network Based Estimation Metric for the Jigsaw Puzzle Problem
3
a probabilistic puzzle solver that could handle up to 432 pieces, given some a priori
knowledge of the puzzle. Their results were improved a year later by Yang et al. [26]
who presented a particle filter-based solver. Furthermore, Pomeranz et al. [20] introduced that year, for the first time, a fully automated square jigsaw puzzle solver
that could handle puzzles of up to 3,000 pieces. Gallagher [10] has further advanced
this by considering a more general variant of the problem, where neither piece orientation nor puzzle dimensions are known. Son et al. [24] improved the accuracy
of the latter variant using so-called “loop-constraints”. Palkin and Tal [19] further
improved the accuracy and handled puzzles with missing pieces. Sholomon et al. [21]
presented a genetic algorithm (GA)-based solver for puzzles of known orientation
which was later generalized to other variants [22,23].
2.1
Compatibility Measures and Estimation Metrics
As stated earlier, most works focus on the compatibility measure and an estimation
metric. A compatibility measure is a function that given two puzzle piece edges (e.g.
the right edge of piece 7 versus the upper edge of piece 12) predicts the likelihood
that these two edges are indeed placed as neighbors in the correct solution. This
measure applies to each possible pair of piece edges. The estimation metric, on the
other hand, predict whether two piece edges are adjacent but may not apply to
many possible pairs. Following is a more detailed review of the efforts made so far
in the field.
Cho et al. [4] surveyed four compatibility measures among which they found dissimilarity the most accurate. Dissimilarity is the sum (over all neighboring pixels)
of squared color differences (over all color bands). Assuming pieces xi , xj are represented in some three-dimensional color space (like RGB or YUV) by a K × K × 3
matrix, where K is the height/width of a piece (in pixels), their dissimilarity, where
xj is to the right of xi , for example, is
v
uK 3
uX X
(xi (k, K, cb) − xj (k, 1, cb))2 ,
(1)
D(xi , xj , r) = t
k=1 cb=1
where cb denotes the color band.
Pomeranz et al. [20] also used the dissimilarity measure but found empirically
that using the (Lp )q norm works better than the usual L2 norm. Moreover, they
presented the high-precision best-buddy metric. Pieces xi and xj are said to bestbuddies if
∀xk ∈ P ieces, C(xi , xj , R1 ) ≥ C(xi , xk , R1 )
and
(2)
∀xp ∈ P ieces, C(xj , xi , R2 ) ≥ C(xj , xp , R2 )
where P ieces is the set of all given image pieces and R1 and R2 are “complementary”
spatial relations (e.g. if R1 = right, then R2 = left and vice versa).
Gallagher [10] proposed yet another compatibility measure, called the Mahalanobis gradient compatibility (MGC) as a preferable compatibility measure to those
4
D. Sholomon, E.O. David, N.S. Netanyahu
used by Pomeranz et al. [20]. The MGC penalizes changes in intensity gradients,
rather than changes in intensity, and learns the covariance of the color channels, using the Mahalanobis distance. Also, Gallagher suggested using dissimilarity ratios.
Absolute distances between potential piece edge matches are sometimes not indicative (for example in smooth surfaces like sea and sky), so considering the absolute
score, divided by the second-best score available seems more indicative.
Son et al. [24] suggested “loop-constraints”, four or more puzzle piece edges where
the compatibility ratio between each pair is in the top ten among all possible pairs of
piece edges in the given puzzle. Palkin and Tal [19] proposed a greedy solver based
on an L1 -norm asymmetric dissimilarity and the best-buddies estimation metric.
3
3.1
DNN-Buddies
Motivation
We propose a novel estimation metric called “DNN-Buddies”. Our goal is to obtain
a classifier which predicts the adjacency likelihood of two puzzle piece edges in the
correct puzzle configuration.
Note that despite the exponential nature of the problem (as there are O((nm)!)
possible arrangements of the pieces, taking into account rotations), the problem can
be solved theoretically by assigning correctly, in a consecutive manner, n × m − 1
piece-edge pairs. (This is reminiscent of finding a minimal spanning tree, as noted
by [10].) Hence, the classifier’s precision is of far greater importance than its recall.
A classifier with perfect precision and a recall of
n×m−1
1
n×m−1
=
<
all possible matches
4 × (n × (m − 1) + (n − 1) × m)
8
(3)
might achieve a perfect solution by itself.
3.2
Challenges
A straight-forward solution might have been to train a neural network against
matching-pairs vs. non-matching ones. However, the issue of a jigsaw puzzle piece
matching is of an imbalanced nature. In each n × m puzzle, there are O(n × m)
matching pairs of piece edges and O(n2 × m2 ) possible nonmatching ones. A thorough review on the challenges and tactics to avoid them can be found in [13].
The trivial approach of random or uninformed undersampling, i.e. randomly
choosing the required number of nonmatching pairs leads to a low-precision and highrecall metric, the very opposite of the goal set beforehand. We believe that the reason
for this shortcoming is that there exist many “easy-to-spot” mismatches but only a
handful of “hard-to-spot” ones. Thus, we resort to informed undersampling, choosing
a subset of “good” mismatching pairs according to some criterion. Nevertheless, we
avoid using any manual feature selection or other sophisticated image-related means.
In the jigsaw puzzle domain, similarly to many other problem domains, the
solver does not actually try to reassemble the original image (as this problem is
Deep Neural Network Based Estimation Metric for the Jigsaw Puzzle Problem
5
not mathematically defined), but rather tries solving a “proxy problem” which is
to achieve an image whose global overall score between abutting-edges is minimal.
Thus, we choose using the compatibility measure as the undersampling criterion.
3.3
Neural Network Training
For training and cross-validation, we use the 2,755 images of size 360 × 480 pixels
from the IAPR TC-12 Benchmark [12]. Each image is first converted to YUV space
followed by the normalization of each channel separately (via z-score normalization).
Next, each (puzzle) image is divided to 12 × 17 tiles, where each tile is of size 28 × 28
pixels (as in all previous works); finally, we create a balanced set of positive and
negative samples of puzzle-piece pairs, using informed undersampling as will be
described below. In the end, we obtain a balanced set of 970,224 pairs overall.
To balance our dataset, we use the most basic compatibility score which is the
dissimilarity between two piece-edges in the YUV color-space, as described in Eq. 1,
as an undersampling criterion. For each puzzle piece edge xi,j (i = 1..n×m, j = 1..4),
we find its most compatible piece edge xk1,l1 and its second most compatible piece
edge xk2,l2 . If the pair of edges xi,j −xk1,l1 is indeed adjacent in the original image, we
add this pair to the pool of positively-labeled samples and toss the pair xi,j − xk2,l2
to the pool of negatively-labeled samples. Otherwise, xi,j − xk1,l1 is added to the
negatively-labeled samples and the other pair is discarded. The latter is done to avoid
training the network on adjacent pieces which happen to be vastly different due to a
significant change of the image scenery in the corresponding region. In other words,
we restrict our interest to highly compatible piece edges that are indeed adjacent.
Since this method leads to more negative samples than positive ones, we eventually
randomly throw some negative samples to balance out the set.
From each image pair we extract the two columns near the edge, i.e. the column
of abutting pixels in each edge and the one next to it. This results is an input of size
(28 × 4 × 3 =) 336 pixels. We use a feed-forward neural network (FFNN) of five fully
connected layers of size 336, 100, 100, 100, and 2. The output is a softmax layer
containing two neurons. We expect (0, 1) for matching pairs and (1, 0) otherwise.
The activation used in all layers is the rectified linear unit (ReLU) function, i.e.
f (x) = max(0, x). Figure 2 depicts the network’s structure.
We trained the network in a supervised manner using Stochastic Gradient Descent that minimizes the negative log likelihood of the error for 100 iterations. The
resulting network reaches 95.04% accuracy on the training set and 94.62% on a
held-out test set.
All dataset preparation and network training was performed using Torch7 [6].
4
Experimental Results
For each piece edge xi,j (i = 1..n × m, j = 1..4), if its most compatible piece edge xk,l
is classified positively using the DNN-Buddies network, we define xk,l to be xi,j ’s
DNN-buddy piece edge. Note that each piece edge can have only a single DNN-buddy;
6
D. Sholomon, E.O. David, N.S. Netanyahu
Fig. 2: Architecture of our DNN-Buddies scheme.
also, some pieces might not have a DNN-buddy at all (if the most compatible piece
is not classified as one by the DNN-Buddies network).
First, we evaluate the precision of the proposed metric, i.e. how many DNNbuddies are indeed adjacent in the original image. Using the well known dataset
presented by Cho et al. [4] of 20 432-piece puzzles, we obtained a precision of 94.83%.
Next, we incorporated the estimation metric (due to the proposed DNN-Buddies
scheme) into the GA-based solver proposed by us previously [23]. Unfortunately,
due to lack of space, no self-contained review of genetic algorithms and the proposed
method can be included in this paper. Nevertheless, the modification required with
respect to the existing GA framework is rather simple; if a DNN-buddy pair appears
in one of the parents, assign this pair in the child. Figure 3 describes the modified
crossover operator in the GA framework according to the above (see Step 2, which
includes the new DNN-buddy phase).
Until
1.
2.
3.
4.
5.
(n − 1) relative relations are assigned do
Try assigning all common relative relations in the parents.
Try assigning all DNN-buddy relative relations in the parents.
Try assigning all best-buddy relative relations in the parents.
Try assigning all existing most-compatible relative relations.
Try assigning random relative relations.
Fig. 3: Crossover overview
We ran the augmented solver on the 432-piece puzzle set and on the two additional datasets proposed by Pomeranz et al. [20] of 540- and 805- piece puzzles.
We evaluated our results according to the neighbor comparison which measures the
fraction of correct neighbors and the number of puzzles perfectly reconstructed for
each set.
Deep Neural Network Based Estimation Metric for the Jigsaw Puzzle Problem
7
Table 1 presents the accuracy results of the same solver with and without the
DNN-Buddies metric. For each dataset we achieve a considerable improvement in
the overall accuracy of the solution, as well as the number of perfectly reconstructed
puzzles. Moreover, our enhanced deep neural network-based scheme appears to outperform the current state-of-the-art results, as it yields accuracy levels of 95.65%,
96.37% and 95.86%, which surpass, respectively, the best results known of 95.4% [19],
94.08% and 94.12% [23].
# of Pieces
432
540
805
GA
Neighbor
Perfect
94.88%
94.08%
94.12%
11
8
6
Our (GA + DNN-Buddies)
Neighbor
perfect
95.65%
96.37%
95.86%
12
11
8
Table 1: Comparison of our accuracy results with and without the new DNN-Buddies
estimation metric.
5
Conclusions
In this paper we presented the first neural network-based estimation metric for the
jigsaw puzzle problem. Unlike previous methods, no manual feature crafting was
employed. The novel method exhibits high precision and when combined with a
real-world puzzle solver, it significantly improves the solution’s accuracy to set a
new state-of-the art standard.
References
1. Altman, T.: Solving the jigsaw puzzle problem in linear time. Applied Artificial Intelligence an International Journal 3(4), 453–462 (1989)
2. Brown, B., Toler-Franklin, C., Nehab, D., Burns, M., Dobkin, D., Vlachopoulos, A.,
Doumas, C., Rusinkiewicz, S., Weyrich, T.: A system for high-volume acquisition and
matching of fresco fragments: Reassembling Theran wall paintings. ACM Transactions
on Graphics 27(3), 84 (2008)
3. Cao, S., Liu, H., Yan, S.: Automated assembly of shredded pieces from multiple photos.
In: IEEE International Conference on Multimedia and Expo. pp. 358–363 (2010)
4. Cho, T., Avidan, S., Freeman, W.: A probabilistic image jigsaw puzzle solver. In: IEEE
Conference on Computer Vision and Pattern Recognition. pp. 183–190 (2010)
5. Cho, T., Butman, M., Avidan, S., Freeman, W.: The patch transform and its applications to image editing. In: IEEE Conference on Computer Vision and Pattern
Recognition. pp. 1–8 (2008)
6. Collobert, R., Kavukcuoglu, K., Farabet, C.: Torch7: A matlab-like environment for
machine learning. In: BigLearn, NIPS Workshop. No. EPFL-CONF-192376 (2011)
7. Deever, A., Gallagher, A.: Semi-automatic assembly of real cross-cut shredded documents. In: ICIP. pp. 233–236 (2012)
8
D. Sholomon, E.O. David, N.S. Netanyahu
8. Demaine, E., Demaine, M.: Jigsaw puzzles, edge matching, and polyomino packing:
Connections and complexity. Graphs and Combinatorics 23, 195–208 (2007)
9. Freeman, H., Garder, L.: Apictorial jigsaw puzzles: The computer solution of a problem
in pattern recognition. IEEE Transactions on Electronic Computers EC-13(2), 118–127
(1964)
10. Gallagher, A.: Jigsaw puzzles with pieces of unknown orientation. In: IEEE Conference
on Computer Vision and Pattern Recognition. pp. 382–389 (2012)
11. Goldberg, D., Malon, C., Bern, M.: A global approach to automatic solution of jigsaw
puzzles. Computational Geometry: Theory and Applications 28(2-3), 165–174 (2004)
12. Grubinger, M., Clough, P., Müller, H., Deselaers, T.: The IAPR TC-12 benchmark:
A new evaluation resource for visual information systems. In: International Workshop
OntoImage. vol. 5, p. 10 (2006)
13. He, H., Garcia, E.A.: Learning from imbalanced data. Knowledge and Data Engineering, IEEE Transactions on 21(9), 1263–1284 (2009)
14. Justino, E., Oliveira, L., Freitas, C.: Reconstructing shredded documents through feature matching. Forensic science international 160(2), 140–147 (2006)
15. Koller, D., Levoy, M.: Computer-aided reconstruction and new matches in the forma
urbis romae. Bullettino Della Commissione Archeologica Comunale di Roma pp. 103–
125 (2006)
16. Marande, W., Burger, G.: Mitochondrial DNA as a genomic jigsaw puzzle. Science
318(5849), 415–415 (2007)
17. Marques, M., Freitas, C.: Reconstructing strip-shredded documents using color as feature matching. In: ACM symposium on Applied Computing. pp. 893–894 (2009)
18. Morton, A.Q., Levison, M.: The computer in literary studies. In: IFIP Congress. pp.
1072–1081 (1968)
19. Paikin, G., Tal, A.: Solving multiple square jigsaw puzzles with missing pieces. In:
Computer Vision and Pattern Recognition, 2015 IEEE Conference on. pp. 4832–4839.
IEEE (2015)
20. Pomeranz, D., Shemesh, M., Ben-Shahar, O.: A fully automated greedy square jigsaw
puzzle solver. In: IEEE Conference on Computer Vision and Pattern Recognition. pp.
9–16 (2011)
21. Sholomon, D., David, O.E., Netanyahu, N.S.: A genetic algorithm-based solver for very
large jigsaw puzzles. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 1767–1774 (2013)
22. Sholomon, D., David, O.E., Netanyahu, N.S.: A generalized genetic algorithm-based
solver for very large jigsaw puzzles of complex types. In: AAAI Conference on Artificial
Intelligence. pp. 2839–2845 (2014)
23. Sholomon, D., David, O.E., Netanyahu, N.S.: Genetic algorithm-based solver for very
large multiple jigsaw puzzles of unknown dimensions and piece orientation. In: ACM
Conference on Genetic and Evolutionary Computation. pp. 1191–1198 (2014)
24. Son, K., Hays, J., Cooper, D.B.: Solving square jigsaw puzzles with loop constraints.
In: European Conference on Computer Vision 2014, pp. 32–46. Springer (2014)
25. Wang, C.S.E.: Determining molecular conformation from distance or density data.
Ph.D. Thesis, Massachusetts Institute of Technology (2000)
26. Yang, X., Adluru, N., Latecki, L.J.: Particle filter with state permutations for solving
image jigsaw puzzles. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 2873–2880 (2011)
27. Zhao, Y., Su, M., Chou, Z., Lee, J.: A puzzle solver and its application in speech
descrambling. In: WSEAS International Conference Computer Engineering and Applications. pp. 171–176 (2007)
| 9cs.NE
|
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH
EMBEDDING DIMENSION THREE
arXiv:1701.00988v1 [math.AC] 4 Jan 2017
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
Abstract. This work extends the results known for the Delta sets of non-symmetric numerical
semigroups with embedding dimension three to the symmetric case. Thus, we have a fast algorithm
to compute the Delta set of any embedding dimension three numerical semigroup. Also, as a
consequence of these resutls, the sets that can be realized as Delta sets of numerical semigroups of
embedding dimension three are fully characterized.
1. Introduction
Delta sets (or sets of distances) were first introduced in [13] as a tool to study factorizations
in non-unique factorization domains. Since then several authors have studied their properties. In
particular, for numerical semigroups, in [3], the firsts results for some special cases of embedding
dimension three numerical semigroups are presented. Delta sets for numerical semigroups are eventually periodic as explained in [5], and thus, if a bound for this periodicity is known, the whole
Delta set of a numerical semigroup can be computed. The bounds given in [5] were improved in
[9]. In [2] a dynamical procedure to compute Delta sets for numerical semigroups is presented, that
makes use of the bound given in [9]; this procedure has been implemented in numericalsgps [7].
Recently, a new procedure to compute Delta sets of any affine semigroup using Göbner basis has
been presented in [11]. Needless to say, all these algorithms and bounds were the consequence of a
better (theoretical) understanding of Delta sets.
In [6] it is shown that the maximum of the Delta set of a numerical semigroup is attained in a Betti
element of the semigroup (indeed, it is shown that this holds for a wider class of atomic monoids).
This does not provide a way to compute the whole set. The minimum was known to be the greatest
common divisor of the Delta set since [13]; however, the elements in the interval determined by this
minimum and maximum element of the Delta set are not known in general. Some realization results
were given in [4], while in [10] the sets that can be realized as the Delta set of a non-symmetric
numerical semigroup with embedding dimension three are completely characterized. In that paper,
the authors present a procedure that strongly reduces the time needed to compute the Delta sets
of non-symmetric embedding dimension three numerical semigroups. In this manuscript, we extend
this algorithm to the symmetric case.
Although the main results yield the same algorithm for both symmetric and non-symmetric
numerical semigroups with embedding dimension three, there are significant differences in the intermediate results and their proofs. These differences stem from the different structure of the Betti
elements of these semigroups (see Proposition 3).
The factorizations of the Betti elements of non-symmetric embedding dimension three numerical
semigroups have been thoroughly studied, and thus provided an easier starting point. However, in
that case, some technical and tedious detours were needed to obtain our result.
Due to the fact that, in the symmetric case, we do not have a unique minimal presentation (in
contraposition to what happens in the non-symmetric case), the first step is to choose the right
2010 Mathematics Subject Classification. 05A17,20M13,20M14.
Key words and phrases. Numerical Semigroups, non-unique factorization, Delta set, Euclid’s Algorithm.
The first and second author are supported by the project FQM-343, MTM2014-55367-P, and FEDER funds.
1
2
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
factorization of the Betti elements to start our algorithm. However, in this case we can work
directly on the Euclid’s set (see Definition 10), whereas this cannot be done in the non-symmetric
case without building a certain theoretical framework. In this way, we can better see the idea behind
the main result, and the technical results are no longer necessary. Moreover, the proofs and the
reading of the paper are more comfortable, and the number of pages is reduced. Unfortunately,
from our reasoning, it emerged that there are no relations between these two cases; we think that
it is not possible to find a common way to deduce both settings together.
This paper, even though is complementary to [10], is self-contained, and therefore can be read
separately from [10]. After some preliminary results, we explain how we can choose the elements
δ1 and δ2 needed to start the algorithm. As we mentioned above, this was not needed in the nonsymmetric case, since in that setting the Delta sets of the Betti elements of the numerical semigroup
are singletons. Then, we prove our main result, which yields an algorithm that works in the same
way as in the non-symmetric case; in particular, notice that Example 20 provided in this paper
gives the same Delta set as in [10, Example 38], because it is obtained from the same δ1 and δ2 .
2. Preliminaries
Let N be the set of non negative integers. Take n1 , n2 , n3 ∈ N with gcd(n1 , n2 , n3 ) = 1, and
define S as the numerical semigroup minimally generated by {n1 , n2 , n3 }, that is,
S = {a1 n1 + a2 n2 + a3 n3 | a1 , a2 , a3 ∈ N}.
The condition gcd(n1 , n2 , n3 ) = 1 is equivalent to say that the set of gaps, G(S) = N \ S, has finitely
many elements. The maximum of Z \ S is called the Frobenius number, and we denote it as F(S).
Definition 1. A numerical semigroup S is symmetric if x ∈ Z \ S implies F(S) − x ∈ S.
The reader interested in numerical semigroups may have a look at [14] and for some applications
to [1].
The set of factorizations of s ∈ S is
Z(s) = {(z1 , z2 , z3 ) ∈ N3 | z1 n1 + z2 n2 + z3 n3 = s}.
We denote the length of a factorization z = (z1 , z2 , z3 ) ∈ Z(s) as
ℓ(z) = z1 + z2 + z3 .
The set of lengths of s ∈ S is
L(s) = {ℓ(z) | z ∈ Z(s)}.
It is easy to see that L(s) ⊂ [0, s], and consequently L(s) is finite. So, it is of the form L(s) =
{l1 , . . . , lk } for some positive integers l1 < l2 < · · · < lk . The set
∆(s) = {li − li−1 | 2 ≤ i ≤ k}
is known as the Delta set of s ∈ S (or set of distances), and the Delta set of S is defined as
∆(S) = ∪s∈S ∆(s).
Now, we define
MS = {v = (v1 , v2 , v3 ) ∈ Z3 | v1 n1 + v2 n2 + v3 n3 = 0}.
We can extend the ℓ function to elements in MS : for v = (v1 , v2 , v3 ) ∈ Z3 , set
ℓ(v) = v1 + v2 + v3 .
Note that ℓ is a linear map.
The aim of this paper is to compute the Delta set of S by using Euclid’s algorithm with two
special elements in this set. These special distances are associated to particular elements of S, called
the Betti elements. The definition of Betti element relies in the construction of a graph associated
to the elements in the semigroup. Let s ∈ S. The graph ∇s is the graph with vertices Z(s), the
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
3
set of factorization of s, and zz′ is an edge if if z · z′ 6= 0, that is, there exists a common nonzero
coordinate in both factorizations.
Definition 2. An element s ∈ S is called a Betti element if and only if ∇s is not connected. The
set of Betti elements will be denoted by Betti(S).
A numerical semigroup with embedding dimension three might have up to three Betti elements.
In [14, Example 8.23] it is shown that Betti(hn1 , n2 , n3 i = {c1 n1 , c2 n2 , c3 n3 }, where ci is the least
positive integer such that ci ni ∈ hnj , nk i, {i, j, k} = {1, 2, 3}.
Proposition 3. [14, Chapter 9] Let S an embedding dimension three numerical then 1 ≤ ♯Betti(S) ≤
3. Moreover, S is non-symmetric if and only if ♯Betti(S) = 3.
Example 4. Consider S = h3, 5, 7i = {0, 3, 5, →} (the arrow means that all integers greater than 5
are in the semigroup). In this case, Betti(S) = {10, 12, 14}, and
Z(10) = {(1, 0, 1), (0, 2, 0)}, L(10) = {2},
∆(10) = ∅,
Z(12) = {(4, 0, 0), (0, 1, 1)}, L(12) = {2, 4}, ∆(12) = {2},
Z(14) = {(3, 1, 0), (0, 0, 2)}, L(14) = {2, 4}, ∆(14) = {2}.
Let S = h6, 8, 11i, with Betti elements 22 and 24.
Z(22) = {(1, 2, 0), (0, 0, 2)}, L(22) = {2, 3}, ∆(22) = {1},
Z(24) = {(4, 0, 0), (0, 3, 0)}, L(24) = {3, 4}, ∆(24) = {1}.
Finally, let S = h6, 10, 15i. Then Betti(S) = {30},
Z(30) = {(5, 0, 0), (0, 3, 0), (0, 0, 2)}, L(30) = {2, 3, 5}, ∆(30) = {1, 2}.
More details on the relation between Betti elements and Delta sets can be found in [6].
It is straightforward to see that from the factorizations of every b ∈ Betti(S), we can construct
{z − z′ ∈ MS | z 6= z′ , z, z′ ∈ Z(b), b ∈ Betti(S)}, which generates MS as a group.
For instance, for S = h6, 8, 11i, (1, 2, −2) and (4, −3, 0) generate as a group the set of integer
solutions of the equation 6x1 + 8x2 + 11x3 = 0.
Since the structure of the Delta set of a non-symmetric numerical semigroup is known, we focus
on the study of the symmetric numerical semigroups.
3. Two cases in the symmetric setting
Let S be a numerical semigroup minimally generated by {n1 , n2 , n3 }, and assume that S is
symmetric. In this Section, we define the integers and vectors that allow us to study MS and
construct ∆(S).
By Proposition 3, #Betti(S) ≤ 2. Now, we going to choose a suitable basis {v1 , v2 } for MS
according to #Betti(S).
(1) A single Betti element.
If #Betti(S) = 1, then (n1 , n2 , n3 ) = (s2 s3 , s1 s3 , s1 s2 ) for some positive pairwise coprime integers s1 > s2 > s3 (see [12]). Moreover, Betti(S) = {s1 s2 s3 }, and the set of factorizations of
s1 s2 s3 is Z(s1 s2 s3 ) = {(s1 , 0, 0), (0, s2 , 0), (0, 0, s3 )}. In this setting, it is easy to see that MS is
the group spanned by {(s1 , −s2 , 0), (0, s2 , −s3 )}. We set v1 = (s1 , −s2 , 0), v2 = (0, s2 , −s3 ).
(2) Two Betti elements.
If #Betti(S) = 2, we have (n1 , n2 , n3 ) = (am1 , am2 , bm1 + cm2 ) with a ≥ 2 and b + c ≥ 2, and
we can also assume that m1 < m2 [14, Theorem 10.6]. In this setting
Betti(S) = {a(bm1 + cm2 ), am1 m2 },
4
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
with
Z(a(bm1 + cm2 )) =
j k
n
j k
b − mb2 m2 , c + mb2 m1 , 0 , . . . , (b − m2 , c + m1 , 0), (b, c, 0),
j k
o
j k
(b + m2 , c − m1 , 0) . . . , b + mc1 m2 , c − mc1 m1 , 0 , (0, 0, a)
and Z(am1 m2 ) = {(m2 , 0, 0), (0, m1 , 0)}. Moreover, MS is spanned by {(m2 , −m1 , 0), (b +
λm2 , c − λm1 , −a)} for any λ ∈ Z. We choose λ ∈ {− ⌊b/m2 ⌋ , . . . , ⌊c/m1 ⌋} such that ℓ((b +
λm2 , c − λm1 , −a)) is minimal. We define v1 = (m2 , −m1 , 0) and v2 = (b + λm2 , c − λm1 , −a).
Now, we define δ1 = ℓ(v1 ) and δ2 = |ℓ(v2 )|. We consider the absolute value for ℓ(v2 ), since it
might happen that ℓ(v2 ) < 0 when there are two Betti elements and a > b + c + λ(m2 − m1 ). In
order to keep trace of the sign of ℓ(v2 ), let sgn be the sign function, and set σ = sgn(ℓ(v2 )). In this
way, we have δ2 = σℓ(v2 ).
The integers δ1 , δ2 just defined are tightly related to ∆(S).
Proposition 5. Let S be a symmetric numerical semigroup, and let δ1 and δ2 be defined as above.
Then
max ∆(S) = max{δ1 , δ2 }.
Proof. We know from [6] that max ∆(S) = max{max ∆(b) | b ∈ Betti(S)}.
If #Betti(S) = 1, since Z(s1 s2 s3 ) = {(s1 , 0, 0), (0, s2 , 0), (0, 0, s3 )}, we get ∆(s1 s2 s3 ) = {s1 −
s2 , s2 − s3 } = {δ1 , δ2 }, and the thesis follows.
If #Betti(S) = 2, we need to study Z(b) with b ∈ Betti(S) = {a(bm1 + cm2 ), am1 m2 }.
• Z(am1 m2 ) = {(m2 , 0, 0), (0, m1 , 0)}. Hence ∆(am1 m2 ) = {δ1 }.
• Z(a(bm1 + cm2 )) = {(0, 0, a)} ∪ {(b + km2 , c − km1 , 0) | k ∈ {− ⌊b/m2 ⌋ , . . . , ⌊c/m1 ⌋}}.
Observe that L(a(bm1 + cm2 )) = {a} ∪ {b + c + k(m2 − m1 ) | k ∈ {− ⌊b/m2 ⌋ , . . . , ⌊c/m1 ⌋}}. Consider the following cases.
(1) If a ≤ b + c − ⌊b/m2 ⌋(m2 − m1 ), then λ = −⌊b/m2 ⌋ and δ2 = b + c − ⌊b/m2 ⌋(m2 − m1 ) − a.
Hence ∆(a(bm1 + cm2 )) equals {δ1 , δ2 }, or {δ1 } if a = b + c − ⌊b/m2 ⌋(m2 − m1 ) (that is, δ2 = 0).
(2) If a ≥ b + c + ⌊c/m1 ⌋(m2 − m1 ), then λ = ⌊c/m1 ⌋, and we argue as in the previous case.
(3) Finally, if b + c − ⌊b/m2 ⌋(m2 − m1 ) < a < b + c + ⌊c/m1 ⌋(m2 − m1 ), there exists k ∈
{− ⌊b/m2 ⌋ , . . . , ⌊c/m1 ⌋ − 1} such that b + c + k(m2 − m1 ) ≤ a < b + c + (k + 1)(m2 − m1 ).
Then λ is either k or k + 1, and ∆(a(bm1 + cm2 )) = {δ1 , δ2 , |δ2 − δ1 |} (unless δ2 = 0 and
∆(a(bm1 + cm2 )) = {δ1 }).
In any case, max ∆(S) = max{δ1 , δ2 }.
The following result is a particular instance of a more general property.
Proposition 6. [6, Corollay 3.1] Let S be a symmetric numerical semigroup. Then
min ∆(S) = gcd{δ1 , δ2 }.
Thus, max ∆(S) is either δ1 or δ2 , while in our setting min ∆(S) = gcd(δ1 , δ2 ), and each element
of ∆(S) is a multiple of this greatest common divisor [13].
Remark 7. Observe that, in both cases, the vector v1 has the first coordinate positive, the second
one negative, and the third coordinate equal to zero. Our choice of λ in the case #Betti(S) = 2
ensures that the vector v2 has first coordinate nonnegative, positive the second, and the third
coordinate negative. We will represent this fact as:
v1 = (+, −, 0) and v2 = (+, +, −).
Proposition 8. Let S be a symmetric numerical semigroup with embedding dimension three. Let
δ1 , δ2 , v1 , v2 and σ be defined as above.
(1) if σ = 1, we have that δ2 v1 − σδ1 v2 = (?, −, +).
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
5
(2) if σ = −1, we have that δ2 v1 − σδ1 v2 = (+, ?, −).
The symbol “ ?” denotes the sign of this coordinate is not determined.
Proof. If σ = 1 and #Betti(S) = 1, we know that v1 = (s1 , −s2 , 0) and v2 = (0, s2 , −s3 ). Then
δ2 v1 − δ1 v2 = (δ2 s1 , −δ2 s2 − δ1 s2 , δ1 s3 ) = (δ2 s1 , −s2 (δ2 + δ1 ), δ1 s3 ) = (+, −, +).
For σ = 1 and #Betti(S) = 2, we have v1 = (m2 , −m1 , 0) and v2 = (b + λm2 , c − λm1 , −a), whence
δ2 v1 − δ1 v2 = (δ2 m2 − δ1 (b + λm2 ), −δ2 m1 − δ1 (c − λm1 ), δ1 a) = (?, −, +).
Finally if σ = −1 and #Betti(S) = 2,
δ2 v1 + δ1 v2 = (δ2 m2 + δ1 (b + λm2 ), −δ2 m1 + δ1 (c − λm1 ), −δ1 a) = (+, ?, −).
4. Euclid’s set and the Delta set
Associated to δ1 and δ2 we are going to define its Euclid’s set as the set of all integers that appear
in the naive implementation of the greatest common divisor algorithm. We will see that precisely
this set is the Delta set of the semigroup except to the zero element.
Proposition 9. Let δ1 , δ2 be two positive integers, and let x ∈ {1, . . . , max{δ1 , δ2 }} ∩ gZ, where
g = gcd(δ1 , δ2 ). Then there exist unique (x1 , x2 ) and (x′1 , x′2 ) in Z2 such that
(1) x = x1 δ1 + x2 δ2 with −δ1 /g < x2 ≤ 0 < x1 ≤ δ2 /g,
(2) x = x′1 δ1 + x′2 δ2 with −δ2 /g < x′1 ≤ 0 < x′2 ≤ δ1 /g.
Moreover (x1 , x2 ) = (x′1 + δ2 /g, x′2 − δ1 /g).
We will denote
• x(δ1 ,δ2 ) = (x1 , x2 ), where 0 < x1 ≤ δ2 /g, −δ1 /g < x2 ≤ 0, and x = x1 δ1 + x2 δ2 ,
• x′(δ1 ,δ2 ) = (x′1 , x′2 ), where −δ2 /g < x′1 ≤ 0, 0 < x2 ≤ δ1 /g, and x = x′1 δ1 + x′2 δ2 .
(δ ,δ )
′(δ ,δ )
′(δ ,δ )
(δ ,δ )
In particular, δ1 1 2 = (1, 0), δ1 1 2 = (1 − δ2 /g, δ1 /g), δ2 1 2 = (0, 1) and δ2 1 2 = (δ2 /g, 1 −
δ1 /g).
We want to depict the set of elements obtained after applying Euclid’s greatest common divisor
algorithm to δ1 and δ2 . We will use the naive approach that uses substraction instead of remainders
of divisions. Our set will be decomposed in subsets corresponding to the algorithm using remainders.
Definition 10. Let δ1 and δ2 be positive integers, and define jη1 = kmax{δ1 , δ2 }, η2 = min{δ1 , δ2 }
ηj
ηj+1 = ηj mod ηj+1 . Let i be
and η3 = η1 mod η2 . In general, for j > 0, define ηj+2 = ηj − ηj+1
the maximum index such that ηi+1 > 0. Define
D(η1 , η2 ) = {η1 , η1 − η2 , . . . , η1 mod η2 = η3 },
D(η2 , η3 ) = {η2 , η2 − η3 , . . . , η2 mod η3 = η4 },
D(η3 , η4 ) = {η3 − η4 , . . . , η3 mod η4 = η5 },
···
D(ηj−1 , ηj ) = {ηj−1 − ηj , . . . , ηj−1 mod ηj = ηj+1 },
···
D(ηi , ηi+1 ) = {ηi − ηi+1 , . . . , ηi mod ηi+1 = ηi+2 = 0}.
S
Let us denote by Euc(δ1 , δ2 ) = i∈I D(ηi , ηi+1 ), where I = {i ∈ N | ηi+1 > 0}.
Observe that η1 ∈ D(η1 , η2 ) and η2 ∈ D(η2 , η3 ), but this is no longer the case for i > 2. This is
because we want the union in the definition of Euc(δ1 , δ2 ) to be disjoint. Also, ηj ∈ D(ηj−2 , ηj−1 )
as ηj = ηj−2 − ⌊ηj−2 /ηj−1 ⌋ηj−1 for any j > 2.
Notice also that Euc(δ1 , δ2 ) ⊂ gZ where g = gcd(δ1 , δ2 ).
Example 11. Let S = hs2 s3 , s1 s3 , s1 s2 i, with s1 = 548, s2 = 155, and s3 = 13. Then v1 =
(548, −155, 0) and v2 = (0, 155, −13). Hence δ1 = 393 and δ2 = 142. In this setting,
6
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
• D(393, 142) = {393, 251, 109},
• D(142, 109) = {142, 33},
• D(109, 33) = {76, 43, 10},
• D(33, 10) = {23, 13, 3},
• D(10, 3) = {7, 4, 1},
• D(3, 1) = {2, 1, 0}.
Then Euc(δ1 , δ2 ) = {0, 1, 2, 3, 4, 7, 10, 13, 23, 33, 43, 76, 109, 142, 251, 393}.
(η ,η )
Remark 12. It is clear that η1 > η2 > · · · > ηi > ηi+1 > ηi+2 = 0, and for ηi 1 2 = (ηi1 , ηi2 ) and
′(η ,η )
′
′
′
′
ηi+11 2 = (ηi+1
1 , ηi+1 2 ), the inequalities ηi+1 1 ηi1 ≤ 0 and ηi+1 2 ηi2 ≤ 0 hold. From these, we have
′
′
too that, for q > 0, |ηi1 | ≤ |ηi1 − qηi+1 1 | and |ηi2 | ≤ |ηi2 − qηi+1
2 | hold. Analogously, considering
′(η ,η )
(η ,η )
′ , η ′ ) and η 1 2 = (η
′
′
′
′
ηi 1 2 = (ηi1
i+1 1 , ηi+1 2 ), we have |ηi1 | ≤ |ηi1 − qηi+1 1 |, |ηi2 | ≤ |ηi2 − qηi+1 2 |
i2
i+1
for q > 0. Observe that we have an equality only when one of the ηhk is zero.
′(η ,η )
(η ,η )
(η ,η )
As η1 1 2 = (1, 0) and η2 1 2 = (0, 1) = (−, +), we obtain η3 1 2 = (1, − ⌊η1 /η2 ⌋) = (+, −)
′(η ,η )
and η4 1 2 = (− ⌊η2 /η3 ⌋, 1 + ⌊η2 /η3 ⌋ ⌊η1 /η2 ⌋) = (−, +), and so on.
This means that in each step in Euc(δ1 , δ2 ), the absolute value of the (η1 , η2 )-coordinates increases,
(and is not decreasing only when xhk = x′21 ). We will use this fact in the next propositions.
Proposition 13. Let the notation and hypotheses be as in Definition 10, and let d ∈ {η3 , η3 +
g, . . . , η1 }. Then d ∈ D(η1 , η2 ) if and only if d1 = 1, with (d1 , d2 ) = d(η1 ,η2 ) (according to the
notation given in Proposition 9).
Proof. Let d ∈ {η3 , η3 + g, . . . , η1 }. If d ∈ D(η1 , η2 ), then d = η1 − kη2 for some k ∈ {0, . . . , ⌊η1 /η2 ⌋}.
Hence d(η1 ,η2 ) = (1, −k). For the converse, if 0 < d = η1 + kη2 ≤ η1 for some k ∈ Z, we obtain
−η1 < kη2 ≤ 0, whence −η1 /η2 < k ≤ 0. Now, as g = gcd(δ1 , δ2 ) ≤ η2 we have −η1 /g ≤ −η1 /η2 < k
and by Proposition 9, k will be equal to d2 , and consequently d ∈ D(η1 , η2 ).
Proposition 14. Let the notation and hypotheses be as in Definition 10.
(1) Let x ∈ {η3 , η3 + g, . . . , η1 } \ (D(η1 , η2 ) ∪ D(η2 , η3 )).
(a) If x(η1 ,η2 ) = (x1 , x2 ) = (+, −), then there exists an integer d < x, d ∈ D(η1 , η2 ), such that
d(η1 ,η2 ) = (d1 , d2 ) with 0 < d1 < x1 and x2 < d2 < 0.
(b) If x′(η1 ,η2 ) = (x′1 , x′2 ) = (−, +), then there exists an integer d < x, d ∈ D(η2 , η3 ), such that
d′(η1 ,η2 ) = (d′1 , d′2 ) with x′1 ≤ d′1 ≤ 0 and 0 < d′2 < x′2 .
The case x′1 = d′1 = 0 corresponds to x = d2 η2 , that is, only when x is a multiple of η2 .
(2) In general, for any x multiple of g such that x < η3 , x ∈
/ Euc(δ1 , δ2 ), consider j + 1 = min{k ∈
I | ηk < x} and then x ∈ {ηj+1 , ηj+1 + g, . . . , ηj } \ (D(ηj−1 , ηj ) ∪ D(ηj , ηj+1 )) = {ηj+1 , ηj+1 +
g, . . . , ηj } \ Euc(δ1 , δ2 ).
(a) If x(η1 ,η2 ) = (x1 , x2 ) = (+, −), then there exists an integer d < x, d ∈ D(ηj−1 , ηj ) ∪
D(ηj , ηj+1 ), such that d(η1 ,η2 ) = (d1 , d2 ) with 0 < d1 < x1 and x2 < d2 < 0.
(b) If x′(η1 ,η2 ) = (x′1 , x′2 ) = (−, +), then there exists an integer d < x, d ∈ D(ηj−1 , ηj ) ∪
D(ηj , ηj+1 ), such that d′(η1 ,η2 ) = (d′1 , d′2 ) with x′1 < d′1 < 0 and 0 < d′2 < x′2 .
Proof. (1) We study each case independently.
(a) As η1 − ⌊ ηη21 ⌋η2 = η3 < x < η1 and x 6∈ D(η1 , η2 ) ∪ D(η2 , η3 ), there exists an integer k
such that η1 + kη2 < x < η1 + (k + 1)η2 and −⌊ ηη12 ⌋ < k < 0. Set d = η1 + kη2 . Then
d ∈ D(η1 , η2 ). Also, η2 ≥ g and δ1 ≤ η1 , thus − δg1 ≤ − ηη21 ≤ ⌈− ηη21 ⌉ = −⌊ ηη21 ⌋ < k. This
implies that d(η1 ,η2 ) = (1, k). It is clear that x1 ≥ 2 by Proposition 13, and as d1 = 1
we have that 0 < d1 < x1 . Also, as x = x1 η1 + x2 η2 and x1 ≥ 2, and we obtain that
2η1 +x2 η2 ≤ x < η1 +(k+1)η2 . Consequently, x2 η2 < η1 +x2 η2 < (k+1)η2 . Thus, x2 < k+1.
Observe that x2 6= k, since otherwise x ≥ 2η1 + kη2 > η1 + η2 + kη2 = η1 + (k + 1)η2 , a
contradiction. Hence x2 < k = d2 < 0.
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
7
(b) For (x′1 , x′2 ) we distinguish two cases.
(i) If x > η2 , it suffices to take d = η2 ∈ D(η2 , η3 ). We have that d′1 = 0 and d′2 = 1. In
light of Proposition 9, it follows that x′1 ≤ d′1 = 0 and 0 < d′2 < x′2 . It is clear that
x′2 6= 1, since otherwise we would have x = x′1 η1 + x′2 η2 = x′1 η1 + η2 ≤ d′1 η1 + η2 = η2 ,
that is, x ≤ η2 , a contradiction.
(ii) If η3 < x < η2 , consider D(η2 , η3 ) = {η2 , η2 − η3 , η2 − 2η3 , . . . , η4 }. As x ∈
/ D(η2 , η3 ),
by denoting x(η2 ,η3 ) = (x2 , x3 ) = (+, −) we can, as above, choose and integer h such
that η2 + hη3 < x < η2 + (h + 1)η3 . Then, taking d = η2 + hη3 and arguing as in (a)
with d = η2 + hη3 and x = x2 η2 + x3 η3 , we have that 0 < 1 < x2 and x3 < h < 0.
Now, as η3 = η1 mod η2 = η1 − ⌊η1 /η2 ⌋ η2 , we can rewrite
x = x2 η2 + x3 (η1 − ⌊η1 /η2 ⌋ η2 ) = x3 η1 + (x2 − x3 ⌊η1 /η2 ⌋)η2 ,
d = η2 + h(η1 − ⌊η1 /η2 ⌋ η2 ) = hη1 + (1 − h ⌊η1 /η2 ⌋)η2 .
Thus, (x′1 , x′2 )(η1 ,η2 ) = (x3 , x2 − x3 ⌊η1 /η2 ⌋) and (d′1 , d′2 )(η1 ,η2 ) = (h, 1 − h ⌊η1 /η2 ⌋).
Moreover, it is clear that x′1 = x3 < h = d′1 < 0. We also have that −x3 ⌊η1 /η2 ⌋ >
−h ⌊η1 /η2 ⌋ > 0 and x2 > 1 > 0. Hence x′2 = x2 − x3 ⌊η1 /η2 ⌋ > 1 − h ⌊η1 /η2 ⌋ = d′2 > 0.
(2) For the general case, we will follow the arguments of the preceding cases. It is clear that for all
suitable j, gcd(ηj , ηj+1 ) = gcd(δ1 , δ2 ) = g. Hence we have the following.
• Observe that j ∈ I implies ηj+1 > 0, so an element d ∈ {ηj+1 , ηj+1 + g, . . . , ηj } is in
(j−1)
(j−1) (j−1)
D(ηj−1 , ηj ) \ {ηj } if and only if d1
= 1, where d(ηj−1 ,ηj ) = (d1
, d2
). This is the
same argument used in Proposition 13, applied for η1 = ηj−1 and for η2 = ηj .
• Working at this (j−1, j) level, it is also clear that, if x ∈ {ηj+1 , ηj+1 +g, . . . , ηj }\(D(ηj−1 , ηj )∪
D(ηj , ηj+1 )), then
(j−1) (j−1)
(a) if x(ηj−1 ,ηj ) = (x1
, x2
) = (+, −), then there exists d < x, d ∈ D(ηj−1 , ηj ), such that
(j−1)
(j−1)
(j−1)
(j−1)
(j−1)
(j−1)
(η
,η
)
d j−1 j = (d1
, d2
) with 0 < d1
< x1
and x2
< d2
< 0; and
′(j−1)
′(j−1)
′(η
,η
)
(b) if x j−1 j = (x1
, x2
) = (−, +), then there exists d < x, d ∈ D(ηj , ηj+1 ), such
′(j−1)
′(j−1)
′(j−1)
′(j−1)
′(j−1)
′(j−1)
′(η
,η
)
that d j−1 j = (d1
, d2
) with x1
≤ d1
≤ 0 and 0 < d2
< d2
.
We only need to apply the preceding cases at this (j − 1, j) level.
It remains to prove that the inequalities for the (j − 1, j) level (denoted by the bracketed
superscript) can be translated to the first level. To this end, we proceed as in the previous
point.
(j−1)
(j−1)
(j−1)
(j−1)
(j−1)
(a) Assume that 0 ≤ d1
< x1
and x2
< d2
< 0. Notice that d1
= 0 if
(j−1)
(j−1)
(j−1)
(j−1)
and only
ηj−1 + x2
ηj , d = d1
ηj−1 + d2
ηj and ηj =
j if dk = ηj . As x = x1
ηj−2 −
ηj−2
ηj−1
ηj−1 , we have that
k
j
ηj−2
(j−1)
ηj−1
ηj−1 + x2
ηj−2 − ηj−1
k
j
(j−1)
(j−1)
(j−1) ηj−2
ηj−1 ,
= x2
ηj−2 + x1
− x2
ηj−1
k
j
(j−1)
(j−1)
η
η
d = d1
ηj−1 + d2
ηj−2 − ηj−2
j−1
j−1
j
k
(j−1)
(j−1)
(j−1) ηj−2
= d2
ηj−2 + d1
− d2
ηj−1 .
ηj−1
(j−1)
x = x1
(j−1)
′(j−2)
(j−1)
′(j−2)
From these equalities we
< d2
= d1
< 0 and
j cank deduce that x1 j = kx2
(j−1)
′(j−2)
(j−1) ηj−2
′(j−2)
(j−1)
(j−1) ηj−2
− d2
> 0, where x′(ηj−2 ,ηj−1 ) =
x2
= x1
− x2
ηj−1 > d1
ηj−1 = d2
′(j−2)
′(j−2)
′(j−2)
′(j−2)
(x1
, x2
) = (−, +), and d′(ηj−2 ,ηj−1 ) = (d1
, d2
ensure that in this setting all the inequalities are strict.
) = (−, +). Notice that we can
8
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
′(j−1)
(b) Assume that 0 ≥ d1
′(j−1)
≥ x1
′(j−1)
and x2
′(j−1)
′(j−1)
> d2j
k> 0. Since x = x1
′(j−1)
′(j−1)
′(j−1)
ηj−2
x2
ηj , d = d1
ηj−1 + d2
ηj and ηj = ηj−2 − ηj−1 ηj−1 we have
ηj−1 +
k
j
′(j−1)
ηj−2
ηj−1
ηj−1 + x2
ηj−2 − ηj−1
j
k
′(j−1)
′(j−1)
′(j−1) ηj−2
= x2
ηj−2 + x1
− x2
ηj−1 ,
ηj−1
k
j
′(j−1)
′(j−1)
η
η
d = d1
ηj−1 + d2
ηj−2 − ηj−2
j−1
j−1
k
j
′(j−1)
′(j−1)
′(j−1) ηj−2
ηj−1 .
= d2
ηj−2 + d1
− d2
ηj−1
′(j−1)
x = x1
(j−2)
′(j−1)
′(j−1)
(j−2)
(j−2)
′(j−1)
Again, jit follows
that x1
=j x2 k > d2
= d1
> 0 and x2
= x1
−
k
′(j−1) ηj−2
′(j−1)
(j−2)
′(j−1) ηj−2
x2
− d2
≤ 0.
ηj−1 < d1
ηj−1 = d2
k
j
′(j−1)
′(j−1)
ηj−2
< d2
. Also
Notice that the inequality is strict because ηj−1 ≥ 1 and we have x2
(j−1)
′(j−1)
′(j−1)
d1
= 0 implies d = d2
ηj−2 , but since d ∈ Euc(δ1 , δ2 ), we deduce d2
= 1, that is, we
always can take d = ηj+1 in the (−, +) case to start the process.
In this way, we can apply alternatively the preceding processes to obtain the desired result.
Notice that only if we start with (b), we can have d′j−1
= 0. Hence, only when j − 1 = 1 (if we
1
perform another step, when applying (a) the inequality becomes strict), we will have that the
(η1 , η2 )-coordinates of d would be (0, 1), that is, d = η2 .
However, as the statement of (3) assumes j > 2, the inequalities will always be strict.
We will refer to d defined in Proposition 14 as the basement of x = (x1 , x2 ), denoted by bsm(x).
Notice that the basement of x may not be unique: however, for our purposes, we will consider as
the basement of x any d satisfying the properties stated in Proposition 14.
Example 15. In the setting of Example 11, consider x = 35. Since 33 = η4 ≤ x ≤ 109 = η3 ,
according to Proposition 14 we can take j = 3. As 35 = −13η3 + 44η4 = 20η3 − 65η4 ; we have
x(η3 ,η4 ) = (20, −65) and x′(η3 ,η4 ) = (−13, 44). Remember that these decompositions are unique using
Proposition 9.
For x′(η3 ,η4 ) = (−13, 44) we know that d′ = η4 = 33 and d′(η3 ,η4 ) = (0, 1); while we look for
d ∈ D(η3 , η4 ) = {76, 43, 10}. As 10 < x = 35 < 43 we take d = 10 and then d(η3 ,η4 ) = (1, −3). If
we translate them to level (δ1 , δ2 ), we have: 35 = −57δ1 + 158δ2 = 85δ1 − 235δ2 , and (x1 , x2 ) =
(85, −235) and (x′1 x′2 ) = (−57, 158). Thus, considering d′ = η4 = 33, as 33 = −η1 + 3η2 , we have
d′(δ1 ,δ2 ) = (d′1 , d′2 ) = (−1, 3). For the other case, by considering d = 10, we obtain 10 = 4η1 − 11η2 ,
that is, d(δ1 ,δ2 ) = (d1 , d2 ) = (4, −11). So we have x′1 < d′1 < 0 < d′2 < x′2 and x2 < d2 < 0 < d1 < x1 .
Observe that 33 is the d′ that yields Proposition 14, but it may happen that there are other
elements in Euc(δ1 , δ2 ) satisfying our purposes too.
It is not difficult to see that for any element in {34, ..., 42} we obtain the same results, that is,
d = 10 and d′ = 33, while if we take a number comprised between 44 and 75, we will obtain d = 43
and d′ = 33.
Let us, now, consider x = 15 which is between η5 = 10 and η4 = 33. In this case j = 4, then we
can write 15 = −5η4 + 18η5 = 5η4 − 15η5 . So, for x′(η4 ,η5 ) = (−5, 18) we know that d′ = η5 = 10
and d′(η4 ,η5 ) = (0, 1); while as d ∈ D(η4 , η5 ) = {23, 13, 3}, we can take d = 13 < x = 15 which
yields d(η4 ,η5 ) = (1, −2), satisfying Proposition 14 as x(η4 ,η5 ) = (5, −15). Observe that, in this case,
when we compute the basements at the first level, we obtain d = 10 because d(δ1 ,δ2 ) = (4, −11),
while x(δ1 ,δ2 ) = (77, −213), and d′ = 33 with d′(δ1 ,δ2 ) = (d′1 , d′2 ) = (−1, 3) for x′(δ1 ,δ2 ) = (−65, 180),
changing the roles of d and d′ from level j = 4 to the first level.
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
9
Our next goal is to associate bsm(x) to a vector in MS . We will have two such vectors per case,
associated to the decompositions in Proposition 9.
As above g = gcd(δ1 , δ2 ). For x ∈ {g, 2g, . . . , max{δ1 , δ2 }}, and {v1 , v2 } the basis for MS chosen
in Section 3, set
(1)
wx = x1 v1 + σx2 v2 ,
wx′ = x′1 v1 + σx′2 v2 ,
with (x1 , x2 ) = x(δ1 ,δ2 ) and (x′1 , x′2 ) = x′(δ1 ,δ2 ) . Notice that in this setting ℓ(wx ) = x1 δ1 + x2 δ2 =
x = x′1 δ1 + x′2 δ2 = ℓ(wx′ ).
In the following, we will use the bracketed subscript to refer to the coordinates of vectors. For
instance, we would have, for a generic vector v, v = (v(1) , v(2) , v(3) ). In particular, we will use the
following notation to refer to the coordinates of the vectors wx and wx′ we just introduced:
wx = (wx(1) , wx(2) , wx(3) ) and wx′ = (wx′ (1) , wx′ (2) , wx′ (3) ).
As x1 > 0 and x2 ≤ 0, following an argument similar to the proof of Proposition 8, it can be shown
that the signs of the coordinates of wx and wx′ are as in Table 1.
σ
wx
wx′
1 (?, −, +) (?, +, −)
-1 (+, ?, −) (−, ?, +)
Table 1. Signs for wx and wx′
Remember that v1 = (+, −, 0) and v2 = (+, +, −).
The following corollary shows that, for x as in Proposition 14, the vector bsm(x) has two coordinates that are in absolute value lower than those of wx or wx′ .
Corollary 16. Let the notations and hypotheses be as above. Let v = a1 v1 + a2 σv2 , with a1 and
a2 integers such that a1 a2 ≤ 0, −δ2 /g < a1 ≤ δ2 /g and −δ1 /g < a2 ≤ δ1 /g. Set x = ℓ(v) ∈
{g, 2g, . . . , max{δ1 , δ2 }} \ Euc(δ1 , δ2 ), and let d = bsm(x).
(i) If a1 > 0 (that is, v = wx ), there exists i ∈ {1, 2} such that |wd(i) | < |wx(i) | and |wd(3) | <
|wx(3) |. Moreover, for such an i we have that wd(i) wd(3) < 0.
(ii) If a1 < 0 (that is, v = wx′ ), there exists i ∈ {1, 2} such that |wd′ (i) | < |wx′ (i) | and |wd′ (3) | <
|wx′ (3) |. Moreover, for such an i we have that wd′ (i) wd′ (3) < 0.
Proof. Notice that x = a1 δ1 + a2 δ2 . According to Proposition 9, if a1 > 0, then (a1 , a2 ) = x(δ1 ,δ2 ) ;
otherwise (a1 , a2 ) = x′(δ1 ,δ2 ) . Hence, in the first case v = wx , while in the second case v = wx′ .
Suppose that a1 > 0. Following Proposition 14, consider d = d1 δ1 + d2 δ2 with a1 > d1 > 0 and
0 > d2 > a2 .
• If σ = 1, wx(3) = a1 v1(3) + a2 v2(3) = a2 v2(3) > d2 v2(3) = d1 v1(3) + d2 v2(3) = wd(3) > 0. We take
i = 2, obtaining wx(2) = a1 v1(2) + a2 v2(2) < d1 v1(2) + d2 v2(2) < 0.
• If σ = −1, wx(3) = a1 v1(3) − a2 v2(3) = −a2 v2(3) < −d2 v2(3) = d1 v1(3) − d2 v2(3) = wd(3) < 0.
Also wx(1) = a1 v1(1) − a2 v2(1) > d1 v1(1) − d2 v2(1) = wd(1) > 0.
Now assume that a1 < 0. Following Proposition 14 consider d = d′1 δ1 + d′2 δ2 with a1 < d′1 ≤ 0
and 0 < d′2 < a2 .
• If σ = 1, 0 > wd′ (3) = d′1 v1(3) + d′2 v2(3) = d′2 v2(3) > a1 v1(3) + a2 v2(3) = wx′ (3) . Now take i = 2.
Then 0 < wd′ (2) = d′1 v1(2) + d′2 v2(2) < a1 v1(2) + a2 v2(2) = wx′ (2) .
• If σ = −1, 0 < wd′ (3) = d′1 v1(3) − d′2 v2(3) = −d′2 v2(3) < −a2 v2(3) = a1 v1(3) − a2 v2(3) = wx′ (3) .
Here we choose i = 1, obtaining 0 > wd′ (1) = d′1 v1(1) − d′2 v2(1) > a1 v1(1) − a2 v2(1) = wx′ (1) .
10
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
This corollary ensures that for every wx or wx′ there exists a wd or wd′ , respectively, with d < x
and such that the two known coordinates of wx or wx′ , according to Table 1, are greater in absolute
value than those of wd or wd′ , respectively.
Now we are going to show how can uniqueness in Proposition 9 extend to the level of the vectors
wx and wx′ . We want to associate to each v ∈ MS such that ℓ(v) ∈ {0, g, 2g, . . . , max{δ1 , δ2 }} two
vectors wx , wx′ such that ℓ(wx ) = ℓ(wx′ ) = ℓ(v).
Next we see that if a vector has length equal to zero, then it is a multiple of the vector appearing
in Proposition 8. Hence, when this vector is added or subtracted to another vector, the length
remains the same.
Lemma 17. With the notation defined in Section 3 and Definition 10, let v ∈ MS with ℓ(v) = 0.
Then
v = α(δ2 /gv1 − σδ1 /gv2 ),
for some α ∈ Z.
Proof. Since {v1 , σv2 } is a basis for MS , we have that there exists λ1 , λ2 ∈ Z such that v =
λ1 v1 + λ2 σv2 . Now 0 = ℓ(v) = λ1 δ1 + λ2 δ2 . Since gcd(δ1 , δ2 ) = g, we deduce that δ1 /g divides λ2 ,
and δ2 /g divides λ1 . So there exists k1 and k2 integers such that λ1 = k1 δ2 /g and λ2 = k2 δ1 /g. We
then have 0 = k1 + k2 and consequently k1 = −k2 . Take α = k1 /g. Hence v = αδ2 v1 − αδ1 σv2 .
We now characterize the vectors in MS with length in {1, . . . , max{δ1 , δ2 }}.
Proposition 18. Under the hypotheses and notations of Section 3 and Definition 10, let v ∈ MS
with 0 < ℓ(v) ≤ max{δ1 , δ2 }. Then
v = a1 v1 + σa2 v2 + α(δ2 /gv1 − σδ1 /gv2 ),
for some a1 , a2 ∈ Z2 \ {(0, 0)} such that a1 a2 ≤ 0, −δ2 /g < a1 ≤ δ2 /g, −δ1 /g < a2 ≤ δ1 /g, where
g = gcd(δ1 , δ2 ) and α ∈ Z.
Furthermore, a1 , a2 and α are unique if we impose that the third coordinates of u = a1 v1 + σa2 v2
and v have the same sign.
Proof. We use once more that {v1 , σv2 } is a generating system for MS ; then there exist unique
λ1 , λ2 ∈ Z such that v = λ1 v1 + λ2 σv2 .
Let us prove that λ1 λ2 ≤ 0. Clearly, if λ1 λ2 = 0, we are done.
Assume that λ1 > 0 (the other case is analogous).
• If δ1 > δ2 , we have 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ max{δ1 , δ2 } = δ1 . Then λ2 δ2 ≤ (1 − λ1 )δ1 ≤ 0,
which implies λ2 ≤ 0.
• If δ2 > δ1 , we have 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ max{δ1 , δ2 } = δ2 . Then 0 < λ1 δ1 ≤ (1 − λ2 )δ2
and 0 < 1 − λ2 , yielding λ2 ≤ 0.
Next, we prove the following assertions.
λ1 > δ2 /g
λ1 ≤ −δ2 /g
−δ2 /g < λ1 ≤ 0
0 < λ1 ≤ δ2 /g
implies
implies
implies
implies
λ2 ≤ −δ1 /g,
λ2 > δ1 /g,
0 < λ2 ≤ δ1 /g,
−δ1 /g < λ2 ≤ 0.
(1) Assume that λ1 > δ2 /g.
• If δ1 > δ2 , we have that 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ δ1 . Hence λ2 δ2 ≤ (1 − λ1 )δ1 . From
λ1 > δ2 /g, we deduce (1 − λ1 ) ≤ −δ2 /g. Therefore, λ2 δ2 ≤ (1 − λ1 )δ1 ≤ −δ1 δ2 /g, and thus
λ2 ≤ −δ1 /g.
• If δ2 > δ1 , we have 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ δ2 , and then λ1 δ1 ≤ (1 − λ2 )δ2 . As δ2 /g < λ1 ,
we get δ2 δ1 /g < λ1 δ1 . This implies δ1 /g < (1 − λ2 ) or equivalently λ2 ≤ −δ1 /g.
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
11
(2) Assume now that λ1 ≤ −δ2 /g. Then 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ −δ2 δ1 /g + λ2 δ2 = (λ2 − δ1 /g)δ2 .
This forces λ2 > δ1 /g.
(3) Suppose that −δ2 /g < λ1 ≤ 0.
• If δ1 > δ2 , then 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ δ1 . As (1 − λ1 ) ≤ δ2 /g, we deduce that
λ2 δ2 ≤ (1 − λ1 )δ1 ≤ δ1 δ2 /g, and consequently λ2 ≤ δ1 /g.
• If δ2 > δ1 , we have ℓ(v) = λ1 δ1 + λ2 δ2 ≤ δ2 . Since −δ2 /g < λ1 , we obtain −δ2 δ1 /g + λ2 δ2 <
λ1 δ1 + λ2 δ2 ≤ δ2 . Hence λ2 − δ1 /g < 1, or equivalently, λ2 ≤ δ1 /g.
(4) Finally, assume that 0 < λ1 ≤ δ2 /g. Then 0 < ℓ(v) = λ1 δ1 + λ2 δ2 ≤ δ2 δ1 /g + λ2 δ2 =
(λ2 + δ1 /g)δ2 . Hence −δ1 /g < λ2 .
In order to conclude the proof it suffices to observe the following.
• If λ1 fulfills the conditions (3) or (4), we are done.
• If λ1 > δ2 /g, take α ∈ Z such that 0 < λ1 − αδ2 /g ≤ δ2 /g. Set a1 = λ1 − αδ2 /g ≤ δ2 /g and
a2 = λ2 + αδ1 /g. Define u = a1 v1 + σa2 v2 . Then ℓ(v) = ℓ(u), and v = u + α(δ2 /gv1 −
σδ1 /gv2 ). Now u fulfills (4), and we are done. Notice that if we take α + 1 instead of α,
we fall in case (3), with u′ = u − (δ2 /gv1 − σδ1 /gv2 ): thus it can be deduced that there are
two possible choices of α fulfilling the conditions of the statement.
• If λ1 < −δ2 /g, we proceed analogously.
Since there are two possible choices of α, but λ1 and λ2 are unique, there exist two possible choices
for (a1 , a2 ). In order to conclude the proof, it suffices to show that the signs of the third coordinates
of u and u′ are different (and thus one of them must be equal to the sign of the third coordinate of
v). Assume that α is such that 0 < a1 ≤ δ2 /g, and thus by (4) applied to u, −δ1 /g < σa2 ≤ 0.
If we write u′ = a′1 v1 + σa′2 v2 then, as a′1 = a1 − δ2 /g > 0, by (3) we have σa′2 > 0. Since a2 and
′
a2 have different signs, we are done.
This proposition will be used later to work with u instead of v.
The goal of this work is to prove the following theorem. For clarity’s sake, we first give an
example; then, we provide the intermediate results needed for its proof, which will be a direct
consequence of Lemmas 22 and 24.
Theorem 19. Let S be a symmetric numerical semigroup with embedding dimension three. With
the notation introduced in Section 3 and Definition 10, we have
∆(S) = Euc(δ1 , δ2 ) \ {0}.
Example 20. Let S = hs2 s3 , s1 s3 , s1 s2 i = h2015, 7124, 84940i, with s1 = 548, s2 = 155, and s3 = 13.
Then v1 = (548, −155, 0) and v2 = (0, 155, −13). Hence δ1 = 393 and δ2 = 142. Extended Euclid’s
Algorithm for δ1 = 393 and δ2 = 142 yields the Table 20, whose cells contain the vectors wd or wd′ ,
as defined in (1), for d ∈ Euc(δ1 , δ2 ).
According to Theorem 19, we would obtain
∆(S) = {1, 2, 3, 4, 7, 10, 13, 23, 33, 43, 76, 109, 142, 251, 393}.
Lemma 21. Let z, z′ ∈ Z(s), with 0 6= s ∈ S. With the notation introduced in Section 3, if
0 < ℓ(z) − ℓ(z′ ) < max{δ1 , δ2 }, then
(1) z − z′ is either of the form (?, +, −) or (?, −, +) when ℓ(v2 ) > 0 (σ = 1), and
(2) z − z′ is of the form (+, ?, −) or (−, ?, +) when ℓ(v2 ) < 0 (σ = −1).
Proof. Recall that z, z′ ∈ N3 . Since z − z′ ∈ MS and 0 < ℓ(z − z′ ) ≤ max{δ1 , δ2 } by Proposition
18 we know that z − z′ = a1 v1 + σa2 v2 + α(δ2 /gv1 − σδ1 /gv2 ). If α > 0, we have that z − z′ =
(αδ2 /g + a1 )v1 + σ(a2 − αδ1 /g)v2 , and clearly (αδ2 /g + a1 ) > 0 and (a2 − αδ1 )/g < 0. According
to Table 1, z − z′ = (?, −, +) if σ = 1, or z − z′ = (+, ?, −) if σ = −1. A similar argument shows
that if α < 0, then z − z′ = (?, +, −) for σ = 1, and z − z′ = (−, ?, +) for σ = −1.
For the case α = 0, as a1 a2 < 0 there are two possibilities.
12
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
wd = (?, −, +)
D(δ1 , δ2 )
wd′ = (?, +, −)
D(δ2 , δ3 )
wd = (?, −, +)
D(δ3 , δ4 )
wd′ = (?, +, −)
D(δ4 , δ5 )
wd = (?, −, +)
D(δ5 , δ6 )
wd′ = (?, +, −)
D(δ6 , δ7 )
w393 = v1
= (548, −155, 0)
δ1 = 393
′
w142
= v2
= (0, 155, −13)
δ2 = 142
w76 = 2v1 − 5v2
= (1096, −1085, 65)
δ3 − δ4 = 76
′
w23
= −5v1 + 14v2
= (−2740, 2945, −182)
δ4 − δ5 = 23
w7 = 17v1 − 47v2
= (9316, −9920, 611)
δ5 − δ6 = 7
w2 = −56v1 + 155v2
= (−30688, 32705, −2015)
δ6 − δ7 = 2
w251 = v1 − v2
= (548, −310, 13)
δ1 − δ2 = 251
′
w33
= −v1 + 3v2
= (−548, 620, −39)
δ2 − δ3 = 33
w43 = 3v1 − 8v2
= (1644, −1705, 104)
δ3 − 2δ4 = 43
′
w13
= −9v1 + 25v2
= (−4932, 5270, −325)
δ4 − 2δ5 = 13
w4 = 30v1 − 83v2
= (16440, −17515, 1079)
δ5 − 2δ6 = 4
w1 = −99v1 + 274v2
= (−54252, 57815, −3562)
δ6 − 2δ7 = 1
w109 = v1 − 2v2
= (548, −465, 26)
δ1 − 2δ2 = 109
w10 = 4v1 − 11v2
= (2192, −2325, 143)
δ3 − 3δ4 = 10
w3′ = −13v1 + 36v2
(−7124, 7595, −468)
δ4 − 3δ5 = 3
w1 = 43v1 − 119v2
= (23564, −25110, 1547)
δ5 − 3δ6 = 1
w0 = −142v1 + 393v2
= (−77816, 82925, −5109)
δ6 − 3δ7 = 0
Table 2. Table for h2015, 7124, 84940i
• If a1 < 0 and a2 > 0, then z − z′ = (?, +, −) for σ = 1, and z − z′ = (−, ?, +) for σ = −1.
• If a1 > 0 and a2 < 0, then z − z′ = (?, −, +) for σ = 1, and z − z′ = (+, ?, −) for σ = −1.
Lemma 22. Let S = hn1 , n2 , n3 i be a symmetric numerical semigroup with embedding dimension
three. With the notation introduced in Section 3 and Definition 10, let z, z′ ∈ Z(s), for some
s ∈ S \ {0} such that x = ℓ(z) − ℓ(z′ ) ∈
/ Euc(δ1 , δ2 ). Then there exists z′′ ∈ Z(s) such that
′
′′
ℓ(z ) < ℓ(z ) < ℓ(z).
′
′ |n , which is an element in S with two
Proof. If z(i) · z(i)
6= 0, we consider s′ = s − |z(i) − z(i)
i
′
′
factorizations z and z with ℓ(z) − ℓ(z ) = x and so that one of them has the ith coordinate equal
P
′ |e (e is
to zero. With this argument, we can assume z · z′ = 0. Later, by adding 3i=1 |z(i) − z(i)
i
i
th
the i row of the 3 × 3 identity matrix) to the factorizations obtained, we will recover the original
factorizations.
Now, from Proposition 18, we can write z − z′ = a1 v1 + σa2 v2 + α(δ2 /gv1 − σδ1 /gv2 ) and
′ ,z
′
′
u = a1 v1 + σa2 v2 in such a way that z − z′ = (z(1) − z(1)
(2) − z(2) , z(3) − z(3) ), u = (u(1) , u(2) , u(3) )
′ = u
and α(δ2 /gv1 − σδ1 /gv2 ) = (α(1) , α(2) , α(3) ). So, we have for j ∈ {1, 2, 3}, z(j) − z(j)
(j) + α(j) .
′
As the sign of the third coordinates of z − z and u is the same and the third coordinate of u is
smaller in absolute value, we necessarily deduce that α(3) has the same sign too. Looking on Table
′ ) = sgn(u ) = sgn(α ) and sgn(z
′
1 and Proposition 8 we have sgn(z(i) − z(i)
(i)
(i)
(j) − z(j) ) = sgn(u(j) ) =
sgn(α(j) ), for some i ∈ {1, 2} and j = 3,.
In light of Proposition 9, if −δ2 /g < a1 ≤ 0 < a2 ≤ δ1 /g, then −δ1 /g < a2 − δ1 /g ≤ 0 <
a1 + δ2 /g ≤ δ2 /g. According to the different cases in Corollary 16, we obtain the following.
(1) If a1 > 0, then v = (v(1) , v(2) , v(3) ) = wd for d < ℓ(z − z′ ) = ℓ(u).
(2) If a1 < 0, then v = (v(1) , v(2) , v(3) ) = wd′ for d < ℓ(z − z′ ) = ℓ(u).
Now we have that 0 < ℓ(v) = d < ℓ(z) − ℓ(z′ ). By adding ℓ(z′ ) we obtain ℓ(z′ ) < ℓ(v) + ℓ(z′ ) <
ℓ(z), and it can be easily deduced that 0 < ℓ(z′ ) < ℓ(z) − ℓ(v) < ℓ(z).
If all the coordinates of v + z′ are positive, or all the coordinates of z − v are positive, then we
would obtain a factorization of s with length comprised between ℓ(z′ ) and ℓ(z): thus we only have
to prove that either z − v or v + z′ has positive coordinates.
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
13
By Proposition 14 and the behavior of the coordinates described in Corollary 16, we can deduce
that two of the three coordinates (the second and third if ℓ(v2 ) > 0, and the first and third if
ℓ(v2 ) < 0) of v are smaller in absolute value than the corresponding coordinates of u. Thus we
have two possible cases, according to the coordinate whose sign we do not control (the first if
ℓ(v2 ) > 0, and the second if ℓ(v2 ) < 0).
Assume first that this coordinate is negative. In this setting, the corresponding coordinate of
z − v is positive, while for the other two we have:
• if v(j) < 0, for some j, as z(j) ≥ 0 we have that the j th coordinate of z − v will be positive,
and
′ ) = sgn(u ) = sgn(α ) and
• if v(j) > 0, we know that u(j) > v(j) > 0 and, as sgn(z(j) − z(j)
(j)
(j)
′
th
z(j) − z(j) − α(j) = u(j) > v(j) we have that z(j) − v(j) > 0, and thus the j coordinate z − v
would be again positive.
If the above unknown coordinate is positive, then we take v + z′ instead of z − v, and make the
same coordinate positive. For the other two coordinates, if v(j) > 0, we do nothing, and if v(j) < 0,
′ +α
′
as 0 < −v(j) < −u(j) = −z(j) + z(j)
(j) ≤ z(j) , remember that in this case α(j) is negative. As
′ is the j-th coordinate of z′ we would again have that z ′ + v
z(j)
(j) is a positive coordinate, and by
(j)
′
′
′
adding ℓ(z ) we obtain that ℓ(z ) < ℓ(v + z ) < ℓ(z).
Example 23. In the setting of Examples 11 and 15, take 11630n2 ∈ S and consider z = (10960, 5, 715)
and z′ = (0, 11603, 0). We have ℓ(z) = 11680 and ℓ(z′ ) = 11630, so ℓ(z − z′ ) = 50 < max{δ1 , δ2 }.
′
As z(2) · z(2)
6= 0, we can consider s′ = 11630n2 − 5n2 as the s′ in the proof of Lemma 22. So,
we are going to work with s′ = 11625n2 , z = (10960, 0, 715) and z′ = (0, 11625, 0), obtaining
z − z′ = (10960, −11625, 715) = 20v1 − 5v2 . In this case α = 0, whence u = 20v1 − 5v2 . As
we showed in Example 15, we can consider wd = w43 = 3v1 − 8v2 = (1644, −1705, 104). Now, if
we consider wd + z′ = (1644, 9920, 104), we obtain a new factorization of 11625n2 with 11630 =
ℓ(z′ ) < ℓ(wd + z′ ) = 11668 < ℓ(z) = 11680, while if we consider z − wd = (9316, 1705, 611), a new
factorization of 11625n2 can be obtained, with ℓ(z − wd ) = 11632 which, again, is between ℓ(z′ ) and
ℓ(z). In this case, both factorizations can be chosen. Finally adding 5 to the second coordinate, we
obtain factorizations of the original element, 11630n2 .
Now if we choose 71300n2 ∈ S, we have two factorizations z = (66856, 0, 4394) and z′ =
(0, 71300, 0) whose lengths are ℓ(z) = 71250 and ℓ(z′ ) = 71300, obtaining again that ℓ(z′ ) − ℓ(z) =
50. In this case u = −122v1 + 338v2 because z′ − z is of the type (?, +, −) so a1 < 0 and
′ = −v + 3v = (−548, 620, −39) to obtain
a2 > 0. According to Example 15, we can pick w33
1
2
′
′
′
wd + z = (66308, 620, 4355) and z − wd = (548, 70680, 39); hence we obtained two factorizations of
71300n2 whose lengths ℓ(wd′ + z) = 71283 and ℓ(z′ − wd′ ) = 71267 are both between ℓ(z) and ℓ(z′ ).
Lemma 24. Let S = hn1 , n2 , n3 i be a symmetric numerical semigroup with embedding dimension
three. With the notation introduced in Section 3 and Definition 10, consider d ∈ Euc(δ1 , δ2 ). Then
there exists s ∈ S and z, z′ ∈ Z(s), with d = ℓ(z) − ℓ(z′ ), and such that for any other z′′ ∈ Z(s)
either ℓ(z′′ ) ≤ ℓ(z′ ) or ℓ(z′′ ) ≥ ℓ(z).
Proof. Let d ∈ Euc(δ1 , δ2 ), and set v = x1 v1 + σx2 v2 , with x1 x2 < 0, and −δ1 /g < x2 ≤ δ1 /g and
−δ2 /g < x1 ≤ δ2 /g, such that ℓ(v) = d. Then
v = (x1 v1(1) + σx2 v2(1) , x1 v1(2) + σx2 v2(2) , −σx2 v2(3) ),
with ℓ(v) = d > 0. Clearly, v has two coordinates with the same sign, and the other with opposite
sign. Let us denote the latter by i.
Set s = |vx(i) |ni , and z = (vx+(1) , vx+(2) , vx+(3) ) and z′ = (vx−(1) , vx−(2) , vx−(3) ), where a+ = (|a| + a)/2 ≥ 0
and a− = (|a| − a)/2 ≥ 0. Then z and z′ are factorizations of s. The rest of the factorizations
of s are of the form z′′ = z′ + a1 v1 + a2 σv2 , with a1 a2 < 0. We are concerned with those such
14
P. A. GARCÍA-SÁNCHEZ, D. LLENA, AND A. MOSCARIELLO
that 0 < a1 δ1 + a2 δ2 < d, in order to find one with length between ℓ(z) and ℓ(z′ ). Denote u =
a1 v1 + a2 σv2 = (u(1) , u(2) , u(3) ).
Next, we prove that |a1 | ≥ |x1 | and |a2 | > |x2 |.
Assume that ℓ(u) ∈ Euc(δ1 , δ2 ). Then, as ℓ(u) < ℓ(v), by construction (see Remark 12) |a1 | ≥ |x1 |
and |a2 | > |x2 |. On the other hand, if ℓ(u) ∈
/ Euc(δ1 , δ2 ), then we consider e = bsm(ℓ(u)) ∈
Euc(δ1 , δ2 ) associated to u (Proposition 14), and then we have |a1 | > |e1 | > |x1 | and |a2 | > |e2 | >
|x2 |, and the assertion is proved.
Thus, Table 3 collects the different possible settings.
x1
x′1
a1
a′1
σ=1
σ = −1
v1 = (+, −, 0), v2 = (+, +, −) v1 = (+, −, 0), σv2 = (−, −, +)
> 0, x2 < 0
v = (?, −, +)
v = (+, ?, −)
v = (?, +, −)
v = (−, ?, +)
< 0, x′2 > 0
> 0, a2 < 0
u = (?, −, +)
u = (+, ?, −)
′
u = (?, +, −)
u = (−, ?, +)
< 0, a2 > 0
Table 3. Signs for v and u
According to Table 3, we have the following cases.
(1) If σ = 1, we obtain that |u(2) | > |v(2) | and |u(3) | > |v(3) |
′
′′ < 0,
(a) If v = (?, −, +), then z(2) = 0, and z(2)
≤ 0. If in addition u(2) < 0, then z(2)
which is a contradiction; and if u(2) > 0 we derive u(3) < 0 and as |u3 | = |a2 ||v2(3) | >
′′ < 0, which is another contradiction.
|x2 ||v2(3) | = |v(3) |, we obtain z(3)
′′ < 0, a contradiction. If u
(b) If v = (?, +, −), then z(3) = 0. If u(3) < 0, then z(3)
(3) >
0, we have that u(2) < 0 and as sgn(a1 v1(2) ) = sgn(a2 v2(2) ), we can assure |u(2) | =
′′ < 0, which is again
|a1 ||v1(2) | + |a2 ||v2(2) | > |x1 ||v1(2) | + |x2 ||v2(2) | = |v(2) |. So we get z(2)
a contradiction.
(2) If σ = −1, we have that |u(1) | > |v(1) | and |u(3) | > |v(3) |.
′′
(a) If v = (+, ?, −), then z(3) = 0. The case u(3) < 0 leads to z(3)
< 0, which is a
contradiction; while from u(3) > 0 we deduce that u(1) < 0, and as sgn(a1 v1(1) ) =
sgn(σa2 v2(1) ), we can assure |u(1) | = |a1 ||v1(1) | + |a2 ||σv2(1) | > |x1 ||v1(1) | + |x2 ||σv2(1) | =
′′ < 0, yielding once more a contradiction.
|v(1) | we obtain z(1)
′′ < 0, yielding a
(b) If v = (−, ?, +), then z(1) = 0. If the inequality u(1) < 0 holds, then z(1)
contradiction. If u(1) > 0, then u(3) < 0, and as |u(3) | = |a2 ||σv2(3) | > |x2 ||σv2(3) | = |v(3) |
′′ < 0, which is also a contradiction.
we derive z(3)
This proves that there is no factorization of s with length between ℓ(z) and ℓ(z′ ).
References
[1] A. Assi, P. A. Garcı́a-Sánchez, Numerical semigroups and applications, RSME Springer series 1, Springer, Switzerland, 2016.
[2] T. Barron, C. O’Neill, R. Pelayo, On dynamic algorithms for factorization invariants in numerical monoids, to
appear in Math of Comp. arXiv:1507.07435.
[3] C. Bowles, S. T. Chapman, N. Kaplan and D. Reiser, On Delta Sets of Numerical Monoids, J. Algebra Appl.5
(2006), 1–24.
[4] S. Colton, N. Kaplan, The Realization Problem for Delta Sets of Numerical Semigroups, to appear in J. Commut.
Algebra. arXiv:1503.08496.
[5] S. T. Chapman, R. Hoyer and N. Kaplan, Delta Sets of Numerical Monoids are Eventually Periodic, Aequationes
Math. 77(2009), 273–279.
DELTA SETS FOR SYMMETRIC NUMERICAL SEMIGROUPS WITH EMBEDDING DIMENSION THREE
15
[6] S. T. Chapman, P. A. Garcı́a-Sánchez, D. Llena, A. Malyshev, D. Steinberg, On the Delta set and the Betti
elements of a BF-monoid, Arab. J. Math. 1 (2012), 53–61.
[7] M. Delgado, P.A. Garcı́a-Sánchez, and J. Morais, NumericalSgps, A package for numerical semigroups, Version
1.0.1 dev (2015), (Refereed GAP package), http://www.gap-system.org/Packages/numericalsgps.html.
[8] The GAP Group, GAP – Groups,
Algorithms,
and Programming,
Version 4.7.5;
2014,
(http://www.gap-system.org).
[9] J.I. Garcı́a-Garcı́a, M.A. Moreno-Frı́as, A Vigneron-Tenorio, Computation of Delta sets of numerical monoids,
Monatsh. Math. 178-3 (2015), 457-472.
[10] P.A. Garcı́a-Sánchez, D. Llena, A. Moscariello, Delta sets for numerical semigroups with embedding dimension
three. arXiv:1504.02116. To Appear in Forum Math.
[11] P.A. Garcı́a-Sánchez, C. O’Neill, G. Webb, On the computation of factorization invariants for affine semigroups,
arXiv:1504.02998.
[12] P. A. Garcı́a-Sánchez, I. Ojeda, J.C. Rosales, Affine semigroups having a unique Betti element, J. Algebra Appl.
12 (2013), 1250177 (11 pages).
[13] A. Geroldinger, On the arithmetic of certain not integrally closed noetherian integral domains, Comm. Algebra
19(1991), 685–698.
[14] J. C. Rosales, P. A. Garcı́a-Sánchez, Numerical Semigroups, Developments in Mathematics 20, Springer New
York (2009).
Departamento de Álgebra and IEMath-GR, Universidad de Granada, 18071 Granada, España
E-mail address: [email protected]
Departamento de Matemáticas, Universidad de Almerı́a, 04120 Almerı́a, España
E-mail address: [email protected]
Dipartimento di Matematica, Universit di Pisa, Largo Bruno Pontecorvo 5, 56127 Pisa, Italy.
E-mail address: [email protected]
| 0math.AC
|
arXiv:1703.08968v1 [math.GR] 27 Mar 2017
The normal closure of big Dehn twists, and plate
spinning with rotating families
François Dahmani
March 28, 2017
Abstract
We study the normal closure of a big power of one or several Dehn
twists in a Mapping Class Group. We prove that it has a presentation
whose relators consists only of commutators between twists of disjoint
support, thus answering a question of Ivanov. Our method is to use
the theory of projection complexes of Bestvina Bromberg and Fujiwara,
together with the theory of rotating families, simultaneously on several
spaces.
Contents
1 Composite projection systems
1.1 Projection systems . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Composite projection systems . . . . . . . . . . . . . . . . . .
1.3 Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
6
8
11
2 Composite rotating families and windmills
2.1 Definition . . . . . . . . . . . . . . . . . . .
2.2 Osculations of two kinds . . . . . . . . . . .
2.3 The unfolding in the different coordinates .
2.4 Towers of windmills, and accessibility . . . .
2.5 End of the proof of the main Theorem 2.2 .
13
13
15
16
24
26
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Conclusion, application to Dehn twists, and Theorem 1
.
.
.
.
.
.
.
.
.
.
26
Introduction
Consider a closed orientable surface Σ of negative Euler characteristic. The
Mapping Class Group of Σ, denoted by MCG(Σ), is the quotient of the group
1
of orientation-preserving homeomorphisms by the path-connected component of the identity. A classical theorem of Dehn and Nielsen indicates a
natural isomorphism between this group and a subgroup of index 2 of the
outer automorphism group of π1 (Σ).
As Riemann uniformisation theorem makes π1 (Σ) act as a lattice on the
hyperbolic plane, one can argue that MCG(Σ) is (in a sense) some hyperbolic
analogue of SL2 (Z) which is of index 2 in the automorphism group of Z2 , a
lattice in the euclidean plane.
However, contrarily to SL2 (Z), some elements of MCG(Σ) have large
centraliser. For instance, consider a simple closed curve α on Σ, a tubular
neighborhood of it α(t) ≃ [−ǫ, ǫ] × α ֒→ Σ and define a (simple) Dehn twist
τ as the identity in Σ \ α(t) , and as a full twist on α(t) , namely, identifying α
(η+ǫ)π
with S 1 , the map [(η, eiθ ) 7→ (η, ei(θ+ ǫ ) )]. A Dehn twist will obviously
commute with any mapping class whose support is disjoint from this tube,
and therefore with a lot of other Dehn twists. By a theorem of Dehn,
MCG(Σ) is generated by Dehn twists around simple closed curves, thus
by an intricate set of generators linked by commutation relations, but also
braid relations and lantern relations. These differences can lead to modify
the expected analogy with the euclidean case in order to include SLn (Z) for
n ≥ 3 (generated by elementary matrices).
Thurston’s classified mapping classes into three cases, those of finite
order, those that are reducible in the sense that they have infinite order
and that some nontrivial power preserves the homotopy class of a simple
closed curve, and finally the pseudo-Anosov. The pseudo-Anosov mapping
classes happen to be the hyperbolic isometries of an action of MCG(Σ) on
an important graph, the curve graph of Σ, which is Gromov hyperbolic
[MM]. They are, in many ways, the witnesses that some phenomena of
rank one happen in MCG(Σ) that are similar to the structure of SL2 (Z),
and its action on the modular tree. On the other hand, Dehn twists are
as reducible as it is possible to be. They are, or should be, the witnesses
of some phenomena of higher rank, similar to the structure of SLn (Z) for
n ≥ 3.
Here is an illustration of the difference of behaviors. If one considers a
finite collection of pseudo-Anosov elements, one can (easily) show that, after
taking powers, the group they generate is free. This is a simple ping-pong
argument in the hyperbolic curve graph. If one considers a finite collection
of Dehn twists around simple closed curves, then Koberda [K] proved the
beautiful ping-pong result that the group generated by some powers of these
Dehn twists is a right angled Artin group: a group whose presentation over
2
the given generating set is a collection of commutators, the obvious ones
(two Dehn twists commute if their curves are disjoint).
The case of normal subgroups is our interest. If n ≥ 3, by Margulis’
normal subgroup theorem, all normal subgroups of SLn (Z) are finite or of
finite index. In SL2 (Z) it is not the case: this group is virtually free, and
has uncountably many non-isomorphic quotients.
It is a natural question to ask whether (and how) these phenomena are
seen in MCG(Σ). What can be the normal closure of a power of a pseudoAnosov, the normal closure of a power of a Dehn twist, and the group
generated by all k-th powers of all simple Dehn twists ? Farb and Ivanov
asked this question in the case of a pseudo-Anosov (respectively [Fa, §2.4]
and [I, §3]), attributing it to Long, McCarthy, and Penner. Ivanov also
asked what he calls the deep relation question [I, §12], that is whether all
relations among certain powers of Dehn twists must derive from ovbious
commutation relations.
In [DGO, §5], we answered the first question: there is an integer N =
N (Σ) such that for all pseudo-Anosov mapping class γ, the normal closure
hhγ N iiMCG(Σ) is free, and consists only of pseudo-Anosov elements and the
identity. This is in line with what happens in SL2 (Z), for all infinite order
element.
We are interested in the question of the closure of a power of a Dehn
twist, and in the group generated by certain powers of all (simple) Dehn
twists, as in Ivanov’s deep relation problem. A naive expectation along
the lines of the analogy with SLn (Z), and the Margulis normal subgroup
theorem, could be to expect such normal subgroups to be a finite index
subgroup. Whereas it is the case for squares of Dehn twists [H], it is not
the case for large powers (see [H], [Fu], [Cou, 6.17], see also [S] and [Mas]
for the case of powers of half-Dehn twists on punctured spheres). Another
expectation could be, in light of the finite-type situation, and ping pong
arguments, to expect infinitely generated right angled Artin groups. Again,
this is not the case in general (see [CLM] and [BM]; Brendle and Margalit
proved restrictions on the automorphism group of certain of these normal
subgroups, that forbid them to be right angled Artin groups). However, we
indeed prove that there is no need of another relation than the obvious ones.
Theorem 1. For every orientable closed surface Σ, there is N such that:
• for any Dehn twist τ , the normal closure of τ N in the Mapping Class
Group of Σ has a partially commutative presentation, built on an infinite set of generators that are conjugates of τ N , so that the relators
3
are commutations between pairs of conjugates of τ N that have disjoint
underlying curves.
• the group generated by all N -th powers of all simple Dehn twists has
a partially commutative presentation, built on an infinite set of generators that are N -th powers of Dehn twists, and whose relators are
commutations between pairs of conjugates of the generators that have
disjoint underlying curves.
The difference with an infinitely generated right angled Artin group is
that some elements in the commutator relators are not in the generating
set, but merely conjugates of elements in the generating set. We recover
that the normal closure is far from being of finite index in MCG(Σ), for
instance because it has abelianisation of infinite rank (the relators being in
the derived subgroup of the free group over the set of generators).
In our point of view, this result above, and its departure from the complexity of normal subgroups of SLn (Z) for n ≥ 3 (granted by Margulis
normal subgroup theorem) reinforce [Fu, Cou] in witnessing a dent in the
analogy between MCG(Σ) and SLn (Z). It also answers Ivanov’s question
on deep relations.
Let us discuss the proof of this theorem.
In [DGO] the structure of the normal closure of a big pseudo-Anosov
was studied with the help of rotating families. Consider G a group acting
by isometries on a space X. A rotating family in G on X is a collection of
subgroups (the rotation groups), that is closed for conjugacy, such that each
of them fix a certain point in X (thus inducing some kind of rotation around
this point). Take ρ in one of these subgroups, fixing c. One may measure
an analogue of the angle of rotation of ρ by taking x at distance 1 from c,
and measuring the infimal length between x and ρx of paths outside the ball
of radius 1 around c. If X is Gromov-hyperbolic (for a small hyperbolicity
constant), if the fixed point of the different rotation groups are sufficiently
far from each other, and if the angles of rotations are sufficiently big, the
group generated by all the rotation groups is a free product of a selection
of them. In [DGO] we applied this theory to the action of MCG(Σ) on a
cone-off of the curve graph of Σ. The rotation groups were the conjugates
of the big pseudo-Anosov considered.
The rotating family argument can be explained as follows. One analyses
the structure of groups generated by more and more rotation groups, to
discover that they arrange as a sequence of free products. Starting from a
quasi-convex set W (that will change over time) that is at first a small ball
4
around a single fixed point of a single rotation group, one make W grow
until it (almost) touches another center of rotation, for some other group.
Then one unfolds W into W ′ by taking its images by the group generated
by the new rotations, and the rotation already with center in W . Because of
hyperbolicity, and of largeness of angles of rotations involved, the resulting
space is still quasi-convex, with the almost the same constant – with a little
repair, it has the same quasi-convexity constant indeed. Actually W ′ has
a structure of tree whose vertices are the images of W by the group, and
the fixed points of the new rotations involved, thus giving by Bass-Serre
duality the structure of free product of the group genertated by rotations
whose center are in W ′ (edge stabilizers are trivial since no element can fix
two different centers of rotation). Then, one takes the new W as W ′ and
do over. In the direct limit, the group generated by all rotations has been
described as a free product of a selection of rotation groups.
In [BBF], Bestvina Bromberg and Fujiwara, using a system of subsurface
projections, discovered that there is a normal finite index subgroup G0 of
MCG(Σ) that acts on some spaces quasi-isometric to trees, and on which
Dehn twists behave like large rotation subgroups. It has been observed by
several people that this implies that the normal closure of a certain power of
a Dehn twist in G0 is free, using the argument of [DGO]. However, it is far
from obvious how to promote this structural feature to the normal closure
in MCG(Σ).
In this paper, we use several quasi-trees as above, one for each left coset
of G0 in MCG(Σ). The group G0 acts on each of them, but its action is
twisted by the automorphism of G0 that is the conjugation by elements
gi , i = 1, . . . , m realising a transversal of G0 in MCG(Σ). If τ N is a Dehn
twist in G0 , the normal closure of τ N in MCG(Σ) equals the normal closure
of the collection {gi τ N gi−1 , i = 1, . . . m} in G0 . Each gi τ N gi−1 is a legitimate
rotation on the quasi-tree associated to gi .
The argument of [DGO] is then performed simultanously on each of the
m quasi-trees. Instead of one convex subset that grows, and gets unfolded
in a hyperbolic space, we have m convex sets W1 , . . . , Wm in the m quasitrees. Each of them is invariant by the group generated by the rotations
around rotation points in all of them. One looks for a rotation point R
that is nearby one of these sets, and in a certain sense, nearby all of them
(although they do not live in the same quasi-trees, this still makes sense
in the framework of projection systems). Then, one unfolds our convex
sets in all coordinates i = 1, . . . , m. A funny phenomenon happens. The
unfolding in the coordinate of R provides a nice tree, as the argument of
[DGO], and the convexity of the result is quantitively very good. This
5
tree gives the structure of the new group by Bass-Serre duality, and reveals
that only commutation relations are involved. There is no reason that the
unfolding in all other coordinates produce something resembling to a tree,
and could in principle destroy the convexity of Wj . However, using the
properties of the projection system, we show that the result is still somehow
convex (less convex than before though). The game is then to unfold in the
different quasitrees at regular intervals of time in the process, and to control
the degradation of the convexity so that the repair can wait until a new
unfolding occurs. It is a game of plate spinning.
The quasi-trees that we will use come from projection complexes defined
in [BBF]. We wrote the argument in this axiomatic language, to avoid dealing with useless hyperbolicity constants. In the end, even if the spaces are
indeed quasi-trees, this fact does not appear in the argument. The axioms
of projection systems are extensively used though, and they contain the information that the geometric space is a quasi-tree. We will thus prove a
similar statement as Theorem 1, namely Theorem 2.2, that gives the structure of groups generated by composite rotating families. There is actually
more information coming from this composite rotating family structure, as
for instance the Greendlinger property (see Definition 2.3), that describes
how an element in the group can be shortened in ome coordinate of the
composite projection system.
1
Composite projection systems
1.1
Projection systems
Let us recall a part of the axiomatic construction of [BBF].
Definition 1.1. ([BBF])
A projection system is a set Y, with a constant θ > 0, and for each
Y ∈ Y, a function (dπY : Y \ {Y } × Y \ {Y } → R+ ) satisfying four or five
axioms:
• symmetry (dπY (X, Z) = dπY (Z, X)),
• triangle inequality (dπY (X, Z) + dπY (Z, W ) ≥ dπY (X, W )),
• Behrstock inequality (min{dπY (X, Z), dπZ (X, Y )} ≤ θ),
• properness ({Y, dπY (X, Z) > θ} is finite).
• In this work one also assume the separation (dπY (Z, Z) ≤ θ).
6
Observe that if the axioms are true for some θ they hold for all larger θ.
From this rudimentary axiomatic set, Bestvina Bromberg and Fujiwara
manage to extract meaningful geometry, by modifying the functions dπY into
some functions dY , that satisfy much more properties, usually encapsulated
in the statement that the projection complex of Y, for a suitable parameter
K is a quasi-tree.
One should think of dY (or dπY ) as an angular measure between X and
Z seen from Y . The axioms fit in this viewpoint: the Behrstock inequality
says that if the angle at Y between X and Z is large, then from the point
of view of Z, the items Y and X look aligned.
Let us review very quickly the procedure of [BBF] to produce the functions dY . Given θ for which the axioms hold, [BBF] define H(X, Z) to be
the set of pairs (X ′ , Z ′ ) such that both dπX and dπZ between them is strictly
larger than 2θ, and one also include the pairs (X, Z ′ ) if dπZ (X, Z ′ ) > 2θ, symmetrically the pairs (X ′ , Z) if dπX (X ′ , Z) > 2θ, and finally the pair (X, Z)
itself.
Then dY (X, Z) is defined to be the infimum of dπY over H(X, Z).
For all K, YK (X, Z) denotes the set {Y, dY (X, Z) ≥ K}.
[BBF, Theorem 3.3] states that there exists Θ and κ ≥ θ depending only
on θ, such that
• (Symmetry) dY (X, Z) = dY (Z, X) for all Y, X, Z
• (Coarse equality) dπY − κ ≤ dY ≤ dπY
• (Coarse triangle inequality) dY (X, Z) + dY (Z, W ) ≥ dY (X, W ) − κ
• (Behrstock inequality) min{dY (X, Z), dX (Y, Z)} ≤ κ
• (Properness) {Y, dY (X, Z) > Θ} is finite
• (Monotonicity) If dY (X, Z) ≥ Θ then both dW (X, Y ), dW (Z, Y ) are
less than dW (X, Z).
˙ such that
• (Order) YΘ (X, Z) ∪ {X, Z} is totally ordered by an order <
˙ 1 <Y
˙ 2 , then
X is lowest, Z is greatest, and if Y0 <Y
dY1 (X, Z) − κ ≤ dY1 (Y0 , Y2 ) ≤ dY1 (X, Z),
and
dY0 (Y1 , Y2 ) ≤ κ,
7
dY2 (Y1 , Y0 ) ≤ κ.
Then choosing K larger than Θ, the projection complex PK (Y) is defined
as follows: it is a graph whose vertices are the elements of Y and where X, Z
span an edge if and only if YK (X, Z) = ∅. Then [BBF, Thm. 3.16] states
that for sufficiently large K, PK (Y) is connected and quasi-isometric to a
tree for its path metric.
1.2
Composite projection systems
In this work, we are concern with a composite situation.
1.2.1
Definitions, and projection complexes
Let Y∗ be the disjoint union of finitely many countable sets Y1 , . . . , Ym .
Their indices i = 1, . . . , m are called the coordinates. Given Y ∈ Y∗ , denote
by i(Y ) its coordinate: Y ∈ Yi(Y ) .
Definition 1.2. A composite projection system on a countable set Y∗ =
⊔m
i=1 Yi is the data of a constant θ > 0, of a family of subsets Act(Y ) ⊂
Y∗ , Y ∈ Y∗ (the active set for Y ) such that Yi(Y ) ⊂ Act(Y ), and of a family
of functions dπY : (Act(Y ) \ {Y } × Act(Y ) \ {Y }) → R+ , satisfying the
symmetry, the triangle inequality, the Behrstock inequality for θ whenever
both quantities are defined, the properness for θ when restricted to each Yi ,
the separation for θ, and also three other properties related to the map Act:
• (symmetry in action) X ∈ Act(Y ) if and only if Y ∈ Act(X),
• (closeness in inaction) if X ∈
/ Act(Z), for all Y ∈ Act(X) ∩ Act(Z),
dπY (X, Z) ≤ θ
• (finite filling) for all Z ⊂ Y∗ , there is a finite collection of elements
Xj in Z such that ∪j Act(Xj ) covers ∪X∈Z Act(X).
The closeness in inaction can be understood as a complement to Behrstock inequality: “if dπY (X, Z) > θ, then dπX (Y, Z) is defined and is less than
θ”.
Applying [BBF] (as recalled in the previous subsection) we get, for each
coordinate i ≤ m, and for a suitable choice of θ, a modified function dY :
Yi × Yi → R+ . This function is unfortunately not defined on Act(Y ) \ Yi ,
but dπY is defined on it, and thus we choose to define d∢
Y (X, Z) to be dY if
π
both X, Z are in Yi and dY otherwise.
We then define YjM (X, Z) = {Y ∈ Yj ∩ Act(X) ∩ Act(Z), d∢
Y (X, Z) ≥
M }. The elements X, Y, Z need not be in the same coordinate.
8
In the following we first choose θ such that the construction of [BBF] applies for all coordinate Yi , and this provides the constants Θ and κ (suitable
for all coordinate).
Then we choose c∗ > 1000(Θ + κ), and ΘP = c∗ + 21mκ. One can choose
K > ΘP . sufficiently large to get quasi-trees in all coordinate, but this is
not important for us.
Finally, choose ΘRot > 2c∗ + 2ΘP + 20(κ + Θ) for later purpose.
To keep track with the constants, it is worth keeping in mind that
ΘRot >> 2ΘP >> 2c∗ >> 20(Θ + κ) >> θ.
1.2.2
Group in the picture
An automorphism of composite projection system is a map ψ : Y∗ → Y∗
• that induces a bijection on each Yi ,
• that sends Act(Y ) to Act(ψ(Y )),
∢
• such that for all Y , and all X, Z ∈ Act(Y ), d∢
Y (X, Z) = dψ(Y ) (ψ(X), ψ(Z)).
A rotation around X ∈ Y∗ in a composite projection system Y∗ is an
automorphism ψ such that ψ(X) = X, and such that for all Y ∈ Y∗ \Act(X),
∢
and for all W, Z ∈ Act(Y ), ψ(Y ) = Y , and d∢
Y (W, Z) = dY (W, ψ(Z)).
Let us now assume that a group G acts on the composite projection
system by automorphisms.
Let us denote by GX the stabilizer of X ∈ Y∗ .
We say that a subgroup ΓX < GX has proper isotropy if for all N > 0
there is a finite subset F (N ) of ΓX such that if γ ∈ ΓX \ F (N ), and if
Y ∈ Act(X), then dπX (Y, γY ) > N .
1.2.3
Betweenness and orbit estimates
Lemma 1.3. (Betweenness is transitive)
∢
∢
If d∢
Y (X, Z) > 2κ and dZ (Y, T ) > 2κ, then Z is in Act(X) and dZ (X, T ) ≥
d∢
Z (Y, T ) − 2κ.
∢
∢
∢
If d∢
Y (X, Z) > 10κ and dZ (X, T ) > 10κ, then dY (X, T ) ≥ dY (X, Z)−2κ.
Proof. By Behrstock inequality, one has d∢
Z (X, Y ) ≤ κ in both cases. For the
∢
∢
first implication, by triangular inequality, d∢
Z (X, T ) ≥ dZ (Y, T )−dZ (X, Y )−
κ.
∢
For the second implication, d∢
Z (Y, T ) is within 2κ from dZ (X, T ). Behr∢
∢
stock inequality gives that dY (Z, T ) ≤ κ and therefore dY (X, T ) ≥ d∢
Y (X, Z)−
2κ.
9
Lemma 1.4. (Orbit estimates, or transfert in a coordinate)
Assume that ΓX has proper isotropy.
For the finite subset F = F (10κ) of ΓX , and for all Y ∈ Act(X), and all
′
X that is either in Act(Y ) or in Act(X), and all γ ∈ ΓX \ F , then either
∢
′
′
d∢
Y (X , X) ≤ κ or dY (γX , X) ≤ κ.
′
Proof. Let us first treat the case of X ′ ∈ Act(Y ). If d∢
Y (X , X) ≤ κ we are
∢
′
done. Assume that dY (X , X) > κ. By closeness in inaction, X ′ ∈ Act(X),
′
and by Behrstock inequality (and because κ ≥ θ), one has d∢
X (X , Y ) ≤ κ.
∢
′
By proper isotropy (and coarse triangle inequality), dX (γX , Y ) > 5κ. Thus,
′
by Behrstock inequality again, d∢
Y (γX , X) ≤ κ.
Now assume that X ′ ∈
/ Act(Y ), but is in Act(X). Since Y ∈ Act(X) we
′ , Y ) and (since Γ
∢
′
can measure d∢
(X
X preserves Act(X)) also dX (γX , Y ).
X
′
′
By proper isotropy, d∢
X (X , γX ) ≥ 10κ and therefore at least one of the
∢
∢
′
quantities dX (X , Y ) and dX (γX ′ , Y ) is larger than 4κ. Assume for instance
∢
′
′
that d∢
X (X , Y ) ≥ 4κ. Then by Behrstock inequality, dY (X , X) ≤ κ.
Using this lemma four times, together with triangle inequality, one gets:
Lemma 1.5. (Orbit estimates for proper isotropy)
Let X1 , X2 , X1′ , X2′ such that X1 , X2 ∈ Act(Y ). Assume that either Xi′
is in Act(Y ) or in Act(Xi ).
If the group ΓX1 and ΓX2 have proper isotropy, then for almost all elements γ1 ∈ ΓX1 and γ2 ∈ ΓX2 , one has
′
′
∢
′
′
∢
d∢
Y (γ1 (X1 ), γ2 (X2 )) − 4κ ≤ dY (X1 , X2 ) ≤ dY (γ1 (X1 ), γ2 (X2 )) + 4κ.
Recall that we chose K > 2Θ + κ.
Proposition 1.6. (Ellipticity)
Given X ∈ Y∗ , and any j ≤ m, the group GX has an orbit in PK (Yj )
of diameter at most 1.
Proof. If j = i(X), and more generally, if GX fixes an element Y ∈ Yj , it is
obvious. Assume then that Yj ⊂ Act(X).
The group GX preserves the set {Z ∈ Yj , YjK0 (X, Z) = ∅} for any K0
hence for K0 = (K − κ)/2 ≥ Θ. Consider Za , Zb in this set, we claim
that YjK (Za , Zb ) is empty. Assume Y ∈ YjK (Za , Zb ). Since Y ∈ Act(X) we
∢
∢
can consider d∢
Y (Za , X) and dY (Zb , X). By triangle inequality, dY (Za , X) +
∢
∢
dY (Zb , X) ≥ dY (Za , Zb ) − κ ≥ K − κ. Thus, one of them needs to be larger
than (K − κ)/2 hence Y is either in YK0 (X, Za ) or in YK0 (X, Zb ), and this
is a contradiction to our assumption.
10
Proposition 1.7. (Induced orders)
Consider X, Z ∈ Y∗ , with Z ∈ Act(X). Assume that ΓX , ΓZ are subgroups of GX , GZ with proper isotropy.
For all i ≤ m, for all M ≥ Θ + 12κ, the set YiM (X, Z) is finite, and
˙ that is given by the order of YiM −4κ (γX (X i ), γZ Z i ),
carries a partial order <,
for arbitrary X i , Z i , in Yi , and almost all γX ∈ ΓX and γZ ∈ ΓZ .
Proof. Let us first check that the set is finite. We may assume that there are
X i ∈ Act(X) ∩ Yi and Z i ∈ Act(Z) ∩ Yi , otherwise YiM (X, Z) is empty. By
Lemma 1.4, there exists γX ∈ ΓX , γZ ∈ ΓZ such that each Y ∈ YiM (X, Z)
is in either one of the four sets YiM −3κ (ηX X i , ηZ Z i ) for ηX ∈ {1, γX } and
ηZ ∈ {1, γZ }. The union of these four sets is finite by properness axiom.
We now need to check that the order on YiM −4κ (γX (X i ), γZ Z i ) includes
all YiM (X, Z) and does not depend on the choice of the points X i , Z i . By the
Lemma 1.5, for arbitrary choice of points, and for any Y ∈ YiM (X, Z), there
is a finite set of ΓX and of ΓZ such that for all elements γX , γZ outside
these finite sets, Y ∈ YiM −4κ (γX X i , γZ Z i ) (the finite sets depend on the
choice of X i , Z i though). Since YiM (X, Z) is finite, we may find a finite
set of ΓX and ΓY suitable for all of them. Thus, for almost all γX , γZ , all
YiM −4κ (γX (X i ), γZ (Z i )) is ordered, and the order, once chosen the points
X i , Z i , does not depend on γX , γZ .
Assume that for two different choices of points X i , Z i , namely (Xai , Zai )
˙ a Y2 for
and (Xbi , Zbi ), the orders are different, and take Y1 , Y2 such that Y1 <
˙ b Y1 for the other.
the first order, and Y2 <
˙ a Y2 means that dY1 (Y2 , γZ (Zai )) ≤ κ. By the orbit estimate, d∢
Y1 <
Y1 (Y2 , Z) ≤
5κ for suitable γZ .
˙ b Y1 means that dY1 (Y2 , γX (Xbi )) ≤ κ, and by the orbit estimate,
Y2 <
∢
dY1 (Y2 , X) ≤ 5κ. Finally, by coarse triangular inequality, d∢
Y1 (Z, X) ≤ 11κ,
i
contradicting the assumption that Y1 is in YM (X, Z).
1.3
Convexity
Definition 1.8. (Convexity)
Let L > 10κ. We say that a subset W ⊂ Y∗ is L-convex if: for all i, for
all X, Z ∈ W ∩ Yi , for all j, the set YjL (X, Z) is a subset of W.
Let now L = (L(1), . . . , L(m)) be a m-tuple of positive numbers. A
subset W of Y∗ is said L-convex if for all X, Z ∈ W, of same coordinate
i(X) = i(Z), and for all j, the set YjL(j) (X, Z) is a subset W.
11
Note that being L-convex, for L > 0 is equivalent to being (L, . . . , L)convex.
Definition 1.9. Let W ⊂ Y∗ non-empty, and R ∈ Y∗ \W for which Act(R)∩
W is non-empty. Let L ≥ 10κ. Define YL (W, R) as the set of Y ∈ Y∗
satisfying the following.
• Y ∈ Act(R)
• Y ∈
/W
• W ∩ Act(R) ∩ Act(Y ) is non-empty, and for all X ∈ W ∩ Act(R) ∩
i(Y )
Act(Y ), one has Y ∈ YL (X, R).
Proposition 1.10. If L ≥ Θ + 12κ, then for all R for which it is defined,
the set YL (W, R) is finite.
S T
Proof. From the definition, YL (W, R) ⊂ i X∈Act(R)∩W (YiL (X, R) ∪ (Yi \
Act(X))). By finite filling assumption on the projection system, there is a
finite collection of elements Xj ∈ W ∩ Act(R) such that ∪j Act(Xj ) covers
∪W∩Act(R) Act(X).
In particular, YL (W, R) is inside a finite union of sets of the form
i
YL (Xj , R) which are finite by Proposition 1.7.
Proposition 1.11. Assume that for all X ∈ W, W is invariant by an
infinite group ΓX of rotations around X, with proper isotropy. Let L ≥
Θ + 12κ.
If W is (L−6κ)-convex, and if S ∈ YL (W, R) then YL (W, S) ⊂ YL−2κ (W, R).
Moreover, if W ′ contains W, then YL (W ′ , R) ⊂ YL (W, R).
Proof. Let Y ∈ YL (W, S) in coordinate i. There exists X ∈ W ∩ Act(Y ) ∩
Act(S) such that d∢
Y (X, S) ≥ L.
Assume that X̃ ∈ Act(R) ∩ Act(Y ) ∩ W. If it is not in Act(S), then
∢
d∢
Y (X̃, S) < κ and dY (X̃, X) > L − 2κ. Transfering X̃ in the coordinate of
X (by invariance under ΓX̃ ), one has d∢
Y (X̃i(X) , X) > L − 6κ. By convexity,
Y ∈ W though we assumed otherwise. Therefore, X̃ ∈ Act(S). Therefore,
∢
by definition of YL (W, S), one has d∢
Y (X̃, S) ≥ L, but also dS (X̃, R) ≥ L. It
follows by transitivity of betweenness (Lemma 1.3) that d∢
Y (X̃, R) ≥ L − 2κ.
The second assertion is a direct consequence of the definition.
Proposition 1.12. If Act(R)∩ W is not empty, for all L ≥ (2m + 12)κ+ Θ,
there exists Z ∈ YL (W, R) such that YL−2mκ (W, Z) = ∅.
12
Proof. Let us say that R has k L-links to W if {i, YL (R, W) ∩ Yi 6= ∅} has
k elements.
For any such index i, take a minimal item Zi in YL (R, W) ∩ Yi for
the order of Proposition 1.7. Then, by Proposition 1.11, YL−2κ (W, Zi ) is
included in YL (R, W), thus Zi has at most (k − 1) (L − 2κ)-links to W.
Iterating this choice at most m times, we find an element Z that has no
(L − 2mκ)-links to W. Therefore YL−2mκ (W, Z) = ∅.
Proposition 1.13. Let L ≥ Θ + 12κ. Consider W, and assume it is Lconvex, and that for all X ∈ W, there is ΓX < GX , infinite, that leaves W
invariant and that has proper isotropy.
If YL′ (W, R) is well defined and empty, then W ∪ {R} is (L + L′ + 5κ)convex.
Proof. If W ∩ Yi(R) is empty, there is nothing to prove. We assume it is
non-empty. Consider Y ∈ YL+L′ +5κ (R, X) for some X ∈ W ∩ Yi(R) , and
assume that Y ∈
/ W. Notice that Y ∈ Act(R) though, and X ∈ Act(R)
since they have same coordinate. Hence, X ∈ W ∩ Act(R) ∩ Act(Y ).
Let X ′ be any other element of W ∩ Act(R) ∩ Act(Y ). Transfert X ′ in
the coordinate i = i(R), inside W, by ΓX ′ . There exists Xi′ ∈ Yi ∩ W such
∢
′
′
′
′
that d∢
Y (X , Xi ) ≤ κ. But, W being L-convex, one has dY (X , Xi ) ≤ L. It
′
′
follows by triangular inequality, that d∢
Y (R, X ) ≥ L + 2κ. Since this is true
′
for all X as above, it follows that Y ∈ YL′ +2κ (W, R), contradicting our
assumption.
2
Composite rotating families and windmills
We proceed to adapt the rotating families study of [DGO] to the context of
composite projection systems.
2.1
Definition
Definition 2.1. (Composite rotating family)
A composite rotating family on a composite projection system, endowed
with an action of a group G by isomorphisms, is a family of subgroups
ΓY , Y ∈ Y∗ such that
• for all X ∈ Y∗ , ΓX < GX = StabG (X), is an infinite group of rotation
around X
13
• for all g ∈ G, and all X ∈ Y∗ , one has ΓgX = gΓX g−1
• if X ∈
/ Act(Z) then ΓX and ΓZ commute,
• for all i, for all X, Y, Z ∈ Yi , if dY (X, Z) ≤ ΘP then for all g ∈
ΓY \ {1}, dY (X, gZ) ≥ ΘRot .
We will show the following.
Theorem 2.2. Consider Y∗ a composite projection system. If {ΓY , Y ∈ Y∗ }
is a composite rotating family for sufficiently large ΘRot , then the group ΓRot
generated by the union of all ΓY (for Y ∈ Y∗ ) has a partially commutative
presentation.
More precisely, there
S exists S ⊂ Y∗ , and R a Sset of words of the form
′
−1
[s, ws w ] with w ∈ h S ΓS i such that ΓRot ≃ h S ΓS | Ri and such that
[s, ws′ w−1 ] ∈ R if and only if s ∈
/ Act(ws′ ).
Moreover, for all γ ∈ ΓRot \ {1}, there is i(γ) ≤ m, and R ∈ Yi(γ) such
that for all X ∈ Yi(γ) , dR (X, γX) > ΘRot − 2ΘP − κ, and there is γs ∈ ΓR
such that dR (X, γs γX) ≤ 2ΘP + 3κ (the pair (R, γs ) is called a shortening
pair for γ).
The last property is, in our point of view, an incarnation of the Greendlinger
lemma, from the small cancellation theories. If one consider a relation γ of
the quotient group, one can find in it a large part of a defining relation γs .
Compare to [DGO, §5.1.3]
A major tool for analysing rotating families was the concept of windmills.
We are going to use composite windmills.
Let us fix L the m-tuple
L = (c∗ + 20(m − 1)κ, c∗ + 20(m − 2)κ, . . . , c∗ + 20κ, c∗ ).
Let σ be the cyclic shift on Z/mZ: σ(i) = (i − 1), and define Lj = σ j−1 (L)
obtained by shifting the coordinates of the m-tuple.
Thus Li reaches its maximum c∗ + 20(m − 1)κ on the coordinate i,
minimal value c∗ at i − 1. Note that the maximum of L is less than ΘP − κ.
Definition 2.3. (Composite windmills)
A composite windmill is a collection (W1 , . . . , Wm , GW , j0 ) in which
• S
GW is the subgroup of G generated by a set of subgroups {ΓY , Y ∈
i∈I∗ Wi } for I∗ either {1, . . . , m} or {1, . . . , m} \ {j0 },
• Wi is a subset of Yi for all i, invariant under GW ,
14
• 1 ≤ j0 ≤ m,
S
• i Wi is Lj0 -convex.
• The group GW has a partially commutative presentation, that is a
presentation of the form
G ≃ hS | Ri
where S is the union over a subset W∗ of W of generating sets for
ΓX , X ∈ W∗ , and R consists of words over the alphabet S ∪ S −1 of the
form [s, ws′ w−1 ] for w a word over S ∪ S −1 . Moreover, if X, X ′ ∈ W∗
and s ∈ ΓX , s′ ∈ ΓX ′ , the word [s, ws′ w−1 ] is in R if and only if
wX ′ ∈
/ Act(X).
• (Greendlinger property) for each γ ∈ GW \ {1} there is i(γ) ≤ m, and
R ∈ Wi(γ) such that for all X ∈ Yi(γ) , dR (X, γX) > ΘRot − 2ΘP − κ.
Moreover, there is γs ∈ ΓR such that dR (X, γs γX) ≤ 2ΘP + 3κ (the
pair (R, γs ) is called a shortening pair for γ).
We say that the composite
S windmill has full group if GW is the subgroup
of G generated by {ΓY , Y ∈ m
i=1 Wi }.
If we do not mention it, our windmills will be full. Only in specific
circumstances do we need non-full windmills. Indeed, we will use the case
of a non-full group only at most one time by coordinate, when initiating the
process in each coordinate.
Since K > max(L), for each i, Wi is connected in PK (Yi ).
We say that a windmill W ′ (with its representative set W∗′ used for the
presentation of the definition) is constructed over W if W ⊂ W ′ and if the
set of representatives W∗′ contains the set of representatives W∗ .
2.2
Osculations of two kinds
• An osculator of type gap of a composite windmill (W1 , . . . , Wm , GW , j0 )
is an element R of Yj0 \ Wj0 such that there exists i ≤ m, Xi , Zi ∈ Wi ,
c∗
that are in Act(R) and such that d∢
R (X, Z) > 2 − 20κ.
• An osculator of type neighbor of a composite windmill (W1 , . . . , Wm , GW , j0 )
is an element R of Yj0 \ Wj0 such that Y c∗ (W, R) = ∅.
2
Lemma 2.4. Consider a composite windmill W = (W1 , . . . , Wm , GW , j0 ),
assume that Wj0 6= ∅, and let R ∈ Yj0 be an osculator of type gap.
Let Y ∈ Yi in Act(R). Then there exists X ∈ Wj0 such that d∢
Y (X, R) ≤
κ.
15
Proof. If R is an osculator of type gap, there are X ′ , Z ′ ∈ Wi , for some i,
′
′
such that d∢
R1 (X , Z ) > c∗ /2 − 20κ.
Let X0 ∈ Wj0 , and consider its orbit under the groups ΓX ′ , and ΓZ ′ ,
which preserves Wj0 . We may use Lemma 1.5 to find X ′(j0 ) , Z ′(j0 ) in these
′(j0 ) , Z ′(j0 ) ) > c /2 − 24κ.
orbits, hence in Wj0 , such that d∢
∗
R1 (X
By the coarse triangle inequality, for at least one point among X ′(j0 ) , Z ′(j0 ) ,
′(j0 ) ) > c /4 − 13κ. Behrstock inequality gives
say X ′(j0 ) , we have d∢
∗
R1 (Y, X
′(j0 ) ) ≤ κ.
(R
,
X
d∢
1
Y
Lemma 2.5. Let W be a composite windmill, and R1 , R3 be two osculators
of W. Assume Wj0 6= ∅, and let X2 ∈ Wj0 .
If R3 is of type neighbor and W is ( c2∗ −20κ)-convex, then dR1 (X2 , R3 ) ≤
c∗ .
If R3 is of type gap, then dR1 (X2 , R3 ) ≤ ΘP .
Proof. If R3 is an osculator of neighbor type, then the result follows from
Proposition 1.13.
If now R3 is an osculator of type gap, the proof is slightly more involved.
There is i, and there are X, Z ∈ Wi such that d∢
R3 (X, Z) > c∗ /2 − 20κ.
Since Wj0 is non-empty, and invariant for ΓX and ΓZ , we can apply Lemma 1.5 and find X (j0 ) , Z (j0 ) ∈ Wj0 such that dR3 (X (j0 ) , Z (j0 ) ) ≥
d∢
R3 (X, Z) − 4κ which is ≥ c∗ /2 − 24κ. By coarse triangular inequality, at
least one of the quantities dR3 (R1 , X (j0 ) ) and dR3 (R2 , Z (j0 ) ) is greater than
c∗ /4 − 13κ. Say it is dR3 (R1 , X (j0 ) ). Behrstock inequality then gives that
dR1 (R3 , X (j0 ) ) ≤ κ, and again coarse triangular inequality gives dR1 (X (j0 ) , X2 ) ≥
dR1 (X2 , R3 ) − κ. Since the first is bounded by the maximal convexity constant of W, the result follows.
2.3
The unfolding in the different coordinates
Given a composite windmill W, we define its unfolding as follows.
An admissible set R is one of the three following possibilities.
Consider Rgap the set of osculators of type gap in Yj0 . If it is non-empty,
we say R = Rgap is admissible, and is the only possible admissible set of
osculators.
If Rgap is empty, and if W is ( c2∗ − 20κ)-convex, then R = {GW R} for
any choice R of osculator (necessarily of type neighbor) is admissible. There
might be several admissible sets, and also there might be none.
16
If Rgap is empty, and if W is not c2∗ -convex, then the set R of admissible
osculators in this coordinate j0 is the empty set R = ∅.
Here is an obvious lemma.
Lemma 2.6. Let R be a choice of an admissible set of osculators of W. If
R is empty, then W ′ = (W1 , . . . , Wm , GW , j0 + 1) is a composite windmill.
We thus concentrate on the case where R is non-empty.
Having made such a choice of set of admissible osculators R =
6 ∅, we
then define, for all i, Wi′ to be the union[
of all the images of Wi by elements
of the group GW ′ generated by GW ∪ { ΓR }. The unfolding of W is then
Ra
′ , G ′ , j + 1).
(W1′ , . . . , Wm
0
W
In the case Wj0 is empty, we include here a convexity result for an
intermediate step in the construction: adding an admissible set of osculators
R, which produces a non-full composite windmill.
Lemma 2.7. Assume that W is a full composite windmill of principal coordinate j0 , with Wj0 = ∅.
Let Wjs0 be a set R of admissible osculators as defined above, assumed
non-empty.
For all other coordinates, let Wis = Wi .
s , G , j ) is a non-full composite windmill
Then W s = (W1s , W2s , . . . , Wm
W 0
of principal coordinate j0 . If moreover R is the orbit of a neighbor osculator,
and if W is ( c2∗ − 20κ)-convex, then W s is B-convex, for B = c2∗ + 10κ ≤
inf L.
Proof. If R = ∅, there is nothing to prove. Consider the case of the orbit
of a neighbor osculator. It suffices to check that Wjs0 (= GW R) is convex in
the sense that for all γ ∈ GW and all i the set YiB (R, γR) is in Wi .
By the Greendlinger Property, given γ, there exists j, and Yj ∈ Wj such
that dYj (R, γR) > ΘRot − 2ΘP − κ, or R = γR (if R is not active for all the
shortening pairs of γ).
Of course we consider only the first case of the alternative.
(i)
Assume that some Y ∈ Yi is in YB (R, γR).
If Y ∈
/ Act(Yi ), then one can use a shortening pair at Yi to reduce the
length of γ in its principal tree, and this shortening pair gives γ ′ such that
(i)
∢
′
′
d∢
Y (R, γR) = dY (R, γ R). Thus, Y ∈ YB (R, γ R) as well, and by performing
this reduction sufficiently many times, we may assume that Y ∈
/ Act(Yi ).
By Lemma 1.4, either R or γR approximates by κ the projection of Yj
on Y .
17
/ Wi , one has d∢
Say that d∢
Y (Yj , R) ≤
Y (γR, Yj ) ≤ κ. By osculation if Y ∈
∢
∢
(R,
Y
)
+
κ
≤ c2∗ + 2κ
Therefore, one has dY (γR, R) ≤ dY (γR, Yj ) + d∢
j
Y
which is less than B.
∢
∢
If now d∢
Y (R, Yj ) ≤ κ, one has dY (R, γR) is within 2κ from dY (Yj , γR),
∢
−1
−1
which equals dγ −1 Y (γ Yj , R). Of course, Y ∈
/ Wi if and only if γ Y ∈
/ Wi ,
c∗
∢
∢
hence, if it is the case, by osculation of R, dY (Yj , γR) ≤ 2 , and dY (γR, R) ≤
c∗
2 + 2κ ≤ B.
In the case where R is the set of gap osculators, the proof is similar.
Indeed, if R1 is a gap between X1 and Z1 , and R2 is a gap between X2 and
Z2 , and if Y is between R1 and R2 , so that d∢
Y (R1 , R2 ) ≥ c∗ + 20(m − 1)κ(=
Lj0 (j0 )), then Y is also between X1 (or Y1 ) and X2 (or Y2 ) so that, say,
d∢
Y (X1 , X2 ) ≥ c∗ + 20(m − 1)κ − 3κ. One can transfert X2 in the coordinate
of X1 by Lemma 1.4, in W (in the ΓX2 -orbit of X1 ). The convexity of W
then shows that Y ∈ W.
c∗
2.
The aim of the next sections is to prove the following.
Proposition 2.8. If the set R of admissible osculators is non-empty, then
W ′ = (W1′ , . . . , Wm , GW ′ , j0 + 1) is a (full) composite windmill, and W∗′ can
be chosen to contain W∗ (in other words, W ′ is constructed over W).
2.3.1
Unfolding a tree
Proposition 2.9. (Principal coordinate tree)
Consider a full composite windmill W, of principal coordinate j0 .
Let R 6= ∅ be an admissible set of osculators as defined in the previous
section. If Wj0 = ∅, let Wjs0 = R, and otherwise let Wjs0 = Wj0 .
There exists a GW ′ -tree T , bipartite, with black and white vertices, with
an equivariant injective map ψ : T → P(Yj0 ) (the set of subsets of Yj0 ) that
sends black vertices on images of osculators by GW ′ , and white vertices on
images of Wjs0 by GW ′ , and that sent the neighbors (in T ) of the preimage
of Wjs0 on R.
Moreover, for any pair of distinct white vertices w1 , w2 , and any black
vertex v in the interval between them (in T ), and any X1 ∈ ψ(w1 ), X2 ∈
ψ(w2 ), one has dψ(v) (X1 , X2 ) ≥ ΘRot − 2ΘP − κ.
Finally, if w1 , w2 are white vertices for which the path from a black vertex
v starts by the same edge, then for any X1 ∈ ψ(w1 ), X2 ∈ ψ(w2 ), one has
dψ(v) (X1 , X2 ) ≤ 2ΘP + 3κ.
18
t
Proof. Take a transversal Rt of R under the action
S of GW . For each R ∈ R ,
let (GW )R the subgroup of GW generated by X∈W\Act(R) ΓX .
Set T to be the Bass-Serre tree of the (abstract) graph of groups whose
vertex groups are GW and the groups ΓR × (GW )R , R ∈ Rt , and the edges
are the pairs (GW , R), R ∈ Rt , and the edge groups are the groups (GW )R .
g
Let G
W ′ the fundamental group of this graph of groups. The group GW ′
is a quotient of this group, since it is generated by GW and the stabilizers
of elements R of Rt , which, by assumption (Definition 2.1), are direct sum
of their rotation group with the groups (GW )R .
g
T is a tree, endowed with a G
W ′ -action, bipartite, and with an equivg
ariant (with respect to GW ′ ։ GW ′ ) map ψ : T → P(Yj ) that sends black
vertices on images of elements of R by GW ′ , and white vertices on images
of Wjs0 by GW ′ .
We need to show that it is injective, and at the same time, we will show
the estimate of the end of the statement.
Consider a path p of T , starting and ending at white vertex. Up to cyclic
permutation, and up to the group action, we may assume that the path p
starts at the vertex fixed by GW , and its second vertex is fixed by some
R1 ∈ Rt , and that its length is even.
Let us denote by p0 , p1 , . . . , pN the consecutive vertices of p, and let X2i
be a choice of a element of ψ(p2i ), and R2i+1 = ψ(p2i+1 ).
The monotonicity property in the coordinate j0 says that if dY (X, Z) ≥
Θ then dW (X, Z) ≥ dW (X, Y ).
We will use it in an induction to establish that for all k odd, and all i in
1 ≤ i ≤ N 2−k and all j in 1 ≤ j ≤ k−1
2 , one has
dRk (Rk−2j , Rk+2i )
≥ Θrot − 2ΘP − κ
∀Xs ∈ ψ(ps ), dRk (Xk−2j+1 , Xk+2i−1 ) ≥ Θrot − 2ΘP − κ
The case i, j = 1 happens as follows. Choose k.
We first show how a black vertex separates two adjacent white vertices.
′
Note that there is Xk+1
∈ ψ(pk+1 ) that equals gXk−1 for some g ∈ ΓRk \{0}.
s
By convexity of Wj0 (ensured by assumption, or by Lemma 2.7 in case Wj0
′
) ≤ ΘP . And by assumption on the rotating
is empty), dRk (Xk+1 , Xk+1
′
groups, dRk (Xk−1 , Xk+1 ) ≥ ΘRot . Thus, dRk (Xk−1 , Xk+1 ) ≥ ΘRot − ΘP − κ,
the second inequality.
By Lemma 2.5, dRk (Xk+1 , Rk+2 ) ≤ ΘP and dRk (Xk−1 , Rk−2 ) ≤ ΘP . By
triangle inequality, we get dRk (Rk−2 , Rk+2 ) ≥ Θrot − 2ΘP − κ. We have both
inequalities.
19
Assume that the inequalities are proven for all (i, j) such that i + j ≤ i0
(and for all k), and let us choose k and (i, j) with i + j ≤ i0 , and prove the
inequality for for (i + 1, j).
Set Y = Rk+2i , and and W = Rk . In the following we set either Z =
R2i+k+2 or X2i+k+1 , and either X = Rk−2j or X = Xk−2j+1 .
By the inductive assumption for k′ = k + 2i, i′ = 1, j ′ = i, one has
dY (W, Z) ≥ Θrot − 2ΘP − κ.
Also for k, i and j the induction gives dW (Y, X) ≥ Θrot − 2ΘP − κ.
Behrstock inequality then provides dY (W, X) ≤ κ and therefore dY (X, Z) ≥
Θrot − 2ΘP − 3κ. This is still far above Θ. One thus may apply the monotonicity property and obtain dW (X, Z) ≥ dW (X, Y ). In other words,
dRk (Rk−2j , Rk+2i+2 ) ≥ Θrot − 2ΘP − κ.
The inequality is also proven for (i, j + 1) in the same manner, symmetrically. This finishes the induction.
In the end, we have obtained for i = N/2−1, and k = 1, dR1 (X0 , RN −1 ) ≥
ΘRot − ΘP , and it follows that dR1 (X0 , XN ) ≥ ΘRot − 2 × ΘP − κ, which is
the estimate of the statement.
If we assume that p is mapped to a loop, Wj0 contains both X0 and XN ,
and not R1 (it is an osculator), the convexity of Wj0 imposes ΘRot − 2 ×
ΘP − κ ≤ ΘP , meaning ΘRot ≤ ΘP + κ. and this contradicts our choice of
ΘRot .
It also follows from this analysis that if w1 , w2 are white vertices of T
and v is a black vertex between then, then dψ(v) (X1 , X2 ) ≥ ΘRot − 2ΘP − κ
(in our induction above). A final use of Behrstock inequality provides that
whenever the paths from v to a white vertex w1 has more than three edges,
then if v ′ is the first black vertex after v on this path, and if X1 ∈ ψ(w1 ),
then dψ(v) (X1 , ψ(v ′ )) ≤ κ. It follows from that and Lemma 2.5 that if
w2 is another white vertex w1 whose path from v starts at the same edge,
dψ(v) (X1 , ψ(v ′ )) ≤ 2ΘP + 3κ.
The former proposition allows to define, for each element γ of GW ′ ,
its principal coordinate, and its principal tree. Indeed, if γ ∈ GW ′ not
conjugated to GW , the proposition shows that it is either loxodromic or
the stabilizer of a black vertex on the tree T . Then we define its principal
coordinate as j0 and its principal tree as T . If it is in GW , or conjugate
in it, its principal coordinate and its principal tree are defined inductively,
according to the process of unfoldings of composite windmills.
20
2.3.2
Preservation of convexity
Proposition 2.10. (Convexity of W ′ )
Let W = (W1 , . . . , Wm , GW , j0 ) be a composite windmill (possibly nonfull).
Assume that R is an admissible set of osculators, and consider the sets
′ ).
Wi′ defined above. Let W ′ = (W1′ , . . . Wm
If R consists of the orbit of a neighbor, then W ′ is c∗ -convex.
If R consists of gap osculators, then W ′ is Lj0 +1 -convex.
The case of R = ∅ is trivial, so we assume it is not empty.
If R consists of the orbit of a neighbor, let Aj = c∗ for all j. If R consists
of gaps, let Aj = Lj0 (j) + 20κ (which is less than Lj0 (j + 1)).
Let X, Z ∈ Wi′ , consider Y ∈ YjA(j) (X, Z).
Here is our main claim.
We will show that Y is a GW ′ -translate of one of the following type of
elements:
• Y ′ for which there exists Xf , Zf ∈ Wj0 such that d∢
Y ′ (Xf , Zf ) ≥ A(j)−
10κ;
• Y ′ for which there exists Xf ∈ Wj0 , and R an osculator of W in Wj′ 0
such that d∢
Y ′ (Xf , R) ≥ A(j) − 10κ;
• Y ′ for which there exists R1 , R2 osculators of W in Wj′ 0 such that
d∢
Y ′ (R1 , R2 ) ≥ A(j) − 10κ
We will then finish the proof with this claim established, but before that
we will prove the claim.
Transfert of X and Z to Yj0 . In W ′ , the groups ΓX and ΓZ preserve
′
Wj0 which is not empty (it contains R). Therefore, by Lemma 1.5 there are
(j0 ) , Z (j0 ) ) ≥ A(j) − 4κ.
X (j0 ) , Z (j0 ) in Wj′ 0 such that d∢
Y (X
The interval in T . Taking ψ −1 of X (j0 ) and of Z (j0 ) produce two vertices
in the principal coordinate tree T of Proposition 2.9. More precisely, either
one of X (j0 ) , Z (j0 ) is the image of a black vertex of T , or in the image of
a white vertex of T . This thus give two vertices of T that we (slightly
abusively) denote by ψ −1 (X (j0 ) ), ψ −1 (Z (j0 ) ).
If these vertices are adjacent, we have achieved the second point of the
claim. If these vertices are the same, we have achieved the first point of the
claim. If these vertices are different, both black with only one white vertex
in the interval, we have achieved the third point of the claim.
21
Thus, we may assume that there is at least one black vertex of T in the
open interval (ψ −1 (X (j0 ) ), ψ −1 (Z (j0 ) )). Let R1 , . . . , RN the images by ψ of
these black vertices, in order starting from the side of ψ −1 (X (j0 ) ).
By Proposition 2.9, we have for all i, dRi (X (j0 ) , Z (j0 ) ) > ΘRot − 2ΘP − κ,
which is > 50κ.
Reduction to the case where Ri ∈ Act(Y )
If Y is equal to one of the Ri then we fall in the first possibility of the
main claim. Thus, let us assume that Y is different from all the Ri .
We may assume that Y is in Act(Ri ) for all i. Indeed if it was not,
one could use an element of ΓRi to reduce the length of the path p, without
(j0 ) , Z (j0 ) ) since Γ
changing the value of the projection distance d∢
Ri leaves
Y (X
π
dY invariant.
Transfert of Y in Yj0 . We may apply Lemma 1.4 again, and find an
element Y (j0 ) in Yj0 (far in an orbit of ΓY ) such that, for all i, one has
(j0 ) ) ≤ 4κ.
d∢
Ri (Y, Y
Position of Y (j0 ) in the order. Fix 0 < i ≤ N . Since dRi (X (j0 ) , Z (j0 ) ) >
50κ, either dRi (X (j0 ) , Y (j0 ) ) or dRi (Y (j0 ) , Z (j0 ) ) is larger than 24κ.
All Ri are in Y50κ (X (j0 ) , Y (j0 ) ) therefore they satisfy the order property
in this set, which coincide with the ordering of their indices. By this order
property and Behrstock inequality, if for some i one has dRi (Y (j0 ) , X (j0 ) ) >
5κ, then for all i′ < i, one still has dRi (Y (j0 ) , X (j0 ) ) > 5κ. Similarily if
dRi (Y (j0 ) , Z (j0 ) ) > 5κ then for all greater i′′ the same holds.
Therefore we have three cases.
Either dR1 (Y (j0 ) , X (j0 ) ) ≤ 5κ or dRN (Y (j0 ) , Z (j0 ) ) ≤ 5κ, or there exists
i ≥ 1, largest such that dR1 (Y (j0 ) , X (j0 ) ) > 5κ and i < N .
By symmetry, and translation by an element of GW ′ the first and second
case have same resolution. Let us treat the first one. By triangle inequality,
dR1 (Z (j0 ) , Y (j0 ) ) > ΘRot − 10κ − 2ΘP which is still greater than 20κ.
(j0 ) , Y ) > 16κ. By Behrstock inequality, d∢ (Z (j0 ) , R ) <
Going back to Y : d∢
1
Y
R1 (Z
(j
)
∢
0
κ, and finally by triangle inequality, dY (X , R1 ) ≥ A(j) − 2κ. We are in
the second point of the claim if X (j0 ) is in a white vertex, and in the third
point if it is a black vertex.
We thus turn to the case where there exists i ≥ 1, largest such that
dR1 (Y (j0 ) , X (j0 ) ) > 5κ and i < N .
One has
dRi+1 (Y (j0 ) , Z (j0 ) ) > ΘRot − 2ΘP − 10κ
(j0 ) )
d∢
> ΘRot − 2ΘP − 14κ
Ri+1 (Y, Z
∢
)
(j
0
dY (Ri+1 , Z )
≤ κ
22
and
dRi (Y (j0 ) , X (j0 ) ) ≥ 5κ
(j0 ) )
≥ κ
d∢
Ri (Y, X
∢
(j
)
0
dY (Ri , X )
≤ κ
So, d∢
Y (Ri , Ri+1 ) ≥ A(j) − 4κ. We have the third point of the claim, and
the claim is established.
We need to finish the proof of the lemma. There are several cases to
treat. The easiest is when the first case of the claim occurs.
In that case, if j = j0 , Y ′ is actually a gap osculator, hence in Wj′ 0 . If
j 6= j0 , by convexity of W, it is in Wj .
Assume now that the second case occurs.
If R is of type neighbor, it simply contradicts Proposition 1.13.
If R is an osculator of type gap between X0 , X1 , and j = j0 , one easily
gets that R is an osculator of type gap between Xf and either X0 or X1
(any one for which dR (Y ′ , Xǫ ) is larger than κ, and by triangular inequality,
there must be at least one). If j 6= j0 , we may use the same argument.
′
Y ′ ∈ Act(R) therefore d∢
R (Y , Xǫ ) is larger than κ for either ǫ = 0 or 1. Then,
∢
d∢
Y ′ (R, Xǫ ) < κ and by triangular inequality, dY ′ (Xf , Xǫ ) ≥ A(j) − 12κ(>
′
Lj0 (j)). It follows by convexity of W that Y ∈ Wj .
Finally, assume that the third case occurs.
Assume that R2 is an osculator of type gap, between X0 , X1 . Then,
again with the same reasoning, Y ′ ∈ Act(R2 ) and there is ǫ for which it is
∢
in Act(Xǫ ) and d∢
Y ′ (R2 , Xǫ ) is less than κ. Thus dY ′ (R1 , Xǫ ) ≥ A(j) − 12κ,
and we are back to the case 2 of the claim, with a slightly lower constant.
The proof goes nevertheless through, and the desired conclusion holds.
Finally, assume that R2 is of type neighbor. Then both R1 , R2 are of
type neighbor, and R2 = γR1 for some γ ∈ ΓW . Let us rename R1 = R,
call i = i(Y ′ ), and j the principal coordinate of γ (for the Greendlinger
property). Let Z ∈ Wj be the vertex of a shortening pair for γ for which
Z ∈ Act(Y ′ ) ∩ Act(R) (there exists one, otherwise one can reduce the lenght
of γ in its principal tree by a shortening pair at Z). Thus, d∢
Z (RγR) >
ΘRot − 2ΘP − 2κ.
Suppose d∢
Y ′ (R, γR) > c∗ − 10κ. Then, there are two possible cases.
c∗
∢
Either dY ′ (R, Z) > c2∗ − 6κ or d∢
Y ′ (γR, Z) > 2 − 6κ (or both).
∢
∢
′
In the first case, dZ (R, Y ) ≤ κ. Thus dZ (Y ′ , γR) > κ, and so d∢
Y ′ (γR, Z) <
κ.
Recall that Z ∈ Act(R) ∩ Act(Y ′ ). Thus d∢
Y ′ (Z, R) > c∗ − 2κ, and
′
Y ∈ Yc∗ −2κ (Z, R). Now let Z ′ any other element of W in Act(R) ∩ Act(Y ′ ).
c∗
′
By ( c2∗ − 20κ)-convexity of W, one has d∢
Y ′ (Z, Z ) ≤ 2 − 20κ and therefore
23
Y ′ ∈ Yc∗ −2κ− c2∗ +21κ (Z ′ , R). In other words, Y ′ ∈ Y c2∗ +19κ (W, R) and this
contradicts the fact that R is a neighbor.
In the second case, the situation is similar after composing by the automorphism γ −1 .
2.3.3
The unfolding is a windmill
Proposition 2.11. If W = (W1 , . . . , Wm , GW , j0 ) is a composite windmill,
′ , G ′ , j + 1) is defined as above (where j + 1 is
and if W ′ = (W1′ , . . . , Wm
0
0
W
taken modulo m in {1, . . . , m}), then W ′ is a composite windmill.
Moreover, the set W∗′ of the fifth point of the definition can be assumed
to contain the set W∗ (in other words, W ′ is constructed over W).
The first three points follow by construction. The fourth point (convexity) is the result of Proposition 2.10. The sixth point is a consequence of
Proposition 2.9. The same proposition introduces an action of GW ′ on a
tree T which is Bass-Serre dual to a presentation of GW ′ as the fundamental
group of a graph of group, with one vertex v0 carrying the group GW and
the other vertices v[R] , [R] ∈ R/GW , adjacent to a single edge whose other
end is v0 , carrying the group ΓR × (GW )R , if R is a representative of the
orbit [R].
2.4
2.4.1
Towers of windmills, and accessibility
Starting point
We start the process by selecting W(0) to be a maximal collection of mutually inactive elements in Y∗ . Thus, whenever W(0)j 6= ∅, it is reduced to a
single point.
We choose j0 = 1. It is clear that W(0) defines a composite windmill
where for all i, W(0)i is either empty or a singleton, and where GW is the
direct product of the groups GX , for X ∈ W(0) (there are at most m direct
factors).
W(0) is κ-convex, and for all R, by maximality of W(0), Act(R)∩W(0) 6=
∅. Recall that by choice, c∗ > 25κ + 2Θ, hence by Proposition 1.12,
there exists a neighbor osculator in Y c2∗ +2mκ (W(0), R). We set a counter
N (0) = 0 ∈ N, which will count the number of times we resort to a neighbor
osculator.
24
2.4.2
The process
We need a process for choosing consistently among the possible neighbor
osculators. Recall that we assumed Y∗ to be countable. Let us choose once
and for all (Xn )n∈N , an infinitely redundent enumeration of elements of Y∗
(in the sense that every X ∈ Y∗ is equal to infinitely many Xn ).
We will define W(k) for k a countable ordinal (not necessarily a number).
We take the notation
W(k) = (W(k)1 , . . . , W(k)m , GW (k) , jk ).
The general process for defining W(k + 1) from W(k) is the following.
If in the coordinate jk , there is a gap osculator of W(k) in Yjk one defines
the set of these gap osculators as R(k) the set of admissible osculators at step
k, and one unfolds W(k) into W ′ (k) = (W ′ (k)1 , . . . , W ′ (k)m , GW ′ (k) , jk + 1)
(where (jk + 1) is taken modulo m, in {1, . . . , m}), which we call W(k + 1).
We set N (k + 1) = N (k).
If in the coordinate jk , there is no gap osculator and if W(k) is not
c∗
( 2 − 20κ)-convex, then W(k + 1) = (W(k)1 , . . . , W(k)m , GW (k) , jk + 1). We
set N (k + 1) = N (k).
If in the coordinate jk , there is no gap osculator and if W(k) is ( c2∗ −20κ)convex, then we consider any neighbor osculator R in Y c2∗ +2mκ (W(k), Xn ) ∪
{Xn } for n the minimal value above N (k) for which Xn ∈
/ W. We then set
N (k + 1) = N (k) + 1, R = GW R, then first
W(k + 1) = (W(k)1 , . . . , W(k)m , GW (k) , i(R))
and then and one unfolds W(k + 2) into W ′ (k + 1).
Lemma 2.12. For each k, one and only one (except for the freedom in the
choice of the neighbor) of these operations is possible.
Proof. It follows from the maximality of W(0), that for all R, Act(R) ∩
W(0) 6= ∅, hence also Act(R)∩W(0) 6= ∅. Thus, assuming W(k) is ( c2∗ −20κ)convex, one can find a neighbor osculator in Y c∗ +2mκ (W(k), Xn ) ∪ {Xn } if
2
Xn ∈
/ W(k).
Lemma 2.13. In all these cases, W(k + 1) is still a composite windmill,
constructed over W(k).
Proof. This follows from Proposition 2.8 if the set of osculators is non-empty,
and from Lemma 2.6 otherwise.
25
Let us Zorn things up. If for all number s, W(k + s) is not ( c2∗ −
20κ)-convex, then for countably many steps in the process, one never use a
neighbor. Let us define W(k + ω) the direct union of the W(k + s), s ∈ N. It
is a composite windmill, constructed over all W(k + s) for s any integer. By
Zorn Lemma, there is an ordinal η such that W(k + η) is ( c2∗ − 20κ)-convex.
2.4.3
Accessibility
Lemma 2.14. For all R ∈ Y∗ , R is eventually in W(η).
Proof. Indeed, assume it is never in Y∗ . Each time W(k) is ( c2∗ −20κ)-convex
(and by the previous observation, this happens infinitely many times), we
use a neighbor that is in Y c2∗ +2mκ (W(k), Xn )∪{Xn } for n the minimal value
above N (k) for which Xn ∈
/ W (hence an integer). Because the sequence
(Xn )N contains infinitely many times the element R (assumed to never be in
W(η)), we will pick an element of Y c2∗ +2mκ (W(k), R) infinitely many times.
However, Y c2∗ +2mκ (W(k), R) ⊂ Y c2∗ +2mκ (W(0), R) and has at least one element less each times it is used by Proposition 1.11. Since by Proposition
1.10, Y c∗ +2mκ (W(0), R) is finite, this is a contradiction.
2
2.5
End of the proof of the main Theorem 2.2
Applying the unfolding operation for each countable ordinal (while taking direct limits for each countable limit ordinal) allows to obtain a tower
of windmills W(k) (indexed by countable ordinals k) with representatives
W(k)∗ such that if k < k′ then W(k′ ) is constructed over W(k). Applying
Lemma 2.14 countably many times (for each element of Y∗ ), there exists
ktop that is countable and that is the smallest ordinal such that Y ⊂ W(k).
Considering GW (ktop ) , we obtain the structure theorem desired, which is
Theorem 2.2.
3
Conclusion, application to Dehn twists, and Theorem 1
Let Σ be an orientable closed surface of genus greater than 2. Consider
MCG(Σ) its Mapping Class Group.
Bestvina Bromberg and Fujiwara produced a finite coloring of the set of
simple closed curves of Σ such that two curves of same color intersect, and
a finite-index normal subgroup G0 of MCG(Σ) that preserves the coloring.
G0 is called the color preserving group. After refinement of the colors, we
26
actually may assume that the colors are in correspondance with the cosets
of G0 . We denote the colors by {1, . . . , m}.
Let c and c′ be a simple closed curves. If they intersect, the projection
of c′ on c is the family of elements in the arc complex of the annulus around
c (that is the cover of Σ associated to c) that come from lifts of c′ . They
are all disjoint. If c′′ is another simple closed curve intersecting c, dπc (c′ , c′′ )
is the diameter in the curve graph of the union of the projections of c′ and
c′′ on the annulus around c.
dπ defines a composite projection system on the set of all (homotopy
classes of) simple closed curves. Indeed, let Act(c) be the set of curves
intersecting c. Clearly dπc is symetric, and satisfies the separation. The
symetry in action, and the closeness in inaction are also direct consequences
of definitions. The finite filling property is a consequence that all sequences
of subsurfaces up to isotopy, increasing for inclusion, is eventually stationnary. dπc satisfies the triangle inequality since it is a diameter of projections,
and the Behrstock inequality [B], see also [Man] [Man2]. The properness is
ensured by [BBF, Lemma 5.3]
We can now define two composite projection systems with composite
rotating families. The first one is defined on Y∗ is the set S of all homotopy
classes of simple closed curves of Σ.
Let us define Yi to be the subset of this set of simple closed curve of
color i in the Bestvina-Bromberg-Fujiwara coloring, and Y∗ their union.
It is, as we just said, a composite projection system on which G0 acts by
automorphisms.
Performing the construction of [BBF] and the choices as after Definition
1.2, we have constants Θ, κ, c∗ , ΘP , ΘRot .
We select N1 such that all N1 -powers of Dehn twists in MCG(Σ) are in
G0 . This is possible since there are only finitely many MCG(Σ)-orbits of
simple closed curves in Σ, and G0 has finite index. Then we select N2 a
multiple of N1 such that for all simple closed curve c, the Dehn twist τcN2
around c satisfies that dc (c′ , τ N2 c′ ) > ΘRot + 2ΘP if c′ is a curve of the
same color than c (hence intersecting c). Since dc is comparable with dπc , by
definition of the latter, there exists such an exponent N2 . Then it follows
that, for all k ∈ N, the collection {Γc = hτckN2 i, c ∈ S }, is a composite
rotating family for all k.
The second composite projection system is a sub-system, invariant for
G0 , provided by the MCG(Σ)-orbit of a since simple closed curve c0 ∈ S .
Namely, the composite rotating family is the collection {Γc , c ∈ MCG(Σ)c0 ⊂
S }.
It is straightforward that both families are composite rotating families.
27
One can then apply Theorem 2.2. In the first case, one obtains that
the group generated by the kN2 -th powers of all Dehn twists has a partially
commutative presentation, which is the second point of Theorem 1. In the
case of the second composite rotating family, one obtains that the group generated by all kN2 -th powers of all Dehn twists that are MCG(Σ)-conjugated
to τc0 has a partially commutative presentation. This latter group is the
2 in MCG(Σ). We therefore obtained Theorem 1.
normal closure of τckN
0
Acknowledgements
The present work has been mostly developed during the visit of the author
to the Mathematical Science Research Institute in Berkeley, during the thematic semester on Geometric Group Theory of the Fall 2016. The author is
supported by the Institut Universitaire de France.
I wish to thank M. Bestvina, K. Bromberg, K. Fujiwara, J. Mangahas,
J. Manning, and A. Sisto for discussions, and J. Tao, and S. Dowdall for
organising an influencial working seminar in MSRI. After I talked in MSRI
on a first version of this work, which was performed on cone-offs of the
blown-up projection complexes of [BBF], and was more in line with [DGO,
§5], J. Mangahas suggested that I work directly in the language of projection
complexes, which is indeed more natural for this situation, and allows a
similar argument; this choice is in line with one of her work in progress with
M. Clay and D. Margalit. I thank T. Brendle and D. Margalit for suggesting
relevant references.
References
[B] J. Behrstock, Asymptotic geometry of the mapping class group and Teichmller
space. Geom. Topol. 10 (2006), 1523–1578.
[BBF] M. Bestvina, K. Bromberg, K. Fujiwara, Constructing group actions on
quasi-trees and applications to mapping class groups. Publ. Math. Inst. Hautes
Études Sci. 122 (2015), 1–64.
[BM] T. Brendle, D. Margalit, Personnal communication, 2017.
[CLM] M. Clay, C. Leininger, D. Margalit, Abstract commensurators of right angled Artin groups and Mapping Class Groups, Math Research Letters, 21 (3)
461–467 (2014).
[Cou] R. Coulon, Partial periodic quotients of groups acting on a hyperbolic space,
Ann. Inst. Fourier, Grenoble, 66, 5 (2016) 1774–1857.
28
[DGO] F. Dahmani, V. Guirardel, D. Osin, Hyperbolically embedded subgroups
and rotating families in groups acting on hyperbolic spaces. Mem. Amer. Math.
Soc. 245 (2017), no. 1156, v+152 pp.
[Fa] B. Farb, Some problems on Mapping Class Groups and Moduli Spaces, in
Problems on Mapping Class Groups and Related Topics, B. Farb Editor, Proceedings of Symposia in Pure Mathematics, vol.74, Amer. Math. Soc (2006)
11–56.
[Fu] L. Funar, On the TQFT representations of the mapping class groups, Pacific
J. Math. 188 (1999), no. 2, 251–274.
[H] S. Humphries, Normal closures of powers of Dehn twists in Mapping Class
Groups, Glasgow, Math. J. 34 (1992) 313–317.
[I]
N. Ivanov, Fifteen problems about Mapping Class Groups, in Problems on
Mapping Class Groups and Related Topics, B. Farb Editor, Proceedings of
Symposia in Pure Mathematics, vol.74, Amer. Math. Soc (2006) 71–80.
[K] T. Koberda, Right-angled Artin groups and a generalized isomorphism problem for finitely generated subgroups of mapping class groups. Geom. Funct.
Anal. 22 (2012), no. 6, 1541–1590.
[Man] J. Mangahas, Uniform uniform exponential growth of subgroups of the mapping class group, Geom. Funct. Anal., 19 (2010), 1468–1480.
[Man2] J. Mangahas, A recipe for short-word pseudo-Anosovs, Am. J. Math., 135
(2013), 1087–1116.
[Mas] G. Masbaum, On powers of half-twists in M (0, 2n), arXiv:1608.08449.
[MM] H. A. Masur, Y. N. Minsky, Geometry of the complex of curves. I. Hyperbolicity, Invent. Math., 138 (1999), 103–149.
[S] C. Stylianakis, The normal closure of a power of a half Dehn twist has infinite
index in the mapping class group of a punctured sphere, arXiv:1511.02912.
Institut Fourier, Univ. Grenoble Alpes, cnrs, 38000 Grenoble, France
e-mail: [email protected]
29
| 4math.GR
|
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT
MATRICES: IMPROVED BOUNDS
arXiv:1608.06498v2 [cs.IT] 26 Dec 2017
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
Abstract. We consider the problem of encoding a finite set of vectors into a
small number of bits while approximately retaining information on the angular
distances between the vectors. By deriving improved variance bounds related
to binary Gaussian circulant embeddings, we largely fix a gap in the proof of
the best known fast binary embedding method. Our bounds also show that
well-spreadness assumptions on the data vectors, which were needed in earlier
work on variance bounds, are unnecessary. In addition, we propose a new
binary embedding with a faster running time on sparse data.
1. Introduction
In this paper we are concerned with the problem of encoding a set of vectors
D located on the sphere in a high-dimensional space into a small number of bits.
Our goal is to obtain an efficient encoding which approximately retains information
on the angular distances between the points in the data set. Formally, we want
to construct a δ-binary embedding of D, i.e., a map f : Sn−1 → {−1, 1}m and a
distance d on {−1, 1}m so that
|d(f (p), f (q)) − dSn−1 (p, q)| ≤ δ
for all p, q ∈ D,
where dSn−1 denotes the normalized geodesic distance on the sphere. Recent works
have shown that this goal can be achieved using maps of the form
fA (x) = sgn(Ax),
x ∈ Rn ,
with A ∈ Rm×n . That is, one first embeds the data into a lower-dimensional space
using a linear map and subsequently takes entry-wise signs of the embedded vectors.
Apart from the intrinsic interest in binary embeddings for complexity reduction and
computational purposes, they play an important role in one-bit compressed sensing
[8, 12] and have been considered as computationally cheap layers in deep neural
networks [4].
In many successful approaches, A is a random matrix drawn independently of
the data. For instance, is not hard to show that if A is an m × n standard Gaussian
matrix G and D ⊂ Sn−1 is any finite set of N points, then with probability exceeding
1 − η, fA is a δ-binary embedding for D into ({−1, 1}m, dH ), provided that the
number of bits satisfies m & δ −2 log(N/η) [13, 14]. Here, dH denotes the normalized
Hamming distance and . hides an absolute constant. It is known [14] that this bit
complexity is optimal : any oblivious random map from D into {−1, 1}m which is a
δ-binary embedding with probability at least 1 − η must satisfy m & δ −2 log(N/η).
Although the Gaussian binary embedding achieves the optimal bit complexity
on finite sets, it has a clear computational downside: as a Gaussian random matrix
is densely populated, the implementation involves slow matrix-vector multiplications. A simple idea to decrease the running time, suggested in [14], is to first
2010 Mathematics Subject Classification. 60B20,68Q87.
Key words and phrases. Binary embeddings, Johnson-Lindenstrauss embeddings, circulant
matrices.
1
2
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
reduce the dimensionality of the data by using a fast Johnson-Lindenstrauss transform [2] before applying a Gaussian binary embedding. We will refer to this map
as the accelerated Gaussian binary embedding. For appropriate choices of the involved parameters, this map fA with A = GΦFJL , is a δ-binary embedding into
({−1, 1}m, dH ) with optimal bit complexity and, if log N . δ 2 n1/2 , runs in time
O(n log n). However, outside of this parameter range the running time deteriorates.
To obtain near-linear time and improved space complexity in a wider range of
parameters, focus has recently shifted from purely Gaussian matrices to more structured random matrices. It was experimentally observed [16] that an equally good
embedding into ({−1, 1}m, dH ) can be obtained by replacing the dense Gaussian
matrix A by a subset of m rows of a Gaussian circulant matrix with randomized
column signs. That is, one considers A = RI Cg Dε , where, given I ⊂ [n], RI is
a subsampling operator that restricts a vector to its entries indexed by I, Cg is
a circulant matrix generated by a standard Gaussian vector and Dε is a diagonal matrix with independent random signs on the diagonal. This matrix allows
for matrix-vector multiplication in time O(n log n) by exploiting the fast Fourier
transform. The experiments in [16] indicate that the performance of the embedding
deteriorates if one leaves out the column sign randomization.
In view of these experiments, it is desirable to try to prove that the matrix
A = RI Cg Dε induces a binary embedding with the optimal bit complexity. As
dH (fA (p), fA (q)) is an unbiased estimator of the geodesic distance dSn−1 (p, q), it
would suffice to show that dH (fA (p), fA (q)) concentrates strongly around its mean
for fixed points p, q. Due to the non-linearity of the sign function and the dependencies between the circulant rows, this is a tough problem. In fact, even an
optimal bound for the variance of dH (fA (p), fA (q)) is not known. Nevertheless,
some interesting partial results have recently been established, which can roughly
be grouped in two directions. In both cases, the results have been used to construct new fast binary embeddings by appropriately modifying the matrix A using
additional randomness.
In the first direction, developed by the authors in [15] and later quantitatively
improved in [11], it is shown that the map fA associated to A = RI Cg Dε is a δbinary embedding on a finite set D with large probability, provided that all vectors
in D as well as in D − D are well-spread. One can show that an arbitrary set of
points will have these two properties with high probability after they have been
pre-processed using a randomized Hadamard transform. To be more precise, the
corresponding result of [11] is as follows. Let I be a set of m indices selected
uniformly at random, let g1 , g2 be independent standard Gaussian vectors, Dg1 ,
Dg2 the diagonal matrices with g1 , g2 on the diagonal and let H be an n × n
Hadamard matrix. Consider A = RI Cg Dg1 HDg2 and let D ⊂ Sn−1 be any set of
N points. If
log N . δ 2 (log n)−1 n1/3 ,
m & δ −3 log N,
then with high probability fA is a δ-binary embedding of D into ({−1, 1}m, dH ).
In contrast to the accelerated Gaussian binary embedding, fA (x) can always be
computed in time O(n log n). However, the parameter range in which fA is provably
a δ-binary embedding is more restrictive.
The second direction, pursued by the authors in [14], seeks to prove an optimal
variance bound for dH (fA (p), fA (q)), meaning that the bound decays as O(1/m).
Once this is established, one can stack independent copies of A into a matrix to
obtain a high probability binary embedding. To be precise, [14] introduced the map
fA , with A = ΨΦFJL , where ΦFJL is a n′ × n fast Johnson-Lindenstrauss transform
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
3
and
Ψ(1)
′
Ψ = ... ∈ Rm×n ,
Ψ(B)
consists of B independent m/B × n′ subsampled Gaussian Toeplitz matrices with
randomized column signs. It was claimed in [14] that for appropriate choices of n′
and B, the map fA is a δ-binary embedding into ({−1, 1}m, dmed,B ), where dmed,B is
the median over the B block-wise Hamming distances. The embedding attains the
optimal bit complexity and it runs in O(n log n) when log N . δn1/2 /(log(1/δ))1/2 .
To our knowledge, this is the only embedding with guarantees superior to the ones
for the accelerated Gaussian binary embedding, in the sense that it is simultaneously
a δ-binary embedding with the optimal bit complexity and runs in near-linear time
in a larger parameter range than log N . δ 2 n1/2 . Unfortunately, as we point out in
Remark 3.4, the proof in [14] contains a subtle gap that cannot be fixed in a trivial
way.
The goal of our paper is two-fold. On the one hand, by deriving improved
variance bounds we contribute to the theoretical understanding of the numerical
experiments with circulant binary embeddings from [16]. On the other hand, our
work largely fixes the proof gap in [14]. We start by considering a binary embedding
fA with A = RI Cg . We show, on the one hand, that there are two vectors p, q ∈
S n−1 that cannot be embedded well for any choice of I ⊂ [n]. On the other hand,
we prove that if I consists of m dyadic integers and p, q are m-sparse, then
1
Var dH (fA (p), fA (q)) . .
m
Thus, in general a subsampled Gaussian circulant matrix does not yield a good
binary embedding if one does not randomize the column signs of the matrix. However, by subsampling judiciously, one can still obtain optimal variance decay for
relatively sparse vectors. Next, we consider the situation where we additionally use
column sign randomization, i.e., we set A = RI Cg Dε . In Theorem 2.9 we show
that if I is a set of m indices chosen uniformly at random, then for any p, q ∈ Sn−1
1
1
+√ ,
Var dH (fA (p), fA (q)) .
m
n
√
√
which is optimal for m ≤ n. Up to the 1/ n factor, this is the variance bound
claimed and needed for the proof in [14]. As a consequence, we can resurrect the
aforementioned guarantees for the second binary embedding map from [14], at least
when n & δ −6 . Our result (1) is always superior to the variance bounds based on
well-spreadness conditions√from [15] (i.e., even if p, q are maximally well-spread).
We conjecture that the 1/ n factor in (1) can be removed. This would remove the
condition n & δ −6 .
As a final contribution of our work, we analyze modifications of the two embeddings from [14] that are obtained by replacing the fast Johnson-Lindenstrauss
transform ΦFJL in these maps by a sparse Johnson-Lindenstrauss transform ΦSJL
[6, 3, 9]. The resulting embeddings√again realize the optimal bit complexity and
have near-linear running times on n-sparse vectors under similar conditions on
n, δ and N . In contrast to the ΦFJL -based maps, however, the running times can
improve even further if the vectors in D are very sparse. We refer to Section 3 for
further details.
(1)
1.1. Notation. We start by fixing some notation that will be used throughout. For
any given x = (x1 , . . . , xn ) ∈ Rn , we define Dx ∈ Rn×n to be the diagonal matrix
4
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
with x on its diagonal
x. That is,
x1 0 · · ·
0 x1
0
0
x2
Dx =
0
.
.
..
..
0 ···
and we let Cx ∈ Rn×n be the circulant matrix generated by
0
..
.
..
.
0
0
0
..
.
xn
xn−1
, Cx =
xn−2
..
.
0
x1
xn
x1
xn
xn−1
..
.
x2
x1
xn
..
.
···
···
···
..
.
xn−2
xn−3
xn−4
..
.
x2
x3
···
xn−1
xn−1
xn−2
xn−3
..
.
xn
We use the shorthand notation [n] = {1, . . . , n}. For i, j ∈ [n], we will always take
i + j to mean addition modulo n. If we define the shift operator T : Rn → Rn by
(T x)i = xi+1 ,
then we can write the i-th row of Cx as T n−i x. Closely related to circulant matrices
are Toeplitz matrices. For a vector y ∈ R2n−1 the Toeplitz matrix Ty ∈ Rn×n
generated by y is defined by
y2n−1
y1
y2
· · · yn−2
yn−1
y2n−2 y2n−1
y1
· · · yn−3
yn−2
y2n−3 y2n−2 y2n−1 · · ·
···
yn−3
Ty =
..
..
..
..
..
..
.
.
.
.
.
.
yn
yn+1
yn+2
···
y2n−2
y2n−1 .
Observe that Ty can be viewed as the upper left n × n block of the circulant matrix
Cy ∈ R2n×2n . As a final building block for the construction of binary embeddings
we define, for any given I ⊂ [n], the restriction operator RI : Rn → R|I| by
RI x = (xi )i∈I .
Below we will use several different distances. For x, y ∈ Rn with x, y 6= 0 we define
hx, yi
1
.
dSn−1 (x, y) = arccos
π
kxk2 kyk2
In particular, for two points x, y on the unit sphere, dSn−1 (x, y) denotes the normalized geodesic distance between x and y. The normalization is chosen so that
opposite points on the sphere have distance one. On the discrete cube {−1, 1}m
we consider two different distances. First, we consider the normalized Hamming
distance, i.e.,
m
1 X
1x 6=y .
dH (x, y) =
m i=1 i i
Second, we use a distance considered before in [14]. Let B be a given block size so
that m′ = m/B is an integer. Accordingly, we define dmed,B to be the median of the
block-wise Hamming distances, that is, for x = (x1 , . . . , xB ) and y = (y1 , . . . , yB ),
′
with xi , yi ∈ {−1, 1}m for i ∈ [B], we set
dmed,B (x, y) = median((dH (xi , yi ))B
i=1 ).
1.2. Acknowledgement. The authors would like to thank the reviewers for valuable comments, in particular for a suggestion that substantially shortened the proof
of Lemma 2.3. A. Stollenwerk acknowledges funding by the European Research
Council through ERC Starting Grant StG 258926. S. Dirksen and A. Stollenwerk acknowledge funding by the Deutsche Forschungsgemeinschaft (DFG) through
the project Quantized Compressive Spectrum Sensing (QuaCoSS), which is part
of the priority program SPP 1798 Compressed Sensing in Information Processing
(COSIP).
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
5
2. Variance bounds
In this section we consider an m × n subsampled Gaussian circulant matrix
A = RI Cg , with g an n-dimensional standard Gaussian, as well as a version with
randomized column signs, i.e., A = RI Cg Dε , where ε is a Rademacher vector, i.e.,
a vector of n independent random signs (P(εi = 1) = P(εi = −1) = 1/2). Denote
the i-th row of A by ai . By the invariance of the standard Gaussian distribution
d
under reflections, we have in both cases ai ∼ g. Hence for any p, q ∈ Sn−1
1 X
E(dH (fA (p), fA (q))) =
(2)
P(sgn(hai , pi) 6= sgn(hai , qi))
m
i∈I
= P(sgn(hg, pi) 6= sgn(hg, qi)) = dSn−1 (p, q),
where the final equality is [7, Lemma 3.2]. This shows that dH (fA (p), fA (q)) is an
unbiased estimator of the geodesic distance. To investigate whether these random
matrices induce a δ-binary embedding fA of finite point sets into ({−1, 1}m, dH ),
we first consider a more simple problem and try to bound the quantity
Var dH (fA (p), fA (q)) = Var dH (sgn(Ap), sgn(Aq))
for arbitrary p, q ∈ Sn−1 . Define the indicator random variable
Xi = 1sgn(hai ,pi)6=sgn(hai ,qi) .
We can then write
m
1 X
Var dH (fA (p), fA (q)) = Var
Xi
m i=1
(3)
=
m
2
1 X
Var(Xi ) + 2
2
m i=1
m
X
Cov(Xi , Xj ).
1≤i<j≤m
If A is a standard Gaussian matrix, then the second term vanishes and it is easy
to obtain an optimal bound. In the case of Gaussian circulant matrices, the rows
of A are heavily dependent and it is therefore non-trivial to bound the cross-terms
Cov(Xi , Xj ). This is the main technical challenge addressed in this section.
It is worthwhile to note that the best possible variance decay in m that we can
expect for arbitrary p, q ∈ Sn−1 is
1
(4)
Var dH (fA (p), fA (q)) . .
m
Indeed, by Chebyshev’s inequality, this estimate implies
1
P(|dH (fA (p), fA (q)) − dSn−1 (p, q)| ≥ δ) .
mδ 2
and in particular, on every two-point set {p, q} the map fA is an (oblivious) δbinary embedding with probability 1/2, say, if m & δ −2 . By [14, Theorem 3.1], this
scaling in δ is optimal.
In the first part of this section, we investigate variance bounds for a subsampled
Gaussian circulant matrix. It turns out that there exist vectors p, q for which the
variance does not decay at all. However, one can still get optimal decay for sparse
vectors if one subsamples dyadically (see Proposition 2.7). In the second part of
this section, we consider a subsampled Gaussian circulant matrix with randomized
√
column signs. Our main result, Theorem 2.9, says that (4) holds for m ≤ n if the
set of sampled rows I is chosen uniformly at random. This result is a key ingredient
for our results on fast binary embeddings in Section 3.
For any x ∈ Rn we define the binary random variable
Z(x) = sgn(hg, xi).
6
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
The following is a key observation to control the covariance terms Cov(Xi , Xj ) for
i 6= j.
Theorem 2.1. For any x1 , x2 , y1 , y2 ∈ Sn−1
| Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) )| ≤ 8 max{|hx1 , y1 i|, |hx1 , y2 i|, |hx2 , y1 i|, |hx2 , y2 i|}.
Remark 2.2. In [15, Lemma 6], it was shown that
(5)
| Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) )| ≤ 2 max{kΠy1 k2 , kΠy2 k2 },
where Π is the projection onto Span{x1 , x2 }. If {x1 , z2 } is any orthonormal basis
of Span{x1 , x2 }, then the right hand side is up to constants equal to
max{|hx1 , y1 i|, |hx1 , y2 i|, |hz2 , y1 i|, |hz2 , y2 i|}.
Note that this is, again up to constants, larger than the right hand side in Theorem 2.1. Indeed, as x2 = hx1 , x2 ix1 + hz2 , x2 iz2 , it follows for i = 1, 2,
|hx2 , yi i| ≤ |hx1 , yi i| + |hz2 , yi i|.
In particular, (5) is not a good bound if x1 , x2 are far from orthogonal. For example,
let x1 = e1 , x2 = (cos α, sin α), y1 = e2 and y2 = e3 . The right hand side in
Theorem 2.1 is equal to 10|hx2 , y1 i| = 10| sin α|, which scales as α if α is small. On
the other hand, kΠy1 k2 = 1, so that (5) gives a trivial bound.
Theorem 2.1 can more easily be shown in the case where hx1 , x2 i = 0. In order
to tackle the non-orthogonal case, we use Lemma 2.3 below.
Lemma 2.3. Let g1 , g2 and g3 be independent, standard Gaussian random variables. For a, b ∈ R define
f (a, b) = P(sgn(g1 ) 6= sgn(g1 + ag3 ), sgn(g2 ) 6= sgn(g2 + bg3 )).
Then
1
|ab|.
2π
Proof. By invariance of the standard Gaussian distribution under reflections, i.e.,
(6)
f (a, b) ≤
d
(g1 , g2 , g3 ) ∼ (ε1 g1 , ε2 g2 , ε3 g3 ) for all ε1 , ε2 , ε3 ∈ {±1},
we may assume that a, b ≥ 0. Moreover, using invariance of the Gaussian distribution under reflection several times, we obtain
f (a, b) = P(g1 ≥ 0, g1 + ag3 ≤ 0, sgn(g2 ) 6= sgn(g2 + bg3 ))
+ P(g1 ≤ 0, g1 + ag3 ≥ 0, sgn(g2 ) 6= sgn(g2 + bg3 ))
= P(g1 ≥ 0, g1 + ag3 ≤ 0, sgn(g2 ) 6= sgn(g2 + bg3 ))
+ P(−g1 ≥ 0, −g1 − ag3 ≤ 0, sgn(−g2 ) 6= sgn(−g2 − bg3 ))
= 2P(g1 ≥ 0, g1 + ag3 ≤ 0, sgn(g2 ) 6= sgn(g2 + bg3 ))
= 2 P(g1 ≥ 0, g1 + ag3 ≤ 0, g2 ≥ 0, g2 + bg3 ≤ 0)
+ P(g1 ≥ 0, g1 + ag3 ≤ 0, g2 ≤ 0, g2 + bg3 ≥ 0)
= 2 P(g1 ≥ 0, g1 − ag3 ≤ 0, g2 ≥ 0, g2 − bg3 ≤ 0)
+ P(g1 ≥ 0, g1 − ag3 ≤ 0, g2 ≤ 0, g2 − bg3 ≥ 0) .
Clearly, the second term vanishes and so
f (a, b) = 2P(0 ≤ g1 ≤ ag3 , 0 ≤ g2 ≤ bg3 ).
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
7
Notice that if g is standard Gaussian, then for any t ≥ 0
P(0 ≤ g ≤ t) ≤
√1 t.
2π
Using this inequality we find
P(0 ≤ g1 ≤ ag3 , 0 ≤ g2 ≤ bg3 ) = E(10≤g3 10≤g1 ≤ag3 10≤g2 ≤bg3 )
= Eg3 (10≤g3 Eg1 ,g2 (10≤g1 ≤ag3 10≤g2 ≤bg3 ))
= Eg3 (10≤g3 Eg1 (10≤g1 ≤ag3 )Eg2 (10≤g2 ≤bg3 ))
≤ Eg3 (10≤g3 ( √12π ag3 )( √12π bg3 ))
=
which implies f (a, b) ≤
2
ab
2π E(10≤g3 (g3 ) )
=
ab
4π ,
ab
2π .
1
Remark 2.4. The constant 2π
in inequality (6) of Lemma 2.3 is optimal. Indeed,
consider the symmetric case a = b. From the proof of Lemma 2.3 we see that
f (a, a) = 2P(0 ≤ g1 ≤ ag3 , 0 ≤ g2 ≤ ag3 )
= 2P({0 ≤ g1 ≤ g2 ≤ ag3 } ∪ {0 ≤ g2 ≤ g1 ≤ ag3 })
= 2 P(0 ≤ g1 ≤ g2 ≤ ag3 ) + P(0 ≤ g2 ≤ g1 ≤ ag3 )
= 4P(0 ≤ g1 ≤ g2 ≤ ag3 )
Z ∞
Z ∞
Z ∞
2
1 −x22 /2
1 −x21 /2
1
√ e
√ e
√ e−x3 /2 dx3 .
dx1
dx2
=4
x2
2π
2π
2π
x1
0
a
Observe that
Z ∞
Z ∞
2
2
2
x2
df
4
−x21 /2
e
dx
e−x2 /2 (−e−x2 /2a )(− 2 )dx2
(a, a) =
1
3/2
da
a
(2π)
0
x1
Z ∞
Z ∞ x2 2
a +1
2
2
4
e−x1 /2 dx1
e− 2 ( a2 ) x2 dx2
=
(2π)3/2 a2 0
x1
Z ∞
a2 x21 a2 +1
2
4
e−x1 /2 2
=
e− 2 ( a2 ) dx1
3/2
2
a +1
(2π) a 0
Z ∞ x2 2
1 2a +1
4
e− 2 ( a2 ) dx1
=
3/2
2
(2π) (a + 1) 0
r
1
π
a
a
4
√
√
=
.
=
π (a2 + 1) 2a2 + 1
(2π)3/2 (a2 + 1) 2 2a2 + 1
Hence,
(7)
f (a, a)
= lim
a→0
a→0
a2
lim
df
da (a, a)
2a
= lim
a→0
1
1
1
√
=
.
2π (a2 + 1) 2a2 + 1
2π
In the proof of Theorem 2.1 we will use the following simple estimates.
Lemma 2.5. Suppose x, y ∈ Sn−1 with hx, yi ≥ 0. Then
(8)
1
kx − yk2
dSn−1 (x, y) ≤
23/2
dSn−1 (x, y) ≤
1p
1 − hx, yi2 .
2
and, as a consequence,
(9)
Proof. (9) immediately follows from (8) as
1p
1p
1 p
1
1 − hx, yi ≤
1 − hx, yi2 .
kx − yk2 = 3/2 2 − 2hx, yi =
3/2
2
2
2
2
8
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
To prove (8), define z = hx, yi ∈ [0, 1]. Then dSn−1 (x, y) = π1 arccos(z) and
√
√
1
kx − yk2 = 21 1 − z. We need to show that f (z) = 12 1 − z − π1 arccos(z) ≥ 0
23/2
for z ∈ [0, 1]. Clearly, f (0) = f (1) = 0 and
√
− 41 1 + z + π1
1
1
1
1
′
√
=
.
+ √
f (z) = − √
4 1 − z π 1 − z2
1 − z2
Hence f ′ (z0 ) = 0 for z0 = π162 − 1 ∈ (0, 1). Since f ′ (z) > 0 for z ∈ (0, z0 ) and
f ′ (z) < 0 for z ∈ (z0 , 1), the result follows.
We are now ready to prove Theorem 2.1.
Proof of Theorem 2.1. We may assume that hx1 , x2 i ≥ 0. Indeed, if hx1 , x2 i ≤ 0,
then hx1 , −x2 i ≥ 0 and by using dSn−1 (x1 , x2 ) + dSn−1 (x1 , −x2 ) = 1 and P(Z(y1 ) 6=
Z(y2 )) = dSn−1 (y1 , y2 ) we see
Cov(1Z(x1 )6=Z(−x2 ) , 1Z(y1 )6=Z(y2 ) )
= P Z(x1 ) 6= Z(−x2 ), Z(y1 ) 6= Z(y2 ) − dSn−1 (x1 , −x2 )dSn−1 (y1 , y2 )
= dSn−1 (y1 , y2 ) − P Z(x1 ) 6= Z(x2 ), Z(y1 ) 6= Z(y2 )
− 1 − dSn−1 (x1 , x2 ) dSn−1 (y1 , y2 )
= −P Z(x1 ) 6= Z(x2 ), Z(y1 ) 6= Z(y2 ) + dSn−1 (x1 , x2 )dSn−1 (y1 , y2 )
= − Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) ),
which implies
| Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) )| = | Cov(1Z(x1 )6=Z(−x2 ) , 1Z(y1 )6=Z(y2 ) )|.
From now on, we assume hx1 , x2 i ≥ 0. Clearly, x1 , x2 , y1 , y2 span an at most 4dimensional subspace of Rn . Hence one can construct an orthogonal matrix O,
which maps x1 , x2 , y1 and y2 into R4 × {0}n−4. In fact, we can choose O so that
Ox1 = e1 , Ox2 = a1 e1 + a2 e2 =: a, Oy1 = b1 e1 + b2 e2 + b3 e3 =: b,
Oy2 = c1 e1 + c2 e2 + c3 e3 + c4 e4 =: c
with a2 , b3 , c4 ≥ 0. This can be achieved by
1
1
0
O1 = O˜1 , O2 =
, O3 = 0
0 Õ2
0
defining O = O4 O3 O2 O1 ,
1 0
0 0
0 1
1 0 , O4 =
0 0
0 O˜3
0 0
with Õi ∈ R(n+1−i)×(n+1−i) orthogonal matrices, such that
where
0 0
0 0
1 0
0 O˜4
O1 x1 = e1 , O2 O1 x2 = a, O3 O2 O1 y1 = b and O4 O3 O2 O1 y2 = c.
p
p
Observe that 0 ≤ hx1 , x2 i = he1 , ai = a1 and a2 = 1 − a21 = 1 − hx1 , x2 i2 . We
can assume a2 > 0. Otherwise the statement is trivial, because then hx1 , x2 i = 1,
which implies x1 = x2 and hence Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) ) = 0. Moreover,
p
p
we may assume that γ := c23 + c24 > 0, otherwise 1 = c21 + c22 ≤ |c1 | + |c2 | and
using inequality (9) in Lemma 2.5 yields
| Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) )|
p
1 − hx1 , x2 i2
2
a2 (|c1 | + |c2 |)
|c1 | + |a2 c2 |
|he1 , ci| + |ha, ci − a1 c1 |
a2
≤
≤
=
=
2
2
2
2
2|he1 , ci| + |ha, ci|
2|hx1 , y2 i| + |hx2 , y2 i|
≤
=
.
2
2
≤ P Z(x1 ) 6= Z(x2 ) = dSn−1 (x1 , x2 ) ≤
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
9
The idea is now to replace b, c by suitable approximants b′ , c′ , which are orthogonal
to e1 , a. Set b′ = e3 and c′ = γ1 (c3 e3 + c4 e4 ). Clearly, b′ , c′ ∈ Sn−1 are orthogonal
to e1 , a and we can control their distance to the original vectors b and c. Indeed,
since hb, b′ i = b3 ≥ 0 and hc, c′ i = γ ≥ 0, inequality (9) in Lemma 2.5 yields
q
q
1
1p
1
1
1 − hb, b′ i2 =
1 − b23 =
b21 + b22 ≤ (|b1 | + |b2 |)
dSn−1 (b, b′ ) ≤
2
2
2
2
and
q
q
1
1
1
1p
1 − hc, c′ i2 =
1 − (c23 + c24 ) =
c21 + c22 ≤ (|c1 | + |c2 |).
dSn−1 (c, c′ ) ≤
2
2
2
2
Define
ε = max{|hx1 , y1 i|, |hx1 , y2 i|, |hx2 , y1 i|, |hx2 , y2 i|}.
Then
|b1 | = |he1 , bi| = |hx1 , y1 i| ≤ ε as well as |c1 | = |he1 , ci| = |hx1 , y2 i| ≤ ε.
Similarly we can estimate
(10)
and
(11)
|a2 b2 | = |ha, bi − a1 b1 | ≤ |ha, bi| + |a1 b1 | = |hx2 , y1 i| + |a1 b1 | ≤ 2ε
|a2 c2 | = |ha, ci − a1 c1 | ≤ |ha, ci| + |a1 c1 | = |hx2 , y2 i| + |a1 c1 | ≤ 2ε,
which yields
2ε
2ε
as well as |c2 | ≤ .
a2
a2
In summary we obtain the bounds
ε
ε
ε
ε
(12)
dSn−1 (b, b′ ) ≤ +
and dSn−1 (c, c′ ) ≤ + .
2 a2
2 a2
The rotational invariance of the standard Gaussian distribution implies
|b2 | ≤
Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) ) = Cov(1Z(e1 )6=Z(a) , 1Z(b)6=Z(c) ).
Moreover, since he1 , b′ i = he1 , c′ i = ha, b′ i = ha, c′ i = 0, 1Z(e1 )6=Z(a) and 1Z(b′ )6=Z(c′ )
are independent and in particular
Cov(1Z(e1 )6=Z(a) , 1Z(b′ )6=Z(c′ ) ) = 0.
Using the triangle inequality we now obtain
(13) | Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) )|
= | Cov(1Z(e1 )6=Z(a) , 1Z(b)6=Z(c) ) − Cov(1Z(e1 )6=Z(a) , 1Z(b′ )6=Z(c′ ) )|
≤ |P Z(e1 ) 6= Z(a), Z(b) 6= Z(c) − P Z(e1 ) 6= Z(a), Z(b′ ) 6= Z(c′ ) |
+ |dSn−1 (e1 , a)dSn−1 (b, c) − dSn−1 (e1 , a)dSn−1 (b′ , c′ )|
≤ P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) + P Z(e1 ) 6= Z(a), Z(c) 6= Z(c′ )
+ dSn−1 (e1 , a) dSn−1 (b, b′ ) + dSn−1 (c, c′ ) .
In the last step, we have used
dSn−1 (b, c) ≤ dSn−1 (b, b′ ) + dSn−1 (b′ , c′ ) + dSn−1 (c′ , c),
and that by decomposing the event
{Z(b) 6= Z(c)} = {Z(b) 6= Z(c), Z(b) = Z(b′ ), Z(c) = Z(c′ )}
we obtain
∪ {Z(b) 6= Z(c), Z(b) 6= Z(b′ ) ∨ Z(c) 6= Z(c′ ) },
P Z(e1 ) 6= Z(a), Z(b) 6= Z(c) ≤ P Z(e1 ) 6= Z(a), Z(b′ ) 6= Z(c′ )
+ P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) + P Z(e1 ) 6= Z(a), Z(c) 6= Z(c′ ) .
10
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
It remains to bound all three summands appearing on the far right hand side of
(13). We estimate using (12) and Lemma 2.5
2ε
dSn−1 (e1 , a) dSn−1 (b, b′ ) + dSn−1 (c, c′ ) ≤ dSn−1 (e1 , a) ε +
a2
≤ dSn−1 (e1 , a)ε + ε ≤ 2ε.
Before estimating the other two summands in (13), observe that with
q
q
kb = b22 + b23 and kc = c22 + c23 + c24 ,
we obtain using Lemma 2.5,
(14)
b b 1
3
2
≤
dSn−1 b, 0, ,
kb kb
2
and similarly,
s
1−
b2
2
kb
+
b23 2
ε
1
= |b1 | ≤
kb
2
2
c c c ε
2
3
4
dSn−1 c, 0, , ,
≤ .
kc kc kc
2
(15)
Notice that if kb = 0 (respectively, kc = 0), then b = ±e1 (respectively, c = ±e1 )
and hence ε = 1, so that the result is trivial in this case.
Let us now estimate P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) , the second summand on
the far right hand side of (13) can be bounded in the same fashion. We distinguish
three cases. First, if a22 > 41 , then using (12)
ε
ε
P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) ≤ dSn−1 (b, b′ ) ≤ +
≤ 3ε.
2 a2
Second, if b21 + b22 > 43 , then
1
q
1
1
1
1 − (b21 + b22 ) ≥ arccos(1/2) ≥ ,
arccos(b3 ) = arccos
π
π
π
π
which implies by (12)
1p
1 − he1 , ai2
P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) ≤ dSn−1 (e1 , a) ≤
2
a2
3ε
3πε
=
≤
≤ 3ε.
≤
2
4dSn−1 (b, b′ )
4
dSn−1 (b, b′ ) =
Finally, suppose that a22 ≤
(16)
1
4
and b21 + b22 ≤ 34 . This implies
|a2 b2 |
4
|a2 b2 |
p
=p
≤ √ |a2 b2 | and a1 , b3 > 0.
2
2
2
a1 b 3
3
1 − a2 1 − b 1 − b 2
Using (14) we find
P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ )
= P sgn(g1 ) 6= sgn(a1 g1 + a2 g2 ), sgn(b1 g1 + b2 g2 + b3 g3 ) 6= sgn(g3 )
≤ P sgn(g1 ) 6= sgn(a1 g1 + a2 g2 ), sgn(b2 g2 + b3 g3 ) 6= sgn(g3 )
+ P sgn(b1 g1 + b2 g2 + b3 g3 )) 6= sgn(b2 g2 + b3 g3 )
b
a2
2
≤ P sgn(g1 ) 6= sgn g1 + g2 , sgn(g3 ) 6= sgn
g2 + g3
a1
b3
b b
3
2
+ dSn−1 b, 0, ,
kb kb
b
ε
a2
2
≤ P sgn(g1 ) 6= sgn g1 + g2 , sgn(g3 ) 6= sgn
g2 + g3 + .
a1
b3
2
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
11
Applying Lemma 2.3, (16), and (10) we obtain
a2
b2
P sgn(g1 ) 6= sgn g1 + g2 , sgn(g3 ) 6= sgn g3 + g2
a1
b3
4ε
|a2 b2 |
2|a2 b2 |
≤
≤ √ ≤ √ .
2πa1 b3
π 3
π 3
As a consequence,
ε
4ε
P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) ≤ √ + ≤ 3ε.
2
π 3
In all three cases we find
P Z(e1 ) 6= Z(a), Z(b) 6= Z(b′ ) ≤ 3ε.
To finish the proof, we bound the second summand on the far right hand side of
(13). If either a22 > 14 or c21 + c22 > 34 , then
P Z(e1 ) 6= Z(a), Z(c) 6= Z(c′ ) ≤ 3ε
follows exactly as above, simply by replacing b and b′ everywhere by c and c′ ,
respectively. Let us now assume that a22 ≤ 14 as well as 1 − γ 2 = c21 + c22 ≤ 43 . Then
(17)
|a2 c2 |
4
|a2 c2 |
p
=p
≤ √ |a2 c2 | and a1 , γ > 0.
2
2
2
a1 γ
3
1 − a2 1 − c1 − c2
Using (15) we estimate
P Z(e1 ) 6= Z(a), Z(c) 6= Z(c′ )
= P sgn(g1 ) 6= sgn(a1 g1 + a2 g2 ),
sgn(c1 g1 + c2 g2 + c3 g3 + c4 g4 ) 6= sgn(c3 g3 + c4 g4 )
≤ P sgn(g1 ) 6= sgn(a1 g1 + a2 g2 ), sgn(c2 g2 + c3 g3 + c4 g4 ) 6= sgn(c3 g3 + c4 g4 )
+ P sgn(c1 g1 + c2 g2 + c3 g3 + c4 g4 ) 6= sgn(c2 g2 + c3 g3 + c4 g4 )
= P sgn(g1 ) 6= sgn(a1 g1 + a2 g2 ), sgn(c2 g2 + c3 g3 + c4 g4 ) 6= sgn(c3 g3 + c4 g4 )
c c c
2
3
4
+ dSn−1 c, 0, , ,
kc kc kc
a2
≤ P sgn(g1 ) 6= sgn g1 + g2 ,
a1
c g + c g ε
c
c
g
3 3
4 4
3 3 + c4 g 4
2
g2 +
6= sgn
+
sgn
γ
γ
γ
2
a2
c2
ε
′
′
= P sgn(g1 ) 6= sgn g1 + g2 , sgn
g2 + g 6= sgn(g ) + ,
a1
γ
2
where in the last step we write g ′ := γ1 (c3 g3 + c4 g4 ). Observe that g ′ is a standard
Gaussian random variable which is independent of g1 and g2 . Therefore Lemma 2.3,
(17), and (11) yield
c2
a2
P sgn(g1 ) 6= sgn g1 + g2 , sgn(g ′ ) 6= sgn g ′ + g2
a1
γ
|a2 c2 |
2|a2 c2 |
4ε
≤
≤ √ ≤ √ ,
2πa1 γ
π 3
π 3
which implies that in all cases P Z(e1 ) 6= Z(a), Z(c) 6= Z(c′ ) ≤ 3ε.
Remark 2.6. The dependence on ε = max{|hx1 , y1 i|, |hx1 , y2 i|, |hx2 , y1 i|, |hx2 , y2 i|}
in Theorem 2.1 is optimal up to a constant. We argue by contradiction. Assume
Cov(X, Y ) := Cov(1Z(x1 )6=Z(x2 ) , 1Z(y1 )6=Z(y2 ) ) ≤ Cεs for some constant C > 0
12
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
and s > 1. For a > 0 define x1 = e1 , x2 =
y2 =
√ 1
(e
1+a2 2
+ ae3 ). Then ε = hx2 , y2 i =
a2
1+a2
√ 1
(e
1+a2 1
+ ae3 ), y1 = e2 and
and
Cov(X, Y ) = P sgn(g1 ) 6= sgn(g1 + ag3 ), sgn(g2 ) 6= sgn(g2 + ag3 ) − EXEY
= f (a, a) − g(a)2 ,
where f (a, a) as in Lemma 2.3 and
g(a) = EX = EY = dSn−1 (x1 , x2 ) = dSn−1 (y1 , y2 ) =
We know lima→0
f (a,a)
a2
and as a consequence,
=
1
2π
by (7). Clearly,
′
1
1
arccos( √
) =
2
1
+
a2
1+a
1
1
.
arccos √
π
1 + a2
1
)
arccos2 ( √1+a
2
π 2 g(a)2
=
lim
2
2
a→0
a→0
a
a
1
arccos( √1+a
)
1
2
= lim
= lim
= 1.
2
a→0
a→0 (1 + 3a2 )(1 + a2 )
a(1 + a )
A contradiction now follows by observing that
Cov(X, Y )
1
1
lim
=
− 2 > 0,
2
a→0
a
2π π
but
Cεs
Ca2s
Ca2s−2
=
lim
= 0.
lim 2 = lim
a→0 a
a→0 (1 + a2 )s a2
a→0 (1 + a2 )s
lim
We can now state our main variance bounds for subsampled Gaussian circulant
matrices.
Proposition 2.7. Let I ⊂ [n], let g be an n-dimensional standard Gaussian and
set A = RI Cg .
Pk
Pk
(1) Set n = 2k, p = i=1 e2i−1 , q = i=1 e2i and I ⊂ [n] arbitrary subset with
m elements. Then
1
Var dH (fA (p), fA (q)) = .
4
1
.
In particular, P(|dH (fA (p), fA (q)) − dSn−1 (p, q)| ≥ 1/4) ≥ 36
(2) Let I ⊂ [n] be the set of the first m dyadic integers. Then for any two
s-sparse vectors p, q ∈ Sn−1
s
1
+ 2.
Var dH (fA (p), fA (q)) .
m m
This shows that a subsampled Gaussian circulant matrix in general does not induce a binary embedding with high probability for arbitrary p, q ∈ Sn−1 . However,
if p, q are m-sparse and one subsamples dyadically, then Var dH (fA (p), fA (q))
decays optimally in terms of the number of measurements m.
Proof of Proposition 2.7. First, let us consider case (1). As before, we let ai be the
i-th row of A and use the notation
Xi = 1sgn(hai ,pi)6=sgn(hai ,qi) .
For any i, j ∈ I
Cov(Xi , Xj ) = P sgn(hT n−i g, pi) 6= sgn(hT n−i g, qi),
sgn(hT n−j g, pi) 6= sgn(hT n−j g, qi) − dSn−1 (p, q)2
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
13
k
k
X
X
= P sgn
g2k+2s−i−1 6= sgn
g2k+2s−i ,
s=1
sgn
s=1
k
X
s=1
= P sgn
k
X
s=1
g2k+2s−j−1 6= sgn
g2k+2s−i−1 6= sgn
= dSn−1 (p, q) − dSn−1 (p, q)2 .
k
X
g2k+2s−j
s=1
k
X
g2k+2s−i
s=1
− dSn−1 (p, q)2
− dSn−1 (p, q)2
Here we used that for any pair i, j ∈ [n]
n
sgn
k
X
s=1
=
n
k
X
o
6 sgn
g2k+2s−i−1 =
g2k+2s−i
sgn
s=1
k
X
s=1
g2k+2s−j−1 6= sgn
Since hp, qi = 0, we conclude
k
X
g2k+2s−j
s=1
o
.
1 X
1 X
Var dH (fA (p), fA (q)) = Var
Xi = 2
Cov(Xi , Xj )
m
m
i∈I
i,j∈I
1
1 1
− = .
2 4
4
By the Paley-Zygmund inequality, this implies for any δ ≤ 1/4
= dSn−1 (p, q) − dSn−1 (p, q)2 =
P(|dH (fA (p), fA (q)) − dSn−1 (p, q)| ≥ δ)
≥
(Var(dH (fA (p), fA (q))) − δ 2 )2
1
≥
.
4
E(dH (fA (p), fA (q)) − dSn−1 (p, q))
36
i−1
Let us consider case (2). Here the i-th row of A takes the form ai = T n−2
g and
Xi = 1sgn(hai ,pi)6=sgn(hai ,qi) = 1sgn(hT n−2i−1 g,pi)6=sgn(hT n−2i−1 g,qi) .
Moreover, since the standard Gaussian distribution is invariant under permutation
of coordinates
Cov(Xi , Xj ) = Cov(1sgn(hg,pi)6=sgn(hg,qi) , 1sgn(hg,T 2j−1 −2i−1 pi)6=sgn(hg,T 2j−1 −2i−1 qi) )
= Cov(1Z(p)6=Z(q) , 1Z(T 2j−1 −2i−1 p)6=Z(T 2j−1 −2i−1 q) ).
As a consequence, Theorem 2.1 implies
j−1
| Cov(Xi , Xj )| . |hp, T 2
−2i−1
+ |hq, T
2j−1 −2i−1
Observe that
X
X
j−1
i−1
|hp, T 2 −2 qi| ≤
|pℓ |
1≤i<j≤m
ℓ
j−1
pi| + |hp, T 2
X
−2i−1
pi| + |hq, T
1≤i<j≤m
qi|
2j−1 −2i−1
|qℓ+2j−1 −2i−1 | ≤
qi|.
X
ℓ
√
|pℓ | s ≤ s,
where in the last two inequalities we used Cauchy-Schwarz, kpk2 = kqk2 = 1, the
sparsity of p, q and that {2j−1 − 2i−1 : 1 ≤ i < j ≤ m} is injectively contained in
[n]. Putting our estimates together we find
m
X
1 X
Var(Xi ) + 2
Var dH (fA (p), fA (q)) ≤ 2
| Cov(Xi , Xj )|
m i=1
1≤i<j≤m
14
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
.
1
s
+ 2.
m m
Next, we consider a subsampled Gaussian circulant matrix with randomized
column signs A = RI Cg Dε . In this case,
Xi = 1sgn(hai ,pi)6=sgn(hai ,qi) = 1sgn(h(T n−i g)⊙ε,pi)6=sgn(h(T n−i g)⊙ε,qi) ,
where ⊙ denotes pointwise multiplication of vectors.
Lemma 2.8. Let p, q ∈ Sn−1 . For any i 6= j, |j − i| 6=
| Cov(Xi , Xj )| ≤ 8
p ⊙ T j−i p
2
+ p ⊙ T j−i q
2
n
2
+ q ⊙ T j−i p
2
+ q ⊙ T j−i q
2
.
Proof. We may assume i < j. Using the invariance of the standard Gaussian
distribution under coordinate permutations
Cov(Xi , Xj )
= Cov(1sgn(hT n−i g,ε⊙pi)6=sgn(hT n−i g,ε⊙qi) , 1sgn(hT n−j g,ε⊙pi)6=sgn(hT n−j g,ε⊙qi) )
= Cov(1Z(ε⊙p)6=Z(ε⊙q) , 1Z(T j−i (ε⊙p))6=Z(T j−i (ε⊙q)) ).
Since g and ε are independent
| Cov(Xi , Xj )| = |Eε Cov(1Z(ε⊙p)6=Z(ε⊙q) , 1Z(T j−i (ε⊙p))6=Z(T j−i (ε⊙q)) )|
≤ Eε | Cov(1Z(ε⊙p)6=Z(ε⊙q) , 1Z(T j−i (ε⊙p))6=Z(T j−i (ε⊙q)) )|.
Applying Theorem 2.1 and Jensen’s inequality we find
| Cov(Xi , Xj )| ≤ 8 Eε |hε ⊙ p, T j−i (ε ⊙ p)i| + |hε ⊙ p, T j−i (ε ⊙ q)i|
+ |hε ⊙ q, T j−i (ε ⊙ p)i| + |hε ⊙ q, T j−i (ε ⊙ q)i|
q
q
≤8
E |hε ⊙ p, T j−i (ε ⊙ p)i|2 + E |hε ⊙ p, T j−i (ε ⊙ q)i|2
q
q
+ E |hε ⊙ q, T j−i (ε ⊙ p)i|2 + E |hε ⊙ q, T j−i (ε ⊙ q)i|2 .
For s, t ∈ [n], s 6= t and k ∈ [n − 1]
(
1,
E(εs εs+k εt εt+k ) =
0,
if s = t or s = t + k, t = s + k
else.
Observe that s = t + k, t = s + k implies k = n2 . Set k = j − i ∈ [n − 1], then k 6=
and therefore
n
X
E|hε ⊙ p, T k (ε ⊙ q)i|2 =
ps qs+k pt qt+k E(εs εs+k εt εt+k )
=
s,t=1
n
X
s=1
2
p2s qs+k
= p ⊙ T kq
2
2
n
2
.
Theorem 2.9. Let g be an n-dimensional standard Gaussian vector and let ε be
an independent vector of independent random signs. Let A = RI Cg Dε .
(1) If I = [m], then for any p, q ∈ Sn−1
(18)
1
Var dH (fA (p), fA (q)) . √ .
m
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
15
(2) If I ⊂ [n] is chosen uniformly at random from all subsets of size m, then
for any p, q ∈ Sn−1
1
1
(19)
Var dH (fA (p), fA (q)) .
+√ .
m
n
In [15, Theorem 3] it was shown that if I = [m], then
1
+ρ
Var dH (fA (p), fA (q)) .
m
where ρ = max{kpk∞ , kqk∞ }. This bound is non-trivial if both p and q are wellspread. Note that √1n ≤ ρ, and equality is achieved if and only if p and q are
perfectly spread, i.e. pi = qi = ± √1n for all i ∈ [n]. Our variance bound in (2)
is always better and shows that a well-spreadness assumption is not necessary if
one subsamples uniformly at random. As was
√ discussed at the beginning of this
section the estimate
(19)
is
optimal
for
m
≤
n. We conjecture that it is possible
√
to remove the 1/ n factor altogether.
Proof. We start by proving (1). Since EXi = dSn−1 (p, q), it follows that Var Xi =
dSn−1 (p, q) − dSn−1 (p, q)2 and therefore (see (3))
X
d n−1 (p, q) − dSn−1 (p, q)2
2
+ 2
Cov(Xi , Xj ).
Var dH (fA (p), fA (q)) = S
m
m
1≤i<j≤m
By Lemma 2.8 for i < j with j − i 6=
| Cov(Xi , Xj )| . p ⊙ T j−i p
If j − i =
ities with
n
2
+ p ⊙ T j−i q
2
+ q ⊙ T j−i p
2
+ q ⊙ T j−i q
.
|{1 ≤ i < j ≤ m : j − i = k}| = m − k,
X
1≤i<j≤m
=
| Cov(Xi , Xj )|
m−1
X
X
k=1 {i<j : j−i=k}
.
m−1
X
k=1
(m − k)
| Cov(Xi , Xj )|
p ⊙ T kp
2
+ p ⊙ T kq
n
1{m−1≥ n2 }
+ m−
2
m
X
≤m
p ⊙ T kp 2 + p ⊙ T kq
.m
k=1
3/2
2
2
+ q ⊙ T kp
+ q ⊙ T kp
.
2
2
+ q ⊙ T kq
+ q ⊙ T kq
2
2
k=1 i=1
i=1
+m
For the last inequality we have used Cauchy-Schwarz and
n
n
n X
n
n
X
X
X
X
2
2
2
p2i
qi+k
= 1 for all p, q ∈ Sn−1 .
p2i qi+k
=
p ⊙ T kq 2 =
k=1
2
we use the trivial bound | Cov(Xi , Xj )| ≤ 1. Combining these inequal-
we find
(20)
2
n
2
k=1
We now prove (2). Let I be chosen uniformly at random and let θi ∈ {0, 1} be the
induced selector variables, i.e., θi = 1 if and only if i ∈ I. Note that
1 X
1sgn(hai ,pi)6=sgn(hai ,qi)
Var dH (fA (p), fA (q)) = Var
m
i∈I
16
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
=
Clearly, Var(θi Xi ) =
n
1 X
2
Var(θi Xi ) + 2
2
m i=1
m
m
n dSn−1 (p, q)
−
m2
2
n2 dSn−1 (p, q)
2
X
Cov(θi Xi , θj Xj ).
1≤i<j≤n
and for i 6= j,
m
EXi EXj
n2
m2
m2
m2
≤ 2 E(Xi Xj ) − 2 EXi EXj = 2 Cov(Xi , Xj ),
n
n
n
as θi and θj are negatively correlated. Combining these estimates, we find
X
d n−1 (p, q) dSn−1 (p, q)2
2
Var dH (fA (p), fA (q)) ≤ S
−
+ 2
Cov(Xi , Xj ).
m
n
n
Cov(θi Xi , θj Xj ) = E(θi θj )E(Xi Xj ) −
1≤i<j≤n
Since (20) holds for any m ≤ n, we obtain
X
| Cov(Xi , Xj )| . n3/2 .
1≤i<j≤n
3. Fast binary embeddings
In the following, we are interested to construct a δ-binary embedding for an
arbitrary finite dataset D ⊂ Rn , which achieves both optimal bit complexity
δ −2 log(|D|) and runs in almost linear time n log(n). We consider modifications
of the two binary embeddings from [14], which were discussed in the introduction.
A common step in all the binary embeddings we consider is to first reduce the dimensionality of the data set, while approximately preserving the geodesic distances
between the points. The following observation, from the proof of [14, Lemma 3.6],
says that this can be achieved using a map that preserves the norms of and Euclidean distances between the data points up to a multiplicative error. Let us say
′
that a matrix A ∈ Rn ×n is a δ-isometry on a set P ⊂ Rn if
(21)
(1 − δ)kzk2 ≤ kAzk2 ≤ (1 + δ)kzk2
for all z ∈ P,
that is, A preserves norms of vectors up to a multiplicative error δ.
Lemma 3.1. [14] Let D ⊂ Sn−1 be symmetric and suppose that A is a δ-isometry
on both D and D − D. Then,
|dSn′ −1 (Ax, Ay) − dSn−1 (x, y)| ≤ δ
for all x, y ∈ D.
A substantial amount of research has been devoted to the construction of random
matrices that satisfy (21) with high probability. We will consider two families of
constructions that allow for fast matrix-vector multiplication. The family of fast
Johnson-Lindenstrauss transforms [2, 1, 10, 5] rely on the fast Fourier transform
(FFT) for fast multiplication. One particular construction from this family is the
n′ × n random matrix
r
n
RI HDε ,
ΦFJL =
n′
where I is a subset of n′ indices selected uniformly at random from [n], ε is a
Rademacher vector independent of I and H is the Hadamard transform. Using the
FFT, ΦFJL x can be computed in time O(n log n). By combining the results of [10]
and [5] it follows that ΦFJL satisfies the conditions of Lemma 3.1 on a set D of N
points in Sn−1 with probability at least 1 − η if
n′ & δ −2 log(N/η)(log3 (log(N/η)) log(n) + log(1/η)).
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
17
A different family of ‘fast’ constructions can be obtained by sparsifying a random sign matrix. To be more precise, let σij , i ∈ [n′ ], j ∈ [n] be independent
Rademacher random variables. We consider {0, 1}-valued random variables δij ,
which are independent of the σij , with the following properties:
• For a fixed column j the δij are negatively correlated, i.e.
k
k
Y
Y
s k
∀1 ≤ i1 < i2 < . . . < ik ≤ n′ , E
Eδit ,j =
δit ,j ≤
;
n′
t=1
t=1
Pn′
• For any fixed column j there are exactly s nonzero δij , i.e., i=1 δij = s;
′
• The vectors (δij )ni=1 are independent across different columns 1 ≤ j ≤ n.
The n′ × n sparse Johnson-Lindenstrauss transform ΦSJL with column sparsity s
[6, 3, 9] is defined by
1
(ΦSJL )ij = √ σij δij .
s
One possible concrete implementation [9] is to take the columns independent, and
in each column we choose exactly s locations uniformly at random, without replacement, to specify the δij . For any x ∈ Rn with kxk0 non-zero entries, ΦSJL x can be
computed in time O(skxk0 ). It follows from [9] that ΦSJL satisfies the conditions
of Lemma 3.1 on a set D of N points in Sn−1 with probability at least 1 − η if
n′ & δ −2 log(N/η),
s & δ −1 log(N/η).
The following result under condition (i) was already obtained in [14, Algorithm 3].
Note that both embeddings in Proposition 3.2 achieve the optimal bit complexity.
Proposition 3.2 (Accelerated Gaussian binary embeddings). Let D = {x1 , ..., xN } ⊂
′
Sn−1 . Let G ∈ Rm×n be a standard Gaussian matrix, i.e., its entries are independent standard Gaussians. Set
m & δ −2 log(N/η).
Suppose that one of the two conditions hold:
(i) Φ = ΦFJL is an n′ × n FJLT with
n′ & δ −2 log(N/η)(log3 (log(N/η)) log(n) + log(1/η)).
(ii) Φ = ΦSJL is an n′ × n SJLT with
n′ & δ −2 log(N/η),
s & δ −1 log(N/η).
Set A = GΦ. Then, with probability at least 1 − η, fA is a δ-binary embedding of
D into ({−1, 1}m, dH ), i.e.,
sup |dH (fA (xi ), fA (xj )) − dSn−1 (xi , xj )| ≤ δ.
i,j∈[N ]
′
Proof. For i ∈ [N ] we define yi = Φxi ∈ Rn . As has been discussed before, under
both (i) and (ii) we have with probability 1 − η2
(22)
sup |dSn′ −1 (yi , yj ) − dSn−1 (xi , xj )| ≤ δ/2.
i,j∈[N ]
It is well-known (see e.g. [13] or [14][Proposition 2.2]) that fG is with probability
1 − η2 a δ/2-binary embedding on {y1 , . . . , yN }, i.e.,
sup |dH (fG (xi ), fG (xj )) − dSn−1 (xi , xj )| ≤ δ/2.
i,j∈[N ]
Combining these observations yields the result.
18
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
As was already noted in [14], for any x ∈ Sn−1 one can compute fA (x) in time
O(n log(n)) under condition (i) if
√
log(N/η) . δ 2 n.
(23)
Under condition (ii) one can compute fA (x) in time
O(δ −4 log2 (N/η) + δ −1 log(N/η)kxk0 ).
√
In particular, under (23) the embedding fA runs in linear time on n-sparse vectors.
The running time can even improve further if one is embedding a small set of very
sparse vectors, say.
We next consider two binary embeddings, for which one δ-factor in (23) can
′
be removed, provided that n is large enough. Let g (1) , ..., g (B) ∈ Rn be indepen′
dent standard Gaussian random vectors and ε(1) , ..., ε(B) ∈ {−1, 1}n independent
Rademacher vectors. Let I (1) , ..., I (B) ⊂ [n′ ] be independent, uniformly random
subsets of size m′ = m
B . For s ∈ [B] set
′
′
Ψ(s) = RI (s) Cg(s) Dε(s) ∈ Rm ×n .
We now stack these matrices to obtain the m × n′ matrix
(1)
Ψ
..
Ψ = . .
Ψ(B)
Theorem 3.3 (Median fast binary embeddings). Let D = {x1 , ..., xN } ⊂ Sn−1 .
Suppose that one of the following two conditions hold:
(i) Φ = ΦFJL is an n′ × n FJLT with
n′ & δ −2 log(N/η)(log3 (log(N/η)) log(n) + log(1/η)).
(ii) Φ = ΦSJL is an n′ × n SJLT with
n′ & δ −2 log(N/η),
s & δ −1 log(N/η).
Suppose, moreover, that
m
& δ −2 ,
n′ & δ −4 .
B
Set A = ΨΦ. Then with probability at least 1 − η, fA is a δ-binary embedding of D
into ({−1, 1}m, dmed,B ), i.e.,
B & log(N/η),
n′ ≥
sup |dmed,B (fA (xi ), fA (xj )) − dSn−1 (xi , xj )| ≤ δ.
i,j∈[N ]
Theorem 3.3 under condition (i) is essentially the result claimed in [14, Theorem
3.8]. The proof in [14], however, contains a gap (see Remark 3.4 below). Let us note
that our construction of the Ψ(s) is slightly different from the one in [14]. Instead of a
Gaussian Toeplitz matrices we use Gaussian circulant matrices (although our proof
works for Toeplitz matrices as well) and, more importantly, we use uniform random
subsampling instead of deterministic subsampling, in order to invoke Theorem 2.9.
′
Proof. For i ∈ [N ] we define yi = Φxi ∈ Rn . By the discussion prior to Proposition 3.2, under both (i) and (ii) we have with probability 1 − η2
(24)
sup |dSn−1 (xi , xj ) − dSn′ −1 (yi , yj )| ≤ δ.
i,j∈[N ]
Fix i, j ∈ [N ]. By subsequently applying (2), Markov’s inequality and (19) in
Theorem 2.9, we obtain
P(|dH (sgn(Ψ(s) yi ), sgn(Ψ(s) yj )) − dSn′ −1 (yi , yj )| ≥ δ)
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
19
= P(|dH (sgn(Ψ(s) yi ), sgn(Ψ(s) yj )) − EdH (sgn(Ψ(s) yi ), sgn(Ψ(s) yj ))| ≥ δ)
≤ δ −2 Var dH (sgn(Ψ(s) yi ), sgn(Ψ(s) yj )
1
1 1
√
≤ ,
+
. δ −2
m′
4
n′
where in the last step we use m′ & δ −2 and n′ & δ −4 . Defining
Es = 1{|dH (sgn(Ψ(s) yi ),sgn(Ψ(s) yj ))−d n′ −1 (yi ,yj )|≥δ} ,
S
this translates to EEs ≤
B
nX
Es <
s=1
1
4.
Observe
o
Bo n
⊂ dmed,B (sgn(Ψyi ), sgn(Ψyj )) − dSn′ −1 (yi , yj ) < δ ,
2
which implies using −EEs ≥ − 41
P( dmed,B (sgn(Ψyi ), sgn(Ψyj )) − dSn′ −1 (yi , yj ) ≥ δ)
≤P
≤P
B
X
s=1
Es ≥
B
1 X
B
s=1
B
1 X
B
1
=P
Es ≥
2
B s=1
2
(Es − EEs ) ≥
1
.
4
Since {Es }s∈[B] are independent, Hoeffding’s inequality yields
B
P( dmed,B (sgn(Ψyi ), sgn(Ψyj )) − dSn′ −1 (yi , yj ) ≥ δ) ≤ e− 32 .
Since this holds for any i, j ∈ [N ], a union bound now implies
B
η
(25) P sup dmed,B (sgn(Ψyi ), sgn(Ψyj )) − dSn′ −1 (yi , yj ) ≥ δ ≤ N 2 e− 32 ≤ ,
2
i,j∈[N ]
where in the last step we use B & log(N/η). The triangle inequality using (24) and
(25) yields
sup |dmed,B (fA (xi ), fA (xj )) − dSn−1 (xi , xj )| ≤ 2δ,
i,j∈[N ]
with probability at least 1 − η. A rescaling in δ yields the result.
Remark 3.4. Let us now briefly discuss the subtle proof gap occurring in [14]. As
was noted before, they considered mappings Ψ(s) which used deterministic subsampling (instead of uniform random subsampling) and Gaussian Toeplitz matrices
(instead of Gaussian circulant matrices). In their proof they claimed that
1
Var dH sgn(Ψ(s) yi ), sgn(Ψ(s) yj ) . ′
m
by reasoning as follows. Let ak be the k-th row of Ψ(s) . They first showed (correctly)
′
that sgn(hak , yi i) and sgn(haℓ , yj i) are pairwise independent for any yi , yj ∈ Rn and
k 6= ℓ [14, Lemma 3.7]. They then wrote
′
Var dH sgn(Ψ
(s)
yi ), sgn(Ψ
(s)
m
1 X
1
yj ) = Var
sgn(ha
,y
i)6
=
sgn(ha
,y
i)
i
j
k
k
m′
k=1
=
1
(m′ )2
′
m
X
k=1
Var(1sgn(hak ,yi i)6=sgn(hak ,yj i) ) ≤
1
.
m′
Unfortunately, even though sgn(hak , yi i) and sgn(haℓ , yj i) are pairwise independent,
it is not true that
1sgn(hak ,yi i)6=sgn(hak ,yj i)
and
1sgn(haℓ ,yi i)6=sgn(haℓ ,yj i)
20
SJOERD DIRKSEN AND ALEXANDER STOLLENWERK
are pairwise independent for any yi , yj and k 6= ℓ.
Remark 3.5. Let us compare the running times of the four different binary embeddings for η being a constant. Note first that if
r
n r
o
n
n
3
(26)
log(N ) . δ
min
δ
log(n),
1
,
log(δ −1 )
log(δ −1 )
then the median fast binary embedding with FJLT can be computed in Õ(n log(n)),
where Õ hides log log-factors. Indeed, computation time of the FJLT is O(n log(n))
and matrix-vector multiplication for each of the B blocks can be computed in
O(n′ log(n′ )) = Õ δ −2 log(δ −1 ) max{δ −2 , log(N ) log(n)} .
Hence, total computation time is
O(Bn′ log(n′ ) + n log(n))
= Õ δ −2 log(δ −1 ) log(N ) max{δ −2 , log(N ) log(n)} + n log(n) = Õ(n log(n)),
√ p
if (26) is satisfied. In particular this holds if n & δ −6 and log(N ) . δ n/ log(1/δ).
In comparison, the accelerated Gaussian√ binary embedding with FJLT achieves
running time O(n log(n)), if log(N ) . δ 2 n.
Let us now consider the median fast binary embedding with the SJLT. The
running time of this embedding is
O(Bn′ log(n′ ) + δ −1 log(N )kxk0 )
= Õ(log(δ −1 )(δ −2 log2 (N ) + δ −4 log(N )) + δ −1 log(N )kxk0 ).
Thus, this embedding can achieve a (near-)linear embedding time in a larger range
of N compared
Gaussian binary embedding with√SJLT. Indeed,
√ to the accelerated
δ −6 , then the running time is Õ(n log n + nkxk0 ) which
if log N ≤ δ n and n & √
for the accelerated Gaussian
is near-linear if kxk0 ≤ n. In comparison,
√
√ binary
embedding we required log N ≤ δ 2 n to obtain linear running time on n-sparse
vectors.
In contrast to the running times for the embeddings involving the FJLT, the
running times for the SJLT-based embeddings can be even faster than O(n log n),
for instance if the number of vectors N is small and the vectors are sparse.
Let us finally note that the median fast binary embeddings with FJLT and
−4
−6
SJLT can both be computed
√ in time Õ(n log(n)) if n & δ (instead of n & δ ),
2
if additionally log N ≤ δ n (the condition for the accelerated Gaussian binary
embeddings) holds.
Remark 3.6. In the proof of Theorem 3.3, the optimal 1/m scaling in (19) is essential. In particular, it is not possible to
√ instead use the variance bound in (18).
It would be interesting to remove the 1/ n factor in (19). This would remove the
condition n & δ −6 from the discussion in Remark 3.5.
References
1. N. Ailon and E. Liberty, An almost optimal unrestricted fast Johnson-Lindenstrauss transform, ACM Trans. Algorithms 9 (2013), no. 3, Art. 21, 12. MR 3073515
2. Nir Ailon and Bernard Chazelle, The Fast Johnson–Lindenstrauss Transform and approximate nearest neighbors, SIAM J. Comput. 39 (2009), no. 1, 302–322.
3. Vladimir Braverman, Rafail Ostrovsky, and Yuval Rabani, Rademacher chaos, random Eulerian graphs and the sparse Johnson-Lindenstrauss transform, ArXiv:1011.2590 (2010).
4. Yu Cheng, Felix X. Yu, Rogério Schmidt Feris, Sanjiv Kumar, Alok N. Choudhary, and ShihFu Chang, Fast neural networks with circulant projections, ArXiv:1502.03436 (2015).
5. Mahdi Cheraghchi, Venkatesan Guruswami, and Ameya Velingker, Restricted isometry of
Fourier matrices and list decodability of random linear codes, SIAM J. Comput. 42 (2013),
no. 5, 1888–1914.
FAST BINARY EMBEDDINGS WITH GAUSSIAN CIRCULANT MATRICES
21
6. Anirban Dasgupta, Ravi Kumar, and Tamás Sarlós, A sparse Johnson-Lindenstrauss transform, Proceedings of the 42nd ACM Symposium on Theory of Computing (STOC), 2010,
pp. 341–350.
7. Michel X. Goemans and David P. Williamson, Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming, J. Assoc. Comput. Mach.
42 (1995), no. 6, 1115–1145. MR 1412228
8. Laurent Jacques, Jason N. Laska, Petros T. Boufounos, and Richard G. Baraniuk, Robust 1bit compressive sensing via binary stable embeddings of sparse vectors, IEEE Trans. Inform.
Theory 59 (2013), no. 4, 2082–2102. MR 3043783
9. Daniel M. Kane and Jelani Nelson, Sparser Johnson-Lindenstrauss transforms, Journal of the
ACM 61 (2014), no. 1, 4.
10. F. Krahmer and R. Ward, New and improved Johnson-Lindenstrauss embeddings via the
restricted isometry property, SIAM J. Math. Anal. 43 (2011), no. 3, 1269–1281. MR 2821584
(2012g:15052)
11. Samet Oymak, Near-optimal sample complexity bounds for circulant binary embedding,
ArXiv:1603.03178 (2016).
12. Yaniv Plan and Roman Vershynin, Robust 1-bit compressed sensing and sparse logistic regression: a convex programming approach, IEEE Trans. Inform. Theory 59 (2013), no. 1,
482–494. MR 3008160
13.
, Dimension reduction by random hyperplane tessellations, Discrete Comput. Geom.
51 (2014), no. 2, 438–461. MR 3164174
14. Xinyang Yi, Constantine Caramanis, and Eric Price, Binary embedding: Fundamental limits
and fast algorithm, Proceedings of the 32nd International Conference on Machine Learning,
2015, pp. 2162–2170.
15. Felix X. Yu, Aditya Bhaskara, Sanjiv Kumar, Yunchao Gong, and Shih-Fu Chang, On binary
embedding using circulant matrices, ArXiv:1511.06480 (2015).
16. Felix X. Yu, Sanjiv Kumar, Yunchao Gong, and Shih-Fu Chang, Circulant binary embedding,
Proceedings of the 31st International Conference on Machine Learning, 2014, pp. 946–954.
RWTH Aachen University, Lehrstuhl C für Mathematik (Analysis), Pontdriesch 10,
52062 Aachen, Germany
E-mail address: [email protected]
RWTH Aachen University, Lehrstuhl C für Mathematik (Analysis), Pontdriesch 10,
52062 Aachen, Germany
E-mail address: [email protected]
| 7cs.IT
|
Mantis: Predicting System Performance through Program Analysis and
Modeling
Byung-Gon Chun† , Ling Huang† , Sangmin Lee⋆ , Petros Maniatis† , Mayur Naik†
†
Intel Labs Berkeley, ⋆ University of Texas at Austin
arXiv:1010.0019v1 [cs.PF] 30 Sep 2010
Abstract
diction can answer what if questions — how system behavior changes when input workload changes or system
configuration changes [14, 25, 33].
Despite all these opportunities and demands, prediction
has not been in the mainstream. This is because it is very
difficult to predict metrics with high accuracy for current
practices — analytically modeling the system or treating
the system as a black box and generating a transfer function between input workload and output response. System
execution inherently depends on program semantics (i.e.,
internals of how the program works), thus prediction depends on program semantics. For example, certain metadata of programs (e.g., image resolution and depth) is a
cache of program semantics. One way to obtain program
semantics is to ask its details to its developers. In reality, however, this is not feasible due to the abundance and
complexity of programs. In this work, we aim to automatically extract program semantics without developers of the
program and use them to create better prediction models
for system performance (execution time). In particular, we
focus on predicting with different input workload in the
same environment (i.e., machine).
Mantis is a system that achieves this goal by combining
programming language and machine learning techniques
in a novel way (Section 2). Mantis consists of three key
components: feature instrumentation, model generation,
and code snippets generation for computing feature values when predicting. To capture program semantics without programmer assistance, we begin by extracting a potentially large number of program features that capture the
characteristics of program execution by running programs
instrumented with code analysis (Section 3). Next, we
use machine learning techniques to select important, relevant features to create the prediction models (Section 4).
Finally, program slicing computes small code snippets
that compute features needed by the model for prediction
(Section 5). This component also guides model generation
to choose features that can be computed cheaply.
We evaluate our system by applying the framework to
two applications, Lucene search engine and ImageJ, an
image processing applications (Section 6). We show that
Mantis can achieve more than 93% accuracy with less
than 10% training data set. We compare the model generated by Mantis with blackbox approaches that rely on
workload size and input configurations (e.g., commandline arguments) and show that Mantis can significantly
We present Mantis, a new framework that automatically
predicts program performance with high accuracy. Mantis integrates techniques from programming language and
machine learning for performance modeling, and is a radical departure from traditional approaches. Mantis extracts
program features, which are information about program
execution runs, through program instrumentation. It uses
machine learning techniques to select features relevant to
performance and creates prediction models as a function
of the selected features. Through program analysis, it then
generates compact code slices that compute these feature
values for prediction. Our evaluation shows that Mantis
can achieve more than 93% accuracy with less than 10%
training data set, which is a significant improvement over
models that are oblivious to program features. The system
generates code slices that are cheap to compute feature
values.
1 Introduction
Today’s programs are numerous and become more and
more complex. For example, services running in data centers are often large scale, and perform complicated operations depending on input workload. Predicting how applications will behave for given input workload is key to
helping users and operators better manage those applications.
Predicting metrics (e.g., performance, resource consumption) has great applications in many usage scenarios. First, prediction of execution time of a service request can be used for better workload management [18].
If the request is likely to violate service level agreements, the system can drop the request and allocate resources to other requests. Second, in scheduling applications such as MapReduce [15, 23], if we can predict execution time of tasks, we can then schedule jobs more optimally by considering where to map individual tasks to
candidate resources and perform speculative execution in
a timely fashion without spawning unnecessary processes.
Third, prediction can help with better resource provisioning [11, 12, 35] (e.g., how many servers should I use to
run this job? Should I add more servers?). Fourth, with
prediction, we can detect performance anomaly. If an operation takes much longer than a predicted time, we label
it an anomaly for troubleshooting purposes. Finally, pre1
Figure 1: Mantis architecture.
outperform these models. We explain how slicing can benefit model generation and the overhead of computing selected features. Finally, in Section 7 we discuss related
work, and conclude with future research directions in Section 8.
performance metrics? How do we model performance with relevant features?
3. How do we automatically generate code to compute
feature values for prediction?
We present Mantis, a new prediction architecture that
addresses the three questions above. There are three main
components, each of which addresses a key question.
2 Mantis Overview
2.1 Approach
2.2 Architecture
We address the problem of predicting performance metrics quickly without actually running the entire program.
Traditionally, researchers have taken a stance in two
camps for prediction — modeling systems analytically
(e.g., queuing theory), or treating the system as a black
box and creating a model between input (workload) and
output (performance). However, these approaches do not
work well due to their inherent limitations, i.e., the lack
of knowledge about the program. We take a new whitebox approach to generating prediction models. Unlike traditional approaches, simply put, we extract information
from execution of the program that contains a plethora of
information. In particular, we extract as many features as
possible from programs for given input data if extracting
features incurs little overhead, and rely on machine learning techniques to process the large amount of information dumped out. Machine learning techniques can infer
key features from voluminous information and construct
a robust model that predicts performance based on new
program features. In summary, our approach solves the
prediction system problems by combining programming
language and machine learning techniques in a novel way.
To achieve our goal, we need to address three key questions:
Figure 1 shows the Mantis architecture, a novel prediction
framework that combines programming language techniques with machine learning techniques. This architecture shows the offline part for generating prediction models.
Mantis consists of three major components: feature instrumentation and profiling, prediction model generation,
and feature evaluator generation. The feature instrumentor
analyzes the code of the program and automatically adds
instrumentation code that extracts program features. Then
the profiler runs this instrumented program with sample
input data to collect performance metrics and feature values. This profiling can generate a large number of features
within the budget of instrumentation overhead. Then, the
model generator runs machine learning algorithms to generate a prediction model, i.e., select a subset of key features that are relevant to the performance metrics and create a function of the selected features to predict the performance metrics with high accuracy. To use the model, we
need a way to compute feature values. The feature evaluator generator uses program slicing to automatically extract
small code snippets (which we call feature evaluators) that
compute feature values from the instrumented program.
1. What are good program features? How do we extract Ideally, for a feature evaluator, the technique includes only
program statements that affect the feature value of the
these feature values?
evaluator. Finally, there is a feedback loop from the slicer
2. Among many features, which ones are relevant to to the model generator. We may not be able to use some of
2
rationales of choosing these features below.
Loops When a program repeats computation, the execution time depends on how many times the program repeats. We introduce loop counts to capture this behavior. We instrument all loop constructs (e.g., while and
for) in the program. If there are nested loops, we add a
loop count for each loop. The following example shows a
nested loop. The outer loop performs reading a line from
a file, and the inner loop performs a search operation n
times.
Figure 2: Online predictor.
the features selected by the model generator. If a selected
feature is expensive to compute (e.g., we have to run the
entire program to compute the feature value), we reject
the feature by notifying the cost of computing the feature
value to the model generator. The prediction model generator creates a new model after excluding the rejected feature(s). This loop may run multiple times depending on
scenarios. When the program slicer can generate all the
feature evaluators of the selected features (cheaply), the
entire process ends, and the tool produces final features,
prediction model, and feature evaluators.
Once a prediction model is generated, it is used to forecast a performance metric of interest for a new input as
shown in Figure 2. The example has k feature evaluators.
The new input is sent to each feature evaluator to compute its feature value, and the prediction model computes
an estimate using all the feature values.
In the following, we explain these components in detail
and evaluate the system.
// original code
while(line=readLine()) {
for (int i=0; i<n; ++i)
search(line, i);
}
// instrumented code
while(line=readLine()) {
++mantis_loop_cnt1;
for (int i=0; i<n; ++i) {
++mantis_loop_cnt2;
search(line, i);
}
}
Method invocation Another way to repeat computation
is to use a recursive procedure. The execution time depends on how many times the program invokes recursive
methods. To capture this behavior, we introduce method
invocation counts, each of which is incremented when a
method is invoked. The following example shows an example program that traverses a tree structure and computes an aggregated metric.
3 Feature Instrumentation
We extract program features relevant to performance (e.g.,
execution time). We choose features with the following
goals in mind. First, the features should capture the behavior of program performance. Second, the features should
be accurate and easy to compute. For example, we avoid
relying on inaccurate timer resolution. Third, the features
should be collected with low overhead. We aim to run our
instrumented program to collect feature values and performance metrics at the same time instead of running the
original program to get performance metrics, running the
instrumented program to get feature values, and joining
the data. The latter is not accurate when the program has
non-determinism.
In the following, we first describe what program features we instrument and then present how to create instrumented programs.We use Java programs as examples, but
our techniques are generally applicable to other programming languages.
// original code
process(node n) {
if (cond) return;
process(n.l);
process(n.r);
compute(n);
}
// instrumented code
process(node n) {
++mantis_methodinv_cnt;
if (cond) return;
process(n.l);
process(n.r);
compute(n);
}
Branches Often the execution time changes depending
on which control flow path the program takes. This can
be captured by adding branch information. The follow3.1 Features
ing example shows that depending on the conditional, the
The features we choose are loop counts, branch counts, program takes two different paths with very different exeand variable values in different versions. We discuss the cution times.
3
// original code
if (flag) { lightweightCompute(); }
else { heavyCompute(); }
// instrumented code
if (flag) {
++mantis_branch_cnt1;
lightweightCompute();
} else {
++mantis_branch_cnt2;
heavyCompute();
}
// original code
try { compute(); }
catch (Exception e) { error(e); }
// instrumented code
try { compute(); }
catch (Exception e) {
++mantis_ex_cnt;
error(e);
}
path is not a common case the program takes, so the execution time of the program is likely to change significantly. We add an exception count for each exception handling part of the program to capture this behavior.
To collect these features in multi-threaded objectoriented programs, we need to summarize the features
across objects and across threads. We sum up loop count
and branch count across objects, and also keep a single
array of a variable for all objects created. To handle multiple threads we maintain a separate instrumentation object
that captures features per thread, and merge feature values at the end of program execution. For loop and branch
4 Prediction Modeling
counts, we sum up those counts across threads. For versions of variables, we compute the mean of each version
We add a branch counter for each branch in the pro- across threads.
gram. It counts how many times a particular branch is
taken. For example, if the program takes a particular 3.2 Instrumentor
branch once depending on a conditional, the counter value To instrument program to obtain program features, we
is either 1 or 0. If a branch is taken multiple times, its perform code analysis and transformation. In particular,
value reflects that.
we use source code analysis1 to construct abstract synVariable values We also instrument versions of variable tax trees (ASTs), and manipulate the constructed ASTs
values to characterize the program execution. We focus by adding new nodes representing loop counts, method
on primitive variables (short, int, long, float, double, char, invocation counts, branch counts, exception counts, and
and boolean variables) and collect the first k values when- variable versions to the trees.
We use the instrumented program to capture the exeever a variable is assigned to a value. Our intuition is that
cution
time of the original program as well as to capoften the variable values obtained from input parameters
ture
program
feature values. To achieve low overhead,
and configurations are changing infrequently, and these
we
employ
three
techniques. First, we perform selective
values tend to affect program execution by changing conprofiling
of
programs.
We focus on application programs
trol flow. We track both class field variables and local variand
do
not
instrument
system
libraries (e.g., toString() or
ables.
equals()).
Second,
we
use
a
procedure
that removes instruIn the following example, the execution time of the promentation
from
the
part
that
incurs
high
overhead until the
gram is dominated by the input argument n of compute(),
overall
instrumentation
overhead
is
below
our threshold
which comes from a preprocessed variable, which is done
(e.g.,
5%
of
the
original
program
execution
time). Third,
quickly. We collect the assigned value of n as one of our
to
avoid
synchronization
overhead
of
multiple
threads acprogram features.
cessing the instrumentation variables, we use a thread// original code
local data structure per thread, and merge data structures
n = preprocess();
of the threads at the end of the program execution.
compute(n);
After the instrumentation step, the profiler receives the
// instrumented code
instrumented program with test input data, runs the pron = preprocess();
gram with each input, and collects tuples, each of which is
mantis_n_data[cur_ptr++] = n;
program execution time and feature values for each input
compute(n);
data. The profiler then sends the tuples of (execution time,
Exception counts For certain inputs, the program may feature values) to the prediction model generator that pertake a control flow that throws and handles errors. This forms feature selection and model creation.
We instrument and profile programs to collect many features from program execution runs for modeling the performance metrics of the programs. However, we expect
that a small but relevant set of features may explain the execution time well, and hence seek a compact model, i.e.,
a function of this small set of features, that accurately estimates the execution time of the program. Among all the
1 We can also implement our instrumentation using bytecode analysis. With bytecode, we do not have local variable names we can refer to
unless source code is compiled with the Java compiler debug option.
4
because it tends to create models involving many feature
terms, if not all of them. This does not satisfy us since
we have a lot of features but desire only a small subset of
them to contribute to the model.
Regression with best subset selection finds for each
k ∈ {1, 2, . . . , m} the subset of size k that gives smallest residual sum of squares. However, it is a discrete optimization and is known to be NP-hard [21]. In recent
years a number of approaches based on model regularization have been proposed as efficient alternatives. Their
main idea is to add a regularization term to problem (1)
to control the complexity of the model, and make a tradeoff between the regression error and the number of features used in the model. Among them, a widely used one
is LASSO (Least Absolute Shrinkage
P and Selection Operator) [34], which uses quantity λ m
j=1 |βj | to penalize
problem (1). It effectively enforces many βj ’s to be 0, and
selects a small subset of features (indexed by non-zero
βj ’s) to build the model, which is usually compact and
has better prediction accuracy than models created by ordinary least square regression [21]. Parameter λ controls
the complexity of the model: as λ grows larger, fewer features are selected by the model.
Being a convex optimization problem is the greatest advantage of the LASSO method, and there exist fast algorithms to solve the problem efficiently even with largescale datasets [17, 24]. LASSO also has nice theoretical
and empirical properties, and under suitable assumptions,
4.1 Background
it can recover the true underlying model [16, 34]. In adLeast Square Regression Our feature instrumentation dition, LASSO can be easily extended to create nonlinear
procedure outputs n data samples as tuples of {yi , xi }ni=1 , models (e.g., using polynomial basis functions of the feawhere yi ∈ R denotes the ith observation of execution tures).
time, and xi denotes the ith observation of the vector
of m features. We use regression techniques to model 4.2 Our Procedure
the relationship between y and x, which assumes that We aim to use polynomial functions to model the execuyi ’s are generated from y = f (x, β) + ǫ, where β = tion time, so that we can clearly see what kinds of non[β0 , β1 , β2 , . . .] is a vector of weights to be determined linear terms on which features are important to the exefor the model, and ǫ is the white noise. Least square re- cution time. To capture nonlinear effects of and interacgression is a mathematical procedure for finding the best- tions between multiple features, we expand the features
fitting f (x, β) to a given set of responses yi by minimiz- x = [x1 x2 . . . xk ], k ≤ m to all the terms in the expansion of the degree-d polynomial (1 + x1 + . . . + xk )d , and
ing the sum of the squares of the residuals [21], i.e.,
use them to construct a multivariate polynomial function
n
X
f (x, β) for the regression. For example, using a degree-2
2
(yi − f (xi , β)) .
(1)
min
polynomial with feature vector x = [x1 x2 ], we expand
β
i=1
out (1 + x1 + x2 )2 to get terms 1, x1 , x2 , x21 , x1 x2 , x22 ,
If a linear function f (x, β) is used, we obtain linear least and use them as basis functions to construct the following
square regression, which can be easily extended to cre- function for regression:
ate nonlinear models by using nonlinear (e.g., polynomial,
f (x) = β0 + β1 x1 + β2 x2 + β3 x21 + β4 x1 x2 + β5 x22 .
spline, etc.) basis functions of features x.
information, not all of them are expected to be useful for
the model: Some of them may have no variability across
different inputs, some have very weak or even no correlation to the execution time, and others are redundant to
each other. However, we do not know which features are
useful, but would like to determine a small subset of features that is most relevant to predicting the execution time,
and are willing to sacrifice some of the small details in order to get the “big picture".
To make the problem tractable, we constrain our models to the multivariate polynomial family. We expect that a
good program should have polynomial execution time on
some (combination of) features, and a polynomial model
up to certain degree can approximate well any nonlinear
model (due to Taylor expansion). In addition, a compact
polynomial model that predicts execution time well can
provide an easy-to-understand explanation on what factors are important in determining the execution time of
the program, and then give program developers intuitive
feedback on the performance of the program.
In summary, what we need is an optimal strategy to produce a (nonlinear) model on a small set of features from
thousands of ones collected blindly. We rely on machine
learning techniques (specifically, sparse regression with
multivariate polynomial basis) to automatically infer this
small subset features and construct a compact model to
capture the dominant predictors of execution time.
Sparse Regression While widely used, least square regression has two major drawbacks: 1) When a large number of features exist, least squares tend to create complex
models and overfit the data, resulting in inferior prediction accuracy. 2) It is usually hard to interpret the results,
Because we neither know which features are needed nor
what kinds of nonlinear terms are necessary, an optimal
but naive approach is to expand the degree-d multivariate polynomial with all p features and use all the terms to
construct the regression function. However, this approach
5
// original code
int j;
while(line=readLine()) {
++mantis_loop_cnt1;
j = j + expensive_processing(line);
}
// sliced code on the variable
// mantis_loop_cnt1
while(line=readLine()) {
++mantis_loop_cnt1;
}
gives us m+d
terms, which is large when m is on the
d
order of thousands and even for small d, and will cause
heavy burden on the computing of the regression model.
Complete expansion on all features is not necessary, because many of them have little contribution to the execution time, and many of them are redundant to each other.
For efficient computation, we adopt a 3-step approach
for the feature selection and nonlinear model fitting:
Step 1: Use the linear LASSO algorithm to filter out
(many) features that hardly contribute to the execution
time. Although this step may be suboptimal (mainly due
to the non-linearity in the true underlying model), it is
cheap, fast and scalable, and is provably better than the
traditional feature selection methods that consider individual features one by one.
Step 2: Do degree-d multivariate polynomial expansion
on the features selected in step (1), and use all the terms
from the expansion as the basis functions for the nonlinear
model.
Step 3: Use the LASSO method on the expanded features
to pick out a subset of nonlinear terms to construct the
model.
With these three steps, we have developed an efficient
procedure to select a small set of nonlinear terms to construct a compact and intuitive model. Our experimental
results in Section 6 show our method can construct models to accurately predict execution time for a variety of
applications.
Figure 3: A slicing example.
different from most slicing research work motivated by
debugging (e.g., [32]) whose goal is to highlight as few
statements as possible that will aid the programmer debug
a particular problem. Thus, they elide the constraint in the
original slicing definition that the generated sub-program
be executable. To achieve executability, we need to solve
several engineering issues related to Java language features.
Our slicing algorithm operates on expressions e, which
may be of one of four kinds: a local variable v, a static
field (i.e., a global variable) g, an abstract instance field
hh, f i denoting instance field f of any object allocated at
site h, or an abstract array element h, denoting any element of any array object allocated at site h. Abstractions
of instance fields and array elements are required because
static analysis cannot refer to concrete object addresses.
Our slicing algorithm is not dependent upon the choice
of abstraction, however, can easily be modified to use abstractions besides object allocation sites.2
The slicer takes as input the given program and the slicing criterion c = he, pi, which is an expression e whose
value is desired at program point p, and produces as output
a corresponding slice. In our setting, e is always a static
field g instrumented by us (e.g., a loop counter), and p
is always the exit of a method of the program (e.g., the
program’s main method.)
Our slicing algorithm is based on two algorithms (one
from Horwitz, Reps, and Sagiv [22] and one from Reps,
Horwitz, Sagiv, and Rosay [28]). We summarize four
steps of the algorithm. First, for each method, we construct a Program Dependence Graph (PDG). Then, for the
entire program, we construct a System Dependence Graph
(SDG), which is a set of PDGs where additional edges
are created to capture interprocedural dependencies. We
augment the SDG with summary edges by running the
interprocedural data flow analysis algorithm in [27] to
solve context sensitivity problems. Finally, we run a 2pass reachability algorithm on the augmented SDG. We
explain individual steps more in detail below.
5 Feature Evaluator Generation
In this section we explain our feature evaluator generation component. To generate a feature evaluator, i.e., a
small code snippet that computes the value of a feature,
we use a program slicing algorithm. Given a program and
a slicing criterion, which is a program variable v at a program point p, static slicing [36] computes a slice, which
is an executable sub-program of the given program that
yields the same value of v at p as the given program, on
all inputs. The goal of static slicing is to yield as small
a sub-program as possible. Figure 3 shows a slicing example. The original code performs reading lines from a
file, executes expensive computation on each line, and
accumulates processed values. Suppose we want to extract the code part that affects the computation of the instrumented variable mantis_loop_cnt1. Ideally, the
slicer should produce the sliced code, shown in Figure 3
that captures only code that really affects the variable.
At a high level, our slicer captures intra-procedural and
inter-procedural data dependencies and control dependencies of a slicing criterion. The produced slices must be
executable since in our system the generated sub-program
will be executed online on the given input to obtain the result of the slicing criterion. This is a requirement clearly
2 The
choice of abstraction affects the precision and scalability of the
algorithm, and we found object allocation sites to strike a good tradeoff.
6
PDG and SDG Our slicing algorithm operates on
Joeq [6] quad code, an intermediate representation format based on registers. The vertices of a PDG represent
quad code instructions (e.g., statements and predicates).
The edges of a PDG represent data flow and control dependencies. An SDG also includes inter-procedural dependencies. A method call creates a call vertex and a set
of actual-in and actual-out vertices. Each parameter of a
method call creates an actual-in vertex, and a return value
creates an actual-out vertex. A method entry creates an entry vertex and a set of formal-in and formal-out vertices,
which correspond to arguments and a return value respectively. A call edge is created to connect a call vertex of
an call site to an entry vertex of the matching method. In
addition, a linkage-entry edge is created from an actual-in
vertex to a corresponding formal-in vertex, and a linkageexit edge is created to link a formal-out vertex to an actualout vertex.
the pass goes into the methods associated with the actualouts.
Slicing made practical A set of program statements
identified by the described algorithm may not meet Java
language requirements. This problem needs to be resolved
to create executable slices. We list a few of the engineering issues we addressed for that. First, we need to
handle accesses to static fields and heap locations (instance fields and array elements). Therefore, when building an SDG, we identify all such accesses in a method
and create formal-in vertices for those read and formalout for those written along with corresponding actual-in
and actual-out vertices. Second, there may be uninitialized parameters if they are not included in a slice. We
opt to keep method signatures, hence we initialize them
with default values. Third, there are methods not reachable from a main method but rather called from JVM directly (e.g., class initializers). These methods will not be
included in a slice by the algorithm but still may affect the
slicing criterion. Therefore, we do not slice out such code.
Fourth, when a new object creation is in a slice, a corresponding constructor invocation may not. To address this,
we create a control dependency between object creations
and corresponding constructor invocations to ensure that
they are also in the slice. Fifth, a constructor of a class except the Object class must include a call to a constructor
of its parent class. Hence we include such calls when they
are missing in a slice. Sixth, the first parameter of an instance method call is a reference to the associated object.
Therefore if such a call site is in a slice, the first parameter
has to be in the slice too and we ensure this.
Augmented SDG An SDG does not capture contextsensitivity of method calls. Therefore, if a call site of
a method is included in a slice, other call sites of the
same method may be included even though they do not
affect a slicing criterion. To remedy this problem, we
build an augmented SDG by adding summary edges to an
SDG. A summary edge connects an actual-in vertex to an
actual-out vertex and summarizes the effect of the actualin on the actual-out of a method call. To create summary
edges, we use the backward RHS algorithm [27]. It propagates from formal-out vertices of the method based on
data- and control- dependencies to calculate path edges
of the method. A path edge is of the form hp, e1 i →
hpf ormal−out , e2 i meaning that e1 at program point p affects e2 at pf ormal−out : it always ends with a formal-out
of the method and is created within the method. Therefore,
when there exists a path edge from a formal-in vertex to a
formal-out vertex of a method, a summary edge is created
connecting corresponding actual-in vertex and actual-out
vertex of a call site to the method.
Final step Previous steps we described so far generate
a slice of Joeq quad code. To generate the final Java byte
code we can execute, we translate the Joeq quad code to
Jasmin [4] assembly code and use the Jasmin assembler to
generate Java byte code. We take a simple approach that
translates each quad instruction to a corresponding set of
byte codes. During the process, since we do not have complete information on ordering between basic blocks, we
add an explicit goto instruction at the end of each basic
block. However this may lead to a cycle if a conditional
branch in a loop is sliced out and replaced by goto. We
ensure that no cycle is created by performing a DFS-like
search and choosing a successor as a target of the goto
instruction only if it can reach the exit of the method. Another special case is JSR instruction that pushes the address of the next immediate opcode into an operand stack
as its return address. However the next instruction may
not be the same as one in the original program. Hence we
add an extra goto with an appropriate target after the JSR
operation. Our current translator is not optimized; we plan
to optimize the use of stacks if needed in the future.
2-pass algorithm To identify statements to include in a
slice, we run a 2-pass reachability algorithm on an augmented SDG. The first pass starts from the program point
of a given slicing criterion and goes backwards along all
the edges in the augmented SDG but not along linkageexit edges. As a result, when encountering a call site, the
first pass does not go into the method body but uses summary edges of the call site. The second pass starts from
all actual-out vertices visited in the first pass and traverses
backwards using all the edges but not using linkage-entry
and call edges. This pass covers all methods that correspond to call sites identified in the first pass. In addition,
it may find more call sites while traversing the body of
the methods and use summary edges; this process adds
additional actual-out vertices of the summary edges and
7
Execution Time (sec.)
1.4
Discussion There are static and dynamic program slicing algorithms. They have tradeoffs between input cover1.2
age and slice compactness. Static slicing works for all in1
puts, but it may produce a bigger slice than dynamic slic0.8
ing. Dynamic slicing includes only code that is actually
0.6
executed for given inputs, but it does not cover all inputs.
As a starting point, we chose static slicing as it guarantees
0.4
to work for all inputs, but in the future we plan to explore
0.2
dynamic slicing or hybrid slicing that combines static slic0
ing and dynamic slicing if we need to improve slice com0
0.2
0.4
0.6
0.8
Predicted Time
pactness. In this paper, we tested our static slicing algorithms with simple programs, and we plan to evaluate the
Figure 4: For Lucene, the blackbox approach fails to prescalability of our algorithms with complicated programs.
dict execution time accurately.
6 Evaluation
We have implemented Mantis that works with Java programs by extending existing machine learning and program analysis tools. We built the feature instrumentor atop
Eclipse JDT AST libraries. JDT is a toolkit that provides
APIs to access and manipulate Java source code. We add
visitors to ASTs to add instrumentation code. The basic instrumentation variables are thread local variables.
The instrumentor also introduces static global variables
that summarize feature values across threads and are used
for slicing. We implemented our modeling procedure in
Matlab. Finally, we extended JChord [5], a static and dynamic Java program analysis tool, to implement our program slicing algorithm in Java and Datalog. In the current version, we have to patch models for native library
functions manually to track dependencies inside native library functions. The JChord slicer produces Joeq quadcode slices. To create final executable bytecode slices, we
implemented a translator from quadcode to bytecode using Jasmin [4].
To evaluate our system, we choose two applications,
Lucene Search [8] and ImageJ [3], that involve intensive
computation. We evaluate the prediction accuracy of our
system in terms of prediction error (i.e., prediction accuracy = 1 - prediction error) and compare it with blackbox
approaches. Prediction error is computed using the equation:
prediction error =
Execution Time (sec.)
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
Predicted Time (sec.)
1
1.2
Figure 5: Predicted time vs. execution time of the Lucene
search application. Using 10% of Lucene search data for
training, our system can predict execution time with less
than 7% prediction error.
6.1 Prediction Results for Lucene
After profiling our Lucene search application with various
text input queries over a corpus of the works of Shakespeare and the King James Bible, we obtain a dataset with
3840 samples, each of which consists of 1 execution time,
9 loop features, 29 branch features, and 90 variable features from 18 variables (we record 5 versions of values
for each variable). So we obtain a dataset with 1 column
of execution time and 126 columns of features, subset of
which would hypothetically explain the execution time
well. In the prediction modeling process, we normalize
each column of values into range [0,1], and randomly partition data (row) samples into training set and testing set.
We first evaluate a blackbox approach for predicting execution time. We choose one that can construct compact nonlinear models using the command
line arguments as features (instead of using the feature data from our profiler), which consist of the
following ones: raw, threads, totalqueries,
hitsperpage, repeat, denoted by x1 , x2 , x3 , x4
and x5 , respectively. We build models using either ordinary least-square regression or LASSO with a function us-
|predicted time - actual time|
.
actual time
We show the sensitivity to the size of training data and the
regularization parameter λ, and the results of prediction.
Traditional blackbox approaches fail to predict execution
time with low prediction error, but our system can construct an online predictor that can predict execution time
accurately. Note that, in presenting the results of our system, we use only features that can be compuated cheaply
by iterating over feature selection and program slicing for
rejecting expensive features.
8
2
0
0
0.05
λ
Prediction Error
Number of Features
Prediction Error
4
0.1
0.1
0.1
0.05
0.05
0
0
0.05
λ
0
0
0.1
0.1
0.2
0.3
0.4
0.5
Percentage of Training data
0.6
(a)
(b)
Figure 6: Using Lucene search data, we show in (a) that our approach is insensitive to parameter λ: there are a range
of λ ∈ (0, 0.07] that result in similar models for accurate prediction, and show in (b) that our approach is insensitive
to the size of training data: even using 5% or less of data for training, our system can create models achieving accurate
prediction for execution time.
ing all terms in the expansion of (1 + x1 + x2 + x3 + x4 +
x5 )3 . However, in either case we consistently see more
than 38% prediction error, even when we: 1) (randomly)
sample different portions of data for training, 2) vary the
size of training data (for model regression) from 10% to
40%, 3) use polynomial functions with order higher than
3, and 4) use different subsets of the 5 features. Figure 4
shows predicted execution time vs. actual execution time.
As you can see, this blackbox approach derived from command line features fails to model and predict execution
time accurately.
From our detailed analysis, two features that have the
largest correlation with the execution time are feature-1,
totalqueries, which has a fair amount of correlation with execution time, and feature-2, thread, and the
remaining features are poorly correlated with execution
time. Despite some correlation in feature-1 the model derived from command line features are not enough for predicting execution time accurately: for each value of the
predicted time (on x-axis), there are dramatically different actual execution times (on y-axis) correspond to (an
ideal prediction is a 45 degree line pass through the origin). This result indicates that some other factors that are
not captured by the features should contribute to the execution time. On the contrary, as shown in the following,
our system can automatically select higher quality features from the program, and construct nonlinear models
to predict execution time accurately.
To evaluate our system, we start with its sensitivity to λ,
the parameter for trading off the prediction error with the
number of selected features. We use 10% of data for training (both feature selection and model fitting), and trace a
variety of λ values (which may result in different subset of
selected features, thus different models) using an efficient
algorithm proposed in [17]. We show the result in Figure 6 (a). To our surprise, we see that our method is able
to select 2-4 features (out of 126 in total) that are enough
to build a nonlinear model to predict the execution time
within 7% error. As expected, starting from a very small
λ value and increasing it, our method selects decreasing
number of features (from 4 down to 1), and consequently
results in models with decreasing prediction power. We
clearly see that there is a range of λ values (e.g., (0, 0.07])
that enable our method to select the right set of features
and build models for accurate prediction. Repeating the
experiment with different (random) training samples, and
with 20%, 30% and 40% of data for training, we see very
similar behavior. We conclude that our method is insensitive to the parameter λ, and setting λ ≤ 0.07 allows us to
select right features, and construct compact and accurate
models for predicting execution time.
Fixing λ = 0.033 , we study the sensitivity of our
method to the size of training data, and plot the result in
Figure 6 (b). We see that with different sizes of training
data, prediction errors of the constructed models are fairy
stable, and even using 5% or less training data, our method
is able to produce accurate models for predicting execution time.
To reveal more details of the model, we use λ = 0.03
and 10% of data for training, to investigate which features
are selected and what kinds of models are constructed.
We find that our algorithm usually selects 3-4 features
(depending on which subset of data are sampled for training) and constructs a model with a prediction error around
6.1%. Figure 5 shows predicted execution time vs. actual
execution time of our system. In one instance of modeling,
the following 4 features are selected: 1) loop feature l2 re3 An optimal λ can be determined by a cross-validation approach,
e.g., further partitioning the training data into two sets, one for feature
selection and model regression, and another for testing the model. An
optimal λ is the one giving the smallest testing error (on the part of
training data selected for testing).
9
3
Execution Time (sec.)
2.5
2
1.5
1
0.5
0
0
0.5
1
1.5
2
Predicted Time
2.5
3
Figure 7: Predicted time vs. execution time of the ImageJ
application for a blackbox approach.
3
Execution Time (sec.)
2.5
2
1.5
1
0.5
0
0
0.5
1
1.5
Predicted Time (sec.)
2
2.5
Figure 8: Predicted time vs. execution time of the ImageJ
application. Using 10% of ImageJ data for training, our
system can predict execution time with 5.5% prediction
error.
lated to a while loop for reading keywords from the query
file, 2) variable feature v3 related to totalQueries,
3) variable feature v5 related to hitsPerPage, and 4)
variable feature v9 related to how many query processors
to create per thread. Among them, features l2 and v9 have
the largest weights (indicating they are the most important) and persistently appear when sampling different portions of the data for training. With just these two features,
we do a LASSO sparse regression with all basis functions
of features in the expansion of (1+l2 +v9 )3 . Interestingly,
we are able to construct the following a nonlinear model
f (l2 , v9 ) = 0.1 + 0.52l2 + 0.09v9 − 0.69l22 − 0.07v92 +
+1.16l23 + 0.13l1v92 ,
which can predict execution time with error 6.7% (indicating the rest two selected features v3 and v5 only contribute
to less than 1% of the prediction accuracy).
6.2 Prediction Results for ImageJ
ImageJ [3] is a public domain Java image processing and analysis program. It provides a variety tools
for displaying, editing, analyzing, and processing im10
ages in many formats. We test a dozen of tools
of ImageJ, including Smooth, Find Edges, FFT,
Find Maxima, etc. We choose to profile and predict
the execution time of Find Maxima, because it exhibits
high variance in execution time when processing different
images (even with similar size), making it a challenging
task to model the execution time.
To profile the Find Maxima, we use 3045 images
from popular vision corpus of Caltech 101 [1], Event
Dataset [2] and PASCAL challenge 2008 dataset [9]. The
images vary a lot in size and resolution, and have content
in different scenes (e.g., in the office, on the street, in the
natural environment, etc) and with different object categories (e.g., plan, car, bird, building, etc). After the profiling, we obtain a dataset with 3045 samples, each of which
consists of one execution time, 291 loop features, 2935
branch features, and 2290 variable features from 458 variables (we record five versions of values for each variable).
So we obtain a dataset with one column of execution time
and 5516 columns of features. After removing constant
and redundant columns, we obtain 182 useful features,
(small) subset of which would likely explain the execution
time well. In the experiments, we normalize each column
of values into range [0,1], and randomly partition the data
into training set and testing set.
For a blackbox approach, many methods can be used.
We consider one with the execution time as a nonlinear
function of a simple input parameter – the image size. We
start with a degree 3 polynomial function of the image size
x, and obtain the following prediction model using 20%
of data for training
f (x) = 0.1 + 2.18x − 8.77x2 + 36.6x3 .
Predicting on the remaining 80% test data, we consistently see more 35% of prediction error regardless which
subset of data are sampled as a training set (may result in
slightly different models). We see similar results when we
increase the degree of the polynomial and the percentage
of training data.
We plot the execution time against the predicted execution time obtained from the model in Figure 7. We clearly
see that image size alone is not enough for predicting
execution time regardless of whatever model may come
out: the image size is poorly correlated with the execution
time, and for each value of the image size or the predicted
execution time, there are dramatically different actual execution times corresponding, indicating that some other
factors should contribute to the execution time.
On the contrary, our system can automatically select
two high-quality features from thousands of automatically
instrumented features, and construct a model to predict
execution time accurately. Of the two selected features,
one is the variable feature related to the width of region
of interest (denoted by w); the other is height of the im-
1
2
Features selected by
the model generator
loop features l3 and l7
variable features v3 , v5 , and v9
loop feature l2
variable features v3 , v5 , and v9
1
Rejected
features
loop feature
l3
NONE
0.8
Cumulative probability
Step
Table 1: Iterative procedure of model selection considering the cost of computing feature values. In the example,
loop feature l3 is related to a for loop printing search results, and loop feature l7 is related to a while loop counting how many times queries are executed. We explained
other features in Section 6.1. Computing l3 requires computing the most of the program (i.e., doing actual lookups
of indices), thus it is rejected in step 1. This iterative
procedure stops at step 2 since all selected features are
quickly computable.
age (denoted by h). Although highly correlated to the execution time, neither a single feature (even with nonlinear
model), nor the linear combination of both selected features can predict execution time very well. Instead, using
10% of data for training with λ = 0.03, we do a LASSO
sparse regression using all (nonlinear) terms in the expansion of (1 + w + h)3 , and obtain the following model
f (w, h) = 0.1 + 0.08w + 0.07h + 0.33wh + 0.02h2 ,
which can predict the execution time accurately (around
5.5% prediction error), as shown in Figure 8.
We also study the sensitivity of our method to λ using
10% of data for training (both feature selection and model
fitting), and show the result in Figure 9 (a). Again, we see
that our method is insensitive to λ, and there are a wide
range of λ values (e.g., (0, 0.11]) that allow us to select
right features, and construct compact and accurate models
for predicting execution time.
Fixing λ = 0.03, we study the sensitivity of our method
to the size of training data, and plot the result in Figure 6
(b). Again, we see that with different sizes of training data,
prediction errors of the constructed models are fairy stable, and even using 5% or less training data, our method is
able to produce accurate models for predicting execution
time.
6.3 Benefit of Slicing
In this section, we evaluate the benefit of slicing. We explain how slicing can help choose features that are not
expensive to compute and evaluate the execution times
of feature evaluators of selected features computed manually.
We look at the details on how the Lucene search application chose final features we presented in Section 6.1. Table 6.3 shows the steps taken by the model generator due
11
0.6
0.4
0.2
0
0.001
Entire program
Loop feature l2
0.01
0.1
Execution time (s)
1
10
Figure 10: CDF of execution time of the entire Lucene
search program and time to execute the slice that computes loop feature l2 .
to the feedback from the slicer. In each step, we show the
features selected by the model generator, and the features
rejected by the slicer among the features passed from the
model generator. For this application, at step 2, the slicer
can compute slices that can quickly compute all the features needed by the model, thus it accepts the selected features and the feedback loop from the slicer to the model
generator ends.
Figure 10 shows the cumulative distribution function
(CDF) of execution time of the entire Lucene search program and that of the slice to compute loop feature l2 . We
show only l2 because it is the most expensive selected feature to compute since the slice goes through files to count
keywords. The other variable features are derived by arithmetic operations and assignments of values from inputs.
Computing l2 takes 3 − 4% of the entire program execution time, thus the prediction model can compute an estimate of execution time with low overhead.
7 Related Work
Prediction has been explored in multiple different contexts — database, cluster and cloud, networking, and program complexity modeling. In this paper, we presented a
new performance prediction framework for generic programs by combining programming language and machine
learning techniques. As far as we know, our work is the
first to explore program analysis to extract features, employ machine learning to create accurate models with selected features, and use program slicing to automatically
produce code snippets that compute feature values for prediction.
In the database, researchers explored machine learning algorithms to predict database query execution time.
Gupta, Mehta, and Dayal [20] used a variant of decision
trees to predict time ranges of data warehouse queries.
Ganapathi et al. [18] used KCCA to predict time and resource consumption of database queries (number of I/Os
and messages) using the statistics of query texts and query
plans (e.g., instance count for each possible database op-
4
2
0
0
0.05
0.1
λ
Prediction Error
Number of Features
Prediction Error
6
0.15
0.1
0.1
0.05
0.05
0
0
0.05
0.1
λ
0
0
0.15
0.1
0.2
0.3
0.4
0.5
Percentage of Training data
0.6
(a)
(b)
Figure 9: Using ImageJ data, we show in (a) that our approach is insensitive to λ: there are a range of λ ∈ (0, 0.11] that
result in similar models for accurate prediction, and show in (b) that our approach is insensitive to the size of training
data: even using 5% or less of data for training, our system can create models achieving accurate prediction.
erator).
In resource allocation and provisioning for cluster and
cloud applications, research has been done to forecast how
much resource is required to support given workload to
meet service level agreements [11, 12], or how long it
takes to complete a candidate assignment [31]. The models used resource consumption or workload size for prediction. Xu et al. used console logs — coarse-grained program status reports — to detect anomalies in the Hadoop
Distributed File System and the Darkstar game [37].
For load shedding in network monitoring applications,
Barlet-Ros et al. [10] used a simple linear regression
model of features from five packet header tuples, number of bytes, number of packets to predict CPU resource
usage. This work is specific to packet processing applications. In contrast, our framework is applicable to generic
programs.
In the networking context, multiple projects addressed
the problems of predicting response time changes for
what-if scenarios. WebProphet [25] predicts the impact
of certain optimizations of web services before deploying them by extracting web object dependencies with injected delays and simulating web page loading processes
with web object dependency graphs. WISE [33] predicts
the effects of configuration or deployment changes in
content distribution networks by modeling the network
dependency structure to response-time distribution. Link
Gradients [14] predicts the impact of network latency in
multi-tier systems by doing delay injection and performing spectral analysis.
There have been studies on using information from
execution traces for modeling computational complexity [19], simulating hardware platforms efficiently [29,
30], and finding bugs cooperatively [26]. In contrast to
these, Mantis focuses on creating a model for predicting
program execution time by computing feature values online with slices quickly for new inputs.
12
Trendprof [19] models asymptotic computational complexity by measuring empirical computational complexity. It computes a model that estimates the performance of
a program by modeling basic block execution frequency
in terms of user-specified features (e.g., input size) and
summarizing the program with clusters of basic blocks.
SimPoint [29, 30] finds a subset of execution instruction traces of program for an input for efficient hardware
platform simulation because simulating hardware for the
entire program execution takes too long time. It instruments basic block vectors in each fixed interval and uses
a clustering algorithm to extract a representative subset of
traces from clusters that approximates low-level hardware
metrics such as instructions per cycle, percent RUU occupancy, cache miss rate, branch prediction miss rate, and
address prediction miss rate [30].
Cooperative bug isolation (CBI) [26] used three predicates — branches with four values (always true, always
false, sometimes true and sometimes false, unreachable in
the run), comparisons between all pairs of integer-valued
variables and constants in the program, and comparisons
between integer valued return results of functions and 0.
CBI aims to find which predicates are correlated with
crashes to find bugs, and uses sampling of predicates to
lower runtime overhead since the executed runs are collected from end users of the program.
8 Conclusion
In this paper, we presented Mantis, a new prediction
framework that extracts program features using code analysis, models performance with these features using sparse
regression, and generates code snippets that compute the
feature values. We take a first step towards building such
a framework. Our prototype evaluation shows that Mantis can predict execution time with more than 93% accuracy for the applications we tested, a search engine and an
[16] D. Donoho. For most large underdetermined systems of equations, the minimal 1-norm solution is
the sparsest solution. Communications on Pure and
Applied Mathematics, 59:797âĂŞ829, 2006.
[17] B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani. Least angle regression. Annals of Statistics,
32(2):407–499, 2002.
[18] A. Ganapathi, H. Kuno, U. Dayal, J. L. Wiener,
A. Fox, M. Jordan, and D. Patterson. Predicting multiple metrics for queries: Better decisions enabled by
machine learning. In ICDE, 2009.
[19] S. Goldsmith, A. Aiken, and D. Wilkerson. Measuring empirical computational complexity. In FSE,
2007.
[20] C. Gupta, A. Mehta, and U. Dayal. PQR: Predicting query execution times for autonomous workload
management. In ICAC, 2008.
[21] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Springer, 2009.
[22] S. Horwitz, T. Reps, and D. Binkley. Interprocedural
slicing using dependence graphs. In PLDI, 1988.
[23] M. Isard, V. Prabhakaran, J. Currey, U. Wieder,
References
K. Talwar, and A. Goldberg. Quincy: fair scheduling
for distributed computing clusters. In Proceedings of
[1] Caltech
101
Object
Categories.
SOSP’09, 2009.
www.vision.caltech.edu/Image_Datasets/Caltech101/Caltech101.html.
[24]
S.-J.
Kim, K. Koh, M. Lustig, S. Boyd, and
[2] Event Dataset. vision.stanford.edu/lijiali/event_dataset.
D.
Gorinevsky.
An interior-point method for large[3] ImageJ. rsbweb.nih.gov/ij/.
scale
l1-regularized
least squares. IEEE Journal on
[4] Jasmin. jasmin.sourceforge.net.
Selected
Topics
in
Signal
Processing, 1(4):606–617,
[5] jchord. code.google.com/p/jchord.
2007.
[6] Joeq. joeq.sourceforge.net.
[25] Z. Li, M. Zhang, Z. Zhu, Y. Chen, A. Greenberg, and
[7] LLVM. llvm.org.
Y.-M. Wang. WebProphet: Automating performance
[8] Mahout. lucene.apache.org/mahout.
prediction for web services. In NSDI, 2010.
[9] Visual
Object
Classes
Challenge
2008.
[26]
B. Liblit, M. Naik, A. X. Zheng, A. Aiken, and M. I.
pascallin.ecs.soton.ac.uk/challenges/VOC/voc2008.
Jordan. Scalable statistical bug isolation. In PLDI,
[10] P. Barlet-Ros, G. Iannaccone, J. Sanjuas-Cuxart,
2005.
D. Amores-Lopez, and J. Sole-Pareta. Load shed[27]
T.
Reps, S. Horwitz, and S. Sagiv. Precise interproding in network monitoring applications. In USENIX
cedural
dataflow analysis via graph reachability. In
Annual Tech. Conf., 2007.
POPL,
1995.
[11] P. Bodik, R. Griffith, C. Sutton, A. Fox, M. I. Jordan, and D. A. Patterson. Automatic exploration of [28] T. W. Reps, S. Horwitz, S. Sagiv, and G. Rosay.
Speeding up slicing. In FSE, 1994.
datacenter performance regimes. 2009.
[29]
T. Sherwood, E. Perelman, and B. Calder. Basic
[12] P. Bodik, R. Griffith, C. Sutton, A. Fox, M. I. Jordan,
block distribution analysis to find periodic behavior
and D. A. Patterson. Statistical machine learning
and simulation points in applications. In Parallel Armakes automatic control practical for internet datchitectures and Compilation Techniques, 2002.
acenters. 2009.
[30]
T. Sherwood, E. Perelman, G. Hamerly, and
[13] C. Cadar, D. Dunbar, and D. Engler. KLEE: UnasB. Calder. Automatically characterizing large scale
sisted and automatic generation of high-coverage
program behavior. In ASPLOS, 2002.
tests for complex systems programs. In OSDI, 2008.
[31]
P.
Shivam, S. Babu, and J. S. Chase. Learning appli[14] S. Chen, K. Joshi, M. A. Hiltunen, W. H. Sanders,
cation
models for utility resource planning. In ICAC,
and R. D. Schlichting. Link gradients: Predicting the
2006.
impact of network latency on multitier applications.
[32] M. Sridharan, S. Fink, and R. Bodik. Thin slicing.
In INFOCOM, 2009.
In PLDI, 2007.
[15] J. Dean and S. Ghemawat. MapReduce: Simplified
[33]
M.
Tariq, A. Zeitoun, V. Valancius, N. Feamster, and
data processing on large clusters. In OSDI, 2004.
image processing application, which cannot be achieved
with models without program features. In the future, we
plan to evaluate our system with various complicated applications in terms of accuracy, applicability, and scalability.
Our new approach to prediction presents several exciting research directions we want to explore. First, we want
to extend our model to include environment and to explore more sophisticated features (e.g., feature values that
depend on calling contexts) and more sophisticated slicing algorithms (e.g., algorithms based on dynamic control
flow graphs). Second, we would like to build our framework for C/C++ languages with LLVM [7] since the current prototype works with Java programs. Third, we want
to further extend our framework to apply to networked
systems running on multiple nodes. Our work in this paper
addressed single-machine program execution. Finally, we
also would like to apply the tool to performance debugging (e.g., a tool that generates test cases for performance
debugging similar to KLEE [13] that generates test cases
for correctness).
13
[34]
[35]
[36]
[37]
M. Ammar. Answering what-if deployment and configuration questions with wise. In ACM SIGCOMM,
2008.
R. Tibshirani. Regression shrinkage and selection
via the lasso. J. Royal. Statist. Soc B., 1996.
B. Urgaonkar and A. Chandra. Dynamic provisioning of multi-tier internet applications. In Proceedings of the Second International Conference on
Automatic Computing (ICAC’05), Washington, DC,
2005.
M. Weiser. Program slicing. In ICSE, 1981.
W. Xu, L. Huang, A. Fox, D. Patterson, and M. Jordan. Detecting large-scale system problems by mining console logs. In SOSP, 2009.
14
| 6cs.PL
|
arXiv:1408.6821v3 [math.CO] 19 May 2016
Looking for vertex number one
Alan Frieze∗, Wesley Pegden†
Department of Mathematical Sciences,
Carnegie Mellon University,
Pittsburgh PA 15213,
U.S.A.
May 20, 2016
Abstract
Given an instance of the preferential attachment graph Gn = ([n], En ), we would like
to find vertex 1, using only ‘local’ information about the graph; that is, by exploring the
neighborhoods of small sets of vertices. Borgs et al. gave an an algorithm which runs in
time O(log4 n), which is local in the sense that at each step, it needs only to search the
neighborhood of a set of vertices of size O(log4 n). We give an algorithm to find vertex 1,
which w.h.p. runs in time O(ω log n) and which is local in the strongest sense of operating
only on neighborhoods of single vertices. Here ω = ω(n) is any function that goes to infinity
with n.
1
Introduction
The Preferential Attachment Graph Gn was first discussed by Barabási and Albert [2] and then
rigorously analysed by Bollobás, Riordan, Spencer and Tusnády [3]. It is perhaps the simplest
model of a natural process that produces a graph with a power law degree sequence.
The Preferential Attachment Graph can be viewed as a sequence of random graphs G1 , G2 , . . . , Gn
where Gt+1 is obtained from Gt as follows: Given Gt , we add vertex t + 1 and m random edges
{ei = (t + 1, ui ) : 1 ≤ i ≤ m} incident with vertex t + 1. Here the constant m is a parameter
of the model. The vertices ui are not chosen uniformly from Vt , instead they are chosen with
probabilities proportional to their degrees. This tends to generate some very high degree vertices,
∗
†
Research supported in part by NSF grant CCF1013110
Research supported in part by NSF grant DMS
1
compared with what one would expect in Erdős-Rényi models with the same edge-density. We
refer to u1 , u2 , . . . , um as the left choices of vertex t + 1. We also say that t + 1 is a right neighbor
of ui for i = 1, 2, . . . , m.
We consider the problem of searching through the preferential attachment graph looking for
vertex number 1, using only local information. This was addressed by Borgs, Brautbar, Chayes,
Khanna and Lucier [5] in the context of the Preferential Attachment Graph Gn = (Vn , En ). Here
Vn = [n] = {1, 2, . . . , n}. They present the following local algorithm that searches for vertex 1,
in a graph which may be too large to hold in memory in its entirety.
1: Initialize a list L to contain an arbitrary node u in the graph.
2: while L does not contain node 1 do
3:
Add a node of maximum degree in N(L) to L od;
4: return L
Here for vertex set L, we let N(L) = {w ∈
/ L : ∃v ∈ L s.t. {v, w} ∈ En }.
They show that w.h.p. the algorithm succeeds in reaching vertex 1 in O(log4 n) steps. (We
assume that an algorithm can recognize vertex 1 when it is reached.) In [5], they also show how
a local algorithm to find vertex 1 can be used to give local algorithms for some other problems.
We also note that Brautbar and Kearns [6] considered local algorithms in a more general context.
There the algorithm is allowed to jump to random vertices as well as crawl around the graph in
the search for vertices of high degree and high clustering coefficient.
We should note that, as the maximum degree in Gn is n1/2−o(1) w.h.p., one cannot hope to have a
polylog(n) time algorithm if we have to check the degrees of the neighbors as we progress. Thus
the algorithm above operates on the assumption that we can find the highest-degree neighbor of
a vertex in O(1) time. This would be the case, for example, if the neighborhood of a vertex is
stored as a linked-list which is sorted by degrees. In the same situation, we can also determine the
K highest degree neighbors of a vertex in constant time for any constant K, and in the present
manuscript we assume such a constant-time step is possible. In particular, in this setting, each
of steps 2-7 of the following Degree Climbing Algorithm takes constant time.
We let dn (v) denote the degree of vertex v ∈ Vn .
Algorithm DCA:
The algorithm generates a sequence of vertices v1 , v2 , . . . , until vertex 1 is reached.
Step 1 Carry out a random walk on G until it is mixed; i.e., until the variation distance between
the current vertex and the steady state is o(1). We let v1 be the terminal vertex of the
walk. (See Remark 1.1 for comments on this step.)
Step 2 t ← 1.
2
Step 3 repeat
Step 4
Let Ct = w1 , w2, . . . , wm/2 be the m/2 neighbors of vt of largest degree.
(In the case of ties for the m/2th largest degree, vertices will be placed randomly in Ct in
order to make |Ct | = m/2. Also m is large here and we could replace m/2 by ⌊m/2⌋ if m
is odd without affecting the analysis by very much.)
Step 5
Choose vt+1 randomly from Ct .
Step 6
t ← t + 1.
Step 7 until dn (vt ) ≥
arbitrary.
n1/2
log1/100 n
(SUCCESS) or t > 2ω log4/3 n (FAILURE), where ω → ∞ is
Step 8 Assuming Success, starting from vT , where T is the value of t at this point, do a random
1/2
walk on the vertices of degree at least logn1/20 n until vertex 1 is reached.
Remark 1.1. It is known that w.h.p. the mixing time of a random walk on Gn is O(log n), see
Mihail, Papadimitriou and Saberi [11]. So we can assume that the distribution of v1 is close to
n (v)
the steady state πv = d2mn
.
Note that Algorithm DCA is a local algorithm in a strong sense: the algorithm only requires
access to the current vertex and its neighborhood. (Unlike the algorithm from [5], it does not
need access to the neighborhood of the entire set Pt = {v1 , . . . vt } of vertices visited so far.) Our
main result is the following:
Theorem 1.2. If m is sufficiently large then then w.h.p. Algorithm DCA finds vertex 1 in Gn
in O(ω log n) time.
DCA is thus currently the fastest as well as the “most local” algorithm to find vertex 1. We
conjecture that the factor ω in the running time is unnecessary.
Conjecture 1.3. Algorithm finds vertex 1 in Gn in O(log n) time, w.h.p.
We note that w.h.p. the diameter of Gn is ∼
execution time much below O(log n).
log n
log log n
and so we cannot expect to improve the
The bulk of our proof consists of showing that the execution of Steps 2–7 requires only time
O(ω log n) w.h.p. for any ω = ω(n) → ∞. This analysis requires a careful accounting of conditional probabilities. This is facilitated by the conditional model of the preferential attachment
graph due to Bollobás and Riordan [4]. One contribution of our paper is to recast their model in
terms of sums of independent copies of the rate one exponential random variables; this will be
essential to our analysis.
3
Outline of the paper
In Section 2 we reformulate the construction of Bollobás and Riordan [4] in terms of sums of
independent copies of the exponential random variable of rate one.
Section 3 is the heart of the paper. The aim is to show that if vt is not too small, then the ratio
vt+1 /vt is bounded above by 3/4 in expectation. We deduce from this that w.h.p. the main loop,
Steps 2–7, only takes O(ω log n) rounds. The idea is to determine a degree bound ∆ such that
many of vt ’s left neighbors have degree at least ∆, while only few of vt ’s right neighbors have
degree at least ∆. In this way, vt+1 is likely to be significantly smaller than vt .
Once we find a vertex vT of high enough degree, then we know that w.h.p. vT is not very large
and lies in a small connected subgraph of vertices of high degree that contains vertex one. Then a
simple argument based on the worst-case covertime of a graph suffices to show that only o(log n)
more steps are required.
Our proofs will use various parameters. For convenience, we collect here in table form a dictionary
of some notations, giving a brief (and imprecise) description of the role each plays in our proof,
for later reference.
ω :=
λ0 :=
Definition
Role in proof
O(log log n)
An arbitrarily chosen slowly growing fucntion.
1
log
40/m
n
A (usually valid) lower bound on random variables ηi (cf. Section 2.1).
n1 :=
log1/100 n
W.h.p. the main loop never visits v ≤ n1 .
Pt :=
{v1 , . . . vt }
The set of vertices visited up to time t.
Ψ:=
(log log n)10
Vertices v > Ψvt will not be important in the search for vt+1 .
L:=
m1/5
A large constant, significantly smaller than m.
Notation: We write An ∼ Bn if An = (1 + o(1))Bn as n → ∞. We write α . β in place of
α ≤ o(1) + (1 + o(1))β.
4
2
2.1
Preliminaries
A different model of the preferential attachement graph
Bollobás and Riordan [4] gave an ingenious construction equivalent to the preferential attachment
graph model. We choose x1 , x2 , . . . , x2mn independently and uniformly from [0, 1]. We then let
{ℓi , ri } = {x2i−1 , x2i } where ℓi < ri for i = 1, 2, . . . , mn. We then sort the ri in increasing order
R1 < R2 < · · · < Rmn and let R0 = 0. We then let
Wj = Rmj and wj = Wj − Wj−1 and Ij = (Wj−1 , Wj ]
for j = 1, 2, . . . , n. Given this we can define Gn as follows: It has vertex set Vn = [n] and an
edge {x, y} , x ≤ y for each pair ℓi , ri , where ℓi ∈ Ix and ri ∈ Iy .
We recast the construction of Bollobás and Riordan as follows: we can generate the sequence
R1 , R2 , . . . , Rmn by letting
1/2
Υi
Ri =
,
(1)
Υmn+1
where Υ0 = 0 and
ΥN = ξ1 + ξ2 + · · · + ξN for N ≥ 1
and where ξ1 , ξ2, . . . , ξmn+1 are independent exponential rate one random variables i.e. Pr(ξi ≥
2
are independent and uniform in [0, 1] (as they
x) = e−x for all i. This is because r12 , r22 , . . . , rmn
are each chosen as the maximum of two uniform points) and the order statistics of N independent
uniform [0, 1] random variables can be expressed as the ratios Υi /ΥN +1 for 1 ≤ i ≤ N.
We refer to the distribution of ΥN as ERL(N), as it is known in the literature as the Erlang
distribution.
2.2
Important properties
The advantage of our modification of the variant of the Bollobàs and Riordan construction is
that if we define
ηi := ξ(i−1)m+1 + ξ(i−1)m+2 + · · · + ξim ,
then ηi is closely related to the size of Ii . It can then be used to estimate the degree of vertex i.
This will simplify the analysis since ηi is simply a sum of exponentials.
In this section, we make this claim (along with other more obscure asymptotic properties of this
model) precise. In particular, we let E denote the event that the following properties hold for
Gn . In the appendix, we prove that Gn has all these properties w.h.p.
5
(P1) For Υk,ℓ = Υk − Υℓ , we have
Υk,ℓ
"
1/2
Lθk,ℓ
∈ (k − ℓ) 1 ±
3(k − ℓ)1/2
#
for (k, ℓ) = (mn + 1, 0) or
l=0
1
k−ℓ
2
∈ {ω, ω + 1, . . . , n} and k − l ≥ log n
k ≥ log30 n, l > 0
m
1/300
log
n 0 < l < k < log30 n.
Here, where n0 =
λ20 n
,
ω log2 n
λ0 =
θk,ℓ
1
,
log20/m n
log k
1/2
k
= (k − ℓ)1/2
(k−ℓ)3/2 log n
n1/2
n
ω 3/2 log2 n
Similarly define
θk =
1/2
k3/2
k
log n
1/2
n
n
ω 3/2 log2 n
ω ≤ l < k ≤ log30 n,
ω ≤ k ≤ n2/5 , l = 0,
log30 n < k ≤ n2/5 ,
n2/5 < k ≤ n0 ,
n0 < k.
k ≤ n2/5 ,
n2/5 < k ≤ n0 ,
n0 < k.
#
1/2 "
1/2
1/2
Lθi
i
i
1 ± 1/2 ∼
for ω ≤ i ≤ n.
(P2) Wi ∈
n
i
n
"
#
1/2
ηi
2Lθi
ηi
(P3) wi ∈
1 ± 1/2 1/2 ∼
for ω ≤ i ≤ n.
1/2
2m(in)
m i
2m(in)1/2
(P4) λ0 ≤ ηi ≤ 40m log log n for i ∈ [log30 n].
(P5) ηi ≤ log n for i ∈ [n].
Some properties give asymptotics for intermediate quantities in the Bollobas/Riordan model
(e.g., (P2), (P3)), while the rest give worst-case bounds on parameters in various ranges for
i. The very technical (P1) is just giving constraints on the gaps between the points Υk in the
Bollobas/Riordan model.
2.3
Inequalities
We will use the following inequalities from Hoeffding [9] at several points in the paper. Let
Z = Z1 + Z2 + . . . + ZN be the sum of independent [0, 1] random variables and suppose that
6
µ = E(Z). Then if α > 1 we have
n 2 o
(
exp − α3µ , α ≤ 1.
α2 µ
≤
Pr(Z ≥ (1 + α)µ) ≤ exp −
2 + α/3
exp − αµ
, α > 1.
3
βµ
e
Pr(Z ≥ βµ) ≤ e−µ
, β>1
β
2
α µ
, 0 ≤ α ≤ 1.
Pr(Z ≤ (1 − α)µ) ≤ exp −
2
(2)
(3)
(4)
Our main use for these inequalities is to get a bound on vertex degrees, see Section 2.4.
In addition to these concentration inequalities, we use various inequalities bounding the tails of
the random variable η. We note that the probability density φ(x) of the sum η of m independent
exponential rate one random variables is given by
φ(x) :=
xm−1 e−x
.
(m − 1)!
That is,
Pr(a ≤ η ≤ b) =
Z
b
φ(y)dy.
(5)
a
The equation (5) is a standard result, which can be verified by induction on m (for example, see
exercise 4.14.10 of Grimmett and Stirzaker [8]). Although we will frequently need to bound the
probability (5), this integral cannot be evaluated exactly in general, and thus we will often use
simple bounds on φ(η). We summarise what we need in the following lemma:
Lemma 2.1.
(a)
Pr(η ≤ xm) ≤ m(xe1−x )m
for x ≤ 1 −
1
.
m
(6)
(b)
Pr(η ≤ x) ≤ (1 − e−x )m ≤ xm .
(c)
Pr(η ≥ βm) ≤
eβ
eβ
m
≤ e−3m/10 for β ≥ 2.
(d)
2 m/3
Pr(η ≥ (1 + α)m) ≤ e−α
for 0 < α < 1.
(e)
2 m/2
Pr(η ≤ (1 − α)m) ≤ e−α
7
for 0 < α < 1.
(7)
Proof. (a) φ(η) is maximized at η = m − 1. Taking φ(mx) (x ≤ 1 − 1/m) as an upper bound on
φ(y) for y ∈ [0, mx] and m! ≥ (m/e)m in (5) gives us (6).
Q
(b) Writing η = ξ1 + ξ2 + · · · + ξm we have Pr(η ≤ x) ≤ m
i=1 Pr(ξi ≤ x).
(c) If η = ξ1 + ξ2 + · · · + ξN , then with λ = (β − 1)/β,
Pr(η ≥ βm) = Pr(eλη ≥ eλβm ) ≤ e−λβm E(eλη ) = e−λβm
m
Y
E(eλξi ) =
i=1
−λβm
e
(1 − λ)−m = (βe−(β−1) )m . (8)
(d) Putting β = 1 + α into (8) we see that
2 m/3
Pr(η ≥ (1 + α)m) ≤ ((1 + α)e−α )m ≤ e−α
.
(e) With λ = α/(1 − α) we now have
Pr(η ≤ (1 − α)m) = Pr(e−λη ≥ e−λ(1−α)m ) ≤ eλ(1−α)m E(e−λη ) = eλ(1−α)m
m
Y
E(e−λξi ) =
i=1
λ(1−α)m
e
2.4
(1 + λ)
−m
2 m/2
= ((1 − α)eα )m ≤ e−α
.
Properties of the degree sequence
We will use the following properties of the degree sequence throughout: let
!
1/2
n 1/2
i
5L log log n
ζ(i) =
1−
− 3/4
i
n
ω log n
!
1/2
n 1/2
i
5L log log n
+
.
1−
+ 3/4
ζ (i) =
i
n
ω log n
(9)
(10)
Note that
+
ζ(i) ∼ ζ (i)
ζ(i) ∼
n 1/2
i
2 log log n
.
if i ≤ n 1 −
log n
if i = o(n).
Also, let d¯n (i) denote the expected value of dn (i) in Gn .
Lemma 2.2.
8
(11)
(12)
(a) If E occurs then d¯n − m ∈ [ηi ζ(i), ηi ζ + (i)].
2 η ζ(i)/2
i
(b) Pr(dn (i) − m ≤ (1 − α)ηi ζ(i)) ≤ e−α
for 0 ≤ α ≤ 1.
2 η ζ + (i)/3
i
(c) Pr(dn (i) − m ≥ (1 + α)ηi ζ + (i)) ≤ e−α
(d) Pr(dn (i) − m ≥ βηi ζ + (i)) ≤ (e/β)βηi ζ
+ (i)
for 0 ≤ α ≤ 1.
for β ≥ 2.
(e) W.h.p. ηi ≥ λ0 and ω ≤ i ≤ n1/2 implies that dn (i) ∼ ηi
(f ) W.h.p. ω ≤ i ≤ log30 n implies that dn (i) ∼ ηi
n 1/2
.
i
(g) W.h.p. ω ≤ i ≤ n1/2 implies dn (i) . max {1, ηi }
(h) W.h.p. n1/2 ≤ i ≤ n implies dn (i) ≤ n1/3 .
(i) W.h.p. 1 ≤ i ≤ log1/49 n implies that dn (i) ≥
(j) W.h.p. dn (i) ≥
n
log1/20 n
n 1/2
i
n 1/2
.
i
.
n1/2
.
log1/20 n
implies i ≤ log1/9 n.
Proof. We defer the proof, which is straightforward but tedious, to the appendix.
Remark 2.3. We will for the rest of the paper condition on the occurrence of E. All probabilities include this conditioning. We will omit the conditioning in the text in order to simplify
expressions.
3
Analysis of the main loop
Since the variation distance after Step 1 is o(1), it suffices to prove Theorem 1.2 under the
assumption that we begin Step 2, with v1 chosen randomly, exactly according to the stationary
distribution.
The main loop consists of Steps 2–7. Let v0 = 1 and v1 , v2 , . . . , vs for s ≥ 1 be the sequence of
vertices followed by the algorithm up to time s. Let ρt = vt+1 /vt , and define T1 , T2 by
T1 = min t : vt ≤ log30 n and T2 = T1 + 30ω log4/3 log n and T0 = min 2ω log4/3 n, T2 . (13)
We will prove, see Lemma 3.2, that
E(ρt ) ≤
3
for 1 ≤ t ≤ T0 .
4
(14)
Recalling that T is the time when Step 8 begins, we note that if T < t ≤ T0 then this statement
is meaningless. So, we will keep to the following notational convention: if Xt is some quantity
that depends on t ≤ T and t > T then Xt = XT .
9
Now, roughly speaking, if r = 2 log4/3 n and µ is the number of steps in the main loop, then we
would hope to have
1
1
Pr(µ ≥ r) ≤ Pr ρ0 ρ1 · · · ρr ≥
≤ n E(ρ0 ρ1 · · · ρr ) ≤
n
n
and so w.h.p. the algorithm will complete the main loop within 2 log4/3 n steps. Unfortunately,
we cannot justify Q
the last inequality, seeing as the ρt are not independent. I.e. we cannot replace
E(ρ0 ρ1 · · · ρr ) by ri=0 E(ρi ). We proceed instead as in the next lemma.
Lemma 3.1. Assuming (14) we have the w.h.p. DCA completes the main loop in at most T0
steps with SUCCESS.
Proof. We let s0 denote the number of vertices visited by the main loop, and then define Zs =
ρ0 ρ1 · · · ρs for s ≤ s0 , and Zs = ρ0 ρ1 · · · ρs0 ( 43 )s−s0 for s > s0 .
Suppose first that T1 > ω log4/3 n. Now (14) and Jensen’s inequality implies that for s ≥ 1,
min(s,s0 )
E(log(Zs )) =
X
E(log(ρi )) +
i=0
s
X
log 43
min(s,s0 )+1
≤
min(s,s0 )
X
log E(ρi ) +
i=0
s
X
log 43 ≤ s log(3/4). (15)
min(s,s0 )+1
Now
log(Zs ) ≥ (s − s0 ) log(3/4) − log n ≥ s log(3/4) − log n
(16)
since ρ1 ρ2 . . . ρs0 ≥ 1/n.
Now let
α = Pr(log(Zs ) ≤ (1 − β)s log(3/4))
where α, β are to be determined. Then, (15), (16) imply that
(1 − α)(1 − β)s log(3/4) + α(s log(3/4) − log n) ≤ E(log(Zs )) ≤ s log(3/4).
(17)
Equation (17) then implies that
α≥
βs log(4/3)
.
βs log(4/3) + log n
Now putting s = ω log4/3 n and β = 1/2 we see that (18) becomes
α ≥1−
2
= 1 − o(1).
ω+2
So w.h.p. after at most ω log4/3 n steps, we will have exited the main loop with SUCCESS.
10
(18)
Suppose now that T1 ≤ ω log4/3 n. Using the argument that gave us (18) we obtain
T − T1 ≤ ω log4/3 log30 n w.h.p.
To prove Lemma 3.2, we will use a method of deferred decisions, exposing various parameters of Gn as we proceed. At time t, we will consider all random variables in the model from
Section 2.1 as being exposed if they have affected the algorithm’s trajectory thus far, and condition on their particular evaluation. To reduce the conditioning necessary, we will actually
analyze a modified algorithm, NARROW-DCA(τ ), and then later show that the trajectory of
NARROW-DCA(τ ) is the same as that of the DCA algorithm, w.h.p., when identical sources
of randomness are used.
NARROW-DCA(τ ) is the same as the DCA algorithm, except that for the first τ rounds of
the algorithm, a modified version of Step 4 is used:
Modified Step 4 Let
Ct = w1 , w2 , . . . , wm/2
be the m/2 neighbors of vt of largest degree from {1, . . . , Ψvt } where Ψ := (log log n)10 .
For rounds τ + 1, τ + 2, . . . , the behavior of NARROW-DCA(τ ) is the same as for DCA.
Notice that NARROW-DCA “cheats” by using the indices of the vertices, which we do not
actually expect to be able to use. Nevertheless, we will see later that w.h.p., for τ = 2ω log4/3 n,
the path of this algorithm is the same as for the DCA algorithm, justifying its role in our analysis.
3.1
Analyzing one step
Our analysis of one step of the main loop consists of the following lemma:
Lemma 3.2. Let ρt be the ratio of vt+1 /vt which appears in a run of the algorithm NARROWDCA(t). Then for all t ≤ T0 (see (13)), we have that
E(ρt ) ≤
3
4
and
Pr (ρt ≥ Ψ) ≤
1
.
log2 n
(19)
The first statement ensures that NARROW-DCA(t) makes progress in expectation in the tth
jump. The second part of this statement implies by induction that for any t ≤ ω log n, the
behavior of NARROW-DCA(t) is identical to the behavior of the DCA algorithm for the first
t steps. Thus together these statements give (14).
To prove Lemma 3.2, we will prove a stronger statement which is conditioned on the history of
the algorithm at time t. The history Ht of the process at the end of step t consists of
11
(H1) The sequence v1 , v2 , . . . , vt .
(H2) The left-choices λ(vs , 1), λ(vs , 2), . . . , λ(vs , m), 1 ≤ s < t and the corresponding left neighbors NL (vs ) = {u1,s , u2,s , . . . , um,s }. These are the m ℓ′i s that correspond to the m ri ’s
associated with vs as defined at the beginning of Section 2.1.
(H3) The lists u′1,s , u′2,s , . . . , u′r,s of all vertices u′k,s which have the property that (i) vs ∈ NL (u′k,s)
and (ii) u′k,s ≤ Ψvs for 1 ≤ s < t. (It is important to notice that s < t here.)
(H4) The values ηvi and the intervals Ivi for i = 1, 2, . . . , t.
(H5) The values ηw and the intervals Iw and the degrees deg(w), for w ∈
Here,
St
i=1
N(vi ).
N(v) = NL (v) ∪ NR (v) where NR (v) = {w ≤ Ψv : v ∈ NL (w)} .
We note that at any step t, and for a fixed random sequence used in the NARROW-DCA(t)
algorithm, Ht contains all random variables which have determined the behavior of the algorithm
so far, in the sense that if we modify any random variables from the random graph model
described in Section 2.1 while preserving all values in the history, then the trajectory of the
algorithm will not change. We write Ht to refer to a particular evaluation of the history (so that
we will be conditioning on events of the form Ht = Ht ).
Structure of the proof
The essential structure of our proof of Lemma 3.2 is as follows:
Part 1 We will define the notion of a typical history Ht .
Part 2 We will prove that for t ≤ T0 and any typical history Ht , random variables ηv which are
not explicitly exposed in Ht are essentially unconditioned by the event Ht = Ht (Lemma
3.3).
Part 3 We will prove by induction that Ht is typical w.h.p., for t ≤ T0 .
Part 4 We will use Part 2 and Part 3 to prove that for t ≤ T0 ,
E(ρt | Ht ) ≤
L3
2 21ηvt
+
+ 2
3
mL
m
and
Pr (ρt ≥ Ψ | Ht ) ≤
1
log2 n
(20)
by using using nearly unconditioned distributions of random variables which are not revealed in Ht to estimate the probabilities of various events. Here E(ρt | Ht ) is short for
E(ρt | Ht = Ht ). (Note that ηvt in (20) is simply a real number determined by Ht .) In this
context, we always work under the assumption that Ht is typical.
12
Part 5 We will also prove for t ≤ T0 that
E(ηvt+1 ) ≤ 4m.
(21)
Now the expected value statement in (19) follows from (21) and the first part of (20), by removing
the conditioning on Ht .
Part 1
Let Pt denote the sequence of vertices v1 , v2 , . . . , vt determined by the history Ht . We now
define the notion of a typical history Ht . For this purpose, we consider the reordered values
(t)
(t)
(t)
0 ≤ λ1 < λ2 < · · · < λN (t) where
n
o
(t)
(t)
(t)
(t)
Λ0 = λ1 , λ2 , . . . , λN (t) = {λ(vs , i) : 1 ≤ s ≤ t, 1 ≤ i ≤ m} .
(t)
(t)
Given this we define v = vj to be the index such that λj ∈ Iv and then let
n
o
(t)
(t)
VL = vj : 1 ≤ j ≤ N(t) .
We also define
(t)
VR = {v : v ∈ NR (Pt )}
Now let us reorder
n
o
(t)
(t)
(t)
(t)
(t)
V (t) = x1 < x2 < · · · < xM (t) = VL ∪ VR .
(t)
(t)
We define the extreme points x0 = 0 and xM (t)+1 = n + 1 and define
M (t)+1
(t)
Xj
=
(t)
[xj−1
+
(t)
1, xj
− 1] and X
(t)
=
(t)
[
Xj = [n] \ V (t)
(t)
(t)
and Nj = |Xj |,
j=1
(t)
Uj = [Wx(t)
j−1 +1
, Wx(t) −1 ] and U (t) =
j
N
[
(t)
Uj
(t)
(t)
and Lj = |Uj |.
j=1
A typical history Ht , t ≤ T0 is now one with the following properties:
(S1) There do not exist s1 , s2 ≤ t such that either (i) s1 ≤ t − 2 and vs1 and vs2 are neighbors
or (ii) s1 ≤ t − 3 and there exists a vertex w such that w ∈ N(vs1 ) ∩ N(vs2 ). (We say that
the path is self-avoiding.)
(S2) The points of Λ(t) are well-separated, in the following sense:
(
(t)
log2 n
xj−1 ≥ log30 n.
(t)
(t)
|xj − xj−1 | ≥
log1/400 n Otherwise.
13
(22)
We observe that
(T1) If Ht is typical then vj+1 is chosen from X (j) for all j < t.
(t)
(t)
(T2) Each Uj is the union of intervals Iv , v ∈ Xj .
Part 2
We prove the following:
Lemma 3.3. For any vertex v ∈ X (t) , any interval R ⊆ R, and any typical history Ht , we have
that v ∈
/ Pt ∪ N(Pt ) implies
Pr (ηv ∈ R | Ht ) ∼ Pr (ERL(m) ∈ R) .
(23)
The following lemma is the starting point for the proof of Lemma 3.3.
(t)
Lemma 3.4. Let j ∈ [M(t)+1], let Ht be any typical history, and let X ′ be the value of Xj in Ht .
Then the distribution of the random variables ηv , v ∈ X ′ conditioned on Ht = Ht is equivalent
to
of the random variables ηv , v ∈ X ′ conditioned only on the relationship
P the distribution
2
2
v∈X ′ ηv = A1 − A0 , where A1 , A0 are the values of Wx(t) −1 and Wx(t) +1 , respectively, in Ht .
j
j−1
Proof. Suppose we fix everything except for ηv , v ∈ X ′ . By everything we mean every other ηw
and all of the λ(v, i) and the random bits we use to make our choices in Step 5 of DCA; we
′
′
′
let Ht be the corresponding
Phistory. Suppose now that we replace ηv , v ∈ X with ηv , v ∈ X
without changing the sum v∈X ′ ηv . Then Wx(t) +1 remains the same, as it depends only on ηv
j−1
for v ∈
/ X ′ , and thus Wx(t) −1 remains the same as well, since the difference A21 − A20 is unchanged.
j
In particular, this implies that Ht remains a valid history. We confirm this by induction. Suppose
that Hs , s < t remains valid. We first note that because the λ(vs , i) are unchanged, none of vs′ s
(t)
left neighbors are in Xj . Also, NR (vs ) and the vertex degrees for w ∈ NR (vs ) will not be affected
(t)
by the change, even if vs < min Xj . So Hs+1 will be unchanged, completing the induction.
We are now ready to prove Lemma 3.3.
(t)
(t)
Proof of Lemma 3.3. Suppose that v ∈ X ′ = Xj , then M = Nj
Lemma 3.4 to write
Pr(ηv ≤ x | Ht ) = Pr ηv ≤ x
X
w∈X ′
14
ηw =
A21
−
≥ ζn → ∞. We now use
A20 ,
!
where A1 and A0 are the values of Wx(t) −1 and Wx(t)
j
j−1)+1
, respectively, in Ht , so that A1 − A0 is
(t)
the value of Lj in Ht .
Now from (P1) we have that A := A21 − A22 ∈ [(1 − ε)mM, (1 + ε)mM] for M = |X ′ | w.h.p., for
any ε > 0. Thus we fix any µ ∈ [(1 − ε)mM, (1 + ε)mM] and show that
!
X
ηw = µ = (1 + O(ε)) Pr (ERL(m) ≤ x) .
Pr ηv ≤ x
w∈X ′
The lemma follows since ε is arbitrary.
We write
Pr ηv ≤ x
X
w∈X ′
x
ηw = µ
!
η m−1 e−η (µ − η)(M −1)m−1 e−(µ−η) (Mm − 1)!
·
· M m−1 −µ dη
((M − 1)m − 1)!
µ
e
η=0 (m − 1)!
(M −1)m−1 Q
Z x m−1 −η
1 − µη
eη m
i=1 (Mm − i)
η
e
·
dη
=
µm
η=0 (m − 1)!
2
Z x m−1 −η
m
η
η
η
e
dη
·
1
+
O
· exp η − ((M − 1)m − 1)
+O
=
µ
µ2
M
η=0 (m − 1)!
Z x m−1 −η
η
e
= (1 + O(ε))
dη.
η=0 (m − 1)!
=
Z
Here we used that Ht typical implies that M ≥ log1/400 n → ∞.
Part 3
In the next section we will need a lower bound on vt+1 . Let
(
1
v ≥ log30 n
log3 n
φv =
1
. v < log30 n.
(log log n)3
Lemma 3.5. W.h.p. ρt ≥ φvt for 1 ≤ t ≤ T0 .
Proof. The values of λ(vt , i), i = 1, 2, . . . , m are unconditioned by Ht , see (H2). It then follows
from (P2) that if vt ≥ log30 n then
Pr(vt+1 ≤ φvt vt | Ht ) . m
Wφvt
m
. mφ1/2
.
vt =
Wvt
log3/2 n
There are O(ω log n) choices for t and so this deals with vt ≥ log30 n.
15
(24)
Now there are O(log log n) choices of t ∈ [T1 , T0 ] for which vt ≤ log30 n. In this case we can
replace the RHS of (24) by 1/(log log n)3/2 .
We will also need to bound the size of NR (vt ) for all t.
Lemma 3.6. W.h.p., for all t ≤ T0 ,
|NR (vt )| ≤
(
log3 n
(log log n)20
vt ≥ log30 n.
vt ≤ log30 n.
Proof. The size of NR (v), v = vt is stochastically bounded by Bin(Ψv, ηv /v). This is because if
w ∈ NR (v) then w ≤ Ψv. Also, for any such w, the probability that it has v as a left neighbor
is at most mwv /Ww . ηv /(vw)1/2 ≤ ηv /v. This uses property (S1) to see that the values of
λ(w, i), i = 1, 2, . . . , m are unconditioned by Ht . Thus, if θv = log3 n if v ≥ log30 n and equal to
(log log n)20 otherwise,
θ
ηv θv
Ψv
eΨηv v
Pr(|NR (v)| ≥ θv | Ht ) ≤
≤
.
(25)
θv
v
θv
3
If v ≥ log30 n then the RHS of (25) is at most (e/ log n)log n which is clearly small enough to han20
dle T possible values for t. If v < log30 n then the RHS of (25) is at most (40e/(log log n)9 )(log log n)
which is small enough to handle O(ω log log n) possible values for t such that v < log30 n.
Continuing Part 3, we now show that the DCA walk doesn’t contain cycles.
Lemma 3.7. W.h.p. the path Pt , t ≤ T0 is self avoiding.
Proof. We proceed by induction and assume that the claim of the lemma is valid up to time
t − 1. Now consider the choice of vt .
Case 1: There is an edge vs vt where s ≤ t − 2:
(a): vt ∈ NL (vs ) ∩ NL (vt−1 ).
We bound the probability of this (conditional on E, Ht ) asymptotically by
X X mwv
X X
ηv
.
.
Wvt−1
2(vvt−1 )1/2
s∈[t−2] v∈NL (vs )
(26)
s∈[t−2] v∈NL (vs )
Here, and throughout the proof of Case 1, v denotes a possibility for vt and mwv /Wvt−1 bounds
the probability that vt−1 chooses v. Remember that these choices are still uniform, given the
history.
We split the sum in (26) as
X
X
s∈[t−2] v∈NL (vs )
vs >log30 n
ηv
+
2(vvt−1 )1/2
X
X
s∈[t−2] v∈NL (vs )
vs ≤log30 n
16
ηv
.
2(vvt−1 )1/2
Consider the first sum. There are less than t choices for s; m choices for v and ηv ≤ log n. Now
v ∈ NL (vs ) and Lemma 3.5 implies that v ≥ log27 n. So we can bound the first sum by
1
1
1
.
=o
(# of s) · (# of v) · (max ηv ) · 1/2 ≤ T0 · m · log n ·
v
log11 n
log27/2 n
Summing this estimate over t ≤ T0 gives o(1).
For the second sum, we bound the number of choices of s by O(ω log log n) and ηv by O(log log n),
since v ≤ vs . We use the fact (see Section 3.2) that vt−1 ≥ log1/100 n. So we can therefore bound
the second sum by
1
1
1
(# of s) · (# of v) · (max ηv ) · 1/2 ≤b ω log log n · m · log log n ·
=o
.
log1/200 n
log1/300 n
vt−1
(We use A ≤b B in place of A = O(B).)
There are O(ω log log n) choices for T0 ≥ t > s ≥ T1 and so we can sum this estimate over choices
of t.
(b): vt ∈ NL (vs ) ∩ NR (vt−1 ).
Using vt ∈ NR (vt−1 ), we bound the probability of this asymptotically by
X
X
s∈[t−2] v∈NL (vs )
vs >log30 n
ηvt−1
+
2(vvt−1 )1/2
X
X
s∈[t−2] v∈NL (vs )
vs ≤log30 n
ηvt−1
.
2(vvt−1 )1/2
For the first sum we use the argument of Case (a) without any change, except for bounding ηvt−1
by log n as opposed to bounding ηv by the same. This gives a bound
1
1
1
(# of s) · (# of v) · (max ηvt−1 ) · 1/2 ≤b T0 · m · log n ·
.
=o
v
log11 n
log27/2 n
This is small enough to inflate by the number of choices for t.
For the second sum we split into two cases: (i) vt−1 ≥ log30 n and (ii) vt−1 < log30 n. This enables
us to control ηvt−1 . For the first case we obtain
1
1
1
(# of s) · (# of v) · (max ηvt−1 ) · 1/2 ≤ ω log log n · m · log n ·
=o
.
log15 n
log13 n
vt−1
The RHS is small enough to handle the O(ω log n) choices for t.
For the second case we obtain
(# of s) · (# of v) · (max ηvt−1 ) ·
1
1/2
vt−1
≤ ω log log n · m · log log n ·
1
log1/200 n
The RHS is small enough to handle the O(ω log log n) choices for t.
17
=o
1
log1/300 n
.
(c): vt ∈ NR (vs ) ∩ NL (vt−1 ).
Using vt ∈ NL (vt−1 ), we bound the probability of this asymptotically by
X
X
X
X
ηv
ηv
+
.
1/2
2(vvt−1 )
2(vvt−1 )1/2
s∈[t−2] v∈NR (vs )
vs ≤log29 n
s∈[t−2] v∈NR (vs )
vs >log29 n
For the first sum we use v ≥ vs and the argument of Case (a) without change, but notice we
split over vs > log29 n or not here. This gives a bound of
1
1
1
(# of s) · (# of v) · (max ηv ) · 1/2 ≤ T0 · m · log n ·
=o
.
v
log12 n
log29/2 n
For the second sum we use v ≤ Ψvs to bound v by log30 n. We also use Lemma 3.6 to bound the
number of choices of v by (log log n)20 . This gives a bound of
1
1
1
20
.
(# of s)·(# of v)·(max ηv )· 1/2 ≤b ω log log n·(log log n) ·log log n· 1/200 = o
log
n
log1/300 n
vt−1
(d): vt ∈ NR (vs ) ∩ NR (vt−1 ).
Using vt ∈ NR (vt−1 ), we bound the probability of this asymptotically by
X
X
s∈[t−2] v∈NR (vs )
vs >log30 n
ηvt−1
+
2(vvt−1 )1/2
X
X
s∈[t−2] v∈NR (vs )
vs ≤log30 n
ηvt−1
.
2(vvt−1 )1/2
For the first sum we use v ≥ vs and Lemma 3.6 to bound the number of choices for v and then
we have a bound of
1
1
1
3
.
=o
(# of s) · (# of v) · (max ηvt−1 ) · 1/2 ≤ T0 · log n · log n ·
v
log15 n
log9 n
For the second sum we split into two cases: (i) vt−1 ≥ log30 n and (ii) vt−1 < log30 n. This enables
us to control ηvt−1 . We also use Lemma 3.6 to bound the number of choices for v in each case.
Thus in the first case we have the bound
1
1
1
3
(# of s) · (# of v) · (max ηvt−1 ) · 1/2 ≤b ω log log n · log n · log n ·
=o
.
log15 n
log10 n
vt−1
In the second case we have
(# of s)·(# of v)·(max ηvt−1 )·
1
1/2
vt−1
20
≤b ω log log n·(log log n) ·log log n·
1
log1/200 n
=o
1
log1/300 n
Case 2: There is a path vs , v, vt where s < t.
The calculations that we have done for Case 1 carry through unchanged. We just replace vt−1 by
vt throughout the calculation and treat v as an arbitrary vertex as opposed to a choice of vt .
18
.
(t)
The xj are separated
We now prove that w.h.p. points λi are well-separated. Let
J1 = j : vj ≥ log30 n .
Lemma 3.8. Equation (22) holds w.h.p. for all t ≤ T0 .
Proof. We consider cases.
(t)
(t)
(t)
Case 1: xj−1 , xj ∈ VR .
For this we write
ζv,w =
(
log2 n
min {v, w} ≥ log30 n.
.
log1/300 n Otherwise.
Pr(∃1 ≤ s ≤ t, v ∈ NR (vs ), w ∈ NR (vt ) : |v − w| ≤ ζv,w | E, Ht) ≤
X
X
ηvs ηvt
.
(vs vt vw)1/2
1≤s≤t≤T
0
≤
≤
X
v∈NR (vs ),w∈NR (vt )
|v−w|≤ζv,w
X
ζv,w ηvs ηvt
(v − ζv,w )(vs vt )1/2
(27)
1≤s≤t≤T0 v∈NR (vs )
∗
X ζs,t
ηvs ηvt |NR (vs )|
2
.
(vs vt )1/2
1≤s≤t≤T0
(28)
∗
Here ζs,t
will be a bound on the possible value of ζv,w in (27).
Case 1a: max {vs , vt } ≥ log29 n:
∗
In this case ζs,t
≤ log2 n and we can bound the summand of (28) by
∗
ζs,t
· log2 n · log3 n ·
1
log
29/2
n
=
1
log
15/2
n
.
Multiplying by a bound T02 on the number of summands gives a bound of o(1). Here, and in the
next case, we use Lemma 3.6 to bound |NR (vs )|.
Case 1b: max {vs , vt } < log29 n:
Here we have max {v, w} ≤ Ψ log29 n ≤ log30 n. In this case we can bound the summand of (28)
by
1
(log log n)5
1/300
2
20
log
n · (40m log log n) · (log log n) ·
.
=o
log1/100 n
log1/200
We only have to inflate this by (T0 − T1 )2 = O((ω log log n)2 ). This completes the case where
(t)
(t)
xj−1 , xj ∈ R(t) .
19
(t)
(t)
(t)
Case 2: xj−1 , xj ∈ VL
We first show that the gaps λj − λj−1 are large. Define
β1 =
and
log1/300 n
log15/2 n
and
β
=
.
2
n1/2
n1/2
(
β1
εj =
β2
and
σ1 =
λj = λ(vt , i), vt ∈ J1 .
otherwise.
1
log
15/2
n
and σ2 =
1
log
1/200
n
.
We drop the superscript t for the rest of the lemma.
Claim 3.9.
Pr(∃λj ∈ Λ0 : λj−1 > λj − εj | Ht ) = o(1).
Proof of Claim 3.9. This follows from the fact that
Pr(∃j : λj−1 > λj − εj ) ≤ o(1) + (1 + o(1))(m2 T12 σ1 + m2 (T0 − T1 )(T1 σ1 + (T0 − T1 )σ2 )).
We have fewer than m2 T12 choices for s = τ (j − 1), t = τ (j) ∈ J1 . Assume first that s < t.
Given such a choice, we have that w.h.p. Wvt & log15 n/n1/2 by (P2). Now λj will have been
chosen uniformly from 0 to ≈ Wvt and so the probability it lies in [λj−1, λj−1 + εj ] is at most
≈ β1 /Wvs , which explains the term m2 T12 σ1 . If s > t then we repeat the above argument with
[λj−1 , λj−1 + ε1 ] replaced by [λj − ε1 , λj ]
The term m2 (T0 − T1 )T1 σ1 arises in the same way with j − 1 ∈ J1 , j ∈
/ J1 or vice-versa.
The term m2 (T0 − T1 )2 σ2 arises from the case where j − 1, j ∈
/ J1 . Here we can only assume
1/200
1/2
that Wj & log
n/n . This follows from (P2), (P4) and Lemma 2.2 and the fact that
we exit the main loop with SUCCESS when we see a vertex of degree at least n1/2 / log1/100 n.
Assuming that s < t we see that the probability that λj lies in [λj−1 , λj−1 + ε2 ] is at most
β2 /Wvt ∼ β2 /(log1/200 n/n1/2 ) = o(1).
Given the Claim and (P4), (P5) we have that w.h.p.
(
n
1
β1 − log
1/2 ≥ 2 β1
n
Wv(t) −1 − Wv(t) +1 ≥
log n
j
j−1
≥ 21 β2
β2 − 40mnlog
1/2
j ∈ J1 .
j∈
/ J1
(29)
Now,
Wv(t) −1 − Wv(t)
j
j−1 +1
=
Υm(τ (j)−1)
Υmn+1
1/2
−
Υm(τ (j−1)+1)
Υmn+1
20
1/2
=
Υm(τ (j)−1) − Υm(τ (j−1)+1)
1/2
1/2
1/2
Υmn+1 (Υm(τ (j)−1) + Υm(τ (j−1)+1) )
.
Or,
(
β1 n1/2
1/2
1/2
1/2
ηu = (Wv(t) −1 − Wv(t) +1 )Υmn+1 (Υm(τ (j)−1) + Υm(τ (j−1)+1) ) ≥
j
j−1
β2 n1/2
(t)
j ∈ J1 .
.
j∈
/ J1
X
u∈Xj
It follows that w.h.p.
(t)
|xj−1
(t)
(t)
(t)
−
(t)
xj |
=
(t)
|Xj |
≥
(
β1 n1/2
log n
β2 n1/2
40m log log n
j ∈ J1 .
j∈
/ J1
.
(t)
Case 3: xj ∈ VL , xj−1 ∈ VR
Let θv = β1 , v ≥ log30 n and θv = β2 otherwise. We write
ηvs
wv + 2θv
·
1/2
(vvs )
Wvt
s,t,v,k
X ηv ηv
2n1/2 θv ηvs
s
+
.(30)
.
v(vs vt )1/2
v(vs vt )1/2
s,t,v,k
Pr(∃s < t, v, k : v ∈ NR (vs ), λ(vt , k) ∈ Iv ± θv | E, Ht ) ≤
X
We bound the sum in the RHS of (30) as follows: If max {v, vs } ≥ log30 n then we bound the
first sum by
n
X
1
1
1
2
≤b (ω 2 log2 n) · log2 n · log n ·
= o(1).
(#s, t, k) · log n ·
·
15
v log n
log15 n
v=1
We bound the second sum by
n
X
1
1
1
15/2
(#s, t, k) · 2 log
n · log n ·
·
≤ (ω 2 log2 n) · log15/2 n · log n · log n ·
= o(1).
15
15
v
log
n
log
n
v=1
When max {v, vs , vt } < log30 n we bound the first sum by
log30 n
(#s, t, k) · (40 log log n)2 ·
X 1
1
≤b
·
1/200
v log
n
v=1
(ω log log n)2 · (log log n)2 · log log n ·
1
log
1/200
log
1/200
n
= o(1). (31)
We bound the second sum by
log30 n
(#s, t, k) · log
1/300
n · 40 log log n ·
X 1
1
≤b
·
1/200
v
log
n
v=1
(ω log log n)2 · log1/300 n · log log n ·
1
n
= o(1). (32)
Finally, if max {v, vs } < log30 n ≤ log30 n then we have to replace (#s, t, k) in (31), (32) by
1/2
O(ω 2 log n log log n). But this is compensated by a factor 1/vt ≤ 1/ log15 n.
It follows that (29) holds w.h.p. and the proof continues as for Case 2.
21
Part 4
We now assume t ≤ T0 . We begin by showing that DCA only uncovers a small part of the
distribution of the η’s.
Let Ξt = Pt ∪ N(Pt ) and
St,j =
X
wv .
v∈Ξt
Lemma 3.10. W.h.p., St,j = o(Wj ) for log1/100 n ≤ j and 1 ≤ t ≤ T0 .
Proof. Assume first that j ≥ log30 n. It follows from (P2), (P3), (P5) and Lemma 3.6 that
w.h.p.
max ηvs
T0 log n(m + log3 n)
ω log5 n
St,j ≤ T0 ×
× (m + max |NR (vs )|) .
=O
.
n1/2
2mn1/2
n1/2
10
1/2
log n
j
.
=Ω
Wj ≥ (1 − o(1))
n
n1/2
This completes this case. Now assume that j ≤ log30 n. (P2), (P3), (P4) and Lemma 3.6 that
w.h.p.
1/2
ω log4/3 log n
j
20
St,j . 40m log log n ×
× (m + (log log n) ) ≪ Wj ∼
1/2
2mn
n
for log1/100 n ≤ j ≤ log30 n.
Dealing with left neighbors
The calculation of the ratio ρt takes contributions from two cases: where vt+1 is a left-neighbor
of vt , and where vt+1 is a right-neighbor of vt .
Lemma 3.11.
2
E(ρt 1vt+1 <vt | Ht ) ≤ .
3
Proof. Let D denote the (m/2)th largest degree of a vertex in NR (vt ). We write
X
E(ρt 1vt+1 <vt | Ht | D = d) Pr(D = d)
E(ρt 1vt+1 <vt | Ht ) =
d
ζd
≤
E
vt
d
ζD
,
=E
vt
X
22
Pr(D = d)
where ζd is the index of the smallest degree left neighbor of vt that has degree at least d. We
let ζd = 0 if there are no such left neighbors. We now couple ζ with a random variable that is
independent of the algorithm and can be used in its place.
Going back to Section 2.1 let us associate ℓk for k ≥ ω with an index µk chosen uniformly from
[⌊k/m⌋]. In this way, vertex i ≥ ω is associated with m uniformly chosen vertices ai,1 , ai,2 , . . . , ai,m
in [i − 1]. Furthermore, we can couple these choices so that if NL (i) = {bi,1 , bi,2 , . . . , bi,m } then
given we have (i) Pr(bi,j ≤ ai,j ) ≥ 1 − o(1) and (ii) bi,j ≤ 2ai,j for all i, j. This because
Pr(bi,j ≤ k) ∼ Wk /Wi ∼ (k/i)1/2 (giving (i)) and (k/i)1/2 ≥ k/i and (1 − o(1))(k/i)1/2 ≥ k/2i
(giving (ii)).
So now let µ be the index of the uniform choice associated with the largest degree left neighbor
of vt that has degree at least D. Thus
ζD
µ
1
2
.E
= + o(1) ≤ .
E
vt
vt
2
3
Dealing with right neighbors
It will be more difficult to consider the contribution of right-neighbors. In preparation, for
λ0 ≤ γ ≤ 1 − 1/m we define
∆iγ := m + γmζ(i)
where ζ(i), ζ + (i) are defined in (9), (10) respectively. We note that ηi ζ(i) is a lower bound for
the expected degree of vertex i, i ≥ ω, see Lemma 2.2(a). Note also that ηi ζ + (i) is an upper
bound for the expected degree of vertex i, i ≥ ω.
The parameter ∆iγ is a degree threshold. For a suitable parameter γ, we wish it to be the case
that there should be many left-neighbors but few right-neighbors which have degree greater than
∆iγ . We define
γi∗ = max γ : j ∈ NL (i) : dn (j) ≥ ∆iγ ≥ m/2 .
∆vγtv∗ is a lower bound on the degree needed for vertex j > vt to be considered by DCA as the
t
next vertex; thus we proceed by analyzing the distribution of γv∗t . We first derive upper bounds
for Pr(γv∗t ≤ γ | Ht ).
Lemma 3.12. There exists c1 > 0 such that
1/2
)c1 m + me−c1 γ
Pr(γv∗t ≤ γ | Ht ) . (γ 1/2 e1−γ
1/2
)c 1 m
Pr(γv∗t ≤ 45 | Ht ) . e−c1 m .
Pr(γv∗t ≥ γ | Ht ) . γ −c1 m ,
2
1
, 0≤γ≤ .
8
1
−c1 /γ 1/2
+ me
, 0≤γ≤ .
8
Pr(γv∗t ≤ γ | Ht ) . (γ 1/2 e1−γ
2
5
γ ≥ 10 .
23
1/2 m
(33)
(34)
(35)
(36)
Proof. For j < vt , we define events Aj = ηj ≤ γ 1/2 m and Dj = dn (j) ≤ ∆vγt . We need to
T
estimate Pr
j∈S Dj for subsets S ⊆ NL (vt ) of size m/2. We write
\
j∈S
Dj ⊆
\
(Aj ∪ (Āj ∩ Dj )) ⊆
j∈S
\
Aj ∪
j∈S
[
(Āj ∩ Dj ).
(37)
j∈S
Now, using inequality (6) and equation (23), we see that if 0 ≤ γ ≤ 1/8 then for j < vt ,
Pr(ηj ≤ γ 1/2 m | Ht ) . m(γ 1/2 e1−γ
1/2
)m .
(38)
The RHS of (38) includes a factor of 1 + o(1) due to conditioning on E, Ht .
So,
Pr
\
Aj
j∈S
Ht
!
. (m(γ 1/2 e1−γ
1/2
)m )|S| .
(39)
Furthermore, because j ∈ NL (vt ) implies that i ≥ j and hence ζ(vt ) ≤ ζ(j),
Pr((dn (j) ≤ ∆iγ ) ∧ Āj | Ht ) ≤ Pr (dn (j) − m ≤ γmζ(j) | Ht ) Pr(ηj > γ 1/2 m|Ht )
(1 − γ 1/2 )2 1/2
γ mζ(j) .
. exp −
2
(40)
Explanation of (40): We remark first that the conditioning on E, Ht only adds a (1+o(1)) factor
to the upper bound on our probability estimate. We now apply Lemma 2.2(b) with 1 − α = γ
and ηj ≥ γ 1/2 m.
From (37) (summing over all m/2 subsets of NL (vt )) and (40) (summing over NL (vt )) we obtain
Pr(γv∗t ≤ γ | Ht ) = Pr(| j ∈ NL (vt ) : dn (j) ≤ ∆iγ | ≥ m/2| | Ht ) .
m/2
(1 − γ 1/2 )2 1/2
m
1/2 1−γ 1/2 m
2 (m(γ e
) )
+ m exp −
γ mζ(vt ) . (41)
2
We observe that j ∈ NL (vt ) implies that dn (j) ≥ m + 1. So,
m + 1 < ∆iγ implies ζ(vt ) >
1
.
mγ
(42)
m/2
1/2
2
1/2 1−γ 1/2 m
Using (42) in (41) verifies (34), after bounding 2 (m(γ e
) )
by (γ 1/2 e1−γ )c1 m .
m
From (37) and (40),
Pr(γv∗t ≤ γ | Ht ) . (m(γ 1/2 e1−γ
1/2
)m )m/2 +
(1 − γ 1/2 )2 1/2
γ mζ
+ Pr(A1 | Ht ) + Pr(Ā1 | Ht ) m exp −
2
−1
24
9n
10
(43)
Here
A1 =
9n
NL (vt ) ∩
10
m
≥
4
.
Explanation of (43): The first term is from (39). If Ā1 holds then vt has at least one left
neighbor j ≤ 9n/10. The final term comes from using (40) and ζ(j) ≥ ζ(9n/10). The factor
Pr(Ā1 )−1 handles the conditioning on Ā1 . The factor m is the union bound for choices of j.
Now NL (vt ) ∩ 9n
is dominated by the binomial Bin(m, n/10) and so Pr(A1 | Ht ) ≤ e−d1 m .
10
Now ζ(9n/10) ≥ 1/20 and plugging these facts into (43) yields (33). Here we have absorbed the
1/2
e−d1 m term into me−c1 γ m and we will do so again below.
We continue with the proof of (35). For j ∈ NL (vt ), we observe that if dn (j) ≤ ∆vγt and γ ≤ 54
then
5m
dn (j) − m ≤
ζ(vt ).
4
We now estimate the probability that a uniform random choice of j ∈ NL (vt ) (for fixed Ht , which
determines vt ) has certain properties.
We first observe that
W3i/5
3i
Pr j ≥
Ht . 1 −
∼
5
Wvt
1/2 !
2
3
< .
1−
5
5
(44)
(For this we used (P2).)
Now (6) implies that
Pr(ηj ≤ 0.99m | Ht ) ≤ e−d2 m .
(45)
Moreover, for ηj > 0.99m and j < 3vt /5, we have
ζ(j)
=
ζ(vt )
1/2
vt
j
where
ε=
Thus we have
1−
1−
j 1/2
−ε
n
1/2
vt
−ε
n
!
1/2
vt
≥
j
5L log log n
.
ω 3/4 log n
(46)
1/2
5
E(dn (j) − m | Ht ) ≥ ηj ζ(j) & 0.99m
ζ(vt).
3
Now 0.99 × (5/3)1/2 = 1.278.. > 1.01 × 5/4 and so
ζ(j)ηj
5ζ(vt )
Ht ≤ Pr dn (j) − m ≤
Pr dn (j) − m ≤
4
1.01
25
Ht
≤ e−d3 ηj ζ(j) ≤ e−d4 m
(47)
using Lemma 2.2(b). It follows from (44) and (45) and (47) that
5
m
2
∗
−d2 m
−d4 m
Pr γvt <
≥
≤ e−d3 m .
Ht ≤ Pr Bin m, e
+e
+
4
5
2
This completes the proof of (35).
To deal with (36) we observe that if dn (j) ≥ ∆vγt and γ ≥ 105 then
vt
j ∈ NL (vt ) and j ≤ 1/2
γ
But
or
ηj ≥ γ
1/2
1/2
j
m
≥ γ 1/4 m or dn (j) − m ≥ γ 3/4 ηj ζ(j).
vt
vt
Pr j ∈ NL (vt ) and j ≤ 1/2
γ
Ht
And, using (P3) and γ ≥ 105 ,
Pr ηj ≥ γ
1/4
m Ht
.
vt
X
l=2vt /γ
wl
Wvt
.
Z
Wvt /γ 1/2
1
. 1/4 .
Wvt
γ
∞
ηl =γ 1/4 m
(48)
ηlm e−ηl
dηl
(m − 1)!
1/4
vt
X
e−γ m
.
2(vt l)1/2
l=2vt /γ
. e−γ
1/4 m
.
(49)
Lastly, using (44), (45) and Lemma 2.2(d) and ζ + (j) . ζ(j) for j ≤ 3n/5 we have
2
3/4
Pr dn (j) − m ≥ γ 3/4 ηj ζ(j) | Ht ≤ + e−d2 m + e−γ η ≤ 0.41.
5
(50)
It follows from (48), (49) and (50) that
1
m
∗
−γ 1/4 m
Pr γvt ≥ γ | Ht . Pr Bin m, (1 + o(1))
. e−d3 m .
+e
+ 0.41
≥
γ 1/4
2
This completes the proof of the lemma.
Corollary 3.13. W.h.p. γv∗s ≥ 1/(log log n)2 for s = 1, 2, . . . , T = O(log n).
Proof. The value of γv∗s is determined when vs is first visited and in this case we can apply Lemma
3.12. In which case the result follows directly from (34).
We now have a handle on the distribution of γv∗t . We now put bounds on the expected number
of j > vt that can be considered to be a candidate for vt+1 , conditioned on the value of γv∗t . In
particular, we let
Dγvt = j > vt : dn (j) ≥ ∆vγt .
26
We will bound the size of Dγi by dividing Dγi into many parts bounding each part; in particular,
κ ∈ N we let
h
i
i, i2 ∩ D i
κ = 0.
γ
γ
i
(51)
Jγi,κ = h
i2 1 + κ−1 , i2 1 + κ ∩ D i 1 ≤ κ ≤ 2nγ 2 L .
γ
γ
L
γ
L
i
Note that Jγi,0 = ∅ if γ ≥ 1.
Finally, we let
rγi,κ
:=
|Jγi,κ |
rγi
and
:=
X
rγi,κ
and
siγ
:=
κ≥0
i X
κ + 1 i,κ
j≤ 2
rγ .
1+
γ κ≥0
L
i,κ
X X
(52)
κ≥0 j∈Jγ
Remark 3.14. We have that E m2 v1t svγt | Ht is an upper bound on the expectation of the ratio
, conditioned on the event that vt+1 > vt , since each right neighbor whose index is
ρt = vt+1
vt
included in the sum svγt has probability of at most m2 of being chosen by the algorithm.
3L3
Lemma 3.15. If vt ≤ n 1 − m then
E(rγvt | Ht ) ≤
svγt
E
| Ht
vt
Moreover,
≤
ηvt
L max {0, (1 − γ)} + 7 + 10Le−c2 γL .
γL
(53)
ηvt
L max {0, (1 − γ)} + 13 + 100Le−c2 γL .
3
γ L
If vt ≤ n/5 and κ ≥ (log log n)4 and γ ≥ 1/(log log n)2 then Pr(rγvt ,κ > 0) ≤
(54)
1
.
log2 n
(55)
Note that (55) implies the second inequality in (19).
Proof. Recall from Lemma 3.10 that w.h.p.,
ST,j = o(Wj ) for j ≥ log1/100 n.
(56)
We write
E(rγvt ,κ | Ht )
Z ∞ m−1 −ηj
X
ηj e
mwvt
.
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj
ηj =0 (m − 1)!
vt ,κ Wj − St,j
(57)
j∈Jγ
. ηvt
X
v ,κ
j∈Jγ t
1
2(vt j)1/2
Z
∞
ηj =0
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj .
(m − 1)!
27
(58)
Explanation of (57) and (58): We sum over the relevant j and fix ηj . We multiply by the
density of ηj and integrate. Using (56) we see that
mwvt
ηvt
mwvt
∼
∼
.
Wj − St,j
Wj
2(vt j)1/2
This is asymptotically equal to the expected number of times j chooses vt as a neighbor.
Thus
E(rγvt ,κ | Ht ) .
X
v ,κ
j∈Jγ t
where
(59)
ηjm−1 e−ηj
Ij =
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj ≤ 1.
ηj =0 (m − 1)!
Z
If m is large then
ηvt
Ij
2(vt j)1/2
∞
E rγvt ,0 | Ht
Continuing, for κ ≥ 1, we write:
(
0
. P
ηvt
v
j∈J0 t (γ) 2(vt j)1/2
.
ηvt 1−γ
γ
γ ≥ 1.
γ < 1.
Ij . A1 + A2 + A3 ,
(60)
(61)
where
A1 =
Z
A2 =
Z
A3 =
Z
(1−1/L)γm
ηj =0
(1+1/L)m
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | ηj dηj .
(m − 1)!
ηj =(1−1/L)m
∞
ηj =(1+1/L)m
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | ηj dηj
(m − 1)!
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | ηj dηj .
(m − 1)!
and then write rγvt ,κ = rγvt ,κ,1 + rγvt ,κ,2 + rγvt ,κ,3. Here rγvt ,κ,l is equal to the RHS of (59) with Ij
replaced by Al . The implicit
(1 + o(1)) factor in (61) arises from replacing
Pr dn (j) ≥ ∆vγt | ηj , Ht by Pr dn (j) ≥ ∆vγt | ηj in the integrals, i.e., ignoring the conditioning
due to Ht . Since j > vt , the only effect
of Ht is on Wj through wvt . Here we have that w.h.p.
ηvt
log n
vt 1/2
Wj ∼ n
= o(Wvt ).
and wvt ∼ 2m(vt n)1/2 = O 2m(v
1/2
t n)
Case 1: n1 ≤ vt < n/5:
Note that in this case
1
ζ(vt ) ≥
2
1/2
n
≥ 1.
vt
In the following we use Lemma 2.1 to estimate the integrals over ηj . We observe that
E(rγvt ,κ,1 | Ht )
28
(62)
.
X
ηvt
2(vt j)1/2
X
ηvt
2(vt j)1/2
v ,κ
j∈Jγ t
.
v ,κ
j∈Jγ t
Z (1− 1 )m m−1 −ηj
L
ηj e
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj ,
(m − 1)!
ηj =0
(
!
Z (1− 1 )m m−1 −ηj
L
ηj e
1
1 ≤ κ ≤ 10L,
·
dηj , (63)
(m − 1)!
(e(L/κ)1/2 )d0 γmζ(vt ) κ > 10L
ηj =0
Explanation of (63): We remark first that the conditioning on Ht only adds a (1 + o(1)) factor
to the upper bound on our probability estimate. We will use Lemma 2.2 to bound the probability
that degrees are large. Now with our bound on vt and within the range of integration, the ratio
of ∆vγt − m to the mean of dn (j) − m is
∆vγt − m
ζ + (j)
1/2
1−
γm vnt
=
1/2
ηj nj
1−
vt 1/2
n
1/2
− o(1)
γm j
≥
&
ηj vt
j 1/2
+
o(1)
n
1/2
κ−1
κ 1/2
L
1+
≥
when κ ≥ 10L. (64)
L−1
L
L
We then use (11) and Lemma 2.2(d) with β = (κ/L)1/2 .
Continuing, we observe that
1/2
κ 1/2
κ−1
1
1+
− 1+
≤
L
L
2L
and so
E(rγvt ,κ,1
ηv e−m/(2L
| Ht ) ≤ t
γvt 1/2
2)
ηv e−m/(2L
≤ t
γL
2)
1/2
1/2 !
κ+1
κ−1
vt
vt
×
1+
−
1+
L
L
(
!
1
1 ≤ κ ≤ 10L,
(e(L/κ)1/2 )d0 γmζ(vt ) κ > 10L,
(
!
1
1 ≤ κ ≤ 10L,
(e(L/κ)1/2 )d0 γmζ(vt ) κ > 10L.
(65)
·
(66)
Continuing, it follows from (65) that
X
v ,κ
j∈Jγ t
1
j 1/2
≤
vt 1/2
.
γL
E(rγvt ,κ,2 | Ht )
Z (1+1/L)m
X
ηjm−1 e−ηj
ηvt
vt
Pr
d
(j)
≥
∆
|
H
,
η
dηj
≤
n
t
j
γ
1/2
ηj =(1−1/L)m (m − 1)!
vt ,κ 2(vt j)
j∈Jγ
29
(67)
≤
X
v ,κ
j∈Jγ t
ηvt
2(vt j)1/2
1 n
o
d1 γmζ(vt )
× exp − L2
(e(L/κ)1/2 )d0 γmζ(vt )
1
ηvt −d1 γmζ(vt )/L2
× e
≤
γL
(e(L/κ)1/2 )d0 γmζ(vt )
κ ≤ 3,
4 ≤ κ ≤ 10L,
(68)
κ > 10L,
1 ≤ κ ≤ 3,
4 ≤ κ ≤ 10L,
κ > 10L.
(69)
where we have used (67).
Explanation for (68): We proceed in a similar manner to (64) and use
1/2
vt 1/2
n
vt
1
−
−
o(1)
γm
∆γ − m
vt
n
1
1
m.
=
≥ 1 + if κ ≥ 4, ηj ≥ 1 −
1/2
ζ + (j)
L
L
j 1/2
n
1− n
+ε
ηj j
Then we use Lemma 2.2(c),(d).
Continuing,
E(rγvt ,κ,3
| Ht ) ≤
X
v ,κ
j∈Jγ t
ηvt
2(vt j)1/2
Z
∞
ηj =(1+1/L)m
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj
(m − 1)!
(70)
We bound the integral in (70) by something independent of j and then as above, there is a factor
ηvt /γL arising from the sum over j.
For all 1 ≤ κ ≤ 80L + 1, we simply use the bound
Z ∞
ηjm−1 e−ηj
d2 m
dηj ≤ exp − 2 .
1
L
ηj =m(1+ L
) (m − 1)!
(71)
For κ ≥ 80L + 2, we split the integral from (70) into pieces B1κ , B2κ (whose definition depends on
κ), which we will bound individually.
In particular, we use
B1κ
=
Z
≤
Z
∞
ηj =m(1+ κ−1
L )
1/4
∞
ηj =m(1+ κ−1
L )
1/4
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj
(m − 1)!
ηjm−1 e−ηj
dηj
(m − 1)!
1/4
≤ e−d3 m(κ/L)
(72)
and
30
B2κ
m(1+ κ−1
L )
1/4
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | Ht , ηj dηj
=
1
ηj =m(1+ L
) (m − 1)!
1/4 ! 1/4 d4 γmζ(vt )
κ
−
1
eL
≤ Pr dn (j) ≥ ∆vγt Ht , ηj ≤ m 1 +
≤
(73)
L
κ1/4
Z
to bound the integral in (70) by B1κ + B2κ for all κ ≥ 80L + 2.
Therefore, gathering the many terms together (and using that κ ≤
on m large to allow crude upper bounding, we see that
2nLγ 2
vt
from (51)) and relying
γL
2
E(rγvt | Ht ) . L max {0, (1 − γ)} [from (60)] + 10Le−m/(2L ) [from (66)]
ηvt
2
2nLγ
X/i eL1/2 d0 γmζ(vt )
−d1 γmζ(vt )/L2
−m/(2L2 )
+ 10Le
[from (69)] + 2 + e
[from (66) and (69)]
1/2
κ
κ=10L
d2 m
+ 6[from (69)] + 100L exp − 2 [from (71)]
L
2
1/4 d4 γmζ(vt ) !
2nLγ /i
X
eL
1/4
+
[from (72) and (73)]. (74)
e−d3 m(κ/L) +
1/4
κ
κ=80L+2
We first observe that if vnt < γ102 then the summations κ = 10L, . . . , 2nLγ 2 /vt etc. above are
empty. For larger n/vt we can therefore assume that γm(n/vt )1/2 ≥ m which implies (see (62))
that γmζ(vt ) ≥ m/2 and then we can assume that
2nLγ 2 /vt
X eL1/2 d0 γmζ(vt )
1
≤
and
1/2
κ
1000
κ=10L
2nLγ 2 /vt
X eL1/4 d4 γmζ(vt )
1
≤
.
1/4
κ
1000
κ=80L+1
(75)
Plugging these estimates into (74) and making some simplifications, we obtain (53).
Going back to (52) we have
vt
2sγ
γ 3L
Ht ≤ L max {0, (1 − γ)}
E
ηvt
mi
−m/(2L2 )
+ 200Le
−d1 γmζ(vt )/L2
+ 100Le
2
2nLγ
X/vt 2κ eL1/2 d0 γmζ(vt )
L
κ1/2
κ=10L
1/4 d4 γmζ(vt ) !
eL
1/4
e−d3 m(κ/L) +
.
κ1/4
−m/(2L2 )
+ 2+e
2
2nLγ
X/vt 2κ
d
m
2
+ 12 + 104 L exp − 2 +
L
L
κ=80L+2
Making similar estimates to what we did for (75) gives us (54).
We obtain (55) from (P5), (66), (69), (72) and (73). Indeed, if Jγvt ,κ 6= ∅ then from its definition
2n
1/2
we must have vt ≤ 2Lγ
. Together with vt ≤ n/5 we obtain that ζ(vt ) ≥ 2Lκ1/2 γ . Thus, in this
κ−1
31
case,
eL1/2
κ1/2
d0 γmζ(vt )
eL1/2
(log log n)2
d0 m(log log n)2 /2L1/2
1
≤
=o
.
(76)
log10 n
This deals with the probabilities in (66) and (69). For (69) we rely m large to to show that
1/4
e−d3 m(κ/L) = o(1/ log10 n). Equation (72) is dealt with in a similar manner to (66). Here we
1/4 d4 γmζ(vt )
which is the square root of (76).
have eL
κ1/4
3
Case 2: n/5 ≤ vt ≤ n 1 − 3L
:
m
The upper bound on vt implies that
mζ(vt ) ≥ L3 .
Using the same definitions of rγvt ,κ,l , l = 1, 2, 3 as above:
X
E(rγvt ,κ,1
| Ht ) ≤
X X
≤
X X
κ≥1 j∈Jγvt ,κ
κ≥1
κ≥1 j∈Jγvt ,κ
ηvt
2(vt j)1/2
Z (1− 1 )m m−1 −ηj
L
ηj e
dηj ,
(m − 1)!
ηj =0
ηvt
2
e−m/(2L ) ,
1/2
2(vt j)
from Lemma 2.1(e),
1/2
ηvt n
2
≤
e−m/(2L )
γ vt
51/2 ηvt −m/(2L2 )
e
.
≤
γ
X
E(rγvt ,κ,2 | Ht )
κ≥1
.
X X
ηvt
2(vt j)1/2
Z
X X
ηvt
2(vt j)1/2
Z
κ≥1 j∈Jγvt ,κ
.
κ≥1 j∈Jγvt ,κ
(1+1/L)m
ηj =(1−1/L)m
(1+1/L)m
ηj =(1−1/L)m
2
X ηv −d5 γL
t
4e
≤
d γL3
γL
κ≥1
eL1/2 6
κ1/2
X
E(rγvt ,κ,3
| Ht ) ≤
κ ≤ 3,
4 ≤ κ ≤ 10L,
κ > 10L,
X X
κ≥1 j∈Jγvt ,κ
κ≥1
≤
ηjm−1 e−ηj
Pr dn (j) ≥ ∆vγt | ηj dηj
(m − 1)!
κ ≤ 3,
1 n
o
ηjm−1 e−ηj
d5 γmζ(vt )
4 ≤ κ ≤ 10L,
· exp − L2
(m − 1)!
(e(L/κ)1/2 )d6 γmζ(vt ) κ > 10L,
X X
κ≥1
v ,κ
j∈Jγ t
ηvt
2(vt j)1/2
Z
∞
ηj =(1+1/L)m
ηvt
−m/(3L2 )
e
,
2(vt j)1/2
32
ηjm−1 e−ηj
(m − 1)!
from Lemma 2.1(d),
1/2
ηvt −m/(3L2 ) n
e
≤
γ
vt
1/2
5 ηvt −m/(3L2 )
≤
e
.
γ
The above upper bounds are small enough to give the lemma in this case, without trouble.
We are now in a position to prove (20). We confirmed the second part of the statement (20)
above, using (55), so only the first part remains. The first part follows immediately from Lemma
3.11 and the following, by addition:
Lemma 3.16.
E(ρt 1vt+1 ≥vt | Ht ) ≤
Proof. We consider cases.
Case 1: n1 ≤ vt ≤ n 1 −
3L3
m
21ηvt
L3
+ 2.
mL
m
: Then,
E(ρt 1vt+1 ≥vt | Ht ) ≤ I1 + I2 + I3 + I4 ,
where
5/4
I1 =
Z
≤
Z
γ=1/8
5/4
γ=1/8
≤
E(ρt 1vt+1 ≥vt | γv∗t = γ)d Pr(γv∗t ≤ γ),
1000ηvt
m
−c1 m
2 × 83
× ηvt ×
mL
Z
,
≤ ηvt e
Z 10000
I2 =
γ=5/4
≤
I3 =
7L
+ 13 + 100Le−c2 γL
8
d Pr(γv∗t ≤ γ),
by Remark 3.14, Lemma 3.15,
5/4
γ=1/8
d Pr(γv∗t ≤ γ),
from (35).
(77)
2
× ηvt × (13 + 100Le−c2 γL ) d Pr(γv∗t ≤ γ),
mγ 3 L
20ηvt
.
ZmL
∞
γ=10000
(78)
2
× ηvt × (13 + 100Le−c2 γL ) d Pr(γv∗t ≤ γ)
mγ 3 L
Z ∞
27ηvt
γ −cm dγ,
from (36),
≤ 15
10 Lm γ=100
27ηv
1
= 15 t × 2(cm−1)
.
10 Lm 10
(cm − 1)
Z 1/8
I4 =
E(ρt 1vt+1 ≥vt | γv∗t = γ)d Pr(γv∗t ≤ γ)
γ=0
33
(79)
1/2
≤ e−d0 m
,
(80)
1/2
To obtain the term e−d0 m in (80) we use (33) and (34) to obtain
max γ ∈ [0, 1/8] : γ −2 Pr(γ ∗ ≤ γ) ≤
o
n
2
1/2
2
max γ ∈ [0, 1/8] : (γ 1/2−4/(c1 m ) e1−γ )c1 m +
oo
n
n
−2
−c1 mγ 1/2 −c1 /γ 1/2
max γ ∈ [0, 1/8] : mγ min e
,e
2 )−1/2
≤ (84/(c1 m
1/2
e1−1/8
2
1/2
)c1 m + m2 e−c1 m
.
The first case of the lemma now follows from (77), (78), (79) and (80).
3
:
Case 2: vt > n 1 − 3L
m
3
We observe first that n ≤ vt 1 + 4L
. Then we let Z = dn (vt ) − m be the number of right
m
neighbors of vt . Furthermore,
n
X
wi
.
E(Z | Ht ) .
W
j
j=v +1
t
3
vt 1+ 4L
m
X
j=vt +1
ηvt
. ηvt
2(vt j)1/2
!
1/2
4L3
4L3
1+
− 1 ≤ ηvt
.
m
m
Case 2a: ηvt ≥ 1/L1/2 .
We use (81) and Lemma 2.2(d) to prove
ηvt
1/2
Pr Z ≥
Ht ≤ e−d1 ηvt L ≤ e−d1 L .
L
(81)
(82)
Then we can write
4L3
2ηvt
3ηvt
1/2
E(ρt 1vt+1 ≥vt | Ht ) ≤ 1 +
× e−d1 L +
≤
.
m
Lm
Lm
3
Explanation: ρt will be at most 1 + 4L
if the unlikely event in (82) occurs. Failing this, the
m
chance that ρt > 1 is at most
2Z
m
≤
2ηvt
.
Lm
Case 2b: ηvt < 1/L1/2 .
It follows from (81) that E(Z | Ht ) . 4L5/2 /m. It then follows from Lemma 2.2(d) that
L3
1/2
Ht ≤ e−d2 L .
Pr Z ≥
3m
We then have
E(ρt 1vt+1 ≥vt | Ht ) ≤
4L3
1+
m
34
× e−d2 L
1/2
+
2L3
L3
≤
.
3m2
3m2
Part 5
We now prove (21). To do this, we will obtain a recurrence for E(ηvt+1 | Ht ), and, at the end,
obtain the bound 4m by averaging over the possible histories Ht .
We begin by writing
E(ηvt+1 | Ht ) = E(ηvt+1 1vt+1 <vt | Ht ) + E(ηvt+1 1vt+1 >vt | Ht )
(83)
We consider each term in (83) separately. For the first term, since
ηvt+1 1vt+1 <vt ≤ max {ηl : 1 ≤ l < vt , l ∈ NL (vt )} 1vt+1 <vt ≤ max {ηl : 1 ≤ l < vt , l ∈ NL (vt )} ,
we have that
E(ηvt+1 1vt+1 <vt | Ht ) ≤ E (max {ηl : 1 ≤ l < vt , l ∈ NL (vt )} | Ht )
Z ∞
=
Pr(max {ηl : 1 ≤ l < vt , l ∈ NL (vt )} ≥ η | Ht )dη
η=0
Z ∞
=
Pr(∃1 ≤ l < vt , l ∈ NL (vt ) : ηl ≥ η | Ht )dη
η=0
≤
.
Z
t −1
∞ vX
η=0 l=1
Z ∞ vX
t −1
η=0 l=1
vX
t −1 Z ∞
Pr(l ∈ NL (vt ) and ηl ≥ η | Ht )dη
wl
Wvt
Z ∞
Z
∞
ηl =η
ηlm−1 e−ηl
dηl dη
(m − 1)!
ηlm−1 e−ηl
ηl
dηl dη
1/2 (m − 1)!
2m(lv
t)
η=0
η
=η
l
l=1
Z ∞ Z ∞
ηlm e−ηl
dηl dη
. 2m + (1 + o(1))
η=2m ηl =η (m − 1)!
Z ∞
. 2m +
4e−3η/10 dη, from Lemma 2.1(c),
.
η=2m
−3m/5
. 2m + 20e
≤ 3m.
(84)
We now bound the second term of (83). We consider two cases, according to properties of the
history Ht (which determines vt and ηvt ).
1
Case 1: Ht is such that vt ≤ 1 − ω1/2
n.
In this case, we have that
E(ηvt+1 1vt+1 >vt | Ht )
o
n
≤ E max ηl : vt < l ≤ n, vt+1 ∈ NL (l), dn (l) ≥ ∆vγtv∗ 1vt+1 >vt | Ht
t
35
o
n
≤ E max ηl : vt < l ≤ n, vt+1 ∈ NL (l), dn (l) ≥ ∆vγtv∗ | Ht .
t
So we have that
E(ηvt+1 1vt+1 >vt | Ht )
o
n
vt
≤ E max ηl : vt < l ≤ n, vt+1 ∈ NL (l), dn (l) ≥ ∆γv∗ | Ht
t
Z ∞
o
n
vt
=
Pr(max ηl : vt < l ≤ n, vt ∈ NL (l), dn (l) ≥ ∆γv∗ ≥ η | Ht )dη
t
η=0
Z ∞
=
Pr ∃vt < l ≤ n, vt ∈ NL (l) : ηl ≥ η, dn (l) ≥ ∆vγtv∗ | Ht dη
≤
η=0
n
X
l=vt +1
n
X
t
Z
∞
η=0
Pr (vt ∈ NL (l)) ∧ (ηl ≥ η) ∧ (dn (l) ≥ ∆vγtv∗ ) | Ht , ηl dη
ηvt
.
2(lvt )1/2
l=v +1
t
t
Z
∞
η=0
Z
∞
ηl =η
ηlm−1 e−ηl
Pr dn (l) ≥ ∆vγtv∗ | Ht , ηl dηl dη.
t
(m − 1)!
(85)
Recall that in the final two lines, vt and ηvt are not random variables, but are the actual values of these random variables in the history Ht , so this is a deterministic upper bound on
E(ηvt+1 1vt+1 >vt | Ht ).
We split the sum in the RHS of (85) into E1 + E2 + E3 + E4 according to the ranges of l and η,
and bound each separately. The first part consists of
Z
Z
n
X
ηvt 1l≤4m2 vt /(γv∗t )2 2m ∞ ηlm−1 e−ηl
vt
|
H
,
η
Pr
d
(l)
≥
∆
E1 =
∗
t l dηl dη.
n
γvt
1/2
2(lv
(m
−
1)!
t)
η=0
η
=η
l
l=v +1
t
Even though vt and ηvt are constants (determined by Ht ), we caution that γv∗t and so also E1 are
random variables.
Observe that we have that
Z ∞ m−1 −ηl
Z ∞ m−1 −ηl
ηl e
ηl e
vt
Pr dn (l) ≥ ∆γv∗ | Ht , ηl dηl ≤
dηl ≤ 1
t
ηl =η (m − 1)!
ηl =η (m − 1)!
which allows us to write
E1 1γv∗t ≤5/4 ≤ 1γv∗t ≤5/4
n
X
2mηvt 1l≤4m2 vt /(γv∗t )2
l=vt +1
2(lvt )1/2
≤
5m2 ηvt
1γv∗t ≤5/4 .
γv∗t
(86)
We will use this expression when we take the expectation over γv∗t ≤ 5/4.
We also have that
Z ∞ m−1 −ηl
ηl e
Pr (dn (l) ≥ ∆vγt∗ ) ∧ (γv∗t > 5/4) | Ht , ηl dηl ≤ I1 + I2 + I3 ,
ηl =η (m − 1)!
36
(87)
where
I1 =
Z
7m/8
ηl =η
Z ∞
ηlm−1 e−ηl
dηl ≤ m
(m − 1)!
7 1/8
e
8
m
≤ e−d0 m ,
from Lemma 2.1(a),
ηlm−1 e−ηl
dηl ≤ e−d1 m , from Lemma 2.1(d).
I2 =
ηl =9m/8 (m − 1)!
Z 9m/8
ηlm−1 e−ηl
∗
∗
Pr dn (l) − m ≥ γvt mζ(vt ) ∧ (γvt > 5/4) Ht , ηl dηl ,
I3 =
ηl =7m/8 (m − 1)!
Z 9m/8
ηlm−1 e−ηl
5
≤
Pr dn (l) − m ≥ mζ(vt ) Ht , ηl dηl .
4
ηl =7m/8 (m − 1)!
(88)
We bound I3 with two subcases:
Subcase 1a: ζ(l) > 0.
Z 9m/8
8
10ζ(vt)
ηlm−1 e−ηl
I3 ≤
Pr dn (l) − m ≥
ηl ζ(l) Ht , ηl dηl [since m ≥ ηl ],
9ζ(l)
9
ηl =7m/8 (m − 1)!
n
o
2
(10ζ(vt )−9ζ(l)) ηl
Z 9m/8
10ζ(vt) ≤ 18ζ(l)
ηlm−1 e−ηl exp n−
81ζ(l)
o
[from (2) and ℓ ≥ vt + 1],
≤
(10ζ(vt )−9ζ(l))ηl
10ζ(vt) > 18ζ(l)
ηl =7m/8 (m − 1)! exp −
27
≤
Z
9m/8
ηl =7m/8
n
o
exp − 7mζ(vt )
10ζ(vt ) ≤ 18ζ(l)
648
n
o
(m − 1)! exp − 7mζ(vt )
10ζ(vt ) > 18ζ(l)
216
ηlm−1 e−ηl
≤ e−mζ(vt )/100 .
(89)
Subcase 1b: ζ(l) ≤ 0.
In this case, we go back to (88) and use ζ + (l) in place of ζ(l), see (10).
Z 9m/8
10ζ(vt) +
ηlm−1 e−ηl
Pr dn (l) − m ≥
ηl ζ (l) Ht , ηl dηl ,
I3 ≤
9ζ + (l)
ηl =7m/8 (m − 1)!
(90)
For ε as in (46) we see that ζ(l) ≤ 0 implies that l ≥ n(1 − ε)2 . In which case
ζ + (l) ≤
On the other hand, vt ≤ 1 −
1
ω 1/2
2ε
≤ 3ε.
1−ε
(91)
implies that
ζ(vt ) ≥
1
ω 1/2
− 2ε ≥
1
.
2ω 1/2
t)
≥
Comparing (91) and (92), we see that ζ(vt ) ≫ ζ + (l). From this and (3) with β = 10ζ(v
9ζ + (l)
we deduce that
10ζ(vt ) +
ηl ζ (l) Ht , ηl ≤ (6εω 1/2)10ηl ζ(vt ) ≤ (6εω 1/2 )35mζ(vt )/36 .
Pr dn (l) − m ≥
9ζ + (l)
37
(92)
1
6εω 1/2
Plugging this estimate into (90) we obtain something stronger than (89), finishing Subcase 1b
and giving that I3 ≤ e−mζ(vt )/100 in all cases.
Having bounded the three terms in (87), we then have that
E1 1γv∗t >5/4 ≤
n
X
ηvt 1l≤4m2 vt /(γv∗t )2
(lvt )1/2
l=vt +1
5m
e−d2 m ∗
γvt
≤ ηvt
−d2 m
e−d2 m + e−mζ(vt )/100
n
e−mζ(vt )/100 X 1
+
vt 1/2
l1/2
l=v +1
t
−mζ(vt )/100
≤ ηvt 4me
+e
!
(n + 1)1/2 − (vt + 1)1/2
·
vt 1/2
≤ ηvt (4me−d2 m + 2ζ(vt)e−mζ(vt )/100 )
200
−d2 m
.
+
≤ ηvt 4me
m
(93)
It follows from (86) and (93) that
2
200
5m
−d2 m
.
1γ ∗ ≤5/4 + 4me
+
E1 ≤ ηvt
γv∗t vt
m
We continue with the other parts of the RHS of (85):
Z
Z ∞ m−1 −ηl
n
X
ηvt 1l≤4m2 vt /(γv∗t )2 ∞
ηl e
E2 =
Pr(dn (l) ≥ ∆vγt∗ | Ht , ηl )dηl dη
1/2
2(lv
)
(m
−
1)!
t
η=2m ηl =η
l=v +1
t
4m2 vt /(γv∗t )2
≤
X
ηvt
2(lvt )1/2
Z
∞
X
ηvt
2(lvt )1/2
Z
∞
X
ηvt
×m
2(lvt )1/2
l=vt +1
4m2 vt /(γv∗t )2
≤
l=vt +1
4m2 vt /(γv∗t )2
≤
l=vt +1
4m2 vt /(γv∗t )2
=
≤
X
l=vt +1
−d3 m
e
ηvt
.
γv∗t
η=2m
Z
∞
eη/m
eη/m
ηl =η
η=2m
Z
ηlm−1 e−ηl
dηl dη
(m − 1)!
m
dη
from Lemma 2.1(c),
∞
e−3mx/10 dx
x=2
10ηvt e−3m/5
6(lvt )1/2
(94)
Note that we aborbed an O(m) factor into the expression in (94). This is valid because m is
large. We continue to do this where possible.
Z
Z
n
X
ηvt 1l>4m2 vt /(γv∗t )2 γv∗t (l/vt )1/2 ∞ ηlm−1 e−ηl
Pr(dn (l) ≥ ∆vγtv∗ | Ht )dηl dη
E3 =
1/2
t
2(lv
)
(m
−
1)!
t
η=0
ηl =η
l=v +1
t
38
n
X
γv∗t (l/vt )1/2
∞
ηlm−1 e−ηl
dηl dη
∗ (l/v )1/2 (m − 1)!
η=0
η
=γ
t
l
2
∗
2
v
t
l=4m vt /(γvt )
(
)
Z γv∗ (l/vt )1/2
n
X
t
3γv∗t (l/vt )1/2
ηvt
exp −
dη from Lemma 2.1(c),
≤
2(lvt )1/2 η=0
10
2
∗
2
l=4m vt /(γvt )
)
(
n
X
3γv∗t (l/vt )1/2
ηvt
≤
,
exp −
2(lvt )1/2
10
2
∗
2
l=4m vt /(γvt )
)
(
Z
ηvt γv∗t n
3γv∗t (x/vt )1/2
≤
dx,
exp −
vt
10
x=4m2 vt /(γv∗t )2
Z ∞
ηvt γv∗t
8vt
≤
× ∗ 2
ye−3y/5 dy,
vt
(γvt ) y=m
≤
ηvt
2(lvt )1/2
Z
Z
ηvt e−d4 m
=
.
γv∗t
Z
Z ∞ m−1 −ηl
n
X
ηvt 1l>4m2 vt /(γv∗t )2 ∞
ηl e
Pr(dn (l) ≥ ∆vγtv∗ | Ht , ηl )dηl dη
E4 =
1/2
t
2(lvt )
η=γv∗t (l/vt )1/2 ηl =η (m − 1)!
l=i
Z ∞
Z ∞ m−1 −ηl
n
X
ηl e
ηvt
dηl dη
≤
1/2
2(lvt )
η=γv∗ (l/vt )1/2 ηl =η (m − 1)!
2
∗ 2
l=4m vt /(γvt )
≤
n
X
l=4m2 vt /(γv∗t )2
t
ηvt
2(lvt )1/2
Z
∞
e−3η/10 dη
η=γv∗t (l/vt )1/2
)
(
3γv∗t (l/vt )1/2
5ηvt
,
≤
exp −
3(lvt )1/2
10
2
2
∗
l=4m vt /(γvt )
)
(
Z
1/2
∗
(x/j)
3γ
2ηvt ∞
≤ 1/2
dx,
x−1/2 exp − vt
i
10
2
∗
2
x=4m vt /(γvt )
Z
4ηvt ∞ −3y/10
e
dy,
= ∗
γvt y=2m
n
X
≤
ηvt e−d5 m
.
γv∗t
Thus,
E(ηvt+1 1vt+1 >vt
5m∗ 2 + e−d∗6 m ηvt ≤
γvt
γvt
| Ht ) ≤ E1 + E2 + E3 + E4 ≤ −d
m
7
200
e
ηvt
+
∗
γvt
m
7m2
η
γv∗t vt
γv∗t ≤ 5/4
γv∗t > 5/4.
We now integrate with respect to the value of γv∗t . (Note that γv∗t is actually a discrete random
variable, so that Pr(γv∗t ≤ γ | Ht ) is discontinuous, but one can view this as a Riemann-Stieltjes
integral. We write Pr† (γv∗t ≤ γ) below in place of Pr(γv∗t ≤ γ | Ht ).) Using Lemma 3.12 we see
39
that if m is large then integrating over γ,
E(ηvt+1 1vt+1 >vt | Ht )
!
Z ∞
Z 5/4
−d7 m
e
200
7m2
d Pr† (γv∗t ≤ γ) +
d Pr† (γv∗t ≤ γ) +
≤ ηvt
γ
γ
m
γ=5/4
γ=0
!
Z 5/4
7m2
200
≤ ηvt
d Pr† (γv∗t ≤ γ) + e−d8 m +
γ
m
γ=0
!
Z 1/m
Z 5/4
2
7m2
200
7m
= ηvt
d Pr† (γv∗t ≤ γ) +
d Pr† (γv∗t ≤ γ) + e−d8 m +
γ
γ
m
1/2
γ=0
γ=1/m
1/m Z 1/m1/2
2
7m2
7m
† ∗
+
Pr (γvt ≤ γ)
Pr† (γv∗t ≤ γ)dγ
≤ ηvt
2
γ
γ
γ=0
0
!
Z 5/4
7m2
200
+
d Pr† (γv∗t ≤ γ) + e−d8 m +
γ
m
γ=1/m
Z 1/m
7m2
−d9 m1/2
Pr† (γv∗t ≤ γ)dγ+
+
≤ ηvt e
2
γ
γ=0
!
Z 5/4
7m3
200
,
from (34),
d Pr† (γv∗t ≤ γ) + e−d8 m +
m
γ=1/m γ
200
1
†
∗
−d8 m
−d9 m1/2
−d10 m1/2
4
,
≤ γvt ≤ 5/4 + e
+
+e
+ 7m Pr
≤ ηvt e
m
m
200
−d9 m1/4
−d10 m1/2
4 −c1 m
−d8 m
+e
+ 7m e
+e
+
≤ ηvt e
,
from (35)
m
200
−d12 m1/4
+
≤ ηvt e
.
m
(95)
Combining (84) and (95) via (83), we have that
E(ηvt+1
200
−cm1/4
ηvt .
| Ht ) ≤ 3m + e
+
m
This completes Case 1. Case 2 is much shorter.
1
Case 2: Ht is such that vt > 1 − ω1/2
n.
E(ηvt+1 | Ht ) .
∼
∼
n
X
l=vt +1
n
X
l=vt +1
n
X
ηvt
2(lvt )1/2
Z
ηvt
2(lvt )1/2
Z
mηvt
2(lvt )1/2
l=v +1
t
40
∞
η=0
Z
∞
ηl =η
∞
−η
e
η=0
ηlm−1 e−ηl
dηl dη
(m − 1)!
m
X
i=1
η m−i
dη
(m − i)!
(96)
. mηvt
≤
(n + 1)1/2 − (vt + 1)1/2
(vt + 1)1/2
mηvt
.
ω 1/2
This completes Case 2. In particular, for sufficiently large n we see that for any typical Ht (i.e.,
in both Case 1 and Case 2), the bound from (96) is valid. Putting
Et = E ∩ {Ht is typical}
we deduce from (96) that
E(ηvt+1
200
−cm1/4
| Et ) ≤ 3m + e
+
E(ηvt | Et )
m
200
−cm1/4
E(ηvt | Et−1 ).
. 3m + e
+
m
(97)
(98)
We obtain (98) from (97) because Et ⊆ Et−1 and so
E(ηvt | Et−1 ) ≥ E(ηvt | Et ) Pr(Et | Et−1 ) ∼ E(ηvt | Et ).
Because m is large, (21) will follow by induction once we have shown that
E(ηv1 ) ≤ 3m.
(99)
Here we will use the assumption that v1 is chosen exacty according to the stationary distribution
for a simple random walk on Gn . In particular, we have
!
n
X
dn (i)
Pr(ηv1 ≥ η) ≤ E
1η ≥η ,
2mn v1
i=1
and Lemma 2.2 implies that if η ≥ 2m
E((dn (i) − m)1ηv1 ≥η ) .
n 1/2 Z
i
∞
η=2m
n 1/2
η m e−η
dη .
× 4me−η/2 .
(m − 1)!
i
Furthermore,
E(m1ηvt ≥η ) = m Pr(ηv1 ≥ η) ≤ 5me−η/2 .
So, if η ≥ 2m, then
Pr(ηv1
Therefore,
E(ηv1 ) ≤ 2m +
Z
n
9e−η/2 X 1
≤ 10e−η/2 .
≥ η) .
2n1/2 i=1 i1/2
∞
Pr(ηv1 ≥ η)dη ≤ 2m + 10
η=2m
Z
∞
η=2m
and (99) follows.
41
e−η/2 dη
3.2
Exiting the main loop with success
In summary, it follows that w.h.p. DCA reaches Step 7 in O(ω log n) time. Also, at this time
vT ≤ log1/49 n. This follows from Lemma 2.2(g), 2.2(h) and (P4). Furthermore, this justifies
using n1 as a lower bound on vertices visited during the main loop. The random walk of Step
8 will w.h.p. take place on [log1/9 n]. This follows from Lemma 2.2(j). Vertex 1 will be in the
1/2
same component as vt in the subgraph of Gn induced by vertices of degree at least logn1/20 n . This
is because there is a path from vT to vertex 1 through vertices in [vT ] only and furthermore it
1/2
follows from Lemma 2.2(i) that w.h.p. every vertex on this path has degree at least logn1/20 n .
The expected time to visit all vertices of a graph with ν vertices is O(ν 3 ), see for example
Aleliunas, Karp, Lipton, Lovász and Rackoff [1]. Consequently, vertex 1 will be reached in a
further O((log1/9 n)3 ) = o(log n) steps w.h.p, completing the proof of Theorem 1.2.
4
Concluding remarks
We have described an algorithm that finds a distinguished vertex quickly and which is local in a
strong sense. There are some natural questions that are left unanswered:
• Can the running time be improved from O(ω log n) to O(log n)?
• Can we get polylog expected running time for DCA if m = 2?
• Can we extend the analysis to other more general models of web graphs e.g. Cooper and
Frieze [7]. In this case, we would not be able to use the model described in Section 2.
As a final observation, the algorithm DCA could be used to find the vertex of largest degree: if
we replace Step 8 by “Do the random walk for log n steps and output the vertex of largest degree
encountered” then w.h.p. this will produce a vertex of highest degree. This is because log n will
be enough time to visit all vertices v ≤ log1/39 n, where the maximum degree vertex lies.
References
[1] R. Aleliunas, R.M. Karp, R.J. Lipton, L. Lovász and C. Rackoff, Random Walks, Universal
Traversal Sequences, and the Complexity of Maze Problems. Proceedings of the 20th Annual
IEEE Symposium on Foundations of Computer Science (1979) 218-223.
[2] A. Barabási and R. Albert, Emergence of scaling in random networks, Science 286 (1999)
509-512.
[3] B. Bollobás, O. Riordan, J. Spencer and G. Tusnády, The degree sequence of a scale-free
random graph process, Random Structures and Algorithms 18 (2001) 279-290.
42
[4] B. Bollobás and O. Riordan, The Diameter of a Scale-Free Random Graph, Combinatorica
24 (2004) 5-34.
[5] C. Borgs, M. Brautbar, J. Chayes, S. Khanna and B. Lucier, The Power of Local Information
in Social Networks, http://arxiv.org/abs/1212.0884.
[6] M. Brautbar and M. Kearns, Local Algorithms for Finding Intersting Individuals in Large
Networks, in Innovations in Computer Science (ITCS) (2010) 188-199.
[7] C. Cooper and A.M. Frieze, On a general model of web graphs, Random Structures and
Algorithms 22 (2003) 311-335.
[8] G. Grimmett and D. Stirzaker, Probability and Random Processes, Third Edition, Oxford
University Press, Oxford UK, 2001.
[9] W. Hoeffding, Probability inequalities for sums of bounded random variables, Journal of the
American Statistical Association 58 (1963) 13-30.
[10] M. Jerrum and A. Sinclair. The Markov chain Monte Carlo method: an approach to approximate counting and integration. In Approximation Algorithms for NP-hard Problems.
(D. Hochbaum ed.) PWS (1996) 482-520.
[11] M. Mihail, C. Papadimitriou and A. Sabeeri, On Certain Connectivity Properties of the
Internet Topology, Journal of Computer and System Sciences 72 (2006) 239251.
A
Proofs of properties (P1)–(P5)
In this section we give proofs of (P1)–(P5), which we list here for convenience.
(P1) For Υk,ℓ = Υk − Υℓ , we have
Υk,ℓ
"
1/2
Lθk,ℓ
∈ (k − ℓ) 1 ±
3(k − ℓ)1/2
#
for (k, ℓ) = (mn + 1, 0) or
l=0
1
k−ℓ
2
∈ {ω, ω + 1, . . . , n} and k − l ≥ log n
k ≥ log30 n, l > 0
m
1/300
log
n 0 < l < k < log30 n.
43
Here n0 =
λ20 n
,
ω log2 n
λ0 =
1
,
log20/m n
log k
1/2
k
= (k − ℓ)1/2
(k−ℓ)3/2 log n
n1/2
n
θk,ℓ
ω ≤ l < k ≤ log30 n, l > 1
ω ≤ k ≤ n2/5 , l = 0
log30 n < k ≤ n2/5
n2/5 < k ≤ n0 .
n0 < k.
ω 3/2 log2 n
#
1/2 "
1/2
1/2
i
Lθi
i
(P2) Wi ∈
1 ± 1/2 ∼
for ω ≤ i ≤ n.
n
i
n
(P3) wi ∼
ηi
for ω ≤ i ≤ n.
2m(in)1/2
(P4) λ0 ≤ ηi ≤ 40m log log n for i ∈ [log30 n].
(P5) ηi ≤ log n for i ∈ [n].
Proof of (P1)
Applying Lemma 2.1(d),(e) to (1) for i ≥ 1 we see that
Pr(¬(P1))
n
X
L2 θk,0
≤2
exp −
+2
27
k=ω
n2/5
L2 k 1/2
exp −
=2
27
k=ω
X
30
n
X
k−ℓ=log1/300 n
+2
n0
X
k=n2/5 +1
L2 θk,ℓ
exp −
27
L2 θmn+1,0
+ 2 exp −
27
L2 k 3/2 log n
exp −
27n1/2
+2
n+1
X
k=n0
L2 n
exp −
27ω 3/2 log2 n
+1
n2/5
X
L2 log k
L2 (k − ℓ)1/2
exp −
+2
+2
exp −
+
27
27
30
1/300
k−ℓ=log n
k−ℓ=log
n
n
n
0
X
X
L2 n
L2 (k − ℓ)3/2 log n
+2
exp −
2
exp −
27n1/2
27ω 3/2 log2 n
2/5
k−ℓ=n +1
log
k−ℓ=n
Xn
+1
0
= o(1).
Proof of (P2)
For this we use
Wi =
Υmi
Υmn+1
44
1/2
.
Then,
#
1/2 "
1/2
Lθi
i
Wi ∈
/
1 ± 1/2
n
i
implies that either
Υmn+1
"
1/2
Lθi
∈
/ (mn + 1) 1 ±
3(mn + 1)1/2
#
or Υmi
"
1/2
Lθi
∈
/ mi 1 ± 1/2
3i
#
.
These events are ruled out w.h.p. by (P1).
Proof of (P3)
We use (1 + x)1/2 ≤ 1 +
wi =
x
2
for 0 ≤ |x| ≤ 1. Then,
Υmi
Υmn+1
1/2
−
Υm(i−1)
Υmn+1
1/2
!
1/2
Υm(i−1) 1/2
ηi
=
1+
−1
Υmn+1
Υm(i−1)
1/2
1/2
Lθi
m(i − 1) 1 + 3m1/2 (i−1)1/2
η
i
≤
1/2
1/2
Lθi
2m(i − 1) 1 −
(mn + 1) 1 −
1/2
3(mn+1)
ηi
≤
2m(in)1/2
1/2
2Lθ
1 + 1/2i 1/2
m i
!
1/2
Lθi
3m1/2 (i−1)1/2
.
A similar calculation gives
1/2
ηi
wi ≥
2m(in)1/2
2Lθ
1 − 1/2i 1/2
m i
!
.
Proof of (P4)
The upper bound follows from Lemma 2.1(c). For the lower bound, we observe by (7) that the
expected number of i ≤ log30 n with ηi ≤ λ0 is at most log30 n × λm
0 = o(1).
Proof of (P5)
This follows from Lemma 2.1(c).
45
B
Proof of Lemma 2.2
We restate the lemma for convenience.
Lemma B.1.
(a) If E occurs then d¯n − m ∈ [ηi ζ(i), ηi ζ + (i)].
2 η ζ(i)/2
i
(b) Pr(dn (i) − m ≤ (1 − α)mζ(j)) ≤ e−α
for 0 ≤ α ≤ 1.
2 η ζ + (i)/3
i
(c) Pr(dn (i) − m ≥ (1 + α)mζ + (j)) ≤ e−α
(d) Pr(dn (i) − m ≥ βmζ + (j)) ≤ (e/β)βηi ζ
+ (i)
for 0 ≤ α ≤ 1.
for β ≥ 2.
(e) W.h.p. ηi ≥ λ0 and ω ≤ i ≤ n1/2 implies that dn (i) ∼ ηi
(f ) W.h.p. ω ≤ i ≤ log30 n implies that dn (i) ∼ ηi
n 1/2
.
i
(g) W.h.p. ω ≤ i ≤ n1/2 implies that dn (i) . max {1, ηi }
(h) W.h.p. n1/2 ≤ i ≤ n implies dn (i) ≤ n1/3 .
(i) W.h.p. 1 ≤ i ≤ log1/49 n implies that dn (i) ≥
(j) W.h.p. dn (i) ≥
n
log1/20 n
n 1/2
.
i
n 1/2
i
.
n1/2
.
log1/20 n
implies i ≤ log1/9 n.
Proof. (a) Suppose that we fix the values for W1 , W2 , . . . , Wn . Then the degree dn (i) of vertex i
can be expressed
n X
m
X
dn (i) = m +
ζj,k
j=i k=1
where the ζj,k are independent Bernouilli random variables such that
wi wi
Pr(ζj,k = 1) ∈
,
.
Wj Wj−1
So, putting
d¯n (i) = E(dn (i))
we have
mwi
n
n
X
X
1
1
≤ d¯n (i) − m ≤ mwi
.
W
W
j
j
j=i−1
j=i
Now assuming that (P2), we have for ω ≤ i ≤ n,
n 1/2
n
X
X
n
1
≥
Wj
j
j=i
j=i
46
1/2
2Lθj
1 − 1/2
j
!
.
But
1/2
n
X
θj
j=ω
j
2/5
n
X
1
≤
+
3/4
j
j=ω
n0
X
j=n2/5 +1
n
X
n1/2
log1/2 n
+
n1/4 j 1/4 j=n +1 jω 3/4 log n
0
3n1/2 log log n
4n1/2
+
3ω 3/4 log n
ω 3/4 log n
4n1/2 log log n
≤
.
ω 3/4 log n
≤ 4n1/10 +
It follows that
9Ln1/2 log log n
1/2
1/2
1/2
¯
2(n − (i + 1) ) −
dn (i) ≥ m + mwi n
ω 3/4 log n
!
1/2
n 1/2
i
9L log log n
1
1−
,
− 1/2 1/2 −
≥ m + ηi
i
n
n i
2ω 3/4 log n
after using (P3).
A similar calculation gives a similar upper bound for d¯n (i) and this proves that
"
#
1/2
n 1/2
i
5L
log
log
n
i ≥ ω implies that d¯n (i) ∈ m + ηi
1−
.
± 3/4
i
n
ω log n
It follows from (2) and (4) that
L2 ηi n1/2
Pr dn (i) − m ≤ (1 − α)ηi ζ(i) ηi ≤ exp − 1/2 1/2 .
4i ω
L2 ηi n1/2
Pr dn (i) − m ≥ (1 + α)ηi ζ(i) ηi ≤ exp − 1/2 1/2 .
4i ω
(a) For ηi ≥ λ0 and ω ≤ i ≤ n0 we have
L2 ηi n1/2
2
exp − 1/2 1/2 ≤ e−L log n/4 .
4i ω
(b) This follows from (a) and (4).
(c) This follows from (a) and (2).
(d) This follows from (a) and (3).
(e) This follows from (a), (b), (c) and (12).
(f) This follows from (e) and (P4).
47
(g) This follows from (c) and (12).
(h) The degree of i ≥ n1/2 is stochastically dominated by the degree of n1/2 . Also, the probability
that dn (n1/2 ) exceeds the stated upper bound is o(1/n). So (h) follows from (g).
(i) For ω ≤ i ≤ log1/49 n, this follows from (f) and (P4). For 1 ≤ i < ω we can use (b) with
ηi ≥ λ0 and α = n−1/10 .
(j) This follows from (e), (f) and (g) and (P4).
48
| 8cs.DS
|
On Game-Theoretic Risk Management
(Part Three)
Modeling and Applications
Stefan Rass ∗
[email protected]
arXiv:1711.00708v1 [q-fin.EC] 2 Nov 2017
November 3, 2017
Abstract
The game-theoretic risk management framework put forth in the precursor reports “Towards a Theory of Games with Payoffs that are ProbabilityDistributions” (arXiv:1506.07368 [q-fin.EC]) and “Algorithms to Compute
Nash-Equilibria in Games with Distributions as Payoffs” (arXiv:1511.08591v1
[q-fin.EC]) is herein concluded by discussing how to integrate the previously
developed theory into risk management processes. To this end, we discuss
how loss models (primarily but not exclusively non-parametric) can be constructed from data. Furthermore, hints are given on how a meaningful game
theoretic model can be set up, and how it can be used in various stages of
the ISO 27000 risk management process. Examples related to advanced persistent threats and social engineering are given. We conclude by a discussion
on the meaning and practical use of (mixed) Nash equilibria equilibria for
risk management.
∗
Universität Klagenfurt, Institute of Applied Informatics, System Security Group, Universitätsstrasse
65-67, 9020 Klagenfurt, Austria. This work has been done in the course of consultancy for the EU
Project HyRiM (Hybrid Risk Management for Utility Networks; see https://hyrim.net), led by the
Austrian Institute of Technology (AIT; www.ait.ac.at). See the acknowledgement section.
1
Contents
1 Introduction
3
2 Game-Theory based Risk Management
4
3 Modelling Losses
3.1 Parametric Loss Models . . . . . . . . . .
3.2 Nonparametric Loss Models . . . . . . . .
3.3 Collecting Data from Experts . . . . . . .
3.4 Refining the Expert Survey . . . . . . . .
3.5 Outlier Elimination . . . . . . . . . . . . .
3.6 Harmonizing Risk Attitudes . . . . . . . .
3.7 Using the Bandwidth Parameter to Model
3.8 Defining Risk Assessment Categories . . .
3.9 Using Continuous Scales . . . . . . . . . .
3.10 Preparing the Loss Model . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
Answer Uncertainty
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Risk Prioritization
11
11
12
12
15
15
15
15
16
16
18
20
5 Examples of Game-Theoretic Modeling
22
5.1 Modeling APTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.2 Modeling Social Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6 Working with Game-Theoretic Risk Measures
6.1 Choosing the Cutoff Point . . . . . . . . . . . . . . . . . . .
6.2 Interpretation and Meaning of Equilibria . . . . . . . . . . .
6.3 Optimizing an Infrastructure’s Resilience (Risk Treatment)
6.4 Software Support . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Playing the Equilibrium Strategies . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
32
32
33
35
38
40
7 On Model Validation and Verification
41
8 Conclusion and Outlook
44
2
1 Introduction
With algorithmic matters of game-theoretic risk management being covered in [37], it
remains to discuss a few (among more existing) possibilities of how game models can be
used in risk management.
First, observe that matrix games directly cover plain optimization as the special case
of an either n × 1 or 1 × m game. Both have applications in risk management, such as
helping with the following common subtasks:
• If the current security configuration is to be assessed against a number m of (new)
threats, we can think of the defender having only 1 strategy (the current state).
The equilibrium in terms of the -ordering is then the most severe threat (since
the attacker maximizes).
• Likewise, if several options for mitigating a particular threat are available, then
the equilibrium being the -minimum determined by the defender, is the action
leaving the -least damage when being mounted.
• The general case of n > 1 and m > 1 strategies for both, the defender and attacker,
is discussed in the remainder of this article.
Unique Selling Points: The method described in the following offers a variety of advantages relative to the “conventional” approach to managing risks, which is usually tied
to intensive discussion (meetings), and also up to difficult consensus finding. On the
contrary, we propose a method that is based on questioning experts individually, separately, asynchronously and anonymously, which entails the following features, among
well known improvement of the so-obtained data quality [31]:
1. Distributed expert interviews that do not require to be at a certain room at a
certain time (no meetings), and thus allow provide input in between the normal
workflow (asynchronously to the input of other experts).
2. Since the questioning is done individually, it can be done anonymized. This avoids
social or cultural effects that may change a person’s statement spoken out loud in
presence of certain other people (superiors, subordinates, etc.)
3. Exploitation of matrix organization: it may well be the case that experts are
well informed about certain aspects of a problem, but have no reliable clue on
some other aspects. Polling people by online questionnaires in the privacy of
their own office allows them to answer only those parts of the survey that they
can offer input for, while leaving them a safe way of refraining from answering
other questions. Since our method works on probability distributions, the resulting
dataset from which these are compiled may be richer or sparser, depending on how
many informed opinions are available. In any case, however, asking people face-toface in a meeting can bring up an uninformed guess just to have said something,
so that the overall data quality is not necessarily as good as in a distributed and
individual interview.
3
4. Enlargement of the opinion pool: without the need for a personal (and confidential)
meeting, people even outside the company can be included in the risk assessment.
For example, matters of reputation can more reliably be assessed by customers
(which are typically not invited to internal management meetings).
5. The use of distributions also avoids problems of consensus finding or opinion pooling (see, e.g., [5, 20] to mention only two out of many publications in this area)
towards a representative number. It is in that sense preserving all information,
since all opinions (all available data) goes into the decision process with equal
weight.
As a final pro that a game theoretic risk assessment allows, it is even possible to offer
the game solution algorithms as a webservice, where the modeling and threat assessment can be left to the customer enterprise. For example, if the client has identified a
set of threats and countermeasures, it will prefer to not disclose this information to a
subcontractor in charge of risk management. Since the game model and solution can be
formulated using abstract names only, it is possible to anonymize the data by letting
the threats be only named “T1”, “T2”, . . . , “Tn”, as well as the countermeasures be
called “C1”, “C2”, . . . , “Cm”. The real meaning of threats and countermeasures can
then remain private information of the company, whereas the game model being given
in terms of these abstract identifiers remains solvable by any third party contractor.
This allows the method to be offered as a (web-)service, without running into troubles
of unwanted information disclose (even the number of threats and countermeasures can
be disguised, if a company adds dummy copies of actions to the list to make it longer
than it actually is).
2 Game-Theory based Risk Management
An eloquent and detailed comparison of how game theory fits into and aids the classical
risk management process has been given by [34]. We follow their presentation hereafter,
while instantiating and adapting the specific steps to the concrete setting outlined in
the precursor parts to this work [35, 37].To get started, consider the classical ISO risk
management process as depicted in Figure 1, and compare it to the workflow to be
completed when a game-theoretic model is to be set up; shown in Figure 2.
The workflow in Figure 2 needs some explanation, in order to establish a mapping
to the risk management process. Perhaps the most important difference between risk
management and game theory is the former being about minimization of losses caused
by a not necessarily rational opponent (nature, but possibly also a hostile party that
has explicit intentions). Contrary to this, game theory in any case assumes a rational
opponent, whose goal is maximizing the own revenue. A conflict/competition arises if the
revenues for both players are negatively correlated, in the extreme case culminating in the
well-known zero-sum competition, meaning “my gain is your loss” (and vice versa). This
is the scenario that we also assume for risk management based on distribution-valued
game theory, although bearing in mind that the incidents that the risk management
4
Establish Context
Risk Analysis
Risk Identification
Risk Estimation
Risk Evaluation
Risk Decision Point 1
Assessment satisfactory?
Risk Monitoring and Review
Risk Communication
Risk Assessment
Risk Treatment
Risk Decision Point 2
Accept risks?
Risk Acceptance
Figure 1: ISO/IEC 27005 Risk Management Process [16]
1. investigate the scenario
2. identify the players
3.2. determine the
strategies
for each player
3.1. available information
for decision making
3.3. Identify preferences
3.3. Represent by Payoff/
Utility
3.4. Define Scale &
Weight
4. fill in the matrix (normal form)
5. compute the equilibrium
Figure 2: Workflow of setting up a Game Theoretic Model (cf. [34])
5
refers to may not always follow a hidden rationale or agenda. Nevertheless, the zerosum assumption (even though perhaps unrealistic) provides us with a valid worst-case
assessment, and reality can only look better than predicted for the risk manager. The
players engaging in the risk management process may be diverse and many, depending
on the variety of threats to be considered. In mapping this to a game-theoretic model,
we collect all physically existing opponents into a single adversary acting as “player 2”,
against the risk manager, which is player 1. As for the adversary, “player 1” is here an
abbreviation and means the entirety of people engaged in the practicalities (“do”-phase
of the ISO PDCA-cycle) of the risk management. We will hereafter call this player the
“defender”, to ease our wording. Specifically the “do” phase is the one where game
theory can help, since it requires:
• The selection of controls: this is an action undertaken in the final “risk treatment”
phase in Figure 1, and the point where game theory is applied.
• The implementation of controls: while the selection of controls can be based on the
knowledge of an equilibrium, the enforcement thereof (i.e., playing the equilibrium)
corresponds to an optimized implementation scheme. We will revisit this issue later
in sections 6.4 and 6.3.
• The definition of measures to check their effectiveness: the measures per goal are
the assurances defined in [35, Def.4.1].
So, all these actions can be supported by game theory. To make this precise, let us look
at the steps numbered in Figure 2, and to be completed for each player:
3.1: each player has some a-priori or current knowledge when a decision is made. In
classical game theory, the next action depends only on the current state of the
game (in a generalization to stochastic or sequential games, a dependency on past
game iterations is included; we discuss one such example in Section 5.1). In the
reality of risk management, external need to be considered, and the player’s action
is hardly dependent on the current state of the system only (at least because the
system state may not even be known precisely at all times). For this reason,
the theory developed in [35],unlike classical game theory, allows uncertainty to be
explicitly modeled at this stage, and incarnate through the payoffs of an action
that we describe below.
3.2: strategies for a player refer to everything that can be done in the current position.
In fact, if the game play covers several stages until the payoff is received, then the
strategy is an exact prescription of what is to be done at each step. In that sense,
it is comparable to a “recipe” that the player can follow to accomplish the desired
goal. Mapping this to a practical setting, the strategy may be named “patch
machine X”, whereas its details relate to all the steps taken from the current state
of machine X until the point where the patch has been installed and the machine
is put back to work. Likewise, a strategy for the opponent (player 2, adversary)
6
may be high-level named “hack machine X”, where the details of this strategy may
include a sequence of steps such as “sending a phishing mail” → “connect to the
malware” → . . .
3.3: Identify the interests of a player and define measures of “fulfilment” of these needs.
In our case, this process refers to the identification of goals that are object of the
risk management. Examples include (but are not limited to): economic, reputation, people, information, capability, etc. (see [45]). The degree of achievement
in each of these goals must be measured on a scale that makes the different goals
comparable, and for technical reasons also arithmetically compatible. The easiest
way of assuring this is to define a common set of discrete risk categories, with individually specific meanings per security goal. This has several advantages beyond
pure theoretical reasons, as it “equalizes” the understanding of risk valuations and
the taxonomies in which risk and outcome of actions is expressed. This is the
fundament to the next step 3.4.
3.4: Preferences of the players are found by asking how the players valuate an outcome.
For general games, this has to be done for each player. In our case (and every
zero sum game), it suffices to ask only player 1 (the defender) for this valuation,
in each scenario. By construction, we specifically allow an outcome of a specific
scenario of strategies for the defender and the attacker to be rated in various respects, i.e., in terms of each security goal. Having a common vocabulary (set of
risk categories) in which the risk in each goal is expressed, the underlying theory
of multi-criteria optimization meaningfully applies and helps optimizing the actions towards maximizing the security goal fulfilment (equivalently, minimize the
residual risk).
3.5: the representation of preferences by a utility function is –in classical game theory
– done by specifying a function ui : P S1 × P S2 → R for the i-th player, so
that each scenario of defense (action from the list P S1 of available options) and
attack (action from P S2 , i.e., possible exploits) is rated in some real-valued score.
The construction in [35]deviates from this definition in allowing the outcome to
be not crisp but a random probability distribution, thus the function ui takes
the form ui : P S1 × P S2 → F, where F is the set of all probability distributions
(more precisely their density functions) that satisfy the regularity conditions (lower
bounded by 1, absolute continuity w.r.t. Lebesgue or counting measure; see [35]).
The crux of this modification is that:
• Letting ui be valued in terms of probability distributions offers a powerful
model to capture uncertainty (cf. our remarks in step 3.1 above).
• The specific definition of ui can be made based on empirical statistics; that is,
we can simply collect many domain expert opinions on a specific scenario from
(d, a) ∈ P S1 × P S2 , and define the value u1 (d, a) as the empirical histogram
of this expert survey (we expand this approach below in section 3.3). This
has a neat effect, since it:
7
– Preserves all information provided by the experts
– Avoids a consensus problem that would normally arise from the need
to agree on a single representative opinion about the risk. In practice,
people may be unsure and disagreeing to the opinion of others, so that
conflicts and aggregation of opinions may be required. In letting the risk
valuation be an entire histogram, each input goes into the assessment
with the same weight, so that no domain expert is “overruled” or has less
influence than any other.
If the system response dynamics is known, on the other hand side, then there
may not even be a need to poll experts, and a simulation of the outcomes
under the specific scenario (d, a) could be imaginable. Percolation theory [22]
offers one way to do this (among other possibilities).
In assuming a zero-sum competition and a two-player game, the risk management
process outlined here is asset-centric. This is an important core philosophy of the
entire process, since the overall goal is not about preventing all possible threats, but
about making an attack non-economic for the attacker. This is the main point of applying
optimization (i.e., game theory) here, since we seek to minimize our own losses (measured
in terms of the values of our assets), against whatever an attacker may do.
The mapping sketched in Table 1 has been adapted from [34], but needs a bit of
tailoring towards defining an analogous process based on distribution-valued games (and
the theory thereabout). Specifically, the risk assessment phase (Figure 1), comprising
risk identification, risk estimation and risk evaluation, refers to the current state of the
system, whose valuation may trigger further action (implementation of controls) in later
phases of the process (namely the risk treatment).
In Table 1, the process of determining the current state is only briefly mentioned as
adding a strategy “do nothing” to the game, which merely models the possibility of the
situation at hand being satisfying already. Here, we will resemble the classical process of
risk management to a wide extent using only the idea of an outcome valuation in terms
of probability distributions, which is made possible by the -ordering on these objects
that was invented in earlier stages of the project. This extends up to the point where
risks are evaluated, which is normally done using a 2-dimensional risk matrix, typically
with colored entries like shown in Figure 3. The (specification of the) elliptic region
displayed on the upper right corner is the main output of the risk evaluation, which
establishes a priority list of risks and divides the threats into those that demand action
(i.e., which are put into the critical region based on their impact and likelihood) and
those who do not demand immediate actions (i.e., which lie outside the critical region).
In the original ISO/IEC process, the risk value is computed as the product risk =
likelihood × impact. It is popular to define a categorical scale for both, the likelihood
and the impact (in tabular form explaining what exactly is meant by a “medium likelihood” or a “medium impact”; section 3.8 gives examples). Multiplying the ranks (i.e.,
category numbers1 ) then gives the risk score and delivers the coloring of the risk matrix
1
it is advisable to avoid using the number zero as a category number or rank, as it would cancel out
8
Context
establishment
Risk identification
ISO/IEC Process/Terminology
Setting the basic criteria Defining the scope
and boundaries Organization of an information security risk management (ISRM)
Identification of assets
Identification of existing controls
Identification of vulnerabilities
Identification of consequences
Risk estimation
Assessment of consequences
Assessment of incident likelihoods
Level of risk estimation (list of risks with
value levels assigned)
Risk evaluation
Risk treatment
List of risks prioritized
Risk treatment options are risk reduction, retention, avoidance and transfer
Residual risks
Risk acceptance
Risk communication
Risk monitoring and
review
List of accepted risks based on the organization criteria
Continual understanding of the organization’s
ISRM process and results
Monitoring and review of risk factors Risk
management monitoring, reviewing and improving
not included
not included
not included
Game theoretic step/terminology
Scenario investigation (scope definition and
asset identification) Player identification
(mostly assigning the role of the defender).
Included in the scenario investigation
Identify implemented controls, i.e. “do nothing” option for the defender.
Options that can be exploited by threats. Included while determining the strategies of the
attacker (player 2). We denote this list as
P S2 .
Identify how players value multiple aspects of
outcomes. Identify preferences (i.e., priorities
among different security goals)
Define a common scale and ranking scheme
for all relevant outcomes.
Computed likelihoods for each strategy for
both players.
Expected outcome for each scenario. This expected value (a real number) is in HyRiM replaced by an entire distribution function (thus
avoiding information loss due to a “representative” centrality measure like the average).
Prioritize the expected outcome for both of
the players
Strategies (control measures) for the defender
can be categorized into “static” (changes applied to the system that have a permanent
effect, e.g., installation of intrusion detection
systems) and “dynamic” (actions that need to
be repeated in order to retain their effect on
security, e.g., security awareness trainings).
The entirety of controls available is denoted
as P S1 .
Expected outcome of the game. In HyRiM,
this is the equilibrium value distribution, from
which all statistically meaningful quantities
can be computed. For example, the mean of
this outcome distribution would correspond
to the classical quantitative understanding of
risk as the product of likelihood and impact.
Since HyRiM admits multiple goals to be optimized at the same time, the residual risk is
returned per security goal. The theory coined
this output artefact “assurance”, and it is one
probability distribution for the losses in each
relevant security goal.
Strategies of the defender (based on the organization criteria)
Strategies of the defender
The process is repeated as the player’s options
and their outcome valuation may change
Information gained by the opponent
Beliefs and incentives of the opponent
Optimization of the strategies
Table 1: Mapping between ISO/IEC 27005 Risk Management and Game Theoretic Approach (cf. [34])
9
low
med
likelihood
low
low
low
low
medium med hi
low
med
low
med
low
med
low
high
high
medium med hi
high
medium med hi
med hi
low
med
low
med
medium med hi
medium medium
impact
threats located in this
region demand action
Figure 3: Risk Matrix (example)
as exemplified in Figure 3. Note that loss valuations based on crowd sourcing (expert
surveys, detailed in section 3.3) in fact embody both, the impact (as is directly asked
for), and the likelihood (as the relative frequency of answers provided by the set of experts polled) in one object. That is, the classical two-dimensional matrix depicted in
Figure 3 would boil down to a 1-dimensional list of threats, which can plainly be sorted
in -ascending order. Thus, the risk evaluation is even simplified in our setting here.
It may nonetheless be advisable to resemble the classical process of risk management a
little closer by asking for both, the impact and the likelihood, and construct empirical
distributions for both. This may be useful in letting the expert express his beliefs more
detailed than just asking for a possibility (and not for a probability too). We revisit this
issue later in section 4, after having established the basics of this process first.
In generalizing this classical approach to risk management to a more sophisticated
technique based on games and distributions as risk representing objects, our main target
in the next section will be modeling the losses as probability distributions, and how
to prepare the game theoretic models for the risk treatment phase, when the games
are to be solved and “played” in practice. We will exemplify the modeling based on
two examples, which are advanced persistent threats (APTs), covered in section 5.1,
and social engineering, discussed in section 5.2. Once the cycle in Figure 1 has been
completed for once, any repetition entailing another round of risk estimation and risk
evaluation can be supported by exactly the same kind of game-theoretic models that were
used in the previous risk treatment phase (in fact, the equilibrium strategy enforced as
result of the last risk treatment phase is the “do nothing” strategy in the next risk
assessment; cf. Table 1). Examples of potentially suitable models are described in the
sections to follow. Section 6.3 discusses how to use the game analysis results for effective
improvements and control selection.
any rating in the other aspect. Indeed, we have adopted such an assumption quite explicitly already
by declaring loss random variables (or categories) to be ≥ 1 in any case; see [36].
10
3 Modelling Losses
The proposed method of risk modeling is here based on non-parametric loss models, for
their conservation of information and absence of perhaps difficult to verify assumptions.
Nonetheless, actuarial science knows important applications of parametric loss models,
which we briefly discuss in the next section.
3.1 Parametric Loss Models
The weight of a distribution’s tails is what makes it (in)appropriate for risk management,
where distributions with heavy, fat or long tails are common choices. In the continuous
case, extreme value distributions (Gumbel, Frechet or Weibull) are suitable choices,
as well as stable distributions. The latter, despite not having analytically expressible
densities except for special cases, can nevertheless be ordered upon using sequences of
truncations or moment sequences (if they exist) [37].
In the discrete case, the (a, b, 0) or (a, b, 1) class of distributions may be considered,
with a density in the former class being defined recursively as Pr(X = k) = Pr(X =
k − 1) · (a + kb ) for k = 1, 2, 3, . . .. The (a, b, 1) class is – roughly speaking – the truncated
version of an (a, b, 0)-density excluding the possibility of the event X = 0. It can be
shown that the former class includes exactly three families of distributions, which are
Poisson, Binomial and the Negative Binomial distribution.
A general issue with parametric losses is their representation of an arbitrary amount
of information by a fixed number of parameters. This inevitably incurs a loss of information, and calls for partly sophisticated methods of parameter fitting or similar. On
the contrary, nonparametric losses like (the previously proposed) kernel densities come
with the appeal of preserving all information upon which they are constructed, as well
as offering the flexibility of allowing for adjustments to model uncertainty in the expert’s
answers more explicitly. An example of this will be sketched in section 3.7.
Nonetheless, if a parametric model should be used, then one should bear in mind
that risk is intrinsically a latent variable, and the best we can do is relate it to some
observable variables. Unfortunately, the usual assumptions on the latent variable being
dependent on the hidden variable, so that an inference towards risk is possible, may not
directly apply to a risk assessment process. Risk is a property of some general entity,
action, intention, or similar. As such, subjective assessments of it are made by humans
whose risk perception may be correlated (to a degree that depends on the skill level of
the person in the respective regard) but not directly influenced by the true underlying
objective value of the risk variable. Still, algorithms like expectation-maximization [27]
could (and should) be assessed for the extent to which they can deliver a useful risk
model. Indeed, the result of an EM-algorithm being a distribution model over latent
variables together with a point estimate on its parameters is a perfectly suitable input
for the decision models put forth in the precursor parts of this report.
11
3.2 Nonparametric Loss Models
Parametric models come at the cost of information loss due to representing the entire
data by a (preferably small) set of parameters to the distribution. This is the price for
analytic elegance and the ease of many matters related to working with such models
practically.
In light of the scarceness and inconsistencies in the data to be expected in a risk
management application, we may thus look at nonparametric loss models as a potentially interesting alternative. The method of choice here will follow the proposals made
in [37].Specifically, we will start from an empirical distribution compiled from expert
interview data, and use a kernel density estimator for our loss model. Moreover, we
will focus on discrete risk assessment scales, which lead to discrete (in fact categorical)
empirical distributions. For a kernel density estimator in this context, it is somewhat
surprising that the literature is unexpectedly thin on discrete kernel density estimates.
Besides only a few proposals found in [6, 24, 33], the problem of “good” kernel density
estimation seems to be as much of an art (on top of science) as it is in the continuous
case. Hence, we will continue defining our own discrete kernel proposal based on Gaussian densities, but postpone the details until later, when we have specified the process
of data collection, which obviously goes first.
3.3 Collecting Data from Experts
As for every empirical study, the first step is fixing the details of the scenario about
which experts are to be interviewed. In our context, this may entail the identification of
a particular threat and related countermeasures. The high level description of a threat,
say “unauthorized access to server X”, must here be followed by a sequence of possible
detailed suggestions on how the respective attack could be mounted. For the aforementioned example, possibilities include the exploitation of (known) vulnerabilities (up to
zero-day exploits) in the server itself, social engineering, theft of access credentials, or
similar. Normally, each of these possibilities is matched with a respective countermeasure. Risk management standards like ISO 27000 [15, 17] or related ones [4] provide an
indispensable source of threats and countermeasures, which can be used in this step.
Abstractly, let us think of this process having brought up a list P S1 of countermeasures, opposing a list P S2 of possible threat scenarios (the acronym P S means “pure
strategy”, as a reminder that we are approaching a game-theoretic model here).
The lists P S1 and P S2 can be assumed to be quite short (though they must be
comprehensive; ideally exhaustive) in practice, and each scenario (d, a) ∈ P S1 × P S2
of “defense (d)-vs-attack (a)” can be put to its own individual review (not necessarily
independent of other scenarios, but the design of questionnaires and the amount of
context specification in there is a different and nontrivial story of empirical science not
subject of this report).
Remark 3.1 (Simultaneous Occurrences of Multiple Scenarios) Note that the modeling so far implicitly prescribes the assessment to be done relative to a specific single
threat. This restriction can be dropped in presenting the expert a set of threats and
12
explicitly allowing for several of them to occur at the same time. At first glance, this
would exponentially enlarge the list P S2 , this combinatorial explosion can be avoided
by allowing the expert to name multiple possible loss categories with different likelihoods.
That is, the expert may center the thoughts around a specific threat, but may take further
considerations of coincidental other incidents into account in saying that, for example,
• losses of category x are most likely,
• while losses of the larger category y > x remain possible if two or more incidents
occur at roughly the same time.
If several such possibilities are uttered, they are most conveniently described by a distribution supported on (at least) the anticipated loss categories x and y.
To set up the empirical game theoretic model, let us therefore assume that one scenario
(d1 , d2 ) has been specified in detail, say (for illustration)
d1 : perform periodic updates
d2 : exploit some software vulnerability
It goes without saying that this specific (example) scenario appears (if at all) in the
middle of a real-life APT, as the earlier stages are usually matters of social engineering
to make an initial contact and infection. We will go into details of this in section 5.1,
and keep our example abstract here only for the sake of illustration.
The goal of data collection is getting a payoff matrix composed from loss distributions
that can be used with the game theoretic framework defined in the preceding reports
[36, 37].Figure 4 displays the workflow to fill in one cell in the payoff matrix, which is
basically done along four steps:
1. Selection and specification of a scenario (di , dj ) ∈ P S1 × P S2 as a questionnaire
presentable to experts (the wording and style of the questions is itself a highly nontrivial matter, and should be done w.r.t. the subsequent implementation strategy
for the optimal defense. We will revisit this issue later in section 6.5),
2. Doing an expert survey on the effectiveness of countermeasure di against attack
dj . It is crucial for this survey to clearly define at least the following items:
• the context of the risk assessment (that is, the aspects that are relevant and
those that are irrelevant for the risk assessment)
• a clear definition of the scale in which the risk is quantified (in the style of
Table 2 shown in section 3.8 or similar).
3. Collecting as much expert input as possible to define an empirical loss distribution
L̂ij over the categories as specified in the survey.
4. Preparing the loss distribution for a subsequent game-theoretic analysis.
13
game matrix (to be defined)
define the context
(action spaces)
step 2
step 1
action space PS2
j
...
...
...
m
?
...
n
will probably work; damage at most medium
...
...
...
...
...
very likely to work; low damage expected
...
...
i
Question: how effective is
countermeasure i […]
against attack j […]
in the following scenario: ...
...
...
...
...
...
...
action space PS1
1 2 3
1
2
ask experts for their
assessment of each relevant
scenario
may not work; damage could be large
...
action i may be not doable/applicable at all;
high damage might nevertheless occur
action space PS2
...
survey data
collection
^L
ij
...
...
prepare the
loss models
and payoff
structure
no. of opinions
(absolute frequency)
^L
ij
...
...
...
n
...
...
...
i
m
...
...
...
step 3
j
...
...
...
...
...
action space PS1
1 2 3
1
2
step 4
...
low
medium
high
damage
Figure 4: Setting up a Matrix Game with Uncertain Payoffs from Expert Questionnaires
The details of step 1 are individually dependent on the application context, i.e., are
specific for the system or infrastructure under investigation. This step is recommended
to be done according to established standard procedures as described by ISO 27000 or
its relatives. Step 2 is a matter of empirical research and questionnaire design. The
literature on this is rich, and this task should be left with experienced staff educated in
empirical research and statistics. On the technical level, collecting the data is relatively
simple, since it is an easy task to set up an online survey, displaying a sequence of
questions asking the expert to give her/his risk rating on each described scenario. Things
may be even made more efficient by showing the expert the entire matrix, asking to enter
a risk assessment in each cell (showing yet another combination of defense and attack),
and to leave all fields blank for which there is a lack of domain knowledge or no justified
opinion can be expressed.
Step 3 deserves some attention, as this involves an a-priori agreement on the loss
categories to be used in the survey. We give details on this in section 3.8.
The final preparation of loss distributions in step 4 is a matter of kernel smoothing,
and described in section 3.10.
14
3.4 Refining the Expert Survey
Note that the expert survey – in the form described – only asks for possibilities and
not for probabilities. That is, the expert is only questioned to state an expectation of
outcome, rather than telling how likely s/he feels it to be. If the survey is refined to ask
for a likelihood in addition (as is prescribed in the conventional ISO/IEC risk management standards), then we end up with two probability distributions, one for the impact,
the other for the likelihood. A “multiplication” of these two objects, to resemble the
usual formula “risk = impact × likelihood” underlying quantitative risk management,
is theoretically possible (as a multiplication of the hyperreal representatives), but not
practically meaningful. If the two distributions are available, then a lexicographic comparison of (impact, likelihood) or (likelihood, impact) may be the more reasonable way
to go. However, an explicit advantage of the expert surveys and using distributions
as representative objects is them embodying both, impact and likelihoods within the
same loss distribution object, thus the expert interviews are greatly simplified over the
standard risk management process.
3.5 Outlier Elimination
It is important to note that any survey data should be cleaned from outliers, and there
has to be a consensus on the treatment of missing values. Either is up to a variety of
statistical methods and the particular method should be chosen in light of the given
application.
The discussion here is only meant to bring this point to the attention of the reader,
and will not go into details of how this could be done.
3.6 Harmonizing Risk Attitudes
After having cleaned outliers, it remains to “harmonize” different kinds of answers depending on the individual personalities and risk attitudes. Persons known to be risk
averse will tend to overestimate the risk, while risk seekers will tend to underestimate.
It is beyond the scope of this report to discuss concrete methods to correct risk estimates
based on a respective classification of the individual, but it is important to bear in mind
methods of statistical classification as a potential toolbox to help in this regard.
3.7 Using the Bandwidth Parameter to Model Answer Uncertainty
In some occasions, it may happen that an expert is unsure about whether or not certain
circumstances enable certain damages. For example, if the question of whether an incident at one point in the system can cause damage at another point in the system can
be answered with “generally no, except for some rare cases”, then the risk assessment
related to the incident will have its modal value at a low category, but – due to the
possibility of the incident being nonetheless severe – extends the distribution up to the
full range of loss categories. The bandwidth parameter of the kernel density estimate
can be increased to let the density put more weight on far away categories, or be chosen
15
smaller when the certainty about the assessment (rareness of the incident), is better. In
any case, the choice of bandwidth for the kernel density estimates gains another dimension of importance as being a parameter to control/describe the expert’s certainty in the
data.
3.8 Defining Risk Assessment Categories
It is crucial (not only for technical reasons) that all loss distributions for all goals in a
multi-goal security game must be defined on the same categories. This can be justified
by technical but also interpretative reasons:
Technical reason: the procedure to numerically compute multi-goal security strategies
(MGSS) relies on casting the multi-criteria objective function (vector-valued) into
a scalar that is a weighted sum. This transformation requires “compatible” objects
to be weighted and added (see [37]), which calls for the same underlying scale in
all goals.
Interpretation: comparing the severity of damages in two security goals is only meaningful if the goals are quantified in the same terms. That is, the understanding
of, say “high” damage has to be fixed for all goals, and must not be left to a
subjective idea that is individual for each expert (otherwise, the outcome of any
such assessment is useless). This is especially relevant for non-numeric goals like
the reputation of an enterprise, customer trust, or similar.
Especially to address the latter, it appears advisable to underpin the survey/questionnaire by an a priori fixed definition of categories in which the risk assessment shall be
made. This can be done in tabular form where each goal is assigned a column, with rows
corresponding to the categories, and cell entries describe the meaning of a risk category
specifically for each goal. Table 2 shows an example, whereas it must be stressed that
the concrete content of the table must be adapted/tailored to the practical situation at
hand.
Such a table should be displayed together with the survey, to equalize the expert’s
individual understandings of the risk categories, and to harmonize the resulting data.
The loss categories actually used with the model are simply the integers 1, 2, 3, . . ., noting
that the number 0 is precluded as a category (in order not to violate the assumptions
made in [36]).
3.9 Using Continuous Scales
If the loss is measured in continuous terms, then a common categorization like outlined
above must be replaced by a “meaningful” common (continuous) risk range. The exact definition of “meaningful” must herein be made dependent on the context of the
problem, so that all security goals (losses) homogeneously cover the range without being
concentrated in disjoint regions. We illustrate the issue with two examples, one showing
how it should be done, the other illustrating the problem:
16
risk
category
(numerical
representative)
negligible (1)
noticeable (2)
loss of intellectual property
< 500 e
between 500 e
and 10.000 e
loss category
damage to reputation
not noticeable
noticeable loss of customers (product substituted)
significant loss of customers
low (3)
> 10.000 e and 50.000 e
medium (4)
> 50.000 e and 200.000 e
high (5)
> 200.000 e
and 1 Mio. e
noticeable loss of market
share (loss of share value
at stock market)
potential loss of marked
(lead)
critical (6)
> 1 Mio. e
chances of bankrupt
harm to customers
...
none
inconvenience experienced
but no physical damage
caused
damages to customers’
property, but nobody injured
reported and confirmed
incidents involving light
injuries
reported incidents involving at least one severe injury but with chances of
total recovery
unrecoverable
harm
caused to at least one
customer
...
...
...
...
...
...
Table 2: Potential (Example) Definition of Risk Categories for different Security Goals
Example 3.2 (well chosen loss range) Consider two loss variables, with are monetary loss due to theft of intellectual property, and (monetary) investments to protect
these assets. Since security is primarily not about making an attack impossible but only
about making it non-economic, it can be expected that both losses range in roughly the
same numeric region. Formally, let R1 , R2 be the two random loss variables, and let
ε > 0 be a small value for which we truncate both distributions F1 , F2 at their respective
(1 − ε)-quantiles, denoted as q(1−ε) (Fi ) for i = 1, 2. Call the so-obtained loss ranges
I1 = [1, q(1−ε) (F1 )] and I2 = [1, q(1−ε) (F2 )]. If I1 ≈ I2 (i.e., q(1−ε) (F1 ) ≈ q(1−ε) (F2 )),
then we may take the convex hull of I1 ∪ I2 as the common loss range, and be sure that
the comparativeness of the two loss variables is retained and reasonable.
Example 3.3 (badly chosen loss range) Let R1 ∼ N (5, 1), R2 ∼ N (900, 100) be two
random losses with Gaussian distributions. As before, if the modeler chooses the range
as the (convex hull) of the union of both (truncated) ranges at some (1 − ε)-quantile,
the game may be defined over losses within the common range [1, 1000] (for ε = 0.027,
the range µ ± 3σ covers ≈ 99.73% of the cases for the Gaussian distribution N (µ, σ);
our cut off here is at µ + 5σ, and thus covers more than 99.9%). Since R2 , however,
assigns most of its mass in the region [800, 1000], it will always (at least up to reasonable
numerical precision) be -larger than R1 , so the optimization is pointless.
To avoid situations like sketched in Example 3.3, (at least) two options are available:
1. Rescaling of all loss ranges to a common region. Continuing example 3.3, this
1
would mean replacing R2 by 90
R2 ∼ N (10, 100/(902 )) ≈ N (10, 0.012346).
2. Defining a common continuous scale with the same intended usage as the discrete
categories. That is, if a continuous loss rating is permitted, we may define it within
a common interval from [1, 10], allowing any value to be picked by the modeler, as
17
long as it is within the range. This may be the method of choice if the modeling
interface displays a “slider” where a user can drag the gauge at any position on a
continuous scale to express the (subjective) belief about the loss.
3.10 Preparing the Loss Model
In [37], the theoretical possibility of convergence issues in the game’s analysis (by fictitious play) was anticipated, occurring when the loss distributions do not share the
same support (this adds to the technical justification stated in the previous section 3.8).
Practically, this is highly likely to occur due to missing data. If some categories are
simply not being used (either because of an unsuitable definition or because the scale
is fine-grained so that not all levels are being used by the experts), then the loss model
may be empirically correct, but not useful with a subsequent numeric analysis.
The solution in the continuous case was a kernel smoothing, technically a convolution
with a Gaussian density, to extend all loss distributions until a common end of their
support (previously called the cutoff point) is reached. We will do the same thing in
the discrete case, using a kernel √obtained from discretizing the Gaussian distribution.
To this end, define f (x) := (1/ 2π) exp(− 12 x2 ) as the density of a standard normal
distribution, and for every h > 0, define the kernel function Kh : Z → R as
Kh (n) :=
1
h
Z
n+ 12
f (t/h)dt,
n− 12
where h is the bandwidth
parameter (asR familiar from continuous kernel density estiP
1 ∞
mates). Obviously, ∞
n=−∞ Kh (n) = h −∞ f (t/h)dt = 1 and K ≥ 0, so K defines a
discrete probability mass function on Z. Moreover, it is easy to see that for every h > 0,
the function Kh resembles a discrete version of a Gaussian density, so that by letting
h → 0, Kh degenerates into a discrete Dirac mass,
1, if n = 0;
Kh (n) →
(1)
0, otherwise.
If we call F̂n a general empirical distribution function (defined on any subset of Z)
obtained from n data points (answers in the expert survey). The convolution F̂n ∗ Kh
is another distribution function supported on all Z, i.e., (F̂ ∗ Kh )(n) > 0 for all n ∈ Z.
Moreover, in letting h → 0, we have the pointwise convergence (F̂ ∗ Kh )(n) → F̂ (n) for
all n ∈ Z, so the estimate is asymptotically correct (note that in contrast to Nadaraja’s
theorem, we do not even need to assume a specific speed of decay when letting h → 0).
This convergence easily follows from (1). The same result equivalently holds if we replace
f by a truncated version fˆ thereof (being supported on [1, a] for some integer a > 1),
leading by disretization to the truncated kernel K̂h . In the following, we thus consider
the smoothed empirical distribution F̂n,h := F̂n ∗ K̂h , noting that this also pointwise
converges to F̂ and that the support (strictly) covers the full interval [1, a]. However
now, we can make a stronger convergence statement: let F be the (unknown) distribution
18
of the loss assessment, which is approximated by the expert data2 , which is F . We can
write
F̂n,h − F̂n + F̂n − F ≤ F̂n,h − F̂n + F̂n − F
(2)
To the first term, we can apply a uniform bound thanks to the supports of F̂n,h and F̂n
being all finite, which is
sup F̂n,h − F̂n → 0
as
h → 0,
as is an easy consequence of the equivalence of all norms on R. The other term in (2),
we can apply Glivenko-Cantelli’s theorem to conclude the convergence
sup F̂n − F → 0,
as
n → ∞,
so that sup F̂n,h − F → 0 in the limit h → 0 and n → ∞. Note that this argument is
only good for plausibility of our approach, but cannot be taken as a proof of correctness,
since practically, it still relies on an infinitude of data, which – more crucially – must be
objectively sampled from the real loss variable (having the distribution F ).
On the bright side, however, the proposed smoothing enjoys a nice intuitive justification as accounting for uncertainty in the assessment.
Example 3.4 Suppose an expert utters the opinion that, on a scale from 1 to 5, the risk
is 3. Despite that a middle assessment is an implicit statement of uncertainty already3 ,
there statement “the risk score is 3” is not equivalent to the statement that the risk
cannot be anything else. In other words, there may be admitted chances of the damage
being higher or lower than the average told by the expert. To express such uncertainty
more detailed, the expert may admit possible outcomes on the entire range 1, . . . , 5, with
the initial assessment being just the most confident outcome.
Of course, it will not be feasible to ask experts to provide entire probability distributions over a scoring scale, but the smoothing of the empirical distribution by convolution
with Kh achieves the same thing. Viewing Kh as a set of weights associated with all
possibilities around the mean value x, the expert assessment x ∈ Z receives the maximal weight (as being the modal value of Kh ), whereas all other possibilities receive a
nonzero weight that decays with the distance to the value x. Graphically, this process
thus “fills” all empty bins in the empirical histogram (say, if a category has never been
2
We somewhat sloppily assume here that the experts provide us with “observations” about the real
random loss X having the distribution function F . This assumption is clearly not correct, but still
the best that can be done, given that we cannot simply wait for losses to occur, as this would probably
kill the enterprise much before a decent lot of data about the true loss distribution could have been
obtained. Thus, we have to live with overly trusted experts providing us with “objective” samples of
the unobservable loss variable distributed like F .
3
This is usually avoided by choosing a risk scale with an even number of categories to avoid having the
median on the scale.
19
0.30
0.15
0.35
0.4
2
3
4
5
6
7
8
9
10
(a) Original empirical histogram
0.00
0.00
0.05
0.10
0.05
0.15
0.20
0.10
0.25
0.3
0.2
0.1
0.0
1
1
2
3
4
5
6
7
8
9
10
(b) Smoothing with h = 1
1
2
3
4
5
6
7
8
9
10
(c) Smoothing with h =
0.25
Figure 5: Effect of smoothing the histogram to fill empty categories and extend the
support
assigned during the expert survey), and additionally extends the distribution up to the
full and identical support for all loss distributions. Hence, besides satisfying the technical constraints imposed, the smoothing naturally accounts for the uncertainty in the
assessment; the lower the parameter h is, the more confident we are in the assessment
(i.e., the closer the smoothed approximation F̂n,h approaches F̂n ). Figure 5 displays an
example of the original empirical histogram (Figure 5c), and its smoothed version with
h = 1, having its gaps filled in by the superposition of smoothing kernels (expressing the
uncertainty around the modal values given in Figure 5c), and another smoothed version
of the histogram with a much smaller bandwidth parameter h = 0.25. This plot visually
illustrates (confirms) the formal claim of convergence as stated previously.
The quality of the kernel density estimation, as in general with this nonparameteric
method, is the art of choosing the bandwidth parameter h. Various rules of thumb (e.g.,
Silverman’s rule [43]) or cross-validation techniques may be applied. Since there is no
generally “best” way to choose this parameter depending on the information at hand,
we refer to the literature on nonparametric statistics for concrete methods [8, 47, 49].
4 Risk Prioritization
Let us assume that a total of m threats T1 , . . . , Tm has been identified. If, say n, experts
provide their input on the impact and likelihoods (both artefacts obtained by surveys
as outlined in section 3.3), then it is a simple matter to compile an empirical histogram
(a distribution FI ) for the impact and another empirical histogram (a distribution FL )
for the likelihood. In doing so for every threat T1 , . . . , Tm , we end up with 2m such
distributions, which we denote here as F( k, I), F( k, L) for the k-th threat Tk in the list.
To resemble the risk matrix familiar from the standard process of risk management,
we can use the -ordering on the distributions (remember that it is a total order), to
20
ranking (w.r.t. ˜-ascending order)
...
1 2
Impact:
Likelihood:
ri
...
T2
T2
...
rj
T1
m
T3
T1
T3
likelihood ˜-ranking
m
ri
T1
T3
T2
1
1
impact ˜-ranking
rj
m
threats located in this
region demand action
Figure 6: Risk Matrix using -Orderings
separately sort the threats T1 , . . . , Tm in ascending -order according to their impact
and likelihood. This gives two lists, in which each threat gets a rank assigned among the
total of m threats, say the k-th threat has rank ri on the impact ranking and position rj
on the likelihood ranking. The risk matrix familiar from the ISO processes (cf. Figure
3) is then re-established by placing each threat into a 2D-coordinate system with the
x-coordinate being the rank on the impact scale and the y-coordinate being the rank
on the likelihood scale. This puts each threat to a particular position on the grid, and
we may then proceed “as usual” by coloring the area as we would do for a normal risk
matrix. Likewise, the “critical region” can be defined as the area inside which threats
fall that need to be addressed by the subsequent risk management process. Figure 6
displays an example.
The main difference to the standard risk matrix is herein twofold:
1. While a standard risk matrix has axes scaled in categories, the so adapted risk
matrix has axes scaled in ranks that go from 1 to the total number m of threats.
2. The consensus problem of assigning a threat a single quantification in terms of likelihood and impact is avoided. Since many experts can utter disagreeing opinions,
all of which go into the impact- and likelihood-distributions FI and FL with equal
importance. The total ordering then assures somewhat like a “base-democratic”
ranking, since one threat outranks the other in the -ordering, if more people classify the impact, respectively the likelihood, as high (see [37] for the full detailed
effect of this ordering).
If the assessment considers multiple criteria, say, if the impact is not only measured
in money but also in reputation, then the respective assessments are made separately
21
(not necessarily independently) from one another. For the example of budget-impact
and reputation impact, we could think of the one distribution FI as being replaced by
two new distributions FBI and FRI (for budget and reputation). Likewise, if two threats
have assessments in these terms, given as F(BI,1) , F(BI,2) and F(RI,1) , F(RI,2) , then we
ought to rank the two in terms of two criteria. A canonical way of doing that is offered
by defining a game with two goals (“budget” and “impact”), and using one (dummy)
strategy for player 1, and letting the two threats being the strategies for the opponent
(player 2). It is, however, important to stress that we cannot directly go ahead and set
up a multi-criteria (1 × n)-game to get the most severe threat via an equilibrium, since
the gameplay is defined to be among n + 1 players, where each opponent (corresponding
to a goal) plays independently of all others. The equilibrium would then return a worstcase threat identified per goal, which is possibly not what we seek here for a multi-criteria
risk optimization. The theory, however, remains applicable when the game is reduced
to a two-player (1 × n) game with scalar(ized) payoffs per player. In that case, the
equilibrium is necessarily pure and indicates the worst case threat for player 1. The R
package implements this method in a designated procedure preference:
# ranking if there is only one goal (say, FBI1, vs. FBI2)
> preference(FBI1, FBI2)
2 # this means that the second parameter (FBI1) is preferable
# ranking, if there are multiple goals (of equal importance)
> preference(list(FBI1, FRI1), list(FBI2, FRI2))
#
#
#
>
Let us assign twice as high priority to the "reputation"
goal by supplying the parameter weight=c(1,2), i.e. "goal" has
priority 1, and "reputation" has the (double) priority 2.
preference(list(FBI1, FRI1), list(FBI2, FRI2), weights=c(1,2))
The results directly gives the desired ranking by telling that either the first (output
“1”) or second (output “2”) threat is less severe; if the function returns zero, then the
distributions are identical and the decision is indifferent. For three or more threats, the
procedure can be repeated pairwise to rank a whole set of threats in the way as described
above.
5 Examples of Game-Theoretic Modeling
5.1 Modeling APTs
The investigation of many examples of advanced persistent threats (APTs) reported in
the past reveal a common structure underlying an APT, even though the details thereof
may be highly different. Quoting the taxonomy of [7], an APT roughly proceeds along
the following steps:
1. Initial infection:
22
a) reconnaissance: information gathering,
b) development: design of a made to measure malware,
c) weaponization: preparing the trojan and droppers,
d) delivery: transmission into the victim infrastructure, e.g., by a phishing email,
or similar.
2. Learning and propagation: repeated sequence of
a) exploitation: to get deeper into the system
b) installation: to leave artifacts and backdoors for an easy return later, and
destroy footprints of the attack
3. Damage:
a) command and control: interaction with the victim system’s compromised
resources via previously left artifacts
b) actions on the target: causing the actual damage
The specific actions taken in each of these phases depend on the target infrastructure and no general description is possible due to the diversity of such infrastructures.
However, specific examples can be “read off” reported prominent incidents, such as including Stuxnet [10], Duqu [2], Flame [28], and Aurora [23]. A common element in all
these appears to be the human factor, which we discuss separately in section 5.2.
Game theory has been applied in various ways to model attacks and cyber risks [9,25,
50] and proposals (independent of game theory) include moving target defenses [18, 19],
trust mechanisms [48], and defense-in-depth techniques [30]. Here, we will review a
recent proposal to model APTs as a sequence of games, each of which is tailored to the
specific nature and details of the respective phase.
That is, an APT usually begins with a harmlessly looking email or “lost” USB stick,
on which malware enters the system. Alas, nowadays attackers have joint forces into
an entire illegal business sector covering the entire supply chain of cyber crime in a
spectrum of independently offered services (often referred to as “cybercrime as a service”
in alignment to cloud computing terminology). This means that the person identifying
a weakness is not necessarily the same who is writing the exploit for it. Likewise, the
author of the exploit kit is not necessarily applying it anywhere, but merely sells it to
the actual attacker. Similarly, infections with malware are not automatically intended to
cause immediate damage. The term “botnet-as-a-service” describes the business model
of infecting a large number of machines and offering to deploy malware on these zombies
upon request and for a (smaller or larger) fee. In this way, the attacker can simply “buy”
access to a potentially large number of infected machines instantly. For the victim, this
has the unpleasant effect of the infection remaining stealthy and inactive, until some time
later, when the outbreaks causes noticeable damage. At this time, however, it is most
likely that no connection between a past email and the current incident is recognized.
The exact time window between an infection and its activity is also dependent on the
23
technical countermeasures adopted in the specific company. For example, signaturebased malware recognition (like classical anti-virus software used to work) may take a
couple of days until the malware is recognized as such and the respective signature is
shipped with the next update. On the contrary, cloud based malware detection that
is based on recognizing a huge lot of “identical” email (attachments) suddenly flowing
through the internet can have a much shorter time (even a few hours) until a malware
is suspected and classified. So, the time window between the infection and the exploit
may, in some cases, be closed quite fast (by a good malware recognition system).
To model a so-structured APT in terms of games, the first step is listing all potential
ways into the system, such as include (but are not limited to):
• (spear) phishing, whaling,
• waterholing,
• tailgating,
• etc.
The entirety of these possibilities makes up the set of strategies P S2 for the attacker.
Typically, an entry in P S2 is thus not constrained to have a specific form or structure
(in particular, an element x ∈ P S2 is not a purely mathematical object but rather a
description of how an attack would be launched according to this particular strategy x).
In a different view, P S2 can be considered as a list of threats, to which corresponding
countermeasures can be defined (appealing to catalogues found in standards like ISO
27000, etc.). The list of countermeasures is P S1 .
Note that P S1 and P S2 should be defined bearing in mind the steps and structure
of the first phase of the APT, which comprises information gathering, development,
weaponization and delivery. Consequently, example defense measures in P S1 may relate
to periodic (and random) re-configurations of the system in order to thwart the attacker’s
respective next steps, either by invalidating so-far collected information or by removing
malware by coincidence, say if a computer is reinstalled or privileges of the victim user
have changed (were revoked).
In any case, let the game describing the initial infection be a matrix game G1 (where
the subscript is a reminder of this being the first phase of the APT). It appears fair to
let this be a matrix game (static and repeated), since if the game is expectedly repeated
either in case of an attack failed, or to infect further parts of the system, if an attack
was successful.
Let the equilibrium computed for the infection game G1 be the pair of distributions
(F1∗ , F2∗ ), where the optimal defense against the initial infection is the distribution F1∗
(the distribution F2∗ , more specifically, its probability mass function can be taken as a
non-unique indicator set of neuralgic points in the infrastructure).
To model the next phase of learning and propagation, we divide the (physical) infrastructure in stages that the adversary needs to reach one by one in order to get to
the inner target asset. If the infrastructure is a network graph G(V, E), then the target
asset may be a(ny) fixed node v0 ∈ V , and the i-th stage can (but does not need to) be
24
ftp(0,2)
ftp(0,1)
execute(0)
ftp_rhosts(0,2)
sshd_bof(0,1)
ftp_rhosts(0,1)
rsh(0,2)
trust(2,0)
rsh(0,1)
trust(2,0)
rsh(0,2)
ssh(0,1)
rsh(0,1)
ftp(1,2)
ftp, rsh, ssh
execute(1)
ftp_rhosts(1,2)
rsh(1,2)
trust(2,0)
File Server
(Machine 1)
rsh(1,2)
execute(2)
Workstation
Machine 0
Firewall
local_bof(2)
Router
full_access(2)
Initial attacker‘s capability
ftp, rsh
precondition
Database Server
(Machine 2)
(a) Example Infrastructure [44]
attacker‘s exploit
final result
(b) Example Attack Graph [44]
Figure 7: Topological Vulnerability Analysis Example
defined as the set of nodes at distance i to the node v0 . That is, the stages are “concentric
circles” around the asset, and on each stage, the adversary plays a (different) game to
get to the next stage. A single such stage is penetrated by the aforementioned steps
of exploitation and installation, continuing with exploitation again on the next stage
(only in a different setting there). This modeling is practically supported by topological
vulnerability analysis (TVA), which can cook up an attack graph for an infrastructure.
In the following, we will illustrate our thoughts based on Example 5.1.
Example 5.1 (based on [44]; see also [40]) Consider a system as shown in Figure
7b, composed from three devices, with various ports opened and services enabled. Based
on this information, the attacker can consider several exploits, such as:
• FTP- or RSH-connections from a node x to a remote host y, hereafter denoted as
ftp rhosts(x,y), and rsh(x,y), respectively.
• a secure shell buffer overflow at node y, remotely initiated from node x, hereafter
denoted as sshd bof(x,y).
25
Table 3: APT scenarios (adversary’s action set P S2 , based on Figure 7)
1
2
3
4
5
6
7
8
execute(0) → ftp rhosts(0,1) → rsh(0,1) → ftp rhosts(1,2) → sshd bof(0,1)
rsh(1,2) → local bof(2) → full access(2)
execute(0) → ftp rhosts(0,1) → rsh(0,1) → rsh(1,2) → local bof(2)
full access(2)
execute(0) → ftp rhosts(0,2) → rsh(0,2) → local bof(2) → full access(2)
execute(0) → rsh(0,1) → ftp rhosts(1,2) → sshd bof(0,1) → rsh(1,2)
local bof(2) → full access(2)
execute(0) → rsh(0,1) → rsh(1,2) → local bof(2) → full access(2)
execute(0) → rsh(0,2) → local bof(2) → full access(2)
execute(0) → sshd bof(0,1) → ftp rhosts(1,2) → rsh(0,1) → rsh(1,2)
local bof(2) → full access(2)
execute(0) → sshd bof(0,1) → rsh(1,2) → local bof(2) → full access(2)
→
→
→
→
Table 4: Security controls (selection) – subset of P S1
countermeasure
deactivation of services (FTP, RSH,
SSH)
software patches
reinstalling entire machines
organizational precautions
comment
these may not be permanently disabled, but could be temporarily turned off or
be requested on demand (provided that either is feasible in the organizational
structure and its workflows)
this may catch known vulnerabilities (but not necessarily all of them), but can
be done only if a patch is currently available
this surely removes all unknown malware but comes at the cost of a temporary
outage of a machine (thus, causing potential trouble with the overall system
services)
for example, repeated security training for the employees. These may also have
only a temporary effect, since the security awareness is raised during the training, but the effect decays over time, which makes a repetition of the training
necessary to have a permanent effect.
• local buffer overflows in node x, hereafter denoted as local bof(x).
Each of these may establish a trust relation between two nodes x and y, which we denote
as trust(x,y). The list of attack strategies to penetrate all the stages until full access
to machine 2 is a matter of plain path enumeration in the attack graph, whose results
are shown in table 4.
A selection of respective countermeasures is given in table 4.
The game to model the penetration can then be defined per stage by defining the
machines 0, 1 and 2 as stages, where machine 2 is the inner assert (node v0 in our
previous wording), and a stage is the game played to establish a trust relation between a
machine at distance i and one at distance i − 1 to machine 2.
In the (quite simple) infrastructure of Figure 7a, the stages would thus be:
• stage 1 (distance 1 to machine 2): {router}
• stage 2 (distance 2 to machine 2): {file server (machine 1), firewall}
• stage 3 (distance 3 to machine 2): {workstation (machine 0)}
The game played at stage 3, accordingly, has strategies equal to all exploits that can be
mounted on the workstation (machine 0), which can be read off the TVA attack tree
(Figure 7b) as P S2 = {ftp rhosts(0,1), ftp rhosts(0,1), sshd bot(0,1)}. The
26
corresponding set P S1 comprises all countermeasures that can be implemented (such as
virus checks, temporary disabling of services, but also non-technical ones like repeated
security training, etc.). The strategy spaces P S1 and P S2 identified in this way then
define the shape of the respective 3rd stage game G3 , whose payoff structure is to be
defined following the procedure outlined in section 3.3; Figure 4.
The games for the other stages are constructed analogously.
Given a game for each stage, we can connect them into an overall model for phase
two of the APT by adopting a high level perspective. In each stage, the adversary has
basically two options, which are:
1. penetrate: this means launching an attack as identified based on previously gathered information (cf. example 5.1), or
2. stay, to collect more information while remaining stealthy → learning.
Both options can be modeled using their own (distinct) game models, and the connection
between the games over all stages is established by considering that in the n-th stage
game (whether we go for penetrating or staying), the outcome is one of the following:
• if the attacker decides to penetrate, then
– it may succeed, in which case it enters game Gn−1 , or
– it may fail, in which case it has to repeat game Gn once more.
• if the attacker decides to stay, then
– it may succeed and gain further information, but this leaves the attacker in
this stage n,
– it may fail, in which case the overall game terminates, and the entire investment of the attacker is lost (we model this as a negative gain for the attacker).
Likewise, the defender has the option of defending (without any guarantee of the
defense being successful), or not defending. The latter strategy is implicitly played
whenever the defender “pauses” in its defense, say, if the security guard is being sent
elsewhere and misses the attack by this unfortunate coincidence. In any case, “do not
defend” is never a dominating strategy and occurs only due to resource limitations and
the inability to defend everywhere at all times against everything.
Writing I(n) for the payoff (to the adversary) in the n-th stage, the above modeling
yields a 2 × 2 zero-sum matrix game structured as shown in Figure 8 (cf. [41]). In this
model, we have additional quantities p(n) and q(n) that depend on the current stage,
and quantify the likelihood for each action to be successful. These values can be either
defined directly, or themselves be derived as saddle point values of classical 0-1-valued
matrix games in each stage (in this case, the risk assessment is done in binary terms
only asking the expert for whether or not the attack will be successful).
It is reasonable to assume a circular structure in this game (cf. also [1]), which amounts
to a single unique equilibrium, for the following reasons:
27
```
```
attacker
```
penetrate
```
defender
defend
p(n) · I(n − 1) + (1 − p(n)) · I(n)
do not defend
I(n − 1)
stay
q(n) · I(n) + (1 − q(n)) · (−I(n))
I(n)
Figure 8: Sequential 2-Player Game Model for Advanced Persistent Threats
• if “defend” is a dominating (row) strategy, then either “stay” or “penetrate” will
complete this into an equilibrium, and there is nothing to be optimized by game
theory here on this (higher) level.
• if “do not defend” is a dominating (row) strategy, then there is no need to do any
active security here, as the infrastructure is already secure anyway.
• if “penetrate” is a dominating (column) strategy, then the obvious best choice is
to defend, which again degenerates the game into a trivial matter.
• if “stay” would be a dominating (column) strategy, then there is no need to defend
anything, since the attacker is not trying to get to the asset anyway.
In any case, it is a simple (yet laborious) matter of working out the game matrix for
each stage, and determine an equilibrium for it (say, using fictitious play as described
in [37]). The particular equilibrium obtained for the n-th stage is the distribution I(n),
telling us the distribution of damage in that stage, accompanied by optimal defenses in
this stage (obtained from the 2 × 2 game and the inner sub-games played for penetration
and to collect information during a stay).
Concluding the idea, the decomposition of the learning-and-penetration phase in the
APT life cycle into stages and corresponding games played therein delivers an in-depth
defense action plan (individual randomized defense actions being taken on each part in
the infrastructure), as well as a higher-level risk assessment I(1), I(2), . . . that refers to
each stage. The distribution I(n) for the n-th such stage then indicates the likelihood
of damage suffered at the n-th “protective layer” around the asset of interest.
The final game modeling the Damage-phase of the APT can then be defined similar to
the initial infection game G1 , as the attacker may simply try and retry causing damage.
The process and steps to identify possible actions and countermeasures must again be
supported by special purpose (and application specific) tools and expertise, but the game
theoretic treatment and computation of risk metrics remains the same. Thus, we will
not repeat the details here.
A Static Game Model for the Penetration Phase In a more simplified view towards
a substitute of the sequential phase two game, a static game model can be considered
as an alternative (being easier to model and more efficient to analyze computationally).
This simplification is bought at the cost of getting a more coarse-grained model, since
the attack and defense strategies are defined more “high-level” and not specific for each
stage in the graph representation (of the infrastructure or the attack tree).
28
As before, we can think of the attacker working its way through the stages, while
occasionally being sent back or kicked out if a security officer (perhaps unknowingly)
closes some of the backdoors established previously. The pure strategy set P S1 for the
defender is thus the set of all nodes/components in the system on which spot checks (e.g.,
malware scans, configuration changes, updates, patches, etc.) can be done. Different
to the per-stage game modeling from before, the defenses now correspond to high-level
attack strategies P S2 that describe only techniques but are not specifically tailored to a
particular machine. That is, P S2 would be composed from a more generic set of threats
like
• buffer overflow exploits,
• cross-site scripting,
• code injections,
• etc.,
but unlike in Example 5.1, these attacks would not be considered on specific machines.
Rather, the attack scenario is described as a general code injection that may be tried on
any machine in the network (if possible). The description of the attack/defense scenario,
as well as the possibilities to consider for a risk assessment may be more complex in this
kind of modeling, as Example 5.2 shall illustrate.
Example 5.2 The expert survey would ask something like:
What is the expected damage if an attacker attempts a code injection attack
in our infrastructure, while machine 1 (see Figure 7a) is being re-installed at
roughly the same time.
To answer this question towards a loss estimate, the expert may consider the following
aspects:
• Which machines could be vulnerable to such an attack (say, which of them are
accessible by a web interface or maybe have an insufficient patch level, etc.)
• Depending on which machine is attacked, the damage may be more or less (thus
allowing the expert to utter multiple possibilities; substantiating the construction
of a categorial loss distribution once again).
• Depending on where the spot check is done (say, on machine 1 in Example 5.1),
three cases are possible:
– Machine 1 has so far not been reached by the attacker (more precisely the
attack path), so the reinstallation has no effect on the APT at this stage.
– Machine 1 is exactly the one currently targeted by the attacker. This may
render the results of the APT learning phase on this machine useless, since
the configuration has changed. Consequently, the attacker is sent back one
stage and has to restart learning from here.
29
– Machine 1 has already been infected with malware, so that the adversary’s
backdoor path goes through it. In that case, the attacker’s secret backdoor
may be closed by the reinstallment and the attacker is again sent back to a
previous stage.
Since there is uncertainty on where the attack may be mounted, and also on the current
position of the adversary, the expert may utter several possibilities in one of two ways:
• Tell about a set of possibilities, expressing the likelihood for each of them individually. For example, the expert may say something about the most likely outcome,
but in addition also say that more or less extreme results are possible with certain
other likelihoods.
• Give a most likely outcome r, and express the uncertainty about it: this in fact
corresponds to the aforementioned kernel density smoothing, since the uncertain
variation around the most likely value (as expressed by the expert) may be well
expressed by a bell-shaped curve centered around the value r. This is nothing else
than a kernel density, where the uncertainty is “quantified” by the parameter h.
In repeating the procedure exemplified in Example 5.2 for each (high-level) defense/attack scenario (di , aj ) ∈ P S1 × P S2 , we end up with a standard matrix game defined with
probability-distribution valued payoffs, which are now defined not on the risk categories,
but on the graph theoretic distances, or more specifically, the stages between the initial
infection and the inner asset. Picking up Example 5.1, there are three stages which
the attacker has to proceed through, and the loss distribution defined per defense/attack assigns likelihoods to each of these stages; Figure 9 displays an example. The
game-theoretic optimization then goes for pushing the probability mass towards “more
remote” nodes in the network (or attack graph). That is, we will try maximizing the
distance between the attacker and the asset, measured as the number of hops (= stages
in our wording) that need to be taken to reach the goal. The stage numbering should,
for consistency with the minimization, be increasing from 1 up to the stage where the
asset is, i.e., stage 1 should comprise the outermost perimeter, followed by the adjacent
inner nodes, until stage N > 1 being the set of final nodes adjacent to the target asset.
5.2 Modeling Social Engineering
Social engineering is a good example of a case where the distribution-valued gametheoretic framework perfectly fits. Since there is hardly any technical countermeasure
against such attacks, and the exploits are based on psychological principles of human
behavior, there is also an element of “forgetting” that makes security awareness decay
over time. For this reason, security training, information campaigns and similar must
be repeated from time to time, and there is never a guarantee that any of these precautions has any or even durable effect. Both of these reasons render matrix games with
distribution-valued payoffs into a nice model, since:
• the necessity of repeating social engineering awareness training corresponds with
the modeling assumption on the matrix game to be repeated.
30
0.2
0.0
0.1
probability
0.3
0.4
loss(reinstall machine 1 / injection attacks)
1
2
3
stage
Figure 9: Example of loss distribution on stages (phase 2 of an APT)
• the finiteness of matrix games corresponds to the limited set of possible countermeasures known against such attacks. That is, we cannot ask an employee just to
become “creative” in how social engineering may be detected, and the best we can
hope for is a feasibly small set of recommendations that a person can remember
to avoid falling victim to social engineering.
• the inevitable element of human error renders the outcome of a social engineering
scenario in any case up to randomness. Hence, our specification as a distributionvalued matrix game appears as a good fit, as it allows us to classify social engineering countermeasures as being “variably effective” (taking into account the
differences in people’s personality, daily mood, current workload, technical skills,
time since the last security awareness training, and many more).
For a pure model of social engineering attack/defense scenarios, the questionnaire
outlined in section 3.3 may be adapted to poll people about their awareness against
social engineering, describing the particular attack scenario in the description of the
question, and asking the employee on how he/she would behave in this scenario. The
results then equally well compile into the sought loss distributions, if each answer in the
multiple-choice survey has a background association with a predefined loss category (cf.
section 3.8).
A different approach to account for social engineering when it comes to loss distributions is considering these attacks as methods to establish the initial infection in an APT
model as outlined in section 5.1. Here, we would include social engineering techniques
in the attacker’s strategy space P S2 , and the specification of the respective success rates
can be done using the same kind of survey as before. The resulting game then already
models the first phase of an APT infection, in giving the probability for a social engineering attack to succeed under the given security awareness campaign (which is a mixed
strategy over P S1 in the sense of repeated randomly selected trainings, information
broadcast, etc.).
The second phase of the APT, the penetration, can as well use social engineering
31
techniques to overcome barriers within the system. Suppose that a subnetwork, for
security reasons, does not have any physical or logical connection to the outside or
any other intranet within the company. Then a malware can jump over this physical
separation by a bring-your-own-device incident. That is, if the malware gets into the
system by someone connecting a virulent USB stick to the inner separated network, the
infection has effectively overcome the logical separation. Even more, this scenario may
start from within the company’s perimeter, since the infection of the USB stick may
indeed happen on an employees’ computer, which has a connection to the internet and
got infected in the first phase.
The respective loss distributions associated with such an infection outbreak can effectively be constructed by simulation, as is eloquently outlined in [22]. We leave the
details aside here.
6 Working with Game-Theoretic Risk Measures
The description in the following is based on the previous explanations about the model,
and therefore only discusses practical matters of choosing the parameters and hints on
how to interpret the results. A description on how to do the calculations with aid of the
R statistical software suite is given in section 6.4.
6.1 Choosing the Cutoff Point
So far, we mentioned the necessity of truncating distributions only as a technical matter
of assuring convergence. As such, the point a > 1 at which the payoff distributions are
truncated (for a compact support) also influences the outcome of the game, since the
equilibria depend on it. Indeed, the practical choice of a can be made in light of how a
determines the -relation among the payoffs. Informally, Lemma 4.4 in [37] tells that
is decided based on how the payoff densities behave in a right neighborhood of a. That is,
the setting of the value a controls the range in which damages are considered as relevant
for , whereas damages far lower than a become less and less relevant for the -relation.
This means that the choice of a can be made to implement a risk prioritization in the
model in a sense that is perhaps best illustrated by an example (cf. also [38] for a related
yet different illustration):
Example 6.1 Suppose an enterprise has backup capacities to bear losses less than 50.000 e.
Then, we may set the truncation point a := 50.000, so that the entire probability mass
assigned to damages > a is “squeezed” underneath the distribution on the interval [1, a].
Consequently, if a loss model F1 admits highly likely losses > 50.000 e (i.e., has in that
sense a fat tail) will result in a higher value of f1 (a) than maybe the alternative loss distribution F2 , assigning smaller likelihood to such incidents (and in turn coming out with
a smaller density f2 (a)). Thus, f1 (a) > f2 (a) will make F2 -preferable over F1 [37].
The dependence of the resulting equilibrium is best visualized by taking another example of a 2 × 2-game with continuous payoff distributions as shown in Figure 10. Taking
32
the cutoff point a = 10 results in a pure equilibrium (x∗ , y ∗ ) = ((1, 0), (1, 0)), while
truncating the distributions at a = 6 changes the equilibrium into a mixed one, namely,
(x∗ , y ∗ ) = ((0.751, 0.249), (0.531, 0.469)). Both of these have been obtained using fictitious play as described in [37] taking 1000 iterations.
6.2 Interpretation and Meaning of Equilibria
The difference in the equilibria computed for the previous examples is immediately
understood by looking at the way how depends on the truncated distribution’s tails
(cf. [36, 37]): if the cutoff point a is chosen so large that the likelihoods assigned to
events around a are approximately equal, then there is no point in frequently switching
strategies for either player, since the outcome under any of the four scenarios admitted
by the example 2 × 2-game is roughly the same. Hence, the equilibrium is “purified”
by the choice of the cutoff point. Conversely, if the cutoff point a is chosen so that the
masses that accumulate around it are significantly different, then mixed equilibria may
arise.
This has a twofold and partly positive consequence for the risk manager, since:
• the practical choice of the cutoff point a can be made to mark the region in which
risks are considered as relevant. That is, if there is an agreement that damages
beyond some threshold value are highly relevant, while everything below it is only
of secondary interest, then the cutoff point a can be set to exactly this threshold
value. More importantly, any such setting is not even to be taken as sharp, since
what determines the -relation is the left neighborhood of a, so the transition from
the domain of acceptable risks to the area of critical risks is somewhat smooth.
For example, if the risk management is done in terms of a CVSS scoring, then the
risk manager may fix the convention of not caring about vulnerability scores in the
range around 2, but certainly action is demanded for scores above 7. In that case,
we would set the cutoff point to a = 7, in which case the game optimization focuses
on scorings of 7 or (slightly) less, while the region around scores of 2 becomes
relevant only upon equiprobable ratings of all defense actions on the range (2, 7].
This effect holds in general and is not limited to the example games here. In fact,
the cutoff point creates a qualitative difference to conventional optimization, where
the numeric magnitude of the goal function is irrelevant for the maximization. This
may not practically be the case, but can be accounted for by setting the cutoff point
accordingly.
• An equilibrium must in any case be interpreted relative to the cutoff point. If a can
reasonably/meaningfully be chosen so that pure (or approximately pure) equilibria
arise, then these greatly ease matters of playing the equilibrium. That is, if the
risk prioritization results in almost all actions ending up with approximately equal
likelihoods in the relevant neighborhood of a, then practically playing the optimal
behavior is easy, since the best response is no longer randomized.
33
goal
a2
0.3
0.2
0.15
0.0
0.05
0.1
0.10
d1
0.20
0.4
0.25
a1
2
4
6
8
10
12
14
2
4
8
10
12
0.30
0.20
0.25
0.16
0.15
0.14
0.05
0.10
0.10
0.12
d2
6
loss(1,2)
0.18
loss(1,1)
2
4
6
8
10
12
14
2
4
6
loss(2,1)
8
10
12
14
loss(2,2)
(a) Cutoff at a = 10
goal
a2
0.5
0.4
0.30
0.15
0.2
0.20
0.3
0.25
d1
0.35
0.40
0.6
a1
2
4
6
8
10
12
14
2
4
6
8
10
12
loss(1,2)
0.30
0.25
0.22
0.15
0.16
0.20
0.18
0.20
d2
0.24
0.35
0.26
0.40
0.28
loss(1,1)
2
4
6
8
10
12
14
2
4
loss(2,1)
6
8
10
12
14
loss(2,2)
(b) Cutoff at a = 6
Figure 10: Payoff structures with different cutoff points
34
Risk vs. Goal Prioritization: The role of the cutoff point in risk prioritization must be
distinguished from the role of the weights (discussed in [37, Sec.4.5.1]) assigned to reflect
the priorities of different goals in an MGSS. In fact, this is a different degree of freedom,
and goal priorities (the weights αi for i = 1, 2, . . . , d goals can be chosen independently
of the cutoff parameter a).
6.3 Optimizing an Infrastructure’s Resilience (Risk Treatment)
If the line between threats that need to be addressed and those that can be left aside
(for the moment) has been drawn, i.e., the risk evaluation phase is completed, the risk
treatment concerns the selection of controls to mitigate the threats.
Given one or more threats to be addressed, risk treatment is about selecting controls
against them. The particular selection of controls and resource allocation is a matter
of optimization, and this is where game theory comes into play. Specifically, it helps to
determine the optimal style of implementing controls towards minimizing the risks. The
procedure, on a high level, is defining a game matrix that has a list P S2 of adversary’s
actions (typically the threat or a set of threats to be considered), and another list P S1 of
controls that address the threats. Note that the setup of the game matrix, specifically the
definition of P S1 and P S2 should be made with care: it may well be the case that some
controls in P S1 may not address certain threats in P S2 and vice versa. For example, if
we have the countermeasures
• d1 ∈ P S1 : “installing an intrusion detection system”
• d2 ∈ P S1 : “install more fire extinguishers”
And the threat list P S2 includes
• T1 ∈ P S2 : fire outbreak
• T2 ∈ P S2 : hacking attack
Then it is obvious that d1 is pointless against T1 and d2 has no mitigating effect on
T2 . In such cases, one may consider a clustering of threats and controls based on their
mutual relevance. This amounts to selecting specific controls against specific threats,
and allows an optimized control selection “per threat” or “threat group” (if a control is
effective against more than one threat). Once the lists P S1 and P S2 have been specified,
it pays to sub-classify the actions in P S1 , as there are:
Static controls: these have a permanent effect in being a change to the system structure.
Their selection must be based on the goal of finding the minimal set of controls
that covers the maximal spectrum of threats. If several candidate controls are
available, then the problem is “game-theoretic” in the sense of being a humble
optimization for the defender having n strategies against a specific single strategy
of the attacker (that is the threat to be mitigated). If all threats shall be mitigated
at the same time, then the problem of control selection can be set up as a matrix
game, where n controls are available against m threats. The resulting equilibrium
consists of three output artefacts:
35
1. A probability distribution x∗ over the n controls
2. A probability distribution y ∗ over the m threats
3. An equilibrium payoff distribution V
Each of these is meaning- and useful, because:
• All controls having a nonzero probability assigned by the equilibrium distribution x∗ must be played for risk minimization. Thus, these controls need
to be implemented. Note that the magnitude of the probability is herein not
relevant, but can be taken as an indication of importance of the respective
countermeasure. For example, if the games equilibrium prescribes to “play”
the strategy “install another firewall” with likelihood 0.3, then it is obviously
meaningless to install a firewall only to an extent of 30%. Nonetheless, it
tells us that this control must be installed at all. However, if the firewall
installation gets a likelihood of 0.3, and another control, say “change passwords” gets a likelihood of 0.6, then this indicates that both of these measures
must be implemented, but changing the access credentials (or enforcing this)
should go first, as it is the more frequent (thus more important) strategy in
the equilibrium.
• The likelihoods of different threats can be taken as an indication of “how
likely” a threat is in the worst case. This indication must, however, be interpreted in light of the non-uniqueness of equilibria, which entails that the
computed distribution y ∗ is indeed a worst case scenario, but there may be
(many) others too.
• The equilibrium payoff distribution V quantifies the residual risk remaining
after the controls have been implemented. If, say the average (mean value)
of V turns out to be not satisfying, then the entire selection process may be
repeated (in Figure 1, this is risk decision point 2, where the process may be
restarted from the “establish context” phase.
Alternatively, we can ask for the smallest set of countermeasures that are simultaneously effective against all threats. Given the information of which defense
measure is effective against which threat (in the form of a set Si of candidate
countermeasures for the i-th threat), the problem is to compute a hitting set for
the family {S1 , S2 , . . .}. That is a well-known problem and can be solved using
linear optimization, or more direct algorithms found in the literature.
Dynamic controls: these have a volatile effect and need to be repeated, such as patching
and security awareness trainings. The latter is also an example that justifies the
approach of letting the outcome of a control being uncertain and in fact quantified
by a probability distribution. In case of a security awareness training, this distribution could reflect the (relative) amount of people that are (afterwards) highly
aware, medium aware or even remain unaware despite the training. In any case,
the outcome is never certain and depends on the person and its background. The
second example of patching equally well justifies the modeling of outcomes with
36
probability distributions: patching may be prescribed, but the patch may be unavailable at the current time, or it may not help against the right vulnerability. So,
the outcome of “patching” is also uncertain and the model put forth in [36] explicitly accounts for this. The selection and implementation of such dynamic controls
works as for the static controls described before, but with an important difference
in the use of the artefact distribution x∗ : this is a randomized prescription telling
the frequency of how often a control is repeated. That is, if a control, say “patching”, gets assigned a probability of 60% by x∗ , then the implementation is done
as follows: on each day, we toss a biased coin, coming up heads with a chance
of 60%. If it does so, the administrator is told to look for and install a patch.
The other days, no patching is done. This example is of course oversimplified, but
shall illustrate the “security-by-randomness” approach that game theory enforces:
security against the attacker is gained by it not knowing the current patch level
reliably (as the patching can happen every day), so anything that the attacker has
been learned so far may soon be invalidated, and hence the system is more secure.
A Purely Combinatorial Selection Technique: A different selection method for risk
controls is offered by the methods of model-based diagnosis (MBD) [42]. In a nutshell,
MBD (as put forth in [13, 42]) offers a systematic way of testing the effectiveness of
subsets of P S1 against subsets of P S2 . The goal is finding a minimal set of actions in P S1
that address all threats in P S2 . More formally, let there be a relation ⊆ P S1 × P S2 ,
with the semantics that
d1
d2 , if and only if defense d1 is working against threat d2 .
Conversely, if some defense action d1 is useless against threat d2 , then (d1 , d2 ) ∈
/ .
Using this relation, we can associate each threat with countermeasures relevant to it,
giving a family of sets as
P := {{di ∈ P S2 : di
t} : t ∈ P S1 } ,
i.e., an entry in the family P is a set Ci := {di1 , di2 , . . . , dik } of all k countermeasures
that are effective against threat di ∈ P S2 .
The sought optimal selection of controls, a diagnosis in the terminology of [42], is a
minimal hitting-set D for P. Minimality can here be understood by cardinality or in
terms of ⊆-relation, so that no smaller set than D is a hitting set (this is the usually
preferred setup). By definition, a hitting set D has the property of intersecting every
element in P (formally, D ∩ Ci 6= ∅ for all Ci ∈ P). Again, interpreting this in our
context, this means that the selected security defense measures in D effectively address
all threats in P S2 (this is the hitting-set property), and any proper subset of these
controls would leave at least one threat untouched (this is the ⊆-minimality).
Computing hitting sets is a simple matter of setting up a linear integer constrained
optimization problem, or possible by direct computation and heuristics. We leave the
computational details aside here, except for the final remark that the problem of computing minimal hitting sets is NP-hard in general (see [21]). This is why a game-theoretic
method may be more efficient and hence preferable here.
37
6.4 Software Support
The entire theory outlined up to this point has been implemented as a package for
the R-system [32], named HyRiM [39]. Plots and results given to illustrate matters in
the following have all been obtained using this implementation, and we will give the
respective call sequence (on the R shell) here in a condensed form that may serve as an
example-based tutorial for applying the theory elsewhere.
It must be emphasized that the implementation mentioned here cannot be a substitute
for the entire risk management process as outlined in section 2. Duties of context
definition, infrastructure modeling, threat identification and listing of defense actions
should be subject of extensive use of established (relevant) software tools and standards
in this context. Vulnerability scanners like Nessus [46], reporting tools like Cauldron [26],
or tool support to work with the ISO or BSI standards family are strongly recommended
in these initial phases of the risk management process.
1. Assume that the expert opinions are available as a numeric vector obsij containing
all the answers for the scenario (di , dj ) ∈ P S1 × P S2 . That is, obsij is a vector
of integers in the range {1, 2, . . . , M }, where M is the maximum value (for our
examples, we took M = 10).
Given a distinct such vector obsij for each i, j, we can compile a loss distribution
by the call:
fij <- lossDistribution(obsij)
This produces a continuous loss model as displayed in Figure 10. A visualization
as in this figure is obtained by either plotting a single loss distribution
plot(f11, cutoff=6)
(in which the parameter cutoff is optional and defaults to the maximum observation
in the data from which the loss distribution has been constructed), or by plotting
the entire game matrix by calling:
plot(G, cutoff=6)
where the parameter cutoff is again optional and, if supplied, is passed to the inner
plots of the payoff distributions.
Assume that the previous step has been repeated to construct continuous loss
distributions f11, f12, f21 and f22. Discrete loss models (with smoothing) can be
obtained by telling lossDistribution to come up with a discrete model. This is
done by adding an according flag, and (optionally) supply a bandwidth parameter
(which if omitted, defaults to Silverman’s rule of thumb, applied to the data is if
it were continuous):
fij <- lossDistribution(obsij,discrete=TRUE,supp=c(1,10))
38
The parameter supp is supplied as a reminder (also for the user) that all loss
distributions need to have the same support (otherwise, the construction of games
upon a set of distributions with different supports will fail with an error message
reported).
2. From the (continuous) loss models l11, l12, l21 and l22, we compile the MGSS
by an invocation of:
G <- mosg(n = 2, m = 2,
losses = list(f11, f12, f21, f22),
goals=1,
defensesDescr = c("d1", "d2"),
attacksDescr = c("a1", "a2"),
goalDescriptions = c("goal"))
in which the textual descriptions of the strategy spaces P S1 and P S2 are optional
and insignificant for the subsequent analysis (they only show up in plots).
3. Given the game object, we can call mgss to give us a security strategy. This
function takes the cutoff point as a direct parameter, which if omitted, defaults to
the right end of the (common) support of all distributions (no truncation):
mgss(G,T=1000,cutOff=6)
Here, the parameter T is the number of iterations to the FP algorithm.
The final result returned by mgss is an object of class mosg.equilibrium, and contains
the following fields:
• $optimalDefense: an n-dimensional column vector with rows labeled by the
names of the entries in P S1 (parameter defensesDescr in the previous call to
mosg).
• $optimalAttacks: an m-dimensional column vector with rows labeled by the
names of the entries in P S2 (parameter attacksDescr in the previous call to
mosg).
• $assurances: a list of mosg.lossDistribution objects (similar as returned by
the lossDistribution function used to construct the game structure). They
can be plotted (see Figure 10 for examples), printed in detail (using the summary
function), and related statistical quantities can be worked out such as:
– probabilities for individual attacks (probability mass function can be evaluated by the generic function density)
– the cumulative distribution function cdf
39
– moments (of given order, returned by the function moment), mean, variance
and quantiles (at given levels, returned by the generic function quantile).
For example, after having computed an equilibrium in our generalized setting, a simple
numeric risk measure according to the rule “risk = damage × likelihood” can be recovered
by invoking
mean(eq$assurances$goal)
which gives the assured expected loss (= usual quantitative risk) in the first security goal
(g1) of the game. If, besides the mean, the risk manager is interested in the expected
variation around the mean, then it is an easy matter of computing the variance of the
assured loss distribution as
variance(eq$assurances$goal)
Finally, if the risk manager is interested in the chances of suffering losses of categories
“high” or above, he/she may invoke the cdf function in the proper way. For example, if
category “high” has the number 4, then the chances of getting such a loss or more are
given by
1 - cdf(eq$assurances$goal, 3)
Conversely, computing quantiles tells the losses expected up to a, say 95%, chance. That
is, calling
quantile(eq$assurances$goal, 0.95)
tells us the maximal loss (category) that occurs in 95% of the cases.
6.5 Playing the Equilibrium Strategies
To cast frequencies from the game into a practical project plan that tells the security
officer when to take which action, let us fix a particular action and call p its equilibrium
frequency, given the time unit (e.g., one month). The number of actions per time interval
T is therefore p · T and in fact is a Poisson distributed variable with rate parameter p.
In turn, the time between two actions is exponentially distributed with parameter 1/p.
So, if we are at time t0 , the next repetition of the action should be at time t0 + ∆, where
∆ ∼ Exp(1/p) is an exponentially distributed random variable and measured in the unit
of time that was fixed before. Sampling exponentially distributed variables is easy using
uniformly distributed random numbers in the unit interval. Given such a random value
U ∼ U nif orm(0, 1), we simply compute ∆ = −p · log (U ) to have ∆ ∼ Exp(1/p).
Note that this method yields the desired fraction of actions (events) per time unit,
which is only correct if the gameplay is presumed with a periodicity taking time T .
Thus, the rating of strategies (the questionnaire; cf. section 3.3) should be done bearing
in mind the game period T .
For example, if the defense is a security awareness training, and the unit of time T is
one month, then the threat assessment could ask something like:
40
Consider an employee whose last security awareness training happened no
more than a month ago. How likely do you think is this person plugging a
virulent USB stick into her/his computer at work?
A game-theoretic treatment of risks using actions that need repetition will return
likelihoods for each action to occur (in an optimal defense strategy). Repeatable actions
correspond to repeatable and independent game repetitions. To make this semantically
sound, we thus need to make the assessment in light of the game’s periodicity, to decide
whether or not it is necessary to take a particular action again. Returning to the previous
example of security training, we certainly will seek to avoid having such training every
month, but the optimal periodicity of the action is far from obvious. A game-theoretic
analysis will deliver a frequency of training, counted over the repetitions of the game. If
we – hypothetically – let the game duration be one month before the next (hypothetical)
iteration starts, then the question of the effectiveness of a security training must refer
to the current period, which lasts no more than one month up back in time. Hence, we
get the proposed question sketched above.
7 On Model Validation and Verification
It lies in the nature of risk management that there is no such thing as a “rollback” that
could reset a company into the state before an incident, to verify if the here proposed
decision framework or another would have delivered the better outcomes. The same
problem is very well known from other fields of science, and clinical studies tackle the
issue by introducing control groups to be able to clearly relate observed effects to the
treatment rather than coincidence (or a placebo effect).
Things are no different in the area of risk management, except perhaps at increased
complexity and difficulty. In a famous quote, George Box is telling us that essentially
all models are wrong, but some are still useful. His point is definitely right, but it must
not be taken as an excuse to not ask for the quality of a decision model. While it is
certainly true that incorrect model parameters lead to incorrect model output, judging
the parameterizations as wrong because of the model performing badly would be logically
flawed (as it incorrectly reverses the implication).
Instead, the model validation must be up to long term studies with control groups,
followed by well-designed statistical tests to reject the hypothesis of the model performing identically to a chosen alternative. The goal of rejecting the hypothesis is crucial
here, since a statistical test can never be a formal proof, but only provide an empirical counterexample against some claim. Thus, to verify the claim of the here-proposed
method to be a useful risk decision supporting framework, we need to refute the claim
of it performing equal to a competing technique. This setting is very well known from
clinical studies incurs some challenges for risk management, since security incidents are
random events, and cannot be considered as independent or isolated from one another
(advanced persistent threats are only one hidden cause that can manifest itself in a
long sequence of seemingly independent but nevertheless cumulating events up to a final
catastrophe). Moreover, there are no such things as “lab conditions”, under which the
41
effect of a decision can be tested against a different one. Thus, a control group in our
setting can only be another company, but incidents – particularly APTs – are focused
and tailored to the company, so comparing data from one company to that of another
is most likely not too meaningful, since incidents happen to both companies, but in
crucially different forms.
This substantiates the aforementioned long-run of studying the effects of one decision
making against another. Statistical tests are a natural tool here, but the selection of an
appropriate one requires care for several reasons:
1. Parametric tests come with assumptions on the distributions of the sample sets
under investigation. For example, assuming Gaussian distributions here could
be flawed, since we talk about losses that are typically modeled by heavy tailed
distributions.
2. The tested hypothesis must be considered relative to how the risk management
works. For example, tests on equality of means appear quite well-aligned to risk
management that minimizes expected damages, i.e. means. That is, if the alternative risk management against which the proposed method is to be tested
considers only means (for example, if risk is defined as the product of likelihood
and impact), while our framework allows for a disregard of this quantity (since
the hyperreal ordering remains unchanged upon disregarding any finite number
of moments, which can include the first moment), the test may unintendedly be
biased towards favouring the alternative hypothesis.
Similar issues apply to the selection of the “control group”, because the benchmark
figures that an enterprise uses to measure its own performance may be highly different
between companies and may be highly specific for the business area. One benchmark
figure that may have wide applicability and – in the end of the day – may be the only
high-level figure that counts for risk management, is the number of severe incidents. This
number depends on the understanding of “severe”, and the threshold at which losses are
counted as severe can be defined individually for each enterprise.
Towards a method for validating our risk management decision framework, we will
thus refrain from specifying a particular risk threshold here, and go for verifying the
purely qualitative intuition that:
A good risk management framework should minimize the number of severe
incidents.
Consequently, to assess the “goodness” of the decision framework based on game theory,
the proposed procedure is as follows, for any particular (i.e., fixed) enterprise:
1. Define a control group: since data coming from a different company may be biased
and relate to different endo- and exogenous processes, the control group should be
the enterprise itself. This also fixes the risk management that we test our method
against, since it simply is the current decision making framework that we want to
judge against the new (here proposed) framework.
42
2. Fix a meaning of “severe incident” in the sense of specifying a threshold above
which damages or losses count as severe. Given a series of historic records on
losses that occurred in the past, one can compile a past loss distribution Fpast ,
and, for example, define the threshold using a quantile of Fpast . In any case, it
is important to fix this threshold a-priori (still, we need to bear in mind that
the thresholds may need adaption over the lifetime of the enterprise to remain
meaningful), and independently of the value implied by [36, Thm.2.14].(this is to
avoid a bias in favour of the here proposed decision making technique). Call the
threshold t0 , and let us think of losses > t0 as being severe.
3. Collect the number n0 of incidents whose losses were ≥ t0 , i.e., which have been
severe in this understanding. Based on the distribution Fpast of past losses, the
likelihood for such an incident to occur would be p = Pr(L ≥ t0 ), if L measures
the random loss. Assuming independence of events, the number N of severe past
incidents is Poisson-distributed with a rate parameter λpast that depends on p.
For the data collection as such, the independence assumption entails that common
underlying causes of seemingly independent events must be identified before feeding
this data into a test statistic. Several events with a single common cause must then
be compiled into a single event (representing the underlying cause) with a total loss
incurred. For example, if many events can be related to an APT, the typical nature
of these symptoms is to stay undetected, so they would probably not count into
the set of “severe” incidents. Nonetheless, if the APT manifests itself in several
quite severe damage events, then these must be considered as one single incident
(being the APT, only exhibiting multiple damages) with accumulated effects.
4. At some point in time, the new risk management framework (not necessarily the
one described in this report) becomes installed, and a new recording of “severe”
events (with the same definition as was used in the past) must start. Call the
number of these M , and observe that the new loss distribution Fnew is as well
Possonian with a rate parameter λnew , which depends on the mass p that Fnew
assigns to the region ≥ t0 .
5. Fix equal time periods for data collection: For the new collection of losses, it is
crucial that the time period over which M was counted equals the period over which
N was taken. Otherwise, the test would be biased, since if the past period is much
shorter than the current one, then the number of past incidents is necessarily much
smaller than the number of current incidents. Likewise, if the past period was very
long, and the new period is relatively short, then the number of incidents N will
necessarily be much higher than the number M of currently observed incidents.
6. A final assumption that must nevertheless be emphasized is both rates λpast , λnew
being taken as constant over time (otherwise, the events would not be Poisson
distributed any more).
Given the past and current samples with sizes n and m, respectively, the distributions
of N and M are N ∼ Poisson(nλpast ) and M ∼ Poisson(mλnew ). The hypothesis to be
43
refuted is the past risk management performing better or equal to the new method, in
which case the null hypothesis becomes
H0 :
λpast
≤ 1,
λnew
which expresses that the past risk management reduces severe incidents at least as good
as the current (new) one, hence the new risk management is no better than the previous.
The statistical test of choice is a one-sided rate ratio test [14, 29]. These references
discuss various methods to test Poisson rates, and one exact test is directly available in R
via the package rateratio.test [11,12], which can be invoked by issuing the command:
rateratio.test(c(n·λpast , m·λnew ), c(m, n), alternative = "greater")
for our case. The null-hypothesis is rejected if the p-value is less than the pre-specified
1 − α (that defaults to α =conf.level= 0.95 in the above call).
A general validation or verification of the method against competing ones is thus a
matter of several companies running long-time evaluations and deciding on their own
if the decision framework proposed here performs better or worse than their competing
candidate. Extending the investigation towards checking whether or not the stochastic order makes up a good model of decision making under bounded rationality is an
independent challenge and widely extended line of research that we will not go farther
into here. Instead, the last test is meant to complete the theory developed here, and we
close the report by a discussion in favour of the qualitative advantages that this form of
decision making may offer for practitioners.
8 Conclusion and Outlook
The most important pros of the method can be roughly summarized as follows:
• No consensus problem: The approach of working with entire distributions rather
than single values avoids issues of consensus finding among experts. That is, if there
are disagreeing opinions among a group of people, every opinion goes into the risk
analysis at equal weight. Thus, nobody has to be overruled or otherwise accept a
compromise that deviates from the own thoughts. If, subsequently, damage occurs
nonetheless, arguments of people claiming to have foreseen this incident are useless,
since any such warning was considered from the beginning on already.
• Vague assessments allowed : Even for a single expert, there is no need to confine
oneself to a single risk assessment, perhaps at the risk of being mistakenly understood as also saying what the risk “is not”. In other words, if a security expert
utters a risk of level “medium”, then this is not the implicit additional statement
that the risk is “not high”. In fact, in allowing the expert to add uncertainty
44
to her/his assessment4 , the intrinsic fuzziness of perceived security risks can be
(conveniently) expressed.
This degree of freedom also helps to treat scenarios in which several threats occur
jointly. If a threat scenario is such that most likely other attacks are mounted
along, then the expert may still focus on the “main objective” (main threat) of
the assessment, while explicitly stating and rating the possibility of more severe
damages upon the simultaneous occurrence of other threats by assigning some
likelihood to the other alternatives (cf. remark 3.1).
• Enforced systematization: The need to set up a game matrix for all combinations of
defenses and attacks enforces a systematic approach to the entire risk management
problem. The method itself is herein merely embedded and rests on a comprehensive threat analysis and identification of countermeasures. The method of risk
assessment as outlined in section 3.3 is actually an application of crowdsourcing;
see [3].
Perhaps the most important advantage over more conventional quantitative methods
of risk assessment is the following: none of the information items (except the average)
returned by the call sequences at the end of Section 6.4, can be obtained by standard
techniques of risk management that are entirely based on expected losses as risk. Hence,
the preservation of information throughout the entire process from the modeling up
to the final risk result interpretation is an important benefit over other methods. In
brief: the here proposed method is preserving information of all experts over the entire
risk management process, and avoids loss of information through consensus finding and
opinion or risk aggregation.
Acknowledgment
This work was supported by the European Commission’s Project No. 608090, HyRiM
(Hybrid Risk Management for Utility Networks) under the 7th Framework Programme
(FP7-SEC-2013-1). The author is indebted to the whole consortium of HyRiM, and
particular gratitude is expressed to Sandra König and Stefan Schauer from the Austrian Institute of Technology for invaluable discussions and suggestions along this entire
research.
References
[1] Rudolf Avenhaus, Bernhard von Stengel, and Shmuel Zamir. Inspection games.
In R.J. Aumann and S. Hart, editors, Handbook of Game Theory with Economic
Applications, volume 3, pages 1947–1987. Elsevier, 2002.
4
either explicitly by giving the loss distribution directly (cf. Example 5.2), or implicitly by adjusting
the bandwidth parameter of the kernel smoothing (see Section 3.10 and Example 5.2) to assign
likelihoods to other possibilities around the actual assessment.
45
[2] Boldizsár Bencsáth, Gábor Pék, Levente Buttyán, and Mark Felegyhazi. The
cousins of stuxnet: Duqu, flame, and gauss. Future Internet, 4(4):971–1003, 2012.
[3] David Bratvold.
What is crowdsourcing – daily crowdsource.
https:
//dailycrowdsource.com/training/crowdsourcing/what-is-crowdsourcing,
2016. [retrieved on September 23, 2016].
[4] Bundesamt für Sicherheit in der Informationstechnik. BSI-Standard 100-2: ITGrundschutz Methodology. https://www.bsi.bund.de/cln_156/ContentBSI/
grundschutz/intl/intl.html, May 2008. Version 2.0, english.
[5] Arthur Carvalho and Kate Larson. A consensual linear opinion pool, 2013.
[6] Peter Chesson and Charlotte T. Lee. Families of discrete kernels for modeling
dispersal. Theoretical Population Biology, pages 241–256, 2005.
[7] Dell
SecureWorks.
Advanced
threat
protection
with
dell
secureworks.
https://www.secureworks.com/resources/
sb-advanced-threat-protection-with-dell-secureworks, retrieved:
July
27th, 2016, 2014.
[8] Luc Devroye and László Györfi. Nonparametric Density Estimation - The L1 View.
Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, New
York, 1985.
[9] Marten Dijk, Ari Juels, Alina Oprea, and Ronald L. Rivest. Flipit: The game of
”stealthy takeover”. J. Cryptol., 26(4):655–713, October 2013.
[10] James P Farwell and Rafal Rohozinski. Stuxnet and the future of cyber war. Survival, 53(1):23–40, 2011.
[11] Michael Fay. Two-sided exact tests and matching confidence intervals for discrete
data. R Journal, 2(1):53–58, 2010.
[12] Michael Fay. rateratio.test: Exact rate ratio test, 2015. https://cran.r-project.
org/package=rateratio.test [retrieved Dec. 19, 2016].
[13] Russell Greiner, Barbara A. Smith, and Ralph W. Wilkerson. A correction to the
algorithm in reiter’s theory of diagnosis. Artificial Intelligence, 41(1):79–88, 1989.
[14] Kangxia Gu, Hon Keung Tony Ng, Man Lai Tang, and William R. Schucany. Testing the ratio of two poisson rates. Biometrical journal. Biometrische Zeitschrift,
50(2):283–298, 2008.
[15] International Standards Organisation (ISO). ISO/IEC 31000 - Risk management –
Principles and guidelines. http://www.iso.org/iso/home/store/catalogue_tc/
catalogue_detail.htm?csnumber=43170, 2009. (accessed: April 11, 2016).
46
[16] International Standards Organisation (ISO).
ISO/IEC 27005 - Information technology – Security techniques – Information security risk management.
http://www.iso.org/iso/home/store/catalogue_tc/catalogue_
detail.htm?csnumber=56742, 2011. (accessed: April 11, 2016).
[17] International Standards Organisation (ISO). ISO/IEC 27000. http://www.27000.
org/, 2016.
[18] Sushil Jajodia, Anup K. Ghosh, V. S. Subrahmanian, Vipin Swarup, Cliff Wang,
and Xiaoyang Sean Wang, editors. Moving Target Defense II – Application of Game
Theory and Adversarial Modeling, volume 100 of Advances in Information Security.
Springer, 2013.
[19] Sushil Jajodia, Anup K. Ghosh, Vipin Swarup, Cliff Wang, and Xiaoyang Sean
Wang, editors. Moving Target Defense - Creating Asymmetric Uncertainty for Cyber
Threats, volume 54 of Advances in Information Security. Springer, 2011.
[20] Ashutosh Garg Jayram, Ashutosh Garg, T. S. Jayram, Shivakumar Vaithyanathan,
and Huaiyu Zhu. Generalized opinion pooling. In In Proceedings of the 8th Intl.
Symp. on Artificial Intelligence and Mathematics, pages 79–86, 2004.
[21] Richard M. Karp. Reducibility among combinatorial problems. In R. E. Miller
and J. W. Thatcher, editors, Complexity of Computer Computations, pages 85–103,
New York: Plenum, 1972.
[22] Sandra König. Error propagation through a network with non-uniform failure.
arXiv:1604.03558 [stat.AP], 2016.
[23] George Kurtz. Operation “aurora” hit google & others.
http://siblog. mcafee. com/cto/operation-% E2, 80, 2010.
Published online at
[24] Q. Li and J.S. Racine. Nonparametric estimation of distributions with categorical
and continuous data. Journal of Multivariate Analysis, 86:266–292, 2003.
[25] Mohammad Hossein Manshaei, Quanyan Zhu, Tansu Alpcan, Tamer Bacşar, and
Jean-Pierre Hubaux. Game theory meets network security and privacy. ACM Comput. Surv., 45(3):25:1–25:39, July 2013.
[26] Combinatorial analysis utilizing logical dependencies residing on networks (CAULDRON), 2008. http://ait.gmu.edu/~csis/.
[27] Geoffrey J. McLachlan and Thriyambakam Krishnan. The EM algorithm and extensions. Wiley Series in Probability and Statistics. Wiley-Interscience, Hoboken,
NJ, 2. ed. edition, 2008.
[28] Kate Munro. Deconstructing flame: the limitations of traditional defences. Computer Fraud & Security, 2012(10):8–11, 2012.
[29] NCSS Software. PASS 14 Documentation, 2016. [retrieved Dec. 19, 2016].
47
[30] Jeffrey Pawlick, Sadegh Farhang, and Quanyan Zhu. Flip the Cloud: Cyber-Physical
Signaling Games in the Presence of Advanced Persistent Threats, pages 289–308.
Springer International Publishing, Cham, 2015.
[31] William D. Perreault and Laurence E. Leigh. Reliability of nominal data based on
qualitative judgments. Journal of Marketing Research, 26(2):135, 1989.
[32] R Development Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2016. ISBN
3-900051-07-0.
[33] Balaji Rajagopalan and Upmanu Lall. A kernel estimator for discrete distributions.
4:409–426, 1995. Gordon and Breach Science Publishers SA.
[34] Lisa Rajbhandari and Einar Arthur Snekkenes. Mapping between Classical Risk
Management and Game Theoretical Approaches, pages 147–154. Springer, Berlin,
Heidelberg, 2011.
[35] S. Rass. Game-theoretic risk management – part one: Security strategies in nondeterministic games, 2015. internal report of the HyRiM Project.
[36] S. Rass. On Game-Theoretic Risk Management (Part One) – Towards a Theory of
Games with Payoffs that are Probability-Distributions. ArXiv e-prints, June 2015.
http://arxiv.org/abs/1506.07368.
[37] S. Rass. On game-theoretic risk management (part two) – algorithms to compute
nash-equilibria in games with distributions as payoffs. arXiv:1511.08591, 2015.
[38] S. Rass, S. König, and S. Schauer. Uncertainty in games: Using probability distributions as payoffs. In MHR Khouzani, Emmanouil Panaousis, and George Theodorakopoulos, editors, Decision and Game Theory for Security, 6th International Conference, GameSec 2015, LNCS 9406. Springer, 2015.
[39] Stefan Rass and Sandra König. R Package ’hyrim’: Multicriteria risk management
using zero-sum games with vector-valued payoffs that are probability distributions.
https://hyrim.net/software/, 2017. (current stable release as of Oct.2017; ongoing
development).
[40] Stefan Rass, Sandra Konig, and Stefan Schauer. Defending Against Advanced Persistent Threats Using Game-Theory. PLoS ONE, 12(1):e0168675, 2017. Journal
Article.
[41] Stefan Rass and Quanyan Zhu. GADAPT: A Sequential Game-Theoretic Framework for Designing Defense-in-Depth Strategies Against Advanced Persistent
Threats. In Quanyan Zhu, Tansu Alpcan, Emmanouil Panaousis, Milind Tambe,
and William Casey, editors, Decision and Game Theory for Security, volume 9996
of Lecture Notes in Computer Science, pages 314–326. Springer International Publishing, Cham, 2016.
48
[42] Raymond Reiter. A theory of diagnosis from first prinicples. Artificial Intelligence,
32(1):57–95, April 1987.
[43] B.W. Silverman. Density Estimation for Statistics and Data Analysis. Chapman &
Hall/CRC, 1998.
[44] Anoop Singhal and Ximming Ou. Security risk analysis of enterprise networks using
probabilistic attack graphs. Technical report, National Institute of Standards and
Technology (NIST), 2011. NIST Interagency Report 7788.
[45] Julian Talbot.
Whats right with risk matrices?
– an great tool
for
risk
managers...
https://31000risk.wordpress.com/article/
what-s-right-with-risk-matrices-3dksezemjiq54-4/., 2011.
[retrieved
Sep.29th, 2016].
[46] Tenable Network Security R . NESSUS – the vulnerability scanner. www.nessus.
org, 2011.
[47] Alexandre B. Tsybakov. Introduction to Nonparametric Estimation. Springer Series
in Statistics. Springer, 2009.
[48] Roberto Uzal, Narayan C Debnath, Daniel Riesco, and German Montejano. Trust
in cyberspace: New information security paradigm. Managing trust in cyberspace,
page 405, 2013.
[49] Larry Wasserman. All of Nonparametric Statistics: A Concise Course in Nonparametric Statistical Inference. Springer, 2007.
[50] Quanyan Zhu and Tamer Başar. Game-theoretic approach to feedback-driven multistage moving target defense. In 4th International Conference on Decision and Game
Theory for Security - Volume 8252, GameSec 2013, pages 246–263, New York, NY,
USA, 2013. Springer-Verlag New York, Inc.
49
| 10math.ST
|
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
arXiv:1410.6121v2 [cond-mat.str-el] 9 Dec 2014
International Journal of Modern Physics B
c World Scientific Publishing Company
THE NONEQUILIBRIUM QUANTUM MANY-BODY PROBLEM
AS A PARADIGM FOR EXTREME DATA SCIENCE
J. K. Freericks
Department of Physiocs, Georgetown University, 37th and O Sts. NW
Washington, District of Columbia 20057, U. S. A.
[email protected]
B. K. Nikolić
Department of Physics and Astronomy, University of Delaware, Newark, DE 19716, U. S. A.
[email protected]
O. Frieder
Department of Computer Science, Georgetown University, 37th and O Sts. NW
Washington, District of Columbia 20057, U. S. A.
[email protected]
Received Day Month Year
Revised Day Month Year
Generating big data pervades much of physics. But some problems, which we call extreme
data problems, are too large to be treated within big data science. The nonequilibrium
quantum many-body problem on a lattice is just such a problem, where the Hilbert
space grows exponentially with system size and rapidly becomes too large to fit on any
computer (and can be effectively thought of as an infinite-sized data set). Nevertheless,
much progress has been made with computational methods on this problem, which serve
as a paradigm for how one can approach and attack extreme data problems. In addition, viewing these physics problems from a computer-science perspective leads to new
approaches that can be tried to solve them more accurately and for longer times. We
review a number of these different ideas here.
Keywords: Nonequilibrium quantum many-body problem; extreme data science
1. Extreme data science
Conventional forms of big data science and data mining involve the creation of
large finite data sets and the analysis of that data to produce insight and to predict
future behavior of the system. Many physics problems are being analyzed in this big
data paradigm, such as high-energy-physics scattering experiments, which routinely
create petabytes of data. Other fields of science also involve big data science, with
large datasets being analyzed in fields as diverse as weather forecasting, genomics,
and neuroscience. Many algorithms and data analysis tools are used to interact with
such datasets; often they are created specifically for a given data set, even if they
1
December 10, 2014
2
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
employ more general purpose algorithms and strategies for the data analysis.
We discuss physics problems for which the datasets are so large they can effectively be treated as infinite in size, which we call extreme data science. In general,
such problems are impossible to solve unless at any given time, only a small subset of
the large dataspace is physically active, and that the system maintains a reasonably
small footprint within the data space for future times as well.
One technique which has been developed in the previous decade for dealing with
big data in computer science is the technique of compressive sensing. This method
represents a multidimensional signal with a much lower dimensional model. Instead
of employing conventional signal compression techniques, which aim to capture the
whole signal, compressive sensing attempts to directly capture the sparse signal1,2 .
One can understand how this works in two ways: (1) most data in real world problems are overdetermined and hence can be represented in a more sparse fashion and
(2) from a mathematical standpoint, a time series originating from a Fourier series
or from a Fourier transform look similar on a finite time interval, but the series
can be represented sparsely, while the transform cannot. This latter feature is now
being exploited in a wide range of physics problems.
The unprecedented growth of data has also challenged the most affordable computational procedures within big data science—namely numerical linear algebra.
Even partial eigenvalue or singular value decompositions become too expensive for
matrices employed in graph mining of large networks, of information retrieval, and
so on. As an alternative, randomized or stochastic processes have emerged as a
promising new approach to these problems3,4,5 . Regardless of these developments
within big data science, the field of physics has been grappling with these issues for
decades, and have come up with a number of ideas for these types of problems. We
believe a cross-fertilization of ideas from computer science, applied mathematics,
and physics, will enable much more rapid growth in the ability to solve extreme
data science problems.
One physics-based example that fits into the extreme data paradigm is the
nonequilibrium quantum many-body problem. This problem governs many different behaviors in physical systems. Some examples include the following: (1) the
current-voltage characteristic for transport of charge through a device made from
strongly correlated materials; (2) the response functions for pump/probe experiments, where intense lasers are shone onto strongly correlated materials; (3) the
response of a cold atom system following the quench of an interaction generated
by a Feshbach resonance; (4) the response of materials to shock compression; (5)
the evolution and thermalization of the quark-gluon plasma in heavy ion collisions;
(6) particle production and the reheating/thermalization of the universe after inflation; (7) photo-ionization of atoms or molecules; and (8) the dynamical behavior of
charged plasmas created by intense lasers.
It is well known that the Hilbert space for the many-body problem grows exponentially with the number of lattices sites in the problem. This growth is so rapid,
that supercomputers can only handle problems that have a few tens of lattice sites
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
3
before the memory is completely exhausted. Such small systems often are too small
to be able to reliably predict properties of 1023 particles. Fortunately, because of
physical principles like those found in the renormalization group, many equilibrium
many-body problems are governed by effective low-energy field theories, where the
high-energy degrees of freedom are frozen out, and do not change as the temperature is further lowered. Using these ideas, the numerical renormalization group
showed how one could numerically solve the Kondo problem, which is one of the
most important many-body problems in condensed matter physics. A large number
of other techniques exist for treating equilibrium many-body problems, and they often are governed by the ideas of the renormalization group, via energy minimization
methods where the quantum problem can be mapped onto an equivalent statistical
problem that can be solved using importance sampling methods like the Metropolis
algorithm, or via variational techniques chosen to minimize the loss of information
like the density matrix renormalization group.
In the case of nonequilibrium, problems become more complicated, because energy minimization is often no longer a valid physical principle, and the driving of
the system pushes it out of the low-energy effective field theory regime (of course,
reducing information loss is probably still a valid principle). At first glance, it then
seems hopeless to try to solve such problems, but significant progress has been made
recently with a number of different algorithms, even if they all are currently limited
by how far out in time they can evolve the system. In most cases this time scale is
too short to fully determine the long-time behavior. Nevertheless, it is highly likely
that even a nonequilibrium system will only evolve through a very small target subspace of the infinite dataspace (especially if entanglement growth is bounded), and
hence there is a good chance that appropriate algorithms can be developed to describe the behavior of these systems. In fact, there is a proof6 that states that if the
Hamiltonian involves local interactions (or interactions cut off after a short range),
then the time-dependent evolution of the system can only expand into an exponentially small subspace of the full Hilbert space in polynomial time. Techniques
that can focus in on just this subspace will be the most efficient in simulating the
nonequilibrium many-body problem.
When viewed from the computer-science perspective, extreme data science is
based on four principals: (i) encoding the active target subspace of the dataspace; (ii)
evolving the target subspace forward in time; (iii) discarding nonphysical dimensions
of the dataspace where the target has negligible weight; and (iv) extrapolating
(or forecasting) the behavior of the target space for future times from the past
history. Algorithms that solve the nonequilibrium quantum many-body problem
often employ these extreme data science principles in formulating their solution.
For example, since one cannot use all of the basis vectors of the Hilbert space to
represent a state vector, one must find an efficient method to encode the many-body
wavefunction which can describe the system with a small set of basis functions
initially, and can evolve to remain efficient in describing the system as it moves
forward in time (similar in spirit to compressive sensing). All algorithms formally
December 10, 2014
4
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
deal with the evolution of the system in time, as given by the quantum evolution
operator
Z
i t 0
0
dt H(t )
(1)
U (t, t0 ) = T exp −
~ t0
n Z t
Z tn−1
Z t1
Z t2
∞
X
i
=
dtn H(t1 )H(t2 ) · · · H(tn )
−
dt1
dt2
dt3 · · ·
~
t0
t0
t0
t0
n=0
which is a time-ordered exponential of the time-dependent Hamiltonian H(t), because this is how one ultimately determines the Green’s function, or the evolution
of the wavefunction. Third, algorithms often include some form of truncation step,
where the system is projected onto a subsystem that includes the most important
terms. Finally, since it is difficult to reach long times with most nonequilibrium
algorithms, one often attempts to extrapolate results either to try to stabilize the
algorithm so it can run further in time, or to try to forecast future behavior from
the prior knowledge.
In the remainder of this review, we describe a wide range of different algorithms
that have been used for the quantum many-body problem, describe their successes
or failures for the nonequilibrium systems, and illustrate a new attack motivated by
extreme data science to create new algorithms that might have a higher chance for
success. We group the algorithms into those that simplify the many-body problem
to solve it approximately, those that employ stochastic methods and the challenges
with generalizing them to the nonequilibrium case, algorithms based on renormalization group ideas, and algorithms that directly evaluate the evolution operator.
Note that while we have made every attempt to include a wide number of references to the original literature, this review is intended to be illustrative and not
exhaustive, so many appropriate references have not been included. We apologize
to any authors for whom we did not reference their relevant work due to length
restrictions. In addition, we have not made an exhaustive review of big data science
ideas as applied to the many-body problem. For example, machine learning ideas
have recently been employed to accelerate the solution of the problem7 . Instead, we
concentrate on the new extreme data science aspect of the problem.
We emphasize that the techniques reviewed here focus on methods that work
directly or indirectly in Hilbert space. Recently, Gaussian phase space methods8,9
have been proposed to evade the exponential growth of the problem size in Hilbert
space. They work by mapping the quantum evolution in real or imaginary time onto
a set of stochastic differential equations with a drastically reduced dimensionality.
Much excitement arose for this method early on, particularly with the anticipation
that phase-space methods could solve the fermionic sign problem10 . But a careful
analysis11,12 shows that the nonvanishing of the distribution function evolved by
stochastic differential equations at the boundaries, introduces systematic errors such
as spiking trajectories or rapid growth of the sampling error which currently prevent
extracting reliable results for fermionic systems in the limits of strong interaction,
large system sizes, or long times (the approach has been more successful for bosonic
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
5
systems though13 ). In any case, we focus the remainder of our discussion on Hilbertspace-based methods.
2. Algorithms that truncate the dataspace or simplify the
many-body problem
One strategy to take for an extreme data problem is to truncate or simplify it so that
it results in a finite dataspace that can fit onto the computer and be solved. Within
the many-body problem, there are a few different approaches that fall into this category: (i) exact diagonalization of finite systems; (ii) configuration-interaction-based
methods; (iii) time-dependent density-functional theory; and (iv) the Gutzwiller
variational method. We describe these approaches next.
Exact diagonalization is one of the first methods14 tried for many-body problems like the Hubbard model15 . Early work by the Falicov group16 or the work by
Heilmann and Lieb17 that investigated the properties of the benzene molecule, focused on determining energy eigenvalues and eigenstates after taking into account
all of the symmetries of the system. The technique has evolved significantly since
then, and now has been applied to many nonequilibrium problems where strongly
correlated materials are driven by high intensity laser fields18,19 .
For equilibrium problems, one can handle systems with large Hilbert spaces because the Hamiltonian is sparse, and diagonalization methods like the Davidson20
or Lanczos21 methods are iterative and require only the action of the Hamiltonian
on a state vector in the Hilbert space (in most of these calculations a number of
vectors in the Hilbert space are stored in the computer memory). This can often
be calculated without even storing the Hamiltonian matrix, when one uses the operator form and acts it on representative basis states. These techniques can only
determine eigenvalues (and associated eigenvectors) within a finite energy interval,
and usually focus on the low-energy states (in principle, long-run Lanczos methods
that can compute all eigenvalues, or compute eigenvalues within specified energy
windows, do exist22 , but they require too much storage to be efficient methods for
fully diagonalizing large sparse matrices). By using time evolution methods, such
as evaluating the evolution operator via a Trotter expansion23 (or more sophisticated methods like commutator-free exponential time expansions24 ) or using the
Crank-Nicolson algorithm25 (which preserves unitarity through second order in the
discretization time step), one can evolve these finite systems forward in time and
solve nonequilibrium problems. This has been looked at recently by a number of
different groups18,19 .
The exact diagonalization approaches work with discrete eigenvalues and are
accurate in equilibrium when the temperature is large relative to the average level
spacing. For nonequilibrium, the time evolution is usually very accurate for short
times, but then starts to deviate from the correct behavior when it displays recurrences, which occur for times such that the time multiplied by an energy difference is
equal to a multiple of 2π. Other issues arise with finite-size effects if a disturbance
December 10, 2014
6
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
moves a longer distance than half the distance to the edge of the finite system.
While this approach can provide much insight about the nonequilibrium problem,
it is unlikely that it will ever be able to successfully describe generic systems in the
long-time limit.
Another approach, which involves truncating the Hilbert space, is a configuration
interaction-based approach, which is widely used in chemistry and atomic physics,
but is not so widely used for lattice systems. In this class of methods, the wavefunction is described by the sum of a finite number of Slater determinants, with the
coefficients and the determinants adjusted to produce the best variational energy.
These techniques are also used for time-dependent problems, primarily in atomic
physics26 where the wavefunction is represented as a time-dependent function with
an expansion basis that includes a description of the active excited electron and
additional single-pair particle-hole excitations or uses a so-called restricted active
Hilbert space27 . This creates a finite expansion for time-dependent coefficients of
the wavefunction, which are then solved by recasting the problem as a differential
equation for the different coefficients and solving those differential equations. In
practice, many tricks need to be implemented to solve these problems accurately
and efficiently. Within lattice problems, there have been some attempts to try to
adopt these approaches in nonequilibrium. One approach is called the multiconfiguration time-dependent Hartree impurity solver28 , which employs a tensor-based
structure for the time-dependent coefficients of a wavefunction that is expanded in
terms of a time-varying single-particle basis set which is truncated to include just
a small number of single-particle basis functions. While this approach shows some
promise, much more work needs to be done to see if it can be an effective impurity
solver.
Another class of approaches is based on density functional theory and is called
time-dependent density functional theory (TD-DFT). Density functional theory is
based on the Hohenberg-Kohn theorem29 which states that the wavefunction is a
unique functional of the electron density, and is implemented by mapping the system
onto a noninteracting problem via the Kohn-Sham framework30 . TD-DFT has its
origins in the Runge-Gross theorem31 , which applies to time-dependent potentials.
In practice, because density-functional theory cannot be solved exactly, it ends up
being an effective mean-field theory approach to the problem which is highly accurate for real materials calculations in weakly correlated systems. The TD-DFT
needs modification to handle the full time dependence of the coupling to an electromagnetic field, and currently has not been applied to too many problems in this
category. Current implementations also use functionals that typically limit the applicability to finite closed quantum systems. It remains the most accurate approach
for real systems calculations, but is known to fail when correlations become too
strong, and hence will also not be able to solve the most general nonequilibrium
many-body problems. A recent review32 , provides a nice introduction to this topic.
We will have more to say about effective mean-field theory descriptions below.
The final method in this approximate approach is the Gutzwiller approach33,34 ,
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
7
which uses a variational form of the wavefunction for strong correlations, and
projects out the double-occupancies in the wavefunction. This ansatz is then used as
a constrained wavefunction to approximately solve the time-dependent Schrödinger
equation via a minimization of the Schrödinger equation when taken as an expectation value with respect to the time-dependent wavefunction. This approach has been
applied primarily to quench problems, where the Hamiltonian during the evolution
of the system is time independent, the initial state is just not an eigenstate of this
Hamiltonian, so it evolves in time. This technique is an uncontrolled approximation,
since it is not a formal expansion in powers of a small parameter, but it does appear
to produce physically reasonable results for the cases it has been applied to.
In all of these approaches, the many-body problem has been simplified in some
fashion, allowing for the subsequent dynamics of the simplified problem to be computed exactly. These approaches produce physically reasonable results, and often
are quite accurate for short times, but they also are difficult to gauge their accuracy
unless they can be compared to more exact methods. In general, they are not methods that can be easily extended to treat the problem with systematically improving
approximations, so that it can eventually be solved in the general case. However,
they often are a good first approach to these problems.
3. Algorithms that use stochastic methods
We next turn to stochastic methods that solve the many-body problem via Monte
Carlo techniques. We begin by discussing how these algorithms work in equilibrium,
before discussing issues that arise for nonequilibrium algorithms. We also discuss
only a few categories of these methods due to space limitations. Our focus will be
primarily with many-body approaches for the dynamical mean-field theory technique, which provides an exact solution to the many-body problem in the limit of
large spatial dimensions. If one were to look at electrons moving on a lattice in real
space, and were to focus on what happens at a specific lattice site as a function of
time, one would see electrons hopping onto and off of the site in time. The dynamical mean-field theory approach maps this lattice problem onto a single-site problem
with a time-dependent field that represents the hopping onto and off of the site.
By solving the impurity problem in a self-consistently determined time-dependent
field that makes the impurity Green’s function identical to the local lattice Green’s
function, the problem is solved exactly in the large dimensional limit (the reason
why it is not exact in finite dimensions is that the self-energy for the impurity
problem has no momentum dependence, and hence is an approximate solution in
finite dimensions, but becomes exact in infinite dimensions because one can prove
the self-energy has no momentum dependence in that limit).
Impurity problems have been studied in physics for years. The single impurity
Anderson model35 is the pardigmatic impurity problem, and it corresponds to a
localized electron, which has a strong Coulomb interaction with itself, embedded in
a conduction electron host, that is noninteracting (in the strong-interaction limit,
December 10, 2014
8
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
this model reduces to the so-called Kondo problem36 ). These impurity problems
are best represented by an action, which takes into account the time-dependence
of the field that describes the hopping onto and off of the lattice site. Due to the
interaction of the impurity electron with itself, there is a four-fermion operator in
the action. To illustrate how this representation works, we must first start with the
definition of the contour-ordered Green’s function on the lattice, which is defined
via
Gcijσ (t, t0 ) = −iTrTc e−βH(t̄→−∞) ciσ (t)c†jσ (t0 )/Z
(2)
where the fermionic destruction operator on site i with spin σ ciσ and the fermionic
creation operator on site j with spin σ c†jσ are both written in the Heisenberg
representation with respect to the time-dependent Hamiltonian H(t), β = 1/T is
the inverse temperature of the initial equilibrium configuration of the system, and
the partition function is Z = Tr exp[−βH(t̄ → −∞)]. The time-ordering operator
orders times along the contour, which will be described next.
The
Heisenberg
representation
of
an
operator
satisfies ciσ (t) = U † (t, −∞)ciσ U (t, −∞). Substituting this into the definition of the
Green’s function and assuming t > t0 yields
Gcijσ (t, t0 ) = −iTre−βH(t̄→−∞) U (−∞, t)ciσ U (t, t0 )c†jσ U (t0 , −∞)/Z,
(3)
where we note that we used the fact that U † (t, −∞) = U −1 (t, −∞) = U (−∞, t)
and U (t, t1 )U (t1 , t0 ) = U (t, t0 ). Now, starting from the earliest time as t̄ = −∞,
which corresponds to the right most part of the above expression, we evolve the
system forward in time from −∞ to t0 , then we operate with a c†jσ , evolve further
to time t, operate with ciσ and then evolve backwards in time from t to −∞ before
operating with the density matrix operator. This can be summarized graphically by
evaluating the operator average on the so-called Kadanoff-Baym-Keldysh contour
(see Fig. 1), to which we have added an imaginary spur from −∞ to −∞ − iβ, and
the density matrix term can be thought of as
" Z
#
−∞−iβ
e−βH(t̄→−∞) = exp −i
dt̄H(t̄) ,
(4)
−∞
so that the evolution operator extends along the entire contour. We originally
thought of t and t0 both being real times, but now that we have the evolution operator evolving on the contour, we can generalize the definition to allow the times
t and t0 to be any two times chosen on the contour itself, and the formula for the
Green’s function can be written compactly as
Z
c
0
Gijσ (t, t ) = −iTrTc exp −i dt̄H(t̄) ciσ;t c†jσ;t0 /Z
(5)
c
where the contour now extends from −∞ to ∞ along the upper real branch, from ∞
to −∞ along the lower real branch, and from −∞ to −∞ − iβ along the imaginary
spur. The integral is over the full contour, and the subscript t and t0 indicate where
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
9
Fig. 1. The Kadanoff-Baym-Keldysh contour running from tmin to tmax along the real axis, and
then back from tmax to tmin along the real axis, and finally along the negative imaginary axis a
length −iβ. We take the limit tmin → −∞ and tmax → ∞. The system starts in equilibrium at
a temperature T = 1/β at time tmin and then is driven out of equilibrium via a quench, or an
applied field, or some other perturbation.
the operators act along the contour, with the time-ordering operator along the
contour distributing the (time-ordered) exponential factors in the right locations
between the operators.
In dynamical mean-field theory, we will equate the local Green’s function (i = j)
to the impurity Green’s function. The impurity Green’s function depends on the
dynamical mean field, which is often described by the symbol λσ (t, t0 ) with either
time on the contour. Since the dynamical mean field determines the effect of the
hopping on the lattice, we define the local part of the Hamiltonian on the lattice via
P
H0i such that the total Hamiltonian on the lattice is equal to H(t) = i H0i +K where
K is the kinetic energy term of the Hamiltonian. The local part of the Hamiltonian
has the same functional form on each lattice site; the i superscript simply denotes
that we construct it from the operators ciσ and c†iσ . We then take the impurity
Hamiltonian to be H0 constructed from the impurity creation and annihilation
operators c†σ and cσ , respectively. All operators for the impurity problem are taken
in the Heisenberg representation with respect to the impurity Hamiltonian. Then
the impurity Green’s function is written as
Gcσ;imp (t, t0 ) = −iTrTc e−βH0 (t̄→−∞) e
P R
R
dt̄0 c†σ (t̄)λσ (t̄,t̄0 )cσ (t̄0 )
cσ (t)c†σ (t0 )/Zimp
(6)
R 0 †
P R
0
0
where Zimp = TrTc exp[−βH0 (t̄ → ∞)] exp[ σ c dt̄ c dt̄ cσ (t̄)λσ (t̄, t̄ )cσ (t̄ )], which
can be restricted to integrating only over the imaginary part of the contour for the
partition function because the integration over the real parts cancels.
The dynamical mean field λσ must be determined self-consistently in the dynamical mean-field theory algorithm (see Fig. 2). The procedure is as follows: (1)
make a guess for the local self-energy; (2) use Dyson’s equation in momentum space
on the lattice to determine the momentum-dependent Green’s function from the
self-energy; (3) sum the momentum dependent Green’s function over all momenta
to determine the local Green’s function of the lattice; (4) use Dyson’s equation for
the local Green’s function to extract the dynamical mean field; (5) solve the impurity problem in the dynamical mean field for the impurity Green’s function; (6) use
σ
c
dt̄
c
December 10, 2014
10
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
Fig. 2. Schematic of the iterative loop for dynamical mean-field theory. We start with a guess for
the self-energy and then use Dyson’s equation on the lattice to calculate the local Green’s function
(by summing over all momenta). Dyson’s equation is used on the local Green’s function to extract
the effective medium (which depends on the dynamical mean field). The impurity problem is then
solved with this effective medium (or λ field). Dyson’s equation for the impurity is used to extract
the self-energy and one checks to see if the iterative loop has converged. If it has, the computation
ends. If not, the new self-energy is used in the loop and the process continues.
Dyson’s equation to extract the impurity self-energy from the Green’s function and
the dynamical mean field; (7) use the impurity self-energy as the new guess for step
(1) and iterate steps (2–7) until one reaches a fixed point and the Green’s functions
do not change from one iteration to the next. All of the steps of this algorithm are
straightforward to implement (even if we have not given the detailed equations)
except for step (5) which is what requires a sophisticated impurity solver. The iterative approach is the same for equilibrium or nonequilibrium, it is just that in
nonequilibrium one usually works entirely in a time-based formalism37 . Note that
if one is solving an equilibrium problem, one usually restricts the contour to just
the imaginary spur, and then uses an analytic continuation method like maximum
entropy38 to numerically analytically continue the results from the imaginary axis to
the real axis. This simplifies the impurity problem solver, but requires an ill-posed
and uncontrolled numerical analytic continuation procedure. This is often viewed
as a superior way of performing the calculation than simulating the problem on
the full contour, for reasons that will be described below, which are related to the
fermion sign (or phase) problem.
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
11
We next focus on two different algorithms that solve the impurity problem on
the imaginary spur of the contour—the Hirsch-Fye algorithm39 and the continuoustime Monte Carlo algorithm40,41 . To begin, we show how one evaluates the partition
function on the imaginary-time spur of the contour within the Hirsch-Fye algorithm
Zimp = TrTτ e−βH0 e−
P Rβ
σ
0
dτ
Rβ
0
dτ 0 c†σ (τ )λσ (τ,τ 0 )cσ (τ )
.
(7)
The imaginary time axis is discretized into N time steps of length ∆τ = β/N , with
tj = −∞ − i∆τ j for 0 ≤ j ≤ N , which we write as −∞ − iτj . Then we have
Zimp = TrTτ e−∆τ
PN
Hsing
(τj ) −∆τ
0
e
PN
V(τj ) −(∆τ )2
e
P PN
c†σ (τi )λσ (τi ,τj )cσ (τj )
,
(8)
where H0sing is the quadratic (single-particle) part of the impurity Hamiltonian
P
(which we will take to be H0sing (τ ) = −µ σ c†σ cσ for concreteness) and V =
H0 − H0sing is the interacting part of the Hamiltonian [which we take to be
V(τ ) = U (c†↑ c↑ − 1/2)(c†↓ c↓ − 1/2) for the Hubbard model, for concreteness]. Note
that we have included an artificial time dependence to the impurity Hamiltonian
even though it is a constant, independent of time, which is notationally convenient.
The challenge with evaluating this partition function comes from the fact that the
interaction V is not a quadratic function of the fermionic operators. It can be made
quadratic by introducing an auxiliary field via the Hirsch-Hubbard-Stratanovich
transformation42,43,44
†
†
e−∆τ U/4 X αsj (c†↑ c↑ −c†↓ c↓ )
e
,
(9)
e−V(τj ) = e−∆τ U (c↑ c↑ −1/2)(c↓ c↓ −1/2) =
2
s =±1
j=0
j=0
σ
i,j=0
j
where cosh α = exp[∆τ U/2]. This introduces a set of Ising variables at each τj time
step which converts the quartic interaction term into a quadratic term. One can
now evaluate the partition function using standard techniques that convert into a
Grassmann path integral and then evaluate the trace with coherent states to yield
X
−1
Zimp =
det G−1
(10)
↑{si } det G↓{si }
{si }
where
2
G−1
σ{si } (j, k) = δjk − δj−1k [1 + ∆τ µ + αsj σ] − (∆τ ) λσ (τj , τk ).
(11)
So, the evaluation of this sum over all of the Ising spins will give the partition
function. It turns out that averaging the Green’s function G using the weights
for the partition function, will yield the impurity Green’s function. The Monte
Carlo method starts with a configuration for the Ising spins, and proposes the
change of the Ising spin at a particular time slice. One calculates the ratio of the
determinants for the previous G and the current G, using a Metropolis algorithm45
to decide whether or not to accept the update move. If it is accepted, the Green’s
function and determinants need to be updated. One of the hallmarks of the HirschFye algorithm is that the updating of the Green’s function and of the determinant
is very fast, because only a few matrix elements are changed from the previous
December 10, 2014
12
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
configuration matrix. Hence, because one need not recalculate the full determinant
from scratch, nor recalculate the new contribution to Green’s function (for the new
Ising configuration) from scratch, the algorithm can work fast to sum over the Ising
fields via importance sampling.
Since it is a statistical algorithm, it must deal with probabilities that satisfy
a detailed balance procedure, so that every configuration is equally likely to be
reached by the algorithm updating procedure. This requires (among other things)
that the products of the determinants be nonnegative. Unfortunately, although they
are manifestly real, they often change sign as the system is updated. One uses the
absolute value for the weights, but one then must keep track of the average sign
during the updating procedure and divide all statistical averages by the average
sign. When this average sign becomes too small in magnitude, the algorithm breaks
down. This is the origin of the famous sign problem in quantum Monte Carlo, and it
limits how low in temperature one can run the simulation because the sign problem
becomes worse as the temperature is lowered. In addition, since one is using a fixed
grid in imaginary time, it turns out one needs to make ∆τ small at low temperature
as well, because the Green’s function has a large slope near τ = 0 and τ = β and
one needs to determine it accurately in this region. This also limits how low in
temperature one can proceed with the Hirsch-Fye algorithm, since a small ∆τ and
a large β implies the size of the matrices is large.
We next discuss the continuous time quantum Monte Carlo algorithm in equilibrium and based on a weak-coupling expansion40 . The starting point for the
continuous-time algorithm is the formula for the impurity partition function, written
in the interaction representation with respect to H0sing :
sing
Zimp = TrTτ e−βH0
e−
Rβ
0
dτ VI (τ ) −
e
P Rβ
σ
0
dτ
Rβ
0
dτ 0 c†σ (τ )λσ (τ,τ 0 )cσ (τ 0 )
(12)
where the time ordering properly orders both of the time-dependent exponential factors. Here, the interaction picture interaction satisfies VI (τ ) =
exp[τ H0sing ]V exp[−τ H0sing ] and similarly for the fermionic creation and annihilation operators. We assume the concrete forms for the single-particle piece of the
impurity Hamiltonian and for the interaction piece as we did before, to describe the
Hubbard model. The next step in the derivation is to expand the exponential of the
interaction in a power series
Z
Z β
∞
Rβ
X
P Rβ
sing
0 †
0
0
(−U )k β
dτ1 . . .
dτk TrTτ e−βH0 e− σ 0 dτ 0 dτ cσ (τ )λσ (τ,τ )cσ (τ )
Zimp =
k!
0
0
×
k=0
c†↑ (τ1 )c↑ (τ1 ) . . . c†↑ (τk )c↑ (τk )c†↓ (τ1 )c↓ (τ1 ) . . . c†↓ (τk )c↓ (τk ).
(13)
Because this average over the product of the fermionic creation and annihilation
operators is with respect to a quadratic action (given by H0sing and the λ-field
term), one can evaluate the average via Wick’s theorem, which expresses the result
in terms of products of the Green’s function corresponding to the quadratic action.
In total, there are k! different ways to form these products, but they can all be
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
13
Fig. 3. (Color online) Example of one diagram in imaginary time for the equilibrium implementation of the continuous time algorithm from weak coupling. The narrow lines are Green’s functions
connecting different time points (red for spin up and blue for spin down). The Ising variable
associated with the auxiliary field at the interaction vertex has been suppressed.
summarized via the determinant of an appropriate matrix. Note that here we have
0
0
0
G−1
0σ (τ, τ ) = (−∂τ + µ)δ(τ − τ ) − λσ (τ, τ ).
We define the determinant Dσ (k) via
G0σ (τ1 , τ1 ) G0σ (τ1 , τ2 ) · · ·
G (τ , τ ) G (τ , τ ) · · ·
0σ 2 2
0σ 2 1
G (τ , τ ) G (τ , τ ) · · ·
0σ 3 1
0σ 3 2
Dσ (k) = Det
·
·
·
·
·
·
·
·
·
G0σ (τk , τ1 ) G0σ (τk , τ2 ) · · ·
G0σ (τ1 , τk )
G0σ (τ2 , τk )
G0σ (τ3 , τk )
·
·
·
G0σ (τk , τk )
(14)
(15)
with the time arguments given by the k different time values that will be taken
in the kth multiple integral (see Fig. 3 for an example diagram). The stochastic
aspect of the problem is now clear. One uses Monte Carlo methods to sample the
different τi values for each multidimensional integral, and since the error decreases
as the number of random points chosen in the integrand independent of dimension,
one can, in principle, evaluate each order of the perturbation theory expansion with
an equivalent amount of computer time. This, of course, is not the case since the
integrand involves the determinants, and those determinants are more difficult to
evaluate for larger k. In an actual implementation of the algorithm, the Monte
Carlo step allows for the system to increase or decrease k by one unit (adding a
vertex or removing a vertex), and the acceptance criteria for the Monte Carlo step
comes from evaluating the ratio of the determinants for the two sets of diagrams
evaluated. Fast updating methods are used to calculate the new determinants from
the old ones without having to diagonalize the full matrix, which is what makes
the algorithm efficient. The algorithm will actually determine the most important
orders of the perturbation theory via the importance sampling.
Once again, there is a sign problem that occurs if the determinant changes
sign, which often happens. To handle this, the interaction is modified from the
December 10, 2014
14
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
conventional Hubbard interaction to
1
1
U X †
1
1
†
†
†
U c↑ c↑ −
c↓ c↓ −
=
c c − + sα
c↓ c↓ − − sα ,
2
2
2 s=± ↑ ↑ 2
2
(16)
where α ≥ 1/2 seems to fix the sign problem for the equilibrium calculations. Of
course, this does require the additional summation over Ising variables at each time
step for a given diagram (and a rederivation of the above formalism using the new
interaction). The summation is also done via stochastic importance sampling.
Calculating the Green’s function is a bit more involved for the continuous-time
algorithm. The formula for the Green’s function includes two more creation and
annihilation operators than the partition function. This requires one more Wick
contraction for each diagram of “order” k than used in the partition-function calculation. Otherwise, the technique is identical. By manipulating the Dyson equation,
one can show that the term that is actually summed in the algorithm is a reducible
self-energy, which can readily be employed to determine the full Green’s function.
One might ask about the relationship between these two methods. It turns out
that they have a simple relationship46 . If one were to restrict the continuous time
algorithm to include only diagrams of order N = β/∆τ and allow the vertices to be
restricted to the time slices used in the Hirsch-Fye algorithm, then the two methods
calculate the same set of diagrams. By extrapolating to the N → ∞ limit, one can
then show that the extrapolated results for these two methods will be identical,
because the continuous time method will require diagrams of effectively one order
only, due to the central limit theorem. It is also clear, then, that the continuous time
algorithm should be superior to the Hirsch-Fye algorithm for truncated calculations,
because it effectively samples a wider range of diagrammatic terms. The continuoustime approach has now become the standard technique used by most dynamical
mean-field theory calculations in equilibrium.
There is a second form for the continuous-time algorithm, which performs the
expansion of the perturbation series from a strong-coupling approach called the
hybridization expansion41 . We do not have the space here to describe the derivation
of this approach. While similar to the weak-coupling technique, the derivation is
more complicated because the atomic Hamiltonian is not quadratic, and hence one
cannot use Wick’s theorem, but instead one has to employ some more complex
techniques.
Since the continuous-time algorithm is a diagrammatic method, which sums
diagrams to high order, one might wonder whether or not there might be a better
way to organize the diagrams, so that an infinite class of diagrams can be summed
initially, and the Monte Carlo is used to sum the remaining ones. This idea is called
the bold quantum Monte Carlo method47,48 . The idea behind the bold method,
is that if one can identify the dominant terms in the summation that provide the
physical behavior of the system, then summing them up front can greatly reduce
the order that one needs to extend the quantum Monte Carlo summation. There
are additional challenges to employing the bold technique, as one now needs to
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
15
identify and remove terms in the diagrammatic summation that already appear in
the infinite summation of the starting point. This can make the diagram selection
and the Monte Carlo integration more complicated, but this is usually more than
compensated by requiring a much lower order in terms of the number of diagrams
to keep in the expansion. Usually the latter is more important than the former,
especially in cases where there is a sign problem to deal with.
Finally, we need to discuss how one generalizes these approaches to nonequilibrium. In principle, the procedure is straightforward, where we simply extend the
contour along the real axis and proceed as before. The main problem with this is
that the matrix elements for the Green’s functions now become manifestly complex,
and the weight factors also become complex. One can once again use the modulus of
the complex numbers for the weights, but then one needs to sum the average complex phase of the summations and divide by the average phase. The average phase
rapidly goes to zero as the real branch of the contour increases in size, and this
poses the most severe limitation on these methods. The first attempt in this area
was to generalize the Hirsch-Fye algorithm along these lines49 . The continuous-time
algorithm has been tried50,51 , and it can only be stabilized for quite short times,
where only specialized problems can be studied that have their transient evolution
take place over a short period of time. The bold technique has also been applied52
and shows promise, being able to be evolved about twice as long as the conventional algorithm. But, generically, more stable algorithms that can evolve further
in time are needed within this class of impurity solvers. It is possible that making
the algorithm manifestly unitary can help with the phase problem.
The weak convergence or divergence of a diagrammatic series is usually handled
by techniques like Borel, Cesáro-Riesz, or Lindelöf resummation53 . A more general
approach to doing this is offered by resurgence theory54 , which expresses f (x) as
a triple power series (called a trans-series) in x, exp[−1/x], and ln[x], employs
analytic properties in the vicinity of all poles and branch cuts to determine the
relevant coefficients of the expansion, and automatically includes nonperturbative
effects. It is possible that employing this method within a bold-like quantum Monte
Carlo approach could allow calculations to proceed to even longer times55 . We
discuss these techniques in more detail below when we examine extrapolation of
nonequilibrium results to longer times.
There also have been some recent proposals for extensions of variational quantum
Monte Carlo approaches to nonequilibrium. So far, this has been done only for
bosonic systems56,57 , but there is promise that one might be able to extend these
ideas to fermionic systems, and at the very least, the bosonic methods might benefit
from being re-examined along the extreme data science perspective.
4. Algorithms based on renormalization group ideas
The next class of algorithms we discuss are those related to the renormalization
group of Wilson58 . The original numerical renormalization group was designed to
December 10, 2014
16
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
treat the low-energy scale that develops in a system of noninteracting conduction
electrons (the bath electrons) that hybridize with an impurity atom that has strong
Coulomb interactions for electrons on the impurity (the impurity electrons). It was
well known that this model had complex behavior with a low energy scale that
develops in the system at low temperature and that the low-energy model displays
universal behavior. Numerically calculating such results turned out to be problematic. Wilson’s idea was to discretize the frequency space into a logarithmic energy
grid, and replace the continuum of states within each energy interval by one discrete degree of freedom. As one moves in energy closer to the Fermi energy, the
energy intervals become smaller and smaller. Hence, the Hamiltonian for the single
impurity Anderson model is described by
HSIAM = Ef
X
fσ† fσ +U f↑† f↑ f↓† f↓ +
σ
N X
X
N X
X
Vα fσ† cασ + c†ασ fσ +
ωα c†ασ cασ
α=0 σ
α=0 σ
(17)
where 0 ≤ α ≤ N labels the bath states that have energy which we choose for
illustration to satisfy ωα = 1/Λα (the parameter Λ is commonly chosen between
1.5 and 3), Vα is the hybridization between the impurity and the αth bath state,
U is the Coulomb interaction, and Ef is the energy of the impurity level. In this
form, the Hamiltonian is written as an impurity coupled to each bath site, which
is schematically illustrated in Fig. 4(a). One can integrate out the conduction electrons, which produces an impurity problem interacting with a time-dependent field
which is our old λ field, but written in real frequency instead of time. One finds
λσ (ω) =
N
X
|Vα |2
.
ω − ωα + i0+
α=0
(18)
The negative of the imaginary part of this field is called the hybridization function
∆σ (ω) = −Imλσ (ω). Given a dynamical mean-field λσ in frequency space, one
can immediately construct the hybridization and energy for each bath state that
approximates the imaginary part of the dynamical mean-field as a sum over a set
of delta functions. The procedure is as follows: for the αth energy interval, which
runs from ωαmin to ωαmax , the frequency ωα is chosen to be the weighted average of
the hybridization function over the interval
Z ωαmax
1
ωα = max
dωω∆(ω)
(19)
ωα − ωαmin ωαmin
and the weight of the delta function that describes the effective discrete degree of
freedom yields the hybridization
s
Z ωαmax
1
dω∆(ω),
(20)
Vα =
ωαmax − ωαmin ωαmin
which is always real, because the hybridization function is always nonnegative [see
Fig. 4(b) for a schematic of the process].
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
17
Fig. 4. (Color online) In this schematic of the single impurity Anderson model, we take eight bath
states indicated by the circles which couple to the central impurity site indicated by the square,
where the interaction takes place [panel (a)]. In addition, we show the traditional logarithmic
grid in frequency space (upper blue) and how one determines the weight and frequency associated
with each bath site (lower green) for the dynamical mean field of a Hubbard model with strong
interaction and a filling of 0.9 [panel (b)]. The black curve is the original hybridization ∆(ω) =
−Imλ(ω) and the red curve is the approximation using Λ = 1.4 and 25 points. The red curve
is a reconstruction of the dynamical mean field by using a logarithmic broadening of the delta
functions. Note how it fits poorly for large frequency but is much more accurate near ω = 0. Note
also that there is a large deviation at the smallest frequencies because there are no discretized
states there. Increasing the number of states would fix that disagreement.
Wilson’s next step is to construct the Wilson chain. This is done by starting
with the impurity Hamiltonian written in terms of the f operators only, and then
constructing the linear combination of bath states that directly couple to the impurity
v
u N
N
uX
X
|Vα0 |2
(21)
a1σ =
Vα cασ /t
α=0
α0 =0
where the constant term in the denominator is chosen to guarantee that the a
operator satisfies the canonical fermionic anticommutation relations. We now want
to define a new set of chain fermions (with a0σ = fσ ) and construct the Wilson
chain Hamiltonian (see Fig. 5), which takes the form
chain
HSIAM
= Ef
X
a†0σ a0σ + U a†0↑ a0↑ a†0↓ a0↓ +
σ
+
∞ X
X
i=1
∞ X
X
i=1
ti a†i−1σ aiσ + a†iσ ai−1σ (22)
σ
i a†iσ aiσ
σ
where each term in the Wilson chain is determined recursively by forcing the original
single-impurity Anderson model into the above form. The Coulomb interaction U
appears only on the zeroth lattice site, and the hopping matrix elements decay
December 10, 2014
18
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
Fig. 5. (Color online) The Wilson chain is constructed similar to a Lanczos procedure, where
the impurity is coupled to a linear combination of bath sites, which we denote as the first chain
site. The chain site is then coupled in turn to new chain sites, with the coupling between chains
decreasing down the chain (indicated by the shrinking widths of the lines denoting the hopping).
exponentially as one moves down the chain. One can view this transformation as
a tridiagonalization of the SIAM Hamiltonian in the same way that one writes the
Lanczos algorithm.
Wilson’s numerical renormalization group works with the problem in the Wilson chain form. One starts with the impurity site and then adds one chain site and
another, and so on, until the size of the matrix is just bigger than a predetermined
size m (typically m ranges from 2000 to 5000) which is going to determine the number of kept states. One then diagonalizes the full Hamiltonian, and truncates the
basis, keeping the m lowest energy eigenvalues. The process is then iterated further,
with another chain site added, the Hamiltonian constructed for the enlarged chain,
then diagonalized, and finally truncated to the low-energy space. As this process
is iterated, eventually it stops at a fixed point, where the system is described by
the renormalized Fermi liquid. By keeping account of the states discarded at each
step of the iteration, one has a complete set of states for the problem that describes
the high-energy degrees of freedom (which become frozen in at high temperature)
and the low-energy degrees of freedom, which correspond to the active m states
on the current Wilson chain. Since one has both eigenvalues and eigenvectors, one
can compute many different quantities of interest, like the single-particle Green’s
function (via the Lehmann representation), the self-energy, magnetic and charge
susceptibilities, etc. Spectral properties are also recorded as a series of delta functions which need to be broadened to obtain a spectral function. The broadening
scheme needs to be chosen for a given problem, but usually one chooses a logarithmic scaled Lorentzian or Gaussian. This choice does well at low frequencies, but
does not properly produce the bandwidth of the system. Since one can input the
hybridization function, and output the self-energy, this method can be used to solve
dynamical mean-field theory, although care needs to be taken in how this is done
to ensure the iterated equations will converge.
One subtle point to keep in mind is that the effective temperature of the system is
determined by the energy scale of the chain where the calculation ends. Hence, to get
to the zero-temperature limit, one performs the calculation on a chain that is long
enough that the results stop changing. Higher temperatures are found on smaller
chains, and it should become clear that this technique fails if the temperature is
too high, because there are not enough degrees of freedom to properly describe the
behavior (in fact, for dynamical mean-field theory, it also fails at low temperature,
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
19
Fig. 6. (Color online) (a) Particle in a box problem which motivated the density matrix renormalization group approach. The ground state wavefunction in the smaller boxes (black curves)
do not allow for a simple way to construct the ground state wavefunction of the larger box (red
curve), although they do have some weight in that construction. (b) Schematic of the original
block of size L, the additional site, the final block of size L + 1 and the reflection of the system to
construct a superblock.
as the imaginary part of the self-energy often has the wrong sign for frequencies
near the Fermi energy at low temperature, below the low-energy coherence scale).
While the numerical renormalization group approach has been extremely successful in solving some of the paradigmatic many-body problems, it turned out to
be limited in scope, in that similar ideas applied to strongly correlated lattice problems failed to work effectively. The reason for this was twofold: (i) it isn’t necessarily
true that the low-energy eigenfunctions for a smaller system are the most important
ones to keep for finding the low-energy eigenfunctions of a larger system and (ii)
the choice for keeping the lowest energy eigenvalues turns out to significantly lose
information about the quantum state as the system size grows. In the 1990’s, White
developed an alternative formulation to the renormalization group, which he called
density matrix renormalization group, that solved both of these problems and allowed a wide range of one-dimensional strongly correlated problems to be solved59 .
A quantum information analysis of this algorithm shows that it works well as long
as the entanglement entropy does not grow too rapidly. This usually holds in onedimensional systems, and rarely does for higher dimensional ones. We describe the
density matrix renormalization group algorithm next.
The motivation for the density matrix renormalization group problem was the
challenge of finding a way to apply the renormalization group ideas to the case of
December 10, 2014
20
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
expanding the size of a one-dimensional quantum system and maintain an appropriate basis that projects primarily onto the low-energy states of the larger system.
Consider a particle in a box of length L in the ground state. We want to double the
size of the system to 2L and find the states in the low-energy manifold. One might
think that they should consist of the states that have low energy in the smaller box,
but a simple calculation shows that smaller box ground state projects onto states of
all energies in the larger box, and hence choosing states to keep by minimizing the
energy is not the correct thing to do [see Fig. 6(a)]. In fact, the entire renormalization group procedure needs to be modified. White chose to increase the system of
size L to one of size L + 1 through a complicated procedure involving a superblock
of size 2L + 2. One starts with the block of size L, reflects it to another block of
size L and adds two sites in the middle. If we let α and β denote the states in the
left and right blocks of size L + 1, then the wavefunction of the superblock is given
by the matrix-product state expression
X
|ψi2L+2 =
ψαβ |αi ⊗ |βi
(23)
αβ
where ψαβ are the expansion coefficients (numbers) and |αi and |βi are the states
that form the basis in the left and right block [see Fig. 6(b)]. Using a singular value
decomposition, one can transform this matrix into its diagonal form
X
λᾱ |ᾱileft ⊗ |ᾱiright
(24)
|ψi2L+2 =
ᾱ
where the barred symbols are the rotated basis, and the same states are coupled
together in the direct products for the left and right subblocks. The reduced density
matrix for the left block is then defined to be
X
X
ρlef t =
|ψihψ| =
λ2ᾱ |ᾱileft left hᾱ|.
(25)
ᾱright
ᾱ
The density-matrix renormalization group algorithm says that one takes the m
largest eigenvectors of the reduced density matrix as the truncated basis for the
left block of size L + 1. So the procedure is to start with a block of size L that has
already been written in the effective basis according to the maximal eigenvalues of
the reduced density matrix. Construct a superblock of size 2L + 2 and diagonalize
the Hamiltonian to give the ground state via a Davidson or Lanczos procedure, take
the matrix that expresses the ground state wavefunction in the original basis and
diagonalize via a singular value decomposition and transform the matrix to this
new basis. Construct the reduced density matrix by tracing over the right degrees
of freedom and pick the m states with the largest eigenvalues for the basis states of
the L + 1 block. Proceeding in this fashion, one can keep growing the system until
the energy stops changing. This then gives the results for the ground state of the
equilibrium and infinite-sized system. One can examine multiple low energy states
either by separately calculating for the different individual states, or by averaging
the density matrix over the different eigenstates one is computing.
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
21
There also is a finite-sized system algorithm that also creates superblocks, but
the size of the system is fixed at 2L say, so one increases the left block while
decreasing the right and vice versa, zipping through the system to the left and then
to the right. This approach is usually more accurate for determining the energy,
especially if one uses finite size scaling to the large system limit, than the infinite
system algorithm.
The density-matrix renormalization group algorithm is a variational algorithm,
and the choice of using the reduced density matrix rather than the energy eigenvalues in choosing the appropriate basis to describe the different blocks turns out
to be the choice that minimizes the information loss in the system as one increases
the system size, and is precisely what is needed to produce an accurate and efficient algorithm. In a more modern language, we like to think of the density matrix
renormalization group ansatz for the wavefunction as part of a general matrix product state ansatz60 . Then, quantum information theory tells us that this approach
is a form of a quantum compression algorithm61,62,63 and one can quantify the
computational cost and the accuracy via studying the associated entanglement and
entanglement entropy. Recent developments have also shown how the density-matrix
renormalization group can be employed to benchmark density functional theory in
one dimension64 .
So why does the approach fail for higher dimensions? It turns out that the representation of the ground state in terms of the direct products with a truncated
basis is an efficient way to expand wavefunctions that have limited entanglement.
As entanglement grows in the system, more complex ways to represent the wavefunction are needed, and generalizations from the matrix product state approach of
the density matrix renormalization group to projected entangled pair states65 , and
to multi-scale entanglement renormalization ansatzs66 , show promise in being able
to handle more entanglement in the wavefunction; they have already solved a number of difficult problems67,68,69 . In addition, progress has been made with solving
long-range interactions, which can lead to efficient simulation of higher dimensional
systems70 .
There has been progress on solving time-dependent problems within the numerical renormalization group framework71 , but so far these problems have focused on
quantum quenches, or steady-state properties of current flowing through a quantum
dot attached to leads at two different voltages. The basic idea to carry this out is
that one can calculate the time-dependent expectation value of an operator from
the time dependent density matrix. Under the assumption that the system starts
in a Hamiltonian H0 for early times, and is in equilibrium at some temperature,
then the initial density matrix is just ρ0 = exp[−βH0 ]. At time t = 0 the system is
suddenly switched to a new Hamiltonian H1 , and that Hamiltonian is responsible
for all future time evolution of the system. The time-dependent density matrix then
satisfies
ρ(t) = eiH1 t ρ0 e−iH1 t .
(26)
December 10, 2014
22
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
So the Wilson chain is constructed from the final Hamiltonian, and the evolution of
the system is found in the long-time limit by averaging the time evolution, which
corresponds to a dropping of the off-diagonal elements when expressed in an eigenbasis with respect to the final Hamiltonian. For current transport problems, it is
important to formulate the system in terms of the scattering states of the noninteracting system, which is somewhat different from the conventional way of developing
the numerical renormalization group because the impurity operators are part of the
scattering states operators. Generalizations to include multiple quenches have also
been made72 . One of the issues with this approach is that the exponential decay of
the hopping matrix elements projects the system to lower and lower energy shells of
the final Hamiltonian, and it isn’t clear that the nonequilibrium system is projected
onto those low energy states. Hence, a new hybrid approach has been constructed
where the numerical renormalization group approach is initially taken, and then
one ends with a density matrix approach for the time-dependent problem73 . This
then allows for more accuracy at longer times when one does not continue to project
onto lower energy shells.
The nonequilibrium dynamical mean-field theory problem requires a more complex approach to solving it because driving a system with a field, requires a more
general nonequilibrium impurity solver than one that can handle quenches only.
One has to first grapple with the so-called mapping problem, which involves determining the hybridization and the site-energy for each bath state that couples
to the impurity74 . Unlike the construction for the real frequency case, where integration over the individual frequency intervals directly yielded the hybridization
and bath energy, here, because of the more complex time dependence, one cannot
easily determine the hybridization or bath energy. The problem has, however, been
solved in principle, and it requires one to construct two sets of bath states, one
called the + bath and one the − bath. The + bath is constructed from the initial equilibrium mapping and the mixed real and imaginary time dynamical mean
field. There is a unique and well defined strategy to determine these hybridization
functions (which now depend on time). Once they are found, they determine the
hybridization function on the real axis. But since this is most likely not the hybridization function of the nonequilibrium system, we need to subtract this result,
and construct a second hybridization function and a second set of bath functions
from the difference Λ− = λ − Λ+ . Assuming this second dynamical mean field has a
positive representation in terms of hybridization functions, one can then construct
it. One should imagine that the original imaginary axis dynamical mean field is
given by the Λ+ field, and that as time evolves, the initial correlations, given by
the mixed real/imaginary dynamical mean field die off as the real time gets large,
and the system evolves from the Λ+ field to the Λ− field as time moves forward,
as illustrated in Fig. 7. In any case, this approach has a constructive way to determine the effective hybridizations and bath energies, even though it is numerically
challenging to solve these equations.
The problem for developing a complete nonequilibrium numerical renormaliza-
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
23
Fig. 7. (Color online) If we take the contour, shown at the top, and stretch it to a straight line,
we can represent a matrix function that has each time lying on the contour on either the upper
real branch (+), the lower real branch (-), or the imaginary branch (i). The Λ+ field is largest
in the equilibrium part of the matrix, closest to the imaginary spur, and fades away as one gets
farther from that region (indicated by the orange color). The other field Λ− grows the farther one
gets from the spur (indicated by the purple color). For large real times, the field is primarily Λ− .
tion group approach, if that even is possible, has not yet been completed. Since the
driven system need not be restricted to a low energy subspace, nor do high energy
degrees of freedom need to be frozen in at low temperature, the renormalization aspect of the algorithm may not hold anymore. Numerical attempts have been made
recently28,75 , but more work needs to be done to make these approaches more efficient. These ideas provide an interesting starting point for algorithms that focus
on the causal nature of the many-body problem and are described in more detail
below.
We conclude this section be describing the time-dependent density matrix renormalization group approach76,77 . Here, one wants to find the wavefunction as a function of time, and it will change in time, according to how the Hamiltonian evolves,
by simply following the equation of motion. Since we consider a Hamiltonian that
December 10, 2014
24
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
changes as a function of time, a simple Trotterization of the evolution operator is
appropriate, and so the algorithm needs to be modified from calculating the ground
state wavefunction to determining the action of exp[−iH(t)∆t] on the wavefunction
at time t. It turns out this can also be calculated using sparse matrix techniques
based on either the Lanczos algorithm or the Crank-Nicolson algorithm. It is best to
perform these calculations for the finite-system algorithm, because then one can zip
to the left and then to the right for each time step and update the system accordingly. Of course, the set of m kept basis state functions will now change as a function
of time, and they will continue to be determined by the reduced density matrix of
the Hamiltonian at the current time instant. It is not obvious that this produces
the optimal basis to work in, but it is how current calculations are performed. Once
again, if the time evolution of the system creates significant entanglement growth,
or if the disturbance in the system moves to the boundaries where it will reflect,
one can no longer evolve the equations further in time to determine the properties
in the long-time limit. In any case, this algorithm has proved to be one of the best
approaches to the nonequilibrium problem in one dimensional systems, and it is
likely that any improvements to the approach need to adopt a similar time evolution of the active subspace within which the current quantum state lies, as it evolves
forwards in time.
There have been some interesting new developments that incorporate Chebyshev
polynomials for expansions of the many-body density of states in equilibrium78 . The
density of states is found by recursively determining moments of the Chebyshev
expansion, which involves a matrix-vector multiplication to determine the recursion.
The full matrix is truncated by employing a matrix-product-state representation
for the wavefunction used in the recursion relation. It is possible ideas represented
in this work will be useful also for nonequilibrium problems and for deciding on
appropriate compression schemes for representing the data for the wavefunction.
5. Algorithms based on the evaluating the evolution operator
In our final technical section, we discuss the the so-called nonequilibrium Green’s
function approach, which works on effective quadratic Hamiltonians (which are
either noninteracting, or describe mean-field couplings, including low-temperature
superconductivity). Quadratic Hamiltonians have the benefit of being able to be
diagonalized exactly, and hence, a calculation of the evolution operator can be
performed exactly, and from the evolution operator, one can directly find the Green’s
function. (For example, the retarded Green’s function is given by the evolution
operator, while the lesser Green’s function, which depends on the full history of the
system, can be constructed from the evolution operator as well, but is a much more
complicated object.) Hence, these approaches effectively are focused on determining
the evolution operator directly.
The nonequilibrium Green’s function approach79,80 , is used to solve for the
nonequilibrium transport in mesoscopic systems that are connected to ideal (bal-
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
25
listic metal) leads. The leads are maintained at fixed temperatures and voltages,
and by projecting the Dyson equation onto the active interacting part of the device,
where the mesoscopic behavior occurs, one can describe the leads via effective selfenergy contributions determined by the surface Green’s functions of the leads. This
then maps an infinite system onto an effective finite-sized system that can then
be treated numerically. In the case where the many-body dynamics is restricted
to a mean-field treatment, or is described by simple perturbation theory, one can
immediately solve these problems, although the numerical effort can be huge, and
one has to pay particular detail to conservation laws, gauge invariance, electronic
charge reconstruction, and self-consistency in solving both the quantum mechanics
and the electrodynamics of the driven system81,82 .
Recently, there has been a breakthrough on the formal/numerical side by mapping the nonequilibrium Green’s function approach onto a simple wavefunction
based approach83 . If one decomposes the retarded Green’s function into a sum
over products of wavefunctions, and examines the equation of motion for the wavefunctions, one finds they satisfy the time-dependent Schrödinger equation, with the
full time-dependent Hamiltonian, plus an additional source term that arises from
the currents being driven by the leads. Since the Green’s function is composed from
the wavefunctions, essentially as the sum over outer products, one can reduce the
complex matrix-based numerics for solving the Green’s functions by vector-based
numerics for solving the wavefunctions. Here, one of the critical issues one needs
to preserve in solving these problems is to maintain unitarity of the time evolution, which can rapidly be lost in conventional integrators of wavefunctions (or
Green’s functions), but there are algorithms to do this. The main challenge with
this wave-function-based approach, is that it is not easily generalized to full manybody interacting problems, but it does hint at the fact that one can find much more
efficient algorithms for solving these types of problems by determining the correct
way to represent the Green’s functions. In particular, if it is possible to represent
the full many-body problem in terms of a wavefunction-based representation with
a simple Schrödinger-like equation with an additional source term, then one can
extend these wavefunction-based methods much more broadly.
Motivated by the success of this approach, one can consider other methods which
have not been exploited too much to try to tackle the problem of determining the
evolution operator. One method to consider is to factorize the evolution operator
into a sequence of factors which might make it easier to evaluate the full operator
numerically. This approach is well known, as it forms the basis for the interactionpicture representation of the evolution operator, but often the factorization stops
there. In cases where the Hamiltonian does not commute with itself at different
times, but its commutator does (i.e., [H(t), H(t0 )] commutes with H(t00 )), one can
find the exact evolution operator by breaking the Hamiltonian into a series of three
factors, as is often used in solving for the driven harmonic oscillator in quantum
mechanics textbooks84,85 . By employing clever decompositions of the Hamiltonian,
December 10, 2014
26
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
as is done with the density-matrix renormalization group method, one might find
new ways to factorize the evolution operator that can make computation much
simpler with it. We feel this has been an underexplored area for the nonequilibrium
many-body problem.
6. Computer Science ideas for extreme data science
From a computer science perspective, one would like to determine an efficient way to
encode the large complex data set, construct a small database of how that set evolves
as a function of time, and use the current information to forecast future information.
In principle, this should be possible, because the structure of the formula for the
Green’s function is a trace over initial states that have high thermodynamic weight.
At low temperatures, this is restricted to a small set of states, which becomes just the
ground state at T = 0. The evolution of this state through the Hilbert space traces
out a one-dimensional path, and hence it always is a small subspace of the full space.
The problem arises when we try to represent this state in a particular basis. As the
wavefunction becomes more entangled, or involves mixtures of more eigenstates, it
can take a large number of states to represent it. One key element within this is to
employ causality as the unifying principle for how to evolve the system forward in
time. For the quantum many-body problem, this entails determining an equation of
motion and solving it. One can do so by determining the evolution operator, or a
projection of the evolution operator onto the active subspace of the Hilbert space,
or via a direct integration of the wavefunction. To do either of these, requires an
efficient way to encode the evolution operator or the wavefunction, for otherwise,
the representation of these objects will grow too rapidly to be able represent it for
any reasonable amount of time. Since the operation needed at any given time step
is the action of the evolution operator on the particular wavefunction that one is
following, it is likely that one can evolve the wavefunction efficiently.
The simplest method for encoding the basis vectors is to pick a single-particle
basis and encode the state as a binary with a one indicating a particle in that state
and a zero indicating no particle in that state. As the density-matrix renormalization
group studies have shown, clever ways of encoding the wavefunction, via tensor
products or other types of data structures, can prove to be much more efficient ways
of encoding the wavefunction and allowing it to grow in complexity as a function of
time. More research and more ideas of how to create, manipulate, and calculate with
such data structures is imperative to make progress with finding efficient encoding
schemes. It is also likely that one will need to change the bases and let them evolve
as a function of time for more efficient implementations. Clearly, computer science
expertise in appropriate data structures will be helpful in making progress with this
approach.
In addition to novel data structures, appropriate computing architectures and
paradigms are needed to support extreme data science computations. Dataflow and
computing models that display extreme parallelism are among such possibilities to
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
27
be considered86 . Dataflow computing arises from situations where storing the entire
data space is impossible (which encompasses extreme data science by definition),
so the algorithms that direct computation are based on the availability and flow of
the data and are able to employ parallelism at a high level. Developing this type of
computation is the foundational approach to solving extreme data science problems.
Furthermore, computing paradigms such as lazy computation87 and speculative
computing88 must also be considered. Lazy computation delays the actual computation until absolutely necessary. In our focus, such an approach supports the
projection of our global data space to only what is immediately needed and can be
maintained (the time-evolving target data subspace). Speculative computation supports simultaneous multiple path solution space explorations providing the ability
to sustain approximating boundary conditions as needed.
The most numerically challenging scheme will be to determine how to either
solve for the evolution operator of the system, or how to integrate the wavefunction
as a function of time. Ideas along these lines will be presented in the next section.
Forecasting falls into two realms. One is extrapolating the average time to longer
times and the other involves extrapolating the relative time to longer times. It turns
out that the latter is often much easier than the former to accomplish. For equilibrium problems, the time translation invariance of the dynamical mean field, often
implies that the relative time dependence can be determined by the determinant
of a large Toeplitz matrix89 . Using Szego’s theorem and the elegant mathematics
of the Wiener-Hopf approach90,91,92 , one can construct an exact formula for the
asymptotic exponential decay of the Green’s function as a function of relative time.
If the spectral function has power-law singularities, then the approach needs to be
modified to the Fisher-Hartwig conjecture93 , which will determine the power-law
decay of the Green’s function in relative time. It is only in the unlikely case of a
delta function in the spectral function, that one encounters a relative time dependence that does not decay. For the nonequilibrium problem, if sufficient data have
been generated for long enough times, one can extract the appropriate decay forms
for longer times, under the assumption that the decaying behavior as a function of
relative time can be extrapolated from the short-time results. The expectation is
that the Green’s function always decays with a complex exponential behavior for
large relative times.
Extrapolation as a function of average time requires some form of an ansatz for
the behavior of the system at long times. Since the retarded Green’s function determines the quantum states of the system, it is expected to rapidly approach its longtime behavior, while the lesser Green’s function, which determines how those states
are filled, often takes much longer to reach its asymptotic limit. But in cases where
it is described by an effective fluctuation-dissipation theorem with a time-dependent
temperature, the extrapolation can be carried out quite effectively94 . In other cases,
if one knows the analytic behavior of the system, one can extrapolate summations
of diagrams to capture the long-time limits as well95 . Techniques along these lines
December 10, 2014
28
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
have been used within numerical approaches for describing asymptotic expansions,
but much more can be done by enforcing properties of analyticity, where they apply.
They generically fall into the mathematical theory of resurgence54 , which has been
applied to quantum mechanics, quantum field theory and string theory96,97 , and
has great potential to be applied to the nonequilibrium many-body problem.
It is also advantageous to use reduced dynamics techniques, which re-express
long-time dynamics in terms of a short-time memory kernel, which generically decays much faster in time than the Green’s function98 . The short-term kernel can
be determined by computational methods for nonequilibrium that are accurate for
short times, and then can be input into the reduced dynamics formalism and evaluated in the long-time limit. This approach has been successfully applied to a number
of different problems. One example is the application to spin relaxation in the Kondo
regime with the bold continuous time quantum Monte Carlo method99 .
7. Future thoughts and conclusions
We examined the behavior of a number of different algorithms for calculating properties of many body systems in equilibrium and in nonequilibrium. Clearly, the
nonequilibrium problem is much more complicated than the equilibrium one, and
it may require the development of new approaches to be able to reach long times
with a general purpose algorithm. One result which remains clear, is that stochastic
methods, which map the quantum problem onto an effective statistical problem,
will suffer from the dynamical phase problems, which will greatly limit their ability
to perform accurate calculations for long times. This can only be fixed by finding
alternative representations that can forestall the phase problem from occurring. The
bold method shows promise in this area, but still has much development to go before it will be useful for general problems. If we leave these stochastic methods off
the table as we develop novel approaches that will likely work for nonequilibrium,
we are left with focusing on either determining the evolution operator directly, as
a function of time (and perhaps projected onto the active subspace), or evolving
the wavefunction directly in time. These methods are closely related to exact diagonalization techniques and to density matrix renormalization group methods, which
both use this approach. For dynamical mean-field theory applications, we always
need to solve an effective single impurity Anderson model, but we believe that it
must be solved in alternative ways than the conventional time-dependent numerical
renormalization group, because one need not project onto an effective low energy
space for a driven system. We call these time-driven methods, methods based on
causality, since we will utilize the causal nature of the wavefunctions or Green’s
functions in solving the problem. Recent ideas along these lines can be found in
an approach that focuses on the evolution of expectation values in the Heisenberg
picture for the single impurity Anderson model100 .
We want to discuss the application of the four computer science ideas within this
many-body context. First, we have the issue of how to encode the data. Here, the
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
29
density matrix renormalization group is probably the most sophisticated approach,
as it tries to create wavefunctions that can contain a certain amount of entanglement in their representation. Determining the most efficient way to represent the
wavefunction is one of the unsolved goals, but there has been some new work along
these lines in the context of the numerical renormalization group in equilibrium101
which could point in a new direction for how to formulate the solution of this problem. One of the key elements is to focus on how to efficiently update the basis that is
used from one time step to another, and new ideas within this context are certainly
going to be needed.
The evolution of the system boils down to one of two techniques. The first is to
solve for the evolution operator (which can then subsequently be used to find the
Green’s functions). One approach that has not been used too much is to factorize the
evolution operator. While everyone learns the simplest factorization as seen with the
interaction representation, it might be possible to factorize further via a Magnuslike expansion102 , where subsequent factorizations are possible if the potential in
the interaction picture can be broken up into two pieces, whose commutator at two
different times commutes with all operators. The second approach is to directly solve
for the wavefunction as a function of time. This approach, via a direct integration
of the Schrödinger equation, is often the most efficient way to proceed, because
it requires one to operate just the Hamiltonian on the wavefunction at each time
instant, but many algorithms that do this suffer from loss of unitarity, and hence
the algorithm needs to be designed with care to guarantee it will preserve the total
probability.
None of the evolution methods in time will work without having an ability
to truncate the results into an active subspace of the whole Hilbert space. The
density matrix renormalization group uses the reduced density matrix to determine
this optimal basis, while the recent work on the numerical renormalization group
by Gunnarsson and collaborators101 used a simple truncation scheme to trim the
number of basis vectors used by their amplitude to appear in the wavefunction.
Both methods show great promise in finding ways to reduce the active subspace so
that the problem does not grow to be too large as a function of time, and to adapt
the target subspace as time moves forward. Efficient coding of the states will be
needed to determine the full history of the evolution.
Finally, the ideas for forecasting in this problem rely strongly on the analytic
properties of the functions being calculated, and by using as much of these analytic properties as possible, one can, perhaps, evolve these systems much further
in time than one thought they could be evolved simply by extrapolating with the
appropriate analytic functions for the long-time behavior.
The basic ideas of extreme data science, the ability to work with data sets that
are so large they cannot be constructed or stored on any computer, seems to be
impossible to deal with at first glance. By carefully analyzing these problems, especially for the class of problems where the active subspace remains small as the
system evolves in time, it is clear that much progress can be made. The nonequilib-
December 10, 2014
30
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
Freericks, Nikolić and Frieder
rium many-body problem can serve as a unique paradigm where a cross fertilization
of ideas from physics, computer science and applied mathematics can help solve just
such a problem and potentially lay the ground work for how one solves similar classes
of problems in other fields. Doing so could revolutionize the field of big data science
and also the scientific fields that can be successfully analyzed within these contexts.
Acknowledgments
We acknowledge useful conversations with John Arrington, Guy Cohen, Peter
Drummond, Adrian Feiguin, Ushma Kriplani, Hector Mera, Roman Orus, Andreas Stathopoulos, and Xavier Waintal. J. K. F. acknowledges the support of
the Department of Energy Office of Basic Energy Science under grant number
DE-FG02-08ER46542 and the McDevitt bequest at Georgetown University. B. K.
N. acknowledges support of the National Science Foundation under grant number
ECCS-1202069. O. F. acknowledges the McDevitt bequest at Georgetown University. J. K. F. also thanks the Aspen Center for Physics and the National Science
Foundation under Grant Number 1066293 for hospitality during the initial stages
of developing this review.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
E. Candès, J. Romberg, and T. Tao, IEEE Trans. Inform. Theory 52, 489 (2006).
D. Donobo, IEE Trans. Inform. Theory 52, 1289 (2006).
N. Halko, P. G. Martinsson, and J. A. Tropp, SIAM Rev. 53 217 (2011).
M. W. Mahoney and P. Drineas, Proc. Nat. Acad. Sci. USA 106, 697 (2009).
H. Avron, P. Maymounkov, and S. Toledo, SIAM J. Sci. Comp. 32, 1217 (2010).
D. Poulin, A. Qarry, R. Somma, and F. Verstraete, Phys. Rev. Lett. 106, 170501
(2011).
L.-F. Arsenault, A. Lopez-Bezanilla, O. A. von Lilienfeld, and A. J. Millis, Phys. Rev.
B 90, 155136 (2014).
J. F. Corney and P. D.Drummond, Phys. Rev. Lett. 93 260401 (2004).
J. F. Corney and P. D. Drummond, Phys. Rev. B 73, 125112 (2006).
M. Troyer and U.-J. Wiese, Phys. Rev. Lett. 94, 170201 (2005).
P. Corboz, M. Troyer, A. Kleine, I. P. McCilloch, U. Schollwöck, and F. F. Assaad,
Phys. Rev. B 77, 085108 (2008).
M. Ögren, K. V. Kheruntsyan, and J. F. Corney, Comp. Phys. Commun. 182, 1999
(2011).
P. Deuar and P. D. Drummond, Phys. Rev. A 66, 033812 (2002).
A. B. Harris and R. V. Lang, Phys. Rev. 157, 295 (1967).
J. Hubbard, Proc. R. Soc. London Ser. A 276, 238 (1963); 277, 237 (1964); 281, 401
(1964); 285, 542 (1965); 296, 82 (1967); 296, 100 (1967).
L. M. Falicov and R. A. Harris, J. Chem. Phys. 51, 3153 (1969).
O. J. Heilmann and E. H. Lieb, Trans. N. Y. Acad. Sci. 33, 116 (1971).
M. Mierzejewski, L. Vidmar, J. Bonca, and P. Prelovsek, Phys. Rev. Lett. 106, 196401
(2011).
H. Lu, S. Sota, H. Matsueda, J. Bonca, and T. Tohyama, Phys. Rev. Lett. 109 197401
(2012).
E. R. Davidson, J. Comp. Phys. 17, 87 (1975).
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
31
C. Lanczos, J. Res. Nat. Bur. Stand. 45, 255 (1950).
H.-R. Fang and Y. Saad, SIAM J. Sci Comput. 34, A2220 (2012).
H. Trotter, Proc. Amer. Math. Soc. 10, 545 (1959).
A. Alverman and H. Feske, J. Comput. Phys. 230, 5930 (2011); A. Alverman, H.
Feske, and P. B. Littlewood, New J. Phys. 14, 105008 (2012).
J. Crank and P. Nicolson, Proc. Camb. Phil. Soc. 43, 50 (1947).
L. Greenman, P. J. Ho, S. Pabst, E. Kamarchik, D. A. Mazziotti, and R. Santra, Phys.
Rev. A 82, 023406 (2010).
D. Hochstuhl and M. Bonitz, Phys. Rev. A 86, 053424 (2012).
K. Balzer, Z. Li, O. Vendrell, and M. Eckstein, preprint, arxiv:1407.6578 (2014).
P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).
W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
E. Runge and E. K. U. Gross, Phys. Rev. Lett. 52, 997 (1984).
C. A. Ulrich and Z.-H. Yang, Braz. J. Phys. 44, 154 (2014).
M. Schiró and M. Fabrizio, Phys. Rev. Lett. 105, 076401 (2010); M. Schiró and M.
Fabrizio, Phys. Rev. B 83, 165105 (2011).
A. Jreissaty, J. Carrasquilla, F. A. Wolf, and M. Rigol, Phys. Rev. A 84, 043610
(2011).
P. Anderson, Phys. Rev. 124, 41 (1961).
J. Kondo, Prog. Theor. Phys. 32, 37 (1964).
J. K. Freericks, V. M. Turkowski, and V. Zlatić, Phys. Rev. Lett. 97, 266408 (2006).
R. N. Silver, D. S. Sivia, and J. E. Gubernatis, Phys. Rev. B 41, 2380 (1990); M.
Jarrell and J. E. Gubernatis, Phys. Rep. 269, 133 (1996).
J. E. Hirsch and R. M. Fye, Phys. Rev. Lett. 56, 2521 (1986).
A. N. Rubtsov, V. V. Savkin, and A. I. Lichtenstein, Phys. Rev. B 72, 035122 (2005);
E. Gull, A. J. Millis, A. I. Lichtenstein, A. N. Rubtsov, M. Troyer, and Ph. Werner,
Rev. Mod. Phys. 83, 349 (2011).
Ph. Werner, A. Comanac, L. de Medici, M. Troyer, and A. J. Millis, Phys. Rev. Lett.
97, 076405 (2006).
J. E. Hirsch, Phys. Rev. B 28, 4059(R) (1983).
J. Hubbard, Phys. Rev. Lett. 3, 77 (1959).
R. L. Stratonovich, Sov. Phys. Doklady 2, 416 (1958).
N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, J.
Chem. Phys. 21, 1087 (1953).
K. Mikelsons, A. Macridin, and M. Jarrell, Phys. Rev. E 79, 057701 (2009).
K. van Houcke, F. Werner, E. Kozik, N. Prokof’ev, B. Svistunov, M. J. H. Ku, A.
T. Sommer, L. W. Cheuk, A. Schirotzek, and M. W. Zwierlein, Nature Phys. 8, 366
(2012).
E. Gull, D. R. Reichman, and A. J. Millis, Phys. Rev. B 84, 085134 (2011).
L. Mühlbacher and E. Rabani, Phys. Rev. Lett. 100, 176403 (2008).
M. Schiró and M. Fabrizio, Phys. Rev. B 79, 153302 (2009); M. Schiró, Phys. Rev. B
81, 085126 (2010).
M. Eckstein, M. Kollar, and Ph. Werner, Phys. Rev. Lett. 103, 056403 (2009).
G. Cohen, D. R. Reichman, A. J. Millis, and E. Gull, Phys. Rev. B 89, 115139 (2014).
N. Prokof’ev and B. Svistunov, Phys. Rev. B 77, 020408(R) (2008).
J. Ecalle, Les Fonctions Resurgentes, Vols. I−III. (Orsay, Publ. Math. 1981).
O. Costin, R. D. Costin, and J. L. Lebowitz, J. Stat. Phys. 116, 283 (2004).
G. Carleo, F. Becca, M. Schiró, and M. Fabrizio, Scientific Rep. 2, 243 (2011).
G. Carleo, F. Becca, L. Sanchez-Palencia, S. Sorella, and M. Fabrizio, Phys. Rev. A
89, 031602(R) (2014).
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
32
Freericks, Nikolić and Frieder
58.
59.
60.
61.
62.
63.
64.
K. Wilson, Rev. Mod. Phys. 47, 773 (1975).
S. R. White, Phys. Rev. Lett. 69, 2863 (1992).
S. Östlund and S. Rommer, Phys. Rev. Lett. 75, 3537 (1995.
G. Vidal, Phys. Rev. Lett. 91, 147902 (2003).
G. Vidal, Phys. Rev. Lett. 93, 040502 (2004).
F. Verstraete, D. Porras, and J. I. Cirac, Phys. Rev. Lett. 93, 227205 (2004).
E. M. Stoudenmire, L. O. Wagner, S. R. White, and K. Burke, Phys. Rev. Lett. 109,
056402 (2012).
J. Jordan, R. Orus, G. Vidal, F. Verstraete, and J. I. Cirac, Phys. Rev. Lett. 101,
250602 (2008).
G. Vidal, Phys. Rev. Lett. 99, 220405 (2007); G. Evenbly and G. Vidal, Phys. Rev. B
79, 144108 (2009); G. Evenbly and G. Vidal, Phys. Rev. Lett. 102, 180406 (2009).
S. Dusuel, M. Kamfor, R. Orus, K. P. Schmidt, and J. Vidal, Phys. Rev. Lett. 106,
107203 (2011).
Ph. Corboz, T. M. Rice, and M. Troyer, Phys. Rev. Lett. 113, 046402 (2014).
Ph. Corboz and Frederic Mila, Phys. Rev. Lett. 112, 147203 (2014).
M. P. Zaletel, R. S. K. Mong, Ch. Karrasch, J. E. Moore, and F. Pollmann, preprint,
arXiv:1407.1832 (2014).
F. B. Anders and A. Schiller, Phys. Rev. Lett. 95, 196801 (2005); F. B. Anders and
S. Schmitt, J. Phys: Confer. Ser. 220, 012021 (2010).
H. T. M. Nghiem and T. A. Costi, Phys. Rev. B 90, 035129 (2014).
F. Guettge, F. B. Anders, U. Schollwoeck, E. Eidelstein, and A. Schiller, Phys. Rev.
B 87, 115115 (2013).
Ch. Gramsch, K. Balzer, M. Eckstein, and M. Kollar, Phys. Rev. B 88, 235106 (2013).
F. A. Wolf, I. P. McCulloch and U. Schollwöck, preprint arxiv:1410.3342 (2014).
M. A. Cazalilla and J. B. Marston, Phys. Rev. Lett. 88, 256403 (2002); M. A. Cazalilla
and J. B. Marston, Phys. Rev. Lett. 91, 049702 (2003).
S. R. White and A. E. Feiguin, Phys. Rev. Lett. 93, 076401 (2004).
F. A. Wolf, I. P. McCulloch, O. Parcollet, and U. Schollwöck, Phys. Rev. B 90, 115124
(2014).
Y. Meir and N. S. Wingreen, Phys. Rev. Lett. 68, 2512 (1992); N. S. Wingreen, A.-P.
Jauho, and Y. Meir, Phys. Rev. B 48, 8487 (1993).
S. Datta, Electronic transport in mesoscopic systems (Cambridge University Press,
Cambridge, 1995).
G. Stefannuci and R. van Leuween, Nonequilibrium many-body theory of quantum
systems: A modern introduction (Cambridge University Press, Cambridge, 1995).
F. Mahouzi and B. K. Nikolić, Phys. Rev. B 90, 045115 (2014).
B. Gaury, J. Weston, M. Santin, M. Houzet, Ch. Groth, and X. Waintal, Phys. Rep.
534, 1 (2014).
L. D. Landau and E. M. Lifshitz, Quantum Mechanics, Third ed. (Burlington, MA,
Elsevier Butterworth-Heinemann, 1977).
K. Göttfried, Quantum Mechanics, Vol. 1. Fundamentals (New York, Benjamin, 1966).
A. R. Hurson and K. M. Kavi, Dataflow Computers: Their history and future, in Wiley
Encyclopedia of Computer Science and Engineering (New York, John Wiley and Sons,
2007).
P. Henderson and J. H. Morris, Conference Record of the Third ACM symposium on
principles of programming languages (1976); P. Hudak, ACM Computing Surveys 21,
383 (1989).
M. Carlsson, Proceedings of the seventh ACM SIGPLAN international conference on
functional programming (2002).
65.
66.
67.
68.
69.
70.
71.
72.
73.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
85.
86.
87.
88.
December 10, 2014
1:18
WSPC/INSTRUCTION FILE
extreme˙data˙v6
The nonequilibrium quantum many-body problem as a . . .
33
89. A. M. Shvaika and J. K. Freericks, Condens. Matter Phys. 11, 425 (2008).
90. N. Wiener and E. Hopf, Sitzber. Preuss. Akad. Wiss. Berlin, Kl. Math. Phys. Tech.,
696 (1931).
91. M. G. Krein, Uspekhi. Mat. Nauk. 13, 3 (1958) [Am. Math. Soc. Transl. 22, 163
(1962)].
92. B. M. MCoy and T. T. Wu, The two-dimensional Ising model, (Cambridge, MA,
Harvard University Press, 1973).
93. M. E. Fisher and R. E. Hartwig, Adv. Chem. Phys. 15 333 (1968).
94. H. Fotso, K. Mikelsons, and J. K. Freericks, Scientific Rep. 4, 4699 (2014).
95. H. Mera, T. G. Pedersen, and B. K. Nikolić, preprint, arXiv:1405.7956 (2014).
96. E. Delabaere, H. Dillinger, and F. Pham, Ann. Inst. Fourier 43, Fascicule 1, 163
(1993).
97. M. Mariño, Fortschr. Phys. 62, 455 (2014).
98. G. Cohen, E. Y. Wilner, and E. Rabani, New J. Phys. 15, 073018 (2013).
99. G. Cohen, E. Gull, D. R. Reichman, A. J. Millis, and E. Rabani, Phys. Rev. B 87,
195108 (2013).
100. P. Wang, G. Cohen, and S. Xu, preprint, arxiv:1410.1480 (2014).
101. Y. Lu, M. Höppner, O. Gunnarsson, and M. W. Haverkort, Phys. Rev. B 90, 085102
(2014).
102. W. Magnus, Commun. Pure Appl. Math. VII, 649 (1954).
| 5cs.CE
|
1
Active Anomaly Detection in Heterogeneous
Processes
arXiv:1704.00766v2 [cs.IT] 27 Jun 2017
Boshuang Huang, Kobi Cohen, Qing Zhao
Abstract— An active inference problem of detecting an anomalous process among M heterogeneous processes is considered. At
each time, a subset of processes can be probed. The objective is to
design a sequential probing strategy that dynamically determines
which processes to observe at each time and when to terminate
the search so that the expected detection time is minimized
under a constraint on the probability of misclassifying any
process. This problem falls into the general setting of sequential
design of experiments pioneered by Chernoff in 1959, in which
a randomized strategy, referred to as the Chernoff test, was
proposed and shown to be asymptotically optimal as the error
probability approaches zero. For the problem considered in this
paper, a low-complexity deterministic test is shown to enjoy the
same asymptotic optimality while offering significantly better
performance in the finite regime and faster convergence to the
optimal rate function, especially when the number of processes
is large. Furthermore, the proposed test offers considerable
reduction in implementation complexity. Extensions to detecting
multiple anomalous processes are also discussed.
Index Terms— Active hypothesis testing, sequential design
of experiments, anomaly detection, dynamic search.
I. I NTRODUCTION
We consider the problem of detecting an anomalous process
(referred to as the target) among M heterogeneous processes
(referred to as the cells). At each time, K (1 ≤ K < M ) cells
can be probed simultaneously to search for the target. Each
search of cell i generates a noisy observation drawn i.i.d. over
time from two different distributions fi and gi , depending on
whether the target is absent or present. The objective is to
designa sequential search strategy that dynamically determines
which cells to probe at each time and when to terminate the
search so that the expected detection time is minimized under
a constraint on the probability of declaring a wrong location
of the target.
The above problem is prototypical of searching for rare
events in a large number of data streams or a large system. The
rare events could be opportunities (e.g., financial trading opportunities or transmission opportunities in dynamic spectrum
access [1]), unusual activities in surveillance feedings, frauds
in financial transactions, attacks and intrusions in communication and computer networks, anomalies in infrastructures such
as bridges, buildings, and the power grid that may indicate
Boshuang Huang and Qing Zhao are with the School of Electrical and
Computer Engineering, Cornell University, Ithaca, NY, 14853, USA. Email:
{bh467, qz16}@cornell.edu
Kobi Cohen is with the Department of Electrical and Computer Engineering, Ben-Gurion University of the Negev, Beer-Sheva 84105, Israel. Email:
[email protected]
This work was supported by the National Science Foundation under Grant
CCF-1534957.
catastrophes. Depending on the application, a cell may refer
to an autonomous data stream with a continuous data flow or
a system component that only generates data when probed.
A. Main Results
The anomaly detection problem considered in this paper
is a special case of active hypothesis testing originated from
Chernoff’s seminal work on sequential design of experiments
in 1959 [2]. Compared with the classic passive sequential
hypothesis testing pioneered by Wald [3], where the observation model under each hypothesis is predetermined, active
hypothesis testing has a control aspect that allows the decision maker to choose the experiment to be conducted at
each time. Different experiments generate observations from
different distributions under each hypothesis. Intuitively, as
more observations are gathered, the decision maker becomes
more certain about the true hypothesis, which in turn leads to
better choices of experiments.
In [2], Chernoff proposed a randomized strategy, referred
to as the Chernoff test, and established its asmyptotic (as the
error probability diminishes) optimality1 . This randomized test
chooses, at each time, a probability distribution that governs
the selection of the experiment to be carried out at this time.
This distribution is obtained by solving a minimax problem so
that the next observation generated under the random action
can best differentiate the current maximum likelihood estimate
of the true hypothesis (using all past observations) from its
closest alternative, where the closeness is measured by the
Kullback-Liebler (KL) divergence. Due to the complexity in
solving this minimax problem at each time, the Chernoff test
can be expensive to compute and cumbersome to implement,
especially when the number of hypotheses or the number of
experiments is large.
It is not difficult to see that the problem at hand is a special
case of the general active hypothesis testing problem. Specically, the available experiments are in the form of different
subsets of K cells to probe, and the number of experiments
is M
K . Under each hypothesis that cell m (m = 1, ..., M )
is the target, the distribution of the next observation (a vector
of dimension K) depends on which K cells are chosen. The
Chernoff test thus directly apply. Unfortunately, with the large
number of hypotheses and the large number of experiments,
it can be computationally prohibitive to obtain the Chernoff
test.
1 The asymptotic optimality of the Chernoff test was shown under the
assumption that the hypotheses are distinguishable under every experiment.
2
In this paper, we show that the anomaly detection problem
considered here exhibits sufficient structures to admit a lowcomplexity deterministic policy with strong performance. In
particular, we develop a deterministic test that explicitly specifies which K cells to search at each given time and show
that this test enjoys the same asymptotic optimality as the
Chernoff test2 . Furthermore, extensive simulation examples
have demonstrated a significant performance gain over the
Chernoff test in the finite regime and faster convergence to the
optimal rate function, especially when M is large. In contrast
to the Chernoff test, the proposed test requires little offline
or online computation. The test can also be extended to cases
with multiple targets as discussed in Section V. Its asymptotic
optimality is preserved for K = 1.
We point out that proving the asymptotic optimality of the
deterministic policy is much more involved comparing with
the Chernoff test, due to the time dependency in the test
statistics, namely, the log-likelihood ratios (LLRs), introduced
by deterministic actions. In particular, since the distribution of
the random action chosen by the Chernoff test depends only on
the current maximimum likelihood estimate of the underlying
hypothesis which becomes time-invariant after an initial phase
with a bounded duration, the stochastic behaviors of the
LLRs are independent over time, resulting in a much easier
analysis of the detection delay. The deterministic actions of the
proposed test, however, lead to complex time dependencies in
LLRs that make the analysis much more involved.
B. Related Work
Chernoff’s pioneering work on sequential design of experiments focuses on sequential binary composite hypothesis
testing [2]. Variations and extensions of the problem were
studied in [4]–[9], where the problem was referred to as
controlled sensing for hypothesis testing in [5]–[7] and active
hypothesis testing in [8], [9]. As variants of the Chernoff test,
the tests developed in [4]–[9] are all randomized tests.
There is an extensive literature on dynamic search and target
whereabout problems under various scenarios, most of them
focusing on homogeneous processes. We discuss here existing
studies within the sequential inference setting, which is the
most relevant to this work. In [10], the problem of searching
among Gaussian signals with rare mean and variance values
was studied and an adaptive group sampling strategy was
developed. In [11], searching over homogeneous Poisson point
processes with unknown rates was investigated and an asymptotically optimal randomized test was developed. In [12], the
problem of tracking a target that moves as a Markov Chain in a
finite discrete environment is studied and a search strategy that
provides the most confident estimate is developed. In [13], an
important case of multichannel sequential change detection is
studied and an asymptotic framework in which the number
of sensors tends to infinity was proposed. Asymptotically
optimal search policies over homogeneous processes were
established in [14] under a non-parametric setting with finite
discrete distributions and in [15] under a parametric composite
2 The asymptotic optimality of the proposed test holds for all but at most
three singular values of K (see Lemma 1 and Theorem 1).
hypothesis setting with continuous distributions. In [16], [17],
the problem of quickly detecting anomalous components under
the objective of minimizing system-wide cost incurred by
all anomalous components was studied. The objective of
minimizing operational cost as opposed to detection delay led
to a different problem from the one considered in this paper.
Other related work on quickest search over multiple processes
under various models and formulations includes [18]–[21]
and references therein. Sequential spectrum sensing within
both the passive and active hypothesis testing frameworks has
also received extensive attention in the application domain
of cognitive radio networks (see, for example, [22]–[25] and
references therein). The readers are also referred to [26] for
a comprehensive survey on the problem of detecting outlying
sequences.
A prior study by Cohen and Zhao considered the problem
for homogeneous processes (i.e., fi ≡ f and gi ≡ g) [27]. This
work builds upon this prior work and addresses the problem
in heterogeneous systems where the absence distribution fi
and the presence distribution gi are different across processes.
Allowing heterogeneity significantly complicates the design
of the test and the establishment of asymptotic optimality.
Specifically, since each process has different observation distributions, the rate at which the state of a cell can be inferred is different across processes. Hence, the decision maker
must balance the search time effectively among the observed
processes, which makes both the algorithm design and the
performance analysis much more involved under the heterogeneous case. In terms of algorithm design, when dealing with
homogeneous processes, the search strategy is often static in
nature [11], [14], [18], [27]. In contrast, the asymptotically
optimal search strategy developed here for heterogeneous processes dynamically changes based on the current belief about
the location of the target. In terms of performance analysis,
handling heterogeneity adds new challenges and difficulties
for establishing asymptotic optimality. When searching over
homogeneous processes, the resulting rate function (which
is inversely proportional to the search time) always obeys a
certain averaging over the KL divergences between normal and
abnormal distributions of all process. This observation follows
from the fact that the decision maker completes gathering
the required information from all the processes at approximately the same time due to the homogeneity. In contrast,
when searching over heterogeneous processes, the overall rate
function does not always obey a simple averaging across the
KL divergences of all processes. In Section IV, we show that
the search time can be analyzed by considering two separate
scenarios, referred to as the balanced and the unbalanced
cases. The balanced case holds when the a judicious allocation
of probing resources can ensure the information gathering
from all the processes be completed at approximately the
same time, in which case the rate function is a weighted
average among the heterogeneous processes. The unbalanced
case occurs when there is a process with a sufficiently small
KL divergence that it dominates the overall rate function of
the search. We establish asymptotic optimality by analyzing
the sum LLR dynamics of the heterogeneous processes under
these two cases which adds significant technical difficulties as
3
compared to the homogeneous case as detailed in Section IV.
Besides the active inference approach to anomaly detection
considered in this paper, there is a growing body of literature
on various approaches to the general problem of anomaly
detection. We refer the readers to [28], [29] for comprehensive
surveys on this topic.
II. P ROBLEM F ORMULATION
We consider the problem of detecting a single target located
in one of M cells. If the target is in cell m, we say that
hypothesis Hm is true. The aPpriori probability that Hm is
M
true is denoted by πm , where m=1 πm = 1. To avoid trivial
solutions, it is assumed that 0 < πm < 1 for all m.
When cell m is observed at time n, an observation ym (n)
is drawn, independent of previous observations. If cell m
contains a target, ym (n) follows distribution gm (y). Otherwise, ym (n) follows distribution fm (y). Let Pm be the
probability measure under hypothesis Hm and Em the operator
of expectation with respect to the measure Pm .
An active search strategy Γ consists of a stopping rule τ
governing when to terminate the search, a decision rule δ for
determining the location of the target at the time of stopping,
and a sequence of selection rules {φ(n)}n≥1 governing which
K cells to probed at each time n. Let y(n) be the set of all
cell selections and observations up to time n. A deterministic
selection rule φ(n) at time n is a mapping from y(n − 1)
K
to {1, 2, ..., M } . A randomized selection rule φ(n) is a
mapping from y(n − 1) to probability mass functions over
K
{1, 2, ..., M } .
PThe error probability under policy Γ is defined as Pe (Γ) =
m πm αm (Γ), where αm (Γ) = Pm (δ 6= m|Γ) is the
probabilityPof declaring δ 6= m when Hm is true. Let
M
E(τ |Γ) = m=1 πm Em (τ |Γ) be the average detection delay
under Γ.
We adopt a Bayesian approach as in Chernoff’s original
study [2] by assigning a cost of c for each observation and
a loss of 1 for a wrong declaration. Note that c represents
the ratio of the sampling cost to the cost of wrong detections.
The Bayes risk under strategy Γ when hypothesis Hm is true
is given by:
Rm (Γ) , αm (Γ) + cEm (τ |Γ).
(1)
The average Bayes risk is given by:
R(Γ) =
M
X
πm Rm (Γ) = Pe (Γ) + cE(τ |Γ).
(2)
m=1
The objective is to find a strategy Γ that minimizes the
Bayes risk R(Γ):
inf R(Γ).
(3)
Fig. 1: Typical sample paths of sum LLRs.
III. T HE D ETERMINISTIC DGF I P OLICY
In this section we propose a deterministic policy, referred
to as the DGFi policy.
A. DGFi policy for K = 1
We first consider the case where only a single process can
be observed at a time, i.e., K = 1.
Let 1m (n) be the indicator function, where 1m (n) = 1 if
cell m is observed at time n, and 1m (n) = 0 otherwise. Let
`m (n) , log
and
Sm (n) ,
gm (ym (n))
,
fm (ym (n))
n
X
`m (t)1m (t)
(6)
(7)
t=1
be the log-likelihood ratio (LLR) and the observed sum LLRs
of cell m at time n, respectively. Let D(g||f ) denote the KL
divergence between two distributions g and f given by
Z ∞
g(x)
D(g||f ) ,
log
g(x) dx.
(8)
f
(x)
−∞
Illustrated in Fig. 1 are typical sample paths of the sum LLRs
of M = 4 cells, where, without loss of generality, we assume
that cell 1 is the target. Note that the sum LLR of cell 1 is a
random walk with a positive expected increment D(g1 ||f1 ),
whereas the sum LLR of cell i is a random walk with a
negative expected increment −D(fi ||gi ) for i = 2, 3, 4. Thus,
when the gap between the largest sum LLR and the second
largest sum LLR is sufficiently large, we can declare with
sufficient accuracy that the cell with the largest sum LLR is
the target. This is the intuition behind the stopping rule and
the decision rule. Specifically, we define m(i) (n) as the index
of the cell with the ith largest observed sum LLRs at time n.
Let
∆S(n) , Sm(1) (n) (n) − Sm(2) (n) (n)
(9)
Γ
∗
A strategy Γ is asymptotically optimal if
lim
c→0
R(Γ∗ )
= 1,
inf Γ R(Γ)
(4)
which is denoted as
R(Γ∗ ) ∼ inf R(Γ).
Γ
denote the difference between the largest and the second
largest observed sum LLRs at time n. The stopping rule and
the decision rule under the DGFi policy are given by:
τ = inf {n : ∆S(n) ≥ − log c} ,
(10)
δ = m(1) (τ ) .
(11)
and
(5)
4
We now specify the selection rule of the DGFi policy. The
intuition behind the selection rule is to select a cell from
which the observation can increase ∆S(n) at the fastest rate.
The selection rule is thus given by comparing the rate at
which Sm(1) (n) increases with the rate at which Sm(2) (n) (n)
decreases. If Sm(1) (n) is expected to increase faster than
Sm(2) (n) (n) decreases, cell m(1) (n) is chosen. Otherwise, cell
m(2) (n) is chosen. This leads to the following selection rule:
(1)
m (n) if D(gm(1) (n) ||fm(1) (n) ) ≥ Fm(1) (n)
,
φ(n) =
m(2) (n) otherwise
(12)
where
1
.
Fm , P
(13)
1
j6=m D(fj ||gj )
The selection rule in (12) can be intuitively understood by
noticing that D(gm(1) (n) ||fm(1) (n) ) is the asymptotic increasing rate of Sm(1) (n) when cell m(1) is probed at each time.
This is due to the fact that m(1) (n) is the true target after an
initial phase (defined by the last passage time that m(1) (n)
is an empty cell) which can be shown to have a bounded
expected duration. Similarly, even though much more involved
to prove, Fm(1) (n) is the asymptotic rate at which Sm(2) (n) (n)
decreases when cell m(2) (n) is probed at each time. To see
the expression of Fm for any m as given in (13), consider
the following analogy. Consider M − 1 cars being driven
by a single driver from 0 to −∞. Car j (j = 1, . . . , M ,
j 6= m) has a constant speed of D(fj ||gj ). At each time,
the car closest to the origin is chosen by the driver and driven
by one unit of time. We are interested in the average moving
speed of the position of the closest car to the origin. It is not
difficult to see that it is given by Fm in (13). This analogy,
concerned with deterministic processes, only serves as an
intuitive explanation for the expression of Fm . As detailed
in Sec. IV, proving Fm(1) (n) to be the asymptotic decreasing
rate of Sm(1) (n) (n) requires analyzing the trajectories of the
M sum LLRs {Sm (n)}M
m=1 , which are stochastic processes
with complex dependencies both in time and across processes.
B. DGFi under multiple simultaneous observations
Next we extend the DGFi policy to the case where multiple
simultaneous observations are allowed, i.e., K > 1.
The stopping rule and the decision rule remains the same
as given in (10), (11), whereas the selection rule requires
significant modification. The main reason is that when K cells
can be observed simultaneously, the asymptotic increasing
rate of Sm(1) (n) (n) and the asymptotic decreasing rate of
Sm(2) (n) (n) are much more involved to analyze.
The selection rule is as follows. At each time n, the selection
rule φ(n), as given in (14), chooses either the K cells with
(
φ(n) =
m(1) (n), m(2) (n), ..., m(K) (n)
m(2) (n), m(3) (n), ..., m(K+1) (n)
the top K largest sum LLRs or those with the second to the
(K + 1)th largest sums LLRs, where
Fm (K) , min{ P
K
1
j6=m D(fj ||gj )
, min D(fj ||gj )}.
j6=m
(15)
Similar to the case with K = 1, the intuition behind the
selection rule is to select K cells from which the observations
increase ∆S(n) at the fastest rate. Specifically, Fm(1) (n) (K) is
the asymptotic decreasing rate of Sm(2) (n) (n) when K cells
with the second largest to the (K + 1)th largest sum LLRs
are probed each time. The expression of Fm (K) for any m
as given in (15) can be explained with the same car analogy,
except now there are K > 1 drivers. It is not difficult to see
that Fm (K) is upper bounded by the speed minj6=m D(fj ||gj )
of the slowest car among the M − 1 cars. In particular, when
the speed of the slowest car is sufficiently small, this car
always lags behind even with a dedicated driver. We refer
to this case as the unbalanced case, which presents the most
challenge in proving the asymptotic optimality of DGFi (see
Theorem 1 and Appendix B). With this intuitive understanding
of Fm (K), we can see that the asymptotic increasing rate
of ∆S(n) is D(gm(1) (n) ||fm(1) (n) ) + Fm(1) (n) (K − 1) when
the cells with the top K largest sum LLRs are probed each
time, where D(gm(1) (n) ||fm(1) (n) ) is the asymptotic increasing
rate of Sm(1) (n) (n) and Fm(1) (n) (K − 1) is the asymptotic
decreasing rate of Sm(2) (n) (n).
It is easy to see that when K = 1, the policy reduces to the
one described in section III-A.
IV. P ERFORMANCE A NALYSIS
In this section, we establish the asymptotic optimality of the
DGFi policy. While the intuitive exposition of DGFi given in
Sec. III may make its asymptotic optimality seem expected,
constructing a proof is much more involved. In particular,
bounding the detection time of DGFi requires analyzing the
trajectories of the M stochastic processes {Sm (n)}M
m=1 which
exhibit complex dependencies both over time and across
processes as induced by the deterministic selection rule.
We first state the following assumption.
Assumption 1: Under hypothesis Hm , assume that
u∗m , arg max uD(gm ||fm ) + Fm (K − u)
u∈[0,1]
(16)
takes value of either 0 or 1, where we allow the domain of
Fm (·) defined in (15) to be all real numbers.
In the following lemma, we give an explicit characterization
on when Assumption 1 is violated and at what values of K.
Lemma 1: For any given {D(gi ||fi ), D(fi ||gi )}M
i=1 , we
have the following statements.
if D(gm(1) (n) ||fm(1) (n) ) + Fm(1) (n) (K − 1) ≥ Fm(1) (n) (K)
otherwise
(14)
5
1) For each m = 1, . . . , M , Assumption 1 holds if at least
one of the following two statements is true:
P
minj6=m D(fj ||gj )
(a)
is an integer,
j6=m
D(fj ||gj )
1
P
(b) D(gm ||fm ) ≥
.
1
j6=m D(fj ||gj )
If neither is true, then Assumption 1 does not hold for a
single value of K, denoted as K̃m , as given below.
X minj6=m D(fj ||gj )
(17)
K̃m =
D(fj ||gj )
j6=m
2) All {K̃m }M
m=1 take at most three distinct values.
Proof: See Appendix A.
From Lemma 1, we conclude that for every given M
and {D(gi ||fi ), D(fi ||gi )}M
i=1 , Assumption 1 always holds for
K = 1. For the general case and with an arbitrarily large
M , Assumption 1 holds for all but at most three values of
K ∈ {1, 2, . . . , M }.
Next, we establish the asymptotic optimality of DGFi under
Assumption 1. Define
Im , max{D(gm ||fm ) + Fm (K − 1), Fm (K)},
(18)
which is the increasing rate of ∆S(n) under hypothesis
Hm . For a given a priori distribution {πm }M
m=1 of the true
hypothesis, define
I ∗ , PM
1
πm
m=1 Im
.
(19)
As shown in Theorem 1 below, I ∗ is the optimal rate function
of the Bayes risk.
Theorem 1: Let R∗ and R(Γ) be the Bayes risks under the
DGFi policy and an arbitrary policy Γ, respectively. Assume
that Assumption 1 holds for all m = 1, . . . , M . Then,
R∗ ∼
−c log c
∼ inf R(Γ)
Γ
I∗
small such that the increasing rate of ∆S(n) is given by a
certain averaging among the heterogeneous
P processes.
The unbalanced case is when K/ j6=m D(fj1||gj ) >
minj6=m D(fj ||gj ). In this case, there is a process with a
sufficiently small information acquisition rate D(fj ||gj ) such
that it becomes the bottleneck of the detection process and
determines the asymptotic increasing rate of ∆S(n). Directly
bounding the dynamic range of all sum LLR trajectories
is no longer tractable. Instead, the proof is built upon the
analysis of the trajectory of the sum LLR with the smallest
expected increment. In particular, we recognize that the key
in handling the imbalance in the information acquisition rates
among empty cells is to define a last passage time as the last
time at which the empty cell with the smallest D(fj ||gj ) is
not probed and then analyze, separately, the detection process
before and after this last passage time.
c
is an asymptotic lower bound
Next, we show that −cIlog
∗
on the Bayes risk. This is done by first proving that if the
Bayes risk is sufficiently small under strategy Γ, i.e., R(Γ) =
O(−c log c), the difference between the largest sum LLRs and
the second largest sum LLRs must be sufficiently large when
the test terminates, i.e. ∆S(τ ) = Ω(− log c). Otherwise, it
is not possible to achieve a risk O(−c log c) due to a large
error probability. We then show that in order to make ∆S(n)
sufficiently large, the sample size must be large enough, i.e.,
c
E[τ |Γ] ≥ − Ilog
. Since each sample costs c, the total risk will
∗
c
be lower bounded by −cIlog
as desired.
∗
(20)
Proof: Here we provide a sketch of the proof. The detailed
proof can be found in Appendix B.
We first show that the proposed DGFi policy achieves a
Bayes risk −c log c/I ∗ asymptotically. First, we show that
when ∆S(τ ) is large, the probability of error is small, i.e.
Pe = O(c). As a result, by the definition of the Bayes risk,
it suffices to show that the detection time is upper bounded
by − log c/I ∗ . By the definition of I ∗ in (19), it suffices to
show that the detection time is upper bounded by − log c/Im
under hypothesis Hm . Since the decision maker might not
complete to gather the required information from all the cells
at the same time, we carry out the analysis by considering the
balanced
P and the unbalanced cases separately. In particular,
if K/ j6=m D(fj1||gj ) < minj6=m D(fj ||gj ), we refer to this
case as the balanced case. Otherwise, we refer to this case as
the unbalanced case.
P
The balanced case is when K/ j6=m D(fj1||gj ) ≤
minj6=m D(fj ||gj ). The key to bounding the detection time in
this case is to show that the dynamic range of the M − 1 sum
LLRs corresponding to the M − 1 empty cells are sufficiently
V. E XTENSION TO D ETECTING M ULTIPLE TARGETS
In this section we extend the DGFi policy to the case with
L > 1 targets. The number of hypotheses in this case is M
L .
We consider first K = 1. The DGFi policy can be extended
to detect multiple targets as follows. The stopping rule and
decision rule are given below, similar in principle to those for
L = 1 as described in Section III:
τ = inf {n : ∆SL (n) ≥ − log c} ,
(21)
δ = {m(1) (τ ), m(2) (τ ), ..., m(L) (τ )},
(22)
∆SL (n) , Sm(L) (n) (n) − Sm(L+1) (n) (n)
(23)
where
denotes the difference between the Lth and the (L + 1)th
largest observed sum LLRs at time n.
For the selection rule, define, for a given set D ⊂
{1, 2, . . . , M } with |D| = L,
FD , P
1
1
j ∈D
/ D(fj ||gj )
.
(24)
Similar to Fm defined in (13), FD can be viewed as the
asymptotic increasing rate of ∆SL (n) when the L targets are
given by set D and we probe the cell with the (L+1)th largest
sum LLR. We also define
1
GD , P
,
(25)
1
j∈D D(gj ||fj )
6
which can be viewed as the asymptotic increasing rate for
∆SL (n) when we probe the cell with the Lth largest sum
LLR.
The selection rule follows the same design principle of
maximizing the asymptotic increasing rate of ∆SL (n), and
is given by
(L)
m (n)
if GD(n) ≥ FD(n)
φ(n) =
,
(26)
m(L+1) (n) otherwise
where
D(n) = {m(1) (n), m(2) (n), ..., m(L) (n)}.
The asymptotic optimality of DGFi for L > 1 and K > 1
remains open. We have, however, strong belief of the following
conjecture.
Conjecture 1: The DGFi policy preserves its asymptotic
optimality if
u∗D , arg max FD (K − u) + GD (u)
where
u∈[0,K]
D(n) = {m(1) (n), m(2) (n), ..., m(L) (n)}.
(27)
It is not difficult to see that when L = 1, the policy reduces
to the one described in Section III.
Next, we establish the asymptotic optimality of the DGFi
policy for L > 1 and K = 1. Let D denote a subset of L cells
and πD the prior probability of hypothesis HD (i.e, the target
cells are given by D). Define
ID , max{FD , GD },
1
IL∗ , P πD ,
(28)
D ID
where IL∗ is again the optimal rate function of the Bayes risk
as shown in the theorem below, and reduces to the one defined
in (19) when L = 1.
∗
Theorem 2: Let RL
and RL (Γ) be the Bayes risks under
the DGFi policy and an arbitrary policy Γ, respectively. Then,
for K = 1,
−c log c
∼ inf R(Γ) .
Γ
IL∗
R∗ ∼
(29)
Proof: See Appendix C.
When K > 1, the stopping rule and the decision rule remain
the same. For the selection rule, define
K
FD (K) , min{ P
1
j ∈D
/ D(fj ||gj )
, min D(fj ||gj )}.
j ∈D
/
K
1
j∈D D(gj ||fj )
, min D(gj ||fj )},
j∈D
(31)
which can be viewed as the asymptotic increasing rate of
∆SL (n) when we probe the cells with the (L − K + 1)th
to the Lth largest sum LLR.
Let
∗
, arg max FD (K − k) + GD (k),
kD
0≤k≤K
(32)
which can be interpreted as the optimal number of target
cells that should be probed at each time for maximizing the
asymptotic increasing rate of ∆SL (n). The selection rule of
DGFi is thus given by
∗
∗
(35)
is an integer for all D, where we allow the domain of FD (·)
and GD (·) to be real numbers.
VI. C OMPARISON WITH THE C HERNOFF T EST
In this section, we compare the performance of the proposed
DGFi policy and the Chernoff test in terms of both computational complexity and sample complexity.
A. The Chernoff Test
The Chernoff test has a randomized selection rule. Specifically, let q = (q1 , ..., qκ ) be a probability mass function over
κ
a set of κ available experiments {ui }i=1 that the decision
maker can choose from, where qi is the probability of choosing
experiment ui . Note that in our case, κ = M
K . For a
general M -ary active hypothesis testing problem, the action
at time n under the Chernoff test is drawn from a distribution
q ∗ (n) = (q1∗ (n), ..., qκ∗ (n)) that depends on the past actions
and observations:
X
i
q ∗ (n) = arg max
min
qi D(puî(n)
||puj i ) , (36)
q
j∈M\{î(n)} u
i
where M is the set of the M hypotheses, î(n) is the ML
estimate of the true hypothesis at time n based on past actions
and observations, and puj i is the observation distribution under
hypothesis j when action ui is taken. The stopping rule and
the decision rule are the same as in (10), (11).
(30)
Similar to Fm (K) defined in (15), FD (K) can be viewed as
the asymptotic increasing rate of ∆SL (n) when the L targets
are given by set D and we probe those K cells with the (L +
1)th to the (L + K)th largest sum LLR. Similarly,
GD (K) , min{ P
(34)
φ(n) = {m(L−kD(n) +1) (n), · · · , m(L−kD(n) +K) (n)}, (33)
B. Comparison in computational complexity
Here we compare the computational complexity of the
proposed DGFi policy with the Chernoff test. We show that
the Chernoff test can be expensive to compute especially when
the number of processes or the number of experiments is large.
In contrast to the Chernoff test, the DGFi policy requires little
computaion.
For the case of detecting a single target (L = 1), computing
the selection rule of Chernoff test defined in (36) requires solving M minimax problems, each corresponding to a particular
value of the ML estimate î(n) ∈ {1, . . . , M }. One efficient
way of solving minimax problems is through linear programming which takes polynomial time with respect to the number
of variables and constraints.
For this problem, however, the
number of variables is M
K , which is not polynomial and can
be exponential in M in the worst case.
The only computation involved in the selection rule of DGFi
is (15), which requires M summations each with M − 1
elements. As a result, the compuational time is O(M 2 ), which
is polynomial in M and independent of K.
7
(m)
= 9 +
Fig. 4: Performance comparison (L = 2, K = 1, λg
(m)
m, λf = 0.0188, c = 10−5 ).
(m)
= 9 +
Fig. 5: Performance comparison (M = 20, L = 2, K =
(m)
(m)
2, λg = 50 + 0.1m, λf = 2).
Fig. 2: Performance comparison (K = 1, λg
(m)
m, λf = 0.0188, c = 10−5 ).
Fig. 3: Performance comparison (K = 2, λg
(m)
m, λf = 0.0188, c = 10−5 ).
C. Comparison in sample complextity
Although both the Chernoff test and the DGFi policy
are asymptotically optimal3 , we show below via simulation
examples the significant performance gain of DGFi over the
Chernoff test in the finite regime (i.e., when the sample cost
c is bounded away from 0).
Consider a uniform prior and exponentially distributed ob(m)
(m)
servations: fm ∼ exp(λf ) and gm ∼ exp(λg ). The KL
divergences can be easily computed as follows.
(m)
(m)
D(gm ||fm ) = log(λ(m)
g ) − log(λf ) +
λf
(m)
−1,
λg
(m)
(m)
D(fm ||gm ) = log(λf ) − log(λ(m)
g )+
λg
(m)
−1.
λf
Shown in Fig. 2 is the performance comparison between
DGFi policy and Chernoff test for L = 1 and K = 1.
The figure clearly demonstrates the significant reduction in
detection delay offered by the DGFi policy as compared with
the Chernoff test. The performance gain increases drastically
as M increases. A similar comparison is observed in Fig. 3
3 While the assumption of positive KL divergence between every pair of
hypotheses under every probing action as required in Chernoff’s proof of
asymptotic optimality does not hold here, it can be shown that Chernoff test
preserves its asymptotic optimality for the problem at hand.
(m)
= 9+
with K = 2. The performance comparison for a case with
multiple targets is shown in Fig. 4 with L = 2, K = 1.
In Fig. 5, we consider a case of L = 2 and K = 2 and
examine the performance of DGFi as c approaches 0. Also
c
plotted in Fig. 5 is the asymptotic lower bound − Ilog
on the
∗
L
detection delay which increases linearly with − log c with rate
1/IL∗ as given in (28). We observe that the increasing rate of
the detection delay offered by DGFi quickly converges to that
of the lower bound, which supports the conjecture that DGFi
preserves its asymptotic optimality for the case with L > 1
and K > 1. Besides showing the same level of reduction
in the finite regime compared with the Chernoff test, Fig. 5
also reveals a significantly faster convergence to the optimal
rate function IL∗ with the detection delay of the Chernoff test
increasing at a faster rate even at c = 10−10 .
Next, we provide an intuition argument for the better finitetime performance of DGFi. Consider a special case where
K = 1 and all fi and gi are identical, i.e., fi ≡ f and
gi ≡ g and we assume D(f ||g) > (M − 1)D(g||f ). In
this case, the DGFi policy chooses, at each time, the cell
with the second largest sum LLR whereas the Chernoff test
randomly and uniformly chooses a cell from all but the one
with the largest sum LLR at each time. Consider a short
horizon scenario where the sampling cost c is sufficiently high
such that D(f ||g) > − log c. This means each empty cell only
need one observation (with high probability) to distinguish
8
from the true cell. We can formulate this as coupon collectors
problem, where each empty cell is a coupon and the goal is
to collect all M − 1 coupons.
Since Chernoff test employs a randomized strategy that
chooses empty cells with equal probability, based on results
in coupon collectors problem, the expected probing time will
be roughly M log M . However, the proposed DGFi policy is
deterministic and guaranteed to collect a new coupon at each
time, therefore the expected probing time will only be M .
VII. C ONCLUSION
The problem of detecting anomalies among a large number
of heterogeneous processes was considered. A low-complexity
deterministic test was developed and shown to be asymptotically optimal. Its finite-time performance and computational
complexity were shown to be superior to the classic Chernoff
test for active hypothesis testing, especially when the problem
size is large.
A PPENDIX A: PROOF OF L EMMA 1
Define
hm (u) = uD(gm ||fm ) + Fm (K̃m − u).
(37)
If u∗m takes value other than 0 or 1, i.e., u∗m ∈ (0, 1), then
h0m (u) > 0 for u ∈ (0, u∗m ) and h0m (u) < 0 for u ∈ (u∗m , 1).
By taking the derivative of hm (u), we have
0
h0m (u) = D(gm ||fm ) − Fm
(K̃m − u),
(38)
where
0
Fm
(u)
=
P
1
1
j6=m D(fj ||gj )
, if u <
P
minj6=m D(fj ||gj )
D(fj ||gj )
if u >
P
minj6=m D(fj ||gj )
.
D(fj ||gj )
0,
j6=m
j6=m
(39)
0
(u) is piecewise constant with a breakpoint
Since Fm
P
minj6=m D(fj ||gj )
, h0m (u) is piecewise constant with a
j6=m
D(fj ||gj )
P
minj6=m D(fj ||gj )
breakpoint K̃m − j6=m
. Therefore,
D(fj ||gj )
K̃m = u∗m +
X minj6=m D(fj ||gj )
.
D(fj ||gj )
(40)
j6=m
Since
h0m (u)
.
(41)
Note that u∗m ∈ (0, 1) and K̃m is an integer. Such K̃m exists
only if neither (a) or (b) holds and we have
X minj6=m D(fj ||gj )
.
K̃m =
(42)
D(fj ||gj )
j6=m
Next we show that there are only three possible values of K.
Let j 0 = arg minj D(fj ||gj ). Since there is only one possible
K̃j 0 as proved above. It remains to show that there are only
two possible values of Km when m 6= j 0 . Let
M
X
D(fj 0 ||gj 0 )
j=1
≤ 1, we have
X minj6=m D(fj ||gj )
D(fj 0 ||gj 0 )
=V −
∈ [V − 1, V ]
D(fj ||gj )
D(fm ||gm )
j6=m
for all m 6= j 0 . Combining (42) implies that K̃m , m 6= j 0 can
only take 2 possible integers as desired.
A PPENDIX B: PROOF OF T HEOREM 1
The structure of the proof is as follows. In subsection A,
we show that −c log c/I ∗ is an asymptotic upper bound on
the Bayes risk that DGFi achieves. Specifically, the asymptotic
optimality property of DGFi is based on Lemma 8, showing
that the asymptotic expected search time is upper bounded
by − log c/I ∗ , while the error probability is O(c) following
Lemma 2. In subsection B, we provide the sum LLR analysis
of the heterogeneous empty cells. The analysis is based on
studying two cases, referred to as balanced and unbalanced
cases. For the balanced case, the decision maker can balance
the remaining information required to be gathered among the
processes. For the unbalanced case, there is a process with a
very small KL divergence so that it dominates the increasing
rate. Finally, in subsection C we show that the asymptotic
Bayes risk that can be achieved by any policy is lower bounded
by −c log c/I ∗ , in which together with Appendix A completes
the proof.
Throughout the this section, we use the following notations.
Let
n
X
Nj (n) ,
1j (t)
(43)
t=1
be the number of times that cell j has been observed up to
time n. Let
∆Sm,j (n) , Sm (n) − Sj (n)
(44)
be the difference between the observed sum of LLRs of cells
m and j. We also define
∆Sm (n) , min ∆Sm,j (n) .
j6=m
(45)
∆S(n) = Sm(1) (n) (n) − Sm(2) (n) (n) = max ∆Sm (n) . (46)
1
1
j6=m D(fj ||gj )
V ,
D(fj 0 ||gj 0 )
D(fm ||gm )
As a result, we have:
< 0 for u ∈ (u∗m , 1),
D(gm ||fm ) < P
Since 0 ≤
D(fj ||gj )
.
m
Without loss of generality we prove the theorem when
hypothesis m is true. We define
`k (i) − D(gk ||fk ) , if k = m,
`˜k (i) =
(47)
`k (i) + D(fk ||gk ) , if k 6= m,
which is a zero-mean r.v under hypothesis Hm .
A. The Asymptotic Upper Bound on the Bayes Risk under
DGFi
In this subsection we show that the Bayes risk obtained by
DGFi policy is upper bounded by −c log c/I ∗ as c approaches
zero..
9
Lemma 2: If DGFi policy is used, then the error probability
is upper bounded by:
Pe ≤ (M − 1)c .
(48)
Proof:
P Let αm,j = Pm (δ = j) for all j 6= m. Thus,
αm =
j6=m αm,j . By the definition of the stopping rule
under DGFi (see (10)), accepting Hj is done when ∆Sj (n) ≥
− log c which implies ∆Sj,m ≥ − log c. Hence, for all j 6= m
we have:
αm,j = Pm (δ = j)
≤ Pm (∆Sj,m (τ ) ≥ − log c)
(49)
≤ cPj (∆Sj,m (τ ) ≥ − log c) ≤ c ,
where changing the measure in the second inequality follows
by the fact that ∆Sj,m (τ ) ≥ − log c. As a result,
X
αm =
αm,j ≤ (M − 1)c .
j6=m
Lemma 3: There exist constants C > 0 and γ > 0 such
that for any fixed 0 < q < 1, under any arbitrary policy, the
following statements hold:
Pm (Sj (n) ≥ Sm (n), Nj (n) ≥ qn) ≤ Ce−γn ,
(50)
Pm (Sj (n) ≥ Sm (n), Nm (n) ≥ qn) ≤ Ce−γn ,
(51)
and
for m = 1, 2, ..., M and j 6= m.
Pm (Sj (n) ≥ Sm (n), Nj (n) ≥ qn)
n
X
Pm
≤
n
X
r
X
`j (i) +
i=1
r=dqne k=0
n h
X
k
X
ir h
r=dqne
n
X
k=0
≤
∞
X X
Pm (Sj (t) ≥ Sm (t)) .
j6=m t=n
(55)
Following (55), it suffices to show that there exist C > 0 and
γ > 0 such that Pm (Sj (n) ≥ Sm (n)) ≤ Ce−γn .
(52)
Em es(−`m (1))
e−γ1 k ≤ Ce−γn .
maxj6=m D(fj ||gj )
,
minj6=m D(fj ||gj )
= arg min D(fj ||gj ),
km =
−`m (i) ≥ 0
ik
jm
ρm =
for all s > 0.
Since a moment generating function (MGF) is equal to one
at s = 0 and Em (`j (1)) = −D(fj ||gj ) < 0, Em (−`m (1)) =
−D(gm ||fm ) < 0 are strictly negative, differentiating the
MGFs of `j (1), `m (1) with respect to s yields strictly negative
derivatives at s = 0. As
a result, there exist
s > 0 and γ1 > 0
such that Em es`j (1) , Em es(−`m (1)) are strictly less than
e−γ1 < 1. Hence, there exist C > 0 and γ = γ1 · q > 0 such
that
Pm (Sj (n) − Sm (n) ≥ 0, Nj (n) ≥ qn)
e−γ1 r
Proof: We focus on proving for M > 2. Proving for
M = 2 is straightforward. Note that the event τ1 > n implies
that there exists a time instant t for t ≥ n in which for some
j 6= m, Sj (t) > Sm (t). Hence,
Pm (τ1 > n) ≤ Pm max sup (Sj (t) − Sm (t)) ≥ 0
We next establish the required exponential decay. Let
r=dqne k=0
n
X
(54)
for m = 1, 2, ..., M .
!
i=1
Em es`j (1)
≤
Lemma 4: If the selection rule of DGFi is implemented
indefinitely, there exist C > 0 and γ > 0 such that
j6=m t≥n
Proof: We start with proving (50). Note that
Nj (n), Nm (n) can take integer values Nj (n)
=
dqne, dqne + 1, ...n, and Nm (n) = 0, ..., n. Applying
the Chernoff bound and using the i.i.d. property of the
observations across time yield:
n
X
Note that τ1 is not a stopping time since it depends on the
future. τ1 is a last passage time. It is the last time in which
Sm (n) crosses Sj (n) for all j 6= m. In Lemma 4 we show that
the probability that τ1 is greater than n decreases exponentially
with n. This result will be used when evaluating the asymptotic
expected search time to show that it is not affected by τ1 .
Remark 1: We often analyze the dynamic of the sum LLRs
according to the selection rule of DGfi in the asymptotic
regime. Thus, when we say that the selection rule of DGFi
policy is implemented indefinitely we mean that we probe the
cells according to the selection rule of DGFi as given in (14)
indefinitely, while the stopping rule is disregarded.
Pm (τ1 > n) ≤ Ce−γn ,
Hence, (48) follows.
≤
For proving (51) we can use the Chernoff bound with minor
modifications.
Definition 1: τ1 is defined as the smallest integer such that
Sm (n) > Sj (n) for all j 6= m for all n ≥ τ1 .
(53)
j6=m
(56)
1
.
8(km + 1)(M − 2)
Note that 0 < ρm ≤ 1/16. Thus, we can write
Pm (Sj (n) ≥ Sm (n))
≤ Pm (Sj (n) ≥ Sm (n), Nj (n) < ρm n, Nm (n) < ρm n)
+Pm (Sj (n) ≥ Sm (n), Nj (n) ≥ ρm n)
+Pm (Sj (n) ≥ Sm (n), Nm (n) ≥ ρm n)
(57)
The second and the third terms on the RHS of (57) decay
exponentially with n by Lemma 3. Thus, it remains to show
that the first term decays exponentially with n as well. Note
that the event (Nj (n) < ρm n, Nm (n) < ρm n) implies that at
least ñ = n − Nj (n) − Nm (n) ≥ n (1 − 2ρm ) times cells j, m
10
er (n) as the number of times in
are not probed. We define N
which cell r 6= j, m has been probed and cells j, m have not
been probed by time n. There exists a cell r 6= j, m such that
er (n) ≥ ñ = n(1−2ρm ) . Hence, we can upper bound (57)
N
M −2
M −2
as follows:
Pm (Sj (n) ≥ Sm (n))
X
n(1 − 2ρm )
, Nj (n) < ρm n,
≤
Pm Ñr (n) >
M −2
r6=j,m
Nm (n) < ρm n + 2De−γ1 n
for all n0 ≤ ρm n. Therefore,
(58)
i=1
ζn+q
X
n0 =0
i=1
q=0
`r (i) +
i=1
=
ζn+q
X
i=1
−`j (i)
i=1
i=1
n
X
≤
i=1
`˜r (i) +
ζn+q
X
`˜r (i) +
n
X
−`˜j (i) ≥ C1 (n + q + n0 )
i=1
0
×e−sC1 (n+q+n )
h
˜
iζn+q
= Em es(`r (1)−C1 )
(65)
for all s > 0.
Since Em (`˜r (1)−C1 ) = −C1 < 0 and Em (−`˜j (1)−C1 ) =
−C1 < 0 are strictly negative, by applying a similar argument
as at the end of the proof
of Lemma
0 and
3, there
exist s >
(s`˜r (1)−C1 )
s(−`˜j (1)−C1 )
, Em e
and
γ2 > 0 such that Em e
e−sC1 are strictly less than e−γ2 < 1. Hence,
0
ζn+q
n
X
X
0
(66)
Pm
`j (i) ≤
`r (i) ≤ e−γ2 (n+q+n ) ,
i=1
0
n−ζn
mn
X ρX
Pm
n0 =0
(61)
in (56) and (59), we have
≤ e−γ2 n
0
n
X
`j (i) ≤
i=1
n−ζn
X
the ease of presentation, throughout the proof we assume that ζn,
ρm n are integers. This assumption does not affect the exponential decay
of the Chernoff bound but only the exact value of C > 0 in (54) (since
αn − 1 ≤ bαnc ≤ dαne ≤ αn + 1 holds for all α ≥ 0 for all n = 0, 1, ...).
ζn+q
X
`r (i)
i=1
e−γ2 q
ρ
mn
X
(67)
0
e−γ2 n ≤ C2 e−γ2 n ,
n0 =0
q=0
−2
4 For
0
i=1
−`˜j (i) − D(fjm ||gjm ) (ζn + q − km n0 ) ,
and by the definitions of ζ, km , ρm
(64)
h
iζn+q h
in0
˜
˜
≤ Em es`r (1)
Em es(−`j (1))
q=0
i=1
> 0.
and
i=1
n0
X
4(M − 2)
i=1
i=1
−`˜j (i)
−D(fr ||gr ) (ζn + q) + D(fn0 ||gn0 )n
ζn+q
X
D(fjm ||gjm )
Applying the Chernoff bound yields:
0
ζn+q
n
X
X
Pm
`j (i) ≤
`r (i)
0
`˜r (i) +
(63)
in0
h
˜
× e−sC1 (n−ζn) .
× Em es(−`j (1)−C1 )
0
n
X
(62)
i=1
≤ Pm
(60)
For upper bounding the first term on the RHS of (60) we write
the sum LLRs as follows:
ζn+q
X
−`˜j (i) ≥ C1 (n + q + n0 ) ,
C1 =
i=1
i=1
`˜r (i) +
n
X
where
Sr (t̃rζn )
4
0
n−ζn
ζn+q
n
mn
X ρX
X
X
+
Pm
`m (i) ≤
`r (i) .
n0 =0
−`j (i) ≥ 0
i=1
0
i=1
(59)
Sm (t̃rζn )
n
X
implies
Note that the event
≤
or
≤
must occur (otherwise, cell j or m will be probed). Hence ,
m)
Pm Ñr (n) > n(1−2ρ
,
N
(n)
<
ρ
n,
N
(n)
<
ρ
n
j
m
m
m
M −2
0
n−ζn
ζn+q
n
mn
X ρX
X
X
=
Pm
`j (i) ≤
`r (i)
q=0
`r (i) +
i=1
1 − 2ρm
.
ζ,
2(M − 2)
Sr (t̃rζn )
0
ζn+q
X
where the second and third terms on the RHS of (57) are
upper bounded by De−γ1 n (there exist such D > 0, γ1 > 0
by Lemma 3), and the first term on the RHS of (57) is upper
bounded by the first term (i.e., the summation term) on the
RHS of (58). Next, we show that each term in the summation
decays exponentially with n to get the desired result.
Let t̃r1 , t̃r2 , ..., t̃rÑ (n) be the indices for the time instants in
r
which cell r 6= j, m has been probed and cells j, m have not
been probed by time n. Let
Sj (t̃rζn )
ζn + q − km n0 ≥ ζn + q − km n0 − (km + 1) (ρm n − n0 )
1
n + q + n0
= n (ζ − (km + 1)ρm ) + q + n0 ≥
4(M − 2)
1
≥
(n + q + n0 ) ,
4(M − 2)
where C2 = (1 − e−γ2 )
.
A similar Chernoff bounding technique can be applied to
upper bound the second term on the RHS of (60).
11
Next, we consider two cases:
1) the balanced case, referring to the case when
1
minj6=m D(fj ||gj ) ≥ P
;
j6=m 1/D(fj ||gj )
2) the unbalanced case, when the above inequality is reversed.
The reason for referring to the first case as the balanced
case is that DGFi policy is able to balance the detection
time so that the difference between the largest sum LLR
and the sum LLRs of any other cell exceeds the threshold
− log c approximately at the same time. As a result, the rate
function is determined by a certain averaging among the KL
divergences of the heterogeneous processes. On the other hand,
when the smallest KL divergence is too small, then too many
measurements are required to be gathered from that cell. In
that case, the difference between the largest sum LLR and
the sum LLR gathered from the cell with the smallest KL
divergence exceeds the threshold − log c significantly after the
difference between the largest sum LLR and the sum LLR
gathered from any other cell does. As a result, the rate function
is dominated by the smallest KL divergence.
For the unbalanced case, the proof follows directly from
subsection B.2. Thus, here it remains to show the proof for
the balanced case.
2: τ2 is defined as the smallest integer such that
PDefinition
n
i=τ1 +1 `jn (i)1jn (i) ≤ log c for some jn 6= m for all n ≥
τ2 ≥ τ1 .
We also define n2 , τ2 − τ1 as the total amount of time
between τ1 and τ2 .
In Lemma 5 we show that the probability that n2 is greater
than n decays exponentially with n when n is greater than
− log c/Im . Later, we will show that the asymptotic search
time is dominated by n2 , which together with Lemma 5 yields
the desired search time − log c/Im under hypothesis Hm .
Lemma 5: If the selection rule of DGFi is implemented
indefinitely, then for every fixed > 0 there exist C > 0 and
γ > 0 such that
Pm (n2 > n) ≤ Ce−γn
∀n > −(1 + ) log c/Im ,
(68)
for all m = 1, 2, ..., M .
Proof: First, we consider the case where Im >
D(gm ||fm ). Note that cell m isPnot observed for all n ≥ τ1 in
τ1 +t
1 (i) and j ∗ (τ1 +t) =
this case. Define Nj0 (τ1 +t) = i=τ
1 +1 j
0
arg maxj Nj (τ1 + t)D(fj ||gj ). Thus,
(69)
t≥n i=τ +1
1
Since Kt is the total number of obeservation from τ1 to τ1 +t,
by the definition of j ∗ (t) we have
Kt =
j6=m
≤
j6=m
`j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i) − log c
i=τ1 +1
τX
1 +t
`˜j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i)
=
i=τ1 +1
−Nj0 ∗ (τ1 +t) (τ1 + t)D(fj ∗ (τ1 +t) ||gj ∗ (τ1 +t) ) − log c
τX
1 +t
≤
`˜j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i)
i=τ1 +1
Kt
− log c
1/D(f
j ||gj )
j6=m
−P
≤
≤
τX
1 +t
i=τ1 +1
τX
1 +t
`˜j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i) − tIm + tIm /(1 + )
`˜j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i) − t1
i=τ1 +1
(72)
for all t ≥ n > −(1 + ) log c/Im . By applying the Chernoff
bound,
Pτit+tcan be shown that there exists γ1 > 0 such that
Pm ( τ11 +1 −`˜j ∗ (τ1 +t) (i) ≥ t1 ) < e−γ1 t for all t ≥ n >
−(1 + ) log c/Im . Hence, there exist C1 > 0 and γ1 > 0 such
that Pm (n2 > n) ≤ C1 e−γ1 n for all n > −(1 + ) log c/Im .
A similar argument applies for case where Im ≤ D(gm ||fm ).
In what follows we define a r.v. DR(t) as the dynamic
range between sum LLRs of empty cells, maxj6=m Sj (t) −
minj6=m Sj (t). Note that the dynamic range at time τ2 (which
is the time in which sufficient information has been gathered
to distinguish Hm from at least one false hypothesis) can be
viewed as a measure of the amount of information remains
to gather in order to distinguish Hm from any other false
hypothesis. Lemma 6 below shows that the dynamic range at
time τ2 is sufficiently small.
Definition 3: The dynamic range between sum LLRs of
empty cells at time t is defined as follows:
DR(t) , max Sj (t) − min Sj (t) .
j6=m
(73)
Lemma 6: If the selection rule of DGFi is implemented
indefinitely. Then, for every fixed 1 > 0, 2 > 0 there exist
C > 0 and γ > 0 such that
Pm (DR(τ2 ) > 1 n) ≤ Ce−γn
(74)
for all m = 1, 2, ..., M .
j6=m
D(fj ||gj )
τX
1 +t
∀n > −(1 + 2 ) log c/Im ,
X Nj0 (τ1 + t)D(fj ||gj )
Nj0 (τ1 + t) =
D(fj ||gj )
X Nj0 ∗ (τ1 +t) (τ1 + t)D(fj ∗ (τ1 +t) ||gj ∗ (τ1 +t) )
Then,
j6=m
Pm (n2 > n)
τX
1 +t
`j ∗ (τ1 +t) (i)1j ∗ (τ1 +t) (i) ≥ log c).
≤ Pm (sup
X
P
Let 1 = Im /(1 + ). Since Im = j6=m K/D(fj ||gj ), we
have
K
P
.
(71)
1 =
(1 + ) j6=m 1/D(fj ||gj )
(70)
.
Proof: The proof follows directly by applying Lemma 5
and substituting τ2 in subsection B.2.
12
Definition 4: τ3j is defined as the smallest integer such that
Sj (n) ≥ − log c for all n ≥ τ2 . We also define τ3 = maxj τ3j .
Note that by the definitions of the last passage times we
have τ3j ≥ τ2 .
Remark 2: Using some algebraic manipulations, it can
be verified that ∆Sm,j (n) ≥ − log c for all j 6= m for
all n ≥ τ3j . Since τ3 = maxj6=m τ3j we have ∆S(n) =
Sm (n) − Sm(2) (n) (n) ≥ − log c for all n ≥ τ3 . Recall that
DGFi stops the test once ∆S(n) first occurs. Thus, in the
sequel we will use τ3 as an upper bound on the actual stopping
time τ .
Definition 5: n3 , τ3 − τ2 is defined as the total amount of
time between τ2 and τ3 .
In Lemma 7 we show that n3 is sufficiently small with high
probability. We will use this result to show that the asymptotic
expected search time is not affected by n3 .
Lemma 7: If the selection rule of DGFi is implemented
indefinitely, then for every fixed > 0 there exist C > 0 and
γ > 0 such that
Pm (n3 > n) ≤ Ce−γn
∀n > − log c/Im ,
(75)
for all m = 1, 2, ..., M .
Proof: To prove the Lemma, we first define N3j as the
total number of observations that the decisionP
maker collected
from cell j between τ2 and τ3j . Since n3 ≤ j N3j , we only
need to show that Pm (N3j > n) decays exponentially with n.
We can write Pm (N3j > n) as follows:
minj D(fj ||gj )
)
2
(76)
minj D(fj ||gj )
j
)
+Pm (N3 > n|DR(τ2 ) ≤ n
2
Lemma 6 provides the desired decay for the first term on the
RHS. We next show the desired decay for the second term.
Let t1 , t2 , · · · denote the time indices when cell j is observed
between τ2 and τ3j . We can write:
P(N3j > n) ≤ Pm (DR(τ2 ) > n
minj D(fj ||gj )
Pm (N3j > n|DR(τ2 ) ≤ n
)
2
r
X
minj D(fj ||gj )
≤ Pm ( inf
−`j (ti ) < n
)
r>n
2
i=1
r
X
minj D(fj ||gj )
≤ Pm (
`˜j (ti ) > r
).
2
i=1
(77)
Em (τ ) ≤ − (1 + o(1))
log(c)
,
Im
(78)
for some C3 , γ3 which completes the proof.
The following Lemma provides an upper bound on the
detection time when DGFi policy implemented (i.e., the cells
are probed based on the selection rule and the test terminates
based on the stopping rule).
(79)
for m = 1, ..., M .
Proof: Since the actual detection time under DGFi is
upper bounded by: τ ≤ τ3 = τ1 + n2 + n3 , combining
Lemmas 4, 5 and 7 proves the statement.
B. Analyzing the Dynamic of Empty Cells under DGFi
In this appendix we analyze the sum LLRs dynamics
at empty cells under DGFi, used to prove the theorem in
subsection A. To analyze the sum LLR of empty cells, we
introduce the following (slightly different) active hypothesis
testing problem. It should be noted that in what follows we
slightly change the notations for the new problem setting for
convenience and differentiating it from the original problem.
At each time, only k cells can be observed from cells
1, 2, · · · , m. When cell j is observed at time n, an observation xj (n) is drawn independently from previous times and
xj (n) follows distribution fj . We also assume that gj , j =
1, 2, · · · , m are m known distributions.
For the ease of the presentation when analyzing the sum
LLRs of empty cells. we remove the subscript m from Pm ()
when referring to the probability measure.
Let 1j (n) be the indicator function, where 1j (n) = 1 if cell
j is observed at time n, and 1j (n) = 0 otherwise.
Let
fj (xj (n))
(80)
lj (n) , log
gj (xj (n))
and
Sj (n) ,
n
X
lj (t)1j (t)
(81)
t=1
Note that we are now focusing on empty cells. Thus, for
convenience the LLR is defined as negative LLR defined in
the original problem. The sum LLR is defined accordingly.
We know that
Z
fj (x)
E[lj (n)] =
fj (x) log
dx = D(fj ||gj ).
(82)
gj (x)
x
Thus, here the sum LLR of an empty cell j is a random walk
with a positive increment D(fj ||gj ). Similarly, we define
˜lj (n) , lj (n) − D(fj ||gj ),
Using the Chernoff bound and the i.i.d. property of `˜j (ti )
yields:
n
X
minj D(fj ||gj )
Pm (
`˜j (ti ) > n
) < C3 e−γn
2
i=1
Lemma 8: If DGFi policy is implemented, then the expected
detection time τ is upper bounded by:
(83)
which is a zero mean r.v.. Without loss of generality, we
assume D(f1 ||g1 ) ≤ D(f2 ||g2 ) ≤ · · · ≤ D(fm ||gm ). We also
define
1
v̄ , P
.
(84)
1/D(f
j ||gj )
j
Let r(i) (n) denotes the cell index with the ith smallest sum
LLR collected from this cell up to time n. We define the
following random variables:
U (n) , max Sj (n), L(n) , min Sj (n),
j
j
(85)
13
DRjk (n) , Sr(k) (n) (n) − Sr(j) (n) (n),
(86)
DR(n) , DR1m (n) = U (n) − L(n),
(87)
and
Since kt is the total number of observations by time t, by the
definition of j(t) we have
kt =
Nj (t) =
1j (i),
Nj (t) =
X Nj (t)D(fj ||gj )
D(fj ||gj )
j
j
X Nj(t) (t)D(fj(t) ||gj(t) )
≥
.
D(fj ||gj )
j
Also, let
t
X
X
(88)
(94)
i=1
Hence,
j(t) = arg min Nj (t)D(fj ||gj ),
(89)
j̄(t) = arg max Nj (t)D(fj ||gj ).
(90)
j
j
1
= t · kv̄. (95)
1/D(f
j ||gj )
j
Nj(t) (t)D(fj(t) ||gj(t) ) ≤ kt · P
Therefore,
Sj(t) (t) > t · kv̄ + n
Remark 3: Note that we defined the LLR in this appendix as
the negative LLR which was defined in the original problem.
Thus, the corresponding DGFi policy in this appendix chooses
the k cells with smallest sum LLRs (in contrast to the selection
of the empty cells with the top sum LLRs as done in the
original problem).
Definition 6: The modified selection rule of DGFi for the
active hypothesis testing problem defined in this appendix is
given by: φ(n) = {r(1) (n), r(2) (n), · · · , r(k) (n)}.
Next, we provide the outline for the next lemmas. Lemma 9
states that the smallest observed sum LLR is sufficiently small
as requireed with high probability. Lemma 10 states that the
largest observed sum LLR is sufficiently large as required with
high probability. Lemma 11 shows that under (the modified)
DGFi policy, the difference between the largest sum LLR and
the (m − k + 1)th largest sum LLR is sufficiently small as
required with high probability. Whether the smallest sum LLR
is approximately equal to the largest sum LLR depends on
which one of balanced or unbalanced cases is valid. For the
balanced case, Lemma 12 claims that the dynamic range is
small under DGFi policy. Hence, DGFi can balance the search
time among all the processes so that the search time is a
certain averaging of their KL divergences. For the unbalanced
case, Lemma 13 states that the sum LLRs of the cell with
the smallest KL divergence cannot be too small (which will
determine the rate function function for the search in this case)
with high probability. Lemma 14 shows that the sum LLR of
other cells are larger than that of the cell with the smallest KL
divergence. Finally, Lemma 15 upper bounds the asymptotic
search time.
Lemma 9: For any selection rule, ∀t, ∀ > 0, there exist
C, γ > 0 such that
P(L(t) > t · kv̄ + n) < Ce
−γn
∀n > t.
(96)
implies
t
X
˜lj(t) (i)1j(t) (i) > n.
(97)
i=1
Since ˜lj(t) (t) is a zero mean r.v. with a bounded moment generating function, applying the Chernoff inequality completes
the proof.
Lemma 10: For any selection rule, ∀t, ∀, there exist C, γ >
0 such that
P(U (t) < t · kv̄ − n) < Ce−γn
∀n > t.
(98)
Proof: Note that
P(U (t) > t · kv̄ − n) ≤ P(Sj̄(t) (t) < t · kv̄ − n),
(99)
and
Sj̄(t) (t) = Nj̄(t) (t)D(fj̄(t) ||gj̄(t) )+
t
X
˜lj̄(t) (i)1j̄(t) (i). (100)
i=1
Since kt is the total number of observations by time t, by the
definition of j̄(t) we have
X
X Nj (t)D(fj ||gj )
kt =
Nj (t) =
D(fj ||gj )
j
j
(101)
X Nj̄(t) (t)D(fj̄(t) ||gj̄(t) )
.
≤
D(fj ||gj )
j
Hence,
1
= t·kv̄. (102)
j 1/D(fj ||gj )
Nj̄(t) (t)D(fj̄(t) ||gj̄(t) ) ≥ kt· P
Therefore,
Sj̄(t) (t) < t · kv̄ − n
(91)
(103)
implies
t
X
Proof: Note that
P(L(t) > t · kv̄ + n) ≤ P(Sj(t) (t) > t · kv̄ + n),
(92)
and
Sj(t) (t) = Nj(t) (t)D(fj(t) ||gj(t) ) +
t
X
i=1
˜lj(t) (i)1j(t) (i). (93)
˜lj(t) (i)1j(t) (i) < −n.
(104)
i=1
Since ˜lj(t) (t) is a zero mean r.v. with a bounded moment generating function, applying the Chernoff inequality completes
the proof.
14
Lemma 11: For DGFi selection rule, ∀t, ∀, there exist
C, γ > 0 such that
P(DRkm (t)
> D(fm ||gm ) + n) < Ce
−γn
∀n > t. (105)
Proof: We prove by induction with respect to t. When
t = 1, DRkm (1) > D(fm ||gm ) + n indicates that
3) ∀t, ∀, there exist C, γ > 0 such that
P(DR(t) > k · D(fm ||gm ) + n) < Ce−γn
∀n > t.
(113)
Proof: We prove by induction with respect to k. For k =
1, statement 3 follows directly from Lemma 11. For statement
1,
P(U (t) > t · v̄ + D(fm ||gm ) + n)
D(fm ||gm ) + n < lj (1) = ˜lj (1) + D(fj ||gj )
≤ ˜lj (1) + D(fm ||gm )
(106)
≤P(L(t) > t · v̄ + n or DR(t) > D(fm ||gm ) + n)
≤P(L(t) > t · v̄ + n) + P(DR(t) > D(fm ||gm ) + n),
(114)
which indicates that
˜lj (1) > n.
Using the Chernoff bound completes the induction base.
If the statement is true for t − 1, then for t we have
P(DRkm (t) > D(fm ||gm ) + n)
=P(DRkm (t) > D(fm ||gm ) + n, r(m) (t) = r(m) (t − 1))
+ P(DRkm (t) > D(fm ||gm ) + n, r(m) (t) 6= r(m) (t − 1)).
(107)
which can be bounded by Lemma 9 and 11. Similarly, we can
prove statement 2 using Lemmas 10 and 11.
If the statement is true for k − 1, for k we first prove
statement 3. For any fixed t, let j̄ = arg minj Sj (t), and let
t0 be the smallest integer such that j̄ ∈ φ(τ ), ∀t0 < τ ≤ t.
From Lemma 11 it follows that ∀, there exist C, γ > 0 such
that
P(DRkm (t0 ) > D(fm ||gm ) + n) < Ce−γn ,
∀n > t0 .
Since j̄ ∈
/ φ(t0 ), we have:
For the first term on the RHS, we have
P(DRkm (t) > D(fm ||gm ) + n, r(m) (t) = r(m) (t − 1))
n
≤ P(DRkm (t − 1) > D(fm ||gm ) + , r(m) (t) = r(m) (t − 1)
2
n
or lr(k) (t−1) (t) < − , r(m) (t) = r(m) (t − 1))
2
n
≤ P(DRkm (t − 1) > D(fm ||gm ) + )
2
n
+P(lr(k) (t−1) (t) < − ) ≤ C1 e−γ1 n ,
2
(108)
where the first term can be bounded using assumptions on
t − 1 and the second term can be bounded using the Chernoff
bound.
P(Sj̄ (t0 ) − U (t0 ) < −D(fm ||gm ) − n) < Ce−γn , ∀n > t0 .
(115)
Since that j̄ ∈ φ(τ ), ∀t0 < τ ≤ t we have:
P(Sj̄ (t) − U (t0 ) < (t − t0 )D(fj̄ ||gj̄ ) − D(fm ||gm ) − n)
< Ce−γn ,
∀n > t
(116)
by applying the Chernoff bound.
Next, consider a subproblem where cell j̄ is removed and
only k − 1 cells can be selected. Let Sj0 (n) be the sum LLR
in this subproblem. Then, by statement 1 with assumption on
k − 1 we have that ∀, there exist C, γ > 0 such that
For the second term on the RHS of (107), we have
P (U (t − t0 ) > (t − t0 ) · (k − 1)v̄ 0
P(DRkm (t) > D(fm ||gm ) + n, r(m) (t) 6= r(m) (t − 1))
≤P(lr(m) (t) (t) > D(fm ||gm ) + n)
≤P(˜lr(m) (t) (t) > n) < C2 e−γ2 n
+ (k − 1)D(fm ||gm ) + n)
(117)
< Ce−γn , ∀n > t − t0
(109)
where
1
.
1/D(f
j ||gj )
j6=j̄
v̄ 0 = P
Combining (107), (108), (109) completes the proof.
Now, for the original problem, we have Sj (t) = Sj (t0 ) +
Sj0 (t − t0 ) ≤ U (t0 ) + Sj0 (t − t0 ). By (117) we have
1) The Balanced Case:
Lemma 12: Under the DGFi selection rule, if
D(f1 ||g1 ) ≥ k · v̄
P(U (t) − U (t0 ) > (t − t0 ) · (k − 1)v̄ 0
+(k − 1)D(fm ||gm ) + n)
(110)
holds, then we have the following statements:
1) ∀t, ∀, there exist C, γ > 0 such that
P(U (t) > t·kv̄ +kD(fm ||gm )+n) < Ce−γn
(118)
∀n > t
since D(f1 ||g1 ) ≥ kv̄ implies D(fj̄ ||gj̄ ) ≥ (k−1)v̄ 0 . By (116)
and (118) we have that ∀, there exist C, γ > 0 such that
∀n > t.
(111)
2) ∀t, ∀, there exist C, γ > 0 such that
P(L(t) < t · kv̄ − kD(fm ||gm ) − n) < Ce−γn
< Ce
−γn
∀n > t.
(112)
P(U (t) − Sj̄ (t) > k · D(fm ||gm ) + n) < Ce−γn
∀n > t,
(119)
which proves statement 3 for k as desired. Then, statements
1 and 2 can be proved using Lemma 9 and 10 with statement
3 similar to the case with k = 1.
15
which indicates that the LHS has positive means. By applying
the Chernoff bound and using the i.i.d. property of ˜lj (ti ) we
have:
2) The Unbalanced Case:
Lemma 13: Under the DGFi selection rule, if
D(f1 ||g1 ) < k · v̄
(120)
then ∀t, ∀, there exist C, γ > 0 such that
P(S1 (t) < tD(f1 ||g1 ) − n) < Ce−γn
∀n > t.
(121)
Proof: Define t0 as the smallest integer such that cell 1
is observed at time i for all t0 < i ≤ t. Then, by our selection
rule, cell 1 is the one of the top m − k sum LLRs at time t0 .
Then, by applying t = t0 to Lemma 11 we have
P(U (t0 ) − S1 (t0 ) > n) < C1 e−γ1 n
∀n > t0
(122)
for some C1 , γ1 . Substituting t = t0 in Lemma 10 we have:
P(U (t0 ) < t0 · kv̄ − n) < C2 e−γ2 n
∀n > t0
(123)
∀n > t0
(124)
for some C2 , γ2 . Hence,
P(S1 (t0 ) < t0 · kv̄ − n) < C3 e−γ3 n
for some C3 , γ3 . Then, by the definition of t0 and using the
Chernoff bound we have
P(S1 (t) − S1 (t0 ) < (t − t0 )D(f1 ||g1 ) − n) < C4 e−γ4 n
∀n > (t − t0 ).
(125)
P(S1 (t) < tD(f1 ||g1 ) − n) < C5 e−γ5 n
∀n > t
Lemma 14: For every fixed > 0, there exists C > 0 and
γ > 0, such that for all j we have:
P(S1 (τ̃2 ) − Sj (τ̃2 ) > n) ≤ Ce−γn ,
∀n > τ̃2 .
(127)
tj0
Proof: For fixed j, define
as the smallest integer
such that S1 (n) > Sj (n) for all tj0 < i ≤ τ̃2 . By definition, S1 (tj0 ) ≤ Sj (tj0 ). Then, by our selection rule, for
all tj0 < i ≤ τ̃2 , whenever cell 1 is observed, cell j must
be observed based on their ranking of sum LLRs. Note that
D(f1 ||g1 ) ≤ D(fj ||gj ). Thus,
lj (i)1j (i) −
τ̃2
X
l1 (i)11 (i)
˜lj (i)1j (i) −
˜l1 (i)11 (i)
τ̃2
X
1j (i) − D(f1 ||g1 )
i=tj0
i=tj0
,
(130)
∀n > τ̃2
as desired.
Definition 8: Define τ̃3j as the smallest integer such that
Sj (n) ≥ − log c for all n ≥ τ̃3j . We also define τ̃3 = maxj τ̃3j .
Definition 9: ñ3 , τ̃3 − τ̃2 denotes the total amount of time
between τ̃2 and τ̃3 .
Lemma 15: For every fixed > 0, there exists C > 0 and
γ > 0 such that
P(ñ3 > n) < Ce−γn ,
∀n > − log c/D(f1 ||g1 ). (131)
Proof: By substituting t = τ̃2 in Lemma 13 we have:
P(S1 (τ̃2 ) < − log c − n) < C1 e−γ1 n
∀n > τ̃2
(132)
for some C1 , γ1 . By applying Lemma 14, we have:
∀n > τ̃2 , j = 1, 2, · · · , m
(133)
˜lj (i)1j (i) −
Let N3j denote that total number of observations,
taken from
P j
cell j between τ̃2 and τ̃3j . Since ñ3 ≤
N3 , it suffices to
show that P(N3j > n) decays exponentially with n. Note that
P(N3j > n)
D(f1 ||g1 )
≤ P Sj (τ̃2 ) < − log c − n
2
D(f1 ||g1 )
+P N3j > n|Sj (τ̃2 ) ≥ − log c − n
.
2
τ̃2
X
i=tj0
τ̃2
X
i=tj0
˜l1 (i)11 (i),
(128)
11 (i)
(134)
By (133) it remains to show that the second term decays
exponentially with n. Let t1 , t2 , · · · denote the time indices
when cell j is observed between τ̃2 and τ̃3j . Then,
r
X
D(f1 ||g1 )
2
i=1
!
r
X
D(f
||g
)
1
1
˜lj (ti ) > r
≤P
.
2
i=1
≤P
i=tj0
+D(fj ||gj )
τ̃2
X
≤ Ce
P(N3j > n|Sj (τ̃2 ) ≥ − log c − n
i=tj0
τ̃2
X
i=tj0
≥
−γn
for some C2 , γ2 > 0.
Definition 7: Define τ̃2 = − log c/D(f1 ||g1 ).
=
≤ P(S1 (τ̃2 ) − S1 (tj0 ) − (Sj (τ̃2 ) − Sj (tj0 )) > n)
(126)
as desired.
i=tj0
τ̃2
X
P(S1 (τ̃2 ) − Sj (τ̃2 ) > n)
P(Sj (τ̃2 ) < − log c − n) < C2 e−γ2 n
Since kv̄ > D(f1 ||g1 ), we have:
τ̃2
X
P(S1 (τ̃2 ) − S1 (tj0 ) − (Sj (τ̃2 ) − Sj (tj0 )) > n) ≤ Ce−γn ,
(129)
for some C, γ. Since S1 (tj0 ) ≤ Sj (tj0 ), we have:
inf
r>n
D(f1 ||g1 )
)
! 2
lj (ti ) < n
Applying the Chernoff bound and using the i.i.d. property of
˜lj (ti ) across time we have
!
r
X
D(f
||g
)
1
1
˜lj (ti ) > r
P
< C3 e−γn
(135)
2
i=1
16
is monotonically increasing with t for 0 ≤ t ≤ n.
Proof: Differentiation d(t) with respect to t yields:
for some C3 , γ3 which completes the proof.
C. The Asymptotic Lower Bound on the Bayes Risk
In this appendix, we show that the asymptotic Bayes risk
that can be achieved by any policy is lower bounded by
−c log c/I ∗ .
Lemma 16: Assume that αj (Γ) = O(−c log c) for all j =
1, ..., M . Let 0 < < 1. Then:
∂d(t)
1
= D(gm ||fm ) − P
≥0,
1
∂t
j6=m D(fj ||gj )
which completes the proof.
For the next lemma we define
j ∗ (t) , arg min Nj (t)D(fj ||gj ),
Pm (∆Sm (τ ) < − (1 − ) log c | Γ) = O(−c log c) , (136)
j6=m
(144)
and
for all m = 1, ..., M .
Proof: Note that:
∗
Wm
(t) ,
Pm (∆Sm (τ ) < − (1 − ) log c|Γ)
t
X
`˜m (i)1m (i) −
i=1
t
X
`˜j ∗ (t) (i)1j ∗ (t) (i),
(145)
i=1
= Pm (∆Sm (τ ) < − (1 − ) log c , δ = m|Γ)
which is a sum of zero-mean r.v.
+Pm (∆Sm (τ ) < − (1 − ) log c , δ 6= m|Γ)
Lemma 18: For every fixed > 0 there exist C > 0 and
γ > 0 such that
∗
Pm max Wm
(t) ≥ n|Γ ≤ Ce−γn
(146)
≤ Pm (∆Sm (τ ) < − (1 − ) log c , δ = m|Γ) + αm (Γ),
(137)
where αm (Γ) = O(−c log c) by assumption. In what follows,
we upper bound
Pm (∆Sm (τ ) < − (1 − ) log c , δ = m|Γ) .
Similar to [2, Lemma 4] we can show that for all j 6= m
there exists G > 0 such that:
−Gc log c ≥ Pj (δ 6= j|Γ) ≥ Pj (δ = m|Γ)
≥ Pj (∆Sm,j (τ ) ≤ −(1 − ) log c , δ = m|Γ)
≥ c1− Pm (∆Sm,j (τ ) < − (1 − ) log c , δ = m|Γ) ,
(138)
where the last inequality holds by changing the measure as in
[2, Lemma 4]. Thus,
1≤t≤n
for all m = 1, ..., M and for any policy Γ.
Proof: We upper bound (146) by summing over any
possible values that Nm (t), Nj ∗ (t) (t) can take and using the
Chernoff bound:
∗
Pm max Wm (t) ≥ n|Γ
1≤t≤n
=
t X
t
n X
X
t=1 i=0 j=0
Pm
t
X
`˜m (r)1m (r) +
r=1
−`˜j ∗ (t) (r)1j ∗ (t) (r) ≥ n,
r=1
Nm (t) = i, Nj ∗ (t) = j|Γ
Pm (∆Sm,j (τ ) < − (1 − ) log c , δ = m|Γ)
= O (−c log c) ∀j 6= m .
t
X
(139)
t h
t X
n X
ii
X
˜
Em es(`m (1)−/2)
≤
×
As a result,
t=1 i=0 j=0
h
Pm (∆Sm (τ ) < − (1 − ) log c , δ = m|Γ)
X
≤
Pm (∆Sm,j (τ ) < − (1 − ) log c , δ = m|Γ)
j6=m
= O (−c log c) .
(140)
Finally,
Pm (∆Sm (τ ) < − (1 − ) log c|Γ) = O (−c log c) .
(141)
ij
˜
Em es(−`j∗ (t) (1)−/2)
×
o
n
exp −s (2n − i − j) ,
2
(147)
for all s > 0.
Since Em (`˜m (1) − /2) = −/2 < 0 and Em (−`˜j ∗ (t) (1) −
/2) = −/2 < 0 are strictly negative, using a similar
argument as at the end of the proof of Lemma
3, there
˜
exist s > 0 and γ 0 > 0 such that Em es(`m (1)−/2) ,
˜
Em es(−`j∗ (t) (1)−/2) and e−s/2 are strictly less than
0
e−γ < 1. Since 2n − i − j ≥ 0, there exist C > 0 and
γ > 0, such that summing over t, i, j yields (146).
Lemma 17: Assume that
1
D(gm ||fm ) ≥ P
1
j6=m D(fj ||gj )
Then, the function:
"
.
(142)
Lemma 19: For any fixed > 0,
Pm max ∆Sm (t) ≥ n (Im + ) | Γ → 0
1≤t≤n
d(t) , t D(gm ||fm ) + P
n
t
−1
1
j6=m D(fj ||gj )
#
(143)
as n → ∞ ,
(148)
17
for all m = 1, ..., M and for any policy Γ.
Proof: We next show exponential decay of (148) (which
is stronger than the polynomial decay shown under the binary
composite hypothesis testing case in [2, Lemma 5]). Let
∗
∆Sm
(t) , Sm (t) − Sj ∗ (t) (t).
Proof: For any > 0 let nc = −(1 − )
log c
. Note that
Im +
Pm (τ ≤ nc | Γ)
= Pm (τ ≤ nc , ∆Sm (τ ) ≥ − (1 − ) log c | Γ)
∗
∆Sm
(t)
Note that ∆Sm (t) ≤
for all m and t. As a result,
Pm max ∆Sm (t) ≥ n (Im + ) |Γ
1≤t≤n
(149)
∗
≤ Pm max ∆Sm (t) ≥ n (Im + ) |Γ .
+Pm (τ ≤ nc , ∆Sm (τ ) < − (1 − ) log c | Γ)
≤ Pm max ∆Sm (t) ≥ − (1 − ) log c | Γ
(155)
t≤nc
+Pm (∆Sm (τ ) < − (1 − ) log c | Γ) .
1≤t≤n
We next prove the lemma for the case where Im = Fm (K).
Proving the lemma for the cases where Im = D(gm ||fm ) +
Fm (K − 1) applies with minor modifications.
Note that:
∗
∗
∆Sm
(t) = Wm
(t) + Nm (t)D(gm ||fm )
Both terms on the RHS approaches zero as c → 0 by
Lemmas 16, 19. Hence,
Em (τ |Γ) ≥
∞
X
nPm (τ = n|Γ)
(156)
n=nc +1
≥ nc Pm (τ ≥ nc + 1|Γ) → nc as c → 0
+Nj ∗ (t) (t)D(fj ∗ (t) ||gj ∗ (t) )
1
1/D(f
j ||gj )
j6=m
∗
≤ Wm
(t) + Nm (t) · P
(150)
Since > 0 is arbitrarily small we have Em (τ |Γ) ≥
− (1 + o(1)) log(c)/Im . As a result, Rm (Γ) ≥ cEm (τ |Γ) ≥
− (1 + o(1)) c log(c)/Im .
+Nj ∗ (t) (t)D(fj ∗ (t) ||gj ∗ (t) ).
Since that j ∗ (t) = arg minj6=m Nj (t)D(fj ||gj ) and Kt −
Nm (t) is the total number of observations taken from M − 1
cells j 6= m, we have:
X Nj ∗ (t) D(fj ∗ (t) ||gj ∗ (t) )
≤ Kt − Nm (t) ≤ Kn − Nm (t).
D(fj ||gj )
j6=m
(151)
Hence,
∗
∗
(t) + Kn P
(t) ≤ Wm
∆Sm
j6=m
1
1/D(fj ||gj )
∗
(t) + nIm .
= Wm
(152)
Therefore,
∗
(t) ≥ n (Im + )
∆Sm
implies
∗
Wm
(t) ≥ n.
By Lemma 18 we have:
Pm max ∆Sm (t) ≥ n (Im + )
1≤t≤n
∗
≤ Pm max Wm (t) ≥ n
(153)
1≤t≤n
Finally, we show that the Bayes risk cannot be made smaller
than −c Ilog(c)
:
m
Lemma 20: Any policy Γ that satisfies Rj (Γ) = O(−c log c)
for all j = 1, ..., M must satisfy:
for all m = 1, ..., M .
c log(c)
.
Im
We now focus on proving asymptotic optimality for L > 1,
and K = 1. For L > 1, we define τ1 as the smallest integer
such that Sm (n) > Sj (n) for all m ∈ D, j 6= D and n ≥ τ1 .
Note that when K = 1 and n ≥ τ1 the decision maker always
probe the consistent cell (target or not depending on the order
of GD and FD ) for making the difference between the Lth
and (L + 1)th largest sum LLRs greater than the threshold
− log c. As a result, the decision maker can always balance
the detection time so that the difference between the largest
sum LLR and the sum LLRs of any other cell exceeds the
threshold − log c approximately at the same time as c → 0.
Thus, proving the asymptotic optimality of DGFi for L > 1
and K = 1 follows similar arguments as in the balanced case
in the proof of Theorem 1 given in Appendix B, and we focus
here only on the key modifications. Let
∆SD (n) ,
≤ Ce−γn → 0 as n → ∞.
Rm (Γ) ≥ − (1 + o(1))
A PPENDIX C: PROOF OF T HEOREM 2
(154)
min
m∈D,j ∈D
/
∆Sm,j (n),
(157)
where ∆Sm,j (n) is defined in (44). Without loss of generality
we prove the theorem when set D contains all the targets. We
define
`k (i) − D(gk ||fk ) , if k ∈ D,
`˜k (i) =
(158)
`k (i) + D(fk ||gk ) , if k ∈
/ D,
which is a zero-mean r.v.
We start with sowing the upper bound on the Bayes risk
obtained by DGFi. Similar to Lemma 2, we can show that
the error probability under DGFi is O(c). Specifically, we can
show that the error probability is upper bounded by:
Pe ≤ (M − L)L · c .
(159)
18
We can show this by letting αD = PD (δ 6= D) and αD,j =
PD (j ∈ δ) for all j ∈
/ D, where the subscript D denotes
the
measure
when
set
D
contains all the targets. Thus, αD ≤
P
j ∈D
/ αD,j . By the stopping rule, accepting j ∈ δ implies
∆Sj,m ≥ − log c for some m ∈ D. Hence, for all j ∈
/ D we
have:
αD,j = PD (j ∈ D)
P
≤ m∈D PD (∆Sj,m (τ ) ≥ − log c)
X
≤
cPD∪j\m (∆Sj,m (τ ) ≥ − log c) ≤ L · c,
(160)
m∈D
where we changed the measure in the second inequality. As a
result,
X
αD ≤
αD,j ≤ (M − L)L · c,
which yields (159).
≤ Ce−γn
(167)
and
GD
D(gm ||fm )−
Proof: For each j, define tj (n) as the time when cell j
is observed for the nth time. By DGFi selection rule, if cell
j is observed at time t, then there exists m ∈ D such that
Sj (t) ≥ Sm (t). Hence,
PD (Nj (n) > qn)
(168)
for all t < n.
Since (167) can be shown similarly as in (163), it remains to
show (168). By the definition of GD , if Nm (t) > D(gmG||fDm )− ·
n, there exists m0 ∈ D and 0 > 0 such that Nm0 (t) <
GD
0 ||f 0 )+0 · t. Hence,
D(gm
m
GD
PD Nm (t) >
· n,
D(gm ||fm ) −
∀m0 ∈ D : Sm0 (t) > Sm (t)
≤
X
PD Nm (t) >
m0 ∈D
PD (Nj (t) > qn, ∃m ∈ D : Sj (t) > Sm (t))
(162)
Sm0 (t) >
t=1
0
Sm (t), Nm
(t)
×PD (t (dqne) = t).
It suffices to show that there exist constants C, γ such that
PD (Nj (t) > qn, ∃m ∈ D : Sj (t) > Sm (t)) ≤ Ce−γn (163)
for all t ≤ n.
First we have
(164)
≤
m∈D
Fix m, then we have
≤
n
X
PD
r=dqne k=0
n
X
i=1
n
X
bs2 tc
X
PD
`j (i) +
k
X
≤
bs2 tc
X
PD
r=ds1 ne k=0
!
−`m (i) ≥ 0
n
X
(165)
k=1
+
≤ Cm e−γm n .
The last inequality can be shown using the Chernoff bound as
in Lemma 3.
P
To show (163), we let C = m Cm , γ = minm γm , which
completes the proof.
k
X
r
X
i=1
r=ds1 ne k=0
PD (Nj (t) > qn, Sj (t) > Sm (t))
GD
<
·t .
0 ||f 0 ) + 0
D(gm
m
Fix m0 , and let s1 = D(gmG||fDm )− , s2 =
Then, we have
GD
PD Nm (t) >
· n,
D(gm ||fm ) −
0
Sm0 (t) > Sm (t), Nm
(t) <
PD (Nj (t) > qn, ∃m ∈ D : Sj (t) > Sm (t))
X
≤
PD (Nj (t) > qn, Sj (t) > Sm (t)).
r
X
(169)
GD
· n,
D(gm ||fm ) −
j
n
X
· n, ∀m0 ∈ D : Sm0 (t) > Sm (t)
≤ Ce−γn
Here we consider the case where ID = GD , the case ID =
FD applies with minor modifications. For showing that τ1 is
sufficiently small we need to show first the following Lemmas:
Lemma 21: For all j ∈
/ D, ∀0 < q < 1, there exist C, γ > 0
such that
PD (Nj (n) > qn) < Ce−γn
(161)
n
X
Proof: For each m, define tm (n) as the time when cell
m is observed for the nth time. By DGFi selection rule, if cell
m is observed at time t, either there exists j ∈
/ D such that
Sj (n) > Sm (n) or Sm0 (n) > Sm (n) for all m0 ∈ D. Similar
to (162), it suffices to show that
/ D : Sj (t) > Sm (t)
PD Nm (t) > D(gmG||fDm )− · n, ∃j ∈
PD Nm (t) >
j ∈D
/
≤
Lemma 22: For all m ∈ D, and > 0, there exist C, γ > 0
such that
GD
PD Nm (n) >
· n ≤ Ce−γn
(166)
D(gm ||fm ) −
GD
D(gm0 ||fm0 )+ .
GD
·t
D(gm0 ||fm0 ) + 0
−`m (i) +
k
X
!
`m0 (i) ≥ 0
i=1
D(gm ||fm ) − − `m (i)
i=1
!
`m0 (i) − D(gm0 ||fm0 ) − 0 ≥ 0
i=1
≤
n
X
bs2 tc h
X
ir h
ik
0
˜
ED es(−`m (1)−)
ED es(`m0 (1)− )
r=ds1 ne k=0
≤ Cm0 e−γm0 n
(170)
19
The last inequality can be shown using the Chernoff
bound
P
as in Lemma 3. To show (169), we let C = m0 Cm0 , γ =
minm0 γm0 , which completes the proof.
Lemma 23: For all m ∈ D, ∀ > 0, there exist C, γ > 0
such that
GD
PD Nm (n) < (
)n ≤ Ce−γn .
(171)
2D(gm ||fm )
0m
Proof: By choosing
qj Pand
in Lemma 21 and
P
0
Lemma 22 such that j qj + m 0m = 2D(gGmD||fm ) , we have
GD
PD Nm (n) < (
)n
2D(gm ||fm )
X
≤
PD (Nj (n) > qj n)
j ∈D
/
+
X
0
P D Nm
(n) > (
m0 ∈D
GD
+ 0m )n
0 ||f 0 )
D(gm
m
for all D and m ∈ D. Then, we define:
j ∗ (t) , arg min Nj (t)D(fj ||gj ),
(177)
m∗ (t) , arg min Nm∗ (t) (t)D(gm ||fm ),
(178)
j ∈D
/
m∈D
and
WD∗ (t) ,
t
X
`˜m∗ (t) (i)1m∗ (t) (i) −
i=1
≤ Cm0 e
−γn
(172)
Next, similar to Lemma 4, we can show that the probability
that τ1 is greater than n decreases exponentially with n. This
result is used when evaluating the asymptotic expected search
time to show that it is not affected by τ1 . We can show this
by noting that
PD (τ1 > n) ≤ PD
max sup (Sj (t) − Sm (t)) ≥ 0
i=1
(179)
where WD∗ (t) is a sum of zero-mean r.v. Using these definitions, similar to Lemma 18, we can show that for every fixed
> 0 there exist C > 0 and γ > 0 such that
PD max WD∗ (t) ≥ n|Γ ≤ Ce−γn
(180)
for all D and for any policy Γ.
Next, similar to Lemma 19 we can show that for any fixed
> 0,
PD max ∆SD (t) ≥ n (ID + ) | Γ → 0
1≤t≤n
as n → ∞ ,
(181)
for all D and for any policy Γ.
Finally, similar to Lemma 20, we can show that any policy
Γ that satisfies RD (Γ) = O(−c log c) for all D must satisfy:
j ∈D,m∈D
/
t≥n
≤
∞
X
RD (Γ) ≥ − (1 + o(1))
PD (Sj (t) ≥ Sm (t)) .
(173)
Following (173), it suffices to show that PD (Sj (n) ≥ Sm (n))
decays exponentially with n. Note that
GD
≤ PD Sj (n) ≥ Sm (n), Nm (n) ≥ (
)n
2D(g
m ||fm )
GD
)n
+PD Nm (n) < (
2D(gm ||fm )
(174)
The first term decays exponentially with n by Lemma 3 (with
minor modifications). The second term decays exponentially
with n by Lemma 23.
Note that we obtained that the expectation of τ1 is bounded,
and we can use similar arguments as in the balanced case of
Theorem 1 in Appendix B to obtain the detection rate ID
for n ≥ τ1 . Combining these results yields that the expected
detection time τ under the DGFi policy is upper bounded by:
ED (τ ) ≤ − (1 + o(1))
log(c)
,
ID
c log(c)
.
ID
(182)
for all D.
t=n
j ∈D,m∈D
/
PD (Sj(n) ≥ Sm (n))
`˜j ∗ (t) (i)1j ∗ (t) (i),
1≤t≤n
as desired.
X
t
X
(175)
for m = 1, ..., M .
Finally, showing that the asymptotic Bayes risk is lower
bounded by −c log c/IL∗ follows a similar outline as in Appendix B. Specifically, similar to Lemma 16, if αD (Γ) =
O(−c log c) for all D, and we let 0 < < 1, then:
PD (∆Sm (τ ) < − (1 − ) log c | Γ) = O(−c log c) , (176)
R EFERENCES
[1] Q. Zhao and B. M. Sadler, “A survey of dynamic spectrum access,”
IEEE signal processing magazine, vol. 24, no. 3, pp. 79–89, 2007.
[2] H. Chernoff, “Sequential design of experiments,” The Annals of Mathematical Statistics, vol. 30, no. 3, pp. 755–770, 1959.
[3] A. Wald, “Sequential analysis. 1947,” Zbl0029, vol. 15805, 1947.
[4] S. A. Bessler, “Theory and applications of the sequential design of
experiments, k-actions and infinitely many experiments. part i. theory,”
tech. rep., DTIC Document, 1960.
[5] S. Nitinawarat, G. K. Atia, and V. V. Veeravalli, “Controlled sensing
for hypothesis testing,” in 2012 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pp. 5277–5280,
IEEE, 2012.
[6] S. Nitinawarat, G. K. Atia, and V. V. Veeravalli, “Controlled sensing
for multihypothesis testing,” IEEE Transactions on Automatic Control,
vol. 58, no. 10, pp. 2451–2464, 2013.
[7] S. Nitinawarat and V. V. Veeravalli, “Controlled sensing for sequential
multihypothesis testing with controlled markovian observations and nonuniform control cost,” Sequential Analysis, vol. 34, no. 1, pp. 1–24,
2015.
[8] M. Naghshvar and T. Javidi, “Active sequential hypothesis testing,” The
Annals of Statistics, vol. 41, no. 6, pp. 2703–2738, 2013.
[9] M. Naghshvar and T. Javidi, “Sequentiality and adaptivity gains in
active hypothesis testing,” IEEE Journal of Selected Topics in Signal
Processing, vol. 7, no. 5, pp. 768–782, 2013.
[10] A. Tajer and H. V. Poor, “Quick search for rare events,” IEEE Transactions on Information Theory, vol. 59, no. 7, pp. 4462–4481, 2013.
[11] N. K. Vaidhiyan and R. Sundaresan, “Learning to detect an oddball
target,” arXiv preprint arXiv:1508.05572, 2015.
[12] K. Leahy and M. Schwager, “Always choose second best: Tracking
a moving target on a graph with a noisy binary sensor,” in Control
Conference (ECC), 2016 European, pp. 1715–1721, IEEE, 2016.
20
[13] G. Fellouris, G. V. Moustakides, and V. V. Veeravalli, “Multistream
quickest change detection: Asymptotic optimality under a sparse signal,”
in Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE
International Conference on, pp. 6444–6447, IEEE, 2017.
[14] S. Nitinawarat and V. V. Veeravalli, “Universal scheme for optimal
search and stop,” in Information Theory and Applications Workshop
(ITA), 2015, pp. 322–328, IEEE, 2015.
[15] B. Hemo, K. Cohen, and Q. Zhao, “Asymptotically optimal search of
unknown anomalies,” in Proc. of the 16th IEEE Symposium on Signal
Processing and Information Technology (ISSPIT), (Limassol, Cyprus),
Dec. 2016.
[16] K. Cohen and Q. Zhao, “Asymptotically optimal anomaly detection via
sequential testing,” IEEE Transactions on Signal Processing, vol. 63,
no. 11, pp. 2929–2941, 2015.
[17] K. Cohen, Q. Zhao, and A. Swami, “Optimal index policies for anomaly
localization in resource-constrained cyber systems,” IEEE Transactions
on Signal Processing, vol. 62, no. 16, pp. 4224–4236, 2014.
[18] D. A. Castanon, “Optimal search strategies in dynamic hypothesis
testing,” IEEE transactions on systems, man, and cybernetics, vol. 25,
no. 7, pp. 1130–1138, 1995.
[19] L. Lai, H. V. Poor, Y. Xin, and G. Georgiadis, “Quickest search over
multiple sequences,” IEEE Transactions on Information Theory, vol. 57,
no. 8, pp. 5375–5386, 2011.
[20] M. L. Malloy, G. Tang, and R. D. Nowak, “Quickest search for a rare
distribution,” in Information Sciences and Systems (CISS), 2012 46th
Annual Conference on, pp. 1–6, IEEE, 2012.
[21] J. Heydari, A. Tajer, and H. V. Poor, “Quickest linear search over correlated sequences,” IEEE Transactions on Information Theory, vol. 62,
no. 10, pp. 5786–5808, 2016.
[22] Y. Pei, Y.-C. Liang, K. C. Teh, and K. H. Li, “Energy-efficient design of
sequential channel sensing in cognitive radio networks: optimal sensing
strategy, power allocation, and sensing order,” IEEE Journal on Selected
Areas in Communications, vol. 29, no. 8, pp. 1648–1659, 2011.
[23] R. Caromi, Y. Xin, and L. Lai, “Fast multiband spectrum scanning
for cognitive radio systems,” IEEE Transactions on Communications,
vol. 61, no. 1, pp. 63–75, 2013.
[24] L. Ferrari, Q. Zhao, and A. Scaglione, “Utility maximizing sequential
sensing over a finite horizon,” IEEE Transactions on Signal Processing,
vol. 65, no. 13, pp. 3430–3445, 2017.
[25] M. Egan, J.-M. Gorce, and L. Cardoso, “Fast initialization of cognitive
radio systems,” in IEEE International Workshop on Signal Processing
Advances in Wireless Communications, 2017.
[26] A. Tajer, V. V. Veeravalli, and H. V. Poor, “Outlying sequence detection
in large data sets: A data-driven approach,” IEEE Signal Processing
Magazine, vol. 31, no. 5, pp. 44–56, 2014.
[27] K. Cohen and Q. Zhao, “Active hypothesis testing for anomaly detection,” IEEE Transactions on Information Theory, vol. 61, no. 3,
pp. 1432–1450, 2015.
[28] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection: A survey,”
ACM computing surveys (CSUR), vol. 41, no. 3, p. 15, 2009.
[29] M. H. Bhuyan, D. K. Bhattacharyya, and J. K. Kalita, “Network anomaly
detection: methods, systems and tools,” IEEE Communications Surveys
& Tutorials, vol. 16, no. 1, pp. 303–336, 2014.
| 7cs.IT
|
arXiv:1507.07778v3 [math.GT] 6 Nov 2017
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS
AND MAPS OF NON-ZERO DEGREE
CHRISTOFOROS NEOFYTIDIS
A BSTRACT. We define a new class of irreducible groups, called groups not infinite-index presentable by products or not IIPP. We prove that certain aspherical manifolds with fundamental
groups not IIPP do not admit maps of non-zero degree from direct products. This extends previous results of Kotschick and Löh, providing new classes of aspherical manifolds – beyond those
non-positively curved ones which were predicted by Gromov – that do not admit maps of non-zero
degree from direct products.
A sample application is that an aspherical geometric 4-manifold admits a map of non-zero degree
from a direct product if and only if it is a virtual product itself. This completes a characterization of
the product geometries due to Hillman. Along the way we prove that for certain groups the property
IIPP is a criterion for reducibility. This especially implies the vanishing of the simplicial volume
of the corresponding aspherical manifolds. It is shown that aspherical manifolds with reducible
fundamental groups do always admit maps of non-zero degree from direct products.
1. I NTRODUCTION
A fundamental topic in topology is the study of maps of non-zero degree between manifolds of
the same dimension. The existence of a map of non-zero degree defines a transitive relation, called
domination relation, on the homotopy types of closed oriented manifolds. Whenever such a map
M −→ N exists we say that M dominates N and denote this by M ≥ N. Gromov suggested investigating the domination relation as defining a partial order of manifolds and formulated several
conjectures for candidate classes that might (not) be comparable under ≥ [14, 13, 15, 4, 38, 20].
A particular case of the domination question is when the domain is a non-trivial direct product.
That case was raised in Gromov’s theory of functorial semi-norms (such as the simplicial volume,
see also [25]) and of topological rigidity, where Gromov predicted that the fundamental classes of
certain aspherical manifolds with large universal covers are not dominated by products. Furthermore, the domination-by-products question has its own independent interest, being a special case
of Steenrod’s classical problem on the realization of homology classes by manifolds [7].
The homotopy types of aspherical manifolds are determined by their fundamental groups. Longstanding rigidity conjectures state that their homeomorphism types are determined by their fundamental groups as well. Several related questions concern the decomposition of the fundamental
group of an aspherical manifold as a direct product and the realization of finitely presented Poincaré
duality groups as fundamental groups of aspherical manifolds. An open question in this context is
Date: November 7, 2017; c C. Neofytidis 2015–2017.
2010 Mathematics Subject Classification. 57M05, 57M10, 57M50, 55M25, 22E25, 57N65, 53C23, 20F34.
1
2
CHRISTOFOROS NEOFYTIDIS
whether every closed aspherical manifold M with fundamental group Γ1 × Γ2 can be decomposed
as a product of closed manifolds with fundamental groups Γ1 and Γ2 respectively. In particular, it
is an open question whether aspherical manifolds with reducible fundamental groups (that is, virtual products of two infinite groups) are finitely covered – and therefore dominated – by products.
Lück showed how to obtain an affirmative answer in dimensions higher than four, relying on very
strong assumptions concerning the Farrell-Jones conjecture and the cohomological dimensions of
the involved groups [24]. For non-positively curved manifolds an affirmative answer is given by
Gromoll-Wolf’s isometric splitting theorem [12].
1.1. Summary of results. Before stating the main results in detail, we give a general overview of
the content of this paper. First, applying Thom’s emphatic answer (in rational homology) [33] of
Steenrod’s realization problem, we will show that every closed aspherical manifold with reducible
fundamental group is indeed dominated by products (Theorem A). Therefore, in this paper we are
interested to investigate whether aspherical closed manifolds with irreducible (i.e. not reducible)
fundamental group admit arbitrary maps of non-zero degree by non-trivial direct products (clearly
such maps cannot be homotopic to coverings).
Our goal is to introduce a new algebraic obstruction to domination by products for aspherical
manifolds whose fundamental groups have non-trivial center. This property will be termed “groups
not infinite-index presentable by products” or “not IIPP” for short (cf. Definition 1.1). We will
obtain new classes of aspherical manifolds that are not dominated by direct products, extending
a previous obstruction of Kotschick and Löh [20] (called “groups not presentable by products”),
and expanding Gromov’s predictions (about irreducible, non-positively curved manifolds [14])
to aspherical manifolds with non-trivial center. In particular, we will show that large classes of
circle bundles with fundamental groups not IIPP are not dominated by products (Theorem B). In
many cases, we will prove that the existence of a map of non-zero degree from a direct product is
equivalent to the existence of a finite covering of the same product type (Theorem C).
By definition, a group not presentable by products is not IIPP and a group not IIPP is irreducible,
but none of those implications can be reversed (cf. Remark 1.2). Nevertheless, the characterization
of Theorem C contains a special case of equivalence between “irreducible” and “not IIPP” for
arbitrary (i.e. not necessarily finitely generated or torsion-free) groups: If the quotient of a group Γ
by its center C(Γ) is not presentable by products, then Γ is reducible if and only Γ is IPPP (Theorem
D). This result fails when Γ/C(Γ) is presentable by products. For example, the 5-dimensional
Heisenberg group H5 – whose quotient by its center is Z4 and thus presentable by products – is IIPP
but irreducible (cf. Example 1.7). Dimension five is the sharp dimension in which this phenomenon
occurs, since “IIPP” and “reducible” are equivalent for fundamental groups of aspherical geometric
manifolds (in the sense of Thurston) in dimensions ≤ 4; cf. Section 6.2, in particular Theorem 6.7.
As a sample application of this study we deduce (combining Theorems A and D) the vanishing
of the simplicial volume of certain aspherical manifolds (Corollary 1.5). Moreover, our results
combined with Hillman’s work [17] show that a 4-dimensional geometric manifold is dominated by
a product if and only if it is covered by a product (Theorem F). Further applications about ordering
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
3
manifolds using maps of non-zero degree [4, 38] and the monotonicity of Kodaira dimensions
with respect to the existence of maps of non-zero degree [39] will be presented in a subsequent
paper [28].
1.2. Main theorems. In the development of the theory of bounded cohomology, Gromov [13]
conjectured that the fundamental classes of irreducible, locally symmetric spaces of non-compact
type cannot be represented by non-trivial products (of surfaces). Kotschick and Löh [20] verified
Gromov’s conjecture, by finding an algebraic obstruction to domination by products for rationally
essential manifolds. A closed oriented connected n-dimensional manifold M is called rationally
essential if the classifying map of the universal covering cM : M −→ Bπ1 (M) sends the fundamental class of M to a non-trivial element in Hn (Bπ1 (M); Q); see [14, 13]. The non-domination
criterion of [20], given in Theorem 2.4 below, reads as follows: An infinite group Γ is called not
presentable by products if, for every homomorphism ϕ : Γ1 ×Γ2 −→ Γ onto a finite index subgroup
of Γ one of the factors Γi has finite image ϕ(Γi ) ⊂ Γ.
As we shall see in Section 2.2, the proof of Theorem 2.4 was obtained by showing that the
existence of a map of non-zero degree f : X1 × X2 −→ M, with M rationally essential, implies
a presentation by products Γ1 × Γ2 −→ π1 (M), where Γi := im(π1 (f Xi )). However, the proof
of that statement does not give any insight on the index of the presenting factors Γ1 , Γ2 in π1 (M).
In fact, all the possibilities for the indices [π1 (M) : Γi ] may occur, as we will see in Example 4.1.
The targets of that example are however direct products of (aspherical) manifolds. Hence, a main
question is to understand how close to “reducible” must the fundamental group of an aspherical
manifold be, in order this manifold to be dominated by products. Indeed, the case of reducible
fundamental groups has a complete affirmative answer:
Theorem A. A closed aspherical manifold with reducible fundamental group is dominated by a
non-trivial direct product of closed oriented manifolds.
Thus our purpose in this article is to examine aspherical manifolds with irreducible fundamental
groups. We extend the notion “group not presentable by products”, including groups with (virtually) infinite center:
Definition 1.1. An infinite group Γ is called not infinite-index presentable by products or not IIPP
if, for every homomorphism ϕ : Γ1 × Γ2 −→ Γ onto a finite index subgroup of Γ at least one of
the images ϕ(Γi ) has finite index in Γ.
Otherwise, if such homomorphism ϕ exists with [Γ : ϕ(Γi )] = ∞ for both i, then Γ is called
infinite-index presentable by products or IIPP.
Remark 1.2. The following inclusions hold immediately by definition:
{groups not presentable by products} ⊂ {groups not IIPP} ⊂ {irreducible groups}.
However, none of the inverse inclusions hold: For the first inclusion, the infinite cyclic group is
a trivial example of a group presentable by products, but not IIPP. For the second inclusion, the
5-dimensional Heisenberg group is an example of an IIPP irreducible group; cf. Example 1.7.
4
CHRISTOFOROS NEOFYTIDIS
The strong feature of the property “not IIPP” is that it detects at once all the possible dimensions
of the factors of a product that dominates (with a π1 -surjective map) a rationally essential manifold
with torsion-free fundamental group. For aspherical manifolds one of the factors can be taken as
simple as possible:
Up to finite covers, if an aspherical manifold M with fundamental group not IIPP is dominated
by a product X1 × X2 , then X1 × T dim X2 ≥ M, where dim X2 ≤ rankC(π1 (M)).
Our first main non-domination result deals with circle bundles:
Theorem B. Let M be a circle bundle over a closed oriented aspherical manifold B, so that
π1 (M) is not IIPP and its center remains infinite cyclic in finite covers. Then M is not dominated
by any non-trivial direct product of closed oriented manifolds.
Example 1.3. Closed Nil4 -manifolds fulfill the conditions of Theorem B and therefore are never
dominated by products; cf. Propositions 6.10 and 6.22. Examples of such manifolds can easily be
constructed as mapping tori of suitable self-homeomorphisms of T 3 ; see Remark 6.12.
The fundamental group of M in Theorem B is presentable by products having infinite (cyclic)
center. By Eberlein’s works [5, 6] or, more generally, by the stirring work of Farb and Weinberger [9], irreducible manifolds of dimension higher than one with fundamental groups with
center do not admit metrics of non-positive sectional curvature. The non-domination results of
Kotschick and Löh [20] deal mostly with non-positively curved manifolds. More precisely, they
show that non-positively curved manifolds are dominated by products if and only if they are virtually diffeomorphic to products. Equivalently, the fundamental groups of those manifolds are
reducible if and only if they are presentable by products. In the case of fiber bundles (which
includes manifolds that do not admit metrics of non-positive sectional curvature [18]), the results
of [20] deal with targets whose fiber and base have fundamental groups not presentable by products
and with targets with positive simplicial volume.
A circle bundle M over an aspherical manifold has vanishing simplicial volume [13, 23]. Moreover, the infinite cyclic fundamental group of the S 1 -fiber is presentable by products and, as mentioned above, central in π1 (M), which means that π1 (M) is presentable by products. However,
as noted in Remark 1.2, the property “presentable by products” is not generally equivalent to
“reducible”, and this applies as well to fundamental groups of many circle bundles; e.g. the fundamental group of any circle bundle over a closed surface with non-zero rational Euler class is
irreducible and presentable by products (but it is not IIPP; see Proposition 6.3). In fact, when
the base of the circle bundle has fundamental group not presentable by products, then π1 (M) is
reducible if and only if it is IIPP. The following result characterizes circle bundles with the latter
property, and is a partial converse to Theorem B:
π
Theorem C. Let M −→ B be a circle bundle over a closed aspherical manifold B whose fundamental group π1 (B) is not presentable by products. Then the following are equivalent:
(1) M is dominated by a non-trivial product of closed oriented manifolds;
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
5
(2) M is finitely covered by a product S 1 × B ′ , for some finite cover B ′ −→ B;
(3) π1 (M) is reducible;
(4) π1 (M) is IIPP.
Example 1.4. A closed 4-manifold M carrying the geometry Sol14 is virtually a circle bundle
over a closed oriented Sol3 -manifold, and π1 (M) is not IIPP (see Propositions 6.15 and 6.24
respectively). Since Sol3 -manifold groups are not presentable by products (cf. Proposition 6.2),
Theorem C implies that M is not dominated by products. Actually, a circle bundle over a closed
oriented Sol3 -manifold is dominated by a product if and only if it possesses the geometry Sol3 × R
(see also Theorem 6.1).
The idea of Theorem C is that the center of π1 (B) remains trivial in finite covers, being not
presentable by products and torsion-free. In fact, the equivalence between (3) and (4) holds in its
greatest generality with no assumptions on finite generation, torsion freeness or the virtual rank of
the center:
Theorem D. Let Γ be a group with center C(Γ) such that the quotient Γ/C(Γ) is not presentable
by products. Then, Γ is reducible if and only if it is IIPP.
In the light of Theorem A we obtain the following consequence of Theorem D:
Corollary 1.5. Let M be a closed oriented aspherical manifold such that π1 (M) is IIPP and
π1 (M)/C(π1 (M)) is not presentable by products. Then there exists a closed oriented manifold
N such that T k × N ≥ M, where T k is the k-dimensional torus with k < rankC(π1 (M)). In
particular, M has zero simplicial volume.
Remark 1.6. It is a long-standing conjecture that the simplicial volume of a closed aspherical manifold whose fundamental group contains a non-trivial amenable normal subgroup vanishes [23]. The vanishing result in Corollary 1.5 is a straightforward consequence of the domination T k × N ≥ M (applying Theorems D and A). Alternatively, after showing that π1 (M) is a
virtual product with an Abelian factor (by Theorem D), the vanishing of the simplicial volume of
M follows as well by Gromov’s isometry theorem with respect to the simplicial ℓ1 -norm [13].
The property IIPP truly recognizes reducibility of a group Γ whenever the quotient Γ/C(Γ) is
not presentable by products. Indeed, IIPP is not anymore a criterion for a group Γ to be reducible
when the quotient Γ/C(Γ) is presentable by products:
Example 1.7. The 5-dimensional Heisenberg group H5 is irreducible and IIPP, with H5 /C(H5 )
isomorphic to Z4 ; cf. Lemma 8.1. Also, H5 is realizable by a non-trivial circle bundle over T 4 .
In Theorem 8.2, we will prove that this manifold is not dominated by products. This seems to be
the first example of an aspherical manifold whose fundamental group admits a presentation by a
product of two non-trivial subgroups of infinite index, but this manifold is still not dominated by
products. Previously, examples of rationally essential, but not aspherical, manifolds with reducible
fundamental groups that are not dominated by products were given in [20].
6
CHRISTOFOROS NEOFYTIDIS
Also, this example shows that the converse of Theorem B does not hold in general. In particular,
the assumptions of Theorem C (which is a partial converse of Theorem B) cannot be removed.
The Heisenberg group H5 has cohomological dimension five which allows enough space for the
existence of two infinite commuting subgroups Γi ⊂ H5 such that H5 is presented by Γ1 × Γ2 and
[H5 : Γi ] = ∞. This is not true in lower dimensions as we shall see in Theorem E below.
1.3. Examples and applications. We now characterize the fundamental groups of certain aspherical manifolds, giving non-trivial examples of groups presentable by products, but not IIPP. Then,
we apply our main results to show that an aspherical 4-manifold possessing a Thurston geometry
is dominated by a direct product if and only if it is virtually a direct product.
1.3.1. Non-trivial examples of groups not IIPP. The condition “not IIPP” is the crucial property for the non-existence results of this paper. The following result gathers together some nonelementary examples of groups not IIPP.
Theorem E. Irreducible fundamental groups of aspherical manifolds that possess a Thurston solvable geometry in dimensions ≤ 4 are not IIPP.
The above statement contains solvable groups that are not presentable by products as well,
4
namely Sol3 -, Sol04 - and Solm6
=n -manifold groups. Theorem E says, roughly, that in low dimensions the properties “reducible” and “IIPP” are actually equivalent. The requirement on the cohomological dimension being at most four is crucial in the above theorem, because as we have seen
in Example 1.7 the 5-dimensional Heisenberg group H5 is irreducible and IIPP.
1.3.2. Domination by products for geometric 4-manifolds. After [20], it is natural to ask to what
extent the condition “presentable by products” on the fundamental group of a rationally essential
manifold M would be sufficient for X1 × X2 ≥ M. Theorem A says that reducibility suffices
for aspherical manifolds. A complete answer is known in dimension three [22], where Kotschick
and the author proved that a closed 3-manifold is dominated by products if and only if it is either
a virtual product or a virtual connected sum of copies of S 2 × S 1 . Thus, in particular, non-trivial
circle bundles over closed oriented aspherical surfaces are never dominated by products, although
their fundamental groups are presentable by products, having infinite center. It is well known [34,
g2 if and only if it is
30] that a closed 3-manifold possesses one of the geometries Nil3 or SL
virtually a non-trivial circle bundle over a closed aspherical surface (torus or a hyperbolic surface
respectively). A main application of Theorems B and C, together with the examples of groups not
IIPP given in Theorem E, is the following characterization in dimension four:
Theorem F. A closed oriented aspherical geometric 4-manifold M is dominated by a non-trivial
product if and only if it is finitely covered by a product. Equivalently, M carries one of the product
geometries X3 × R or the reducible H2 × H2 geometry.
The existence of finite coverings of (diffeomorphism) type N × S 1 for manifolds modeled on
product geometries X3 × R is due to Hillman [17] (note that domination by products alone follows
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
7
by Theorem A as well). The above theorem says that only geometric aspherical 4-manifolds that
are virtual products admit maps of non-zero degree from direct products.
Remark 1.8. We note that we could have included non-aspherical geometries as well in the above
statement, however those geometries are not interesting for the domination-by-products question,
either because they are products themselves or because their representatives are simply connected.
The latter geometries were contained as trivial examples in [26], where we constructed maps from
products to every simply connected 4-manifold.
Acknowledgements. Most of the content of this paper stems by a large part of author’s thesis,
which was defended in Munich. The author is grateful to his advisor, Dieter Kotschick, for his wise
guidance and support. Among several other people, the author is particularly grateful to Clara Löh
and to Shicheng Wang for useful discussions and many suggestions; Shicheng Wang’s previous
numerous works on maps of non-zero degree have been a great source of inspiration and have
strongly influenced this research project. The financial support of the Deutscher Akademischer
Austausch Dienst (DAAD) is also gratefully acknowledged.
2. BACKGROUND
AND
P RELIMINARIES
In this section, we give a short description of the background of the topic of this paper, explain
the notation, and give some elementary properties of groups presentable by products.
2.1. The domination relation. In the early 1940s, Steenrod raised the question of whether every n-dimensional integral homology class can be realized as the image of the fundamental class
of a closed oriented n-dimensional manifold under a continuous map [7, Problem 25]. About a
decade later, Thom answered affirmatively Steenrod’s question in degrees up to six, and found a
7-dimensional integral homology class which is not realizable by a closed manifold. (Since then,
other non-realizability results have been obtained.) Nevertheless, Thom proved that, in all degrees,
some multiple of each integral homology class can be realized by a closed smooth manifold [33].
In this paper, we are interested in the realization of fundamental classes of closed oriented manifolds (especially by direct products of manifolds), and therefore we deal with the notion of the
degree of a continuous map. Namely, suppose that f : M −→ N is a continuous map between two
closed oriented n-dimensional manifolds. The degree of f is defined to be the integer d so that
Hn (f ; Z)([M]) = d · [N], where [M] ∈ Hn (M; Z) and [N] ∈ Hn (N; Z) denote the fundamental
classes of M and N respectively. Whenever d is not zero, we say that M dominates N (or that M
d-dominates N), and write M ≥ N (or M ≥d N). The degree of f is denoted by deg(f ). Unless
otherwise stated, in this paper we consider continuous maps between the homotopy types of closed
oriented connected manifolds.
Gromov suggested to investigate the domination relation as defining an ordering of manifolds
and as a tool to understand the values of functorial semi-norms on homology, most notably the
simplicial volume [14, 13, 4]. The simplicial volume of a closed manifold M is completely determined by the classifying space of the fundamental group, Bπ1 (M), because the classifying map
8
CHRISTOFOROS NEOFYTIDIS
of the universal covering, cM : M −→ Bπ1 (M), induces an isometry H∗ (cM ; Q) : H∗ (M; Q) −→
H∗ (Bπ1 (M); Q) with respect to the simplicial ℓ1 -norm [13]. This gives rise to the following definition:
Definition 2.1 ([13]). A closed oriented connected n-dimensional manifold M is called rationally
essential if Hn (cM ; Q)([M]) 6= 0 ∈ Hn (Bπ1 (M); Q), where cM : M −→ Bπ1 (M) classifies the
universal covering of M. Otherwise, M is called rationally inessential.
Clearly, every closed aspherical manifold is rationally essential. In fact, the notion of essentialness expands widely the class of aspherical manifolds, because for example every connected sum
containing a rationally essential summand is rationally essential itself, and every manifold with
non-zero simplicial volume is rationally essential [13].
2.2. Kotschick-Löh’s non-domination criterion. Gromov conjectured that there might exist certain classes of (rationally essential) manifolds which are not dominated by products, pointing out
irreducible locally symmetric spaces of non-compact type as potential candidates; cf. [14, Chapter 5G+ ]. Kotschick and Löh [20] verified Gromov’s suggestion, by finding a condition on the
fundamental groups of rationally essential manifolds that are dominated by products:
Definition 2.2 ([20]). An infinite group Γ is called presentable by products if there is a homomorphism ϕ : Γ1 × Γ2 −→ Γ onto a finite index subgroup of Γ so that the restriction of ϕ to each factor
Γi has infinite image ϕ(Γi ).
The property of being (not) presentable by products is clearly preserved under passing to finite
index subgroups.
Example 2.3.
(1) A reducible group is obviously presentable by products, being a virtual product of two
infinite groups.
(2) Let Γ be a group which contains a finite index subgroup Γ with infinite center. Then Γ is
presentable by products through the multiplication homomorphism C(Γ) × Γ −→ Γ.
These two examples include every torsion-free group presentable by products [20, Prop. 3.2].
Suppose now that M is a rationally essential n-dimensional manifold and let f : X1 ×X2 −→ M
be a map of non-zero degree, where the Xi are closed oriented manifolds of positive dimensions.
Consider the induced map π1 (f ) : π1 (X1 ) × π1 (X2 ) −→ π1 (M) and set
Γ := im(π1 (f )) ⊂ π1 (M) and Γi := im(π1 (f |Xi )) ⊂ Γ
for the image of π1 (f ) and the images under π1 (f ) of the restrictions of f to the two factors
Xi respectively. The multiplication map ϕ : Γ1 × Γ2 −→ Γ is then a well-defined surjective
homomorphism, because the Γi commute element-wise and Γ1 ∪ Γ2 generates Γ. Moreover, the
outer commutative diagram in Figure 1 implies that X1 × X2 is rationally essential [20].
Let cXi : Xi −→ Bπ1 (Xi ) be the classifying maps of the universal coverings of the Xi and
Bπ1 (f |Xi ) : Bπ1 (Xi ) −→ BΓi be the maps induced by π1 (f |Xi ) on the level of classifying spaces.
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
f
X1 × X2
cX1 ×X2
Bϕ
❤❤❤❤
❤❤❤❤
❤
❤
❤
❤
❤❤❤❤
❤❤❤❤
M
(Bπ1 (f |X1 )◦cX1 )×(Bπ1 (f |X2 )◦cX2 )
BΓ1 × BΓ2
33
Bπ1 (X1 × X2 )
//
9
cM
//
66
Bπ1 (M)
Bπ1 (f )
F IGURE 1. Domination by products on the level of classifying spaces.
Also, let Bϕ : BΓ1 × BΓ2 −→ BΓ be the map induced by ϕ; here we apply the homotopy
equivalence BΓ1 ×BΓ2 ≃ B(Γ1 ×Γ2 ). We then have for i = 1, 2 the maps Bπ1 (f |Xi )◦cXi : Xi −→
BΓi , and the corresponding rational homology classes
(1)
αi := Hdim Xi (Bπ1 (f |Xi ) ◦ cXi )([Xi ]) ∈ Hdim Xi (BΓi ; Q),
where [Xi ] denote the fundamental classes of the factors Xi . According to this notation, the key
observation of Kotschick and Löh [20], shown in the commutative rectangle of Figure 1, is
0 6= deg(f ) · Hn (cM )([M]) = Hn (Bϕ)(α1 × α2 ).
This means that the αi are not trivial and therefore the Γi are both infinite. In particular, Γ is
presented by the product ϕ : Γ1 × Γ2 −→ Γ. This proves the following:
Theorem 2.4 (Kotschick-Löh [20]). If M is a rationally essential manifold and π1 (M) is not
presentable by products, then M is not dominated by products.
Remark 2.5. A consequence of Theorem 2.4 is that Gromov’s prediction was indeed correct.
Namely, a locally symmetric space of non-compact type is dominated by a product if and only if it
is virtually (isometric to) a product; cf. [20, Cor. 4.2].
2.3. Groups presentable by products. We end this preliminary section with some elementary
properties of groups presentable by products, mainly as introduced in [20, Section 3].
If a group Γ is presentable by a product through a homomorphism ϕ : Γ1 × Γ2 −→ Γ, then
the images ϕ(Γi ) commute with each other and ϕ(Γ1 ) ∪ ϕ(Γ2 ) generates im(ϕ). This means that,
whenever a group Γ is presented by a product ϕ : Γ1 × Γ2 −→ Γ, we can replace each Γi by its
image ϕ(Γi ), Γ by its finite index subgroup im(ϕ) and ϕ by the multiplication map. Therefore we
may always assume that Γ can be presented by two element-wise commuting subgroups Γi through
the multiplication map. The following properties can be easily verified:
10
CHRISTOFOROS NEOFYTIDIS
Lemma 2.6 ([20, Lemma 3.3]). Suppose Γ1 , Γ2 are element-wise commuting subgroups of Γ so
that Γ1 ∪Γ2 generates Γ. Then the multiplication map ϕ : Γ1 ×Γ2 −→ Γ is a well-defined surjective
homomorphism and the following statements hold:
(1) the intersection Γ1 ∩ Γ2 is a subgroup of the center C(Γ);
(2) the kernel of ϕ is isomorphic to the Abelian group Γ1 ∩ Γ2 .
In particular, there exists a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ Γ −→ 1,
(2)
where the isomorphism between Γ1 ∩ Γ2 and the kernel of ϕ is given by the antidiagonal. For
groups with finitely generated center we moreover observe the following:
Lemma 2.7. Let Γ be a finitely generated group with finitely generated center. Assume that Γ is
presented by a product Γ1 × Γ2 as in Lemma 2.6. Then each of the factors Γi is finitely generated.
Proof. For i, j ∈ {1, 2}, i 6= j, there exist (two) short exact sequences
1 −→ Γi −→ Γ −→ Γj /(Γ1 ∩ Γ2 ) −→ 1,
(3)
where Γ −→ Γj /(Γ1 ∩Γ2 ) is obtained by composing the isomorphism Γ ∼
= (Γ1 ×Γ2 )/(Γ1 ∩Γ2 ) (cf.
sequence (2)) with the homomorphism induced by the projection from Γ1 ×Γ2 to Γj (see also [21]).
Since Γ is finitely generated, the short exact sequence (3) implies that the group Γj /(Γ1 ∩ Γ2 )
is also finitely generated. Moreover, the center C(Γ) is finitely generated Abelian and thus the
intersection Γ1 ∩ Γ2 is also finitely generated Abelian by Lemma 2.6. This shows that Γj is also
finitely generated.
3. T HE
CASE OF REDUCIBLE GROUPS ( PROOF OF
T HEOREM A)
If M is a closed aspherical n-dimensional manifold whose fundamental group π1 (M) is reducible, then there exists a finite cover of M with fundamental group isomorphic to a direct product Γ1 × Γ2 . Thus, up to finite covers, we may identify π1 (M) with Γ1 × Γ2 . Then BΓ1 × BΓ2
is homotopy equivalent to M. In particular, there exists a non-trivial class α ∈ Hn (BΓ1 × BΓ2 )
mapping to the fundamental class [M] ∈ Hn (M).
Since each of the Γi has infinite index in π1 (M) and M is aspherical manifold of dimension n,
a theorem of Strebel [32] implies that the cohomological dimensions of each of the Γi is less than
n. Thus, the Künneth formula (with rational coefficients) in degree n for the product BΓ1 × BΓ2
implies that there exist non-trivial homology classes a1 ∈ Hk (BΓ1 ) and a2 ∈ Hn−k (BΓ2 ), where
0 < k < n, such that α = a1 ⊗ a2 .
By Thom’s theorem [33], there exist two closed smooth manifolds X1 and X2 of dimensions
k and n − k respectively, together with continuous maps gi : Xi −→ BΓi , i = 1, 2, such that
H∗ (gi )([Xi ]) = di · ai , for some non-zero integers di . Finally, the product map
g1 ×g2
X1 × X2 −→ BΓ1 × BΓ2 ≃ M
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
11
is continuous and in homology of degree n maps the fundamental class [X1 × X2 ] to a non-zero
multiple of [M]. This finishes the proof of Theorem A.
4. C IRCLE
BUNDLES WITH FUNDAMENTAL GROUPS NOT
( PROOFS
OF
T HEOREMS B
AND
IIPP
C)
The purpose of this section is to introduce (in more detail) the property IIPP and prove Theorems
B and C.
4.1. Motivation and definition of the property IIPP. Two basic examples of groups presentable
by products are given by the reducible ones and by groups containing a finite index subgroup with
infinite center; cf. Example 2.3. If follows by Lemma 2.6 that these two - not generally distinct classes contain all torsion-free groups presentable by products [20, Prop. 3.2].
A reducible group Γ can always be presented (being a virtual product) by a product Γ1 × Γ2
so that both subgroups Γi have infinite index in Γ, whereas a group with infinite center does not
generally have this property; a trivial example is given by the infinite cyclic group. On the topological level, Theorem 2.4 states that whenever a rationally essential manifold M is dominated by a
product, its fundamental group is presentable by products. However, (the proof of) that result does
not give any additional information on the index of the factors of a product presenting π1 (M). The
following example shows that all the possibilities can actually occur:
Example 4.1.
(1) Let M be a closed oriented manifold of positive dimension and infinite fundamental group.
The identity map idM ×M of the product M × M is obviously π1 -surjective of degree one
and both subgroups im(π1 (idM )) = π1 (M) have infinite index in π1 (M × M).
(2) For g ≥ 1, let Σg+1 = Σg #(Sa1 × Sb1 ) be a closed oriented surface of genus g + 1. Let the
composition
(4)
q
id∨p
Σg #(Sa1 × Sb1 ) −→ Σg ∨ (Sa1 × Sb1 ) −→ Σg ∨ Sb1 ,
where q is the quotient map pinching to a point the essential circle defining the connected
sum Σg #(Sa1 × Sb1 ), id is the identity map of Σg and the map p pinches to a point the
meridian of the torus Sa1 × Sb1 ; cf. Figure 2. Denote by h the composition (id ∨ p) ◦ q. Now,
let the composite map
h×id
g
Σg+1 × Sc1 −→c (Σg ∨ Sb1 ) × Sc1 −→ Σg × Sc1 ,
where idc is the identity map of Sc1 and g is the identity on Σg and Sc1 , and sends the
generator b of Sb1 to the generator c of Sc1 . Let f := g ◦ (h × idc ) : Σg+1 × Sc1 −→ Σg × Sc1 .
Then H3 (f )([Σg+1 × S 1 ]) = [Σg × S 1 ], i.e. deg(f ) = 1. By the definition of f , we obtain
an index-one subgroup of π1 (Σg × S 1 ), namely im(π1 (f |Σg+1 )) = π1 (Σg × S 1 ), and the
infinite-index subgroup im(π1 (f |S 1 )) = π1 (S 1 ) ⊂ π1 (Σg × S 1 ).
12
CHRISTOFOROS NEOFYTIDIS
q
id∨p
−→
−→
F IGURE 2. The map (id ∨ p) ◦ q : Σg+1 −→ Σg ∨ Sb1
(3) Let two copies of a closed oriented surface of genus three,
Σ3 = (Sg11 × Sg12 )#(Sa11 × Sa12 )#(Sb11 × Sb12 )
Σ′3 = (Sg11′ × Sg12′ )#(Sa1′1 × Sa1′2 )#(Sb1′1 × Sb1′2 ).
As in the previous example (cf. Figure 2), define h : Σ3 −→ (Sg11 × Sg12 ) ∨ Sa12 ∨ Sb12 as the
composition
q
id∨p∨p
Σ3 −→ (Sg11 × Sg12 ) ∨ (Sa11 × Sa12 ) ∨ (Sb11 × Sb12 ) −→ (Sg11 × Sg12 ) ∨ Sa12 ∨ Sb12
(see above for the notation). Now, let the composition
g
h×h′
Σ3 × Σ′3 −→ ((Sg11 × Sg12 ) ∨ Sa12 ∨ Sb12 ) × ((Sg11′ × Sg12′ ) ∨ Sa1′2 ∨ Sb1′2 ) −→ Sg11 × Sg12 × Sg11′ × Sg12′ ,
where h, h′ are defined above, and g restricts to the identity map on each Sg1j and Sg1′ , and
j
is given as follows on the rest of the circles:
a2 7→ g1′ , b2 7→ g2′ , a′2 7→ g1 , b′2 7→ g2 .
We define f : Σ3 × Σ′3 −→ T 4 to be the composition g ◦ (h × h′ ). Again, f is a degree
one map. However, both im(π1 (f |Σ3 )) and im(π1 (f |Σ′3 )) are now of index one in π1 (T 4 ).
We note that this construction cannot be generalized when the target is not a product of
two tori, T 2 × T 2 , because the generators of higher genus surfaces do not commute with
each other. Actually, it will be transparent by the discussion in the upcoming subsection
(cf. Lemma 4.3), that, if an n-dimensional aspherical manifold M admits a map f : X1 ×
X2 −→ M so that both subgroups im(π1 (f |Xi )) ⊂ π1 (M) are of finite index, then M is a
virtual n-dimensional torus T n .
In this paper, we analyze groups presentable by products by adding a constraint on the index of
the presenting factors. More precisely, we introduce the following class of groups:
Definition 4.2. An infinite group Γ is called infinite-index presentable by products (IIPP) if there is
a homomorphism ϕ : Γ1 ×Γ2 −→ Γ onto a finite index subgroup of Γ so that for both factors Γi the
images ϕ(Γi ) ⊂ Γ are of infinite index in Γ. Otherwise, Γ is called not infinite-index presentable
by products (not IIPP).
In the upcoming subsection, we will show that an aspherical manifold with fundamental group
not IIPP can be dominated by a product only if it is dominated by a product containing a torus
factor. This will imply that large classes of aspherical manifolds with fundamental groups not IIPP
cannot be dominated by products. In the case of circle bundles, we will prove that, under a certain
assumption on the fundamental group of the base, the condition “IIPP” characterizes aspherical
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
13
circle bundles that are dominated by products. Without that additional assumption on the base, this
characterization does not generally hold (as we have already mentioned in Example 1.7).
4.2. Not IIPP as a non-domination criterion (proof of Theorem B). We first extend the nonexistence results of [20] to certain rationally essential manifolds with fundamental groups presentable by products, but not IIPP. The strong feature of such torsion-free groups is that one of the
presenting subgroups must be Abelian:
Lemma 4.3. Let Γ be a finitely generated torsion-free group with finitely generated center. Suppose
that there exist element-wise commuting subgroups Γ1 , Γ2 ⊂ Γ so that Γ1 ∪ Γ2 generates Γ. If Γ is
not IIPP, then one of the Γi is isomorphic to Zk for some k ≤ rankC(Γ).
Proof. By Lemma 2.6, there is a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ Γ −→ 1,
where ϕ is the multiplication map and the intersection Γ1 ∩ Γ2 is contained in the finitely generated
center C(Γ). Since Γ is torsion-free, we have that Γ1 ∩ Γ2 is isomorphic to Zk for some k ≤
rankC(Γ). Moreover, each Γi is finitely generated by Lemma 2.7.
Because Γ is not IIPP, one of the Γi , say Γ1 , must have finite index in Γ. This means that Γ2 is
virtually Γ1 ∩ Γ2 , and so it is virtually Abelian. Moreover, the intersection Γ1 ∩ Γ2 is central in Γ2 ,
which implies that Γ2 /C(Γ2 ) is finite (because Γ2 is virtually Γ1 ∩ Γ2 ). By Schur’s theorem [29],
we conclude that the commutator [Γ2 , Γ2 ] is also finite and so trivial, because Γ2 is torsion-free.
This shows that Γ2 is Abelian itself and thus isomorphic to Zk .
As a warm-up, we observe that, in the torsion-free case, Lemma 4.3 yields the following dimension restrictions on the factors of a product that dominates a rationally essential manifold:
Proposition 4.4. Let M be a rationally essential manifold so that π1 (M) is torsion-free and
rankC(π1 (M)) = r. If π1 (M) is not IIPP, then there is no π1 -surjective non-zero degree map
X1 × X2 −→ M, whenever min{dim X1 , dim X2 } > r.
Proof. Suppose that there exist X1 , X2 closed oriented manifolds of positive dimensions and a
π1 -surjective non-zero degree map f : X1 × X2 −→ M. Then there is a short exact sequence
(5)
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ π1 (M) −→ 1,
where ϕ is the multiplication map, Γi := π1 (f |Xi )(π1 (Xi )) and Γ1 ∩ Γ2 ⊂ C(π1 (M)); see Section
2.2. In particular, Γ1 ∩ Γ2 is isomorphic to Zk , for some k ≤ r = rankC(π1 (M)), because it
is torsion-free. We moreover observe that k ≥ 1, otherwise π1 (M) would be isomorphic to the
product Γ1 × Γ2 by (5) and so IIPP.
We now apply Lemma 4.3 to π1 (M) to conclude that one of the Γi , say Γ2 , is isomorphic to Zk .
This means that BΓ2 ≃ T k and by the non-vanishing (cf. Section 2.2) of
α2 := Hdim X2 (Bπ1 (f |X2 ) ◦ cX2 )([X2 ]) ∈ Hdim X2 (T k ; Q),
we deduce that dim X2 ≤ k. This is possible only if min{dim X1 , dim X2 } ≤ r.
14
CHRISTOFOROS NEOFYTIDIS
In the case where M is aspherical, then the above proposition says that, if f : X1 × X2 −→ M
is a π1 -surjective map of non-zero degree, then there is a finite cover M of M such that f factors
through the map Bϕ : M × T k −→ M which is induced by the multiplication homomorphism
ϕ : Γ1 × Γ2 −→ π1 (M), for some k ≤ rankC(π1 (M)); cf. Figure 1. In particular, there exist two non-trivial homology classes α1 ∈ Hdim X1 (M; Q) and α2 ∈ Hdim X2 (T k ; Q) such that
Hn (Bϕ)(α1 × α2 ) = deg(f ) · [M]. Thus X1 × T m ≥ M, where m = dim X2 ≤ k.
Example 4.5 (1-domination). Let M be a closed aspherical manifold. If π1 (M) has infinite cyclic
center and it is not IIPP, then M can admit a degree one map by a product X1 × X2 only if one of
the Xi is a circle. (Recall that a map of degree one is π1 -surjective.)
Aspherical manifolds whose fundamental groups have infinite cyclic center are of special interest, in particular with respect to the study of circle bundles and circle actions; see [3] and the
references there. We begin with two general facts about finite coverings of circle bundles:
π
Lemma 4.6. Let M −→ B be a circle bundle over a closed oriented manifold B.
p′
p
(1) Every finite cover M −→ M is a circle bundle over a finite cover B ′ −→ B. If moreover
B is aspherical and π1 (B) is not presentable by products, then π1 (M ) and π1 (M) have
infinite cyclic center.
(2) ([2]) If the Euler class of M is torsion, then M is a virtually trivial circle bundle over a
finite cover of B.
Proof. (1) Since π1 (p)(π1 (M )) has finite index in π1 (M) and π1 (π)(π1 (M)) = π1 (B), the imp′
age H := π1 (π ◦ p)(π1 (M )) has finite index in π1 (B). Let B ′ −→ B be the finite covering
π′
corresponding to H. Then π ◦ p lifts to M −→ B ′ , which is the desired circle bundle.
If B is aspherical, then the S 1 fiber is central in the fundamental group of M. If, in addition,
π1 (B) is not presentable by products, then it has trivial center (because it is torsion-free), and
so the center of π1 (M) is infinite cyclic. Now π1 (B ′ ) has finite index in π1 (B), and so it is not
presentable by products as well and therefore the center of π1 (M ) is also infinite cyclic.
(2) Consider the Abelianization H1 (B) = π1 (B)/[π1 (B), π1 (B)]. Since the Euler class of M is
torsion, the Universal Coefficient Theorem implies that the torsion part of H1 (B) is not trivial. Let
now the composition
π1 (B) −→ H1 (B) −→ TorH1 (B),
where the first map is the quotient map and the second is the projection to the torsion of H1 (B).
p′
If B ′ −→ B is the finite covering corresponding to the kernel of the above composition, then the
pullback bundle (p′ )∗ (M) is the product S 1 × B ′ ; see [2, Prop. 3] for more details.
p
Remark 4.7. Conversely to part (2) of the above lemma, let M = S 1 × B ′ −→ M be a finite
p′
cover, where B ′ −→ B is a finite covering between the bases (the map p′ is covered by p). The
Euler class of M is trivial, that is eM = H 2 (p′ ; Z)(eM ) = 0 ∈ H 2 (B ′ , Z), where eM ∈ H 2 (B; Z)
is the Euler class of M. By the fact that H 2 (p′ ; Q) is injective (since deg(p′ ) 6= 0), we conclude
that eM is torsion.
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
15
A basic ingredient of our proof is the following lemma which generalizes [22, Lemma 1]:
π
Lemma 4.8 (Factorization Lemma). Let M −→ B be a non-trivial n-dimensional circle bundle
over a closed oriented aspherical manifold B. Suppose that the Euler class of M is not torsion
and that the center of π1 (M) remains infinite cyclic in finite covers. Then X × S 1 M for any
closed oriented manifold X.
Proof. Since M is a non-trivial circle bundle whose integer Euler class eM ∈ H 2 (B; Z) is not
torsion, the rational Euler class of M is not trivial as well. The same property holds for every (fiber
preserving) finite cover of M, by Lemma 4.6 and Remark 4.7. By Poincaré duality, there exists
a non-trivial class α ∈ H n−3(B; Q) so that eM ⌣ α is a non-zero multiple of the cohomology
fundamental class ωB of B. Since H n−1 (B; Q) = Q, the Gysin sequence
H n−1 (π)
⌣e
M
· · · −→ H n−3(B; Q) −→
H n−1(B; Q) −→ H n−1(M; Q) −→ · · ·
implies that ker(H n−1(π)) = im(⌣ eM ) = H n−1 (B; Q). Therefore H n−1 (π) = 0.
Suppose now that there exists a non-zero degree map f : X × S 1 −→ M. After passing to a
finite cover, if necessary, we may assume that f is π1 -surjective and that the center of π1 (M) is
infinite cyclic. The latter means that the circle fiber of M represents (up to multiples) the only
central factor in π1 (M). By the surjectivity of π1 (f ), we deduce that the composite map π ◦ f
kills the homotopy class of the S 1 factor of the product X × S 1 , because this factor is central in
π1 (X × S 1 ). Since B is aspherical, we conclude that π ◦ f factors up to homotopy through the
projection p1 : X × S 1 −→ X. In particular, there is a continuous map g : X −→ B, so that
π ◦ f = g ◦ p1 up to homotopy. (We note that X is not necessarily aspherical. It is, however,
rationally essential, because f has non-zero degree and M is aspherical.)
Let ωX be the cohomology fundamental class of X. Since H n−1 (p1 ; Q)(ωX ) = ωX ∈ H n−1(X×
S 1 ; Q) and H n−1 (π; Q)(ωB ) = 0 ∈ H n−1(M; Q), the homotopy equation π ◦ f = g ◦ p1 implies
that g must be of zero degree. Let now the pullback of M under g:
g ∗M = {(x, y) ∈ X × M | g(x) = π(y)} .
The map f : X × S 1 −→ M factors through g ∗M as follows:
X × S 1 −→
g ∗M
π
2
−→
M
(x, t) 7→ (x, f (x, t)) 7→ f (x, t) .
However, the degree of the pullback map π2 : g ∗ M −→ M is zero, being equal to the degree of g,
which contradicts our assumption on deg(f ). This completes the proof.
We now finish the proof of Theorem B:
Proof of Theorem B. Since π1 (M) is not IIPP, M is a non-trivial circle bundle and, moreover, its
Euler class is not torsion by Lemma 4.6 (2). After passing to a finite cover, if necessary, suppose
16
CHRISTOFOROS NEOFYTIDIS
that there is a π1 -surjective non-zero degree map f : X1 × X2 −→ M, where dim Xi > 0 and
C(π1 (M)) = Z; cf. Lemma 4.6 (1). As before, there is a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ π1 (M) −→ 1,
where Γi := im(π1 (f |Xi )) ⊂ π1 (M), ϕ is the multiplication map and Γ1 ∩ Γ2 ⊂ C(π1 (M)) = Z,
see Section 2.2.
Lemma 4.3 implies that one of the Γi , say Γ2 , must be infinite cyclic, because π1 (M) is not IIPP
and torsion-free. Therefore, BΓ2 ≃ S 1 and because the rational class
α2 := Hdim X2 (Bπ1 (f |X2 ) ◦ cX2 )([X2 ]) ∈ Hdim X2 (S 1 ; Q)
is not trivial we conclude that dim X2 = 1, i.e. X2 = S 1 . Now, we have a π1 -surjective dominant
map X1 × S 1 −→ M, where C(π1 (M)) = Z. The proof follows by Lemma 4.8.
4.3. A characterization for circle bundles (proof of Theorem C). A main motivation for Theorem 2.4 was to show that non-positively curved manifolds which are not virtual products are not
dominated by products. Actually, the property “fundamental group presentable by products” suffices for domination by products for non-positively curved manifolds (of dimension higher than
one) and it is equivalent to “reducible”; cf. [20, Theorem 4.1]. Another consequence of the results
of [20], which moreover deals with manifolds that do not admit any metric of non-positive sectional curvature (cf. [18]), concerns fibrations whose fiber and base have fundamental groups not
presentable by products:
π
Theorem 4.9 ([20, Theorem 5.1]). Let F −→ M −→ B be a fiber bundle whose fiber F and base
B are closed oriented aspherical manifolds with fundamental groups not presentable by products.
Then M is dominated by products if and only if it is a virtual product F ′ × B ′ , where F ′ and B ′
are finite covers of F and B respectively.
Corollary 4.10 ([20, Cor. 5.3]). Let M be a closed oriented 4-manifold which is the total space of
a surface bundle whose fiber F and base B are both hyperbolic surfaces. Then the following are
equivalent:
(1)
(2)
(3)
(4)
M is dominated by a non-trivial product of closed oriented manifolds;
M is virtually diffeomorphic to a trivial surface bundle;
π1 (M) is reducible;
π1 (M) is presentable by products.
Remark 4.11. The fact that hyperbolic groups are not presentable by products is proved in [20].
Moreover, we note that 4-manifolds satisfying one (and therefore every) property in the above
corollary constitute the class of closed reducible H2 × H2 -manifolds; cf. Section 6.1.
Now, if we replace the fiber F by S 1 in Theorem 4.9, then π1 (M) is presentable by products
having infinite cyclic center. Theorem C, which is a partial converse of Theorem B, says that the
conclusion of Theorem 4.9 still holds when the fiber is S 1 . However, domination by products is
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
17
now equivalent to the conditions “π1 (M) IIPP” and “π1 (M) reducible”; compare with the equivalence between (3) and (4) of Corollary 4.10.
Proof of Theorem C. Since B is aspherical, M is also aspherical and its fundamental group fits
into a short exact sequence
(6)
π1 (π)
1 −→ π1 (S 1 ) −→ π1 (M) −→ π1 (B) −→ 1,
where π1 (S 1 ) is in the center of π1 (M). Moreover, π1 (B) has trivial center, because it is torsionfree and not presentable by products. Thus C(π1 (M)) = π1 (S 1 ) = Z.
Suppose that there is a non-zero degree map f : X1 × X2 −→ M. After passing to a finite cover,
if necessary, we may assume that f is π1 -surjective. (The finite cover of M is a circle bundle with
infinite cyclic center, by Lemma 4.6 (1).) As before, we have a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ π1 (M) −→ 1,
where Γi := im(π1 (f |Xi )) ⊂ π1 (M) and Γ1 ∩ Γ2 ⊂ C(π1 (M)) = Z. Moreover, we obtain two
non-trivial rational homology classes
αi := Hdim Xi (Bπ1 (f |Xi ) ◦ cXi )([Xi ]) 6= 0 ∈ Hdim Xi (BΓi ; Q),
see Sections 2.2 and 2.3 for the details.
The composite homomorphism
π1 (π)
ϕ
Γ1 × Γ2 −→ π1 (M) −→ π1 (B) ∼
= π1 (M)/π1 (S 1 )
maps one of the Γi , say Γ1 , to the neutral element of π1 (B), because π1 (B) is not presentable by
products and torsion-free. This means that Γ1 is contained in C(π1 (M)) = π1 (S 1 ) = Z and it is
therefore isomorphic to C(π1 (M)) = Z. In particular, BΓ1 ≃ BZ = S 1 and so the non-vanishing
of α1 ∈ Hdim X1 (S 1 ; Q) implies that dim X1 ≤ 1. Since dim X1 > 0, we have that X1 = S 1 , i.e.
S 1 × X2 ≥ M. It follows by Lemma 4.8 that M is a virtual product and, more precisely, that it is
finitely covered by a product S 1 × B ′ for some finite cover B ′ −→ B. Thus (1) implies (2). The
converse is trivially true and so (1) is equivalent to (2).
Next, we show that the properties (2) and (3) are equivalent. Obviously (2) implies (3). Assume
now that π1 (M) is reducible. Then there exists a finite cover M ′ −→ M so that π1 (M ′ ) is isomorphic to a direct product ∆1 × ∆2 , where ∆i are non-trivial (and therefore infinite) subgroups
of π1 (M ′ ). The cover M ′ is a circle bundle over a finite cover B ′ of B, where π1 (B ′ ) is not presentable by products being a finite index subgroup of π1 (B); cf. Lemma 4.6 (1). We therefore
obtain a short exact sequence
(7)
1 −→ π1 (S 1 ) −→ ∆1 × ∆2 −→ π1 (B ′ ) −→ 1,
where π1 (S 1 ) = C(π1 (M ′ )) ∼
= C(∆1 ) × C(∆2 ). Since π1 (B ′ ) is not presentable by products
and torsion-free, one of the ∆i , say ∆1 , maps trivially to π1 (B ′ ) ∼
= π1 (M ′ )/π1 (S 1 ) in (7). Thus
∆1 ⊂ π1 (S 1 ) (and so ∆1 is isomorphic to Z) and ∆2 surjects onto π1 (B ′ ). Moreover, π1 (S 1 ) maps
trivially to ∆2 , otherwise ∆2 would have finite index in π1 (M ′ ), which is impossible, because
18
CHRISTOFOROS NEOFYTIDIS
π1 (M ′ ) ∼
= ∆1 × ∆2 and both ∆i are infinite. Therefore ∆2 maps isomorphically onto π1 (B ′ ). We
have now proved that π1 (M ′ ) ∼
= π1 (S 1 × B ′ ) and so M ′ is homotopy equivalent to S 1 × B ′ . Thus
(3) implies (2).
Finally, the equivalence between (3) and (4) follows from the more general group-theoretic
Theorem D, whose proof is given in the upcoming section.
Remark 4.12. An alternative argument for the last step in the proof of the implication (3) ⇒ (2)
is the following: Having that ∆1 maps trivially to π1 (B ′ ) ∼
= π1 (M ′ )/π1 (S 1 ), we conclude that
π1 (S 1 ) maps trivially to ∆2 because the center of π1 (M ′ ) ∼
= ∆1 × ∆2 is infinite cyclic, isomorphic
1
to π1 (S ). Actually, taking for granted that the circle fiber of M ′ is the only central factor in
π1 (M ′ ), we can relax the condition “not presentable by products” for the fundamental group of the
base B ′ to “irreducible”.
Also, note that after showing the implication (1) ⇒ (2) and since the implication (2) ⇒ (3) is
trivial, we can deduce the equivalence of (1), (2) and (3) by Theorem A (which gives the implication (3) ⇒ (1)).
This discussion yields a topological example of groups not IIPP in any dimension:
Corollary 4.13. If M is a circle bundle with non-trivial rational Euler class over a closed aspherical manifold B so that π1 (B) is not presentable by products, then π1 (M) is not IIPP.
Proof. Since π1 (B) is not presentable by products, π1 (M) is IIPP if and only if it is reducible, by
the equivalence between (3) and (4) in Theorem C (or by Theorem D below). However, π1 (M) is
not reducible, otherwise M would be covered by S 1 × B ′ , for some finite cover B ′ −→ B (by the
equivalence between (2) and (3) of Theorem C), which is impossible because the Euler class of M
is not torsion; cf. Remark 4.7.
5. T HE IIPP
PROPERTY AS A CRITERION FOR REDUCIBILITY ( PROOF OF
T HEOREM D)
In this section we show that the properties “IIPP” and “reducible” are equivalent for a group Γ,
whenever Γ/C(Γ) is not presentable by products.
Proof of Theorem D. Since every reducible group is IIPP, we only need to show that the converse
is also true when the quotient of our group by its center is not presentable by products.
Let Γ be an IIPP group such that Γ/C(Γ) is not presentable by products. Since Γ is IIPP, there
exists a finite index subgroup Γ ⊂ Γ and two element-wise commuting subgroups Γ1 , Γ2 ⊂ Γ of
infinite index such that Γ = Γ1 Γ2 . Thus we obtain the following composite surjective homomorphism
Γ1 × Γ2 −→ Γ −→ Γ/Γ ∩ C(Γ),
whose image Γ/Γ ∩ C(Γ) is of finite index in Γ/C(Γ). In particular, Γ/Γ ∩ C(Γ) is not presentable
by products. Thus, for one of the Γi , say for Γ2 , the quotient Γ2 /Γ2 ∩C(Γ) is finite. Since Γ2 ∩C(Γ)
is central in Γ2 , we deduce that Γ2 is virtually C(Γ2 ), in particular it is virtually Abelian.
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
19
We claim that [Γ : Γ1 C(Γ2 )] < ∞. First, we have that
Γ = Γ1 Γ2 = Γ1 C(Γ2 )Γ2 .
(8)
Next, we observe that
(9)
Γ1 C(Γ2 ) ∩ Γ2 = C(Γ2 ).
Indeed, on the one hand it is clear that C(Γ2 ) ⊂ Γ1 C(Γ2 ) ∩ Γ2 . On the other hand, every element
in Γ1 C(Γ2 ) ∩ Γ2 is central in Γ2 , because it belongs to Γ2 and the Γi commute element-wise. By
(8), (9) and since Γ2 /C(Γ2 ) is finite, we obtain
[Γ : Γ1 C(Γ2 )] = [Γ2 : Γ1 C(Γ2 ) ∩ Γ2 ] = [Γ2 : C(Γ2 )] < ∞,
as claimed.
Let now the presentation of Γ1 C(Γ2 ) by
(10)
Γ1 ∩ C(Γ2 ) −→ Γ1 × C(Γ2 ) −→ Γ1 C(Γ2 ).
Since [Γ : Γ1 ] = ∞ and [Γ : Γ1 C(Γ2 )] < ∞, we conclude that
[C(Γ2 ) : Γ1 ∩ C(Γ2 )] = [Γ1 C(Γ2 ) : Γ1 ] = ∞.
Thus Γ1 C(Γ2 ) is isomorphic to the product Γ1 × (C(Γ2 )/Γ1 ∩ C(Γ2 )), where the quotient group
C(Γ2 )/Γ1 ∩ C(Γ2 ) is Abelian of positive rank.
Theorem D proves in particular the equivalence between (3) and (4) in Theorem C, for circle
bundles over aspherical manifolds with fundamental groups not presentable by products.
As pointed out in the introduction, the property IIPP on a group Γ is not anymore a criterion for
reducibility of Γ if the quotient Γ/C(Γ) is presentable by products. More precisely, we have seen
in Example 1.7 that the 5-dimensional Heisenberg group H5 – whose quotient H5 /C(H5 ) is Z4 –
is irreducible and not IIPP. This nilpotent group is realized as the fundamental group of a circle
bundle over T 4 . As we shall see in the next section (cf. Theorem E), dimension four is the sharp
dimension in which irreducible fundamental groups of solvable manifolds are not IIPP.
We note that the center of Γ in Theorem D can be assumed to be infinite, otherwise Γ is trivially
not presentable by products. At the other end, for groups whose every finite index subgroup has
finite center, the notions “presentable by products” and “IIPP” are equivalent:
Proposition 5.1. If every subgroup of finite index in Γ has finite center, then Γ is presentable by
products if and only if it is IIPP.
Proof. It suffices to show that presentability by products implies IIPP. Suppose that Γ1 , Γ2 are
commuting infinite subgroups of Γ and that there is a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ Γ −→ 1,
where ϕ is the multiplication homomorphism and Γ1 ∩ Γ2 lies in the center of Γ; cf. Lemma 2.6.
Since C(Γ) is finite, we deduce that Γ1 ∩ Γ2 is also finite and so it has infinite index in both Γi .
The proof now follows by the short exact sequence (3) in the proof of Lemma 2.7.
20
CHRISTOFOROS NEOFYTIDIS
6. F UNDAMENTAL
GROUPS OF GEOMETRIC MANIFOLDS IN LOW DIMENSIONS
( PROOF
OF
T HEOREM E)
After the characterization of groups (not) IIPP in the preceding section (whose topology fits in
the concept of Theorem C), we now give further non-trivial examples of groups presentable by
products but not IIPP, which fit in the concept of Theorem B as well. These examples include
irreducible fundamental groups of low-dimensional solvable manifolds with infinite center. To
this end, we characterize in terms of the IIPP property the fundamental groups of all aspherical
geometric manifolds in dimensions ≤ 4. After Theorem D, the most prominent examples in the
present section (that are not covered by Theorem D) will be irreducible fundamental groups of
nilpotent manifolds, because the quotients of these groups by their center are still presentable by
products (being again nilpotent and torsion-free). We begin this section with a brief review of
Thurston’s geometries.
6.1. Enumeration of the low-dimensional geometries. Let Xn be a complete simply connected
n-dimensional Riemannian manifold. We say that a closed manifold M is an Xn -manifold, or that
M possesses the Xn geometry in the sense of Thurston, if it is diffeomorphic to a quotient of Xn
by a lattice Γ in the group of isometries of Xn (acting effectively and transitively). The group Γ
denotes the fundamental group of M. We say that Xn and Yn are the same geometries if there
exists a diffeomorphism ψ : Xn −→ Yn and an isomorphism Isom(Xn ) −→ Isom(Yn ) mapping
each g ∈ Isom(Xn ) to ψ ◦ g ◦ ψ −1 ∈ Isom(Yn ).
In dimension 1, the circle is the only closed manifold being a quotient of the real line R. In
dimension 2, a closed surface possesses one of the geometries S 2 , R2 or H2 . In dimension 3,
Thurston [34] proved that there exist eight (homotopically unique) geometries, namely the geomeg2 , H2 × R, Nil3 , R3 , S 2 × R and S 3 (see also [30]).
tries H3 , Sol3 , SL
The classification of the 4-dimensional geometries is due to Filipkiewicz [10]. According to
that, there exist eighteen geometries in dimension four with compact representatives. There is an
additional geometry which, however, cannot be realized by any compact 4-manifold. Here, we
deal only with the aspherical geometries, because the non-aspherical ones are not interesting for
domination by products. Namely, the non-aspherical geometries are either products of a sphere
with a non-compact factor (H2 × S 2 , R2 × S 2 , S 3 × R), or compact themselves (S 2 × S 2 , CP2 ,
S 4 ), and all of their representatives are dominated by products [17, 20, 26].
Enumeration of the aspherical geometries in dimension four. We enumerate the aspherical 4dimensional geometries, following Wall’s papers [36] and [37]. Our list is adapted to the dominationby-products question, and this list will be used as an organizing principle; see Table 1.
Hyperbolic geometries. There exist two aspherical irreducible symmetric geometries, namely the
real and the complex hyperbolic, denoted by H4 and H2 (C) respectively.
Product geometries. Seven of the aspherical geometries are products of lower dimensional geg2 × R, Nil3 × R, H2 × R2 , R4 and H2 × H2 . Closed manifolds
ometries: H3 × R, Sol3 × R, SL
possessing a geometry of type X3 × R satisfy the following property:
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
Type
Geometry X4
Hyperbolic
H4 , H2 (C)
Product
Solvable
non-product
21
H3 × R, Sol3 × R,
g2 × R, N il3 × R,
SL
H2 × R2 , R4 ,
H2 × H2
N il4 ,
4
4
Solm6
=n , Sol0
Sol14
TABLE 1. The 4-dimensional aspherical Thurston geometries with compact representatives.
Theorem 6.1 ([17, Sections 8.5 and 9.2]). Let X3 be a 3-dimensional aspherical geometry. A
closed 4-manifold possessing the geometry X3 × R is finitely covered by a product N × S 1 , where
N is a closed oriented 3-manifold carrying the geometry X3 .
The geometry H2 × H2 can be realized both by manifolds that are virtual products of two closed
hyperbolic surfaces and by manifolds that are not even (virtual) surface bundles. These two types
are known as the reducible and the irreducible H2 × H2 geometry respectively; see [17, Section
9.5] for further details and characterizations.
Solvable non-product geometries. Finally, there exist four aspherical non-product geometries of
solvable type. Below, we describe their model Lie groups.
4
3
The nilpotent Lie
group Nil
is defined as the semi-direct product R ⋊ R, where R acts on
0 t 0
3
R by t 7→ exp 0 0 t . The model spaces for the three non-product solvable – but not
0 0 0
nilpotent – geometries are defined as follows:
Let m and n be positive integers and a > b > c reals such that a + b + c = 0 and ea , eb , ec are
4
roots of the equation Pm,n (λ) = λ3 − mλ2 + nλ − 1 = 0. If m 6= n, the Lie group Solm6
=n is
at
e
0 0
defined as R3 ⋊R, where R acts on R3 by t 7→ 0 ebt 0 . We remark that the case m = n
0 0 ect
gives b = 0 and corresponds to the product geometry Sol3 × R.
22
CHRISTOFOROS NEOFYTIDIS
If we require two equal roots of the polynomial Pm,n , then we obtain the model space of the
4
Sol
again defined as R3 ⋊ R, where now the action of R on R3 is given by t 7→
0 geometry,
et 0
0
t
0 .
0 e
0 0 e−2t
The last solvable model space is an extension of R by the 3-dimensional Heisenberg group Nil3 .
4
3
3
Namely,the Lie group Sol
1 is defined as the semi-direct product Nil ⋊ R, where R acts on Nil
1 e−t x z
by t 7→ 0
1
et y .
0
0
1
4
3
Every closed Sol04 - or Solm6
=n -manifold is a mapping torus of a self-homeomorphism of T
and every closed oriented Nil4 - or Sol14 -manifold is a mapping torus of a self-homeomorphism
of a Nil3 -manifold [17, Sections 8.6 and 8.7]. We note that non-orientable closed Nil4 - or
Sol14 -manifolds are not mapping tori of Nil3 -manifolds [17, Theorem 8.9]. Further details about
manifolds possessing a solvable non-product geometry, in particular concerning their fundamental
groups, will be provided while examining each geometry individually.
A crucial property for our study is that the 4-dimensional geometries are homotopically unique
by a result of Wall; cf. [37, Theorem 10.1] and [19, Prop. 1]. In particular, a closed aspherical
geometric 4-manifold M is finitely covered by a closed X4 -manifold if and only if it possesses the
geometry X4 .
6.2. Proof of Theorem E in dimensions ≤ 3. Dimension 1 gives already the first (trivial) example of a solvable group presentable by products but not IIPP, namely the infinite cyclic group,
which is the fundamental group of S 1 . In dimension 2, the only non-trivial solvable fundamental
group of an oriented manifold is the fundamental group of T 2 which is the product Z × Z. The
fundamental groups of higher genus surfaces are (non-elementary) hyperbolic and therefore not
presentable by products as shown in [20].
We now deal with infinite fundamental groups of closed 3-manifolds. First, using Epstein’s
factorization theorem for 3-manifold groups [8] and the fact that not virtually cyclic free products
are not presentable by products [21], we obtain that the fundamental group of a closed 3-manifold
is presentable by products if and only if it has virtually infinite center [22, Theorem 8]. These two
properties are moreover equivalent to M being Seifert fibered (with infinite fundamental group),
by the Seifert fiber space conjecture, which was independently proven by Gabai and by CassonJungreis. Recall that a closed 3-manifold M (possibly with finite fundamental group) is Seifert
fibered if and only if it is virtually a circle bundle over a closed oriented surface, by the works
g2 ,
of Seifert, Thurston and Scott; cf. [34, 30]. Equivalently, M carries one of the geometries SL
H2 × R, Nil3 , R3 , S 2 × R or S 3 . Thus, we have the following consequence of [22, Theorem 8]:
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
23
Proposition 6.2 ([22]). Suppose M is a closed 3-manifold with infinite fundamental group. Then
π1 (M) is presentable by products if and only if M is a virtual circle bundle over a closed oriented
g2 , H2 × R, Nil3 or R3 .
surface. Equivalently, M possesses one of the geometries SL
However, if the circle fiber of M is not (virtually) a direct factor, i.e. if M is not modeled on
H × R or R3 , then π1 (M) cannot be IIPP:
2
Proposition 6.3. The fundamental group of a non-trivial circle bundle M over a closed oriented
aspherical surface Σ is not IIPP.
Proof. If Σ is hyperbolic, then π1 (M) fits into a non-split central extension
1 −→ Z −→ π1 (M) −→ π1 (Σ) −→ 1.
In particular, π1 (M) fulfills the conditions of Theorems C and D, because non-virtually cyclic
hyperbolic groups are not presentable by products [20]. By Epstein’s factorization theorem [8] and
Stallings fibering criterion [31], the fundamental group of a non-trivial circle bundle over a closed
oriented surface is never reducible, and so Theorem D implies that π1 (M) is not IIPP.
The remaining case is when Σ has genus one, i.e. when M is a non-trivial circle bundle over T 2
(and therefore a Nil3 -manifold). In that case, π1 (M) fits into a non-split central extension
1 −→ Z −→ π1 (M) −→ Z2 −→ 1,
where C(π1 (M)) = Z. Suppose that π1 (M) is IIPP. Then we may assume that there exist nontrivial infinite-index commuting subgroups Γ1 , Γ2 ⊂ π1 (M) and a short exact sequence
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ π1 (M) −→ 1,
where ϕ is the multiplication map and Γ1 ∩ Γ2 ⊂ C(π1 (M)). (Note that both Γi are torsionfree, because π1 (M) is torsion-free.) We observe that Γ1 ∩ Γ2 cannot be trivial, otherwise π1 (M)
would be Abelian (equivalently, M would be T 3 ). This means that Γ1 ∩ Γ2 must be isomorphic
to C(π1 (M)) = Z. Moreover, since [π1 (M) : Γi ] = ∞ and π1 (M) has cohomological dimension
three, we conclude that each of the Γi is of cohomological dimension at most two [32]. Now,
Γ1 ∩ Γ2 is central in both Γi which means that the quotients Γi /(Γ1 ∩ Γ2 ) are finitely generated
and virtually free groups Fki , by a result of Bieri [1, Cor. 8.7]. Passing to finite coverings, we may
assume that these quotient groups are free and therefore the central extensions
1 −→ Γ1 ∩ Γ2 −→ Γi −→ Fki −→ 1
split. We have finally reached the absurd conclusion that π1 (M) is virtually isomorphic to a direct
product Z × Fk1 × Fk2 .
Remark 6.4. The case of nilpotent groups in the above proposition could be treated in a different
way, using the Hirsch length, however, the dimensions here were suitable to appeal to Bieri’s [1]
result on central extensions. In fact, the cohomological dimension of a finitely generated torsionfree nilpotent group coincides to its Hirsch length [16]. We will return and use the Hirsch length
of nilpotent groups in Section 6.3.3.
24
CHRISTOFOROS NEOFYTIDIS
Because the fundamental group of S 2 × S 1 is infinite cyclic and therefore not IIPP, we have now
determined all fundamental groups of closed 3-manifolds that are (not) IIPP:
Corollary 6.5. Suppose that the fundamental group of a closed oriented 3-manifold M is infinite.
Then π1 (M) is IIPP if and only if it is reducible. Equivalently, π1 (M) is a virtual product π1 (Σ) ×
Z, where Σ is a closed oriented aspherical surface.
In particular, we have proved Theorem E in dimension three.
6.3. Proof of Theorem E in dimension four. We will prove Theorem E in dimension four in two
steps. First, we will determine which closed aspherical geometric 4-manifolds have fundamental
groups (not) presentable by products:
Theorem 6.6. The fundamental group of a closed aspherical geometric 4-manifold M is presentable by products if and only if M possesses one of the geometries X3 × R, Nil4 , Sol14 or the
reducible H2 × H2 geometry.
Then, Theorem E will follow as a consequence of the next characterization:
Theorem 6.7. The fundamental group of a closed aspherical geometric 4-manifold M is reducible
if and only if it is IIPP. Equivalently, M carries one of the product geometries X3 × R or the
reducible H2 × H2 geometry.
Groups presentable by products: Proof of Theorem 6.6. We proceed by examining case by
case all aspherical geometries, following the enumeration given at the beginning of this section
(cf. Table 1).
6.3.1. Hyperbolic geometries. As we have already seen (and used in the proof of Proposition 6.3),
non-virtually cyclic hyperbolic groups are not presentable by products; we refer to [20, Prop. 3.6]
for the details. Since the fundamental groups of closed 4-manifolds with hyperbolic geometries
H4 or H2 (C) are obviously not (virtually) infinite cyclic, we deduce that they are not presentable
by products.
6.3.2. Product geometries. An equivalent formulation of Theorem 6.1 is the following:
Corollary 6.8. The fundamental group of a closed aspherical 4-manifold M carrying a product
geometry X3 × R is a virtual product π1 (N) × Z, where N is a closed aspherical X3 -manifold. In
particular, π1 (M) is presentable by products and C(π1 (M)) is virtually infinite.
As mentioned previously, the geometry H2 × H2 is an exceptional type among the product
geometries, because not every closed H2 × H2 -manifold has a finite cover which is a product of
two closed hyperbolic surfaces. This property distinguishes closed H2 × H2 -manifolds into two
classes, the reducible and the irreducible ones. Since H2 × H2 -manifolds admit metrics of nonpositive sectional curvature, Theorem 4.1 of [20] implies that irreducible lattices in the group of
isometries of H2 × H2 are not presentable by products:
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
25
Proposition 6.9. The fundamental group of a closed H2 × H2 -manifold M is presentable by products if and only if it is a virtual product of two closed hyperbolic surface groups. Equivalently, M
carries the reducible H2 × H2 geometry.
6.3.3. Solvable non-product geometries. We finally deal with solvable non-product geometries,
4
4
4
i.e. the geometries Nil4 , Solm6
=n , Sol0 and Sol1 . As we shall see, only the fundamental groups of
closed manifolds modeled on the geometries Nil4 or Sol14 are presentable by products.
The geometry Nil4 .
Proposition 6.10. A closed Nil4 -manifold M is a virtual circle bundle over a closed oriented
Nil3 -manifold and the center of π1 (M) remains infinite cyclic in finite covers.
Proof. Let M be a closed Nil4 -manifold. After possibly passing to a double cover, we may assume
that M is oriented and so π1 (M) fits into a short exact sequence
1 −→ π1 (N) −→ π1 (M) −→ Z −→ 1,
where N is a closed oriented Nil3 -manifold and a generator t ∈ Z acts by conjugation on π1 (N);
cf. [17, Sections 8.6 and 8.7]. Passing to another finite cover, if necessary, we may assume that N
is a non-trivial circle bundle over T 2 with fundamental group
π1 (N) = hx, y, z | [x, y] = z, xz = zx, yz = zyi,
where C(π1 (N)) = hzi; cf. [30].
Since M is a Nil4 -manifold, the automorphism of π1 (N)/hzi ∼
= Z2 , induced by the action of
t ∈ Z on π1 (N),
! is given (after possibly passing to another finite cover) by a matrix (conjugate to)
1 k
A=
∈ GL2 (Z), for some k 6= 0; cf. [17, Theorem 8.7]. The relation xm y n = z mn y n xm
0 1
in π1 (N) gives the following presentation of π1 (M) (see also [11] and [35, pg. 522] for further
details):
π1 (M) = hx, y, z, t | txt−1 = x, tyt−1 = xk yz l , tzt−1 = z det A = z, [x, y] = z, xz = zx, yz = zyi,
where C(π1 (M)) = hzi. Thus we have a short exact sequence
(11)
1 −→ hzi −→ π1 (M) −→ Q −→ 1,
where Q = π1 (M)/hzi = hx, y, t | [t, y] = xk , xt = tx, xy = yxi. In particular, the classifying
space BQ is a non-trivial circle bundle over T 2 and thus a Nil3 -manifold. Now, the induced
sequence of the classifying spaces corresponding to (11) implies that M is homotopically a circle
bundle over BQ. Finally, the center of π1 (M) remains infinite cyclic in finite covers, generated by
multiples of z, because k 6= 0.
Remark 6.11. Since every non-trivial nilpotent group has non-trivial center and because the property of being “nilpotent” is closed under subgroups and quotient groups, the proof of the above
26
CHRISTOFOROS NEOFYTIDIS
proposition could be obtained using the fact that every nilpotent group of cohomological dimension three is either Abelian or isomorphic to Q (as in the above proof); see Proposition 6.19,
Lemma 6.20 and Remark 6.21.
Remark 6.12. Note that π1 (M) is (virtually) an extension of Z2 = hy, ti by Z2 = hz, xi, and
so M is (virtually) a T 2 -bundle over T 2 , whose T 2 -fiber contains the S 1 -fiber of the circle bundle
S 1 −→ M −→ BQ of the above proposition. It is a result of Ue [35, Theorem B] that every closed
Nil4 -manifold is a virtual T 2 -bundle over T 2 . We refer to a work of Fukuhara and Sakamoto [11]
for a classification of T 2 -bundles over T 2 .
We furthermore observe that π1 (M) is
(virtually) an extension
of Z = hyi by Z3 = hz, x, ti,
1 −1 −l
3
where the automorphism of Z is given by 0 1 −k and it has infinite order. In particular,
0 0
1
3
M is a (virtual) mapping torus of T ; see also [17, Section 8.6].
4
4
The geometries Solm6
=n and Sol0 .
Proposition 6.13. The fundamental group of a closed 4-manifold possessing one of the geometries
4
4
Solm6
=n or Sol0 is not presentable by products.
In order to prove Proposition 6.13, we will use the concept of “acentral” subgroups which was
introduced in [21]: A subgroup A of a group Γ is called acentral if for every non-trivial element
g ∈ A the centralizer CΓ (g) is contained in A. An acentral extension is an extension of groups
1 −→ N −→ Γ −→ Q −→ 1 such that the normal subgroup N is acentral.
Proposition 6.14 ([21, Prop. 3.2]). If a group contains an infinite acentral subgroup of infinite
index, then it is not presentable by products.
Proposition 6.14 implies that if an extension 1 −→ N −→ Γ −→ Q −→ 1 is acentral, with
N and Q infinite, then Γ is not presentable by products [21, Cor. 3.3]. Also, if Γ is a semi-direct
product N ⋊θ Q, where N is non-trivial Abelian and Q is infinite acting freely outside 0 ∈ N, then
the extension 0 −→ N −→ Γ −→ Q −→ 1 is acentral and N is infinite. In particular, Γ is not
presentable by products [21, Cor. 3.5]. This gives an alternative proof that Sol3 -manifold groups
are not presentable by products [21, Section 3].
4
4
Proof of Proposition 6.13. We will show that the fundamental groups of closed Solm6
=n - or Sol0 manifolds contain acentral subgroups of infinite index.
4
4
Every manifold M with geometry modeled on Solm6
=n or Sol0 is a mapping torus of a selfhomeomorphism of T 3 [17, Section 8.6] (see also [36, 37]) and its fundamental group is a semidirect product Z3 ⋊θ Z, where the automorphism θ of Z3 is induced by the action by conjugation
of a generator t ∈ Z.
4
Now, if M is a Solm6
=n -manifold, then θ has three real distinct eigenvalues and none of them is
equal to ±1, because M is neither nilpotent nor carries the Sol3 × R geometry (which is the case
m = n); cf. [36] and [17, pg. 164/165] (or Section 6.1).
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
27
If M is a Sol04 -manifold, then θ has two complex eigenvalues that are not roots of unity and a
real eigenvalue not equal to ±1; cf. [36] and [17, pg. 164/165] (or Section 6.1).
In both cases we derive that the centralizer Cπ1 (M ) (g) of each g ∈ Z3 \ {0} is contained in Z3 (it
is actually equal to Z3 ). This means that the infinite-index normal subgroup Z3 is acentral and so
π1 (M) is not presentable by products by Proposition 6.14 or [21, Cor 3.3 or Cor 3.5].
The geometry Sol14 .
Proposition 6.15. A closed Sol14 -manifold M is a virtual circle bundle over a mapping torus of T 2
with hyperbolic monodromy. In particular, π1 (M) is presentable by products.
Proof. Let M be a closed Sol14 -manifold. After passing to a double cover, we may assume that M
is oriented, and so its fundamental group fits into a short exact sequence
1 −→ π1 (N) −→ π1 (M) −→ Z −→ 1,
where N is a closed oriented Nil3 -manifold and a generator t ∈ Z acts by conjugation on π1 (N);
cf. [17, Sections 8.6 and 8.7]. If necessary, we pass to another finite cover of M and so we can
assume that the fiber N is a non-trivial circle bundle over T 2 and that its fundamental group has a
presentation
π1 (N) = hx, y, z | [x, y] = z, xz = zx, yz = zyi
with center C(π1 (N))
! = hzi; cf. [30].
a b
Let A =
∈ GL2 (Z) be the automorphism of π1 (N)/hzi ∼
= Z2 induced by the action
c d
of t ∈ Z on π1 (N). The eigenvalues λ1 , λ2 of A satisfy det(A) = λ1 λ2 = ±1. Actually, det A = 1,
because M is oriented. Moreover, λi 6= ±1, because π1 (M) is not nilpotent; see also [17, Theorem
8.7]. We conclude that A is a hyperbolic automorphism.
Now the relation xm y n = z mn y n xm in π1 (N) implies that a presentation of the fundamental
group of M is given by
π1 (M) = hx, y, z, t | txt−1 = xa y c z k , tyt−1 = xb y d z l , tzt−1 = z det A = z,
[x, y] = z, xz = zx, yz = zyi, k, l ∈ Z,
where the infinite cyclic group generated by z is central in π1 (M). Thus we obtain a short exact
sequence
(12)
1 −→ hzi −→ π1 (M) −→ Q −→ 1,
where Q = π1 (M)/hzi = hx, y, t | txt−1 = xa y c, tyt−1 = xb y d , xy = yxi. Clearly, the group Q
fits into an extension
1 −→ Z2 −→ Q −→ Z −→ 1,
!
a
b
where t ∈ Z acts on Z2 by the hyperbolic automorphism A =
. We have now shown
c d
that Q is the fundamental group of a mapping torus of T 2 with hyperbolic monodromy, i.e. BQ is
28
CHRISTOFOROS NEOFYTIDIS
a closed oriented Sol3 -manifold. Therefore, M is homotopically a circle bundle over BQ, by the
induced sequence in homotopy corresponding to the short exact sequence (12).
The proof of Theorem 6.6 is now complete.
Groups not IIPP: Proof of Theorem 6.7. Since reducible groups are IIPP, in order to complete
the proof of Theorem 6.7 we need to show that the fundamental groups of manifolds modeled on
Nil4 or Sol14 are not IIPP.
Closed Nil4 -manifolds. Since torsion-free nilpotent groups have infinite center, an immediate consequence of Proposition 6.10 (see also Proposition 6.19) is the following:
Lemma 6.16. The fundamental group of a closed Nil4 -manifold is not a virtual product.
In the preceding section (Proposition 6.3), we showed that closed Nil3 -manifolds have fundamental groups not IIPP using Bieri’s results [1], since the cohomological dimensions of those
groups (and of their subgroups) were suitable for that purpose. Passing now one dimension higher,
we cannot appeal anymore to those results. However, instead of the cohomological dimension,
we may use the Hirsch length (which is in fact equal to the cohomological dimension for finitely
generated torsion-free nilpotent groups by a theorem of Gruenberg, cf. [16, §8.8]). The Hirsch
length generalizes the notion of the rank of free Abelian groups:
Definition 6.17 ([16]). Let Γ be a (virtually) polycyclic group with a series
Γ = Γ0 ⊃ Γ1 ⊃ · · · ⊃ Γn = 1,
so that the quotients Γi /Γi+1 are cyclic. The sum of the ranks of these quotients is independent of
the choice of the series of groups and is called the Hirsch length of Γ. We denote the Hirsch length
of Γ by h(Γ).
Example 6.18. A finitely generated torsion-free nilpotent group Γ is polycyclic, admitting a central
series Γ = Γ0 ⊃ Γ1 ⊃ · · · ⊃ Γn = 1, so that the quotients Γi /Γi+1 are infinite cyclic for all
i = 0, 1, ..., n − 1. Therefore Γ has a well-defined Hirsch length (equal to n).
The following proposition gives some basic properties of nilpotent groups and their Hirsch
length. For a proof see [27, pg. 75–77].
Proposition 6.19. Let Γ be a finitely generated nilpotent group.
(1) If Γ is torsion-free, then C(Γ) has positive rank and Γ/C(Γ) is torsion-free.
(2) If K is a normal subgroup of Γ, then K and Γ/K are finitely generated nilpotent and
(a) h(Γ) = h(K) + h(Γ/K); (b) h(K) = h(Γ) if and only if [Γ : K] < ∞.
Using Proposition 6.19, we determine all torsion-free nilpotent groups of Hirsch length three.
Lemma 6.20. A torsion-free nilpotent group Γ of Hirsch length three is isomorphic to Gn :=
hx, y, z | zy = yz, zx = xz, [x, y] = z n i, for some n ≥ 0. In particular, Γ is the fundamental
group of a circle bundle over T 2 .
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
29
Proof. First, we observe that Γ is finitely generated, because it is nilpotent of finite Hirsch length.
Moreover, since Γ is torsion-free, we have that its center C(Γ) is free Abelian of positive rank, the
quotient group Q := Γ/C(Γ) is again nilpotent and torsion-free and the short exact sequence
(13)
1 −→ C(Γ) −→ Γ −→ Q −→ 1,
yields that 0 ≤ h(Q) ≤ 2, because h(C(Γ)) ≥ 1; cf Proposition 6.19.
If h(Q) = 0 or 1, then it is easy to see that Γ is free Abelian of rank three. Indeed, this is obvious
if h(Q) = 0. If h(Q) = 1, then Q is infinite cyclic, and therefore the central extension (13) splits.
Suppose, finally, that h(Q) = 2. Since Q is torsion-free nilpotent, it has non-trivial center C(Q).
Therefore, it fits into a short exact sequence
1 −→ C(Q) −→ Q −→ Q/C(Q) −→ 1,
where the quotient Q/C(Q) is again a torsion-free nilpotent group. By the additivity of the Hirsch
length for the above exact sequence, we deduce that h(Q/C(Q) ≤ 1. This finally implies that Q
is free Abelian of rank two. Therefore, the central extension (13) takes the form
1 −→ Z −→ Γ −→ Z2 −→ 1.
Choosing presentations Z = hzi and Z2 = hx, y | [x, y] = 1i, we deduce that Γ is isomorphic to
Gn for some n ≥ 0.
Remark 6.21. In the light of Gruenberg’s theorem [16, §8.8], Lemma 6.20 determines all finitely
generated nilpotent groups of cohomological dimension three. Moreover, it yields another proof of
the fact that closed Nil4 -manifolds are virtual circle bundles over closed oriented Nil3 -manifolds;
compare Proposition 6.10.
We now finish the proof that Nil4 -manifold groups are not IIPP:
Proposition 6.22. The fundamental group of a closed Nil4 -manifold M is not IIPP.
Proof. We know that π1 (M) is presentable by products and that M is virtually a non-trivial circle
bundle over a closed oriented Nil3 -manifold (cf. Proposition 6.10). We need to prove that π1 (M)
cannot be presented by a product of subgroups Γ1 and Γ2 so that both Γi have infinite index in
π1 (M). We proceed again by contradiction. After passing to suitable finite index subgroups,
suppose that there exist two infinite-index commuting subgroups Γi ⊂ π1 (M) and a short exact
sequence
(14)
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ π1 (M) −→ 1,
where ϕ is the multiplication map and Γ1 ∩ Γ2 ⊂ C(π1 (M)); cf. Section 2.3. Also we have
that C(π1 (M)) = Z (by Proposition 6.10) and that the Γi are finitely generated, torsion-free and
nilpotent.
Since π1 (M) is not reducible (Lemma 6.16), we conclude that the intersection Γ1 ∩ Γ2 is not
trivial and so it must be infinite cyclic, as a subgroup of C(π1 (M)) = Z. Also, Proposition 6.19
and Lemma 6.20 imply that the Hirsch length of π1 (M) is h(π1 (M)) = 4. Applying again the
30
CHRISTOFOROS NEOFYTIDIS
additivity of the Hirsch length to the short exact sequence (14), we have that h(Γ1 ×Γ2 ) = 5. Since
both Γi have infinite index in π1 (M), we deduce that h(Γi ) ≤ 3 by Proposition 6.19. Therefore, one
of the Γi must be of Hirsch length three and the other of Hirsch length two (again by Proposition
6.19). Let us assume that h(Γ1 ) = 3 and h(Γ2 ) = 2. Since Γ1 is a torsion-free nilpotent group
of Hirsch length three, Lemma 6.20 implies that Γ1 isomorphic to Gn for some n ≥ 0. Moreover,
Γ2 is isomorphic to Z2 , because it is torsion-free nilpotent of Hirsch length two (see the proof of
Lemma 6.20). We have now reached the conclusion that the rank of the center of Γ1 × Γ2 is at least
three. This is however not possible, according to the next lemma, because Γ1 × Γ2 is an extension
of π1 (M) by Z and C(π1 (M)) = Z.
Lemma 6.23. If a group Γ with finitely generated center C(Γ) fits into a central extension 1 −→
π
Zk −→ Γ −→ Q −→ 1, where Q is torsion-free, then rankC(Γ) ≤ rankC(Q) + k.
Proof. It follows by the fact that if x ∈ C(Γ), then π(x) ∈ C(Q).
Closed Sol14 -manifolds. We finally deal with the Sol14 geometry.
Proposition 6.24. The fundamental group of a closed Sol14 -manifold is not IIPP.
Proof. By Proposition 6.15, a closed Sol14 -manifold M is a virtual circle bundle over a closed
oriented Sol3 -manifold N. In particular, its fundamental group π1 (M) satisfies all the assumptions
of Theorem D (recall that π1 (N) is not presentable by products; see Proposition 6.2). Thus, π1 (M)
is IIPP if and only if it is a virtual product π1 (N)×Z. The latter is impossible by Wall’s uniqueness
theorem [37, Theorem 10.1].
This finishes the proof of Theorem 6.7 and therefore the proof of Theorem E.
7. D OMINATION
BY PRODUCTS IN DIMENSION FOUR ( PROOF OF
T HEOREM F)
In this section we determine which geometric 4-manifolds are (not) dominated by products. To
this end, we combine the algebraic results of the previous section with the topological statements
of Theorem 2.4 and Theorems B and C to prove Theorem F, which completes Hillman’s characterization of product geometries given in Theorem 6.1.
Proof of theorem F. Theorem 6.1 says that closed manifolds possessing one of the product geometries X3 × R are finitely covered by products of type N × S 1 . Moreover, closed manifolds carrying
the reducible H2 × H2 geometry are virtually products of closed hyperbolic surfaces. In particular,
all those manifolds are dominated by products. (Note that domination by products alone follows
by Theorem A as well.)
By Wall’s uniqueness theorem of the 4-dimensional geometries, it suffices to show that closed
4-manifolds possessing either a hyperbolic geometry, the irreducible H2 × H2 geometry, or a nonproduct solvable geometry cannot be dominated by products.
For the hyperbolic geometries H4 and H2 (C), the irreducible H2 ×H2 geometry, and the solvable
4
4
geometries Solm6
=n and Sol0 our claim can be deduced by Theorem 2.4, because the fundamental
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
31
groups of closed 4-manifolds carrying one of those geometries are not presentable by products; see
Section 6.3 for the details.
If now M is a closed Nil4 -manifold, then it is a virtual circle bundle over a closed oriented
Nil3 -manifold and the center of its fundamental group remains infinite cyclic in finite covers; cf.
Proposition 6.10. Moreover, π1 (M) is not IIPP, by Proposition 6.22, and so Theorem B implies
that M cannot be dominated by products.
Finally, if M is a closed Sol14 -manifold, then it is a virtual circle bundle over a closed oriented
Sol3 -manifold and π1 (M) is not IIPP; cf. Propositions 6.15 and 6.24 respectively. Therefore, M
is not dominated by products by Theorem C, because closed Sol3 -manifolds have fundamental
groups not presentable by products. (Equivalently, M is not dominated by products because of the
equivalence between (1) and (2) in Theorem C and by Wall’s uniqueness theorem.)
The proof is now complete.
Combining Theorem E with the characterizations of groups (infinite-index) presentable by products in Section 6, we obtain the following purely algebraic characterization:
Corollary 7.1. A closed oriented aspherical geometric 4-manifold M is dominated by a product
if and only if
(1) either π1 (M) is a virtual product π1 (N) × Z, for some closed aspherical geometric 3manifold N, or
(2) π1 (M) is a virtual product of two closed hyperbolic surface groups.
8. T HE 5- DIMENSIONAL H EISENBERG
MANIFOLD
In this section we give the first example of an aspherical manifold whose fundamental group is
IIPP and irreducible, and show that this manifold is not dominated by products.
Lemma 8.1. The 5-dimensional Heisenberg group
H5 = hx, y, u, v, z | [x, y] = [u, v] = z, all other commutei
is IIPP and irreducible.
Proof. We observe that the groups Γ1 = hx, y, z | [x, y] = z, xz = zx, yz = zyi and Γ2 =
hu, v, z | [u, v] = z, uz = zu, vz = zvi are infinite index subgroups of H5 that commute
element-wise, and Γ1 Γ2 = H5 . In particular, H5 is presented by the product Γ1 × Γ2 through
the multiplication homomorphism Γ1 × Γ2 −→ H5 . Thus H5 is IIPP.
Now we show that H5 is irreducible. First of all, H5 is not a product itself, otherwise its center
would have been of rank greater than 1, because subgroups of nilpotent groups are nilpotent themselves and torsion-free nilpotent groups have infinite center. Second, every finite index subgroup
H5 ⊂ H5 fits into a short exact sequence
1 −→ hz k i −→ H5 −→ Q −→ 1,
32
CHRISTOFOROS NEOFYTIDIS
where Q is a finite index subgroup of Z4 = H5 /hzi. In particular, Q is isomorphic to Z4 , generated
by some powers of x, y, u and v. If H5 were a direct product, then the relations [xm , y n ] = z mn and
[um , v n ] = z mn in H5 would imply that z is a torsion element. However, H5 is torsion-free and the
lemma follows.
The classifying space of H5 is a non-trivial circle bundle over T 4 , called the 5-dimensional
Heisenberg manifold.
Theorem 8.2. The Heisenberg manifold BH5 is not dominated by products.
Proof. Suppose, for contrast, that there exists a map of non-zero degree f : X1 × X2 −→ BH5
which is π1 -surjective, after possibly replacing BH5 by a finite cover; this cover will remain a
non-trivial circle bundle over T 4 with center C(H5 ) ∼
= Z, by Lemma 8.1. We have a short exact
sequence
(15)
ϕ
1 −→ Γ1 ∩ Γ2 −→ Γ1 × Γ2 −→ H5 −→ 1,
where Γi := im(π1 (f |Xi )) ⊂ H5 , Γ1 ∩ Γ2 ⊂ C(H5 ) = Z and ϕ is the multiplication homomorphism; see Sections 2.2 and 2.3 for the details. In particular, Γ1 ∩ Γ2 is isomorphic to Z, because
H5 is irreducible. Since the Γi are finitely generated torsion-free nilpotent groups, the above exact
sequence implies that it suffices (up to order) to examine what happens when the Hirsch length of
Γ1 takes the values one, two and three.
Case I. Suppose first that h(Γ1 ) = 1. Then X1 = S 1 (see the proof of Theorem C), and so
S 1 × X2 ≥ BH5 . The latter is not possible by the Factorization Lemma 4.8.
Case II. Next, suppose that h(Γ1 ) = 2. Then Γ1 is isomorphic to Z2 (see the proof of Lemma 6.20).
Since Γ2 is nilpotent and torsion-free, its center has positive rank. Thus rankC(Γ1 ×Γ2 ) ≥ 3, which
contradicts Lemma 6.23 because Γ1 × Γ2 fits into the short exact sequence (15), where Γ1 ∩ Γ2 and
C(H5 ) are both infinite cyclic.
Case III. The only remaining case is when h(Γ1 ) = 3. In that case, h(Γ2 ) = 3. By Lemma 6.20,
each of the groups Γi is isomorphic to some Gni = hx, y, z | [x, y] = z ni , xz = zx, yz = zyi, for
ni ≥ 0. As in Case II, none of the Γi can be Abelian of rank greater than 1, by Lemma 6.23. We
conclude that ni ≥ 1. According to the proof of Theorem 2.4 (cf. Section 2.2) and because BH5
is aspherical, there exist two non-trivial homology classes αi ∈ Hdim Xi (BΓi ; Q) such that
H5 (Bϕ)(α1 × α2 ) = deg(f ) · [BH5 ].
Moreover, we know by Case I that one of the Xi must have dimension two and the other dimension
three. Without loss of generality, suppose that dim X1 = 3 and dim X2 = 2. Recall that, for n ≥ 1,
every Gn is realizable by a non-trivial circle bundle over T 2 . Therefore the cycle α1 ∈ H3 (BΓ2 ; Q)
is realized by such a nilpotent 3-manifold N and the cycle α2 ∈ H2 (BΓ2 ; Q) is realized by T 2 .
This means that there exists a continuous composite map
T 2 × N −→ BΓ1 × BΓ2 −→ BH5 ,
FUNDAMENTAL GROUPS OF ASPHERICAL MANIFOLDS AND MAPS OF NON-ZERO DEGREE
33
which in degree 5 homology maps the fundamental class [T 2 × N] to a non-trivial multiple of
[BH5 ]. In particular, the product S 1 × (S 1 × N) dominates BH5 which is impossible as we have
seen in Case I.
After having examined all the possibilities for the groups Γ1 and Γ2 , we conclude that BH5
cannot be dominated by products.
R EFERENCES
1. R. Bieri, Homological dimension of discrete groups, Second edition. Queen Mary College Mathematical Notes,
Queen Mary College, Department of Pure Mathematics, London, 1981.
2. J. Bowden, The topology of symplectic circle bundles, Trans. Amer. Math. Soc., 361 (2009), no. 10, 5457–5468.
3. S. Cappell, S. Weinberger and M. Yan, Closed aspherical manifolds with center, J. Topol., 6 (2013), no. 4, 1009–
1018.
4. J. A. Carlson and D. Toledo, Harmonic mappings of Kähler manifolds to locally symmetric spaces, Inst. Hautes
Études Sci. Publ. Math., 69 (1989), 173–201.
5. P. Eberlein, Lattices in spaces of nonpositive curvature, Ann. of Math. (2), 111 (1980), no. 3, 435–476.
6. P. Eberlein, Isometry groups of simply connected manifolds of nonpositive curvature II, Acta Math., 149 (1982),
no. 1-2, 41–69.
7. S. Eilenberg, On the problems of topology, Ann. of Math. (2), 50 (1949), 247–260.
8. D. B. A. Epstein, Factorization of 3-manifolds, Comment. Math. Helv., 36 (1961), 91–102.
9. B. Farb and S. Weinberger, Isometries, rigidity and universal covers, Ann. of Math., 186 (2008), no. 3, 915–940.
10. R. Filipkiewicz, Four-dimensional geometries, PhD thesis, University of Warwick, 1983.
11. S. Fukuhara and K. Sakamoto, Classification of T 2 -bundles over T 2 , Tokyo J. Math., 6 (1983), no. 2, 310–327.
12. D. Gromoll and J. A. Wolf, Some relations between the metric structure and the algebraic structure of the fundamental group in manifolds of nonpositive curvature, Bull. Amer. Math. Soc., 77 (1971), 545–552.
13. M. Gromov, Volume and bounded cohomology, Inst. Hautes Études Sci. Publ. Math., 56 (1982), 5–99.
14. M. Gromov, Metric Structures for Riemannian and Non-Riemannian Spaces, With appendices by M. Katz, P.
Pansu and S. Semmes, translated from the French by S. M. Bates, Progress in Mathematics 152, Birkhäuser
Verlag, 1999.
15. M. Gromov, Singularities, expanders and topology of maps. I. Homology versus volume in the spaces of cycles,
Geom. Funct. Anal., 19 (2009), no. 3, 743–841.
16. K. W. Gruenberg, Cohomological topics in group theory, Lecture Notes in Mathematics 143, Springer-Verlag,
Berlin-New York, 1970.
17. J. A. Hillman, Four-manifolds, geometries and knots, Geometry and Topology Monographs 5, Coventry, 2002.
18. M. Kapovich and B. Leeb, Actions of discrete groups on nonpositively curved spaces, Math. Ann., 306 (1996),
no. 2, 341–352.
19. D. Kotschick, Remarks on geometric structures on compact complex surfaces, Topology, 31 (1992), no. 2, 317–
321.
20. D. Kotschick and C. Löh, Fundamental classes not representable by products, J. London Math. Soc., 79 (2009),
545–561.
21. D. Kotschick and C. Löh, Groups not representable by products, Groups Geom. Dyn., 7 (2013), 181–204.
22. D. Kotschick and C. Neofytidis, On three-manifolds dominated by circle bundles, Math. Z., 274 (2013), 21–32.
23. W. Lück, L2 -invariants: theory and applications to geometry and K-theory, Springer-Verlag, Berlin, 2002.
24. W. Lück, Survey on aspherical manifolds, European Congress of Mathematics, Eur. Math. Soc., 53–82, 2010.
25. J. Milnor and W. Thurston Characteristic numbers of 3-manifolds Enseignement Math. (2), 23 (1977), no. 3-4,
249–254.
34
CHRISTOFOROS NEOFYTIDIS
26. C. Neofytidis, Branched coverings of simply connected manifolds, Topol. Appl., 178 (2014), 360–371.
27. C. Neofytidis, Non-zero degree maps between manifolds and groups presentable by products, Munich thesis,
2014, available online at http://edoc.ub.uni-muenchen.de/17204/.
28. C. Neofytidis, Ordering Thurston’s geometries by maps of non-zero degree, J. Topol. Anal. (to appear).
29. I. Schur, Über die Darstellung der endlichen Gruppen durch gebrochene lineare Substitutionen, Journal für die
reine und angewandte Mathematik, 127 (1904), 20–50.
30. P. Scott, The geometries of 3-manifolds, Bull. London Math. Soc., 15 (1983), 401–487.
31. J. Stallings, On fibering certain 3-manifolds, In Topology of 3-manifolds and related topics (Proc. The Univ. of
Georgia Institute, 1961), 95–100, Prentice-Hall, Englewood Cliffs, N.J., 1962.
32. R. Strebel, A remark on subgroups of infinite index in Poincaré duality groups, Comment. Math. Helv., 52 (1977),
no. 3, 310–314.
33. R. Thom, Quelques propriétés globales des variétés différentiables, Comment. Math. Helv., 28 (1954), 17–86.
34. W. P. Thurston, Three-Dimensional Geometry and Topology, Princeton University Press, 1997.
35. M. Ue, Geometric 4-manifolds in the sense of Thurston and Seifert 4-manifolds I, J. Math. Soc. Japan, 42 (1990),
511–540.
36. C. T. C. Wall, Geometries and geometric structures in real dimension 4 and complex dimension 2, In Geometry
and Topology (College Park, Md., 1983/84), Lecture Notes in Math., 1167, Springer, Berlin, (1985), 268–292.
37. C. T. C. Wall, Geometric structures on compact complex analytic surfaces, Topology, 25 (1986), no. 2, 119–153.
38. S. Wang, The existence of maps of non-zero degree between aspherical 3-manifolds, Math. Z., 208 (1991), 147–
160.
39. W. Zhang, Geometric structures, Gromov norm and Kodaira dimensions, Adv. Math. 308 (2017), 1–35.
D EPARTMENT OF M ATHEMATICAL S CIENCES , SUNY B INGHAMTON , B INGHAMTON , NY 13902-6000, USA
E-mail address: [email protected]
| 4math.GR
|
Distributed Cooperative Decision-Making in Multiarmed Bandits:
Frequentist and Bayesian Algorithms
arXiv:1606.00911v2 [cs.SY] 24 Sep 2016
Peter Landgren, Vaibhav Srivastava, and Naomi Ehrich Leonard
Abstract— We study distributed cooperative decision-making
under the explore-exploit tradeoff in the multiarmed bandit
(MAB) problem. We extend state-of-the-art frequentist and
Bayesian algorithms for single-agent MAB problems to cooperative distributed algorithms for multi-agent MAB problems in
which agents communicate according to a fixed network graph.
We rely on a running consensus algorithm for each agent’s
estimation of mean rewards from its own rewards and the
estimated rewards of its neighbors. We prove the performance
of these algorithms and show that they asymptotically recover
the performance of a centralized agent. Further, we rigorously
characterize the influence of the communication graph structure on the decision-making performance of the group.
I. I NTRODUCTION
Cooperative decision-making under uncertainty is ubiquitous in natural systems as well as in engineering networks.
A fundamental feature of decision-making under uncertainty
is the explore-exploit tradeoff: the decision-making agent
needs to learn the unknown system parameters (exploration),
while maximizing its parameter-dependent decision-making
objective (exploitation).
Multiarmed bandit (MAB) problems are canonical formulations of the explore-exploit tradeoff. In a stochastic MAB
problem a set of options (arms) is given. A stochastic reward
with an unknown mean is associated with each option. A
player can pick only one option at a time, and the player’s
objective is to maximize the cumulative expected reward over
a sequence of choices. In an MAB problem, the player needs
to balance the tradeoff between learning the mean reward at
each arm (exploration), and picking the arm with maximum
mean reward (exploitation).
MAB problems are pervasive across a variety of scientific
communities and have found application in diverse areas
including control and robotics [1], [2], ecology [3], [4], and
communications [5]. Despite the prevalence of the MAB
problem, the research on MAB problems has primarily
focused on policies for a single agent. The increasing importance of networked systems warrants the development
of distributed algorithms for multiple communicating agents
faced with MAB problems. In this paper, we build upon previous work by extending two popular single-agent algorithms
for the stochastic MAB problem to the distributed multiple
This research has been supported by ONR grant N00014-14-1-0635, ARO
grant W911NF-14-1-0431, and the DoD through the NDSEG Program.
P. Landgren and N. E. Leonard are with the Department of Mechanical
and Aerospace Engineering, Princeton University, Princeton, NJ, USA,
{landgren, naomi}@princeton.edu.
V. Srivastava is with the Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI, USA,
[email protected]
agent setting and analyze decision-making performance as a
function of the network structure.
The MAB problem has been extensively studied (see [6]
for a survey). In their seminal work, Lai and Robbins [7]
established a logarithmic lower bound on the expected number of times a sub-optimal arm needs to be selected by an
optimal policy in a frequentist setting. In another seminal
work, Auer et al. [8] developed the upper confidence bound
(UCB) algorithm for the stochastic MAB problem, which
achieves the lower bound in [7] uniformly in time.
The MAB problem has also been studied in the Bayesian
setting. Kaufmann et al. [9] proposed the Bayes-UCB algorithm and showed that it achieves Lai-Robbins’ lower bound
for Bernoulli rewards and uninformative priors. Reverdy et
al. [10] developed and analyzed the upper credible limit
(UCL) algorithm for correlated multiarmed bandits by applying the approach of [9] to the case of Gaussian rewards.
The classical single-agent MAB problem was extended by
Anantharam et al. [11] to the setting of multiple centralized
players. Recently, researchers [12], [13], [5] have studied the
decentralized multi-player MAB problem with no communication among agents. Kar et al. [14] investigated the multiagent MAB problem in a leader-follower setting. Here, we
use a running consensus algorithm [15] for assimilation of information. Running consensus has been used to study related
collective decision-making models in social networks [16].
In the present paper we study the distributed cooperative
MAB problem in which agents are faced with a stochastic
MAB problem and communicate their information with their
neighbors in an undirected and connected communication
graph. We use a set of running consensus algorithms for
cooperative estimation of the mean reward at each arm, and
we design an arm selection heuristic that leads to an orderoptimal performance for the group. The major contributions
of this paper are as follows.
First, we propose and thoroughly analyze the coop-UCB2
and coop-UCL algorithms. We derive bounds on decisionmaking performance for the group and characterize the
influence of the network structure on performance. To predict
nodal performance, we propose a measure of nodal “exploreexploit centrality,” which depends on the location of the node
in the graph.
Second, we demonstrate that the ordering of nodes by
performance predicted by our explore-exploit centrality measure matches the order obtained using numerical simulations.
We also show that the incorporation of priors that are wellinformative about the correlation structure markedly improve
performance.
The remainder of the paper is organized as follows. In
Section II we introduce the cooperative MAB problem. In
Section III we recall a cooperative estimation algorithm. We
review the coop-UCB algorithm in Section IV, and propose
and analyze the improved coop-UCB2 and new coop-UCL
algorithms. We illustrate our analytic results with numerical
examples in Section VI. We conclude in Section VII.
II. C OOPERATIVE M ULTIARMED BANDIT P ROBLEM
Consider an MAB problem with N arms and M decisionmaking agents. The reward associated with arm i ∈
{1, . . . , N } is a random variable with an unknown mean mi .
Let the communication topology of agents be modeled by an
undirected graph G in which each node represents an agent
and edges represent the communication between agents. Let
A ∈ RM ×M be the adjacency matrix associated with G and
let L ∈ RM ×M be the corresponding Laplacian matrix. We
assume that the graph G is connected, i.e., there exists a path
between each pair of nodes.
Let the k-th agent choose arm ik (t) at time t ∈ {1, . . . , T }
and receive a reward rk (t). The objective of each decisionmaker k is to choose using its local information a sequence
of arms {ik (t)}t∈{1,...,T } such that the total expected cumuPM PT
lative reward k=1 t=1 mik (t) is maximized, where T is
the horizon length of the sequential allocation process.
For an MAB problem, the expected regret of agent k
at time t is defined by Rk (t) = mi∗ − mik (t) , where
mi∗ = max{mi | i ∈ {1, . . . , N }}. The collective
objective of the M decision-makers can be equivalently
defined
minimizing the expected
regret defined
PMas P
PM Pcumulative
T
N
k
k
by
R
(t)
=
∆
E[n
i
i (T )], where
k=1
t=1
k=1
i=1
k
ni (T ) is the total cumulative number of times arm i has
been chosen by agent k until time T and ∆i = mi∗ − mi
is the expected regret due to picking arm i instead of arm
i∗ . It is known that the regret of any algorithm for an MAB
problem is asymptotically lower bounded by a logarithmic
function of the horizon length T [7], [11], i.e., no algorithm
can achieve an expected cumulative regret smaller than a
logarithmic function of horizon length as T → ∞.
In this paper, we focus on Gaussian rewards, i.e., the
reward at arm i is sampled from a Gaussian distribution
with mean mi and variance σs2 . We assume that the variance
σs2 is known and is the same at each arm. In the context
of Gaussian rewards, the lower bound [11] on the expected
number of times a suboptimal arm i is selected by a fusion
center that has access to reward for each agent is
M
X
k=1
E[nki (T )] ≥
2σ 2
s
∆2i
+ o(1) ln T.
(1)
In the following, we will design policies that sample a
suboptimal arm i within a constant factor of the above bound.
III. C OOPERATIVE E STIMATION OF M EAN R EWARDS
In this section we recall the algorithm for cooperative estimation of mean rewards proposed in our earlier work [17].
A. Cooperative Estimation Algorithm
For distributed cooperative estimation of the mean reward
at each arm i, we employ two running consensus algorithms
to estimate (i) total reward provided at the arm, and (ii) the
total number of times the arm has been sampled.
Let ŝki (t) and n̂ki (t) be agent k’s estimate of the total
reward provided at arm i per unit agent and the total number
of times arm i has been selected until time t per unit agent,
respectively. Using ŝki (t) and n̂ki (t) agent k can calculate
µ̂ki (t), the estimated empirical mean of arm i at time t as
µ̂ki (t) =
ŝki (t)
.
n̂ki (t)
(2)
Let ik (t) be the arm sampled by agent k at time t and let
ξik (t) = 1(ik (t) = i) where 1(·) is the indicator function,
which is equal to 1 if ik (t) = i and 0 otherwise. Define
rik (t) = rnd · 1(ik (t) = i), where rnd is a random variable
sampled from N (mi , σs2 ). Let P be a row stochastic matrix
given by
κ
P = IM −
L,
(3)
dmax
where IM is the identity matrix of order M , κ ∈ (0, 1]
is a step size parameter [18], dmax = max{deg(k) | k ∈
{1, . . . , M }}, and deg(k) is the degree of node k.
The estimates n̂ki (t) and ŝki (t)are updated locally using
running consensus [15] as follows:
n̂i (t + 1) = P n̂i (t) + P ξi (t),
and ŝi (t + 1) = Pŝi (t) + P ri (t),
(4)
(5)
where n̂i (t), ŝi (t), ξi (t), and ri (t) are vectors of n̂ki (t),
ŝki (t), ξik (t), and rik (t), k ∈ {1, . . . , M }, respectively.
B. Analysis of the Cooperative Estimation Algorithm
We now recall the performance of the
algorithm
Ptestimation
1
>
1
ξ
(τ
) be the
defined by (2–5). Let ncent
(t)
≡
i
i
τ =0 M
M
total number of times arm i has been
selected
per
unit
agent
Pt
1
>
1
r
(τ
)
be
the
total
until time t, and let scent
i
i (t) ≡ M
M
τ =0
reward provided at arm i per unit agent until time t. Also, let
λi denote the i-th largest eigenvalue of P , ui the eigenvector
corresponding to λi , udi the d-th entry of ui , and
√
M
X
|λp |
.
1
−
|λp |
p=2
√
Note that λ1 = 1 and u1 = 1M / M . Let us define
n =
+sum
νpj
=
M
M
X
(6)
udp udj 1((up u>
j )kk ≥ 0)
d=1
-sum
and νpj
=
M
X
udp udj 1((up u>
j )kk ≤ 0).
d=1
We also define
+sum
>
>
νpj (up uj )kk , if λp λj ≥ 0 & (up uj )kk ≥ 0,
-sum
bpj (k) = νpj
(up u>
if λp λj ≥ 0 & (up u>
j )kk ,
j )kk ≤ 0,
max
>
νpj |(up uj )kk |, if λp λj < 0,
(7)
max
-sum
+sum
where νpj
= max {|νpj
|, νpj
}. Furthermore, let
kc = M
M X
M
X
p=1 j=2
|λp λj |
bpj (k).
1 − |λp λj |
We note that both n and kc depend only on the topology
of the communication graph and are measures of distributed
cooperative estimation performance. We now recall the following result from [17].
Proposition 1 (Performance of cooperative estimation):
For the distributed estimation algorithm defined in (2–5),
the following statements hold
(i) the estimate n̂ki (t) satisfies
k
cent
ncent
i (t) − n ≤ n̂i (t) ≤ ni (t) + n ;
(ii) the estimator µ̂ki (t) is unbiased, i.e., E[µ̂ki (t)] = mi ;
(iii) the variance of the estimate µ̂ki (t) satisfies
Var[µ̂ki (t)] ≤
σs2 n̂ki (t) + kc
.
M (n̂ki (t))2
IV. F REQUENTIST C OOPERATIVE D ECISION -M AKING
In this section, we first review the coop-UCB algorithm
proposed in our earlier work [17]. We then improve on this
algorithm with a new algorithm: coop-UCB2. Unlike coopUCB the improved algorithm does not require each agent to
know the global graph structure. Finally, we compute bounds
on the performance of the group for this algorithm as a
function of the graph structure.
A. The coop-UCB Algorithm
The coop-UCB algorithm is analogous to the UCB algorithm [8], and uses a modified decision-making heuristic
that captures the effect of the additional information an agent
receives through communication with other agents as well as
the rate of information propagation through the network.
The coop-UCB algorithm is initialized by each agent
sampling each arm once and proceeds as follows. At each
time t, each agent k updates its cooperative estimate of the
mean reward at each arm using the distributed cooperative
estimation algorithm described in (2–5). Then, at time t each
agent k selects the arm with maximum Qki (t) = µ̂ki (t) +
Cik (t), where
s
n̂k (t) + k ln t
k
Ci (t) = σs 2γ i k c · k .
(8)
M n̂i (t) n̂i (t)
The coop-UCB provides a distributed, cooperative solution
to the MAB problem such that every agent in the network
achieves logarithmic regret. However, the heuristic Qki may
be overly restrictive in the sense that it requires the agent k to
know kc , which depends on the global graph structure. Further, agents with a relatively high kc are essentially forced to
explore more while better positioned agents exploit, leading
to wide disparities in performance across some networks. We
will develop the coop-UCB2 algorithm that addresses these
issues in the next section.
B. The coop-UCB2 Algorithm
The coop-UCB2 algorithm is initialized by each agent
sampling each arm once and proceeds as follows. At each
time t, each agent k updates its cooperative estimate of the
mean reward at each arm using the distributed cooperative
estimation algorithm described in (2–5). Then, at time t each
agent k selects the arm with maximum Qki (t) = µ̂ki (t) +
Cik (t), where
s
n̂k (t) + f (t) ln t
· k ,
(9)
Cik (t) = σs 2γ i
M n̂ki (t)
n̂i (t)
f (t) is an increasing sublogarthmic function, and γ > 1.
Note that the heuristic Qki requires the agent k to know the
total number of agents M , but not the global graph structure.
Theorem 1 (Regret of the coop-UCB2 Algorithm): For
the coop-UCB2 algorithm and the cooperative Gaussian
MAB problem, the number of times a suboptimal arm i is
selected by all agents until time T satisfies
s
M
X
4γσs2
∆2i M f (T )
k
E[ni (T )] ≤
1
+
1
+
ln T
∆2i
2γσs2 ln T
k=1
+M n +
M
X
1 + k
c
k=1
kc
k
c
(t†k ) 1+kc +
(t†k )1−γ
1
1
− k + † ,
γ−1
c
tk
where t†k = f −1 (kc ).
Proof: We proceed similarly to [8]. The number of selections
of a suboptimal arm i by all agents until time T is
M
X
k=1
nki (T ) ≤
M X
T
X
1(Qki (t) ≥ Qki∗ (t))
k=1 t=1
M X
T
X
≤η+
1(Qki (t) ≥ Qki∗ (t), M ncent
≥ η),
i
k=1 t=1
(10)
where η > 0 is a constant that will be chosen later.
At a given time t an individual agent k will choose a
suboptimal arm only if Qki (t) ≥ Qki∗ (t). For this condition
to be true at least one of the following three conditions must
hold:
µ̂i∗ (t) ≤ mi∗ − Cik∗ (t)
(11)
Cik (t)
2Cik (t).
(12)
µ̂i (t) ≥ mi +
mi∗ < mi +
(13)
We now bound the probability that (11) holds:
E[µ̂i∗ (t)] + Cik∗ (t) − mi∗
p
P((11) holds) = P z ≥
Var(µ̂ki∗ (t))
C k∗ (t)
=P z≥ p i
Var(µ̂ki∗ (t))
2
1
(Cik∗ (t))
≤ exp −
,
(14)
2
2Var(µ̂ki∗ (t))
where z is the standard Gaussian random variable and the
last inequality follows from the tail bounds for the error
function [19] and the second statement of Proposition 1.
We now examine the argument of (14). Note that
s
2γ ln t n̂ki∗ (t) + f (t)
1
Cik∗ (t)
p
= σs
k
M
n̂ki∗ (t)
Var(µ̂ki∗ (t))
Var(µ̂i∗ (t))
s
n̂i∗ (t) + f (t)
≥ σs 2γ ln t ·
.
n̂i∗ (t) + kc
Now for t > t†k ,
n̂i∗ (t)+f (t)
n̂i∗ (t)+k
c
≥ 1, and therefore
P((11) holds | t ≥ t†k ) ≤
the group as a function of the graph structure, and the
logarithmic portion of the bound is independent of agent
location. However, the constant factor is dependent on
the agent’s position in the network since it depends on
kc . In this sense, kc can be thought of as a measure of
“explore-exploit centrality,” which indicates that agents with
a higher kc will contribute more to the group’s regret.
1
.
2tγ
If t < t†k , then
(t)
1 − n̂n̂i∗∗+f
1 − 1
t i +kc ≤ t 1+kc .
2
2
P((11) holds | t < t†k ) ≤
It follows analogously that
V. BAYESIAN C OOPERATIVE D ECISION -M AKING
1
P((12) holds | t ≥ t†k ) ≤
2t
1
1 − 1+
†
k
c .
P((12) holds | t < tk ) ≤ t
2
We now examine the event (13).
mi∗ < mi + 2Cik (t)
∆2 M
=⇒ n̂ki (t)2 i 2 − γ n̂ki (t) ln(t) − γf (t) ln(t) < 0. (15)
8σs
The quadratic equation (15) can be solved to find the roots,
and if n̂i (t) is greater than the larger root the inequality will
never hold. Solving the quadratic equation (15), we obtain
that event (13) does not hold if
s
4γσ 2 ln(t) 2
4σs2 γ ln(t)
8σ 2
s
k
n̂i (t) ≥
+
+ 2 s f (t)γ ln(t)
2
2
∆i M
∆i M
∆i M
s
∆2 M f (t)
4σ 2 γ ln t
= s2
1 + 1 + i2
.
∆i M
2σs γ ln t
q
l
m
∆2i M f (T )
4σ 2 γ ln T
Now, we set η = M n + s∆2 (1+ 1 + 2γσ
2 ln T ) .
s
i
It follows from monotonicity of f (t) and ln(t) and statement (i) of Proposition 1 that event (13) does not hold if
M ncent
i (t) > η. Therefore, from (10) we get that
M
X
†
E[nki (T )]
≤η+
k=1
M tX
k −1
X
t=1
k=1
−
t
1
1+k
c
T
X
1
+
. (16)
tγ
†
t=tk
The argument of the first summation of the RHS yields
t†k −1
X
t
−
1
1+k
c
+
t=1
Z
t†k −1
−
t
Z
t=t†k
t=1
≤
1 + kc †
(tk )
kc
T
dt +
k
c
1+k
c
+
1
1
dt + 1 + †
tγ
tk
Qi (t) = νi (t) + σi (t)Φ−1 (1 − 1/t),
(17)
where νi (t) is the posterior mean and σi (t) the posterior
standard deviation of mean reward at time t. Φ−1 (·) is the
standard Gaussian inverse cumulative distribution function.
Let the prior on rewards from each arm be multivariate
Gaussian with mean vector ν0 ∈ RN and covariance matrix
Σ0 ∈ RN ×N . Then, the posterior mean and covariance of
mean reward at time t can be computed using the following
recursive update rule [20]:
(18)
where φ(t) and ν(t) are column vectors of φi (t) and νi (t),
respectively, and φi (t) is the indicator function of selecting
arm i at time t. The update equation (18) can be reduced to
1
1
1
1
− k + †.
†
γ−1
γ − 1 (t )
c
tk
Substituting into (16) then establishes the theorem.
Remark 1 (Asymptotic Regret for coop-UCB2): In
f (t)
limit t → +∞, ln(t)
→ 0+ , and
M
8σ 2 γ
X
s
E[nki (T )] ≤
+ o(1) ln T,
2
∆i
The UCL algorithm developed in [10] applies the approach
of Bayes-UCB [9] to correlated Gaussian bandits. The UCL
algorithm at each time computes the posterior distribution of
mean rewards at each option and then computes the (1−1/t)
upper-credible-limit for each arm, i.e., an upper bound that
holds with probability (1 − 1/t). The algorithm chooses the
arm with highest upper credible limit. For Gaussian rewards,
the (1 − 1/t) upper-credible-limit can be written as
r(t)φ(t)
+ Λ(t − 1)ν(t − 1)
σs2
φ(t)> φ(t)
Λ(t) =
+ Λ(t − 1), Σ(t) = Λ(t)−1
σs2
ν(t) = Σ(t)q(t),
T
X
1
tγ
†
1
1+k
c
A. The UCL Algorithm
q(t) =
t=tk
≤
In this section, we extend the coop-UCB2 algorithm to a
Bayesian setting and develop the coop-UCL algorithm. The
Bayesian setting allows us to model correlated bandits and
incorporate a priori knowledge about reward and correlation
structure in the Bayesian prior. We first recall the UCL algorithm proposed in [10], [9] and extend it to the cooperative
setting. We then analyze the performance of this algorithm
for an uninformative prior.
ν(t) = (Λ0 + Γ(t)−1 )−1 (Γ(t)−1 µ(t) + Λ0 ν0 )
the
k=1
and we recover the upper bound on regret for a centralized
agent as given in (1) within a constant factor.
Remark 2 (Performance of Individual Agents):
Theorem 1 provides bounds on the performance of
Λ(t) = Λ0 + Γ(t)−1 ,
Σ(t) = (Λ(t))−1 ,
(19)
where Λ0 = Σ−1
0 , Γ(t) is a diagonal matrix with entries
σs2
ni (t) , and µ(t) is the vector of µi (t), which is the empirical
mean of rewards from arm i ∈ {1, . . . , N } until time t. Note
that diagonal entries of Σ(t) are (σi (t))2 , i ∈ {1, . . . , N }.
B. The coop-UCL Algorithm
We now extend the UCL algorithm to the distributed
cooperative setting and propose the coop-UCL algorithm. In
the coop-UCL algorithm, each agent first computes an approximate posterior distribution of mean rewards conditioned
on rewards obtained by all the agents. To this end, each
agent uses the approximate frequentist estimator µ̂ki from
Section III in update equation (19).
Let the prior of agent k be a multivariate Gaussian
distribution with mean ν0k and covariance Σk0 . Let Σ̂k (t)
and ν̂ k (t) be the estimated covariance matrix and posterior
mean respectively at time t. Then, the coop-UCL algorithm
performs cooperative approximate Bayesian estimation:
ν̂ k (t) = (Λk0 + Γk (t)−1 )−1 (Γk (t)−1 µ̂k (t) + Λk0 ν0k )
Λ̂k (t) = Λk0 + Γk (t)−1 ,
Σ̂k (t) = (Λ̂k (t))−1 ,
(20)
where Γk (t) is a diagonal matrix with diagonal entries
σs2 /M n̂ki (t), i ∈ {1, . . . , N }, and Λk0 = (Σk0 )−1 .
After computing ν̂ k (t) and Σ̂k (t), the coop-UCL algorithm at time t requires each agent k to choose the option
with maximum (1 − α(t))-upper-credible-limit given by
Qki (t) = ν̂ik (t) + σ̂ik (t)Φ−1 (1 − α(t)),
(21)
where α(t) is defined such that
s
n̂ki (t) + f (t) −1
1
Φ−1 (1 − α(t)) =
,
Φ
1
−
tγ
n̂ki (t)
ν̂ik (t) is the i-th entry of ν̂ k (t), and (σ̂ik (t))2 is the i-th
diagonal entry of Σ̂k (t).
C. Regret of the coop-UCL Algorithm
We now derive bounds on the expected cumulative regret
for each agent using the coop-UCL algorithm for uninformative priors for each agent. For an uninformative prior,
Λk0 = 0, for each k ∈ {1, . . . , M }, and consequently,
ν̂ k (t) = µ̂k (t) and Σ̂k (t) = Γk (t).
Theorem 2 (Regret of the coop-UCL Algorithm): For
the Gaussian MAB problem and the coop-UCL algorithm
with uninformative priors for each agent, the number of
times a suboptimal arm i is selected by all agents until time
T satisfies
M
X
M
X
1 + k
k
c
(t† )1−γ 1 1
− +
kc
γ − 1 kc t†k
k=1
k=1
s
4γσs2 ln T
∆2i M f (t)
†
+max M n+
1+ 1 +
,tmax ,
∆2i
2σs2 γ ln T
E[nki (t)] ≤ 2
c
(t†k ) 1+kc +
where t†max = maxk∈{1,...,M } t†k .
Proof: For uninformative priors, coop-UCL is analogous to
coop-UCB2 with Cik (t) = σ̂ik (t)Φ−1 (1 − α(t)). Similar to
the proof of Theorem 1, we note that
C k∗ (t)
P((11) holds) = P z ≥ p i
Var(µ̂ki∗ )
s
n̂ki (t) + f (t) −1
1
≤P z≥
Φ
1−
. (22)
k
tγ
n̂i (t) + kc
It follows that for t ≥ t†k := f −1 (kc )
P((11) holds | t ≥ t†k ) ≤
1
.
tγ
Similarly, P((12) holds | t ≥ t†k ) ≤ t1γ .
We now examine the event (13). For a standard Gaussian
2
random variable z, P(z ≥ w) ≤ 21 e√−w /2 [19]. It follows
that for α ∈ [0, 0.5], Φ−1 (1 − α) ≤ −2 ln α.
Following the argument in the proof of Theorem 1 and using the upper bound on Φ−1 (·), we obtain that the event (13)
does not hold if
s
4σs2 γ ln t
∆2i M f (t)
k
n̂i (t) ≥
1
+
1
+
.
∆2i M
2σs2 γ ln t
q
m
l
∆2 M
4σ 2 γ ln T
1+ 1 + 2σi 2 γ fln(t)
and η =
We set η̄ = M n + s∆2
T
i
s
max{η̄, t†max }. The theorem follows by proceeding similarly
to the proof of Theorem 1.
VI. N UMERICAL I LLUSTRATIONS
In this section, we elucidate our theoretical analyses from
the previous sections with numerical examples. We first
demonstrate that the ordering of the performance of nodes
obtained through numerical simulations is identical to the
ordering predicted by the nodal explore-exploit centrality
measure: the larger the kc the lower the performance. We
then investigate the effect of the graph connectivity on the
performance of agents in random graphs.
For all simulations we consider a 10-armed bandit problem
with mean rewards drawn from a normal random distribution
with mean 75 and standard deviation 25. The sampling
standard deviation is σs = 30. These parameters were
selected to give illustrative results within the displayed time
horizon, but the relevant conclusions hold across a√wide
variation of parameters. The simulations used f (t) = ln t.
Example 1 (Regret on Fixed Graphs): Consider the set
of agents communicating according to the graph in Fig. 1(a)
and using the coop-UCB2 algorithm to handle the exploreexploit tradeoff in the distributed cooperative MAB problem.
The values of kc for nodes 1, 2, 3, and 4 are 2.31, 2.31, 0,
and 5.43, respectively. As noted in Remark 2, agent 3 should
have the lowest regret, agents 1 and 2 should have equal
and intermediate regret, and agent 4 should have the highest
regret. These predictions are validated in our simulations
shown in Fig. 1(a). The expected cumulative regret in our
simulations is computed using 5000 Monte-Carlo runs.
Fig. 1(b) demonstrates the relative performance differences
between coop-UCB, coop-UCB2, coop-UCL, and single
agent UCB with the same run parameters. Here the coopUCL algorithm is shown with an informative prior and no
correlation structure. Each agent in the coop-UCL simulation
shown here has Σ0 = 625 · IM and ν0 = 75 · 1M . The use
of priors markedly improves performance.
We now explore the effect of kc on the performance of an
agent in an Erdös-Réyni (ER) random graph. ER graphs are
a widely used class of random graphs where any two agents
are connected with a given probability ρ [21].
Expected Cumulative Regret
1200
1000
800
600
400
0
1
2
3
ǫkc
(a)
(b)
(c)
Fig. 1: (a) Simulation results comparing expected cumulative regret for different agents in the fixed network shown, using P as in (3) and κ = d dmax
.
max −1
Note that agents 1 and 2 have nearly identical expected regret. (b) Simulation results of expected cumulative regret for several different MAB algorithms
k
using the fixed network shown in Fig. 1(a). (c) Simulation results of expected cumulative regret as a function of normalized c for nodes in ER graphs at
T = 500. Also shown in red is the best linear fit.
Example 2 (Regret on Random Graphs): Consider a set
of 10 agents communicating according to an ER graph
and using the coop-UCB2 algorithm to handle the exploreexploit tradeoff in the aforementioned MAB problem. In
our simulations, we consider 100 connected ER graphs, and
for each ER graph we compute the expected cumulative
regret of agents using 1000 Monte-Carlo simulations with
ρ = ln(10)/10, P as in (3), and κ = dmax /(dmax − 1).
We show the behavior of the expected cumulative regret of
each agent as a function of the normalized kc in Fig. 1(c).
It is evident that increased kc results in a sharp decrease
in performance. Conversely, low kc is indicative of better
performance. This disparity is due to the relative scarcity of
information at nodes that are in general less “central.”
VII. F INAL R EMARKS
In this paper we used the distributed multi-agent MAB
problem to explore cooperative decision-making in networks.
We designed the coop-UCB2 and coop-UCL algorithms,
which are frequentist and Bayesian distributed algorithms,
respectively, in which agents do not need to know the
graph structure. We proved bounds on performance, showing
order-optimal performance for the group. Additionally, we
investigated the performance of individual agents in the
network as a function of the graph topology, using a proposed
measure of nodal explore-exploit centrality.
Future research directions include rigorously exploring
other communications schemes, which may offer better performance or be more suitable for modeling certain networked
systems. It will be important to consider the tradeoff between
communication frequency and performance as well as the
presence of noisy communications.
R EFERENCES
[1] V. Srivastava, P. Reverdy, and N. E. Leonard. Surveillance in an
abruptly changing world via multiarmed bandits. In IEEE CDC, pages
692–697, 2014.
[2] M. Y. Cheung, J. Leighton, and F. S. Hover. Autonomous mobile
acoustic relay positioning as a multi-armed bandit with switching
costs. In IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, pages
3368–3373.
[3] J. R. Krebs, A. Kacelnik, and P. Taylor. Test of optimal sampling by
foraging great tits. Nature, 275(5675):27–31, 1978.
[4] V. Srivastava, P. Reverdy, and N. E. Leonard. On optimal foraging
and multi-armed bandits. In Allerton Conference on Comm., Control,
and Computing, pages 494–499, 2013.
[5] A. Anandkumar, N. Michael, A. K. Tang, and A. Swami. Distributed
algorithms for learning and cognitive medium access with logarithmic regret. IEEE Journal on Selected Areas in Communications,
29(4):731–745, 2011.
[6] S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic
and nonstochastic multi-armed bandit problems. Machine Learning,
5(1):1–122, 2012.
[7] T. L. Lai and H. Robbins. Asymptotically efficient adaptive allocation
rules. Advances in Applied Mathematics, 6(1):4–22, 1985.
[8] P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time analysis of the
multiarmed bandit problem. Machine Learning, 47(2):235–256, 2002.
[9] E. Kaufmann, O. Cappé, and A. Garivier. On Bayesian upper
confidence bounds for bandit problems. In Int. Conf. on Artificial
Intelligence and Statistics, pages 592–600, 2012.
[10] P. B. Reverdy, V. Srivastava, and N. E. Leonard. Modeling human
decision making in generalized Gaussian multiarmed bandits. Proceedings of the IEEE, 102(4):544–571, 2014.
[11] V. Anantharam, P. Varaiya, and J. Walrand. Asymptotically efficient
allocation rules for the multiarmed bandit problem with multiple playspart I: I.I.D. rewards. IEEE Transactions on Automatic Control,
32(11):968–976, Nov 1987.
[12] D. Kalathil, N. Nayyar, and R. Jain. Decentralized learning for
multiplayer multiarmed bandits. IEEE Transactions on Information
Theory, 60(4):2331–2345, 2014.
[13] Y. Gai and B. Krishnamachari. Distributed stochastic online learning
policies for opportunistic spectrum access. IEEE Transactions on
Signal Processing, 62(23):6184–6193, 2014.
[14] S. Kar, H. V. Poor, and S. Cui. Bandit problems in networks:
Asymptotically efficient distributed allocation rules. In IEEE CDC
and ECC, pages 1771–1778, 2011.
[15] P. Braca, S. Marano, and V. Matta. Enforcing consensus while
monitoring the environment in wireless sensor networks. IEEE
Transactions on Signal Processing, 56(7):3375–3380, 2008.
[16] V. Srivastava and N. E. Leonard. Collective decision-making in ideal
networks: The speed-accuracy trade-off. IEEE Transactions on Control
of Network Systems, 1(1):121–132, 2014.
[17] P. Landgren, V. Srivastava, and N. E. Leonard. On distributed
cooperative decision-making in multiarmed bandits. In European
Control Conference, Aalborg, Denmark, 2016.
[18] R. Olfati-Saber and R. M. Murray. Consensus problems in networks
of agents with switching topology and time-delays. IEEE Transactions
on Automatic Control, 49(9):1520–1533, 2004.
[19] M. Abramowitz and I. A. Stegun, editors. Handbook of Mathematical
Functions: with Formulas, Graphs, and Mathematical Tables. Dover
Publications, 1964.
[20] S. M. Kay. Fundamentals of Statistical Signal Processing, Volume I :
Estimation Theory. Prentice Hall, 1993.
[21] B. Bollobás. Random Graphs. Springer, 1998.
| 3cs.SY
|
Stable Matching with Evolving Preferences∗
Varun Kanade†1 , Nikos Leonardos‡2 , and Frédéric Magniez§3
1
University of Oxford
University of Athens
3
CNRS and IRIF, Univ Paris Diderot, Sorbonne Paris-Cité, France
arXiv:1509.01988v2 [cs.GT] 28 Jun 2016
2
Abstract
We consider the problem of stable matching with dynamic preference lists. At each timestep, the preference list of some player may change by swapping random adjacent members.
The goal of a central agency (algorithm) is to maintain an approximately stable matching,
in terms of number of blocking pairs, at all time-steps. The changes in the preference lists
are not reported to the algorithm, but must instead be probed explicitly. We design an
algorithm that in expectation and with high probability maintains a matching that has at
most O((log n)2 ) blocking pairs.
1
Introduction
In the world of massive and distributed data, it is hardly reasonable to assume that data are
static. Yet, one must design algorithms that maintain a solution for a given problem that is
(approximately) consistent with the requirements, e.g., a permutation that is almost sorted.
Thus, it is important to design algorithms and data structures that are robust to changes in
their input, i.e., they produce an output with some performance guarantee (quickly).
There are a few different dynamic data models that have been considered. The area of
dynamic graph algorithms consists of maintaining some property or structure, such as connectivity, matchings, or spanning trees, even when the underlying graphs are changing [EGI99,
OR10, NS13, GP13]. Here, it is assumed that the changes to the graph may be arbitrary, but
are reported to the algorithm; and the focus is on designing data structures and algorithms that
adapt efficiently (typically in terms of computational time) to changes in the input. The area of
streaming algorithms studies the setting where the data can only be accessed as a stream and
the focus is on producing the desired output with highly space-efficient procedures (typically
poly-logarithmic in the size of the input). In the area of online algorithms, one must design
procedures that, even when data is revealed bit by bit, produce an output that is competitive
with algorithms that see the entire input at once.
Recently, Anagnostopoulos et al. [AKMU11] proposed the evolving data model to take into
account the dynamic aspects of massive data. In this model, the changes to the data are not
∗
Partially supported by the French ANR Blanc project ANR-12-BS02-005 (RDAM) and the European Commission IST STREP project Quantum Algorithms (QALGO) 600700. Most of the work was carried out when
V.K and N. L. were at LIAFA (now IRIF). During this time V.K. was supported by the Fondation Sciences
Mathmatiques de Paris (FSMP).
†
[email protected]
‡
[email protected]
§
[email protected]
1
revealed to the algorithm; instead, an algorithm has query access to the data. However, it
is assumed that the changes to the data are stochastic, not adversarial. In this setting, the
focus is not on computational complexity (which is allowed to be polynomial at each timestep), but query complexity, the number of probes made by the algorithm. Anagnostopoulos
et al. [AKMU11] studied the problem of maintaining the minimum element of a permutation
and an approximately sorted permutation, motivated by questions such as maintaining high
(page)-ranked pages. In their setup, a permutation evolves by choosing a random element and
swapping it with an adjacent element. In later work, Anagnostopoulos et al. [AKM+ 12] studied
evolving graph models and problems such as s-t connectivity and minimum spanning tree.
In this work, our aim is to bring this notion to game theory starting from the basic problem
of computing a stable matching. In other words, we introduce the notion of evolving agents,
who may not report any updates to their strategy (or preferences) without an explicit request.
In the stable matching problem, the input consists of two sets A, B of equal size, and for each
member a total order (preference) over members in the other set. Given a matching between A
and B, a pair (x, y) with x ∈ A and y ∈ B is blocking if they prefer each other to their matches.
A matching is stable if there are no blocking pairs. Gale and Shapley showed that a stable
matching always exists and can be found by an efficient algorithm [GS62]. We consider the
setting where the preference lists evolve over time. The preference lists can evolve over time,
by swapping adjacent elements. More precisely, while the algorithm can perform one query per
time-step, we assume that a total number of α swaps events also occur, where α = Θ(1) is
some fixed parameter, called the evolution rate. This assumption is similar to previous works
and models the critical regime: with less evolution events the input is basically static, and with
more the input evolves too fast. The goal is to maintain a matching that has few blocking pairs.
We summarize our results as follows. All three statements hold in expectation and with
high probability.
(i) Using the results of Anagnostopoulos et al. [AKMU11] for sorting permutations, we
design an algorithm that maintains a matching with at most O(n log n) blocking pairs,
at all time-steps after roughly the first n2 log n steps (Theorem 3.1). Also, we observe
that any analysis that uses their method as a black-box, cannot improve on this bound
(Remark 3.1).1
(ii) In a restricted setting, where only one side, say the B side, has evolving preference lists,
and if the A side has uniform random permutations as preference lists (known to the central
agency), we design an algorithm that maintains a matching with O(log n) blocking pairs
at all time-steps after roughly the first n log n steps (Theorem 3.2).
(iii) Finally, we design an algorithm in the general setting, that maintains a matching with
at most O((log n)2 ) blocking pairs at all time-steps after roughly the first n2 log n steps
(Theorem 4.1).
2
Preliminaries
In the rest of the paper, n ≥ 1 denotes an integer parameter and [n] the set of integers
{1, 2, . . . , n}. For a non-negative random variable X, parametrized by some integer n, we
write “X = O(f (n)) in expectation and with high probability” when for any constant c there
exist constants n0 , c′ , c′′ > 0 such that E[X] ≤ c′ f (n) and P(X > c′′ f (n)) ≤ n−c , for every
integer n ≥ n0 .
1
We don’t rule out the possibility that a more fine-grained analysis of the algorithm may give better bounds;
instead we design new algorithms.
2
2.1
Stable Matching
We only consider the bipartite stable matching problem, also known as stable marriage. There
are two sets of players A and B, with |A| = |B| = n. Each player x ∈ A (y ∈ B) holds a
permutation of B (A), denoted πx : B → [n] (πy : A → [n]) indicating their preferences over
players in the set B (A). Thus, for y ∈ B, πx (y) denotes the rank of y in x’s preference list
(where 1 is the highest rank).
Let M : A → B be a matching (a bijection). A pair (x, y) is said to be blocking if y ≺πx M (x)
and x ≺πy M −1 (y), where z ≺π z ′ indicates that z is ranked higher than z ′ according to
permutation π (i.e., π(z) < π(z ′ )). In words, x prefers y to M (x) and y prefers x to M −1 (y).
A matching M is stable if there are no blocking pairs. Then the stable matching problem
is to find a stable matching given preference lists {πz : z ∈ A ∪ B)}. Gale and Shapley [GS62]
proved that a stable matching always exists, and gave an algorithm that given the preferences
lists as input finds a stable matching in O(n2 ) time.
The Gale-Shapley algorithm is simple to describe. Only players in the set A make proposals.
Initially all players are unmatched. Let M denote a partial matching at some point. If there
is an unmatched player x ∈ A, x makes a proposal to y ∈ B, where y is the highest-ranked
player in πx to whom x has not yet proposed. If y is unmatched, or prefers x to M −1 (y), then
y accepts the proposal and we set M (x) = y. In the latter case, the agent previously matched
to y, i.e., M −1 (y) before M was updated, becomes unmatched once more. Gale and Shapley
showed that this algorithm always results in a stable matching.
Wilson [Wil72] studied the problem where all the preference lists are independent and uniformly random permutations; in this case, he showed that the number of proposals made by
the Gale-Shapley algorithm is O(n log n) in expectation and with high probability (see also
[MR95]). In fact, only the proposing side needs to be random in their statement. We provide
a proof sketch for completeness.
Theorem 2.1 ([Wil72]). If the permutations {πx : x ∈ A} are chosen randomly, the number of
proposals made in the Gale-Shapley algorithm (where only A makes proposals) is O(n log n) in
expectation and with high probability.
Proof sketch. Following the proof in [MR95] (see also [Knu97]), analyze an alternative procedure
where every proposal is uniform over the whole of B. If it happens that x ∈ A proposes to a
y ∈ B that has already rejected x, then a rejection is guaranteed. It is not hard to show that the
number of proposals such an algorithm makes stochastically dominates the number of proposals
of the classical algorithm. Next, by the method of deferred decisions, fix the randomness in
the algorithm only when needed. Then observe that the number of proposals is equal to the
number of coupons chosen in the coupon-collector’s problem.
2.2
Model for evolving input
A general framework for studying dynamic data was introduced in [AKMU11]. Here we are
only concerned with evolving preference lists (or permutations). In our model, we consider
discrete time-steps. In each time-step, the algorithm is competing against nature as follows:
the algorithm can query the input locally, nature lets the input evolve according to one or more
evolution events.
A query to the stable matching input is a triplet (z, u, v) ∈ (A×B 2 )∪(B×A2 ) and the answer
is whether πz (u) < πz (v). One evolution event consists of the following: pick z ∈ A ∪ B and
i ∈ [n − 1] uniformly at random and swap u = πz−1 (i) and v = πz−1 (i + 1) (i.e., set πz (u) = i + 1
and πz (v) = i).
3
While the algorithm can perform one query per time-step, α evolution events also occur,
where α ≥ 1 is some integer called the evolution rate. We further assume that α = Θ(1), meaning
that evolution events occur basically as often than the algorithm probes. We emphasize that
the rate-limiting factor is the queries made by the algorithm. In particular, the algorithm may
perform arbitrary (polynomial-time) computations in between time-steps. We are now ready to
define our problem:
Evolving Stable Matching (ESM): Given query access to an instance of the
stable matching problem with evolution rate α = Θ(1), maintain a matching that
minimizes the number of blocking pairs.
2.3
Sorting evolving permutations
The problem of sorting a single evolving permutation has been already addressed in [AKMU11].
In this context, the evolution rate is still constant, but denotes the evolution rate of this single
permutation. We will use the algorithm Quicksort of [AKMU11]. It is simply the randomized
version of quicksort which is shown to be robust with respect to an evolving input. The first
lemma shows that the running time of quicksort is not affected by evolution events.2
Lemma 2.1 (Proposition 3 in [AKMU11]). The running time of Quicksort is O(n log n) in
expectation and with high probability, for any rate of evolution when the pairs to be swapped are
chosen randomly.
Second, Lemma 6 in [AKMU11] states that Quicksort when run on an evolving permutation π, computes a permutation π̃ in which every element is approximately sorted. At time-step
t, let π t the denote the current permutation, and π̃ t its approximation computed by the algorithm.
Lemma 2.2 (Lemma 6 in [AKMU11]). Let t be the time-step of completion of Quicksort, then
given an element u, the number of pairs (u, v) that the permutations π t and π̃ t rank differently
is O(log n) in expectation and with high probability.
In our setting there are 2n evolving permutations over some set of n elements. Algorithm 1
simply sorts m (out of 2n) permutations, denoted by π1 , . . . , πm using Quicksort one after
another. (We always invoke Algorithm 1 with either n or 2n permutations.)
Algorithm 1 : Sequential sorting
procedure SequentialSort({πj : j = 1, . . . , m})
for j = 1 to m do
3:
π̃j ← Quicksort(πj )
1:
⊲ Only have query access to input
2:
4:
return {π̃j : j = 1, . . . , m}
Using Lemma 2.2 (Lemma 6 of [AKMU11]) we can argue that Algorithm 1 maintains all
permutations approximately sorted. While the evolving rate is still α = Θ(1), there are now 2n
evolving permutations, and the total number of evolution events is α per time-step.
Lemma 2.3. Let t be the time-step when Algorithm 1 terminates. Then, for m ≤ 2n, given
any element u and j ∈ [m], the number of pairs (u, v) that the permutations πjt and π̃jt rank
differently is O(log n) in expectation and with high probability.
2
We remark that Anagnostopoulos et al. [AKMU11] use ‘whp’ to denote events that hold with probability
1 − o(1), rather than the stronger notion we use in this paper. However, their proofs for the results used in our
paper actually prove the stronger bounds. They have other results that do not satisfy the stronger notion and
these are not used in our work.
4
Proof. Fix some j ∈ [n]. Suppose that π̃jt was computed at time-step t′ ≤ t (the time-step when
Quicksort for this particular list terminates). By Lemma 2.2 the statement holds for u at
time-step t′ . Due to Lemma 2.1 we have t − t′ = O(n2 log n) with high probability. During this
time, the number of evolution steps that have swapped u with an adjacent element is O(log n)
with high probability. This follows from a balls-and-bins experiment where we throw O(n2 log n)
balls (corresponding to the evolution steps) into m(n − 1) bins (corresponding to the adjacent
pairs). It is known (see Exercise 3.1 in [MR95]) that in this particular case the number of balls
in every bin is of the order of its mean with high probability. Therefore, during this time, at
most O(log n) more elements may be swapped with u.
2.4
Chernoff Bound with dependent variables
We will require the following extension of the Chernoff bound. It follows from (the more general)
Theorem 3.8 in [McD98].
Theorem 2.2. For i ∈ [n], letQYi be a random variable over some set Yi and Xi be a Boolean
random variable. For any y ∈ ni=1 Yi and k ∈ [n], let Ek (y) denote the event Y1 = y1 , . . . , Yk =
yk . Suppose P[Xk = 1|Ek−1 (y)] ≤ p, for all y and k as above. Then, for any t ≥ 0,
i
hX
Xk ≥ pn + t ≤ exp −
P
3
3t2
.
6pn + 2t
Two simple cases
In this section we present two simple arguments in two different settings. First, we consider
how the original Gale-Shapley algorithm performs when run, without any modification, on lists
produced by running quicksort on the evolving input. We present a simple analysis showing a
bound of O(n log n) on the number of blocking pairs, which is better than the trivial bound of
order n2 . Next, we analyze the Gale-Shapley algorithm when evolution events only occur on one
side and the preference lists are uniformly random permutations; in particular the preference
lists on the A side are chosen uniformly at random, and the preference lists on the B-side are
subject to evolution events.3 We present a simple analysis showing an O(log n) bound on the
number of blocking pairs for this special case.
3.1
A simple algorithm
Our first algorithm ignores evolution of preference lists and runs the standard Gale-Shapley
algorithm to produce a matching. More specifically, it first obtains the preferences lists for all
2n agents using the Quicksort algorithm of [AKMU11] (i.e., using Algorithm 1) and then
produces a matching by running the Gale-Shapley algorithm on these lists (ignoring the fact
that these lists are only approximately correct).
We show that this simple algorithm maintains a matching with at most O(n log n) blocking
pairs. Note that the number of blocking pairs is trivially at most n2 . We further argue that
improving the bound would require new ideas that either go around Lemma 2.3 (Lemma 6 of
[AKMU11]) or improve the analysis in a substantial way.
Algorithm 2 runs in perpetuity. The matching M is maintained as the output until the new
matching based on the newly sorted preference lists can be computed.
3
Note that after sufficiently many time-steps (though still polynomial) the evolution events ensure that all
permutations are uniformly random. This follows from analyzing the mixing time of the corresponding Markov
chain over permutations. See for example the book [LPW09].
5
Algorithm 2 : Simple dynamic stable matching
1: while True do
2:
{π̃z : z ∈ A ∪ B} ← SequentialSort({πz : z ∈ A ∪ B})
⊲ Calling Algorithm 1
3:
return Gale-Shapley matching M on the (approximately) sorted lists {π̃z : z ∈ A ∪ B}
Theorem 3.1. For a sufficiently large constant c0 and any time-step t ≥ c0 n2 log n, Algorithm 2
maintains a matching with O(n log n) blocking pairs in expectation and with high probability.
Proof. We consider the number of blocking pairs at time-step T ≤ t when the current matching
M was computed. In the following discussion, we use the following notation: for x ∈ A, y ∈ B,
if M (x) = y, then M (y) = x (rather than M −1 ). At time-step T , for each z ∈ A ∪ B, define
the indicator function Iz (w) to be 1 when M (z) ≺π̃z w and w ≺πz M (z) and 0 otherwise. (We
don’t explicitly use superscripts on the preference lists π as time-step T is fixed until specified
otherwise.) By Lemma 2.3, in expectation and with high probability,
X
Iz (w) = {w : w ≺πz M (z) and M (z) ≺π̃z w} = O(log n),
(1)
w
If a pair (x, y) is blocking at time-step T , then x ≺πy M (y) and y ≺πx M (x). Assume x ≺π̃y
M (y) and y ≺π̃x M (x). Since y ≺π̃x M (x), x must have proposed to y at some point during the
execution of the Gale-Shapley algorithm. By the properties of the Gale-Shapley algorithm, y
should have been matched to an element of A with rank according to π̃y at least as high as the
rank of x in π̃y . This contradicts x ≺π̃y M (y). It follows that either M (y) ≺π̃y x or M (x) ≺π̃x y.
Define U (x, y) to be 1 when (x, y) is blocking and 0 otherwise. We have argued that
U (x, y) ≤ Ix (y) + Iy (x).
By the union bound, Equation 1 holds for every z ∈ A ∪ B with high probability. Summing
over all pairs (x, y) and applying the union bound again
XX
X
X
XX
U (x, y) ≤
Ix (y) + Iy (x) =
Ix (y) +
Iy (x) = O(n log n),
x,y
x,y
x
y
y
x
in expectation and with high probability.
Next, let t′ = t − T . We need to account for blocking pairs that may have arisen during
these t′ time-steps. First, we observe that by Lemma 2.1 and union bound, with high probability
t′ = O(n2 log n) (as otherwise another matching more up-to-date than the one at time-step T
would be available). During the t′ time-steps from T to t, evolution may create a blocking
pair only if the swap decreases the rank of M (z) in πz , for some z ∈ A ∪ B. Therefore, each
step of the evolution introduces—independently—a new blocking pair with probability at most
α/(n − 1). The expected number of blocking pairs introduced is therefore at most O(n log n)
for α = Θ(1) (and assuming t′ = O(n2 log n)). The result follows by a simple application of the
Chernoff and union bounds.
Remark 3.1. We present a pair of instances to the stable matching problem with preference
lists πz and π̃z for z ∈ A ∪ B, for which the conclusion of Lemma 2.3 is satisfied, i.e., for any z
the number of pairs (i, j) that are ordered differently in πz and π̃z are O(log n). We then show
that a matching that is stable with respect to π̃ as preference lists has Ω(n log n) blocking pairs
with respect to π. Thus, it follows that using the Quicksort algorithm of Anagnostopoulos et
6
1
2
3
4
5
6
7
A-side lists π
3 2 1 4 5 6
4 3 2 5 6 7
5 4 3 6 7 1
6 5 4 7 1 2
7 6 5 1 2 3
1 7 6 2 3 4
2 1 7 3 4 5
7
1
2
3
4
5
6
1
2
3
4
5
6
7
B-side lists
6 7 1 5 4
7 1 2 6 5
1 2 3 7 6
2 3 4 1 7
3 4 5 2 1
4 5 6 3 2
5 6 7 4 3
π
3
4
5
6
7
1
2
2
3
4
5
6
7
1
1
2
3
4
5
6
7
A-side lists π̃
1 2 3 4 5 6
2 3 4 5 6 7
3 4 5 6 7 1
4 5 6 7 1 2
5 6 7 1 2 3
6 7 1 2 3 4
7 1 2 3 4 5
7
1
2
3
4
5
6
1
2
3
4
5
6
7
B-side lists
1 7 6 5 4
2 1 7 6 5
3 2 1 7 6
4 3 2 1 7
5 4 3 2 1
6 5 4 3 2
7 6 5 4 3
π̃
3
4
5
6
7
1
2
2
3
4
5
6
7
1
Figure 1: Instances π and π̃ demonstrating tightness of Algorithm 2.
al. [AKMU11] and its analysis as a blackbox will not result in a stronger result than the one
provided in Theorem 3.1.
First, we define the preference lists π̃z for z ∈ A ∪ B. Let A = B = [n]. Then for x ∈ A,
the preference list (ranking) π̃x is defined as x, x + 1, . . . , n, 1, 2, . . . , x − 1. On the other hand
for y ∈ B, the preference list (ranking) π̃y is defined as y, y − 1, . . . , 1, n, n − 1, . . . , y + 1. The
rankings πz , z ∈ A ∪ B, are now defined as follows: let k be some parameter, πz simply as the
elements at rank 1 and k swapped. Figure 1 shows an example with n = 7 and k = 3. Clearly,
when k = Θ(log n), πz and π̃z satisfy the conclusion of Lemma 2.3. Yet, it is easy to see that
M (i) = i is a stable matching for the preference lists π̃z , z ∈ A ∪ B, and for this matching with
respect to the preference lists πz , z ∈ A ∪ B, every pair (i, j) with 0 < j − i < k is a blocking
pair.
3.2
One-sided evolution
In this section we analyze how the Gale-Shapley algorithm performs when the initial preference
lists are random, but there is no evolution on the lists of the elements in A. Furthermore, we
are going to assume that the algorithm knows each permutation in {πx : x ∈ A}. We call this
setting one-sided evolution.
Algorithm 3 Gale-Shapley Algorithm
1: M ← φ
2: for x ∈ A do
3:
new match ← False
4:
p←x
5:
while new match = False do
6:
y ← first as yet unproposed as per πp
7:
if M (y) not yet set then
8:
M (p) ← y
9:
M (y) ← p
10:
new match = True
11:
else if p ≺πy M (y) then
12:
p′ ← M (y)
13:
M (y) ← p
14:
M (p) ← y
15:
p ← p′
16: return M
⊲ Initialize empty matching
⊲ y prefers p to M (y)
In this setting, the standard Gale-Shapley algorithm is implemented (basic pseudocode is
7
shown in Algorithm 3). Note that the only time the preference lists on the B-side are used is in
Line 11. Thus, it is only for these steps that we need to query the input (since the preference lists
on the A-side are known to the algorithm). Thus, the number of queries made by the algorithm
is bounded by the number of proposals. It was already observed that the number of proposals
made in a random instance of the stable matching problem is O(n log n). The actual algorithm
keeps implementing the Gale-Shapley algorithm from scratch after completion. The matching
from the previous completed run is used as the current matching. We prove the following result
for the one-sided evolution setting.
Theorem 3.2. For a sufficiently large constant c0 and any time-step t ≥ c0 n log n, the GaleShapley algorithm (repeatedly run and using the matching of the last completed run as the
output) under one-sided evolution maintains a matching with at most O(log n) blocking pairs in
expectation and with high probability.
Proof. To prove the bound on the number of blocking pairs, call an evolution event on y’s list
critical if it involves the then match of y. Suppose that after the algorithm terminates, y ∈ B is
involved in k blocking pairs (x1 , y), . . . , (xk , y). We observe that each one of the x1 , . . . , xk was
involved in at least one critical evolution step. To see this note that if (x, y) is blocking, then x
proposed to y during the execution of the algorithm and got rejected subsequently (because πx
didn’t change and y ≺πx M (x)). But since at the end of the execution it forms a blocking pair,
it must ranked higher than M (y). This is only possible if x was swapped with the then match
of y in some evolution event during the execution of the algorithm.
Given this observation, we estimate the number of blocking pairs by estimating the number
of critical evolution steps. Note that an evolution step is critical with probability at most
2α/(n − 1) (at most 2n out of n(n − 1) pairs involve the matching). Let T be a random
variable equal to the number of proposals before the algorithm outputs a matching and label
the corresponding time-steps as 1, 2, . . . , T . For each step k, let Xk be a Boolean random
variable that is equal to 1 if at the time-step labeled k some evolution event was critical.
First note that from a coupon-collecting argument as in Theorem 2.1 it follows that T =
O(n log n) in expectation and with high probability. This is because for that argument the
distribution of {πy : y ∈ B} is irrelevant and {πx : x ∈ A} being random permutations suffices.
Therefore, we may fix an appropriately large constant C so that T ≤ Cn log n with high
probability and let m = Cn log n. As noted above, at any given step and given any information
from the previous steps, an evolution step is critical with probability at most 2α/(n − 1); thus,
by Theorem 2.2,
m
hX
i
P
Xk > 2C log n = O(n−C ),
k=1
for sufficiently large C. We have
T
T
hX
i
hX
i
P
Xk > 2C log n ≤ P T > Cn log n + P
Xk > 2C log n ∧ T ≤ Cn log n
k=1
k=1
m
hX
i
≤ P T > Cn log n + P
Xk > 2C log n .
k=1
By the observation at the beginning of this paragraph the claimed bound holds with high
probability. The bound on the expectation follows by noting that there can be at most n2
blocking pairs.
8
Finally, note that there will be at most O(n log n) time-steps before a new matching is
computed. As in the final paragraph in the proof of Theorem 3.1, one can show that evolution
cannot produce more than O(log n) blocking pairs in these many steps.
4
General Case: Improved algorithm
We now consider the general setting where the preference lists on both sides may be evolving.
We present a modified version of the Gale-Shapley algorithm that takes advantage of Lemma 2.3
(Lemma 6 of [AKMU11]) and maintains a stable matching with at most O((log n)2 ) blocking
pairs. The algorithm consists of two separate processes that run in an interleaved fashion:
the sorting process on even time-steps and the matching process on odd ones. The sorting
process is basically a call to SequentialSort({πx | x ∈ A}) that produces approximately
sorted preference lists on the A side, {π̃x | x ∈ A}. The algorithm runs in perpetuity, in the
sense that as soon as it terminates it restarts, though the copies {π̃x | x ∈ A} from the previous
execution are retained to be used by the stable matching process. Initially, the π̃x are set to be
random permutations, thus, for the first O(n2 log n) steps, until one run of the sorting process
is complete, the matching output by the algorithm will be garbage.
Algorithm 4 : Interleaving Sorting and Matching
1: for t = 1, 2, . . . do
2:
if t is Even then
3:
Perform query for Algorithm 1
4:
else if t is Odd then
5:
Perform query for Algorithm 5
We note that what is counted here is only time-steps; making queries is the bottleneck.
Additional computations required by the algorithm can be performed in between time-steps at
no additional cost.
The sorting process performs queries only during even steps and its purpose it to keep the
preference list of each x ∈ A approximately sorted, where by approximately sorted we meant
that the conclusion of Lemma 2.3 holds.
The matching process performs queries during odd steps. Our stable matching algorithm,
which is a variant of the Gale-Shapley algorithm, is presented in Algorithm 5. Note that the
{π̃x | x ∈ A} used in Algorithm 5 are static and are the output of the latest completed run of
Algorithm 1. However, the comparisons made are all on dynamic data. The difference from the
standard Gale-Shapley algorithm is that whenever some x ∈ A is about to make a proposal,
first the best y ∈ B among the O(log n) highest ranked as per the ranking π̃x that have not
yet rejected x is found. Note however, that the best is with respect to the dynamic (current)
preference list πx (otherwise, it would be trivial since π̃x is static). This operation is basically
the algorithm to find the minimum element, which can be implemented in O(log n) time using
only comparison queries (see Section 3 in [AKMU11]). We don’t need to use any particular
result regarding finding the minimum element; instead, we incorporate the errors that may
have occurred while finding the minimum due to the dynamic nature of the data, into our
stable matching analysis directly.
We first describe the high-level idea of the proof. The sorting process needs O(n2 log n)
comparisons with high probability. The approximations {π̃x : x ∈ A} of {πx : x ∈ A} that are
being computed are used by the modified Gale-Shapley algorithm. By Lemma 2.3 we are able
to claim that for any element u in the preference list of any x ∈ A, the number of pairs (u, v)
that are ordered differently in π̃x and πx are O(log n). Therefore, when x is about to propose it
9
Algorithm 5 : Modified Gale-Shapley Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
M ←∅
for x ∈ A do
new match ← False
p←x
while new match = False do
S ← C log n highest-ranked, not-yet-proposed elements in B per π̃x
y ← best(S)
⊲ Best with respect to dynamic πx
if M (y) not yet set then
M (p) ← y
M (y) ← p
new match ← True
else if p ≺πy M (y) then
p′ ← M (y)
M (y) ← p
M (p) ← y
p ← p′
return M
suffices to look among O(log n) elements in π̃x to find the y to which the proposal will be made.
Since the matching process is expected to make O(n log n) proposals, it is expected to require
O(n(log n)2 ) comparisons. It turns out that, during these steps, evolution creates a blocking
pair with probability at most α/n. Therefore we expect O((log n)2 ) blocking pairs.
We now provide the details of the proof. In order to bound the number of blocking pairs,
it is crucial that during the matching process not too many queries are made, or alternatively
that not too many proposals are made. We therefore need an analog of Theorem 2.1. To apply
the coupon-collecting argument from the proof of that theorem we prove the following lemma.
Lemma 4.1. Provided πx was chosen uniformly at random at time-step 0 and only comparison
queries are made, the element y chosen at line 7 of Algorithm 5 is a random element from the
subset of B to which x has not by that point made any proposals.
Proof. The proof of the lemma relies on the fact that the dynamic quicksort algorithm used to
obtain π̃x for x ∈ A and the procedure used to find the best element in line 7 of Algorithm 5
only use comparison queries.
Let πx be the preference list of x before the first comparison is queried. Fix an arbitrary
sequence of evolution steps that occurs during the computation of y. Suppose that given these
choices of nature, y = πx (k). Then, given the same evolution steps, for any other permutation
πx′ , y = πx′ (k). Since πx is a random permutation, y = πx (k) is a random element of σx .
Remark 4.1. We remark that the requirement on the implementation of Quicksort and
best using only comparison queries is necessary and the lemma does not hold for an arbitrary
algorithm.
Lemma 4.2. The number of proposals during one execution of the matching process is O(n log n)
in expectation and with high probability.
Proof. Suppose x proposes to y and at that point k elements of B are unmatched. Note that
it must be the case that x has not proposed to any of these elements (otherwise, they would
not be currently unmatched). Thus, by the previous lemma, each of these k elements receives
10
a proposal with probability at least 1/n. The stated bound follows from the analysis of coupon
collector’s problem as in Theorem 2.1.
As in the one-sided setting, the analysis will rely on estimating the occurrence of a specific
kind of critical evolution steps. In the present case the definition of a critical evolution step is
a little more involved than its one-sided counterpart.
Definition 4.1. An evolution event on the preference list πz of z ∈ A ∪ B is critical if one of
the following holds:
1. An evolution event involves a swap of the then match of z, M (z).
2. If z ∈ A, the evolution event involves swapping the then best element as per πz to which
z has not yet proposed.
The following claim establishes the link between the critical evolution steps and the number
of blocking pairs.
Claim 4.1. Assume that for the duration of one run of the matching process, the preference
lists {πz | z ∈ A ∪ B} satisfy the conditions of Lemma 2.3, and suppose that (x, y) is a blocking
pair with respect to the returned matching. Then there was a critical evolution event on the
preference list of at least one of x and y during the execution of the matching process.
Proof. First consider the case that x proposed to y during the execution of the matching process.
It follows that y rejected x at some point in favor of some other element. When x was rejected,
the then M (y) satisfies M (y) ≺πy x. Subsequently, M (y) may change but has to become better,
unless there was a swap that involves the then M (y), which is a critical event on πy . Since,
we know that in the end x ≺πy x′ , where x′ is the final match of y, there must have been
some evolution event where x was swapped with the then match of y. Thus, by part 1 of
Definition 4.1, a critical evolution event occurred on πy .
On the other hand, suppose x never proposed to y, and let y ′ be the final match of x.
Suppose that when x proposed to y ′ , y ≺πx y ′ ; it follows that best on line 7 of Algorithm 5
failed to return the best element to which x had not yet proposed. Since, we are assuming that
π̃x is a sufficient approximation of πx , it must be because the actual best element was swapped at
least once while best was being executed. Thus, by part 2 of Definition 4.1, a critical evolution
event occurred on πx . Finally, if when x proposed to y ′ , it was the case that y ′ ≺πx y, but (x, y)
is a blocking pair, it must be that y ′ was involved in a swap subsequently leading to a critical
event involving x’s the then match.
Theorem 4.1. Provided the initial preference lists are drawn randomly,4 for all z ∈ A ∪ B,
for a sufficiently large constant c0 and any time-step t ≥ c0 n2 log n, Algorithm 5 maintains a
matching with at most O((log n)2 ) blocking pairs in expectation and with high probability.
Proof. As a result of Claim 4.1, we can estimate the number of blocking pairs by estimating
the number of critical evolution steps. Let T be a random variable equal to the number of
queries of Algorithm 5 before it outputs a matching and label the corresponding time-steps as
1, 2, . . . , T . For each step k, let Xk be a Boolean random variable that is equal to 1 if at the
time-step labeled k the evolution was critical. Furthermore, denote by E the event that during
these T time-steps the lists were approximately sorted. By Lemma 2.3, the event E occurs with
high probability.
4
This is not actually required, since after sufficiently long (though still polynomial) time, all the preference
lists will be close to random due to a mixing time argument on the set of permutations.
11
By Lemma 4.2 it follows that T = O(n(log n)2 ) in expectation and with high probability,
since we are wasting O(log n) queries per proposal. Therefore, we may fix an appropriately
large constant C so that T ≤ Cn(log n)2 with high probability and let m = Cn(log n)2 . Note
that—given any history of evolution steps—an evolution step is critical with probability at most
O(α/n), since for each z ∈ A ∪ B there is a constant number of elements that evolution has to
swap in order to be critical. Thus, by Theorem 2.2,
m
hX
i
P
Xk > 2C(log n)2 = O(n−C ),
k=1
for sufficiently large C. We have
T
hX
i
¯
P
Xk > 2C(log n)2 ≤ P T > Cn(log n)2 + P[E]
k=1
T
i
hX
+P
Xk > 2C(log n)2 ∧ T ≤ Cn(log n)2 ∧ E
k=1
2
≤ P T > Cn(log n)
¯ +P
+ P[E]
m
hX
i
Xk > 2C(log n)2 .
k=1
It follows that the claimed bound holds with high probability. The bound on the expectation
follows by noting that there can be at most n2 blocking pairs.
Acknowledgments
For earlier discussions, F.M. would like to thank Marcos Kiwi who, among other things, introduced him to the line of works on evolving data and shared preliminary thoughts on the
possibility of computing stable matchings in this context. We also thank a reviewer for (minor)
corrections in our main theorem.
References
[AKM+ 12] A. Anagnostopoulos, R. Kumar, M. Mahdian, E. Upfal, and F. Vandin. Algorithms
on evolving graphs. In Proc. of 3rd Innovations in Theoretical Computer Science,
2012.
[AKMU11] A. Anagnostopoulos, R. Kumar, M. Mahdian, and E. Upfal. Sorting and selection on dynamic data. Theoretical Computer Science, 412(24):2564 – 2576, 2011.
Selected Papers from 36th International Colloquium on Automata, Languages and
Programming.
[EGI99]
D. Eppstein, Z. Galil, and G. F. Italiano. Dynamic graph algorithms. In M. Atallah,
editor, Algorithms and Theory of Computation Handbook, chapter 8. CRC Press,
1999.
[GP13]
M. Gupta and R. Peng. Fully dynamic (1+ e)-approximate matchings. In Proc. of
54th IEEE Foundations of Computer Science, pages 548–557, Oct 2013.
[GS62]
D. Gale and L. S. Shapley. College admissions and the stability of marriage. The
American Mathematical Monthly, 69(1):9–15, 1962.
12
[Knu97]
D. Knuth. Stable Marriage and Its Relation to Other Combinatorial Problems:
An Introduction to the Mathematical Analysis of Algorithms. CRM proceedings &
lecture notes. American Mathematical Society, 1997.
[LPW09]
David Asher Levin, Yuval Peres, and Elizabeth Lee Wilmer. Markov Chains and
Mixing Times. American Mathematical Society, 2009.
[McD98]
C. McDiarmid. Concentration. In M. Habib, C. McDiarmid, J. Ramirez-Alfonsin,
and B. Reed, editors, Probabilistic Methods for Algorithmic Discrete Mathematics,
volume 16 of Algorithms and Combinatorics, pages 195–248. Springer Berlin Heidelberg, 1998.
[MR95]
R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge International
Series on Parallel Computation. Cambridge University Press, 1995.
[NS13]
O. Neiman and S. Solomon. Simple deterministic algorithms for fully dynamic
maximal matching. In Proc. of 45th ACM Symposium on Theory of Computing,
pages 745–754, 2013.
[OR10]
K. Onak and R. Rubinfeld. Maintaining a large matching and a small vertex cover.
In Proc. of 42nd ACM Symposium on Theory of Computing, pages 457–464, 2010.
[Wil72]
L. B. Wilson. An analysis of the stable marriage assignment algorithm. BIT Numerical Mathematics, 12(4):569–575, 1972.
13
| 8cs.DS
|
Modular Construction of Shape-Numeric Analyzers
Bor-Yuh Evan Chang
Xavier Rival
University of Colorado Boulder
INRIA, ENS, and CNRS
[email protected]
[email protected]
The aim of static analysis is to infer invariants about programs that are precise enough to establish
semantic properties, such as the absence of run-time errors. Broadly speaking, there are two major
branches of static analysis for imperative programs. Pointer and shape analyses focus on inferring
properties of pointers, dynamically-allocated memory, and recursive data structures, while numeric
analyses seek to derive invariants on numeric values. Although simultaneous inference of shapenumeric invariants is often needed, this case is especially challenging and is not particularly well
explored. Notably, simultaneous shape-numeric inference raises complex issues in the design of the
static analyzer itself.
In this paper, we study the construction of such shape-numeric, static analyzers. We set up an
abstract interpretation framework that allows us to reason about simultaneous shape-numeric properties by combining shape and numeric abstractions into a modular, expressive abstract domain. Such
a modular structure is highly desirable to make its formalization and implementation easier to do
and get correct. To achieve this, we choose a concrete semantics that can be abstracted step-by-step,
while preserving a high level of expressiveness. The structure of abstract operations (i.e., transfer,
join, and comparison) follows the structure of this semantics. The advantage of this construction is
to divide the analyzer in modules and functors that implement abstractions of distinct features.
1
Introduction
The static analysis of programs written in real-world imperative languages like C or Java are challenging
because of the mix of programming features that the analyzer must handle effectively. On one hand, there
are pointer values (i.e., memory addresses) that can be used to create dynamically-allocated recursive
data structures. On the other hand, there are numeric data values (e.g., integer and floating-point values)
that are integral to the behavior of the program. While it is desirable to use distinct abstract domains to
handle such different families of properties, precise analyses require these abstract domains to exchange
information because the pointer and numeric values are often interdependent. Setting up the structure of
the implementation of such a shape-numeric analyzer can be quite difficult. While maintaining separate
modules with clearly defined interfaces is a cornerstone of software engineering, such boundaries also
impede the easy exchange of semantic information.
In this manuscript, we contribute a modular construction of an abstract domain [10] that layers a
numeric abstraction on a shape abstraction of memory. The construction that we present is parametric
in the numeric abstraction, as well as the shape abstraction. For example, the numeric abstraction may
be instantiated with an abstract domain such such as polyhedra [12] or octagons [27], while the shape
abstraction may be instantiated with domains such as Xisa [5,7] or TVLA [31]. Note that the focus of this
paper is on describing the formalization and construction of the abstract domain. Empirical evaluation
of implementations based on this construction are given elsewhere [5, 7, 8, 22, 29, 36, 37].
We describe our construction in four steps:
1. We define a concrete program semantics for a generic imperative programming language focusing
on the concrete model of mutable memory (Section 2).
A. Banerjee, O. Danvy, K.-G. Doh, J. Hatcliff (Eds):
David A. Schmidt’s 60th Birthday Festschrift
EPTCS 129, 2013, pp. 161–185, doi:10.4204/EPTCS.129.11
c B.-Y. E. Chang and X. Rival
This work is licensed under the
Creative Commons Attribution License.
Modular Construction of Shape-Numeric Analyzers
162
typedef struct s {
struct s ⋆ a; int b; int c;
} t;
void f() {
t y;
t ⋆ x = &y;
y · a = malloc(sizeof(t));
y · b = 24; y · c = 178;
y · a -> a = NULL;
y · a -> b = 70;
y · a -> c = 89;
}
(a)
&x = 0x...a0 0x...b0
&y = &(y · a) = 0x...b0 0x...c0
&(y · b) = 0x...b4
24
178
&(y · c) = 0x...b8
&(y · a -> a) = 0x...c0
&(y · a -> b) = 0x...c4
&(y · a -> c) = 0x...c8
(b)
0x0
70
89
E:
σ:
X −→
x 7−→
y 7−→
A
0x...a0
0x...b0
−→
7−→
7−→
7−→
7−→
7−→
7−→
7−→
V
0x...b0
0x...c0
24
178
0x0
70
89
A
0x...a0
0x...b0
0x...b4
0x...b8
0x...c0
0x...c4
0x...c8
(c)
Figure 1: A concrete memory state consists of an environment E and a store σ shown in (c). This
example state corresponds to the informal box diagram shown in (b) and a state at the return point of the
C-procedure f in (a).
2. We describe a step-by-step abstraction of program states as a cofibered construction of a numeric
abstraction layer on top of a shape abstraction layer (Section 3). In particular, we characterize a
shape abstraction as a combination of an exact abstraction of memory cells along with a summarization operation. Then, we describe how a value abstraction can be applied both globally on
materialized memory locations and locally within summarized regions.
3. We detail the abstract operators necessary to implement an abstract program semantics in terms of
interfaces that a shape abstraction and a value abstraction must implement (Section 4).
4. We overview a modular construction of a shape-numeric static analyzer based on our abstract
operators (Section 5).
2
A concrete semantics
We first define a concrete program semantics for a generic imperative programming language.
2.1 Concrete memory states
We define a “bare metal” model of machine memory. A concrete store is a partial function σ ∈ H =
A ⇀fin V from addresses to values. An address a ∈ A is also considered a value v ∈ V, that is, we assume
that A ⊆ V. For simplicity, we assume that all cells of any store σ have the same size (i.e., word-sized)
and that all addresses are aligned (i.e., word-aligned). For example, we can imagine a standard 32-bit
architecture where all values are 4-bytes and all addresses are 4-byte–aligned. We write for dom(σ ) the
set of addresses at which σ is defined, and we let σ [a ← v] denote the heap obtained after updating the
cell at address a with value v. A concrete environment E ∈ E = X → A maps program variables to their
addresses. That is, we consider all program variables as mutable cells in the concrete store—the concrete
environment E indicates where each variable is allocated. A concrete memory state m simply pairs a
concrete environment and a concrete store: (E, σ ). Thus, the set of memory states M = E × H is the
product of the set of concrete environments and the set of concrete stores.
B.-Y. E. Chang and X. Rival
163
Figure 1(c) shows an example concrete memory state at the return point of the procedure f in (a).
The environment E has two bindings for the variables x and y that are in scope. For concreteness, we
show the concrete store for this example laid out using 32-bit addresses and a C-style layout for struct s.
The figure shown in (b) shows the concrete store as an informal box diagram.
Related work and discussion. Observe that we do not make the distinction between stack and heap
space in a concrete store σ (as in a C-style model), nor have we partitioned a heap on field names (as
in Java-style model). We have intentionally chosen this rather low-level definition of concrete memory
states—essentially an assembly-level model of memory—and leave any abstraction to the definition of
abstract memory states. An advantage of this approach is the ability to use a common concrete model
for combining abstractions that make different choices about the details they wish to expose or hide [22].
For example, Laviron et al. [22] defines an abstract domain that treats precisely C-style aggregates: both
structs and unions with sized-fields and pointer arithmetic. Another abstract domain [36] abstracts
nested structures using a hierarchical abstraction. Rival and Chang [29] defines an abstraction that simultaneously summarizes the stack of activation records and the heap data structures (with a slightly
extended notion of concrete environments), which is useful for analyzing recursive procedures.
2.2 Concrete program semantics
For the most part, we can be agnostic about the particulars of the imperative programming language
of interest. To separate concerns between abstracting memory and control points on which abstract
interpretation collects, all we assume is that a concrete execution state consists of a control state and a
concrete memory state. A shape-numeric abstract domain as we define in Section 3 abstracts the concrete
memory state component.
Definition 1 (Execution states). An execution state s ∈ S consists of a triple (ℓ, E, σ ) where ℓ ∈ L is a
control state, E ∈ E is an concrete environment, and σ ∈ H is a concrete store. The memory component
of an execution state is the pair (E, σ ) ∈ M.
Thus, the set of execution states S = L × E × H ≡ L × M. A program execution is described by a finite
trace, that is, a finite sequence of states hs0 , . . . , sn i. We let T = S⋆ denote the set of finite traces over S.
To make our examples more concrete, we consider a C-like programming language whose syntax is
shown in Figure 2. A location expression loc names a memory cell, which can be a program variable x,
a field offset from another memory location loc1 · f, or the memory location named by a pointer value
⋆exp. We write f ∈ F for a field name and implicitly read any field as an offset, that is, we write a + f for
the address a′ ∈ A obtained by offsetting an address a with field f. To emphasize that we mean C-style
field offset as opposed to Java-style field dereference, we write x · f for what is normally written as x.f
in C. As in C, we write exp -> f for Java-style field dereference, which is a shorthand for (⋆exp) · f. An
expression exp can be a memory location expression loc, an address of a memory location &loc, or any
value literal v, some other n-ary operator ⊕(exp). Like in C, a memory location expression loc used as
an expression (i.e., “r-value”) refers to the contents of the named memory cell, while the &loc converts
the location’s address (i.e., “l-value”) into a pointer “r-value.” We leave the value literals v (e.g., 1) and
expression operators ⊕ (e.g., !, +, ==) unspecified.
An operational semantics: Given a program p, we assume its execution is described by a transition
relation → p ⊆ S × S. This relation defines a small-step operational semantics, which can be defined as
a structured operational semantics judgment s → p s′ . Such a definition is completely standard for our
language, so we do not detail it here.
Modular Construction of Shape-Numeric Analyzers
164
loc (∈ LX ) ::= x
(x ∈ X)
| loc1 · f (loc1 ∈ LX ; f ∈ F)
| ⋆exp
(exp ∈ EX )
p (∈ PX ) ::=
|
|
|
|
|
loc = exp
loc = malloc({f1 , . . . , fn })
free(loc)
p1 ; p2
if (exp) p1 else p1
while (exp) p1
exp (∈ EX ) ::=
|
|
|
⊕ ::=
loc
&loc
v
⊕(exp)
···
(loc ∈ LX ; exp ∈ EX )
(loc ∈ LX ; [f1 , . . . , fn ] ∈ F∗ )
(loc ∈ LX )
(p1 , p2 ∈ PX )
(exp ∈ EX ; p1 , p2 ∈ PX )
(exp ∈ EX ; p1 ∈ PX )
(loc ∈ LX )
(loc ∈ LX )
(v ∈ V)
(exp ∈ EX )
assignment
memory allocation
memory deallocation
sequence
condition test
loop
Figure 2: Abstract syntax for a C-like imperative programming language. A program p consists of
assignment, dynamic memory allocation and deallocation, sequences, condition tests, and loops. An
assignment is specified by a location expression loc that names a memory cell to update and an expression
exp that is evaluated to yield the new contents for the cell. For simplicity, we specify allocation with a
list of field names (i.e., malloc({f1 , . . . , fn })).
As an example rule, consider the D OA SSIGNMENT
case for an assignment loc = exp (ℓ , E, σ ) →
pre
loc = exp (ℓpost , E, σ [LJlocK(E, σ ) ← EJexpK(E, σ )])
where ℓpre and ℓpost are the control
def
def
points before and after the assignLJ⋆expK = EJexpK
LJxK(E, σ ) = E(x)
def
def
ment, respectively. We assume that
EJlocK(E, σ ) = σ ◦ LJlocK(E, σ )
EJ&locK = LJlocK
the semantics of a location expresdef
LJloc · fK(E, σ ) = LJlocK(E, σ ) + f
sion LJlocK is a function from memory states to addresses M → A and
Figure 3: A small-step operational semantics for programs.
that the semantics of an expression
EJexpK is a function from memory states to values M → V. Then, the transition relation for assignment
simply updates the input store σ at the address given by loc with the value given by exp as shown in
Figure 3. The evaluation of locations loc and expressions exp, that is, LJlocK(E, σ ) and EJexpK(E, σ ),
respectively, can be defined by induction on their structure. The environment E is used to lookup the
allocated address for program variables in LJxK. The value for a memory location EJlocK is obtained by
looking up the contents in the store σ . Dereference ⋆exp and &loc mediate between address and value
evaluation, while field offset loc · f is simply an address computation. The evaluation of the remaining
expression forms is completely standard.
Example 1 (Evaluating an assignment). Using the concrete memory state (E, σ ) shown in Figure 1,
the evaluation of the assignment x -> a -> b = y · c proceeds as follows. First, the right-hand side gets
evaluated by noting that E(y) = 0x...b0 and following
EJy · cK(E, σ ) = σ (LJy · cK(E, σ )) = σ (LJyK(E, σ ) + c) = σ (E(y) + c) = σ (0x...b8) = 178 .
Second, the left-hand side gets evaluated by noting that E(x) = 0x...a0 and then following the location
evaluation LJx -> a -> bK(E, σ ) = σ (σ (0x...a0) + a) + b) = σ (0x...b0 + a) + b = 0x...c0 + b = 0x...c4.
Finally, the store is updated at address 0x...c4 with the value 178 with σ [0x...c4 ← 178].
B.-Y. E. Chang and X. Rival
165
Concrete program semantical definitions: Several notions of program semantics can be used as a
basis for static analysis, which each depend on the desired properties and the kinds of invariants needed
to establish them. A semantical definition expressed as the least fixed-point of a continuous function
F over a concrete, complete lattice is particularly well-suited to the design of abstract interpreters [10].
Following this analysis design methodology, an abstract interpretation consists of (1) choosing an abstraction of the concrete lattice (Section 3), (2) designing abstract operators that over-approximate the
effect of the transition relation → p and concrete joins ∪ (Section 4), and (3) applying abstract operators
to over-approximate F using widening (Section 5).
Definition 2 (A concrete domain). Let us fix a form for our concrete domains D to be the powerset of
some set of concrete objects O, that is, let D = P(O). Domain D form a complete lattice with subset
containment ⊆ as the partial order. Hence, concrete joins are simply set union ∪.
For a program p, let ℓpre be its entry point (i.e., its initial control state). A standard definition of
interest is the set of reachable states, which is sufficient for reasoning about safety properties.
Example 2 (Reachable states). We write JpKr for the set of reachable states of program p, that is,
JpKr = {s | (ℓpre , E, σ ) →⋆p s for some E ∈ E and σ ∈ H}
def
where →⋆p is the reflexive-transitive closure of the single-step transition relation →. Alternatively, JpKr
can be defined as lfp Fr , the least-fixed point of Fr , where Fr : P(S) → P(S) is as follows:
Fr (S) = {(ℓpre , E, σ ) | E ∈ E and σ ∈ H} ∪ {s′ | s ∈ S and s → p s′ for some s′ ∈ S} .
def
Note that we have let the concrete objects O be the execution states S in this example.
def
We can also describe the reachable states denotationally [34]—JpKd (E, σ ) = {s | (ℓpre , E, σ ) →⋆p s}—
that enables a compositional way to reason about programs. Here, we let the set of concrete objects be
functions from memory states to sets of states (i.e., M → P(S)).
Related work and discussion. For additional precision or for richer properties, it may be critical to
retain some information about the history of program executions (i.e., how a state can be reached) [30].
In this case, we might choose a trace semantics as a concrete semantics where the concrete objects O are
def
chosen to be traces T. For instance, the finite prefix traces semantics is defined by JpKt = {hs0 , . . . , sn i |
s0 : (ℓpre , E0 , σ0 ) and si → p si+1 for some E0 ∈ E, σ0 ∈ H and for all 0 ≤ i < n}. Or we may to choose to
define a trace semantics denotationally JpKdh : M → P(T) that maps input memory states into traces
starting from them.
In this section, we have left the definition of a control state essentially abstract. A control state is
simply a member of a set of labels on which an interpreter visits. In the intraprocedural setting, the
control state is usually a point in the program text corresponding to a program counter. Since the set
of program points is finite, the control state can be left unabstracted yielding a flow-sensitive analysis.
Meanwhile, richer notions of control states are often needed for interprocedural analysis [26, 35].
3
Abstraction of memory states
In this section, we discuss the abstraction of memory states, including environments and stores, as well
as the values stored in them. A shape abstraction typically abstracts entire stores but only the pointer
values (i.e., addresses) in them. In contrast, a numeric abstraction is typically applied only to the data
Modular Construction of Shape-Numeric Analyzers
166
values stored in program variables (i.e., the part of the store containing the global and local variables).
We defer the abstraction of program executions to Section 5.
Following the abstract interpretation framework [10], an abstraction or abstract domain is a set of
abstract properties D♯ together with a concretization function and sound abstract operators.
Definition 3 (Concretization). A concretization function γ : D♯ → D defines the meaning of D♯ in terms
of a concrete domain D = P(O) for some set of concrete objects O. An abstract inclusion d1♯ ⊑ d2♯ for
abstract elements d1♯ , d2♯ ∈ D♯ should be sound with respect to concrete inclusion: γ (d1♯ ) ⊆ γ (d2♯ ), and
γ should be monotone. For each concrete operation f , we expect a sound abstract counterpart f ♯ ; for
example, an abstract operation f ♯ : D♯ → D♯ is sound with respect to a concrete operation f : D → D if
and only if γ (d ♯ ) ⊆ γ ◦ f ♯ (d ♯ ) for all d ♯ ∈ D♯ .
In this section, we focus on the abstract domains and concretization functions, while the construction
of abstract operations are detailed in Section 4.
3.1 An exact store abstraction based on separating shape graphs
An abstract heap σ ♯ ∈ H♯ should over-approximate a set of concrete heaps with a compact representation.
This set of abstract heaps H♯ form the domain of abstract heaps (or the shape abstract domain). For
simplicity, we first consider an exact abstraction of heaps with no unbounded dynamic data structures.
That is, such an abstraction explicitly enumerates a finite number of memory cells and performs no
summarization. Summarization is considered in Section 3.3.
A heap can be viewed as a set of disjoint cells (cf., Figure 1). At the abstract level, it is convenient
to make disjointness explicit and describe disjoint cells independently. Thus, we write σ0♯ ∗ σ1♯ for the
abstract heap element that denotes all that can be partitioned into a sub-heap satisfying σ0♯ and another
disjoint sub-heap satisfying σ1♯ . This observation about disjointness underlies separation logic [28] and
thus we borrow the separating conjunction operator ∗ from there. An individual cell is described by an
exact points-to predicate of the form α · f 7→ β where α , β are symbolic variables (or, abstract values)
drawn from a set V♯ . The symbolic variable α denotes an address, while β represents the contents at the
memory cell with address α · f (i.e., α offset by a field f). An exact heap abstraction is thus a separating
conjunction of a set of exact points-to predicates.
Such abstract heap predicates can be represented using
αy
αx
βa
δa
a
a
separating shape graphs [8, 22] where nodes are symbolic
variables and edges represent heap predicates. An exact
b
b
βb
δb
points-to predicate α · f 7→ β is denoted by an edge from
c
c
node α to node β with a label for the field offset f. For
βc
δc
example, βa denotes the value corresponding to the C expression y · a.
The concretization γH of a separating shape graph Figure 4: separating shape graph abstracmust account for symbolic variables that denote some con- tion of σ in Figure 1. Symbolics αx and αy
crete values, so it also must yield an instantiation or a denote the address of x and y, respectively.
valuation ν : V♯ → V. Thus, this concretization has type
γH : H♯ → P(H × (V♯ → V)) and is defined as follows (by induction on the structure σ ♯ ):
γH (α · f 7→ β ) = {([ν (α ) + f 7→ ν (β )], ν ) | ν ∈ V♯ → V}
def
/ .
γH (σ0♯ ∗ σ1♯ ) = {(σ0 ⊎ σ1 , ν ) | (σ0 , ν ) ∈ γH (σ0♯ ) and (σ1 , ν ) ∈ γH (σ1♯ ) and dom(σ0 ) ∩ dom(σ1 ) = 0}
def
That is, an exact points-to predicate corresponds to a single cell concrete store under a valuation ν ,
and a separating conjunction of abstract heaps is a concrete store composed of disjoint sub-stores that
B.-Y. E. Chang and X. Rival
typedef struct {
int a;
int b;
} t;
t x;
&x = 0x...
&x = 0x...
167
0
2
8
18
&x = 0x...
&x = 0x...
3
7
10
21
a
βa
αx
b
βb
0 ≤ βa ≤ 10 ∧ βa ≤ 2βb + 1
Figure 5: An example separating shape graph enriched with a numeric constraint (right) with four concrete instances (center) for the C type declaration (left).
are individually abstracted by the conjuncts under the same instantiation (as in separation logic [28]).
Symbolic variables can be viewed as existentially-quantified variables that are bound at the top-level of
the abstraction. The valuation makes this explicit and thus is a bit similar to a concrete environment E.
Related work and discussion. Separating conjunction manifests itself in separating shape graphs as
simply distinct edges. In other words, distinct edges denote disjoint heap regions. Separating shape
graphs are visually quite similar to classical shape and points-to graphs [9, 31] but are actually quite
different semantically. In classical shape and points-to graphs, the nodes represent memory cells, and
typically, a node corresponds to one-or-more concrete cells. Distinct nodes represent disjoint memory
memory regions, and edges express variants of may or must points-to relations between two sets of cells.
In contrast, it is the edges in separating shape graphs that correspond to disjoint memory cells, while
the nodes simply represent values. We have found two main advantages of this approach. First, because
there is no a priori requirement that two nodes be distinct values, we do not need to case split simply
to speak about the contents of cells (e.g., consider two pointer variables x and y and representing to
which objects they point; a classic shape graph must consider two cases where x and y are aliases or
not, while a separating shape graph does not). Limiting case splits is critical to getting good analysis
performance [5]. Second, a separating shape graph is agnostic to the type of values that nodes represent.
Nodes may represent addresses, but they can just as easily represent non-address values, such as integer, Boolean, or floating-point values. We take advantage of this observation to interface with numeric
abstract domains [7], which we discuss further next in Section 3.2.
3.2 Enriching shapes with a numeric abstraction
From Section 3.1, we have an exact heap abstraction based on a separating shape graph with a finite number of exact points-to edges. Intuitively, this abstraction is quite weak, as we have simply enumerated the
memory cells of interest. We have, however, given names to all values—both addresses and contents—of
potential interest. Here, we enrich the abstraction with information about the values contained in data
structures, not just the pointer shape. We focus on scalar numeric values, such as integers or floatingpoint values, but other types of values could be handled similarly. A separating shape graph defines a
set of symbolic variables corresponding to values, so we can abstract the values those symbolic variables
represent. First, we consider a simple example, shown in Figure 5. In Figure 5, we show four concrete
stores such that 0 ≤ x · a ≤ 10 and x · a ≤ 2(x · b) + 1. The separating shape graph on the right clearly
abstracts the shape of the four stores (i.e., two fields a and b off a struct at variable x). The symbolic
variables βa and βb represent the contents of cells x · a and x · b, respectively, so the numeric property
specified above can expressed simply by using a logical formula involving βa and βb (as shown).
In general, a separating shape graph σ ♯ is defined over a set of symbolic variables V♯ [σ ♯ ] where
♯
V [σ ♯ ] ⊆ V♯ . The properties of the values stored in heaps described by σ ♯ can be characterized by
Modular Construction of Shape-Numeric Analyzers
168
σ4♯
bC
b
σ3♯
Dnum hσ4♯ i
b
Dnum hσ3♯ i
bC
b
b
σ1♯
bC
b
b
Dnum hσ1♯ i
b
b
σ2♯
Dnum hσ2♯ i
bC
b
b
b
bC
b
σ0♯
b
b
Dnum hσ0♯ i
b
Figure 6: The combined shape-numeric abstract domain is a cofibered layering of a numeric abstract
domain on a shape abstract domain.
logical formulas over V♯ [σ ♯ ]. Such logical formulas expressing numeric properties can be represented
using a numeric abstract domain Dnum hV♯ [σ ♯ ]i that abstracts functions from V♯ [σ ♯ ] to V, that is, it
comes with concretization function parametrized by a set of symbolic values V♯ [σ ♯ ] of the following
type: γnum hV♯ [σ ♯ ]i : Dnum hV♯ [σ ♯ ]i → P(V♯ [σ ♯ ] → V). For example, the numeric property mentioned
in Figure 5 could be expressed using the convex polyhedra abstract domain [12]. As a shape graph
concretizes into a set of pairs composed of a heap σ and a valuation ν : V♯ [σ ♯ ] → V, such numeric
constraints simply restrict the set of admissible valuations.
The need to combine a shape graph with a numeric constraint suggests using a product abstraction [11] of a shape abstract domain H♯ and a numeric abstract domain Dnum h−i. However, note that
the numeric abstract domain that needs to be used depends on the separating shape graph, as the set of
dimensions is equal to the set of nodes in the separating shape graph. Therefore, the conventional notion
of a symmetric reduced product does not apply here. Instead, we use a different construction known as a
cofibered abstract domain [38] (in reference with the categorical notion underlying this construction).
Definition 4 (Combined shape-numeric abstract domain). Given a shape domain H♯ and a numeric domain Dnum h−i parametrized by a set of symbolic variables. We let N♯ denote the set of numeric abstract
def S
values corresponding to any shape graph (i.e., N♯ = {Dnum hV i | V ⊆ V♯ }), and we define the combined
shape-numeric abstract domain H♯ ⇒ N♯ and its concretization γH♯ ⇒N♯ : (H♯ ⇒ N♯ ) → P(H×(V♯ → V))
as follows:
H♯ ⇒ N♯ = {(σ ♯ , ν ♯ ) | σ ♯ ∈ H♯ and ν ♯ ∈ Dnum hV♯ [σ ♯ ]i}
def
γH♯ ⇒N♯ (σ ♯ , ν ♯ ) = {(σ , ν ) | (σ , ν ) ∈ γH (σ ♯ ) and ν ∈ γnum hV♯ [σ ♯ ]i(ν ♯ )}
def
This product is clearly asymmetric, as the left member defines the abstract lattice to which the right
member belongs. We illustrate this structure in Figure 6. The left part depicts the lattice of abstract
heaps, while the right part illustrates a lattice of numeric lattices. Each element of the lattice of lattices
is an instance of the numeric abstract domain over the symbolic variables defined by the abstract heap,
that is, it is the image of the function σ ♯ 7→ Dnum hV♯ [σ ♯ ]i.
This dependence is not simply theoretical but has practical implications on both the representation
of abstract values and the design of abstract operations in the combined abstract domain. For instance,
B.-Y. E. Chang and X. Rival
169
a
a
αx
β1
∧ β1 = δ1
αx
β0
b
b
δ1
Figure 7: Two abstractions drawn from the combined abstract domain H♯ ⇒ N♯ that have equivalent
concretizations but with non-isomorphic sets of symbolic variables.
αx
a
b
β
−→x·b=x·a
δ
αx
a
b
β
δ
Figure 8: Applying the transfer function for an assignment on a separating shape graph that changes the
set of “live” symbolic variables.
Figure 7 shows two separating shape graphs together with numerical invariants that represent the same set
of concrete stores even though they use two different sets of symbolic variables (even up to α -renaming).
Both of these combined shape-numeric abstract domain elements represent a store with two fields x · a
and x · b such that x · a = x · b. In the right abstract domain element, the contents of both fields are
associated with distinct nodes, and the values denoted by those nodes are constrained to be equal by the
numeric domain. In the left graph, the contents of both fields are associated to the same node, which
implies that they must be equal (without any constraint in the numeric domain).
Now, with respect to the design of abstract operations in the combined abstract domain, the set of
nodes in the shape graph will in general change during the course of the analysis. For instance, the
analysis of an assignment of the value contained into field a to field b from the abstract state shown in
the left produces the one in the right in Figure 8. After this transformation takes place, node δ becomes
“garbage” or irrelevant, as it is not linked anywhere in the shape graph, and no numeric property is
attached to it. This symbolic variable δ should thus be removed or projected from the numeric abstract
domain. Other operations can cause new symbolic variables to be added, and this issue is only magnified
with summaries (cf., Section 3.3). Thus, the combined abstract domain must take great care in ensuring
the consistency of the numeric abstract values with the shape graphs, as well as dealing with graphs
with different sets of nodes. Considering again the diagram in Figure 6, whenever two shape graphs
are ordered σ0♯ ⊑ σ1♯ , there exists a symbolic variable renaming function Φhσ0♯ , σ1♯ i : V♯ [σ1♯ ] → V♯ [σ0♯ ]
that expresses a renaming of the symbolic variables from the weaker shape graph σ1♯ to the stronger
one σ0♯ . For example, the symbolic renaming function Φ for the shape graphs shown in Figure 7 is
[αx 7→ αx , β1 7→ β0 , δ1 7→ β0 ].
Related work and discussion. In practice, the implementation of the shape abstract domain takes the
form of a functor (in the ML programming sense) that takes as input a module implementing a numeric
domain interface (e.g., a wrapper on top of the A PRON library [20]) and outputs another module that
implements the memory abstract domain interface. The construction that we have shown in this section
is general to analyses where the set of symbolic variables is dynamic during the course of the analysis
and where the inference of this set is bound to the inference of cell contents. In other words, it is wellsuited to applying shape analyses for summarizing memory cells and then reasoning about their contents
with another domain. This construction has been used not only in Xisa [7] but also in a TVLA-based
setup [25] and one based on a history of heap updates [6].
Another approach that avoids this construction by performing a sequence of analyses: first, a shape
170
Modular Construction of Shape-Numeric Analyzers
analysis infers the set of symbolic variables; then, a numeric static analysis relies on this set [23, 24].
While less involved, this approach prevents the exchange of information between both analyses, which is
often required to achieve a satisfactory level of precision [7]. This sequencing of heap analysis followed
by value analysis is similar to the application of a pre-pass pointer analysis followed by model checking
over a Boolean abstraction exemplified in SLAM [1] and BLAST [18]
3.3 Enhancing store abstractions with summaries
So far, we have considered very simple abstract heaps described by separating shape graphs where all
concrete memory cells are abstracted by exact points-to edges. To support abstracting a potentially
unbounded number of concrete memory cells via dynamic memory allocation, we must extend abstract
heaps with summarization, that is, a way of providing a compact abstraction for possibly unbounded,
possibly non-contiguous memory regions.
As an example, consider the conαx
&x 0x...
&x 0x...
&x 0x0
crete stores shown in the left part of Figure 9 consisting of a series of linked lists
0x0
0x...
...
...
with 0, 1, and 2 elements. These conβ
crete stores are just instances among inlist
0x0
finitely many ones where x stores a ref...
erence to a list of arbitrary length. Each
of these instances consist of two re- Figure 9: Summarizing linked lists with inductive predicate
gions: the cell corresponding to variable edges in separating shape graphs.
x (green) and the list elements (blue).
To abstract all of these stores in a compact and precise manner, we need to summarize the second region with a predicate. We can define such a predicate for summarizing such a region using an inductive
definition list following the structure of lists: α · list := (emp ∧ α = 0x0) ∨ (α · a 7→ β0 ∗ α · b 7→ β1 ∗
β0 · list ∧ α 6= 0x0). This definition notation is slightly non-standard to match the graphical notation: the
predicate name is list and α is the formal induction parameter. A list memory region is empty if the root
pointer α of the list is null, or otherwise, there is a head list element with two fields a and b such that the
contents of cell α · a called β0 is itself a pointer to a list. Then, in Figure 9, if variable x contains a pointer
value denoted by β , the second region can be summarized by the inductive predicate instance β · list.
Furthermore, the three concrete stores are abstracted by the abstract heap αx 7→ β ∗ β · list (drawn as a
graph to the right). The inductive predicate β · list is drawn as the bold, thick edge from node β .
Materialization: The analyzer must be able to apply transfer functions on summarized regions. However, designing precise transfer functions on arbitrary summaries is extremely difficult. An effective
approach is to define direct transfer functions only on exact predicates and then define transfer functions
on summaries indirectly via materialization [32] of exact predicates from them. In the following, we
focus on the case where summaries are derived from inductive predicates [8] and thus call the materialization operation unfolding. In practice, unfolding should be guided by a specification of the summarized
region where the analyzer needs to perform local reasoning on materialized cells (see Section 4.2). However, from the theoretical point of view, we can let an unfolding operator be defined as some function
♯
♯
, νn−1
).
that replaces one abstract (σ ♯ , ν ♯ ) with a finite set of abstract elements (σ0♯ , ν0♯ ), . . . , (σn−1
Definition 5 (Materialization). Let us write
⊆ (H♯ ⇒ N♯ ) × Pfin (H♯ ⇒ N♯ ) for the unfolding relation.
B.-Y. E. Chang and X. Rival
171
Then, any unfolding of an abstract element should be sound with respect to concretization:
If (σ ♯ , ν ♯ )
♯
♯
, νn−1
) , then γH♯ ⇒N♯ (σ ♯ , ν ♯ ) ⊆
(σ0♯ , ν0♯ ), . . . , (σn−1
[
γH♯ ⇒N♯ (σi♯ , νi♯ ) .
0≤i<n
As seen above, the finite set of abstract elements that results from materialization represents a disjunction
of abstract elements (i.e., materialization is a form of case analysis). For precision, we typically want
an equality instead of inclusion in the conclusion, which motivates a need to represent a disjunction of
abstract elements (cf., Section 3.4).
Example 3 (Unfolding an inductively-defined list). For instance, the abstract element from H♯ ⇒ N♯
depicted in Figure 9 can be unfolded to two elements:
(αx 7→ β ∗ β · list, ⊤)
(αx 7→ β , β = 0x0), (αx 7→ β ∗ β · a 7→ β0 ∗ β · b 7→ β1 ∗ β0 · list, β 6= 0x0) ,
which means that the list pointer β is either a null pointer or points to a list element whose a field contains
a pointer to another list.
Related work and discussion. Historically, the idea of using compact summaries for an unbounded
number of concrete memory cells goes back to at least Jones and Muchnick [21], though the set of
abstract locations was fixed a priori before the analysis. Chase et al. [9] considered dynamic summarization during analysis, while Sagiv et al. [32] introduced materialization. We make note of existing
analysis algorithms that make use of summarization-materialization. TVLA summary nodes [31] represent unbounded sets of concrete memory cells with predicates that express universal properties of all
the concrete cells they denote. The use of three-valued logic enables abstraction beyond a set of exact
points-to constraints (i.e., the separating shape graphs in Section 3.1 are akin to two-valued structures in
TVLA), and summarization is controlled by instrumentation predicates that limits the compaction done
by canonical abstraction. Fixed list segment predicates [2, 14] characterize consecutive chains of list
elements by its first and last pointers. Thus, a predicate of the form ls(α , α ′ ) denotes all chains of list
elements (of any length) starting at α and ending at α ′ . Then, an abstract heap consists of a separating
conjunction of points-to predicates (Section 3.1) and list segments. These predicates can be generalized
to other structure segments. Inductive predicates [7, 8] generalize the list segment predicates in several
ways. First, the abstract domain may be parametrized by a set of user-supplied inductive definitions.
Note that as parameters to the abstract domain and thus the analyzer, the inductive definitions specify
possible templates for summarization. A sound analysis can only infer a summary predicate essentially
if it exhibits an exact instance of the summary. The “correctness” of such inductive definitions are
not assumed, but rather a disconnect between the user’s intent and the meaning an inductive predicate
could lead to unexpected results. Second, inductive predicates can correspond to complete structures
(e.g., a tree that is completely summarized into a single abstract predicate), whereas segments correspond to incomplete structures characterized by a missing sub-structure. Inductive predicates can be
generically lifted to unmaterializable segment summaries [8] or materializable ones [7]. Independently,
array region predicates [15] have been used to describe the contents of zones in arrays. Some analyses
on arrays and containers have used index variables into summaries instead of explicit materialization
operations [13, 16, 17].
3.4 Lifting store abstractions to disjunctive memory state abstractions
At this point, we have described an abstraction framework for concrete stores σ . To complete an abstraction for memory states m : (E, σ ), we need two things: (1) an abstract counterpart to E and (2) a
disjunctive abstraction for when a single abstract heap σ ♯ is insufficient for precisely abstracting the set
of possible concrete stores.
Modular Construction of Shape-Numeric Analyzers
172
Abstract environments: Since the abstract counterpart for addresses are symbolic variables (or nodes)
in shape graphs, an abstract environment E ♯ can simply be a function mapping program variables to
nodes, that is, E ♯ ∈ E♯ = X → V♯ . Now, the memory abstract domain M♯ is defined by M♯ = E♯ × (H♯ ⇒
N♯ ), and its concretization γM : M♯ → P(E × H) can be defined as follows:
γM (E ♯ , (σ ♯ , ν ♯ )) = {(ν ◦ E ♯ , σ ) | (σ , ν ) ∈ γH (σ ♯ ) and ν ∈ γnum hV♯ [σ ♯ ]i(ν ♯ )} .
def
Note that in an abstract memory state m♯ : (E ♯ , σ ♯ ), the abstract environment E ♯ simply gives the symbolic address of program variables, while the abstract heap σ ♯ abstracts all memory cells—just like the
concrete model in Section 2.2.
We let the abstract environment be depicted by node x
y
a
a
αy
labels in the graphical representation of abstract heaps. αx
βa
δa
For instance, the concrete memory state shown in Figure 1
βb
δb
b
b
can be described by the diagram in Figure 10.
c
c
βc
δc
Disjunctive abstraction: Recall that the unfolding operation from Section 3.3 generates a finite disjunction of Figure 10: Depicting a memory abstraction
abstract facts—specifically, combined shape-numeric ab- including the abstract heap from Figure 4
stract elements {. . . , (σi♯ , νi♯ ), . . .} ⊆ H♯ ⇒ N♯ . Thus, a dis- and an abstract environment.
junctive abstraction layer is required regardless of other
analysis reasons (e.g., path-sensitivity). We assume the
disjunctive abstraction is defined by an abstract domain M♯∨ and a concretization function γ∨ : M♯∨ →
P(M). We do not prescribe any specific disjunctive abstraction. A simple choice is to apply a disjunctive completion [11], but further innovations might be possible by taking advantage of being specific to
memory.
Example 4 (Disjunctive completion). For a memory abstract domain M♯ , its disjunctive completion M♯∨
is defined as follows:
M♯∨ = Pfin (M♯ )
def
γ∨ (s♯ ) =
def
[
{γM (m♯ ) | m♯ ∈ s♯ } .
In Figure 11, we sum up the structure of the abstract domain for abstracting memory states M as a
stack of layers, which are typically implemented as ML-style functors. Each layer corresponds to the
abstraction of a different form of concrete semantics (as shown in the diagram).
Related work and discussion. Trace partitioning [30] relies on control-flow history to manage disjunctions, which could be used as an alternative to disjunctive completion. However, it is a rather general
construction and can be instantiated in multiple ways with a large effect on precision and performance.
4
Static analysis operations
In this section, we describe the main abstract operations on the memory abstract domain M♯ and demonstrate how they are computed through the composition of abstract domains discussed in Section 3. Our
presentation describes each kind of operation (i.e., transfer functions for commands like assignment, abstract comparison, and abstract join) one by one and shows how unfolding and folding operations are
triggered by their application. The end result of this discussion is a description of how these domains
implement the interfaces shown in Figure 12. For these interfaces, we let B denote the set of booleans
{true, false} and U denote an undefined value for some functions that may fail to produce a result. We
write XU for X ⊎ {U} for any set X (i.e., an option type).
B.-Y. E. Chang and X. Rival
173
disjunctive abstract domain
γ∨ : M♯∨ → P(M)
memory abstract domain
γM : M♯ → P(M)
M♯ = E♯ × (H♯ ⇒ N♯ )
combined shape-numeric abstract domain
γH♯ ⇒N♯ : (H♯ ⇒ N♯ ) → P(H × (V♯ → V))
shape abstract domain
γH : H♯ → P(H × (V♯ → V))
numeric abstract domain
γnum hV♯ [σ ♯ ]i : Dnum hV♯ [σ ♯ ]i → P(V♯ [σ ♯ ] → V)
Figure 11: Layers of abstract domains to yield a disjunctive memory state abstraction. From an implementation perspective, the edges correspond to inputs for ML-style functor instantiations.
4.1 Assignment over materialized cells
First, we consider the transfer function for assignment. In this subsection, for simplicity, we focus on
the case where none of the locations that appear in either side of the assignment are summarized, and we
defer the case of transfer functions over summarized graph regions to Section 4.2. Because of this simplification, the types of the abstract operators mentioned will not exactly match those given in Figure 12.
At the same time, this transfer function captures the essence of the shape-numeric combination.
Recall that loc ∈ LX and exp ∈ EX are location and value expressions, respectively, in our programming language (cf., Figure 2). The transfer function assignmem : LX × EX × M♯ → M♯ should compute
a sound post-condition for the assignment command loc = exp stated as follows:
Condition 1 (Soundness of assignmem ). If (E, σ ) ∈ γM (m♯ ), then
(E, σ [LJlocK(E, σ ) ← EJexpK(E, σ )]) ∈ γM (assignmem (loc, exp, m♯ )) .
Assignments of the form loc = loc′ . Let us first assume that right hand side of the assignment is a
location expression. As an example, consider the assignment shown in Figure 13 and applying assignmem
to the pre-condition on the left to yield the post-condition on the right. The essence is that loc dictates an
edge that should be updated to point to the node specified by loc′ .
To compute a post-condition in this case, assignmem should update the abstract heap, that is, the preheap σ ♯ ∈ H♯ . An assignmem call should eventually forward the assignment to the heap abstract domain
via the eval[l]shape operation that evaluates a location expression loc to an edge, eval[e]shape that evaluates
a value expression exp to a node, and mutateshape that swings a points-to edge.
The base of a sequence of pointer dereferences is given by a program variable, so the first step consists
of replacing the program variables in the assignment with the symbolic names corresponding to their
addresses using the abstract environment E ♯ . For our example, this results in the call to assigncomb (α0 ->
a · b, α0 · b, (σ ♯ , ν ♯ )) at the combined shape-numeric layer, which should satisfy a soundness condition
similar to that of assignmem (Condition 1). The next step consists of traversing the abstract heap σ ♯
Modular Construction of Shape-Numeric Analyzers
174
• A shape abstract domain H♯
eval[l]shape :
eval[e]shape :
mutateshape :
unfoldshape :
newshape :
delete[n]shape :
delete[e] shape :
compareshape :
joinshape :
widenshape :
LV♯ × H♯
EV♯ × H♯
V♯ × F × V♯ × H♯
(LV♯ × F) × H♯
V♯ × H♯
V♯ × H♯
V♯ × F × H♯
(V♯ → V♯ ) × H♯ × H♯
((V♯ )2 → V♯ ) × H♯ × H♯
((V♯ )2 → V♯ ) × H♯ × H♯
(V♯ × F × H♯)U
(V♯ × H♯ )U
H♯U
Pfin (H♯ × EV♯ )
H♯
H♯
H♯
{false} ⊎ {true} × (V♯ → V♯ )
H♯ × (V♯ → V♯ )2
H♯ × (V♯ → V♯ )2
−→
−→
−→
−→
−→
−→
−→
−→
−→
−→
• A numeric abstract domain over symbolic variables N♯
assignnum :
guardnum :
newnum :
delete[n]num :
renamenum :
comparenum :
joinnum :
widennum :
LV♯ × EV♯ × N♯
EV♯ × N♯
V♯ × N♯
V♯ × N♯
(V♯ → V♯ ) × N♯
N♯ × N♯
N♯ × N♯
N♯ × N♯
−→
−→
−→
−→
−→
−→
−→
−→
N♯
N♯
N♯
N♯
N♯
B
N♯
N♯
−→
−→
−→
−→
−→
−→
−→
−→
Pfin (H♯ ⇒ N♯ )U
Pfin (H♯ ⇒ N♯ )U
Pfin (H♯ ⇒ N♯ )U
Pfin (H♯ ⇒ N♯ )U
Pfin (H♯ ⇒ N♯ )U
{false} ⊎ {true} × (V♯ → V♯ )
(H♯ ⇒ N♯ )
(H♯ ⇒ N♯ )
• A combined shape-numeric abstract domain H♯ ⇒ N♯
assigncomb :
guardcomb :
unfoldcomb :
alloccomb :
freecomb :
comparecomb :
joincomb :
widencomb :
LV♯ × EV♯ × H♯ ⇒ N♯
EV♯ × (H♯ ⇒ N♯ )
LV♯ × H♯ ⇒ N♯
LV♯ × F∗ × (H♯ ⇒ N♯ )
LV♯ × F∗ × (H♯ ⇒ N♯ )
(V♯ → V♯ ) × (H♯ ⇒ N♯ ) × (H♯ ⇒ N♯ )
((V♯ )2 → V♯ ) × (H♯ ⇒ N♯ ) × (H♯ ⇒ N♯ )
((V♯ )2 → V♯ ) × (H♯ ⇒ N♯ ) × (H♯ ⇒ N♯ )
• A memory abstract domain M♯
assignmem :
guardmem :
allocmem :
freemem :
comparemem :
joinmem :
widenmem :
LX × EX × M♯
EX × M♯
LX × F∗ × M♯
LX × F∗ × M♯
M♯ × M♯
M♯ × M♯
M♯ × M♯
−→
−→
−→
−→
−→
−→
−→
Pfin (M♯ )U
Pfin (M♯ )U
Pfin (M♯ )U
Pfin (M♯ )U
B
M♯
M♯
Figure 12: Interfaces for the abstract domain layers shown in Figure 11 (except the disjunctive abstraction
layer).
B.-Y. E. Chang and X. Rival
y
a
α0
b
a
α1
α2
175
b
α3
y · a -> b = y · b
y
a
α0
b
α4
α1
α2
b
a
α3
α4
Figure 13: Applying assignmem to an example assignment of the form loc = loc′ .
L OC A DDRESS
L OC F IELD
L OC VAL
eval[l]shape (loc · g, σ ♯ ) = (α , f + g)
eval[l]shape (⋆exp, σ ♯ ) = (α , 0)
/
eval[l]shape (loc, σ ♯ ) = (α , f)
eval[l]shape (α , σ ♯ ) = (α , 0)
/
VAL D EREFERENCE
eval[l]shape (loc, σ ♯ ) = (α , f)
σ ♯ = σ0♯ ∗ α · f 7→ β
eval[e]shape (loc, σ ♯ ) = β
eval[e]shape (exp, σ ♯ ) = α
VAL L OC
eval[l]shape (loc, σ ♯ ) = (α , 0)
/
eval[e]shape (&loc, σ ♯ ) = α
Figure 14: Evaluating dereferences in an abstract heap.
according to the location expression and the value expression of the assignment. As mentioned above,
this evaluation is performed using the location evaluation function eval[l]shape that yields an edge and the
value expression evaluation function eval[e]shape that yields a node.
Condition 2 (Soundness of eval[l]shape and eval[e]shape ). Let (σ , ν ) ∈ γH (σ ♯ ). Then,
If eval[l]shape (loc, σ ♯ ) = (α , f) , then LJlocK(σ ) = ν (α ) + f .
If eval[e]shape (loc, σ ♯ ) = β , then EJlocK(σ ) = ν (β ) .
In Figure 14, we define eval[l]shape and eval[e]shape following the syntax of location and value expressions (over symbolic variables). We write 0/ for a designated 0-offset field. This abstract evaluation
corresponds directly to the concrete evaluation defined in Figure 3. Note that abstract evaluation is
not necessarily defined for all expressions. For example, an points-to edge may simply not exist for
the computed address in VAL D EREFERENCE. The edge may need to be materialized by unfolding (cf.,
Section 4.2) or otherwise is a potential memory error.
Returning to the example in Figure 13, we get eval[l]shape (α0 -> a · b, σ ♯ ) = (α1 , b)—the cell being
assigned-to corresponds to the exact points-to edge α1 · b 7→ α4 —and eval[e]shape (α0 · b) = α2 —the value
to assign is abstracted by α2 . The abstract post-condition returned by assigncomb should reflect the
swinging of that edge in the shape graph, which is accomplished by the mutateshape function:
mutateshape (α , f, β , (α · f 7→ δ ) ∗ σ ♯ ) = (α · f 7→ β ) ∗ σ ♯ .
This function simply replaces a points-to edge named by the address α and field f with a new one for
the updated contents (and fails if such a points-to edge does not exist in the abstract heap σ ♯ ). The
effect of this assignment can be completely reflected in the abstract heap since the cell corresponding
to the assignment is abstracted by exactly one points-to edge and the new value to store in that cell is
also exactly abstracted by one node. We note that node α4 is no longer reachable in the shape graph,
and thus the value that this node denotes is no longer relevant when concretizing the abstract state. As a
consequence, it can be safely removed both in H♯ (using function delete[n] shape ) and in N♯ (using function
Modular Construction of Shape-Numeric Analyzers
176
a
α0
b
y
c
a
α1
α2
∧ α3 ≤ α2
y·c=y·b+1
α0
y
α1
c
α3
∧ α3 ≤ α2
∧ α4 = α2 + 1
α2
b
α4
α3
Figure 15: Applying assignmem to an example assignment of the form loc = exp.
x
α0
y
α2
list
α1
x
∧ α2 6= 0x0
y = x -> next
α0
α2
y
next
d
α1
α3
list
∧ α2 6= 0x0
α4
Figure 16: Applying assignmem to an example that affects the summarized region α2 · list.
delete[n] num ). Such an existential projection or “garbage collection” step may be viewed as a conversion
operation in the cofibered lattice structure shown in Figure 6.
Assignments of the form loc = exp. In general, the right-hand side of an assignment is not necessarily
a location expression. The evaluation of left-hand side loc proceeds as above, but the evaluation of the
right-hand side expression exp is extended. As an example, consider the assignment shown in Figure 15.
The evaluation of the location expression down to the abstract heap level works as before where
we find that eval[l]shape (α0 · c, σ ♯ ) = (α0 , c). For the right-hand–side expression, it is not obvious what
eval[e]shape (α0 ·b+ 1, σ ♯ ) should return, as no symbolic node is equal to that value in the concretization of
all elements of σ ♯ . It is possible to evaluate sub-expression α0 · b to α2 , but then eval[e]shape (α2 + 1, σ ♯ )
cannot be evaluated any further. The solution is to create a new symbolic variable and constrain it to
represent the value of the right-hand–side expression. Therefore, the evaluation of assigncomb proceeds
as follows: (1) generate a fresh node α4 ; (2) add α4 to the abstract heap σ ♯ and the numeric abstract
value ν ♯ using the function newshape and newnum , respectively; (3) update the numeric abstract value ν ♯
using assignnum (α4 , α2 + 1, ν ♯ ), which over-approximates constraining α4 = α2 + 1; and (4) mutate with
mutateshape with the new node α4 (i.e., mutateshape (α0 , c, α4 , σ ♯ )).
4.2 Unfolding and assignment over summarized cells
We now consider assignmem in the presence of
summary predicates, which intuitively “get in struct list { struct list ⋆ next; int d; };
the way” of evaluating location and value ex- α · list := (emp ∧ α = 0x0)
∨ (α · next 7→ β0 ∗ α · d 7→ β1 ∗ β0 · list ∧ α 6= 0x0)
pressions in a shape graph. For instance, consider trying to apply the assignment shown in
Figure 16. On the left, we have a separating shape graph where α2 is a list described by the inductive definition shown inset. For clarity, we also show the C-style struct definition that corresponds to the layout of
each list element. In applying the assignment, the evaluation of the right-hand–side expression x -> next
fails. While x evaluates to node α2 , there is no points-to edge from α2 . Thus, eval[e]shape (α0 -> next)
fails. It is clear that the reason for this failure is that the memory cell corresponding to the right-hand–
side expression is summarized as part of the α2 · list predicate. To materialize this cell, this predicate
B.-Y. E. Chang and X. Rival
177
should be unfolded; then, the assignment can proceed as in the previous section (Section 4.1). We can
now describe the transfer function for assignment assignmem (loc, exp, (σ ♯ , ν ♯ )) in general:
1. It should call the underlying assigncomb and follow the process described previously in Section 4.1.
If evaluation via eval[l]shape or eval[e]shape fail, then they should return a failure address, which
consists of a pair (β , f) corresponding to the node and field offset that does not have a materialized
points-to edge. In the example in Figure 16, the failure address is (α2 , next). Note that the interface
for evaluation shown in Figure 16 does not show the contents of the failure case for simplicity.
2. Then, assigncomb in the combined domain performs an unfolding of the abstract heap by calling
a function unfoldshape that implements the unfolding relation
with the target points-to edge to
materialize (β , f).
Condition 3 (Soundness of unfoldshape ).
γH (σ ♯ ) ⊆
[
{(σ , ν ) ∈ γH (σu♯ ) | (σu♯ , expu ) ∈ unfoldshape ((β , f), σ ♯ ) and Jexpu K(ν ) = true} .
Note that unfolding of an abstract heap returns pairs consisting of an unfolded abstract heap and
a numeric constraint as an expression expu ∈ EV♯ [σ ♯ ] over the symbolic variables of the unfolded
u
abstract heap. This expression allows a summary to contain constraints not expressible in a shape
graph itself. For instance, in the list inductive definition, each case comes with a nullness or nonnullness condition on the head pointer. Or more interestingly, we can imagine an orderedness
constraint for an inductive definition describing an ordered list. For the example from Figure 16,
unfolding the shape graph at (α2 , next) generates two disjuncts, but the one corresponding to the
empty list can be eliminated due to the constraint that α2 has to be non-null.
3. The numeric constraints should be evaluated in the numeric abstract domain using a condition test
operator guardnum .
Condition 4 (Soundness of guardnum ). Let V ⊆ V♯ , ν ♯ ∈ Dnum hV i, and ν ∈ γnum hV i(ν ♯ ). Then,
If JexpK(ν ) = true , then ν ∈ γnum hV i(guardnum (exp, ν ♯ )) .
Thus, the initial abstract state in the combined domain (σ ♯ , ν ♯ ) ∈ H♯ ⇒ N♯ can be over-approximated
by the following finite set of abstract states:
unfoldcomb (loc, (σ ♯ , ν ♯ )) = {(σu♯ , guardnum (expu , ν ♯ )) | (σu♯ , expu ) ∈ unfoldshape ((β , f), σ ♯ )}
def
4. Finally, assigncomb should perform the same set of operations as described in Section 4.1 to reflect
the assignment on each unfolded heap. The assigncomb returns a finite set of elements because
of potential unfolding (and similarly for assignmem ). The soundness condition for assignmem is
therefore as follows.
Condition 5 (Soundness of assignmem ). Let (E, σ ) ∈ γM (m♯ ). Then,
(E, σ [LJlocK(E, σ ) ← EJexpK(E, σ )]) ∈
[
{γM (m♯u ) | m♯u ∈ assignmem (loc, exp, m♯ )} .
A very similar soundness condition applies to assigncomb .
Figure 16 shows the resulting abstract state for the assignment after unfolding and mutation on the
right. In certain cases, the unfolding process may have to be performed multiple times due to repeated
failures of calling eval[l]shape and eval[e]shape as shown in Chang and Rival [7]. This behavior is expected,
as unfolding may fail to materialize the correct region, and thus, termination should be enforced with a
bound on the number of unfolding steps.
Modular Construction of Shape-Numeric Analyzers
178
x
α0
x
α2
list
∧ α2 6= 0x0
assume(x -> next 6= 0x0)
α0
α2
next
d
α3
list
∧ α2 =
6 0x0
∧ α3 =
6 0x0
α4
Figure 17: Applying the condition test guardmem to an example that affects a summarized region α2 · list.
4.3 Other transfer functions
Unfolding is also the basis for most other transfer functions. Once the points-to edges in question are
materialized, their definition is straightforward as it was for assignment (cf., Section 4.1).
• Condition test. The abstract domain M♯ should define an operator guardmem that takes an expression (of Boolean type) and an abstract value and then returns an abstract value that has taken into
account the effect of the guard expression. Just like with assignment, this function may need to
perform an unfolding and thus returns in general a finite set of abstract states.
Condition 6 (Soundness of guardmem ). Let m ∈ γM (m♯ ). Then,
If JexpK(m) = true , then m ∈
[
{γH (σu♯ ) | σu♯ ∈ γM (guardmem (exp, m♯ ))} .
It applies the transfer function assignnum provided by N♯ satisfying a similar soundness condition,
which is fairly standard (e.g., the A PRON library provides such a function).
• Memory allocation. Transfer function allocmem accounts for the allocation of a fresh memory
block, and the assignment of the address of this block to a given location. Given abstract precondition σ ♯ , the abstract allocation function allocmem (loc, [f1 , . . . , fn ], σ ♯ ) returns a sound abstract
post-condition for the statement loc = malloc({f1 , . . . , fn }).
• Memory deallocation. Similarly, transfer function freemem accounts for freeing the block pointed
to by an instruction such as free. It takes as argument a location pointing to the block being
freed, a list of fields, and the abstract pre-condition. It may also need to perform unfolding to
materialize the location. It calls freecomb in the H♯ ⇒ N♯ level, which then materializes pointsto edges corresponding to the block to remove and deletes them from the graph using function
delete[e] shape defined by delete[e] shape (α , f, α · f 7→ β ∗ σ0♯ ) = σ0♯ . After removing these edges, some
symbolic nodes may become unreachable in the graph and should be removed using delete[n]shape
and delete[n]num .
The analysis of a more full featured programming language would require additional classical transfer
functions, such as support for variable creation and deletion, though this can be supported completely at
the memory abstract domain M♯ layer with the abstract environment E ♯ .
As an example of a condition test, consider applying guardmem in Figure 17. In the same way as
for the example assignment of Figure 16, the first attempt to compute guardcomb (α2 -> next 6= 0x0, σ ♯ )
fails, as there is no points-to edge labeled with next starting from node α2 . Thus guardcomb must first
call unfoldcomb . The unfolding returns a pair of abstract elements, yet the one corresponding to the
case where the list is empty does not need to be considered any further due to the numerical constraint
α2 6= 0x0. Therefore, only the second abstract elements remains, which corresponds to a list with the
first element materialized. At this stage, expression α2 -> next can be evaluated. Finally, the condition
test is reflected by applying guardnum in the numerical abstract domain N♯ .
B.-Y. E. Chang and X. Rival
179
m♯1 : (E1♯ , (σ1♯ , ν1♯ ))
m♯0 : (E0♯ , (σ0♯ , ν0♯ ))
x
α0
α2
next
d
y
α1
α3
α4
next
α5
d
α6
x
list
α7
∧ α3 ≤ α5
∧ α5 ≤ α7
α′0
α′2
⊑
next
d
y
α′1
α′3
α′4
list
α′5
∧ α3′ ≤ α5′
Figure 18: An abstract inclusion that holds and shows the need for a node relation Φ. In both abstract
heaps, variable x points to a list and y points to a number. On the left, the abstract heap describes a list
with at least two elements, while on the right, it describes one with at least one element. The number
pointed to by y is less than or equal to the data field d of the first element in both abstract heaps. The
data field of the first element is less than or equal to the data field of the second in the left abstract heap.
4.4 Abstract comparison
Abstract interpreters make use of inclusion testing operations in many situations, such as checking that
an abstract post-fixed point has been reached in a loop invariant computation or that some, for example,
user-supplied post-condition can be verified with the analysis results. As inclusion is often not decidable,
the comparison function is not required to be complete but should meet a soundness condition:
Condition 7 (Soundness of comparemem ). If comparemem (m♯0 , m♯1 ) = true, then γM (m♯0 ) ⊆ γM (m♯1 )
The implementation of such an operator is complicated by the fact that the underlying abstract heaps
may have distinct sets of symbolic nodes. This issue is a manifestation of the the cofibered abstract
domain construction (Section 3.2). The concretizations of all abstract domains below H♯ ⇒ N♯ make
use of valuations, and thus the inclusion checking operator needs to account for a relation between the
symbolic nodes of the graphs. This relation between nodes in two graphs Φ is computed step-by-step
during the course of the inclusion checking.
The example in Figure 18 illustrates these difficulties. It is quite intuitive that any state in the concretization of m♯0 is also in the concretization of m♯1 . To see the role of the node relation Φ, let us consider
concretizations of m♯0 and m♯1 . Clearly, if concrete state (E, σ ) is in the concretization of m♯0 and in the
concretization of m♯1 , then node α0 in m♯0 and α0′ denotes the address of x. Thus α0 and α0′ denote the
same value, that is, valuations used as part of the concretization should map those two nodes to the same
value. The Φ should relate these two nodes akin to a unification substitution. Similarly, α2 and α2′ both
denote the value stored in variable x, thus should be related in Φ. On the other hand, node α6 of abstract
state m♯0 has no counterpart in m♯1 —it corresponds to a null or non-null address in the region summarized
by the inductive edge.
We notice Φ can be viewed as a map from nodes in m♯1 to nodes of m♯0 and in this example, defined
by Φ(αi′ ) = αi for 0 ≤ i ≤ 5. Also, we notice that mapping Φ can be derived step-by-step, starting from
the abstract environments. Thus, compareshape and comparecomb each take as a parameter a set of pairs of
symbolic nodes that should be related in Φ. We call this initial set the roots, as they are used as a starting
point in the computation of Φ.
We can now describe the steps of computing comparemem (m♯0 : (E0♯ , (σ0♯ , ν0♯ )), m♯1 : (E1♯ , (σ1♯ , ν1♯ ))):
1. First, an initial node mapping Φ : V♯ [σ1♯ ] → V♯ [σ0♯ ] is derived from the abstract environments:
def
Φ = E0♯ ◦(E1♯ )−1 . This definition states that the addresses of the program variables in m♯1 correspond
to the respective addresses of the program variables in m♯0 . It is well-defined, as two distinct
Modular Construction of Shape-Numeric Analyzers
180
variables cannot be allocated at the same physical address.
2. Then, it calls comparecomb (Φ, (σ0♯ , ν0♯ ), (σ1♯ , ν1♯ )) that forwards to a call of compareshape (Φ, σ0♯ , σ1♯ ).
3. The abstract heap comparison function compareshape attempts to match σ0♯ and σ1♯ region-by-region
using a set of local rules:
♯
♯
• (Decomposition) Suppose σ0♯ and σ1♯ can be decomposed as σ0♯ = σ0,0
and σ1♯ =
∗ σ0,1
♯
♯
σ1,0 ∗ σ1,1 . And if the corresponding sub-regions can be shown to satisfy the inclusions
♯
♯
, σ1,0
) = (true, Φ′ ) and
compareshape (Φ, σ0,0
♯
♯
compareshape (Φ′ , σ0,1
, σ1,0
) = (true, Φ′′ ) ,
then the overall inclusion holds—compare shape (Φ, σ0♯ , σ1♯ ) returns (true, Φ′′ );
♯
♯
, σ1♯ = α1 · f 7→ β1 ∗ σ1,r
and Φ(α1 ) = α0 , then
• (Points-to edges) If σ0♯ = α0 · f 7→ β0 ∗ σ0,r
we can conclude inclusion holds locally and extend Φ with Φ(β1 ) = β0 ;
♯
♯
such that compareshape (Φ, σ0♯ , σ1,u
)=
• (Unfolding) If there is an unfolding of σ1♯ called σ1,u
(true, Φ′ ), then compareshape (Φ, σ0♯ , σ1♯ ) = (true, Φ′ ).
4. When compareshape (Φ, σ0♯ , σ1♯ ) succeeds and returns (true, Φ′ ), it means the inclusion holds with
respect to the shape. We, however, still need to check for inclusion with respect to the numeric properties. Recall that the base numeric domain elements ν0♯ ∈ Dnum hV♯ [σ0♯ ]i and ν1♯ ∈
Dnum hV♯ [σ1♯ ]i have incomparable sets of symbolic variables. An inclusion check in the base numeric domain can only be performed after renaming symbolic names so that they are consistent.
The node mapping Φ′ computed by the above is precisely the renaming that is needed. Thus, the
last step to perform to decide inclusion is to compute comparenum (ν0♯ , renamenum (Φ′ , ν1♯ )) and return it as a result for comparecomb (Φ, (σ0♯ , ν0♯ ), (σ1♯ , ν1♯ )). Note that function renamenum should be
sound in the following sense:
∀ν ♯ ∈ Dnum hV i, ∀ν ∈ γnum hV i(ν ♯ ), (ν ◦ Φ) ∈ γnum hΦ(V )i(renamenum (Φ, ν ♯ ))
where Φ(V ) is the set of symbolic variables obtained by applying Φ to set V .
5. If any of the above steps fail, comparemem returns false.
To summarize, the soundness conditions of the inclusion tests for the lower-level domains on which
comparemem relies are as follows:
Condition 8 (Soundness of inclusion tests).
1. If comparenum (ν0♯ , ν1♯ ) = true , then γnum hV i(ν0♯ ) ⊆ γnum hV i(ν1♯ ) .
2. If compareshape (Φ, σ0♯ , σ1♯ ) = (true, Φ′ ) , then (σ , ν ◦ Φ′ ) ∈ γH (σ1♯ ) for all (σ , ν ) ∈ γH (σ0♯ ) .
3. If comparecomb (Φ, (σ0♯ , ν0♯ ), (σ1♯ , ν1♯ )) = (true, Φ′ ) , then (σ , ν ◦ Φ′ ) ∈ γH♯ ⇒N♯ (σ1♯ , ν1♯ )
for all (σ , ν ) ∈ γH (σ0♯ , ν0♯ ) .
Returning to the example in Figure 18, after starting with Φ = [α0′ 7→ α0 , α1′ 7→ α1 ], the compareshape
operation consumes the points-to edges one-by-one extending Φ incrementally, unfolding the inductive
edges in the right argument before concluding that inclusion holds in the shape domain. With the final
mapping Φ′ (αi′ ) = αi for all i, the numeric inclusion simply needs to check that comparenum (α3 ≤ α5 ∧
α5 ≤ α7 , renamenum (Φ′ , α3′ ≤ α5′ )) = comparenum (α3 ≤ α5 ∧ α5 ≤ α7 , α3 ≤ α5 ) = true.
4.5 Join and widening
As is standard, the joinmem operation should satisfy the following:
B.-Y. E. Chang and X. Rival
181
m♯1 : (E1♯ , (σ1♯ , ν1♯ ))
m♯0 : (E0♯ , (σ0♯ , ν0♯ ))
x
α0
α2
next
d
y
α1
α4
list
α5
⊔
x
α′0
y
α′1
α′3
α3
list
α′2
♯
m♯⊔ : (E⊔♯ , (σ⊔
, ν⊔♯ ))
list
x
α′′
0
α′′
2
y
α′′
1
α′′
3
Z=⇒
next
d
α′4
list
list
list
α′5
Figure 19: An abstract join showing the need for different sets of symbolic variables for each of the inputs
and the result. The inputs are the two possible abstract heaps where a possibly-empty and a non-empty
list are pointed to by two non-aliased program variables x and y, so the most precise over-approximation
is the abstract heap where both x and y point to two possibly-empty lists.
Condition 9 (Soundness of joinmem ). For all m♯0 and m♯1 , γM (m♯0 ) ∪ γM (m♯1 ) ⊆ γM (joinmem (m♯0 , m♯1 )).
Like the comparison operator, the join operator takes two abstract heaps that have distinct sets of
symbolic variables as input. Additionally, it generates a new abstract heap, which requires another set of
symbolic variables, as it may not be possible to use the same set as either input. The example shown in
Figure 19 illustrates this situation. In left input m♯0 , variable x points to a non-empty list and y points to a
possibly empty list, whereas right input m♯1 describes the opposite. The most precise over-approximation
of m♯0 and m♯1 corresponds to the case where both x and y point to lists of any length (as shown on
the right side of the figure). These three elements all have distinct sets of nodes (that cannot be put
in a bijection). Thus, the join algorithm uses a slightly different notion of symbolic node mapping Ψ
that binds three-tuples of nodes consisting of one node from each parameter and one node in the output
abstract heap. Conceptually, the output abstract heap is a kind of product construction, so it is composed
of new symbolic variables corresponding to pairs of nodes with one from each input.
Overall, the join algorithm proceeds in a similar way as the inclusion test: the abstract heap join
produces a mapping relating symbolic variables along with a new abstract heap. This mapping is then
used to rename symbolic variables in the base numeric domain elements consistently to then apply the
join in the base domain. Similar to the inclusion test, an initial mapping Ψ is constructed using the
abstract environment at the M♯ level and then extended step-by-step at the H♯ level. For instance, in
Figure 19, the initial mapping is {(α0 , α0′ , α0′′ ), (α1 , α1′ , α1′′ )}, and then pairs (α2 , α2′ , α2′′ ) and (α3 , α3′ , α3′′ )
are added by joinshape . Note that nodes α3 , α4 , α3′ , α4′ have no counterpart in the result.
The local rules abstract heap join rules used in joinshape belong to two main categories:
• (Bijection) When two fragments of each input are isomorphic modulo Ψ, they can be joined into
another such fragment. In the example, the points-to edges α0 7→ α2 and α0′ 7→ α2′ can both be
over-approximated by α0′′ 7→ α2′′ . Applying this rule adds the triple (α2 , α2′ , α2′′ ) to the mapping Ψ.
• (Weakening) When a heap fragment can be shown to be included in a more simple, summary
fragment (in terms of their concretizations), we can over-approximate the original fragment with
the summary. For instance, fragment α2 · next 7→ α3 ∗ α2 · d 7→ α4 ∗ α3 · list can be shown to be
included in α2 · list. The other input can be an effective means for directing the choice of possible
summary fragments [7, 8].
The widening operator widenmem can be defined similarly to joinmem . If the heap join rules enforce
termination (i.e., joinshape can be used as a widening) and joinnum is replaced with a widening operator
widennum , the cofibered domain definition guarantees the resulting operator enforces termination [38].
Modular Construction of Shape-Numeric Analyzers
182
4.6 Disjunctive abstract domain interface
Recall from Sections 3.3 and 4.2 that unfolding returns
partition∨ : C × Pfin(M♯∨ )
−→ M♯∨
a finite set of abstract elements interpreted disjunctively
♯
collapse∨ : C × M∨
−→ M♯∨
and thus justifies the need for a disjunctive abstracassign∨ : C × LX × EX × M♯∨ −→ M♯∨
tion layer—independent of other possible reasons like
−→ M♯∨
guard∨ : C × EX × M♯∨
a desire for path-sensitivity. In this subsection, we dealloc∨ : C × LX × N × M♯∨ −→ M♯∨
scribe the interface for a disjunctive abstraction layer M♯∨
free∨ : C × LX × N × M♯∨ −→ M♯∨
shown in Figure 20 that sits above the memory layer M♯ .
compare∨ : M♯∨ × M♯∨
−→ B
The following discussion completes the picture of the ab♯
♯
−→ M♯∨
join∨ : M∨ × M∨
stract domain interfaces (cf., Figure 11). There are two
♯
♯
widen∨ : M∨ × M∨
−→ M♯∨
main differences in the interface as compared to the one
for M♯ . First, the disjunctive abstract domain should pro- Figure 20: Disjunctive abstraction interface.
vide two additional operations partition∨ and collapse∨
that create and collapse partitions, respectively. A partition represents a disjunctive set of base domain
elements. Second, the transfer functions take an additional context information parameter c ∈ C that can
be used in M♯∨ to tag each disjunct with how it arose in the course of the abstract interpretation.
Condition 10 (Soundness of partition ∨ and collapse∨ ). Let s♯ : M♯∨ and S♯ : Pfin (M♯∨ ).
S
{γ∨ (s♯ ) | s♯ ∈ S♯ } ⊆ γ∨ (partition ∨ (c, S♯ ))
γ∨ (s♯ ) ⊆ γ∨ (collapse∨ (c, s♯ ))
Note that contexts play no role in the concretization, but operations can use them, for example, to decide
which disjuncts to merge using joinmem and which disjuncts to preserve.
Transfer functions assign∨ , guard∨ , alloc∨ , and free∨ all follow the same structure. They first call the
underlying operation on the memory abstract domain M♯ and then apply the partition ∨ partition on the
output. For instance, assign∨ is defined as follows while satisfying the expected soundness condition:
def
assign∨ (c, loc, exp) = partition ∨ (c, {assignmem (loc, exp, m♯ ) | m♯ ∈ s♯ })
(E, σ [LJlocK(E, σ ) ← EJexpK(E, σ )]) ∈ γ∨ (assign∨ (c, loc, exp, s♯ ))
(definition)
(soundness)
Inclusion (compare∨ ), join, and widening operations should satisfy the usual soundness conditions. The
collapse ∨ operator may be used to avoid generating too many disjuncts (and termination of the analysis).
5
A compositional abstract interpreter
In this section, we assemble an abstract interpreter for the language defined in Section 2 using the abstraction set up in Section 3 and the interface of abstract operations described in Section 4.
The abstract semantics of a program p is a function JpK♯ : M♯∨ → M♯∨ , which takes an abstract precondition as input and produces an abstract post-condition as output. Based on an abstract interpretation
of the denotational semantics of programs [33, 34], we can define the abstract semantics by induction
over the syntax of programs as shown in Figure 21 in a completely standard manner. We let C [. . .] stand
for computing some context information based on, for example, the control state ℓ and/or the branch
taken. This context information may be used, for instance, by the disjunctive domain M♯∨ to guide trace
partitioning [30]. The abstract transitions for sequencing, assignment, dynamic memory allocation, and
deallocation are straightforward with the latter three calling the corresponding transfer function in the
top-layer abstract domain M♯∨ . For if, the pre-condition is first constrained by the guard condition via
B.-Y. E. Chang and X. Rival
183
def
def
Jp0 ; p1 K♯ (s♯ ) = Jp1 K♯ ◦ Jp0 K♯ (s♯ )
Jℓ : loc = malloc(n)K♯ (s♯ ) = alloc∨ (C [ℓ], loc, n, s♯ )
def
def
♯
♯
♯
Jℓ : loc = expK (s ) = assign∨ (C [ℓ], loc, exp, s )
Jℓ : free(loc)K♯ (s♯ ) = free∨ (C [ℓ], loc, n, s♯ )
def
Jℓ : if (exp) pt else pf K♯ (s♯ ) = join∨ (Jpt K♯ (guard∨ (C [ℓ, true], exp, s♯ )),
Jpf K♯ (guard∨ (C [ℓ, false], exp = false, s♯ )))
Jℓ : while (exp) pK♯ (s♯ ) = guard∨ (C [ℓ, false], exp = false, lfp♯s♯ F ♯ )
where F ♯ : M♯∨ −→ M♯∨
7−→ JpK♯ (guard∨ (C [ℓ, true], exp, s♯0 ))
s♯0
def
Figure 21: A denotational-style abstract interpreter for the programming language defined in Section 2.2.
guard∨ to interpret the two branches and then the resulting states are joined via join∨ . For while, we
write lfp♯ for an abstract post–fixed-point operator. The lfp♯ operator relies on widen∨ to terminate and
on compare∨ to verify the stability of the abstract post-fixed point. It may also use join∨ to increase the
level of precision when computing the first iterations. We omit a full definition of lfp♯ as there are many
well-known ways to obtain such an operator. The most simple one consists of applying only widen∨
until stabilization can be shown by compare∨ . We simply state its soundness condition:
Condition 11 (Soundness of lfp♯ ). For all concrete transformers F : P(S) → P(S) monotone, all abstract transformers F ♯ : M♯∨ → M♯∨ , and all abstract states s♯ ∈ M♯∨ ,
if F ◦ γ∨ ⊆ γ∨ ◦ F ♯ , then lfpγ∨ (s♯ ) F ⊆ γ∨ (lfp♯s♯ F ♯ ) .
We write lfpS for the least post–fixed point that is at least S and similarly for lfp♯s♯ . Finally, the static
analysis is sound in the following sense:
Theorem 1 (Soundness of the analysis). Let p be a program, and let s♯ ∈ M♯∨ be an abstract precondition. Then, the result of the analysis is sound:
∀s ∈ γ∨ (s♯ ), JpKd (s) ⊆ γ∨ (JpK♯ (s♯ )) .
Soundness can be proven by induction over the syntax of programs and by composing the local soundness
conditions of all abstract operators.
Related work and discussion. An advantage of this iteration strategy, is that it leads to an intuitive
order of application of the abstract equations corresponding to the program [10], eliminating complex
iteration strategies [19]. It also simplifies the choice of widening points [4], as it applies widening
naturally, at loop heads, though it also allows one to make different choices in strategy by, for example,
modifying lfp♯ to unroll loop iterations [3].
6
Conclusion
We have presented a modular construction of a static analysis that is able to reason both about the
shape of data structures and their numeric contents simultaneously. Our construction is parametric in
the desired numeric abstraction, as well as the shape abstraction, making it possible to continuously
substitute improvements for each component or with variants targeted at different classes of programs
or even different programming languages. The main advantage of a modular construction is that it
184
Modular Construction of Shape-Numeric Analyzers
allows one to design, prove, and implement each component of the analysis independently. Modular
construction is a cornerstone of quality software engineering, and our experience has been that this nice
property becomes even more important when dealing with the complexity of creating a static analysis
that simultaneously reasons about shape and numeric properties.
Acknowledgments We are inspired by Dave Schmidt’s continual effort in formalizing the foundations
of static analysis, abstract interpretation, and the semantics of programming languages. We would also
wish to thank Dave for being a pillar in the static analysis community and supporting the research community with often thankless, behind-the-scenes work.
This work has been supported in part by the European Research Council under the FP7 grant agreement 278673, Project MemCAD and the United States National Science Foundation under grant CCF1055066.
References
[1] T. Ball, R. Majumdar, T. D. Millstein & S. K. Rajamani (2001): Automatic Predicate Abstraction of C
Programs. In: PLDI’01, pp. 203–213, doi:10.1145/378795.378846.
[2] J. Berdine, C. Calcagno, B. Cook, D. Distefano, P. O’Hearn, T. Wies & H. Yang (2007): Shape Analysis for
Composite Data Structures. In: CAV’07, pp. 178–192, doi:10.1007/978-3-540-73368-3_22.
[3] B. Blanchet, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, D. Monniaux & X. Rival (2003): A
static analyzer for large safety-critical software. In: PLDI’03, pp. 196–207, doi:10.1145/781131.781153.
[4] F. Bourdoncle (1992): Abstract Interpretation by Dynamic Partitioning. Journal of Functional Programming
2(4), pp. 407–423, doi:10.1017/S0956796800000496.
[5] B.-Y. E. Chang (2008): End-User Program Analysis. Ph.D. thesis, University of California, Berkeley.
[6] B.-Y. E. Chang & R. Leino (2005): Abstract Interpretation with Alien Expressions and Heap Structures. In:
VMCAI’05, pp. 147–163, doi:10.1007/978-3-540-30579-8_11.
[7] B.-Y. E. Chang & X. Rival (2008): Relational inductive shape analysis. In: POPL’08, pp. 247–260,
doi:10.1145/1328438.1328469.
[8] B.-Y. E. Chang, X. Rival & G. Necula (2007): Shape Analysis with Structural Invariant Checkers. In:
SAS’07, pp. 384–401, doi:10.1007/978-3-540-74061-2_24.
[9] D. R. Chase, M. Wegman & F. K. Zadeck (1990): Analysis of Pointers and Structures. In: PLDI’90, pp.
296–310, doi:10.1145/93542.93585.
[10] P. Cousot & R. Cousot (1977): Abstract Interpretation: A Unified Lattice Model for Static Analysis of Programs by Construction or Approximation of Fixpoints.
In: POPL’77, pp. 238–252,
doi:10.1145/512950.512973.
[11] P. Cousot & R. Cousot (1979): Systematic Design of Program Analysis Frameworks. In: POPL’79, pp.
269–282, doi:10.1145/567752.567778.
[12] P. Cousot & N. Halbwachs (1978): Automatic discovery of linear restraints among variables of a program.
In: POPL’78, pp. 84–97, doi:10.1145/512760.512770.
[13] I. Dillig, T. Dillig & A. Aiken (2011): Precise reasoning for programs using containers. In: POPL’11, pp.
187–200, doi:10.1145/1926385.1926407.
[14] D. Distefano, P. O’Hearn & H. Yang (2006): A Local Shape Analysis Based on Separation Logic. In:
TACAS’06, pp. 287–302, doi:10.1007/11691372_19.
[15] D. Gopan, T. W. Reps & M. Sagiv (2005): A framework for numeric analysis of array operations. In:
POPL’05, pp. 338–350, doi:10.1145/1040305.1040333.
B.-Y. E. Chang and X. Rival
185
[16] S. Gulwani, B. McCloskey & A. Tiwari (2008): Lifting abstract interpreters to quantified logical domains.
In: POPL’08, pp. 235–246, doi:10.1145/1328438.1328468.
[17] N. Halbwachs & M. Péron (2008): Discovering properties about arrays in simple programs. In: PLDI’08,
pp. 339–348, doi:10.1145/1375581.1375623.
[18] T. A. Henzinger, R. Jhala, R. Majumdar & G. Sutre (2002): Lazy abstraction. In: POPL’02, pp. 58–70,
doi:10.1145/503272.503279.
[19] S. Horwitz, A. J. Demers & T. Teitelbaum (1987): An Efficient General Iterative Algorithm for Dataflow
Analysis. Acta Informatica 24(6), pp. 679–694, doi:10.1007/BF00282621. Springer.
[20] B. Jeannet & A. Miné (2009): Apron: A Library of Numerical Abstract Domains for Static Analysis. In:
CAV’09, pp. 661–667, doi:10.1007/978-3-642-02658-4_52.
[21] N. Jones & S. Muchnick (1981): Flow Analysis and Optimization of LISP-like Structures. In: Program Flow
Analysis: Theory and Applications, chapter 4, Prentice-Hall, pp. 102–131.
[22] V. Laviron, B.-Y. E. Chang & X. Rival (2010): Separating Shape Graphs. In: ESOP’10, pp. 387–406,
doi:10.1007/978-3-642-11957-6_21.
[23] S. Magill, J. Berdine, E. Clarke & B. Cook (2007): Arithmetic Strengthening for Shape Analysis. In: SAS’07,
pp. 419–436, doi:10.1007/978-3-540-74061-2_26.
[24] S. Magill, M.-H. Tsai, P. Lee & Y.-K. Tsay (2010): Automatic numeric abstractions for heap-manipulating
programs. In: POPL’10, pp. 211–222, doi:10.1145/1706299.1706326.
[25] B. McCloskey, T. Reps & M. Sagiv (2010): Statically Inferring Complex Heap, Array, and Numeric Invariants. In: SAS’10, pp. 71–99, doi:10.1007/978-3-642-15769-1_6.
[26] A. Milanova, A. Rountev & B. G. Ryder (2005): Parameterized object sensitivity for points-to analysis for
Java. TOSEM 14(1), pp. 1–41, doi:10.1145/1044834.1044835.
[27] A. Miné (2006): The octagon abstract domain. HOSC 19(1), pp. 31–100, doi:10.1007/s10990-006-8609-1.
[28] J. Reynolds (2002): Separation Logic: A Logic for Shared Mutable Data Structures. In: LICS’02, pp. 55–74,
doi:10.1109/LICS.2002.1029817.
[29] X. Rival & B.-Y. E. Chang (2011): Calling context abstraction with shapes. In: POPL’11, pp. 173–186,
doi:10.1145/1926385.1926406.
[30] X. Rival & L. Mauborgne (2007): The trace partitioning abstract domain. ACM TOPLAS 29(5), pp. 26–69,
doi:10.1145/1275497.1275501.
[31] M. Sagiv, T. Reps & R. Wilhelm (2002): Parametric shape analysis via 3-valued logic. ACM TOPLAS
24(3), pp. 217–298, doi:10.1145/514188.514190.
[32] M. Sagiv, T. W. Reps & R. Wilhelm (1998): Solving Shape-Analysis Problems in Languages with Destructive
Updating. ACM TOPLAS 20(1), pp. 1–50, doi:10.1145/271510.271517.
[33] D. A. Schmidt (1986): Denotational semantics: a methodology for language development. William C. Brown
Publishers, Dubuque, IA, USA.
[34] D. A. Schmidt (2009): Abstract Interpretation From a Denotational-semantics Perspective. ENTCS 249, pp.
19–37, doi:10.1016/j.entcs.2009.07.082.
[35] M. Sharir & A. Pnueli (1981): Two approaches to interprocedural data flow analysis. In: Program Flow
Analysis: Theory and Applications, chapter 7, Prentice-Hall, pp. 189–233.
[36] P. Sotin & X. Rival (2012): Hierarchical Shape Abstraction of Dynamic Structures in Static Blocks. In:
APLAS’12, pp. 131–147, doi:10.1007/978-3-642-35182-2_10.
[37] A. Toubhans, B.-Y. E. Chang & X. Rival (2013): Reduced Product Combination of Abstract Domains for
Shapes. In: VMCAI’13, pp. 375–395, doi:10.1007/978-3-642-35873-9_23.
[38] A. Venet (1996): Abstract Cofibered Domains: Application to the Alias Analysis of Untyped Programs. In:
SAS’96, pp. 366–382, doi:10.1007/3-540-61739-6_53.
| 6cs.PL
|
Measurement-based adaptation protocol with quantum reinforcement learning
F. Albarrán-Arriagada,1, ∗ J. C. Retamal,1, 2 E. Solano,3, 4, 5 and L. Lamata3
arXiv:1803.05340v1 [quant-ph] 14 Mar 2018
1
Departamento de Fı́sica, Universidad de Santiago de Chile (USACH), Avenida Ecuador 3493, 9170124, Santiago, Chile
2
Center for the Development of Nanoscience and Nanotechnology 9170124, Estación Central, Santiago, Chile
3
Department of Physical Chemistry, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain
4
IKERBASQUE, Basque Foundation for Science, Maria Diaz de Haro 3, 48013 Bilbao, Spain
5
Department of Physics, Shanghai University, 200444 Shanghai, China
(Dated: March 15, 2018)
Machine learning employs dynamical algorithms that mimic the human capacity to learn, where the reinforcement learning ones are among the most similar to humans in this respect. On the other hand, adaptability is an
essential aspect to perform any task efficiently in a changing environment, and it is fundamental for many purposes, such as natural selection. Here, we propose an algorithm based on successive measurements to adapt one
quantum state to a reference unknown state, in the sense of achieving maximum overlap. The protocol naturally
provides many identical copies of the reference state, such that in each measurement iteration more information
about it is obtained. In our protocol, we consider a system composed of three parts, the “environment” system, which provides the reference state copies; the register, which is an auxiliary subsystem that interacts with
the environment to acquire information from it; and the agent, which corresponds to the quantum state that is
adapted by digital feedback with input corresponding to the outcome of the measurements on the register. With
this proposal we can achieve an average fidelity between the environment and the agent of more than 90% with
less than 30 iterations of the protocol. In addition, we extend the formalism to d-dimensional states, reaching
an average fidelity of around 80% in less than 400 iterations for d = 11, for a variety of genuinely quantum as
well as semiclassical states. This work paves the way for the development of quantum reinforcement learning
protocols using quantum data, and the future deployment of semi-autonomous quantum systems.
I.
INTRODUCTION
Machine learning (ML) is an area of artificial intelligence that focuses on the implementation of learning algorithms, and which has undergone great development in recent
years [1–3]. ML can be classified into two broad groups,
namely, learning by means of big data and through interactions. For the first group we have two classes, supervised
learning, which uses previously classified data to train the
learning program, inferring the function of relationship to
classify new data. This is the case, e.g., of pattern recognition problems [4–7]. The other class is unsupervised learning,
which does not require training data, but this paradigm uses
the big data distribution to obtain an optimal way to classify
it using specific characteristics. An example is the clustering
problem [8, 9].
For the second group, learning from interactions, we have
the case of reinforcement learning (RL) [10, 11]. RL is the
more similar paradigm to the human learning process. Its general framework is as follows: we define two basic systems,
an agent A and an environment E, while often it is useful to
define a register R as an auxiliary system. The concept consists of A inferring information by direct interaction with E,
or indirectly, using as a mediator the system R. With the obtained information, A makes a decision to perform a certain
task. If the result of this task is good, then the agent receives
a reward, otherwise a punishment. In addition, the RL algorithms can be divided into three basic parts, the policy, the
reward function (RF) and the value function (VF). The policy
can be subdivided into three stages: first, interaction with the
∗
F. Albarrán-Arriagada
[email protected]
environment. In this stage, the way in which A or R interacts
with E is specified. Second, information extraction, which indicates how A obtains information from E. Finally, action,
where A takes the decision of what to do with the information
of the previous step. RF refers to the criterion to award the
reward or punishment A in each iteration. And VF evaluates
the utility of A referred to the given task. An example of RL
consists in artificial players for go or chess [12, 13].
Other essential aspect of the RL protocols is the
exploitation-exploration relation. Exploitation refers to the
ability to make good decisions, while exploration is the possibility of making different decisions. For example, if we want
to select a gym to do sports, the exploitation is given by the
quality of the gym we tested, while the exploration is the size
of the search area in which we will choose a new gym to test.
In the RL paradigm, a good exploitation-exploration relation
can guarantee the convergence of the learning process, and its
optimization depends on each algorithm.
(a)
(b)
FIG. 1. (a) Bloch representation of the environment state at initial
time, with |0i the state of the agent. (b) Bloch representation of the
environment in the jth iteration, where |0̄ j i is the state of the agent,
which was rotated in the previous iterations.
2
On the other hand, quantum mechanics is known to improve computational tasks [14], so a natural question is: how
are the learning algorithms modified in the quantum domain?
To answer this question the quantum machine learning field
(QML) has emerged. In recent years, QML has been a fruitful area [15–22], in which quantum algorithms have been developed [23–26], that show a possible speedup in certain situations in relation with their classical counterparts [27, 28].
However, these novel works focus mainly on learning from
classical data encoded in quantum systems, processed with a
quantum algorithm and decoded to be read by a classical machine. In this context, the speedup of quantum machine learning is often balanced with the necessary resources to encode
and decode the information, which leads to an unclear quantum supremacy. On the other hand, recent works analyze the
QML paradigm in a purely quantum way [29–31], in which
quantum systems learn quantum data.
In this article, we present a quantum machine learning algorithm based on a reinforcement learning approach, to convert the quantum state of a system (Agent), into an unknown
state encoded, via multiple identical copies, in another system
(Environment), assisted by measurements on a third system
(Register). We propose to use coherent feedback loops, conditioned to the measurements in order to perform the adaptation process without human intervention. In our numerical
calculations we obtain average fidelities of more than 90% for
qubit states after less than 40 measurements, while for qudits
the protocol achieves average fidelities of 80% using 400 iterations with d = 11 dimensions, either for genuinely quantum
or semiclassical states. This proposal can be useful in the way
to implement semi-autonomous quantum devices.
II.
THE QUANTUM ADAPTATION ALGORITHM
Our framework is as follows. We assume a known quantum system called agent (A), and many copies of an unknown
quantum state provided by a system called environment (E).
We also consider an auxiliary system called register (R) which
interacts with E. Then, we obtain information about E by
measuring R, and employ the result as an input to the reward
(RF) function. Finally, we perform a partially-random unitary
transformation on A, which depends on the output of the RF.
The idea is to improve the fidelity between A and E, without
projecting the state of A with measurements.
This protocol differs from quantum state estimation [32–
37] in the fact that we propose a semi-autonomous quantum
agent, that is, the aim is that in the future a quantum agent
will learn the state of the environment without any human intervention. Other authors have considered the inverse problem, an unknown state evolved to a known state assisted by
measurements [38], which deviate from the machine learning paradigm. Therefore, an optimal measurement is not performed in each step, but after a certain number of autonomous
iterations, the agent converges to a large fidelity with the unknown state.
In the rest of the article we use the following notation: the
subscripts A, R and E refer to each subsystem, and the super-
scripts indicate the iteration. For example, O(k)
α refers to the
operator O that acts on the subsystem α during the kth iteration. Moreover, the lack of any of these indices indicates that
we are referring to a general object in the iterations and/or in
the subsystems.
We start with the case where each subsystem is described
by a qubit state. We assume that A(R) is described by |0iA(R) ,
and E by an arbitrary state expressed in the Bloch sphere as
(1)
|EiE = cos(θ(1) /2)|0iE + e−iφ sin(θ(1) /2)|1iE [see Fig. 1 (a)].
The initial state reads
(1)
|ψ(1) i = |0iA |0iR [cos(θ(1) /2)|0iE + eiφ sin(θ(1) /2)|1iE ]. (1)
First of all, we will introduce the general elements of our reinforcement learning protocol, such as policy, the RF and the
VF. For the policy, we perform a Controlled-NOT (CNOT)
NOT
gate (U E,R
) with E as control and R as target (i.e., the interaction with the environment), in order to copy information of
E into R, obtaining
NOT (1)
|Ψ1 i =U E,R
|ψ i
h
i
(1)
=|0iA cos(θ(1) /2)|0iR |0iE + eiφ sin(θ(1) /2)|1iR |1iE . (2)
We then measure the register qubit in the basis {|0i, |1i}, with
(1)
2
2
probability p(1)
0 = cos (θ1 /2), or p1 = sin (θ1 /2), to obtain
the state |0i or |1i respectively (i.e., information extraction).
If the result is |0i, it means that we collapse E into A and
do nothing, but if the result is |1i, it means that we measure
the orthogonal component to A of E, and thus we accordingly
modify the agent. As we do not have additional information
about the environment, we perform a partially-random uniz(1) (1)
x(1) (1)
tary operator on A given by U A(1) (α(1) , β(1) ) = e−iS A α e−iS A β
(action), where α(1) and β(1) are random angles of the form
α(β)(1) = ξα(β) ∆(1) , with ξα(β) ∈ [−1/2, 1/2] a random number,
∆(1) is the range of random angles, α(β)(1) ∈ [−∆(1) /2, ∆(1) /2],
and S k(1)
= S k is the kth spin component. Now, we initialize
A
the register qubit state and employ a new copy of E, obtaining
the next initial state for the second iteration
|ψ(2) i = UA(1) |0iA |0iR |EiE = |0̄i(2)
A |0iR |EiE ,
(3)
h
i
UA(1) = m(1) U A(1) (α(1) , β(1) ) + (1 − m(1) )IA .
(4)
with
Here, m(1) = {0, 1}, is the outcome of the measurement, I is
the identity operator, and we define the new agent state as
(1)
|0̄i(2)
1 = U1 |0i1 .
Now, we define the RF to modify the exploration range of
the kth iteration ∆(k) as
∆(k) = (1 − m(k−1) )R + m(k−1) P ∆(k−1) ,
(5)
where m(k−1) is the outcome of the (k − 1)th iteration, while
R and P are the reward and punishment ratios, respectively.
Equation (5) means, that the value of ∆ is modified by R∆
for the next iteration when the previous outcome is m = 0,
and by P∆ when the outcome is m = 1. In our protocol, we
3
NOT
the gate U E,R
NOT
|Φ(k) i = U E,R
|0̄i(k)
A |0iR |ĒiE
(k)
(k)
= |0̄iA cos(θ(k) /2)|0iR |0iE + eiφ sin(θ(k) /2)|1iR |1iE , (10)
M
FIG. 2. Quantum circuit diagram for the measurement-based adaptation protocol. The box labelled with M indicates the projective
measurement process, and the red lines denote feedback loops.
choose for simplicity R = < 1 and P = 1/ > 1, such
that, every time that the state |0i is measured, the value of ∆
is reduced, and increased in the other case. Also, the fact that
R · P = 1 means that the punishment and the reward have
the same strength, or in other words, if the protocol yields the
same number of outcomes 0 and 1, the exploration range does
not change. Finally, the VF is defined as the value of ∆(n) after
all iterations. Therefore, ∆(n) → 0 if the protocol improves the
fidelity between A and E.
To illustrate the behaviour of the protocol, we consider the
kth iteration. The initial state is given by
|ψi(k) = |0̄i(k)
A |0iR |EiE ,
(6)
(k)
(k)
(k−1) (k−1)
where |0̄i(k)
UA , with U(1)
A = UA |0iA , UA = UA
A = IA ,
z( j) ( j)
x( j) ( j)
and UA( j) given by Eq. (4). Also U A( j) = e−iS A α e−iS A β ,
where we define
1
(|0̄i(Aj) h0̄| − |1̄i(Aj) h1̄|) = UA( j−1)† S z(A j−1) UA( j−1) ,
2
1
= (|0̄i(Aj) h1̄| + |1̄i(Aj) h0̄|) = UA( j−1)† S Ax( j−1) UA( j−1) , (7)
2
S z(A j) =
S Ax( j)
with (Aj) h0̄|1̄i(Aj) = 0. We can write the state of E in the Bloch
representation using |0̄ j i as a reference axis (see Fig. 1 (b)),
and apply the operator U(k)†
E , obtaining for E,
(k)†
(k)
(k)
(k)
(k)
iφ(k)
U(k)†
|Ei
=U
cos(θ
/2)|
0̄i
+
e
sin(θ
/2)|
1̄i
E
E
E
E
E
(k)
=cos(θ(k) /2)|0iE + eiφ sin(θ(k) /2)|1iE = |Ēi(k)
E . (8)
We can write the states |0̄(k) i and |1̄(k) i in terms of the initial
logical states |0i and |1i, and the unknown angles θ(k) , θ(1) , φ(k)
and φ(1) as follows
!
!
θ(1) − θ(k)
θ(1) − θ(k)
(1)
|0̄(k) i =cos
|0i + eiφ sin
|1i
2
2
!
!
θ(1) − θ(k)
θ(1) − θ(k)
(k)
−iφ(k)
i(φ(1) −φ(k) )
sin
|1̄ i =−e
|0i + e
sin
|1i.
2
2
(9)
Therefore, the operator U(k)† performs the necessary rotation
to transform |0̄(k) i → |0i and |1̄(k) i → |1i. Then, we perform
and we measure R, with probabilities p0(k) = cos2 (θ(k) /2), and
2 (k)
(k)
p(k)
= 0 and m(k) = 1,
1 = sin (θ /2), for the outcomes m
respectively. Finally, we apply the RF given by Eq. (5). We
point out that, probabilistically, when p(k)
0 → 1, ∆ → 0, and
when p(k)
→
1,
∆
→
4π.
In
terms
of
exploitation-exploration
1
relation this means that when the exploitation decreases (we
measure |1i often), we increase the exploration (we increase
the value of ∆) to increase the probability of making a beneficial change, and when the exploitation improves (we measure
|0i many times), we reduce the exploration to allow only small
changes in the following iterations. The diagram of this protocol is shown in Fig. (2).
Fig. 3 (a) shows the numerical calculation of mean fidelity
between A and E for the single-qubit case. For this computation we use 2000 random initial states with = 0.1 (blue line),
= 0.3 (red line), = 0.5 (yellow line), = 0.7 (purple line),
and = 0.9 (green line). We can see that the protocol can
reach fidelities over 90% in less than 30 iterations. Fig. 3 (b)
depicts the evolution of the exploration parameter ∆ for each
iteration for the same values of constant . We can see from
Fig. 3, that when the parameter is small, the fidelity between
A and E increases quickly (the learning speed increases), requiring less iterations to reach high fidelities, however, the
maximum value of the average fidelity (maximum learning) is
smaller than when increases. This means that small changes
in the scan parameter ∆ (large ) result in a higher but slower
learning.
III.
MULTILEVEL PROTOCOL
In this section, we extend the previous protocol to the case
where A, R, and E are described by one d-dimensional qudit
state. One of the ingredients in the qubit case is the CNOT
gate. Here, we use the extension of the CNOT gate to mulXOR
tilevel states, also known as the XOR gate [39] (Ua,b
). The
action of this gate is given by
XOR
Ua,b
| jia |kib = | jia | j
kib ,
(11)
where the index a(b) refers to control(target) state, respectively, and
denotes the difference modulo d, with d the
dimension of each subsystem. The CNOT gate has two imNOT
portant properties, namely, (i) Ua,b
is Hermitian, and (ii)
NOT
Ua,b | jia |kib = | jia |0ib if and only if j = k. These two properties are maintained in the XOR gate defined in Eq. (11).
The Policy and VF are essentially the same than in previous case, but now we consider the multiple outcomes (m( j) ∈
{0, 1, . . . , d−1}) that result of measuring R. First, we introduce
|1̄ j iA = |m( j) iA for the definition of S z(A j) and S Ax( j) in Eq. (7).
As in the previous case, we assume the initial state of A to
be |0iA , while R is initialized in |0iR . Moreover, the state of
4
1 (a)
0.9
3
0.8
2
0.7
1
0.6
0
0.85 (a)
4 (b)
100
200
0
4 (b)
3
0.55
2
1
0
100
200
0.25
0
500
0
1000
0
500
1000
FIG. 3. Performance of the measurement-based adaptation protocol
for the qubit case. (a) shows the mean fidelity for 2000 initial random
states, and (b) the value of ∆( j) in each iteration. For both figures, we
have = 0.1 (blue line), = 0.3 (red line), = 0.5 (yellow line),
= 0.7 (purple line),and = 0.9 (green line). The black dashed line
indicates an average fidelity of 90%.
FIG. 4. Performance of the measurement-based adaptation protocol
for a total random state given by Eq. (15). (a) shows the mean fidelity for 2000 initial random states, and (b) the value of ∆ j in each
iteration. For both figures, we have = 0.1 (blue line), = 0.3 (red
line), = 0.5 (yellow line), = 0.7 (purple line),and = 0.9 (green
line). The black dashed line indicates an average fidelity of 80%.
P
E is arbitrary, and expressed as |EiE = d−1
j=0 c j | jiE , where
Pd−1 2
|c
|
=
1,
and
d
is
the
dimension
of
E.
We
can rewrite E
j=0 j
in a more convenient way as
Consider the protocol for three different multilevel examples for the E state. First, we consider a total random state
with d = 11 of the form
(1)
|EiE = cos(θ(1) /2)|0̄(1) iE + eiφ sin(θ(1) /2)|0̄(1)
(12)
⊥ iE ,
P
d−1
where |0̄(1) iE = |0iE , |0̄(1)
⊥ iE = (1/N)
j=1 c j | jiE is the orthogPd−1
2
onal component to |0iE , and N = j=1 |c j |2 . Subsequently,
XOR
we perform the XOR gate U E,R
obtaining
XOR
|Φ0 i =U E,R
|0iA |0iR |EiE
h
i
(1)
= |0iA cos(θ(1) /2)|0iR |0iE + eiφ sin(θ(1) /2)|κiR,E , (13)
P
with |κiR,E = d−1
j=1 (1/N)c j |d − jiR | jiE . As in the previous
case, we measure R, but now we have multiple outcomes.
Therefore, we separated them in two groups. First, the outcome |0i with probability p(1)
= cos2 (θ1 /2). Second, out0
comes | ji with j , 0, and probability to obtain any of them
2 (1)
of p(1)
⊥ = sin (θ /2). As in the previous case, this means that
either we measure in the state of A or in the orthogonal subspace. With this information, we perform a partially-random
z(1) (1)
x(1) (1)
unitary operation on the agent UA(1) = e−iS A α e−iS A β , using the definition (7) with |1̄A i = |m(1) iA , where m(1) = j is the
outcome of the measurement. If m(1) = 0, then UA(1) = IA . The
random angles α(1) and β(1) are defined as in the qubit case.
Now, the RF changes slightly and is given by
∆( j) = δm( j−1) ,0 R + (1 − δm( j−1) ,0 )P ∆( j−1) ,
(14)
where δ j,k is the delta function. Equation (14) means that if we
measure |0i in R, the value of ∆ decreases for the next iteration, and if we measure | ji with j , 0, ∆ increases. Remember
that R = < 1 and P = 1/ > 1. As in the qubit case, the RF
is binary, since all the results | ji with j , 0 are equally nonbeneficial, so we give the same punishment to the agent. For
this reason we use the same policy than in the qubit protocol
for the case of multiple levels. As in the case of a single qubit
state, the parameter plays a fundamental role in the learning
process by handling the speed of learning and the maximum
learning as we will understand in what follows.
|EiE =
10
1 X
ck |kiE ,
N k=0
ck = a + ib,
(15)
where a, b ∈ [0, 1) are random numbers, and N is a normalization factor. Figure 4 shows the numerical calculations for
this case, where (a) gives the average fidelity for 2000 initial
states given by Eq. (15), and (b) the evolution of ∆ in each
iteration. It also shows how this exploration parameter is reduced when the fidelity between E and A grows (increasing
the exploitation). We can see from Fig. 4 (a) that the protocol
can reach mean fidelities of 80% with about 400 iterations, or,
equivalently, the protocol increases the mean fidelity between
A and E in about 0.5 using 400 iterations.
Second, consider the protocol for the coherent state defined
by
|αi = e−|α| /2
2
0.9
(16)
4
0.85
3
0.8
2
0.75
1
0.7
0.65
∞
X
αn
√ |ni
n!
n=0
0
100
200
300
0
0
100
200
300
FIG. 5. Performance of the measurement-based adaptation protocol
for a coherent state of the form in Eq. (16). (a) shows the mean
fidelity for 2000 random pairs {a, b} ∈ [0, 1), where α = a + ib; and
(b) the value of ∆ j in each iteration. For both figures, we have = 0.1
(blue line), = 0.3 (red line), = 0.5 (yellow line), = 0.7 (purple
line), and = 0.9 (green line). The black dashed line indicates an
average fidelity of 85%.
5
0.95 (a)
1 (b)
0.9
0.9
0.85
0.8
0.8
0.7
0.75
0.7
0
250
500
0.6
0
50
100
FIG. 6. Performance of the measurement-based adaptation protocol
for genuinely quantum states. (a) shows the mean fidelity for 2000
cat states of the form in Eq. (17), and (b) the mean fidelity for 2000
√
repetitions of the protocol using the superposition (|0iE + |10iE )/ 2
for the environment. In both figures, we use = 0.1 (blue line),
= 0.3 (red line), c = 0.5 (yellow line), = 0.7 (purple line), and
= 0.9 (green line). The black dashed line indicates an average
fidelity of 85%.
for this case we use α = a + ib, with a and b positive real
random numbers smaller than 1. As |α|2 ≤ 2, we can truncate the sum (16) to n = 10, since the probabilities
√ to 2ob−|α|2 /2 10
10!| ≤
tain√|ni with
n
>
10
are
bounded
by
|e
α
/
√
e−1 210 / 10! u 0.0062. Figure 5 (a) shows the fidelity between A and E for each iteration, reaching values of 85% in
less than 100 iterations. Figure 5 (b) depicts the value of ∆ in
this process. We can also observe that the exploration is reduced when A approaches E (increasing the exploitation), as
in the previous case.
Finally, we consider two quantum states, a cat state of the
form
r
1
|EiE =
|αi + | − αi
(17)
Nα
where |αi is given by Eq. (16) and Nα is a normalization
factor. Additionally, we study the superposition
r
1
|EiE =
|0i + |ni
(18)
2
with n = 10. Figure 6 (a) shows the calculation for cat
states (17). In this case, we reach fidelities over 90% in about
60 measurements. Moreover, Fig. 6 (b) shows results similar
to the qubit case given by Fig. (3), surpassing fidelities of 90%
in less than 40 iterations. The last figure reflects the fact that
for the state in Eq. (18), the protocol is reduced to the qubit
case, given that only two states are involved in the superposition. Thus, all states of the form in Eq. (18) have the same
performance as the qubit case.
We can see from Figs. 4, 5 and 6, that the learning speed is
inversely proportional to the parameter , which means that a
small value of implies a rapid increase in fidelity between R
and E, that is, it increases the speed of learning. On the other
hand, the maximum learning is also directly proportional to
, in other words, a small value of means lower maximum
fidelities between R and E. It is pertinent to emphasize that
our protocol for qubit and multilevel cases employs two-level
operators UA(k) , and each iteration only needs to calculate the
operator U(k) = U (k−1) U(k−1) . Hence, the protocol does not
need to store the complete agent history, which is an advantage in terms of the required resources.
This protocol can be implemented in any platform that enNOT
XOR
ables the logical operator Ua,b
for qubits, or Ua,b
for qudits, and digital feedback loops, as is the case of circuit quantum electrodynamics (cQEDs). This platform takes particular
relevance due to its fast development in quantum computation [40–46]. Current technology in cQEDs allows for digital
quantum feedback loops with elapsed times about 2[µs] and
fidelities around 99% [47, 48], well-controlled one and twoqubits gates with fidelities over 99% in less than 1[µs] [49],
with qubits with coherence times about 100[µs] [50, 51]. This
allows for more than 20 iterations of our protocol, a sufficient number for a feasible implementation. Aditionally, in
the last decade, multilevel gates have been theoretically proposed [52–54], as well as efficient multiqubit gates have recently been proposed using a ML approach [55, 56] providing
all the necessary elements for the experimental implementation of the general framework of this learning protocol.
IV.
CONCLUSIONS
We propose and analyse a quantum reinforcement learning
protocol to adapt a quantum state (the agent) to another, unknown, quantum state (the environment), in the context where
several identical copies of the unknown state are available.
The main goal of our proposal is for the agent to acquire information about the environment in a semi-autonomous way,
namely, in the reinforcement learning spirit. We show that the
fidelity increases rapidly with the number of iterations, reaching for qubit states average fidelities over 90% with less than
30 measurements. Also, for states with dimension d > 2,
we obtain average fidelity over 80% for d = 11, with about
400 measurements. The performance is improved for special
cases such as coherent states (average fidelities of 85% with
less than 100 iterations), cat states (average fidelities of 90%
√
with about 60 iterations) and states of the form (|0i + |ni)/ 2
(average fidelities of 90% with less than 40 iterations).
The performance of the protocol is handled by the value
of the parameter and by the number of states involved in
the superposition of the environment state, E, in the measurement basis. For a small we get a high learning speed and a
reduced maximum learning. Moreover, the number of states
in the superposition is related to the overall performance of
the protocol, that is, a superposition of fewer terms provides
better performance, which increases learning speed as well as
maximum learning, requiring less iterations to obtain high fidelity. These two facts imply that a possible improvement of
the protocol can be achieved by using a dynamic parameter
and a measurement device that can change its measurement
basis throughout the protocol to reduce the number of states
involved in the overlap of the state of E. Besides, since our
protocol increases the fidelity with a small number of iterations, it is useful even when the number of copies of E is limited. Finally, this protocol opens up the door to the implementation of semi-autonomous quantum reinforcement learning, a
6
next step for achieving quantum artificial life.
The authors acknowledge support from CONICYT Doctorado Nacional 21140432, Dirección de Postgrado USACH,
FONDECYT Grant No. 1140194, Ramón y Cajal Grant
RYC-2012-11391, MINECO/FEDER FIS2015-69983-P and
Basque Government IT986-16.
[1] S. Russell and P. Norvig, Artificial Intelligence: A modern approach (Prentice hall, 1995).
[2] R. S. Michalski, J. G. Carbonell, and T. M. Mitchell, Machine
learning: An artificial intelligence approach (Springer Science
& Business Media, 2013).
[3] M. I. Jordan and T. M. Mitchell, Science 349, 255 (2015).
[4] M. Kawagoe and A. Tojo, Pattern Recognition 17, 295 (1984).
[5] R. V. Shannon, F.-G. Zeng, V. Kamath, J. Wygonski, and
M. Ekelid, Science 270, 303 (1995).
[6] A. K. Jain, Nature 449, 38 (2007).
[7] J. Carrasquilla, and R. G. Melko, Nat. Phys. 13, 431 (2017).
[8] A. Fahad, N. Alshatri, Z. Tari, A. Alamri, I. Khalil, A. Y.
Zomaya, S. Foufou, and A. Bouras, IEEE Transactions on
Emerging Topics in Computing 2, 267 (2014).
[9] P. Baldi, P. Sadowski, and D. Whiteson, Nat. Comm. 5, 4308
(2014).
[10] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction (MIT press Cambridge, 1998).
[11] M. L. Littman, Nature 521, 445 (2015).
[12] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou,
A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton,
Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche,
T. Graepel, and D. Hassabis, Nature 550, 354 (2017).
[13] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai,
A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, et al.,
arXiv preprint arXiv:1712.01815 (2017).
[14] M. A. Nielsen and I. L. Chuang, Quantum computation and
quantum information (Cambridge University Press, Cambridge,
UK, 2010).
[15] M. Schuld, I. Sinayskiy, and F. Petruccione, Contemporary
Physics 56, 172 (2015).
[16] J. Adcock, E. Allen, M. Day, S. Frick, J. Hinchliff, M. Johnson,
S. Morley-Short, S. Pallister, A. Price, and S. Stanisic, arXiv
preprint arXiv:1512.02900 (2015).
[17] V. Dunjko, J. M. Taylor, and H. J. Briegel, Phys. Rev. Lett. 117,
130501 (2014).
[18] V. Dunjko and H. J. Briegel, arXiv preprint arXiv:1709.02779
(2017).
[19] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe,
and S. Lloyd, Nature 549, 195 (2017).
[20] R. Biswas, Z. Jiang, K. Kechezhi, S. Knysh, S. Mandrà,
B. O’Gorman, A. Perdomo-Ortiz, A. Petukhov, J. RealpeGómez, E. Rieffel, D. Venturelli, F. Vasko, and Z. Wang, Parallel Computing 64, 81 (2017), high-End Computing for NextGeneration Scientific Discovery.
[21] A. Perdomo-Ortiz, M. Benedetti, J. Realpe-Gómez, and
R. Biswas, arXiv preprint arXiv:1708.09757 (2017).
[22] A. Perdomo-Ortiz, A. Feldman, A. Ozaeta, S. V. Isakov, Z. Zhu,
B. O’Gorman, H. G. Katzgraber, A. Diedrich, H. Neven,
J. de Kleer, et al., arXiv preprint arXiv:1708.09780 (2017).
[23] M. Sasaki and A. Carlini, Phys. Rev. A 66, 022303 (2002).
[24] S. Lloyd, M. Mohseni, and P. Rebentrost, arXiv preprint
arXiv:1307.0411 (2013).
[25] M. Benedetti, J. Realpe-Gómez, R. Biswas, and A. PerdomoOrtiz, Phys. Rev. X 7, 041052 (2017).
[26] M. Benedetti, J. Realpe-Gómez, and A. Perdomo-Ortiz, arXiv
preprint arXiv:1708.09784 (2017).
[27] E. Aı̈meur, G. Brassard, and S. Gambs, Machine Learning 90,
261 (2013).
[28] G. D. Paparo, V. Dunjko, A. Makmal, M. A. Martin-Delgado,
and H. J. Briegel, Phys. Rev. X 4, 031002 (2014).
[29] L. Lamata, Scientific Reports 7, 1609 (2017).
[30] F. Cárdenas-López, L. Lamata, J. C. Retamal, and E. Solano,
arXiv preprint arXiv:1709.07848 (2017).
[31] U. Alvarez-Rodriguez, L. Lamata, P. Escandell-Montero, J. D.
Martı́n-Guerrero, and E. Solano, Scientific Reports 7, 13645
(2017).
[32] R. B. A. Adamson and A. M. Steinberg, Phys. Rev. Lett. 105,
030406 (2010).
[33] H. Sosa-Martinez, N. K. Lysne, C. H. Baldwin, A. Kalev, I. H.
Deutsch, and P. S. Jessen, Phys. Rev. Lett. 119, 150401 (2017).
[34] A. Lumino, E. Polino, A. S. Rab, G. Milani, N. Spagnolo,
N. Wiebe, and F. Sciarrino, arXiv preprint arXiv:1712.07570v1
(2017).
[35] A. Rocchetto, S. Aaronson, S. Severini, G. Carvacho,
D. Poderini, I. Agresti, M. Bentivegna, and F. Sciarrino, arXiv
preprint arXiv:1712.00127 (2017).
[36] G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko,
and G. Carleo, arXiv preprint arXiv:1703.05334 (2017).
[37] L.-L. Sun, Y. Mao, F.-L. Xiong, S. Yu, and Z.-B. Chen, arXiv
preprint arXiv:1802.00140 (2018).
[38] L. Roa, A. Delgado, M. L. Ladrón de Guevara, and A.
B. Klimov, Phys. Rev. A 73, 012322 (2006).
[39] G. Alber, A. Delgado, N. Gisin, and I. Jex, arXiv preprint
quant-ph/0008022 (2000).
[40] A. Blais, R.-S. Huang, A. Wallraff, S. M. Girvin, and R. J.
Schoelkopf, Phys. Rev. A 69, 062320 (2004).
[41] M. H. Devoret, A. Wallraff, and J. M. Martinis, arXiv preprint
cond-mat/0411174 (2004).
[42] M. Hofheinz, E. M. Weig, M. Ansmann, R. C. Bialczak,
E. Lucero, M. Neeley, A. D. O’Connell, H. Wang, J. M. Martinis, and A. N. Cleland, Nature 454, 310 (2008).
[43] M. Hofheinz, H. Wang, M. Ansmann, R. C. Bialczak,
E. Lucero, M. Neeley, A. D. O’Connell, D. Sank, J. Wenner,
J. M. Martinis, and A. N. Cleland, Nature 459, 546 (2009).
[44] L. DiCarlo, J. M. Chow, J. M. Gambetta, L. S. Bishop, B. R.
Johnson, D. I. Schuster, J. Majer, A. Blais, L. Frunzio, S. M.
Girvin, and R. J. Schoelkopf, Nature 460, 240 (2009).
[45] M. H. Devoret and R. J. Schoelkopf, Science 339, 1169 (2013).
[46] J. Otterbach, R. Manenti, N. Alidoust, A. Bestwick, M. Block,
B. Bloom, S. Caldwell, N. Didier, E. S. Fried, S. Hong, et al.,
arXiv preprint arXiv:1712.05771 (2017).
[47] D. Ristè, J. G. van Leeuwen, H.-S. Ku, K. W. Lehnert, and
L. DiCarlo, Phys. Rev. Lett. 109, 050507 (2012).
[48] D. Ristè, C. C. Bultink, K. W. Lehnert, and L. DiCarlo, Phys.
Rev. Lett. 109, 240502 (2012).
[49] R. Barends, A. Shabani, L. Lamata, J. Kelly, A. Mezzacapo,
U. Las Heras, R. Babbush, A. Fowler, B. Campbell, Y. Chen,
Z. Chen, B. Chiaro, A. Dunsworth, E. Jeffrey, E. Lucero,
A. Megrant, J. Y. Mutus, M. Neeley, C. Neil, P. J. J. O’Malley,
C. Quintana, P. Roushan, D. Sank, A. Vainsencher, J. Wenner,
T. C. White, E. Solano, H. Neven, and J. M. Martinis, Nature
7
534, 222 (2016).
[50] H. Paik, D. I. Schuster, L. S. Bishop, G. Kirchmair, G. Catelani,
A. P. Sears, B. R. Johnson, M. J. Reagor, L. Frunzio, L. I. Glazman, S. M. Girvin, M. H. Devoret, and R. J. Schoelkopf, Phys.
Rev. Lett. 107, 240501 (2011).
[51] C. Rigetti, J. M. Gambetta, S. Poletto, B. L. T. Plourde, J. M.
Chow, A. D. Córcoles, J. A. Smolin, S. T. Merkel, J. R. Rozen,
G. A. Keefe, M. B. Rothwell, M. B. Ketchen, and M. Steffen,
Phys. Rev. B 86, 100506 (2012).
[52] F. W. Strauch, Phys. Rev. A 84, 052313 (2011).
[53] B. Mischuck and K. Mølmer, Phys. Rev. A 87, 022341 (2013).
[54] E. O. Kiktenko, A. K. Fedorov, O. V. Man’ko, and V. I.
Man’ko, Phys. Rev. A 91, 042312 (2015).
[55] E. Zahedinejad, J. Ghosh, and B. C. Sanders, Phys. Rev. Lett.
114, 200502 (2015).
[56] E. Zahedinejad, J. Ghosh, and B. C. Sanders, Phys. Rev. Applied 6, 054005 (2016).
| 2cs.AI
|
arXiv:1601.00590v2 [math.GR] 17 Mar 2016
SPINORS AND ESSENTIAL DIMENSION
SKIP GARIBALDI AND ROBERT M. GURALNICK
With an appendix by Alexander Premet
Abstract. We prove that spin groups act generically freely on various spinor
modules, in the sense of group schemes and in a way that does not depend on
the characteristic of the base field. As a consequence, we extend the surprising calculation of the essential dimension of spin groups and half-spin groups
in characteristic zero by Brosnan–Reichstein–Vistoli (Annals of Math., 2010)
and Chernousov–Merkurjev (Algebra & Number Theory, 2014) to fields of
characteristic different from 2. We also complete the determination of generic
stabilizers in spin and half-spin groups of low rank.
1. Introduction
The essential dimension of an algebraic group G is, roughly speaking, the number of parameters needed to specify a G-torsor. Since the notion was introduced in
[BR97] and [RY00], there have been many papers calculating the essential dimension of various groups, such as [KM03], [CS06], [Flo08], [KM08], [GR09], [Mer10],
[BM12], [LMMR13], etc. (See [Mer15a], [Mer13], or [Rei10] for a survey of the
current state of the art.) For connected groups, the essential dimension of G tends
to be less than the dimension of G as a variety; for semisimple adjoint groups this
is well known1. Therefore, the discovery by Brosnan–Reichstein–Vistoli in [BRV10]
that the essential dimension of the spinor group Spinn grows exponentially as a
function of n (whereas dim Spinn is quadratic in n), was startling. Their results,
together with refinements for n divisible by 4 in [Mer09] and [CM14], determined
the essential dimension of Spinn for n > 14 if char k = 0. One goal of the present
paper is to extend this result to all characteristics except 2.
Generically free actions. The source of the characteristic zero hypothesis in
[BRV10] is that the upper bound relies on a fact about the action of spin groups on
spinors that is only available in the literature in case the field k has characteristic
zero. Recall that a group G acting on a vector space V is said to act generically
freely if there is a dense open subset U of V such that, for every K ⊇ k and every
u ∈ U (K), the stabilizer in G of u is the trivial group scheme. We prove:
Theorem 1.1. Suppose n > 14. Then Spinn acts generically freely on the spin
representation if n ≡ 1, 3 mod 4; a half-spin representation if n ≡ 2 mod 4; or
a direct sum of the vector representation and a half-spin representation if n ≡
0 mod 4. Furthermore, if n ≡ 0 mod 4 and n ≥ 20, then HSpinn acts generically
freely on a half-spin representation.
2010 Mathematics Subject Classification. Primary 11E72; Secondary 11E88, 20G10.
Guralnick was partially supported by NSF grants DMS-1265297 and DMS-1302886.
1See [GG15a] for a proof that works regardless of the characteristic of the field.
1
2
SKIP GARIBALDI AND ROBERT M. GURALNICK
(We also compute the stabilizer of a generic vector for the values of n not covered
by Theorem 1.1. See below for precise statements.)
Throughout, we write Spinn for the split spinor group, which is the simply
connected cover (in the sense of linear algebraic groups) of the split group SOn . To
be precise, the vector representation is the map Spinn → SOn , which is uniquely
defined up to equivalence unless n = 8. For n not divisible by 4, the kernel µ2 of
this representation is the unique central µ2 subgroup of Spinn .
For n divisible by 4, the natural action of Spinn on the spinors is a direct sum
of two inequivalent representations, call them V1 and V2 , each of which is called a
half-spin representation. The center of Spinn in this case contains two additional
copies of µ2 , namely the kernels of the half-spin representations Spinn → GL(Vi ),
and we write HSpinn for the image of Spinn (the isomorphism class of which does
not depend on i). For n ≥ 12, HSpinn is not isomorphic to SOn .
Theorem 1.1 is known under the additional hypothesis that char k = 0, see
[AP71, Th. 1] for n ≥ 29 and [Pop88] for n ≥ 15. The proof below is independent
of the characteristic zero results, and so gives an alternative proof.
We note that Guerreiro proved that the generic stabilizer in the Lie algebra
spinn , acting on a (half) spin representation, is central for n = 22 and n ≥ 24, see
Tables 6 and 9 of [Gue97]. At the level of group schemes, this gives the weaker
result that the generic stabilizer is finite étale. Regardless, we recover these cases
quickly, see §3; the longest part of our proof concerns the cases n = 18 and 20.
Generic stabilizer in Spinn for small n. For completeness, we list the stabilizer
in Spinn of a generic vector for 6 ≤ n ≤ 14 in Table 1. The entries for n ≤ 12 and
char k 6= 2 are from [Igu70]; see sections 7–9 below for the remaining cases. The
case n = 14 is particularly important due to its relationship with the structure of
14-dimensional quadratic forms with trivial discriminant and Clifford invariant (see
[Ros99a], [Ros99b], [Gar09], and [Mer15b]), so we calculate the stabilizer in detail
in that case.
n
6
7
8
9
10
char k 6= 2
(SL3 ) · (Ga )3
G2
Spin7
Spin7
(Spin7 ) · (Ga )8
char k = 2
same
same
same
same
same
n char k 6= 2
char k = 2
11
SL5
SL5 ⋊Z/2
12
SL6
SL6 ⋊Z/2
13 SL3 × SL3 (SL3 × SL3 ) ⋊ Z/2
14 G2 × G2
(G2 × G2 ) ⋊ Z/2
Table 1. Stabilizer sub-group-scheme in Spinn of a generic vector
in an irreducible (half) spin representation for small n.
For completeness, we also record the following.
Theorem 1.2. Let k be an algebraically closed field. The stabilizer in HSpin16 of
a generic vector in a half-spin representation is isomorphic to (Z/2)4 × (µ2 )4 .
The proof when char k =
6 2 is short, see Lemma 4.2. The case of char k = 2
is treated in an appendix by Alexander Premet. (Eric Rains has independently
proved this result.)
SPINORS AND ESSENTIAL DIMENSION
3
Essential dimension. We recall the definition of essential dimension. For an
extension K of a field k and an element x in the Galois cohomology set H 1 (K, G), we
define ed(x) to be the minimum of the transcendence degree of K0 /k for k ⊆ K0 ⊆
K such that x is in the image of H 1 (K0 , G) → H 1 (K, G). The essential dimension
of G, denoted ed(G), is defined to be max ed(x) as x varies over all extensions
K/k and all x ∈ H 1 (K, G). There is also a notion of essential p-dimension for
a prime p. The essential p-dimension edp (x) is the minimum of ed(resK ′ /K x) as
K ′ varies over finite extensions of K such that p does not divide [K ′ : K], where
resK ′ /K : H 1 (K, G) → H 1 (K ′ , G) is the natural map. The essential p-dimension
of G, edp (G), is defined to be the minium of edp (x) as K and x vary; trivially,
edp (G) ≤ ed(G) for all p and G, and edp (G) = 0 if for every K every element of
H 1 (K, G) is killed by some finite extension of K of degree not divisible by p.
Our Theorem 1.1 gives upper bounds on the essential dimension of Spinn and
HSpinn regardless of the characteristic of k. Combining these with the results of
[BRV10], [Mer09], [CM14], and [Löt13] quickly gives the following, see §6 for details.
Corollary 1.3. For n > 14 and char k 6= 2,
(n−1)/2
− n(n−1)
2
2
ed2 (Spinn ) = ed(Spinn ) = 2(n−2)/2 − n(n−1)
2
(n−2)/2 n(n−1)
2
− 2 + 2m
if n ≡ 1, 3 mod 4;
if n ≡ 2 mod 4; and
if n ≡ 0 mod 4
where 2m is the largest power of 2 dividing n in the final case. For n ≥ 20 and
divisible by 4,
n(n − 1)
.
2
Although Corollary 1.3 is stated and proved for split groups, it quickly implies
analogous results for non-split forms of these groups, see [Löt13, §4] for details.
Combining the corollary with the calculation of ed(Spinn ) for n ≤ 14 by Markus
Rost in [Ros99a] and [Ros99b] (see also [Gar09]), we find for char k 6= 2:
ed2 (HSpinn ) = ed(HSpinn ) = 2(n−2)/2 −
n
5
ed(Spinn ) 0
6 7
0 4
8 9
5 5
10 11 12 13 14 15 16 17 18 19 20
4 5 6 6 7 23 24 120 103 341 326
Notation. Let G be an affine group scheme of finite type over a field k, which
we assume is algebraically closed and of characteristic different from 2. (If G is
additionally smooth, then we say that G is an algebraic group.) If G acts on a
variety X, the stabilizer Gx of an element x ∈ X(k) is a sub-group-scheme of G
with R-points
Gx (R) = {g ∈ G(R) | gx = x}
for every k-algebra R.
If Lie(G) = 0 then G is finite and étale. If additionally G(k) = 1, then G is the
trivial group scheme Spec k.
For a representation ρ : G → GL(V ) and elements g ∈ G(k) and x ∈ Lie(G), we
denote the fixed spaces by V g := ker(ρ(g) − 1) and V x := ker(dρ(x)).
We use fraktur letters such as g, spinn , etc., for the Lie algebras Lie(G), Lie(Spinn ),
etc.
Acknowledgements. We thank Alexander Merkurjev and Zinovy Reichstein for helpful
comments, and for posing the questions answered in Proposition 9.1 and Theorem 1.1.
4
SKIP GARIBALDI AND ROBERT M. GURALNICK
2. Fixed spaces of elements
Fix some n ≥ 6. Let V be a (half) spin representation for Spinn , of dimension
2⌊(n−1)/2⌋ .
Proposition 2.1. For n ≥ 6:
(1) For all noncentral x ∈ spinn , dim V x ≤ 43 dim V .
(2) For all noncentral g ∈ Spinn , dim V g ≤ 43 dim V .
If n > 8, char k 6= 2, and g ∈ Spinn is noncentral semisimple, then dim V g ≤
5
8 dim V .
In the proof, in case char k 6= 2, we view SOn as the group of matrices
SOn (k) = {A ∈ SLn (k) | SA⊤ S = A−1 },
where S is the matrix 1’s on the “second diagonal”, i.e., Si,n+1−i = 1 and the other
entries of S are zero. The intersection of the diagonal matrices with SOn are a maxi−1
mal torus. For n even, one finds elements of the form (t1 , t2 , . . . , tn/2 , t−1
n/2 , . . . , t1 ),
and we abbreviate these as (t1 , t2 , . . . , tn/2 , . . .). Explicit formulas for a triality automorphism σ of Spin8 of order 3 are given in [Gar98, §1], and for g =
(t1 , t2 , t3 , t4 , . . .) ∈ Spin8 the elements σ(g) and σ 2 (g) have images in SO8
q
q
q
q
t1 t2 t3
t1 t4 t2
t1 t3 t4
t1
ε
(2.2)
and
t4 ,
t3 ,
t2 ,
t2 t3 t4 , . . .
q
q
q
√
t1 t2 t3 t4 , tt13 tt24 , tt21 tt34 , tt21 tt34 , . . . ,
ε
where ε = ±1 is the only impecision in the expression.
Proof. For (1), in the Jordan decomposition x = s + n where s is semisimple, n is
nilpotent, and [s, n] = 0, we have V x ⊆ V s ∩ V n , so it suffices to prove (1) for x
nilpotent or semisimple.
Suppose first that x is a root element. If n = 6, then spinn ∼
= sl4 and V is
the natural representation of sl4 , so we have the desired equality. For n > 6, the
module restricted to son−1 is either irreducible or the direct sum of two half spins
and so the result follows.
If x is nonzero nilpotent, then we may replace x by a root element in the closure
of (Ad G)x. If x is noncentral semisimple, choose a root subgroup Uα of SOn
belonging to a Borel subgroup B such that x lies in Lie(B) and does not commute
with Uα . Then for all y ∈ Lie(Uα ) and all scalars λ, x + λy is in the same Ad(SOn )orbit as x and y is in the closure of the set of such elements; replace x with y. If x
is nonzero nilpotent, then root elements are in the closure of Ad(G)x.
(In case n is divisible by 4, the natural map spinn → hspinn is an isomorphism
on root elements. It follows that for noncentral x ∈ hspinn , dim V x ≤ 43 dim V by
the same argument.)
For (2), we may assume that g is unipotent or semisimple. If g is unipotent,
then by taking closures, we may pass to root elements and argue as for x in the Lie
algebra.
If g is semisimple, we actually prove a slightly stronger result: all eigenspaces
have dimension at most 34 dim V .
Suppose now that n is even. The image of g in SOn can be viewed as an element
of SOn−2 × SO2 , where it has eigenvalues (a, a−1 ) in SO2 . Replacing if necessary
g with a multiple by an element of the center of Spinn , we may assume that g is
SPINORS AND ESSENTIAL DIMENSION
5
in the image of Spinn−2 × Spin2 . Then V = V1 ⊕ V2 where the Vi are distinct half
spin modules for Spinn−2 and the Spin2 acts on each (since they are distinct and
Spin2 commutes with Spinn−2 ). By induction every eigenspace of g has dim at
most 43 dim Vi and the Spin2 component of g acts as a scalar, so this is preserved.
If n is odd, then the image of g in SOn has eigenvalue 1 on the natural module,
so is contained in a SOn−1 subgroup. Replacing if necessary g with gz for some z
in the center of G, we may assume that g is in the image of Spinn−1 and the claim
follows by induction.
For the final claim, view g as an element in the image of (g1 , g2 ) ∈ Spin8 × Spinn−8 .
The result is clear unless in some 8-dimensional image of Spin8 , g1 has diagonal
image (a, 1, 1, 1, . . .) ∈ SO8 . We may assume that the image of g has prime order
in GL(V ).
If g1 has odd order or order 4, then as in (2.2), on the other two 8 dimensional
representations of Spin8 it has no fixed space and each eigenspace of dimension at
most 4. So on the sum of any two of these representations the largest eigenspace is
at most 10-dimensional (out of 16), and the claim follows.
If g1 has order 2, then it maps to an element of order 4 in the other two 8dimensional representations via (2.2) and again the same argument applies.
The proposition will feed into the following elementary lemma, which resembles
[AP71, Lemma 4] and [Gue97, §3.3].
Lemma 2.3. Let V be a representation of a semisimple algebraic group G over an
algebraically closed field k.
(1) If for every unipotent g ∈ G and every noncentral semisimple g ∈ G of
prime order we have
dim V g + dim g G < dim V,
(2.4)
then for generic v ∈ V , Gv (k) is central in G(k).
(2) Suppose char k = p > 0 and let h be a G-invariant subspace of g. If for
every nonzero x ∈ g \ h such that x[p] ∈ {0, x} we have
dim V x + dim(Ad(G)x) < dim V,
(2.5)
then for generic v ∈ V , Lie(Gv ) ⊆ h.
We will apply this to conclude that Gv is the trivial group scheme for generic v,
so the hypothesis on char k in (2) is harmless. When char k = 0, the conclusion of
(1) suffices.
Proof. For (1), see [GG15b, §10] or adjust slightly the following proof of (2). For
x ∈ g, define
[
gV x .
V (x) := {v ∈ V | there is g ∈ G(k) s.t. xgv = 0} =
g∈G(k)
x
Define α : G × V → V by α(g, w) = gw, so the image of α is precisely V (x).
The fiber over gw contains (gc−1 , cw) for Ad(c) fixing x, and so dim V (x) ≤
dim(Ad(G)x) + dim V x .
Let X ⊂ g be the set of nonzero x ∈ g \ h such that x[p] ∈ {0, x}; it is a union of
finitely many G-orbits. (Every toral element — i.e., x with x[p] = x — belongs to
Lie(T ) for a maximal torus T in G by [BS66], and it is obvious that there are only
finitely many conjugacy classes of toral elements in Lie(T ).) Now V (x) depends
6
SKIP GARIBALDI AND ROBERT M. GURALNICK
only on the G-orbit of X (because V Ad(g)x = gV x ), so the union ∪x∈X V (x) is a
finite union. As dim V (x) < dim V by the previous paragraph, the union ∪V (x) is
contained in a proper closed subvariety Z of V , and for every v in the (nonempty,
open) complement of Z, gv does not meet X.
For each v ∈ (V \ Z)(k) and each y ∈ gv , we can write y as
y = yn +
r
X
αi yi ,
[yn , yi ] = [yi , yj ] = 0 for all i, j
i=1
[p]
such that y1 , . . . , yr ∈ gv are toral, and yn ∈ gv satisfies yn = 0, see [SF88,
Th. 3.6(2)]. Thus yn and the y1 , . . . , yr are in h by the previous paragraph.
Note that, in proving Theorem 1.1, we may assume that k is algebraically closed
(and so this hypothesis in Lemma 2.3 is harmless). Indeed, suppose G is an algebraic
group acting on a vector P
space V over a field k. Fix a basis v1 , . . . , vn of V and
consider the element η := ti vi ∈ V ⊗ k(t1 , . . . , tn ) = V ⊗ k(V ) for indeterminates
t1 , . . . , tn ; it is a sort of generic point of V . Certainly, G acts generically freely on
V over k if and only if the stabilizer (G × k(V ))v is the trivial group scheme, and
this statement is unchanged by replacing k with an algebraic closure. That is, G
acts generically freely on V over k if and only if G × K acts generically freely on
V ⊗ K for K an algebraic closure of k.
3. Proof of Theorem 1.1 for n > 20
Suppose n > 2, and put V for a (half) spin representation of Spinn . Recall that
dim Spinn = r(2r − 1) and
dim V = 2r−1
if n = 2r
whereas
dim Spinn = 2r2 + r
and
dim V = 2r
if n = 2r + 1
and in both cases rank Spinn = r. Proposition 2.1 gives an upper bound on dim V g
for noncentral g, and certainly the conjugacy class of g has dimension at most
dim Spinn −r. If we assume n ≥ 21 and apply these, we obtain (2.4) and consequently the stabilizer S of a generic v ∈ V has S(k) central in Spinn (k). Repeating
this with the Lie algebra spinn (and h the center of spinn ) we find that Lie(S)
is central in spinn . For n not divisible by 4, the representation Spinn → GL(V )
restricts to a closed embedding on the center of Spinn , so S is the trivial group
scheme as claimed in Theorem 1.1.
For n divisible by 4, we conclude that HSpinn acts generically freely on V (using
that Proposition 2.11 holds also for hspinn ). As the kernel µ2 of Spinn → HSpinn
acts faithfully on the vector representation W , it follows that Spinn acts generically
freely on V ⊕ W , completing the proof of Theorem 1.1 for n > 20.
4. Proof of Theorem 1.1 for n ≤ 20 and characteristic 6= 2
In this section we assume that char k 6= 2, and in particular the Lie algebra spinn
(and hspinn in case n is divisible by 4) is naturally identified with son .
Case n = 18 or 20. Take V to be a half-spin representation of G = Spinn (if
n = 18) or G = HSpinn (if n = 20). To prove Theorem 1.1 for these n, it suffices to
prove that G acts generically freely on V , which we do by verifying the inequalities
(2.4) and (2.5).
SPINORS AND ESSENTIAL DIMENSION
7
Nilpotents and unipotents. Let x ∈ g with x[p] = 0. The argument for unipotent
elements of G is essentially identical (as we assume char k 6= 2) and we omit it.
If, for a particular x, we find that the centralizer of x has dimension > 89 (if
n = 18) or > 62 (if n = 20), then dim(Ad(G)x) < 41 dim V and we are done by
Proposition 2.1.
For x nilpotent, the most interesting case is where x is has partition (22t , 1n−2t )
for some t. If n = 20, then such a class has centralizer of dimension at least 100,
and we are done. If n = 18, we may assume by similar reasoning that t = 3 or
4. The centralizer of x has dimension ≥ 81, so dim(Ad(G)x) ≤ 72. We claim that
dim V x ≤ 140; it suffices to prove this for an element with t = 3, as the element with
t = 4 specializes to it. View it as an element in the image of so9 × so9 → so18 where
the first factor has partition (24 , 1) and the second has partition (22 , 15 ). Now,
triality on so8 sends elements with partition 24 to elements with partition 24 and
(3, 15 ) — see for example [CM93, p. 97] — consequently the (24 , 1) in so9 acts on
the spin representation of so9 as a (3, 24 , 15 ). Similarly, the (22 , 15 ) acts on the spin
representation of so9 as (24 , 18 ). The action of x on the half-spin representation of
so18 is the tensor product of these, and we find that dim V x ≤ 140 as claimed.
Suppose x is nilpotent and has a Jordan block of size at least 5. An element
with partition (5, 1) in so6 is a regular nilpotent in sl4 with 1-dimensional kernel.
Using the tensor product decomposition as in the proof of Proposition 2.1, we
deduce that an element y ∈ son with partition (5, 1n−5 ) has dim V y ≤ 41 dim V ,
and consequently by specialization dim V x ≤ 14 dim V . As dim(Ad(G)x) ≤ dim G −
rank G < 34 dim V , the inequality is verified for this x.
Now suppose x is nilpotent and all Jordan blocks have size at most 4, so it is a
specialization of (44 , 12 ) if n = 18 or 45 if n = 20. These classes have centralizers
of dimension 41 and 50 respectively, hence dim(Ad(G)x) < 12 dim V . If x has at
least two Jordan blocks of size at least 3, then x specializes to (32 , 1n−6 ); as triality
sends elements with partition (32 , 12 ) to elements with the same partition, we find
dim V x ≤ 12 dim V . We are left with the case where x has partition (3, 22t , 1n−2t−3 )
for some t. If t = 0, then the centralizer of x has dimension 121 or 154 and we are
done. If t > 0, then x specializes to y with partition (3, 22 , 1n−7 ). As triality on
so8 leaves the partition (3, 22 , 1) unchanged, we find dim V x ≤ dim V y ≤ 12 dim V ,
as desired, completing the verification of (2.5) for x nilpotent.
Semisimple elements in Lie(G). For x ∈ son semisimple, the most interesting case
is when x is diagonal with entries (at , (−a)t , 0n−2t ) where exponents denote multiplicity
a∈ k × . The centralizer of x is GLt ×SOn−2t , so dim(Ad(SOn )x) =
2 and
n
n−2t
. This is less than 41 dim V for n = 20, settling that case. For n = 18,
2 −t −
2
if t = 1 or 2, x is in the image of an element (a, −a, 0, 0) or (a/2, a/2, −a/2, −a/2)
in sl4 ∼
= so6 , and the tensor product decomposition gives that dim V x ≤ 21 dim V
and again we are done. If t > 2, we consider a nilpotent y = ( 00 Y0 ) not commuting
with x where Y is 9-by-9 and y specializes to a nilpotent y ′ with partition (24 , 18 ).
′
Such a y ′ acts on V as 16 copies of (3, 24 , 15 ), hence dim V y = 160. By specializing
x
x to y as in the proof of Proposition 2.1, we find dim V ≤ 160 and again we are
done.
Semisimple elements in G. Let g ∈ G(k) be semisimple, non-central, and of prime
order. If n = 20, then dim g G ≤ 180 < 83 dim V and we are done by Proposition
8
SKIP GARIBALDI AND ROBERT M. GURALNICK
2.1. So assume n = 18. If we find that the centralizer of g has dimension > 57,
then dim g G < 38 dim V and we are done by Proposition 2.1.
If g has order 2, then it maps to an element of order 2 in SO18 whose centralizer
is no smaller than SO8 × SO10 of dimension 73, and we are done. So assume g has
odd prime order. We divide into cases depending on the image g ∈ SO18 of g.
If g has at least 5 distinct eigenvalues, then either it has at least 6 distinct eigenvalues a, a−1 , b, b−1 , c, c−1 , or it has 4 distinct eigenvalues that are not equal to 1,
and the remaining eigenvalue is 1. In the latter case set c = 1. View g as the image
of (g1 , g2 ) ∈ Spin6 × Spin12 where g1 maps to a diagonal (a, b, c, c−1 , b−1 , a−1 ) in
SO6 , a regular semisimple element. Therefore, the eigenspaces of the image of g1
under the isomorphism Spin6 ∼
= SL4 are all 1-dimensional and the tensor decomposition argument shows that dim V g ≤ 14 dim V . As dim g G ≤ 144 < 43 dim V , we
are done in this case.
If g has exactly 4 eigenvalues, then the centralizer of g is at least as big as
GL4 × GL5 of dimension 41, so dim g G ≤ 112 < 21 dim V . Viewing g as the image
of (g1 , g2 ) ∈ Spin8 × Spin10 such that the image g 1 of g1 in SO8 exhibits all 4
eigenvalues, then g1 has eigenspaces all of dimension 2 or of dimensions 3, 3, 1, 1.
Considering the possible images of g 1 as in (2.2), each eigenspaces in each of the
8-dimensional representations is at most 4, so dim V g ≤ 12 dim V and this case is
settled.
In the remaining case, g has exactly 2 nontrivial (i.e., not 1) eigenvalues a, a−1 .
If 1 is not an eigenvalue of g, then the centralizer of g is GL9 of dimension 81,
and we are done. If the eigenspaces for the nontrivial eigenvalues are at least 4dimensional, then we can take g to be the image of (g1 , g2 ) ∈ Spin10 × Spin8 where
g1 maps to (a, a, a, a, 1, . . .) ∈ SO10 . The images of (a, a, a, a, . . .) ∈ SO8 as in
(2.2) are (a, a, a, a−1 , . . .) and (a2 , 1, 1, 1, . . .), so the largest eigenspace of g1 on a
half-spin representation is 6, so dim V g ≤ 83 dim V . As the conjugacy class of a
regular element has dimension 144 < 85 dim V , this case is complete. Finally, if g
has eigenspaces of dimension at most 2 for a, a−1 , then dim g G ≤ 58 < 38 dim V
and the n = 18 case is complete.
Case n = 17 or 19. For n = 17 or 19, the spin representation of Spinn can be
viewed as the restriction of a half-spin represenation of the overgroup HSpinn+1 .
We have already proved that this representation of HSpinn+1 is generically free.
Case n = 15 or 16. We use the general fact:
Lemma 4.1. Let G be a quasi-simple algebraic group and H a proper closed subgroup of G and X finite. Then for generic g ∈ G, H ∩ gXg −1 = H ∩ X ∩ Z(G).
Proof. For each x ∈ X \ Z(G), note that W (x) := {g ∈ G | xg ∈ H} is a proper
closed subvariety of G and, since X is finite, ∪W (x) is also proper closed. Thus for
an open subset of g in G, g(X \ Z(G))g −1 does not meet H.
Lemma 4.2. Let G = HSpin16 and V a half-spin representation over an algebraically closed field k of characteristic 6= 2. The stabilizer of a generic vector in
V is isomorphic to (Z/2)8 , as a group scheme.
Proof. Consider Lie(E8 ) = Lie(G) ⊕ V where these are the eigenspaces of an involution in E8 . That involution inverts a maximal torus T of E8 and so there is
maximal Cartan subalgebra t = Lie(T ) on which the involution acts as −1. As E8 is
SPINORS AND ESSENTIAL DIMENSION
9
smooth and adjoint, for a generic element τ ∈ t, the centralizer CE8 (τ ) has identity
component T by [DG70, XIII.6.1(d), XIV.3.18], and in fact equals T by [GG15a,
Prop. 2.3]. Since t misses Lie(G), the annihilator of τ in Lie(G) is 0 as claimed.
Furthermore, Gv (k) = T (k) ∩ G(k), i.e., the elements of T (k) that commute with
the involution, so Gv (k) ∼
= µ2 (k)8 .
Corollary 4.3. If char k 6= 2, then Spin15 acts generically freely on V.
Proof. Of course the Lie algebra does because this is true for Lie(Spin16 ).
For the group, a generic stabilizer is Spin15 ∩X where X is a generic stabilizer
in Spin16 . Now X is finite and meets the center of Spin16 in the kernel of Spin16 →
HSpin16 , whereas Spin15 injects in to HSpin16 . Therefore, by Lemma 4.1 a generic
conjugate of X intersect Spin15 is trivial.
Corollary 4.4. If char k 6= 2, then Spin16 acts generically freely on V ⊕ W , where
V is a half-spin and W is the natural (16-dimensional) module.
Proof. Now the generic stabilizer is already 0 for the Lie algebra on V whence on
V ⊕ W.
In the group Spin16 , a generic stabilizer is conjugate to X g ∩ Spin15 where X
is the finite stabilizer on V and as in the proof of the previous corollary, this is
generically trivial.
5. Proof of Theorem 1.1 for n ≤ 20 and characteristic 2
To complete the proof of Theorem 1.1, it remains to prove, in case char k = 2,
that the following representations G → GL(V ) are generically free:
(1) G = Spin15 , Spin17 , Spin19 and V is a spin representation.
(2) G = Spin18 and V is a half-spin representation.
(3) G = Spin16 or Spin20 and V is a direct sum of the vector representation
and a half-spin representation.
(4) G = HSpin20 and V is a half-spin representation.
Since we are in bad characteristic, the class of unipotent and nilpotent elements
are more complicated. On the other hand, since we are in a fixed small characteristic
and the dimensions of the modules and Lie algebras are relatively small, one can
actually do some computations.
In particular, we check that in each case that there exists a v ∈ V over the field
of 2 elements such that Lie(Gv ) = 0. (This can be done easily in various computer
algebra systems.) It follows that the same is true over any field of characteristic 2.
Since the set of w ∈ V where Lie(Gw ) = 0 is an open subvariety of V , this shows
that Lie(Gw ) is generically 0.
It remains to show that the group of k-points Gv (k) of the stabilizer of a generic
v ∈ V is the trivial group.
First consider G = Spin16 . By Lemma 4.1, it suffices to show that, for generic
w in a half-spin representation W , Gw (k) is finite, which is true by the appendix. Alternatively, the finiteness of Gw (k) was proved in [GLLT16] by working in
Lie(E8 ) = hspin16 ⊕ W and exhibiting a regular nilpotent of Lie(E8 ) in W whose
stabilizer in hspin16 is trivial. Since the set of w where (Spin16 )w (k) is finite is
open, the result follows.
Similarly, Spin15 acts generically freely on the spin representation.
10
SKIP GARIBALDI AND ROBERT M. GURALNICK
As in the previous section, it suffices to show that for G one of HSpin20 and
Spin18 and V a half spin representation, Gv (k) = 1 for generic v ∈ V .
We first consider involutions. We recall that an involution g ∈ SO2n = SO(W )
(in characteristic 2) is essentially determined by the number r of nontrivial Jordan
blocks of g (equivalently r = dim(g − 1)W ) and whether the subspace (g − 1)W is
totally singular or not with r even (and r ≤ n) — see [AS76], [LS12] or see [FGS16,
Sections 5,6] for a quick elementary treatment. If r < n or (g − 1)V is not totally
singular, there is 1 class for each possible pair of invariants. If r = n (and so n is
even) and (g − 1)V is totally singular, then there are two such classes interchanged
by a graph automorphism of order 2.
Lemma 5.1. Suppose char k = 2. Let G = Spin2n , n > 4 and let W be a halfspin representation. If g ∈ G is an involution other than a long root element, then
dim W g ≤ (5/8) dim W .
Proof. By passing to closures, we may assume that r = 4. Thus, g ∈ Spin8 ≤ G.
There is the class of long root elements (which have r = 2). The largest class
is invariant under graph automorphisms and so has a 4-dimensional fixed space
on each of the three 8 dimensional representations. The other three classes are
permuted by the graph automorphisms. Thus, it follows they have a 4 dimensional
fixed space on two of the 8-dimensional representations and a 5-dimensional fixed
space on the third such representation. Since the class of g is invariant under triality,
g has a 4-dimensional fixed space on each of the 8-dimensional representations of
Spin8 . Since the a half-spin representation of Spin10 is a sum of two distinct halfspin representations for Spin8 , the result is true for n = 5.
The result now follows by induction (since W is a direct sum of the two half spin
representations of Spin2n−2 ).
Lemma 5.2. Suppose char k = 2. Let G = Spin18 or HSpin20 with V a half spin
representation of dimension 256 or 512 respectively. Then Gv (k) = 1 for generic
v ∈V.
Proof. This is proved in [GLLT16] but we give a different proof. As in the previous
section, it suffices to show that dim V g + dim g G < dim V for every non-central
g ∈ G with g of prime order. If g has odd prime order (and so is semisimple), then
the argument is exactly the same as in the previous section (indeed, it is even easier
since there are no involutions to consider). Alternatively, since we know the result
in characteristic 0, it follows that generic stabilizers have no nontrivial semisimple
elements as in the proof of [GG15b, Lemma 10.3].
Thus, it suffices to consider g of order 2. Let r be the number of nontrivial
Jordan blocks of g. If g is not a long root element, then dim V g ≤ (5/8) dim V . On
the other hand, dim g G ≤ 99 for n = 10 and 79 for n = 9 by [AS76], [LS12], or
[FGS16]; in either case dim g G < (3/8) dim V .
The remaining case to consider is when g is a long root element. Then dim V g =
(3/4) dim V while dim g G = 34 or 30 respectively and again the inequality holds.
6. Proof of Corollary 1.3
For n not divisible by 4, the (half) spin representation Spinn is generically free
by Theorem 1.1, so by, e.g., [Mer13, Th. 3.13] we have:
ed(Spinn ) ≤ dim V − dim Spinn .
SPINORS AND ESSENTIAL DIMENSION
11
This gives the upper bound on ed(Spinn ) for n not divisible by 4. For n = 16,
we use the same calculation with V the direct sum of the vector representation
of Spin16 and a half-spin representation. For n ≥ 20 and divisible by 4, Theorem
1.1 gives that ed(HSpinn ) is at most the value claimed; with this in hand, the
argument in [CM14, Th. 2.2] (referring now to [Löt13] instead of [BRV10] for the
stacky essential dimension inequality) establishes the upper bound on ed(Spinn ) for
n ≥ 20 and divisible by 4.
It is trivially true that ed2 (Spinn ) ≤ ed(Spinn ). Finally, that ed2 (Spinn ) is at
least the expression on the right side of the display was proved in [BRV10, Th. 33(a)] for n not divisible by 4 and in [Mer09, Th. 4.9] for n divisible by 4; the lower
bound on ed2 (HSpinn ) is from [BRV10, Remark 3-10].
7. Spinn for 6 ≤ n ≤ 12 and characteristic 2
Suppose now that 6 ≤ n ≤ 12 and char k = 2. Let us now calculate the stabilizer
in Spinn of a generic vector v in a (half) spin representation, which will justify
those entries in Table 1. For n = 6, the Spin6 ∼
= SL4 and the representation is the
natural representation. For n = 8, the half-spin representation is indistinguishable
from the vector representation Spin8 → SO8 and again the claim is clear.
For the remaining n, we verify that the k-points (Spinn )v (k) of the generic
stabilizer are as claimed, i.e., that the claimed group scheme is the reduced subgroup-scheme of (Spinn )v . The cases n = 9, 11, 12 are treated in [GLMS97, Lemma
2.11] and the case n = 10 is [Lie87, p. 496].
For n = 7, view Spin7 as the stabilizer of an anisotropic vector in the vector
representation of Spin8 ; it contains a copy of G2 . As a G2 -module, the half-spin
representation of Spin8 is self-dual and has composition factors of dimensions 1, 6,
1, so G2 fixes a vector in V . As G2 is a maximal closed connected subgroup of
Spin7 , it is the identity component of the reduced subgroup of (Spin7 )v .
We have verified that the reduced sub-group-scheme of (Spinn )v agrees with the
corresponding entry, call it S, in Table 1. We now proceed as in §5 and find a w
such that dim(spinn )w = dim S, which shows that (Spinn )v is smooth, completing
the proof of Table 1 for n ≤ 12.
8. Spin13 and Spin14 and characteristic 6= 2
In this section, we determine the stabilizer in Spin14 and Spin13 of a generic
vector in the (half) spin representation V of dimension 64. We assume that char k 6=
2 and k is algebraically closed.
Let C0 denote the trace zero subspace of an octonion algebra with quadratic
norm N . We may view the natural representation of SO14 as a sum C0 ⊕ C0
endowed with the quadratic form N ⊕ −N . This gives an inclusion G2 × G2 ⊂ SO14
that lifts to an inclusion G2 × G2 ⊂ Spin14 . There is an element of order 4 in SO14
such that conjugation by it interchanges the two copies of G2 — the element of
order 2 in the orthogonal group with this property has determinant −1 — so the
normalizer of G2 × G2 in SO14 (k) is isomorphic to ((G2 × G2 )⋊ µ4 )(k) and in Spin14
it is ((G2 × G2 ) ⋊ µ8 )(k).
Viewing V as an internal Chevalley module for Spin14 (arising from the embedding of Spin14 in E8 ), it follows that Spin14 has an open orbit in P(V ). Moreover,
the unique (G2 × G2 )-fixed line kv in V belongs to this open orbit, see [Pop80,
12
SKIP GARIBALDI AND ROBERT M. GURALNICK
p. 225, Prop. 11], [Ros99a], or [Gar09, §21]. That is, for H the reduced sub-groupscheme of (Spin14 )v , H ◦ ⊇ G2 × G2 . By dimension count this is an equality. A
computation analogous to the one in the preceding paragraph shows that the idealizer of Lie(G2 × G2 ) in so14 is Lie(G2 × G2 ) itself, hence Lie((Spin14 )v ) = Lie(H ◦ ),
i.e., (Spin14 )v is smooth. It follows from the construction above that the stabilizer
of kv in Spin14 is all of (G2 × G2 ) ⋊ µ8 (as a group scheme). The element of order 2
in µ8 is in the center of Spin14 and acts as −1 on V , so the stabilizer of v is G2 × G2
as claimed in Table 1.
Now fix a vector (c, c′ ) ∈ C0 ⊕ C0 so that N (c), N (c′ ) and N (c) − N (c′ ) are all
nonzero. The stabilizer of (c, c′ ) in Spin14 is a copy of Spin13 , and the stabilizer of
v in Spin13 is its intersection with G2 × G2 , i.e., the product (G2 )c × (G2 )c′ . Each
term in the product is a copy of SL3 (see for example [KMRT98, p. 507, Exercise
6]), as claimed in Table 1. (On the level of Lie algebras and under the additional
hypothesis that char k = 0, this was shown by Kac and Vinberg in [GV78, §3.2].)
9. Spin13 and Spin14 and characteristic 2
We will calculate the stabilizer in Spinn of a generic vector in an irreducible
(half-)spin representation for n = 13, 14 over a field k of characteristic 2.
Proposition 9.1. The stabilizer in Spin14 (over a field k of characteristic 2) of a
generic vector in a half-spin representation is the group scheme (G2 × G2 ) ⋊ Z/2.
We use the following construction. Let X ⊃ R, V1 , V2 be vector spaces endowed
with quadratic forms qX , qR := qX |R , q1 , q1 such that qR is totally singular; qX , q1 ,
and q2 are nonsingular; R is a maximal totally singular subspace of X; and there
exist isometric embeddings fi : (X, qX ) ֒→ (Vi , qi ). For example, one could take
V1 and V2 to be copies of an octonion algebra C, R to be the span of the identity
element 1C , and X to be a quadratic étale subalgebra of C. There is a natural
quadratic form on the pushout (V1 ⊕ V2 )/(f1 − f2 )(X); if we write Vi ∼
= Vi′ ⊥ fi (X),
′
′
then the quadratic space is isomorphic to V1 ⊥ V2 ⊥ X. We can perform a similar
construction where the role of Vi is played by the codimension-1 subspace fi (R)⊥
and the pushout is (f1 (R)⊥ ⊕ f2 (R)⊥ )/(f1 − f2 )(R), giving a homomorphism of
algebraic groups Bℓ1 × Bℓ2 → Bℓ1 +ℓ2 where 2ℓi + 2 = dim Vi .
Proof of Proposition 9.1. The 7-dimensional Weyl module of the split G2 gives an
embedding G2 ֒→ SO7 . Combining this with the construction in the previous
paragraph gives maps
G2 × G2 → SO7 × SO7 → SO13 → SO14
which lift to maps where every SO is replaced by Spin.
Put V for a half-spin representation of Spin14 . It restricts to the spin representation of Spin13 . Calculating the restriction of the weights of V to Spin7 × Spin7
using the explicit description of the embedding, we see that V is the tensor product
of the 8-dimensional spin representations of Spin7 . By triality, the restriction of
one of the spin representations to G2 is the action of G2 on the octonions C, which
is a uniserial module with 1-dimensional socle S (spanned by the identity element
in C) and 7-dimensional radical, the Weyl module of trace zero octonions. The
restriction of V = C ⊗ C to the first copy of G2 is eight copies of C, so has an
8-dimensional fixed space S ⊗ C. As (S ⊗ C)1×G2 = S ⊗ S, we find that S ⊗ S is
the unique line in V stabilized by G2 × G2 .
SPINORS AND ESSENTIAL DIMENSION
13
We now argue that the Spin14 -orbit of S ⊗ S is open in P(V ). To see this, by
[Röh93], it suffices to verify that G2 × G2 is not contained in the Levi subgroup of
a parabolic subgroup of Spin14 . This is easily verified; the most interesting case is
where the Levi has type A6 , and G2 × G2 cannot be contained in such because the
restriction of V to A6 has composition factors of dimension 1, 7, 21, and 35. We
conclude that every nonzero v ∈ S ⊗ S is a generic vector in V and (Spin14 )v has
dimension 28.
If one constructs on a computer the representation V of the Lie algebra spin14
over a finite field F of characteristic 2, then it is a matter of linear algebra to
calculate the dimension of the stabilizer (spin14 )x of a random vector x ∈ V . One
finds for some x that the stabilizer has dimension 28, which is the minimum possible,
so by semicontinuity of dimension dim((spin14 )v ) = 28 = dim(G2 × G2 ). That
is, (Spin14 )v is smooth with identity component G2 × G2 . Consequently we may
compute (Spin14 )v by determining its K-points for K an algebraic closure of k. The
map Spin14 (K) → SO14 (K) is an isomorphism of concrete groups. The normalizer
of (G2 × G2 )(K) in the latter group is (G2 × G2 )(K) ⋊ Z/2, where the nonidentity
element τ ∈ Z/2 interchanges the two copies of SO7 (K), hence of G2 (K). As τ
normalizes (G2 × G2 )(K), it leaves the fixed subspace S ⊗ S ⊗ K = Kv invariant,
and we find a homomorphism χ : Z/2 → Gm given by τ v = χ(τ )v which must be
trivial because char K = 2.
The above proof, which is somewhat longer than some alternatives, was chosen
because of the details it provides on the embedding of G2 × G2 in Spin14 .
Proposition 9.2. The stabilizer in Spin13 (over a field of characteristic 2) of a
generic vector in the spin representation is the group scheme (SL2 × SL2 ) ⋊ Z/2.
Proof. We imitate the argument used in §8. View Spin13 as (Spin14 )y for an
anisotropic y in the 14-dimensional vector representation of Spin14 . That representation, as a representation of Spin13 , has socle ky and radical y ⊥ . Let v be a
generic element of the spin representation V of Spin13 . Our task is to determine
the group
(9.3)
(Spin13 )v = (Spin14 )y ∩ (Spin14 )v .
The stabilizer (Spin14 )v described above is contained in a copy (Spin14 )e of
Spin13 where ke is the radical of the 13-dimensional quadratic form given by the
pushout construction. As v is generic, y and e are in general position, so tracing
through the pushout construction we see that the intersection (9.3) contains the
product of 2 copies of the stabilizer in G2 of a generic octonion z. The quadratic
étale subalgebra of C generated by z has normalizer SL3 ⋊Z/2 in G2 , hence the
stabilizer of z is SL3 . We conclude that, for K an algebraic closure of k, the group
of K-points of (Spin13 )v equals that of the claimed group, hence the stabilizer has
dimension 16. Calculating with a computer as in the proof for Spin14 , we find that
dim(spin13 )v ≤ 16, and therefore the stabilizer of v is smooth as claimed.
Appendix A. Generic stabilizers associated with a peculiar half-spin
representation
by Alexander Premet
14
SKIP GARIBALDI AND ROBERT M. GURALNICK
A.0. The main theorem. Throughout this appendix we work over an algebraically
closed field k of characteristic 2. Let G = HSpin16 (k) and let V be the natural (halfspin) G-module. The theorem stated below describes the generic stabilizers for the
actions of G and g = Lie(G) on V .
Theorem A.1. The following are true:
(i) There exists a non-empty Zariski open subset U in V such that for every
x ∈ U the stabilizer Gx is isomorphic to (Z/2Z)4 .
(ii) For any x ∈ U the stabilizer gx is a 4-dimensional toral subalgebra of g.
(iii) If x, x′ ∈ U then the stabilizers Gx and Gx′ and the infinitesimal stabilizers
gx and gx′ are G-conjugate.
(iv) The scheme-theoretic stabilizer of any x ∈ U is isomorphic to (Z/2Z)4 ×
(µ2 )4 .
A more precise description of Gx and gx with x ∈ U is given in §A.4. It should
be mentioned here that our Theorem A.1 can also be deduced from more general
invariant-theoretic results recently announced by Eric Rains.
e be a simple algebraic
A.1. Preliminary remarks and recollections. Let G
e
group of type E8 over k and g̃ = Lie(G). The Lie algebra g̃ is simple and carries an (Ad G)-equivariant [p]-th power map x 7→ x[p] . Since p = 2, Jacobson’s
formula for [p]-th powers is surprisingly simple: we have that
(x + y)[2] = x[2] + y [2] + [x, y]
(∀ x, y ∈ g̃) .
e and t = Lie(T ). Write Φ̃ for the root system of
Let T be a maximal torus of G
e
G with respect to T . In what follows we will make essential use of Bourbaki’s
description of roots in Φ̃; see [Bou02, Planche VII]. More precisely, let E be an
8-dimensional Euclidean space over R with orthonormal basis {ε1 , . . . , ε8 }. Then
Φ̃ = Φ̃0 ⊔ Φ̃1 where
Φ̃0 = {±εi ± εj | 1 ≤ i < j ≤ 8}
and
Φ̃1 =
n P
8
1
2
ν(i)
εi |
i=1 (−1)
o
ν(i)
∈
2Z
.
i=1
P8
The roots α1 = 21 (ε1 + ε8 − ε2 − ε3 − ε4 − ε5 − ε6 − ε7 ), α2 = ε1 + ε2 , α3 = ε2
α4 = ε3 − ε2 , α5 = ε4 − ε3 , α6 = ε5 − ε4 , α7 = ε6 − ε5 , α8 = ε7 − ε6 form a
− ε1 ,
basis
of simple roots in Φ̃ which we denote by Π̃. Let ( · | · ) be the scalar product of E.
It is invariant under the action of the Weyl group W (Φ̃) ⊂ GL(E).
e and a
Given α ∈ Φ̃ we denote by Uα and eα the unipotent root subgroup of G
root vector in Lie(Uα ). Let V be the k-span of of all eα with α ∈ Φ̃1 and write G for
e generated by T and all Uα with α ∈ Φ̃0 . It is well known (and
the subgroup of G
straightforward to see) that the algebraic k-group G is isomorphic to HSpin16 (k)
and the G-stable subspace V of g̃ is isomorphic to the natural (half-spin) G-module:
one can choose a Borel subgroup B of G in such a way that the fixed-point space
V Ru (B) is spanned by e−α1 . We write W for the subgroup of W (Φ̃) generated all
orthogonal reflections sα with α ∈ Φ̃0 . Clearly, W ∼
= NG (T )/T is the Weyl group
of G relative to T . Since G has type D8 the group W is a semidirect product
of its subgroup W0 ∼
= S8 acting by permutations of the set {ε1 , . . . , ε8 } and its
abelian normal subgroup A ∼
= (Z/2Z)7 consisting of all maps εi 7→ (±1)i εi with
Q8
i=1 (±1)i = 1; see [Bou02, Planche IV].
SPINORS AND ESSENTIAL DIMENSION
15
We may (and will) assume further that the eα ’s are obtained by base change
from a Chevalley Z-form, g̃Z , of a complex Lie algebra of type E8 . Since the group
e is a simply connected the nonzero elements hα := [eα , e−α ] ∈ t with α ∈ Φ̃ span
G
t. They have the property that [hα , e±α ] = ±2e±α = 0 and hα = h−α for all α ∈ Φ̃.
[2]
[2]
It is well known that eα = 0 and hα = hα for all α ∈ Φ̃. The set {hα | α ∈ Π̃} is a
k-basis of t. Since g̃ is a simple Lie algebra, for every nonzero t ∈ t there is a simple
root β ∈ Π̃ such that (dβ)e (t) 6= 0. This implies that t admits a non-degenerate
W (Φ̃)-invariant symplectic bilinear form h · , · i such that hhα , hβ i = (α|β) mod 2
for all α, β ∈ Φ̃.
A.2. Orthogonal half-spin roots and Hadamard–Sylvester matrices. Following the Wikipedia webpage on Hadamard matrices we define the matrices H2k
of order 2k , where k ∈ Z≥0 , by setting H1 = [1] and
H2k
H2k
H2k+1 =
= H2 ⊗ H2k
H2k −H2k
for k ≥ 0. These Hadamard matrices were first introduced by Sylvester in 1867 and
they have the property that H2k · H2Tk = 2k · I2k for all k. We are mostly interested
in
1
1
1
1
1
1
1
1
1 −1
1 −1
1 −1
1 −1
1
1 −1 −1
1
1 −1 −1
1 −1 −1
1
1 −1 −1
1
.
H 8 = H 2 ⊗ H2 ⊗ H 2 =
1
1
1 −1 −1 −1 −1
1
1 −1
1 −1 −1
1 −1
1
1
1 −1 −1 −1 −1
1
1
1 −1 −1
1 −1
1
1 −1
To each row ri = (ri1 , . . . , ri8 ) of H8 we assign the root γi = 21 (ri1 ε1 + · · · + ri8 ε8 ).
This way we obtain 16 distinct roots ±γ1 , . . . , ±γ8 in Φ̃1 with the property that
(γi |γj ) = 0 for all i 6= j. As ±γi ± γj 6∈ Φ̃ for i 6= j, the semisimple regular subgroup
e generated by T and all U±γi is connected and has type A8 . It is immediate
S of G
1
e that G ∩ S = NG (T ) ∩ NS (T ).
from the Bruhat decomposition in G
Using the explicit
form of the simple roots α1 , . . . , α8 it is routine to determine
the matrix M := (γi |αj ) 1≤i,j≤8 . It has the following form:
−1 1
0
0
0
0
0
0
0 0 −1
1 −1
1 −1
1
0 1
0 −1
0
1
0 −1
1 0 −1
0
1
0 −1
0
.
M =
0
0
0 −1
0
0
0 1
1 0 −1
1 −1
0
1 −1
1 1
0 −1
0
0
0
1
0 0 −1
0
1 −1
1
0
It is then straightforward to check
that M is row-equivalent over the integers to
M1 M2
a block-triangular matrix M ′ =
with M1 , M2 , M3 ∈ Mat4 (Z) and
O4 2M3
det(M1 ) = det(M3 ) = 1. From this it follows that γ1 , . . . , γ8 span E over R and
16
SKIP GARIBALDI AND ROBERT M. GURALNICK
hγ1 , . . . , hγ8 span a maximal (4-dimensional) totally isotropic subspace of the symplectic space t. We call it t0 .
A.3. A dominant morphism. Put Γ = {γ1 , . . . , γ8 } and let r denote the subspace
P8
of V spanned by eγ with γ ∈ ±Γ. If x = i=1 (λi eγi + µi e−γi ) ∈ r then Jacobson’s
P
8
formula shows that x[2] = i=1 (λi µi )hγi ∈ t0 . It follows that
P8
k+1
k
(A.2)
x[2]
= i=1 (λi µi )2 hγi
∀k ≥ 0 .
Our discussion at the end of §A.2 shows that t0 has a basis t1 , . . . , t4 contained in
[2]
[2]
the F2 -span of {hγ | γ ∈ Γ}. Since hα = hα for all roots α, we have that ti = ti
for 1 ≤ i ≤ 4. In view of (A.2) this yields that the subset of r consisting of all x as
k
above such that λi µi 6= λj µj for i 6= j and {x[2] | 1 ≤ k ≤ 4} spans t0 is non-empty
and Zariski open in r. We call this subset r◦ and consider the morphism
ψ : G × r −→ V, (g, x) 7→ (Ad g) · x.
Note that dim(G × r) = 120 + 16 = 136 and dim V = 128. By the theorem on fiber
dimensions of a morphism, in order to show that ψ is dominant it suffices to find a
point (g, x) ∈ G × r such that all components of ψ −1 ((Ad g) · x) containing (g, x)
have dimension ≤ 8.
We take x ∈ r◦ and g = 1Ge . Clearly, ψ −1 (x) ⊂ {(g, y) ∈ G × r | y ∈ (Ad G) · x}.
k
If (g, y) ∈ ψ −1 (x) then y ∈ r and (Ad g)−1 maps the k-span, t(x), of {x[2] | 1 ≤
k
k ≤ 4} onto the k-span, t(y), of {y [2] | 1 ≤ k ≤ 4}. As y [2] ∈ t0 and t0 is a
restricted subalgebra of t, this implies that t(x) = t(y) = t0 . It follows that Ad g
preserves the Lie subalgebra cg (t0 ) of g. The centralizer cg̃ (t0 ) is spanned by t and
all root vectors eα such that hhα , hγi i = (dα)e (hγi ) = 0 for 1 ≤ i ≤ 8. As t0 is a
maximal totally isotropic subspace of the symplectic space t, our concluding remark
in §A.2 shows that cg̃ (t0 ) = Lie(S). Since cg (t0 ) = g ∩ Lie(S) = t we obtain that
g ∈ NG (T ). But then ψ −1 (x) ⊆ {(g, (Ad g)−1 · x) ∈ G × r◦ | g ∈ NG (T )}. Since
dim NG (T ) = dim T = 8, all irreducible components of ψ −1 (x) have dimension ≤ 8.
We thus deduce that the morphism ψ is dominant. As the set G × r◦ is Zariski
open in G × r, the G-saturation of r◦ in V contains a Zariski open subset of V .
P8
P8
A.4. Generic stabilizers. Let x = i=1 λi eγi + i=1 µi e−γi ∈ r◦ . In view of
our discussion in §A.3 we now need to determine the stabilizer Gx . If g ∈ Gx then
i
Ad g fixes t0 = span{x[2] | 1 ≤ i ≤ 4} and hence preserves cg (t0 ) = t. This yields
Gx ⊆ NG (T ). Working over a field of characteristic 2 has some advantages: after
reduction modulo 2 we are no longer affected by the ambiguity in the choice of a
Chevalley basis in g̃Z and the torus T has no elements of order 2. It follows that
NGe (T ) contains a subgroup isomorphic W (Φ̃) which intersects trivially with T . In
the notation of [Ste68, §3] this group is generated by all elements ωα = wα (1) with
α ∈ Φ̃. As a consequence, W embeds into NG (T ) in such a way that NG (T ) =
W ⋉ T.
Our discussion in §A.2 implies that for any α ∈ Π̃ the element 16α ∈ ZΦ̃ lies
e is a group
in the Z-span of γ1 . . . , γ8 . Since T has no elements of order 2 and G
× 8
of adjoint type, it follows that for any collection (t1 , . . . , t8 ) ∈ (k ) there exists a
unique element h = h(t1 , . . . , t8 ) ∈ T with γi (h) = ti for all 1 ≤ i ≤ 8. Conversely,
ex ∩ T = {1 e }. For 1 ≤ i ≤ 8
any element of T has this form. As a consequence, G
G
we set hi := h(1, . . . , µi /λi , . . . , 1), an element of T , where the entry µi /λi occupies
the i-th position. Since Ad sγi permutes e±γi and fixes e±γj with j 6= i, it is
SPINORS AND ESSENTIAL DIMENSION
17
ex . If w0 is the longest element of W (Φ̃)
straightforward to check that sγi hi ∈ G
Q
then it acts on ZΦ̃ as −Id and hence lies in A ⊂ W ֒→ NG (T ). Since w0 = 8i=1 sγi
Q8
we now deduce that n0 := w0
i=1 hi ∈ Gx .
ex ∩ N e (T ) contains an element n = wh, where w ∈ W (Φ̃) and h =
Suppose G
G
h(a1 , . . . , a8 ) ∈ T , such that w(γi ) = γj for i 6= j. Then n(eγi ) = ai eγj and
−1
n(e−γi ) = a−1
i e−γj implying that λj = λi ai and µj = µi ai . But then λj /λi =
µi /µj forcing λi µi = λj µj for i 6= j. Since x ∈ r◦ this is false. As n0 ∈ Gx and
e x ∩ N e (T ) = hni | 1 ≤ i ≤ 8i
w0 (±γi ) = ∓γi for all i, this argument shows that G
G
is isomorphic to an elementary abelian 2-group of order 28 .
k
Let A2k ∼
= (Z/2Z)2 denote the direct product of 2k copies of {±1} ∼
= Z/2Z. The
group operation in A2k is defined componentwise. We write u • v for the product
of u, v ∈ A2k and denote by 12n the identity element of A2k (all components of
12k are equal to 1). The set of rows, R2k , of the Hadamard–Sylvester matrix H2k
may be regarded as a subset A2k and easy induction on k shows that ±R2k is a
subgroup of A2k . In particular, ±R8 is a subgroup of A8 . As mentioned in §A.1
the subgroup W0 ∼
= S8 of the Weyl group W = W (Φ̃0 ) acts on A8 by permuting
components whereas the normal subgroup A ∼
= (Z/2Z)7 of W embeds into A8 and
acts on it by translations.
If n ∈ Gx then n = wh ∈ NG (T ) and w preserves ±R8 setwise. If w = aσ,
where σ ∈ W0 and a ∈ A, then our discussion in the previous paragraph shows
that w(u) = (aσ)(u) = ±u for all u ∈ ±R8 . Taking u = 18 we get σ(18 ) = 18 and
±18 = w(18 ) = a • σ(18 ) = a • 18 = a. This yields a = ±18 implying that w ∈ W0
preserves ±R8 . Also, Gx ∩ A is a cyclic group of order 2 generated by n0 .
We now consider three commuting involutions σ1 = (1, 5)(2, 6)(3, 7)(4, 8), σ2 =
(1, 4)(2, 3)(5, 8)(6, 7) and σ3 = (1, 2)(3, 4)(5, 6)(7, 8) in W ∼
= S8 . One can see by
inspection that each of them maps every r ∈ R8 to ±r. Hence σi ∈ hsγi | 1 ≤ i ≤ 8i.
e x for 1 ≤ i ≤ 8, each σi admits a unique lift in Gx ⊂ NG (T ) which
Since sγi hi ∈ G
will be denoted by ni . The subgroup hni | 0 ≤ i ≤ 3i of Gx is isomorphic to (Z/2Z)4 .
Next we show that any element σh ∈ Gx with σ ∈ W0 ∼
= S8 lies in the subgroup
generated by the ni ’s. Since w maps 18 to ±18 and n0 ∈ Gx we may assume that
w(18 ) = 18 . Since σ maps (14 , −14 ) to ±(14 , −14 ) and n1 ∈ Gx we may also assume
that σ fixes (14 , −14 ). Since σ maps (12 , −12 , 12 , −12 ) to ±(12 , −12 , 12 , −12 ) and
n2 ∈ Gx we may assume that σ fixes (12 , −12 , 12 , −12 ) as well. Finally, since
σ maps (1, −1, 1, −1, 1, −1, 1, −1) to ±(1, −1, 1, −1, 1, −1, 1, −1) and n3 ∈ Gx we
may assume that σ fixes (1, −1, 1, −1, 1, −1, 1, −1). This entails that σ(i) = i for
i ∈ {1, 2, 3, 4}. As σ(r) = ±r for all r ∈ R8 the latter shows that σ = id proving
statement (i) of Theorem A.1.
i
Since gx contains the spanning set {x[2] | 1 ≤ i ≤ 4} of t0 , our remarks in §A.3
show that gx ⊂ t. Since [t, x] = 0 for every t ∈ gx it must be that (dγ)e (t) = 0 for
all γ ∈ Γ. Since (dγ)e (t) = hhγ , ti and t0 is a maximal isotropic subspace of the
symplectic space t, we obtain that t ∈ t0 . As a result, gx = t0 for every x ∈ r◦ .
Statement (ii) follows.
P8
In proving statement (iii) we may assume that x = i=1 (λi eγi + µi e−γi ) and
P8
x′ = i=1 (λ′i eγi + µ′i e−γi ) are two elements of r◦ . Our discussion in the previous
paragraph shows that gx = gx′ = t0 . Let h′i := h(1, . . . , µ′i /λ′i , . . . , 1), where the
entry µ′i /λ′i occupies the i-th position. There is a unique element h = h(b1 , . . . , b8 ) ∈
18
SKIP GARIBALDI AND ROBERT M. GURALNICK
T such that
(1 ≤ i ≤ 8).
h · sγi hi · h−1 = sγi h′i
p
′
′
(We need to take bi = (λi µi )/(λi µi ) ∈ k for all 1 ≤ i ≤ 8.) Our earlier remarks
in this section now show that h · Gx · h−1 = Gx′ . This proves statement (iii).
P8
Remark. We stress that for an element x = i=1 (λi eγi + µi e−γi ) to be in r◦ it is
necessary that λi µi 6= λj µj for all i 6= j. If one removes this condition and only
i
requires that the set {x[2] | 1 ≤ i ≤ 4} ⊂ t is linearly independent, then one obtains
an a priori bigger Zariski open subset, r′ , in r which still has the property that Gx
is a finite group and gx = t0 for every x ∈ r′ . However, it is not immediately clear
that the stabilizers in G of any two elements in r′ are isomorphic. It would be
interesting to investigate this situation in more detail.
e be a reductive group scheme over k with
Scheme-theoretic stabilizers. Let G
e and let G be the regular
root system Φ̃ with respect to a maximal torus T ⊂ G
e
group subscheme of G with root system Φ̃0 . We may assume that T(k) = T ,
e
e and G(k) = G. In this situation, we wish to describe the schemeG(k)
= G,
theoretic stabilizer Gx of x ∈ r◦ , an affine group subscheme of G defined over
k.
Let F be any commutative associative k-algebra with 1. The subscheme NG (T)
of G is smooth and since p = 2 we have an isomorphism NG (T) = W × T of affine
group schemes over k. Arguing as in §A.4 one observes that Gx (F ) is contained in
the group of F -points of NG (T). Since the latter contains Gx = Gx (k) it follows
that Gx (F ) is generated by Gx = (Gx )red and the scheme-theoretic stabilizer Tx .
More precisely,
Gx ∼
= (Z/2Z)4 × Tx
= (Gx )red × Tx ∼
as affine group schemes over k. Our concluding remarks in §A.2 imply that the root
lattice ZΦ̃ contains free Z-submodules Λ1 and Λ2 of rank 4 such that ZΦ̃ = Λ1 ⊕ Λ2
and ZΓ := Zγ1 ⊕ · · · ⊕ Zγ8 = Λ1 ⊕ 2Λ2 . Since T(F ) = HomZ (ZΦ̃, F × ), we have a
short exact sequence
1 → HomZ (Λ2 /2Λ2 , F × ) → T(F ) → HomZ (ZΓ, F × ) → 1
which shows that the groups Tx (F ) and HomZ (Λ2 /2Λ2 , F × ) are isomorphic. Since
Λ2 /2Λ2 ∼
= (Z/2Z)4 and HomZ (Z/2Z, F × ) = µ2 (F ) we have HomZ (Λ2 /2Λ2 , F × ) ∼
=
4
(µ2 ) (F ). Consequently, Tx ∼
= (µ2 )4 as affine group schemes over k. This completes
the proof of Theorem A.1.
References
[AP71]
[AS76]
[BM12]
[Bou02]
[BR97]
[BRV10]
E.M. Andreev and V.L. Popov, Stationary subgroups of points of general position in
the representation space of a semisimple Lie group, Functional Anal. Appl. 5 (1971),
no. 4, 265–271.
M. Aschbacher and G. Seitz, Involutions in Chevalley groups over fields of even order,
Nagoya Math. J. 63 (1976), 1–91.
Sanghoon Baek and Alexander S. Merkurjev, Essential dimension of central simple
algebras, Acta Math. 209 (2012), no. 1, 1–27. MR 2979508
N. Bourbaki, Lie groups and Lie algebras: Chapters 4–6, Springer-Verlag, Berlin,
2002.
J. Buhler and Z. Reichstein, On the essential dimension of a finite group, Compositio
Math. 106 (1997), 159–179.
P. Brosnan, Z. Reichstein, and A. Vistoli, Essential dimension, spinor groups, and
quadratic forms, Ann. of Math. (2) 171 (2010), no. 1, 533–544.
SPINORS AND ESSENTIAL DIMENSION
[BS66]
19
A. Borel and T.A. Springer, Rationality properties of linear algebraic groups, Algebraic Groups and Discontinuous Subgroups (Proc. Sympos. Pure Math., Boulder,
Colo., 1965), Amer. Math. Soc., Providence, R.I., 1966, pp. 26–32.
[CM93]
D. Collingwood and W.M. McGovern, Nilpotent orbits in semisimple Lie algebras,
Van Nostrant Reinhold, New York, 1993.
[CM14]
V. Chernousov and A.S. Merkurjev, Essential dimension of spinor and Clifford
groups, Algebra & Number Theory 8 (2014), no. 2, 457–472.
[CS06]
V. Chernousov and J-P. Serre, Lower bounds for essential dimensions via orthogonal
representations, J. Algebra 305 (2006), 1055–1070.
[DG70]
M. Demazure and A. Grothendieck, Schémas en groupes II: Groupes de type multiplicatif, et structure des schémas en groupes généraux, Lecture Notes in Mathematics,
vol. 152, Springer, 1970.
[FGS16]
J. Fulman, R. Guralnick, and D. Stanton, Asymptotics of the number of involutions
in finite classical groups, arxiv:1602.03611, 2016.
[Flo08]
M. Florence, On the essential dimension of cyclic p-groups, Invent. Math. 171 (2008),
no. 1, 175–189. MR 2358058 (2008i:12006)
[Gar98]
S. Garibaldi, Isotropic trialitarian algebraic groups, J. Algebra 210 (1998), 385–418.
, Cohomological invariants: exceptional groups and spin groups, Memoirs
[Gar09]
Amer. Math. Soc., no. 937, Amer. Math. Soc., Providence, RI, 2009, with an appendix by Detlev W. Hoffmann.
[GG15a]
S. Garibaldi and R.M. Guralnick, Essential dimension of exceptional groups, including
bad characteristic, preprint, 2015, arxiv:1512.09158.
[GG15b]
, Simple groups stabilizing polynomials, Forum of Mathematics: Pi 3 (2015),
e3 (41 pages).
[GLLT16] R.M. Guralnick, R. Lawther, M. Liebeck, and D. Testerman, Generic stabilizers for
actions of simple algebraic groups, in preparation, 2016.
[GLMS97] R.M. Guralnick, M.W. Liebeck, D. Macpherson, and G.M. Seitz, Modules for algebraic
groups with finitely many orbits on subspaces, J. Algebra 196 (1997), 211–250.
[GR09]
P. Gille and Z. Reichstein, A lower bound on the essential dimension of a connected
linear group, Comment. Math. Helv. 84 (2009), no. 1, 189–212.
[Gue97]
M. Guerreiro, Exceptional representations of simple algebraic groups in prime characteristic, Ph.D. thesis, University of Manchester, 1997, arxiv:1210.6919.
[GV78]
V. Gatti and E. Viniberghi, Spinors of 13-dimensional space, Adv. in Math. 30 (1978),
no. 2, 137–155.
[Igu70]
J.-I. Igusa, A classification of spinors up to dimension twelve, Amer. J. Math. 92
(1970), 997–1028.
[KM03]
N. Karpenko and A. Merkurjev, Essential dimension of quadrics, Invent. Math. 153
(2003), no. 2, 361–372. MR 1992016 (2004f:11029)
, Essential dimension of finite p-groups, Invent. Math. 172 (2008), no. 3, 491–
[KM08]
508. MR 2393078 (2009b:12009)
[KMRT98] M.-A. Knus, A.S. Merkurjev, M. Rost, and J.-P. Tignol, The book of involutions,
Colloquium Publications, vol. 44, Amer. Math. Soc., 1998.
[Lie87]
M.W. Liebeck, The affine permutation groups of rank 3, Proc. London Math. Soc. 54
(1987), 477–516.
[LMMR13] Roland Lötscher, Mark MacDonald, Aurel Meyer, and Zinovy Reichstein, Essential
dimension of algebraic tori, J. Reine Angew. Math. 677 (2013), 1–13. MR 3039772
[Löt13]
R. Lötscher, A fiber dimension theorem for essential and canonical dimension, Compositio Math. 149 (2013), no. 1, 148–174.
[LS12]
M. Liebeck and G. Seitz, Unipotent and nilpotent classes in simple algebraic groups
and Lie algebras, Math. Surveys and Monographs, no. 180, Amer. Math. Soc., 2012.
[Mer09]
A. Merkurjev, Essential dimension, Quadratic forms–algebra, arithmetic, and geometry (R. Baeza, W.K. Chan, D.W. Hoffmann, and R. Schulze-Pillot, eds.), Contemp.
Math., vol. 493, 2009, pp. 299–325.
, Essential p-dimension of PGL(p2 ), J. Amer. Math. Soc. 23 (2010), no. 3,
[Mer10]
693–712.
[Mer13]
, Essential dimension: a survey, Transform. Groups 18 (2013), no. 2, 415–481.
, Essential dimension, Séminaire Bourbaki, 67ème année, #1101, June 2015.
[Mer15a]
20
[Mer15b]
[Pop80]
[Pop88]
[Rei10]
[Röh93]
[Ros99a]
[Ros99b]
[RY00]
[SF88]
[Ste68]
SKIP GARIBALDI AND ROBERT M. GURALNICK
, Invariants of algebraic groups and retract rationality of classifying spaces,
preprint, 2015.
V.L. Popov, Classification of spinors of dimension 14, Trans. Moscow Math. Soc.
(1980), no. 1, 181–232.
A.M. Popov, Finite isotropy subgroups in general position of simple linear Lie groups,
Trans. Moscow Math. Soc. (1988), 205–249, [Russian original: Trudy Moskov. Mat.
Obschch. 50 (1987), 209–248, 262].
Z. Reichstein, Essential dimension, Proceedings of the International Congress of
Mathematicians 2010, World Scientific, 2010.
G. Röhrle, On certain stabilizers in algebraic groups, Comm. Algebra 21 (1993), no. 5,
1631–1644.
M. Rost, On 14-dimensional quadratic forms, their spinors, and the difference of two
octonion algebras, unpublished note, March 1999.
, On the Galois cohomology of Spin(14), unpublished note, March 1999.
Z. Reichstein and B. Youssin, Essential dimensions of algebraic groups and a resolution theorem for G-varieties, Canad. J. Math. 52 (2000), no. 5, 1018–1056, with an
appendix by J. Kollár and E. Szabó.
H. Strade and R. Farnsteiner, Modular Lie algebras and their representations, Monographs and textbooks in pure and applied math., vol. 116, Marcel Dekker, New York,
1988.
R. Steinberg, Lectures on Chevalley groups, Yale University, New Haven, Conn., 1968.
| 4math.GR
|
Learning activation functions from data
using cubic spline interpolation
Simone Scardapane, Michele Scarpiniti, Danilo Comminiello, and Aurelio Uncini
arXiv:1605.05509v2 [stat.ML] 11 May 2017
Department of Information Engineering, Electronics and Telecommunications (DIET),
“Sapienza” University of Rome,
Via Eudossiana 18, 00184, Rome.
Email: {simone.scardapane, michele.scarpiniti, danilo.comminiell}@uniroma1.it;
[email protected]
Abstract. Neural networks require a careful design in order to perform properly
on a given task. In particular, selecting a good activation function (possibly in
a data-dependent fashion) is a crucial step, which remains an open problem in
the research community. Despite a large amount of investigations, most current
implementations simply select one fixed function from a small set of candidates,
which is not adapted during training, and is shared among all neurons throughout the different layers. However, neither two of these assumptions can be supposed optimal in practice. In this paper, we present a principled way to have
data-dependent adaptation of the activation functions, which is performed independently for each neuron. This is achieved by leveraging over past and present
advances on cubic spline interpolation, allowing for local adaptation of the functions around their regions of use. The resulting algorithm is relatively cheap to
implement, and overfitting is counterbalanced by the inclusion of a novel damping criterion, which penalizes unwanted oscillations from a predefined shape.
Preliminary experimental results validate the proposal.
Keywords: Neural network, activation function, spline interpolation
1
Introduction
Neural networks (NNs) are extremely powerful tools for approximating complex nonlinear functions [7]. The nonlinear behavior is introduced in the NN architecture by the
elementwise application of a given nonlinearity, called the activation function (AF), at
every layer. Since AFs are crucial to the dynamics and computational power of NNs,
the history of the two over the last decades is deeply connected [15]. As an example,
the use of differentiable AFs was one of the major breakthroughs in NNs, leading directly to the back-propagation algorithm. More recently, progress on piecewise linear
functions was shown to facilitate backward flow of information for training very deep
networks [4]. At the same time, it is somewhat surprising that the vast majority of NNs
only use a small handful of fixed functions, to be hand-chosen by the practitioner before the learning process. Worse, there is no principled reason to believe that a ‘good’
nonlinearity might be the same across all layers of the network, or even across neurons
in the same layer.
2
Scardapane et al.
This is shown clearly in a recent work by Agostinelli et al. [1], where every neuron
in a deep network was endowed with an adaptable piecewise linear function with possibly different parameters, concluding that “the standard one-activation-function-fits-all
approach may be suboptimal” in current practice. Experiments in AF adaptation have a
long history, but they have never met a wide applicability in the field. The simplest approach is to parameterize each sigmoid function in the network by one or more ‘shape’
parameters to be optimized, such as in the seminal 1996 paper by Chen and Chang
[3] or the later work by Trentin [16]. Along a similar line, one may consider the use
of polynomial AFs, wherein each coefficient of the polynomial is adapted by gradient
descent [11]. Additional investigations can be found in [20,5,2,10,9]. One strong drawback of these approaches is that the parameters involved affect the AF globally, such
that a change in one region of the function may be counterproductive on a different,
possibly faraway, region.
Several years ago, an alternative approach was introduced by using spline interpolating functions as AFs [17,6], resulting in what was called a spline AF (SAF). Splines
are an attractive choice for interpolating unknown functions, since they can be described
by a small amount of parameters, yet each parameter has a local effect, and only a fixed
number of them is involved every time an output value is computed [18]. The original
works in [17,6] had two main drawbacks that prevented a wider use of the underlying theory. First, SAFs were only investigated in an online setting, where updates are
computed one sample at a time. Whether an efficient implementation is possible (and
feasible) also for batch (or mini-batch) settings was not shown. Secondly, the obtained
SAFs had a tendency to overfit training data, resulting in oscillatory behaviors which
hindered performance. Inspired by recent successes in the field of nonlinear adaptive
filtering [13,14], our aim in this paper is two-fold. On one hand, we provide a modern
introduction to the use of SAFs in neural networks, with a particular emphasis on their
efficient implementation in the case of batch (or mini-batch) training. Our treatment
clearly shows that the major problem in their implementation, which is evident from
the discussion above, is the design of an efficient way to regularize their control points.
In this sense, as a second contribution we provide a simple (yet effective) ‘damping’ criterion to prevent unwanted oscillations in the testing phase, which penalizes deviations
from the original points in terms of `2 norm. A restricted set of experiments shows that
the resulting formulation is able to achieve a lower test error than a standard NN with
fixed AFs, while at the same time learning non-trivial activations with different shapes
across different neurons.
The rest of the paper is organized as follows. Section 2 presents the basic theory of
SAFs for the case of a single neuron. Section 3 extends the treatment to the case of a NN
with one hidden layer, by deriving the gradient equations for the SAFs parameters in the
internal layer. Then, Section 4 goes over the experimental results, while we conclude
with some final remarks in Section 5.
2
The spline activation function
We begin our treatment of SAFs with the simplest case of a single neuron endowed
with a flexible AF (see [17,13] for additional details). Given a generic input x ∈ RD , the
Learning activation functions from data
3
output of the SAF is computed as:
s = wT x ,
(1)
y = ϕ(s; q) ,
(2)
where w ∈ RD (we suppose that an eventual bias term is added directly to the input vector), and the AF ϕ(·) is parameterized by a vector q ∈ RQ of internal parameters, called
knots. The knots are a sampling of the AF values over Q representative points spanning
the overall function. In particular, we suppose the knots to be uniformly spaced, i.e.
qi+1 = qi + ∆ x, for a fixed ∆ x ∈ R, and symmetrically spaced around the origin. Given
s, the output is computed by spline interpolation over the closest knot and its P rightmost neighbors. The common choice P = 3, which we adopt in this paper, corresponds
to cubic interpolation, and it is generally a good trade-off between locality of the output
and interpolating precision.
Given the index i of the closest knot, we can define the normalized abscissa value
between qi and qi+1 as:
j s k
s
.
(3)
u=
−
∆x
∆x
where b·c is the floor operator. From u we can compute the normalized reference
T
vector u = uP uP−1 . . . u 1 , while from i we can extract the relevant control points
qi = [qi qi+1 . . . qi+P ]T . We refer to the vector qi as the ith span. The output (2) is then
computed as:
(4)
y = ϕ(s) = uT Bqi ,
where B ∈ R(P+1)×(P+1) is called the spline basis matrix. In this work, we use the
Catmull-Rom (CR) spline with P = 3, given by:
−1 3 −3 1
1 2 −5 4 −1
.
(5)
B=
2 −1 0 1 0
0 2 0 0
Different bases give rise to alternative interpolation schemes, e.g. a spline defined by a
CR basis passes through all the control points, but its second derivative is not continuous.
Apart from the locality of the output, SAFs have two additional interesting properties. First, the output in (4) is extremely efficient to compute, involving only vectormatrix products of very small dimensionality. Secondly, derivatives with respect to the
internal parameters are equivalently simple and can be written down in closed form. In
particular, the derivative of the nonlinearity ϕ(s) with respect to the input s is given by:
∂ ϕ(s)
∂ ϕ(s) ∂ u
1
= ϕ 0 (s) =
·
=
u̇Bqi ,
(6)
∂s
∂u
∂s
∆x
where:
u̇ =
T
∂ u P−1
= Pu
(P − 1)uP−2 . . . 1 0 .
∂u
(7)
4
Scardapane et al.
Given this, the derivative of the SAF output y with respect to w is straightforward:
∂ ϕ(s)
∂s
= ϕ 0 (s) ·
= ϕ 0 (s)x ,
∂w
∂w
(8)
∂ ϕ(s)
= BT u .
∂ qi
(9)
Similarly, for qi we obtain:
while we have
3
3.1
∂ ϕ(s)
∂ qk
= 0 for any element qk outside the current span qi .
Designing networks with SAF neurons
Computing outputs and inner derivatives
Now we consider the more elaborate case of a single hidden layer NN, with a Ddimensional input, H neurons in the hidden layer, and O output neurons.1 Every neuron
in the network uses a SAF with possibly different adaptive control points, which are
set independently during the training process. For easiness of computation, we suppose
that the sampling set of the splines is the same for every neuron (i.e., each neuron has
Q points equispaced according to the same ∆ x), and we also have a single shared basis
matrix B. The forward phase of the network is similar to that of a standard network.
In particular, given the input x, we first compute the output of the ith hidden neuron,
i = 1, . . . , H, as:
hi = ϕ(wThi x; qhi ) .
(10)
These are concatenated in a single vector h = [h1 , . . . , hH , 1]T , and the ith output of the
network, i = 1, . . . , O, is given by:
fi (x) = yi = ϕ(wTyi h; qyi ) .
(11)
The derivatives with respect to the parameters wyi , qyi , i = 1, . . . , O can be computed
directly with (8)-(9), substituting x with h. By back-propagation, the derivative of the
ith output with respect to the jth (inner) weight vector wh j is similar to a standard NN:
∂yj
= ϕ 0 (sy j ) · ϕ 0 (shi ) · bwhi c j · x ,
∂ whi
(12)
where with a slight abuse of notation we let sy j denote the activation of the jth output
(and similarly for shi ), b·c j extracts the jth element of its input vector, and the two ϕ 0 (·)
are given by (6). For the derivative of the control points of the ith hidden neuron, denote
by qhi ,k the currently active span, and by uhi the corresponding reference vector. The
derivative with respect to the jth output is then given by:
∂yj
= ϕ 0 (sy j ) · bwhi c j · BT uhi .
∂ qhi ,k
1 We
(13)
note that the following treatment can be extended easily to the case of a network with
more than one hidden layer. However, restricting it to a single layer allow us to keep the discussion
focused on the problems/advantages arising in the use of SAFs. We leave this extension to a future
work.
Learning activation functions from data
3.2
5
Initialization of the control points
An important aspect that we have not discussed yet is how to properly initialize the
control points. One immediate choice is to sample their values from an AF which is
known to work well on the given problem, e.g. a hyperbolic tangent. In this way, the
network is guaranteed to work similarly to a standard NN in the initial phase of learning.
Additionally, we have found good improvements in error by adding Gaussian noise
N (0, σ 2 ) with small variance σ 2 to a randomly chosen subset of control points (around
5% in our experiments). This provides a good variability in the beginning, similarly to
how connections are set close to (but not identically equal to) zero during initialization.
3.3
Choosing a training criterion
Suppose we are provided with a training set of N input/output pairs in the form {xi
, di }Ni=1 .
For simplicity
of notation, we denote by w the concatenation of all weight vectors whi
and wyi , and by q a similar concatenation of all control points. Training can be formulated as the minimization of the following cost function:
J(w, q) =
1 N
∑ L(di , f(xi )) + λw Rw (w) + λq Rq (q) ,
N i=1
(14)
where L(·, ·) is an error function, while Rw (·) and Rq (·) provide meaningful regularization on the two set of parameters. The first two terms are well-known in the neural
network literature [7], and they can be set accordingly. Particularly, in our experiments
we consider a squared error term L(di , f(xi )) = kdi − f(xi )k22 , and `2 regularization on
the weights Rw (w) = kwk22 . The derivatives of L(·, ·) can be computed straightforwardly
with the formulas presented in Section 3.1.
The term Rq (q) is used to avoid overfitted solutions for the control points. In fact,
its presence is the major difference with respect to previous attempts at implementing
SAFs in neural networks [17], wherein overfitting was counterbalanced by choosing a
large value for ∆ x, which in a way goes outside the philosophy of spline interpolation
itself. At the same time, choosing a proper form for the regularization term is non-trivial,
as the term should be cheap to compute, and it should introduce just as much a priori
information as needed, without hindering the training process. Most of the literature
on regularizing w cannot be used here, as the corresponding formulations do not make
sense in the context of spline interpolation. As an example, simply penalizing the `2
norm of q leads to functions close to the zero function, while imposing sparsity is also
meaningless.
For the purpose of this paper, we consider the following ‘damping’ criterion:
Rq (q) = kq − qo k22 ,
(15)
where qo represents the initial values for the control points, as discussed in the previous
section (without considering additional noise). The criterion makes intuitive sense as
follows: while for w we wish to penalize unwanted deviations from very small weights
(which can be justified with arguments from learning theory), in the case of q we are
6
Scardapane et al.
interested in penalizing changes with respect to a ‘good’ function parameterized by the
initial control points qo , namely one of the standard AFs used in NN training. In fact,
setting a value for λq very high essentially deactivates the adaptation of the control
points. Clearly, other choices are possible, and in this sense this paper serves as a starting point for further investigations towards this objective. As an example, we may wish
to penalize first (or second) order derivatives of the splines in order to force a desired
level of smoothness [18].
3.4
Remarks on the implementation
In order to be usable in practice, SAFs require an efficient implementation to compute
outputs and derivatives concurrently for the entire training dataset or, alternatively, for a
properly chosen mini-batch (in the case of stochastic optimization algorithms). To begin
with, we underline that the equations for the reference vector (see (3)) do not depend on
the specific neuron, and for this reason they can easily be vectorized layer-wise on most
numerical algebra libraries to obtain all vectors concurrently. Additionally, the indexes
and relative terms Bqi in (4) can be cached during the forward pass, to be reused during
the computation of the derivatives. In this sense, the outputs of a layer and its derivatives
can be computed by one 4×4 matrix-vector computation, and three 4-dimensional inner
products, which have to be repeated for every pair input/neuron. In our experience,
the cost of a relatively well-optimized implementation does not exceed twice that of a
standard network for medium-sized batches, where the most onerous operation is the
reshaping of the gradients in (9) and (13) into a single vector of gradients relative to the
global vector q.
4
4.1
Experimental results
Experimental setup
To evaluate the preliminary proposal, we consider two simple regression benchmarks
for neural networks, the ‘chemical’ dataset (included among MATLAB’s testbeds for
function fitting), and the ‘California Housing’.2 They have respectively 498 and 20640
examples, and 8 numerical features. Inputs are normalized in the [−1, +1] range, while
outputs are normalized in the [−0.5, +0.5] range. We compare a NN with 5 hidden neurons and tanh(·) AFs (denoted as ‘Standard’ in the results), and a NN with the same
number of neurons and SAF nonlinearities. The weight vector w is initialized with the
method described in [4]. Each SAF is initialized from a tanh(·) nonlinearity, and control
points are defined in the [−2, +2] range with ∆ x = 0.2, which is a good compromise
between locality of the SAFs and the overall number of adaptable parameters. For the
first scenario, λq is kept to a small value of 10−5 . For each experiment, a random 30%
of the dataset is kept for testing, and results are averaged over 15 different splits to average out statistical effects. Error is computed with the Normalized Root Mean-Squared
Error (NRMSE). The optimization problems are solved using a freely available MATLAB implementation of the Polack-Ribiere variant of the nonlinear conjugate gradient
2 http://www.dcc.fc.up.pt/
˜ltorgo/Regression/cal_housing.html
Learning activation functions from data
7
Table 1. Average results for scenario 1 (λw = 1), together with one standard deviation.
Dataset
Nonlinearity Tr. RMSE T.st NRMSE
Chemical
Standard
SAF
1.00 ± 0.00 1.00 ± 0.01
0.29 ± 0.02 0.31 ± 0.02
Calhousing
Standard
SAF
1.02 ± 0.00 1.01 ± 0.01
0.56 ± 0.01 0.57 ± 0.02
optimization algorithm by C.E. Rasmussen. [12].3 The optimization process is allowed
1500 maximum iterations. MATLAB code for the experiments is also available on the
web.4 We briefly remark that the MATLAB library, apart from repeating the experiments presented here, is also designed to handle networks with more than a single
hidden layer, and implements the ADAM algorithm [8] for stochastic training in case
of a larger dataset.
4.2
Scenario 1: strong underfitting
As a first example, we consider a scenario of strong underfitting, wherein the user has
misleadingly selected a very large value of λw = 1, leading in turn to extremely small
values for the elements of w after training. Results in terms of training and test RMSE
are provided in Tab. 1. Since the activations of the NN tend to be very close to 0 (where
the hyperbolic tangent operates in an almost-linear regime), standard NNs have a constant zero output, leading to a RMSE of 1. Nonetheless, SAF networks are able to reach
a very satisfactory level of performance, which in the first case is almost comparable to
that of a fully optimized network (see the following section).
To show the reasons for this, we have plotted four representative nonlinearities after
training in Fig. 1. It is easy to see that the nonlinearities have adapted to act as ‘amplifiers’ for the activations in their operating regime, with mild and strong peaks around 0.
Of particular interest is the fact that the resulting SAFs need not be perfectly centered
around 0 (e.g. Fig. 1c), or even symmetrical around the y-axis (e.g. Fig. 1d). In fact, the
splines are able to efficiently counterbalance a bad setting for the weights, with behaviors which would be very hard (or close to impossible) using standard setups with fixed,
shared, mild nonlinearities.
4.3
Scenario 2: well-optimized parameters
In our second scenario, we consider a similar comparison with respect to before, but
we fine-tune the parameters of the two methods using a grid-search with a 3-fold crossvalidation on the training data as performance measure. Both λw and λq (only for the
3 http://learning.eng.cam.ac.uk/carl/code/minimize/
4 [The
URL has been hidden for the review process.]
8
Scardapane et al.
1
Spline value
Spline value
1
0
-1
-1
-2
0
Activation s
2
-2
0
Activation s
2
-2
0
Activation s
2
2
Spline value
2
Spline value
0
1
0
-1
1
0
-1
-2
0
Activation s
2
Fig. 1. Non-trivial representative SAFs after training for scenario 1.
Table 2. Optimal parameters (averaged over the runs) found by the grid-search procedure for
scenario 2.
Dataset
Nonlinearity λw
Chemical
Standard
SAF
10−3 —
10−2 10−4
Calhousing
Standard
SAF
10−4 —
10−3 10−4
λq
proposed algorithm) are searched in an exponential interval 2 j , with j = −10, . . . , 5.
Optimal parameters found by the grid-search are listed in Table 2, while results in terms
of training and test NRMSE are collected in Table 3.
Overall, we see that the NNs endowed with the SAF nonlinearities are able to surpass by a large margin a standard NN, and the results from the previous scenario. The
only minor drawback evidenced in Table 3 is that the SAF network has some overfitting
occurring in the ‘chemical’ dataset (around 2 points of NRMSE), showing that there is
still some room for improvement in terms of spline optimal regularization.
Also in this case, we plot some representatives SAFs after training (taken among
those which are not trivially identical to the tanh nonlinearity) in Fig. 2. As before, in
general SAFs tend to provide an amplification (with a possible change of sign) of their
activation around some region of operation. It is interesting to observe that, also in this
case, the optimal shape need not be symmetric (e.g. Fig. 2a), and might even be far
Learning activation functions from data
9
Table 3. Average results for scenario 2 (fine-tuning for parameters), together with one standard
deviation.
Dataset
Nonlinearity Tr. RMSE T.st NRMSE
Chemical
Standard
SAF
0.32 ± 0.01 0.32 ± 0.02
0.26 ± 0.01 0.28 ± 0.02
Calhousing
Standard
SAF
0.55 ± 0.01 0.55 ± 0.01
0.51 ± 0.02 0.51 ± 0.02
0
-1
0
-1
-2
0
Activation s
2
0
Activation s
2
-1
0
-1
-2
0
Activation s
2
-2
0
Activation s
2
-2
0
Activation s
2
2
Spline value
Spline value
1
0
0
-1
-2
1
Spline value
1
Spline value
1
Spline value
Spline value
1
1
0
-1
-2
0
Activation s
2
Fig. 2. Non-trivial representative SAFs after training for scenario 2.
from centered around 0 (e.g. Fig. 2c). Resulting nonlinearities can also present some
additional non-trivial behaviors, such as a small region of insensibility around 0 (e.g.
Fig. 2d), or a region of pre-saturation before the actual tanh saturation (e.g. Fig.s 2e-2f).
5
Conclusion
In this paper, we have presented a principled way to adapt the activation functions of a
neural network from training data, locally and independently for each neuron. Particularly, each nonlinearity is implemented with cubic spline interpolation, whose control
points are adapted in the optimization phase. Overfitting is controlled by a novel `2 regularization criterion avoiding unwanted oscillations. Albeit efficient, this criterion does
constrain the shapes of the resulting functions by a certain degree. In this sense, the design of more advanced regularization terms is a promising line of research. Additionally,
10
Scardapane et al.
we plan on exploring the application of SAFs to deeper networks, where it is expected
that the statistics of the neurons’ activations can change significantly layer-wise [4].
References
1. Agostinelli, F., Hoffman, M., Sadowski, P., Baldi, P.: Learning activation functions to improve deep neural networks. arXiv preprint arXiv:1412.6830 (2014)
2. Chandra, P., Singh, Y.: An activation function adapting training algorithm for sigmoidal feedforward networks. Neurocomputing 61, 429–437 (2004)
3. Chen, C.T., Chang, W.D.: A feedforward neural network with function shape autotuning.
Neural networks 9(4), 627–641 (1996)
4. Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Int. conf. on artificial intell. and stat. pp. 249–256 (2010)
5. Goh, S., Mandic, D.: Recurrent neural networks with trainable amplitude of activation functions. Neural Networks 16(8), 1095–1100 (2003)
6. Guarnieri, S., Piazza, F., Uncini, A.: Multilayer feedforward networks with adaptive spline
activation function. IEEE Trans. Neural Netw. 10(3), 672–683 (1999)
7. Haykin, S.: Neural networks and learning machines. Pearson Education, 3rd edn. (2009)
8. Kingma, D., Ba, J.: Adam: A method for stochastic optimization. In: 3rd International Conference for Learning Representations (2015), arXiv preprint arXiv:1412.6980
9. Lin, M., Chen, Q., Yan, S.: Network in network. arXiv preprint arXiv:1312.4400 (2013)
10. Ma, L., Khorasani, K.: Constructive feedforward neural networks using hermite polynomial
activation functions. IEEE Trans. Neural Netw. 16(4), 821–833 (2005)
11. Piazza, F., Uncini, A., Zenobi, M.: Artificial neural networks with adaptive polynomial activation function. In: Int. Joint Conf. on Neural Networks. vol. 2, pp. II–343. IEEE/INNS
(1992)
12. Rasmussen, C.: Gaussian processes for machine learning. MIT Press (2006)
13. Scarpiniti, M., Comminiello, D., Parisi, R., Uncini, A.: Nonlinear spline adaptive filtering.
Signal Process. 93(4), 772–783 (2013)
14. Scarpiniti, M., Comminiello, D., Scarano, G., Parisi, R., Uncini, A.: Steady-state performance of spline adaptive filters. IEEE Trans. Signal Process. 64(4), 816–828 (2016)
15. Schmidhuber, J.: Deep learning in neural networks: An overview. Neural Networks 61, 85–
117 (2015)
16. Trentin, E.: Networks with trainable amplitude of activation functions. Neural Networks
14(4), 471–493 (2001)
17. Vecci, L., Piazza, F., Uncini, A.: Learning and approximation capabilities of adaptive spline
activation function neural networks. Neural Networks 11(2), 259–270 (1998)
18. Wahba, G.: Spline models for observational data. Siam (1990)
19. Wang, Y., Shen, D., Teoh, E.: Lane detection using catmull-rom spline. In: IEEE Int. Conf.
on Intell. Vehicles. pp. 51–57 (1998)
20. Zhang, M., Xu, S., Fulcher, J.: Neuron-adaptive higher order neural-network models for
automated financial data modeling. IEEE Trans. Neural Netw. 13(1), 188–204 (2002)
| 9cs.NE
|
ON SELF-SIMILAR FINITE p-GROUPS
arXiv:1603.04879v1 [math.GR] 15 Mar 2016
AZAM BABAI, KHADIJEH FATHALIKHANI,
GUSTAVO A. FERNÁNDEZ-ALCOBER, AND MATTEO VANNACCI
Abstract. In this paper, we address the following question: when is a
finite p-group G self-similar, i.e. when can G be faithfully represented as
a self-similar group of automorphisms of the p-adic tree? We show that,
if G is a self-similar finite p-group of rank r, then its order is bounded
by a function of p and r. This applies in particular to finite p-groups of
a given coclass. In the particular case of groups of maximal class, that
is, of coclass 1, we can fully answer the question above: a p-group of
maximal class G is self-similar if and only if it contains an elementary
abelian maximal subgroup over which G splits. Furthermore, in that
case the order of G is at most p p+1 , and this bound is sharp.
1. Introduction
A lot of interest has been raised in the past decades around groups of automorphisms of rooted trees, since they provide many examples and counterexamples for significant problems in group theory. Prominent among
these examples are the groups constructed by Grigorchuk [6], and by Gupta
and Sidki [7], which act on the p-adic tree, i.e. a regular rooted tree with
p children at each vertex, where p is a prime number. One of the most
important properties of the Grigorchuk and Gupta-Sidki groups is the fact
that they are self-similar . In other words, for every automorphism g in such
a group G, and for every vertex u of the tree, the section gu also belongs to
G. (See the beginning of Section 2 for the definition of sections and other
relevant concepts from the theory of groups of automorphisms of rooted
trees.)
Given an abstract group G and a prime p, the question arises as to whether
or not G can be faithfully represented as a self-similar group of automorphisms of the p-adic tree. If the answer is positive, then we say that G is
self-similar for the prime p. Nekrashevych and Sidki [14] showed that free
abelian groups of finite rank are self-similar for the prime 2, a result which
can similarly be established for all primes [13, Section 2.9.2]. On the contrary, if a finitely generated nilpotent group G is self-similar for the prime p
and the corresponding action is transitive on the first level of the tree, then
G is either free abelian or a finite p-group, as proved by Berlatto and Sidki
2010 Mathematics Subject Classification. Primary 20E08.
Key words and phrases. finite p-groups, p-adic tree, self-similarity, virtual endomorphisms.
The second author is supported by the University of Kashan, under grant no.
364988/10. The third author is supported by the Spanish Government, grants MTM201128229-C02-02 and MTM2014-53810-C2-2-P, and by the Basque Government, grants
IT753-13 and IT974-16.
1
2 A. BABAI, KH. FATHALIKHANI, G.A. FERNÁNDEZ-ALCOBER, AND M. VANNACCI
in [1, Corollary 2]. Further results about self-similar general abelian groups
can be found in [2]; for example, abelian torsion groups of infinite exponent
are not self-similar with level-transitive action.
Not much is known about self-similarity of finite p-groups. We can mention the work of Šunić [16], who proved that a finite p-group is isomorphic
to a self-similar group of automorphisms of the p-adic tree with abelian first
level stabilizer (so in particular with an abelian maximal subgroup) if and
only if it is a split extension of an elementary abelian group by a cyclic
group of order p. Actually, as we show in Theorem 2.2, this characterization
extends to all finite p-groups with an abelian maximal subgroup, without
needing to specify any properties of its action on the tree. On the other
hand, since a direct power of a self-similar group is again self-similar [13,
Proposition 2.9.3], there are self-similar p-groups without abelian maximal
subgroups.
In this paper we show that self-similarity is a strong condition to impose
on a finite p-group. More specifically, in our first main theorem we prove
that, for every prime p, there are only finitely many self-similar p-groups
of a given rank. This should be compared with the fact that every finite
p-group can be faithfully represented as a group of automorphisms of the
p-adic tree, by using its action on the coset tree corresponding to a chief
series.
Theorem A. Let G be a finite self-similar p-group of rank r. Then the
order of G is bounded by a function of p and r.
As a consequence, the order of a self-similar p-group can be bounded in
terms of the prime p and its coclass. We can give more detailed information
in the case of finite p-groups of coclass 1, i.e. groups of maximal class. As a
matter of fact, in our second main theorem we fully characterize self-similar
p-groups of maximal class. Furthermore we obtain the best possible bound
for the order of such a group.
Theorem B. Let G be a finite p-group of maximal class. Then G is selfsimilar if and only if G possesses an elementary abelian maximal subgroup
over which G splits. If that is the case, then the order of G is at most p p+1 ,
and this bound is sharp.
Thus for groups of maximal class we get the same characterization obtained by Šunić, even if most groups of maximal class do not have abelian
maximal subgroups. Observe that Theorem B gives plenty of examples of
finite p-groups which are not self-similar; in this respect, see Sections 3.1
and 8.2 of the book [11] by Leedham-Green and McKay.
Notation. If G is a finitely generated group, we use d(G) to denote the
minimum number of generators of G. On the other hand, if G is a finite
p-group and i ≥ 0 is an integer, Ωi (G) is the subgroup generated by all
i
elements of G of order at most pi , and Gp is the subgroup generated by all
pi th powers of elements of G.
ON SELF-SIMILAR FINITE p-GROUPS
3
2. Preliminaries
We briefly recall some concepts about the theory of groups of automorphisms of regular rooted trees. Let X be a set of cardinality k. The k-adic
tree T is the graph whose vertices are the elements of the free monoid X ∗
generated by X, and v is a descendant of u if and only if v = ux for some
x ∈ X. The words in X ∗ of length n form the nth level of T . An automorphism g of T is a bijective map from X ∗ to itself that preserves incidence.
The section of g at a vertex u, which we denote by gu , is the automorphism
of T which is defined by the rule
g(uv) = g(u)gu (v),
for every v ∈ X ∗ .
The automorphisms of T form a group Aut T under composition. A subgroup G of Aut T is self-similar if gu ∈ G for every u ∈ X ∗ .
A virtual endomorphism of a group G is a homomorphism φ : H → G,
where H is a subgroup of G of finite index. If |G : H| = k then φ is said to be
a virtual k1 -endomorphism. Also, we say that φ is simple if the only subgroup
of H which is normal in G and φ-invariant (in the sense that φ(H) ⊆ H)
is the trivial subgroup. Every self-similar group G of automorphisms of the
k-adic tree T acting transitively on the first level defines a simple virtual
1
k -endomorphism, by taking H = StabG (x), where x ∈ X, and defining
φ : H → G by φ(h) = hx . The converse is also true: every simple virtual
1
k -endomorphism φ : H → G can be used to define a faithful action of G on
T that identifies G with a self-similar group acting transitively on the first
level, and in which H is the stabilizer of a letter of X. This is accomplished
by using a coset tree; see [12] and [13] for further details.
We are interested in the following question: if p is a prime, when can
a finite p-group G be faithfully represented as a self-similar group of automorphisms of the p-adic tree? One can easily see that a non-trivial finite
p-group with a self-similar action on the p-adic tree is transitive on the
first level. Thus, according to the previous paragraph, the question reduces
to this: given a finite p-group G, when does there exist a simple virtual
endomorphism φ : H → G for some maximal subgroup H of G?
As mentioned in the introduction, Šunić gave an answer to this question
when H is abelian, which corresponds to the action of G having an abelian
first-level stabilizer. In that case, there exists a simple virtual endomorphism
φ : H → G if and only if H is elementary abelian and G splits over H. Next
we want to show that Šunić’s result applies to all finite p-groups with an
abelian maximal subgroup, without any requirements about the action of
the group on the tree. We need the following lemma (see [9, Lemma 4.6 and
its proof]).
Lemma 2.1. Let G be a finite p-group with an abelian maximal subgroup
A. Then the following hold:
(i) For every g ∈ G r A, we have G′ = {[g, a] | a ∈ A}. Thus |G′ | =
|A : CA (g)|.
(ii) If G is not abelian, then |G : Z(G)| = p|G′ |.
4 A. BABAI, KH. FATHALIKHANI, G.A. FERNÁNDEZ-ALCOBER, AND M. VANNACCI
Theorem 2.2. Let G be a finite p-group with an abelian maximal subgroup.
Then G is self-similar if and only if G possesses an elementary abelian
maximal subgroup over which G splits.
Proof. By Šunić’s result, we only have to prove the ‘only if’ part. Assume
then that G is self-similar, and consider a simple virtual endomorphism
φ : H → G, where H is maximal in G. Again by Šunić’s theorem, it suffices
to prove that H is abelian. By way of contradiction, suppose that H is
not abelian. Let A be an abelian maximal subgroup of G, which exists by
hypothesis.
We first prove that φ is injective. If there exists x ∈ ker φ r A then, since
A ∩ H is an abelian maximal subgroup of H, we have H ′ = [x, A ∩ H] by
Lemma 2.1. Thus φ(H ′ ) = 1 and, since φ is simple, we have H ′ = 1. This
is contrary to our assumption. Thus ker φ ⊆ A and A normalizes ker φ. It
follows that ker φ E G, and consequently ker φ = 1, as desired.
Let g ∈ H r A, and observe that G′ = [g, G] by Lemma 2.1. If there exists
a ∈ CA (g) r H then G = haiH and G′ = [g, H]. Hence G′ = H ′ . Since
φ(H ′ ) ⊆ G′ , it follows that G′ = 1, which is a contradiction. Consequently
CA (g) ⊆ H, in other words, Z(G) ⊆ Z(H). Then, again by Lemma 2.1,
|G′ | = |A : CA (g)| = p|A ∩ H : CA (g)| = p|A ∩ H : CA∩H (g)| = p|H ′ |.
Since also |G : Z(G)| = p|G′ | and |H : Z(H)| = p|H ′ |, it follows that
Z(G) = Z(H). Now if we put K = φ(H) then φ−1 defines a simple virtual
1
p -endomorphism from K to G. Thus we also have Z(G) = Z(K). Then
φ(Z(G)) = φ(Z(H)) = Z(K) = Z(G),
which is impossible, since φ is simple. This final contradiction shows that
H must be abelian.
The last theorem naturally raises the following question: if G is a finite
p-group with an abelian maximal subgroup A, is it true that G is self-similar
if and only if A itself is elementary abelian and G splits over A? This is
equivalent to asking whether all abelian maximal subgroups of a self-similar
p-group are elementary abelian. As we next see, this question has a positive
answer for odd primes, but not for the prime 2.
Corollary 2.3. Let G be a finite p-group, where p is an odd prime. If G
has an abelian maximal subgroup A, then G is self-similar if and only if A
is elementary abelian and G splits over A.
Proof. By Theorem 2.2, there exists an elementary abelian maximal subgroup H over which G splits. Then A ∩ H ⊆ Z(G) and G has class at
most 2. Let hgi be a complement to H in G, and write g = ah with a ∈ A
and h ∈ H. Then a = gh−1 is of order p, since G is a regular p-group.
Consequently A = ha, A ∩ Hi is elementary abelian.
On the other hand, note that the dihedral group D8 is self-similar and
possesses a non-elementary abelian maximal subgroup.
ON SELF-SIMILAR FINITE p-GROUPS
5
3. Finite p-groups of a given rank
In this section we prove Theorem A, which shows that there are only
finitely many self-similar finite p-groups of a given rank. Recall that the
(Prüfer) rank of a finite group is the maximum of the values of d(H) as H
ranges over all subgroups of G. The proof of Theorem A relies on the theory
of powerful p-groups. A finite p-group G is called powerful if G′ ≤ Gp for
p > 2, or G′ ≤ G4 for p = 2. The main results about powerful p-groups can
be found in the books [3] and [10].
Given a finite group G, a tuple (a1 , . . . , ad ) is said to be a basis of G
if every g ∈ G can be uniquely written in the form g = an1 1 . . . and d with
0 ≤ ni < |ai | for i = 1, . . . , d. By [3, Chapter 2, Exercise 9], every powerful
p-group has bases. On the other hand, if G = hx1 , . . . , xr i is a powerful
p-group then G = hx1 i . . . hxr i (see [3, Corollary 2.8]). As a consequence, all
bases of G have d(G) elements, and the order of the elements in a basis of
G is irrelevant. In the remainder, when we consider a basis (a1 , . . . , ad ) of a
powerful p-group, we will always assume that |a1 | ≥ · · · ≥ |ad |.
As we next see, a basis of a powerful p-group G can be used to obtain the
subgroup Ω1 (G) if p is odd, and the same result is true for G2 if p = 2.
Lemma 3.1. Let G be a powerful p-group, and let N = G2 . If {a1 , . . . , ad }
is a basis of N , and |ai | = mi for every i = 1, . . . , d, then
mi /p
Ω1 (N ) = hai
| i = 1, . . . , di.
Proof. Since N is powerful and d = d(N ), we have |N : N p | = |N : Φ(N )| =
pd . It follows that |Ω1 (N )| = pd , by [5, Theorem 1]. On the other hand,
since {a1 , . . . , ad } is a basis of N , the subgroup
mi /p
hai
| i = 1, . . . , di
of Ω1 (N ) has cardinality at least pd . This completes the proof.
Of course, G2 coincides with G if p is odd, and stating the last lemma for
is only a trick to avoid a case distinction according as p is odd or p = 2.
G2
We can now prove Theorem A. We use the expression ‘(p, r)-bounded’ as
shorthand for ‘bounded above by a function of p and r’.
Theorem 3.2. Let G be a self-similar finite p-group of rank r. Then the
order of G is (p, r)-bounded.
Proof. Let φ : H → G be a simple virtual endomorphism from a maximal
subgroup H of G. By Proposition 2.12 and Theorem 2.13 of [3], there exists a
normal subgroup P of G of (p, r)-bounded index with the property that every
subgroup of P which is normal in G is powerful. If we put N = (P ∩ H)2 ,
then N has also (p, r)-bounded index in G, and it suffices to show that N
has (p, r)-bounded order.
Let us consider a basis (a1 , . . . , ad ) of the powerful group N , and observe
that d = d(N ) ≤ r. Set |ai | = mi for i = 1, . . . , d, and md+1 = 1. If e denotes
the exponent of G/N , we claim that mi ≤ emi+1 for every i = 1, . . . , d. By
way of contradiction, assume that mj > emj+1 for some j, and for simplicity,
put m = mj+1 . Now, we have
(1)
φ(Ω1 (N em )) ⊆ Ω1 (Gem ) ⊆ Ω1 (N m ),
6 A. BABAI, KH. FATHALIKHANI, G.A. FERNÁNDEZ-ALCOBER, AND M. VANNACCI
since Ge ⊆ N . On the other hand, by [3, Theorem 2.7],
m
m
m
N m = ham
1 i . . . had i = ha1 i . . . haj i,
m
m
and (am
1 , . . . , aj ) is a basis of N . Since |ai | > em for i = 1, . . . , j, we
em
em
similarly obtain that (a1 , . . . , aj ) is a basis of N em . By Lemma 3.1,
m /p
mj /p
Ω1 (N m ) = ha1 1 , . . . , aj
i = Ω1 (N em ).
Thus, by (1), we get φ(Ω1 (N em )) ⊆ Ω1 (N em ). Since φ is simple, this implies
that Ω1 (N em ) = 1. This contradiction proves the claim. It follows that
mi ≤ ed−i+1 for all i = 1, . . . , d, and consequently
|N | = m1 . . . md ≤ ed(d+1)/2 .
Since d ≤ r and e is (p, r)-bounded, we conclude that the order of N is
(p, r)-bounded, as desired.
Corollary 3.3. Let G be a self-similar finite p-group of coclass s. Then the
order of G is (p, s)-bounded.
Proof. According to [15, Theorems 1.2 and 1.5], G contains a powerful characteristic subgroup N with (p, s)-bounded index and at most ps+1 generators. Thus rk(N ) ≤ ps+1 , by [3, Theorem 2.9]. Consequently G has (p, s)bounded rank, and the result follows from Theorem 3.2.
Corollary 3.4. Let G be a self-similar finite d-generator p-group of class
c. Then the order of G is (p, d, c)-bounded.
Proof. This follows immediately from Theorem 3.2, since the rank of a dgenerator nilpotent group of class c is (d, c)-bounded.
Of course, the number d of generators is necessary in the bound of the
previous corollary, since elementary abelian p-groups are all self-similar.
The following result, whose proof is contained in the proof of Theorem
3.2, may be of independent interest.
Proposition 3.5. Let G be a self-similar finite p-group, and let φ : H → G
be a simple virtual endomorphism, with H maximal in G. If U ⊆ H is a
uniform normal subgroup of G then exp(U ) ≤ exp(G/U ) if p is odd, and
exp(U ) ≤ 4 exp(G/U ) if p = 2.
Proof. Put N = U 2 , and use the same notation as in the proof of Theorem
3.2. Since N is uniform, we have exp(N ) = md , and on the other hand,
md ≤ exp(G/N ) by the proof of Theorem 3.2. Thus exp(N ) ≤ exp(G/N ),
and the result follows.
4. p-groups of maximal class
In this final section, we deal with finite p-groups of coclass 1, also known
as p-groups of maximal class. We prove Theorem B, which gives a full
characterization of self-similar p-groups of maximal class and determines the
best bound for the order of a self-similar p-group of maximal class. Observe
that, by Theorem 2.2, we need only consider p-groups of maximal class of
order at least p4 .
ON SELF-SIMILAR FINITE p-GROUPS
7
Before proceeding, we need to introduce some concepts and results from
the theory of groups of maximal class, for which we refer the reader to [4]
and [8, Chapter III, Section 14]. A p-group G of order pn with n ≥ 2 is said
to be of maximal class if its nilpotency class is n − 1. If we put Gi = γi (G)
for every i ≥ 2, this means that |G : G2 | = p2 and |Gi : Gi+1 | = p for every
i = 2, . . . , n − 1. The upper central series of a p-group of maximal class
coincides with its lower central series; in particular, Z(G) is of order p for
n ≥ 3. Actually, it is easy to describe all normal subgroups of G: we only
have to add the p + 1 maximal subgroups to the terms of the lower central
series of G.
For a group of maximal class G of order pn ≥ p4 , we put G1 = CG (G2 /G4 ),
so that |G : G1 | = |G1 : G2 | = p. Then the largest integer ℓ ∈ {0, . . . , n − 3}
such that [Gi , Gj ] ≤ Gi+j+ℓ for all i, j ≥ 1 is called the degree of commutativity of G, and we write ℓ = ℓ(G). A fundamental fact in the theory
of p-groups of maximal class is that ℓ(G/Z(G)) ≥ 1 for all n ≥ 5. As a
consequence, all the so-called two-step centralizers CG (Gi /Gi+2 ) (which are
maximal subgrups of G) coincide with G1 for i = 2, . . . , n − 3, and the remaining two-step centralizer, i.e. CG (Gn−2 ), is equal to G1 if and only if
ℓ(G) ≥ 1. Another key result that we will need is that a finite p-group is
of maximal class if and only if it contains an element s whose centralizer is
of order p2 . The elements satisfying this property are exactly the elements
in the difference G r (G1 ∪ CG (Gn−2 )), and they are called uniform elements of G. If s is a uniform element, then the order of s is p or p2 , and
sp ∈ Z(G). Also, for every 1 ≤ i ≤ n − 2, if x ∈ Gi r Gi+1 then we have
[x, s] ∈ Gi+1 r Gi+2 .
Our analysis of the representation of a p-group of maximal class G as a
self-similar group of automorphisms of the p-adic tree requires a thorough
understanding of the subgroups of G which contain a uniform element. Certainly, these groups are of maximal class again, but much more can be said.
We need the following straightforward lemma.
Lemma 4.1. Let G be a p-group of maximal class of order pn ≥ p4 , and let
K be a subgroup of G which contains a uniform element. If there exists an
element x of K lying in the difference Gt \ Gt+1 , then Gt ≤ K.
Proof. If s ∈ K is a uniform element of G, then [x, s] ∈ Gt+1 \Gt+2 , [x, s, s] ∈
Gt+2 \ Gt+3 , and so on until we get an element in Gn−1 r Gn . Since all these
elements lie in K, we conclude that Gt ≤ K.
Proposition 4.2. Let G be a p-group of maximal class of order pn ≥ p4 ,
and let K be a subgroup of G which contains a uniform element s. If the
order of K is pn−t , then K = hs, Gt+1 i.
Proof. Since G = hsiG1 and s ∈ K, we have K = hsi(K ∩G1 ), by Dedekind’s
law. Hence K ∩ G1 is of order pn−t−1 . The result is obvious for t = n − 1,
so we assume that t ≤ n − 2. Then K ∩ G1 6= 1, and so we can consider
j = min{i | 1 ≤ i ≤ n − 1, K ∩ (Gi r Gi+1 ) 6= ∅}.
Now it is clear from Lemma 4.1 that K ∩ G1 = Gj . Since |Gj | = pn−j , we
conclude that j = t + 1. Thus K = hs, Gt+1 i, as desired.
8 A. BABAI, KH. FATHALIKHANI, G.A. FERNÁNDEZ-ALCOBER, AND M. VANNACCI
Proposition 4.3. Let G be a p-group of maximal class of order at least p4 ,
and let H be a maximal subgroup of G different from G1 . Then, there is no
simple virtual endomorphism from H to G.
Proof. First of all, we observe that H ′ = [H, G2 ], since H/G2 is cyclic. The
commutator [H, G2 ] is a normal subgroup of G which is strictly contained in
G2 , so it must coincide with Gi for some i ≥ 3. Now [H, G2 ] is not contained
in G4 , since otherwise H = G1 . Hence [H, G2 ] = G3 .
By way of contradiction, let us assume that φ : H → G is a simple virtual
endomorphism. Then φ(H) is a proper subgroup of G, and we can consider
a maximal subgroup M of G containing φ(H). Now we have
φ(H ′ ) = φ(H)′ ≤ M ′ = [M, G2 ] ≤ G3 = H ′ ,
which is impossible, since φ is simple and H ′ 6= 1.
At this moment, we can prove the second part of Theorem B, namely the
bound for the order of a self-similar p-group of maximal class. This bound is
sharp, since the wreath product Cp ≀ Cp is a self-similar p-group of maximal
class of order p p+1 .
Corollary 4.4. Let G be a p-group of maximal class. If G is self-similar
then |G| ≤ p p+1 .
Proof. Assume that |G| > p p+1 . By Theorem 4.9 of [4], we have Gp1 = Gp .
Since sp ∈ Z(G) for every s ∈ G r G1 , it follows that Gp = Gp1 6= 1. Now let
φ : H → G be a simple 1p -endomorphism of G, and observe that H = G1 by
Proposition 4.3. Then φ(Gp ) = φ(Gp1 ) ⊆ Gp , which is a contradiction, since
φ is simple.
The following result completes the proof of Theorem B.
Theorem 4.5. Let G be a p-group of maximal class of order at least p4 .
Then G is self-similar if and only if G1 is elementary abelian, and G is a
split extension over G1 .
Proof. By Šunić’s result, we only need to prove the ‘only if’ part of the
statement of the theorem. So assume that G is self-similar, and let φ : H →
G be a simple virtual endomorphism, where H is a maximal subgroup of G.
By Proposition 4.3, H must be G1 .
Let G be of order pn , and suppose first that ℓ(G) ≥ 1. Since φ is simple,
it follows that φ(G1 ) * G1 . If s ∈ φ(G1 ) r G1 , then s is a uniform element
of G, since CG (Gn−2 ) = G1 in this case. Consequently, by Proposition 4.2,
we have φ(G1 ) = hs, Gt+1 i, for some integer t. If |φ(G1 )| ≥ p3 , then the
centre of φ(G1 ) is of order p, since this group is also of maximal class. As a
consequence, we get Z(φ(G1 )) = Z(G). On the other hand, since Z(G1 ) is
a normal subgroup of G, it follows that Z(G) ≤ Z(G1 ). Consequently,
φ(Z(G1 )) ≤ Z(φ(G1 )) = Z(G) ≤ Z(G1 ),
which is a contradiction, since φ is simple. Thus |φ(G1 )| = p or p2 , and then
φ(G′1 ) = φ(G1 )′ = 1. Since φ is simple, this implies that G1 is abelian. By
Šunić’s result, we know that G1 is elementary abelian and G splits over G1 ,
as desired.
ON SELF-SIMILAR FINITE p-GROUPS
9
Assume now that ℓ(G) = 0. Since Z(G1 ) is normal in G, we must have
Z(G1 ) = Gi for some i. But we have [G1 , Gn−2 ] 6= 1, since G1 6= CG (Gn−2 )
in this case. Thus Z(G1 ) = Gn−1 is of order p. If φ is not injective, then
Z(G1 ) ≤ ker φ. Therefore, φ(Z(G1 )) = 1, which is a contradiction, since
φ is simple. On the other hand, if φ is injective, then φ(G1 ) is a maximal
subgroup of G. Since G′1 = [G1 , G2 ] ≤ G4 , we have |G1 : G′1 | ≥ p3 , and
G1 is not of maximal class. So neither φ(G1 ) is of maximal class. But
we know that all maximal subgroups of G other than G1 and CG (Gn−2 )
are of maximal class, and also φ(G1 ) 6= G1 , since φ is simple. Thus we
necessarily have φ(G1 ) = CG (Gn−2 ). This implies that G1 and CG (Gn−2 )
are isomorphic, but notice that |Z(G1 )| = p, while Z(CG (Gn−2 )) contains
Gn−2 , which is of order p2 . This contradiction shows that we cannot have
ℓ(G) = 0, which completes the proof of the theorem.
If G is a p-group of maximal class with G1 elementary abelian, one cannot guarantee that G is a self-similar p-group. Indeed, it may perfectly
happen that all elements in G r G1 (i.e. all uniform elements) are of order
p2 . However, we have the following result.
Corollary 4.6. Let G be a p-group of maximal class of order at most p p+1
in which G1 is abelian. Then G/Z(G) is a self-similar p-group.
Proof. This is an immediate consequence of Theorem 4.5, since by Theorem
4.7 of [4] the quotient group G/Z(G) is of exponent p.
Acknowledgment. We thank Prof. Said Sidki for helpful comments.
References
[1] A. Berlatto, S. Sidki, Virtual endomorphisms of nilpotent groups, Groups Geom. Dyn.
1 (2007), 21–46.
[2] A.M. Brunner, S.N. Sidki, Abelian state-closed subgroups of automorphisms of m-ary
trees, Groups Geom. Dyn. 4 (2010), 455–472.
[3] J.D. Dixon, M.P.F. du Sautoy, A. Mann, D. Segal, Analytic pro-p groups, second edition, Cambridge Studies in Advanced Mathematics 61, Cambridge University Press,
1999.
[4] G.A. Fernández-Alcober, An introduction to finite p-groups: regular p-groups and
groups of maximal class, Mat. Contemp. 20 (2001), 155–226.
[5] G.A. Fernández-Alcober, Omega subgroups of powerful p-groups, Israel J. Math. 162
(2007), 75–79.
[6] R.I. Grigorchuk, On Burnside’s problem on periodic groups, Functional Anal. Appl.
14 (1980), 41–43.
[7] N. Gupta, S. Sidki, On the Burnside problem for periodic groups, Math. Z. 182
(1983), 385–388.
[8] B. Huppert, Endliche Gruppen, I , Grundlehren der mathematischen Wissenschaften
134, Springer, 1967.
[9] I.M. Isaacs, Finite Group Theory, Graduate Studies in Mathematics 92, American
Mathematical Society, 2008.
[10] E.I. Khukhro, p-Automorphisms of finite p-groups, London Mathematical Society
Lecture Note Series 246, Cambridge University Press, 1997.
[11] C.R. Leedham-Green, S. McKay, The Structure of Groups of Prime Power Order ,
London Mathematical Society Monographs New Series 27, Oxford University Press,
2002.
[12] V. Nekrashevych, Virtual endomorphisms of groups, Algebra Discrete Math. 1 (2002),
88–128.
10 A. BABAI, KH. FATHALIKHANI, G.A. FERNÁNDEZ-ALCOBER, AND M. VANNACCI
[13] V. Nekrashevych, Self-similar groups, Mathematical Surveys and Monographs 117,
American Mathematical Society, 2005.
[14] V. Nekrashevych, S. Sidki, Automorphisms of the binary tree: state-closed subgroups
and dynamics of 1/2-endomorphisms, in Groups: Topological, Combinatorial and
Arithmetic Aspects, London Mathematical Society Lecture Note Series 311, Cambridge University Press, 2004, 375–404.
[15] A. Shalev, The structure of finite p-groups: effective proof of the coclass conjectures,
Invent. Math. 115 (194), 315–345.
[16] Z. Šunić, Finite self-similar p-groups with abelian first-level stabilizer, Internat. J.
Algebra Comput. 21 (2011), 355–364.
Department of Mathematics, University of Qom, Qom, Iran
E-mail address: a [email protected]
Department of Pure Mathematics, Faculty of Mathematical Sciences, University of Kashan, Kashan 87317-51167, Iran
E-mail address: [email protected]
Department of Mathematics, University of the Basque Country UPV/EHU,
48080 Bilbao, Spain
E-mail address: [email protected]
Mathematisches Institut der Heinrich-Heine-Universität, Universitätsstr. 1,
40225 Düsseldorf, Germany
E-mail address: [email protected]
| 4math.GR
|
arXiv:1708.09047v1 [math.GR] 29 Aug 2017
Computable groups which do not embed into
groups with decidable conjugacy problem
Arman Darbinyan
1
Abstract
We show the existence of finitely generated torsion-free groups with decidable
word problem that cannot be embedded into groups with decidable conjugacy
problem. This answers a well-known question of Collins from the early 1970’s.
2
Introduction
Two of the most central decision problems associated with finitely generated
groups are word and conjugacy problems. One of the important questions about
these problems is concerning about the relation between them. For example, if
the conjugacy problem is decidable for a finitely generated group G, then the
word problem is decidable as well. However, in general, the inverse is far from
being true, [16, 6, 17, 15, 8].
If G is a finitely generated group and H ≤ G is a subgroup of finite index,
then the word problem in G is decidable if and only if it is decidable for H.
However, it is shown by Goryaga-Kirkinskii, [10], and independently by CollinsMiller, [7], that subgroups of index 2 of some specific finitely generated groups
have decidable (respectively, undecidable) conjugacy problem, while the groups
themselves have undecidable (respectively, decidable) conjugacy problem.
Another important type of questions about word and conjugacy problems
in groups is the following: Is it true that every finitely generated group with
decidable word problem (respectively, conjugacy problem) embeds in a finitely
presented group with decidable word problem (respectively, conjugacy problem)? Both of these questions have positive answer. The answer for the word
problem is obtained by Clapham in 1967, [5], based on the classical embedding
theorem of Higman (see [13]), while the analogous question for the conjugacy
problem was a long-standing open problem until it got positive answer in 2004
by a work of Olshanskii and Sapir. See [18] and also [19].
In light of the aforementioned, a natural question about the connection of
word and conjugacy problems in groups is the following question, asked by
1
Donald Collins in the early 1970s.
Question 1. Can every torsion-free group with solvable word problem
be embedded in a group with solvable conjugacy problem?
This question appears in the 1976 edition of The Kourovka Notebook as Problem 5.21, [12]. Probably, the first source where this problem was posed in a
written form is [3]. For yet another source, see [2].
It was mentioned by Collins in [12] that due to an example by A. Macintyre, there exists a group with torsions which cannot be embedded into a
finitely generated group with decidable conjugacy problem. However, the case
for torsion-free groups remained open until now. Indeed, one of the reasons
why the torsion and torsion-free cases are different is based on the observation
that conjugate elements in a group must have the same order, and since in a
torsion-free group all non trivial elements have the same (infinite) order, one
cannot make use of this observation in order to answer Question 1.
In [19], Olshanskii and Sapir showed the following theorem which gives a
positive answer to Question 1 under the stronger assumption of decidability of
the power problem.
Theorem 1 (Olshanskii-Sapir, [19]). Every countable group with solvable power
and order problems is embeddable into a 2-generated finitely presented group with
solvable conjugacy and power problems.
Note that as it is defined in [19], for a given group G the power problem is
said to be decidable, if there exists an algorithm which for any given pair (g, h)
of elements from G decides whether or not g is equal to some power of h in G.
The order problem is decidable in G if there exists an algorithm which for each
input g ∈ G computes the order of g.
The main result of the current work is the negative answer to Question 1 in
the general case.
Theorem 2. There exists a finitely presented torsion-free group G with decidable word problem such that G cannot be embedded into a group with decidable
conjugacy problem.
A remarkable theorem of Osin (see [20]) says that every torsion-free countable group can be embedded into a two generated group with exactly two conjugacy classes. In the context of this theorem, it is very natural to ask whether
or not every torsion-free countable group with decidable word problem (= computable group) can be embedded into a group with exactly two conjugacy classes
and with decidable word problem. A more relaxed version of this question would
be whether or not every torsion-free countable group with decidable word problem can be embedded in a finitely generated recursively presented group with
finitely many conjugacy classes.
It turns out that a direct consequence of Theorem 2 gives negative answer
to both of these questions.
2
In fact, the decidability of the conjugacy problem for groups with exactly two
conjugacy classes is equivalent to the decidability of the word problem. Even
more, as it is shown in a recent paper of Miasnikov and Schupp [15], a finitely
generated recursively presented group with finitely many conjugacy classes has
decidable conjugacy problem. Therefore, a direct corollary from Theorem 2 is
the following.
Theorem 3. There exists a torsion-free finitely presented group with decidable word problem that cannot be embedded into a finitely generated recursively
presented group with finitely many conjugacy classes.
Proof. Just take the group G from Theorem 2.
Remark 1. In fact, the mentioned result of Miasnikov and Schupp is true not
only for finitely generated recursively presented groups, but for all recursively
presented groups in general. Therefore, Theorem 3 stays true after dropping
the assumption that the group in which the initial group is embedded is finitely
generated. (The exact definition of recursive presentations of groups is given in
the next section.)
Acknowledgements. I would like to thank Alexander Olshanskii for his
thoughtful comments on this work.
3
3.1
Preliminaries
Groups with decidable word problem
A countable group G is said to have recursive presentation, if G can be presented
as G = hX | Ri such that X and R are enumerable by some algorithm (i.e.
Turing machine). See [11]. If in addition, there is an algorithm which for
each pair of words (w, w′ ) from (X ∪ X −1 )∗ verifies whether or not w and w′
represent the same element of G, then the presentation G = hX | Ri is called
computable and in case G possesses such a presentation, the group G itself is
called computable as well. Modulo some slight variances, the original definition
of the concept of computable groups is due to Rabin [21] and Mal’cev [14].
In case the group G is finitely generated (i.e. |X| < ∞) computability property of G is the same as saying that G has decidable word problem. It is not
hard to notice that decidability of the word problem does not depend on the
finite generating sets. From the computability perspective, the last observation
is one of the main advantages of finitely generated groups over countably generated ones, because in case of finitely generated groups decidability of the word
problem is an intrinsic property of a group, rather than of its presentation.
However, in this paper, to keep the notations as uniform as possible, we say
that G has decidable word problem if it is given by a computable presentation.
3
Let G = hx1 , x2 , . . . | r1 , r2 , . . .i, where {x1 , x2 , . . .} and {r1 , r2 , . . .} are
recursive enumerations of X and R, respectively. Then, the embedding constructions of [9] and [19] imply the following theorem.
Theorem 4. If G = hx1 , x2 , . . . | r1 , r2 , . . .i has decidable word problem, then
there exists an embedding Φ : G → H of G into a two generated group H such
that the following holds.
(1). The word problem is decidable in H;
(2). The map i 7→ Φ(xi ) is computable;
(3). An element of H is of finite order if and only if it is conjugate to an image
under Φ of an element of finite order in G.
3.2
HNN-extensions
In the proof of the existence of the group G from Theorem 2 we use some
group theoretical constructions based on HNN-extensions. Therefore, in this
subsection we would like to recall some well-known basic facts about HNNextensions. The basics of the theory of HNN-extensions can also be found in
[13].
Suppose that A, B ≤ H and φ : A → B is a group isomorphism from A to
B. Then the HNN-extension H ′ of H with respect to A and B (and φ) and
with stable letter t is defined as
H ′ = hH, t | t−1 at = φ(a), a ∈ Ai.
In the current text, the isomorphism φ will be clear from the context, hence
we will simply use the notation H ′ = hH, t | t−1 At = Bi.
Clearly, every element h′ of H ′ can be decomposed as a product
h ′ = h 0 tǫ 1 h 1 . . . t ǫ n h n ,
(1)
where ǫi ∈ {±1}, hj ∈ H for 1 ≤ i ≤ n, 0 ≤ j ≤ n.
The decomposition (1) is said to be in reduced form, if it does not contain
subproduct of one of the forms t−1 gi t, gi ∈ A or tgi t−1 , gi ∈ B, for 1 ≤ i ≤ n.
Analogously, if H = hXi, then the word u′ ∈ (X ∪ X −1 ∪ {t±1 })∗ given by
u ′ = u 0 tǫ 1 u 1 tǫ 2 . . . t ǫ n u n ,
where ǫi ∈ {±1}, uj ∈ (X ∪ X −1 )∗ , is said to be a reduced word with respect to
the HNN-extension H ′ if the decomposition h0 tǫ1 h1 . . . tǫn hn is in reduced form,
where hi corresponds to the word ui in H.
The following well-known lemma is attributed to Britton in [13].
Lemma 1 (Britton’s Lemma). If the decomposition (1) is reduced and n ≥ 1,
then h′ 6= 1 in H ′ .
4
Lemma 2 (See Theorem 2.1 in [13]). Let H ′ = hH, t | t−1 At = Bi be an HNNextension with respect to isomorphic subgroups A and B. Then H embeds in H ′
by the maps h 7→ h, h ∈ H.
Lemma 3 (The Torsion Theorem for HNN-extensions. See Theorem 2.4 in
[13]). Let H ′ = hH, t | t−1 At = Bi be an HNN-extension. Then every element
of finite order in H ′ is a conjugate of an element of finite order in the base H.
Thus H ′ has elements of finite order n if an only if H has elements of order n.
4
Proof of Theorem 2
In order to show the existence of G from Theorem 2, first, we will construct a
special countable group Ġ with decidable word problem, then G will be defined
as a group in which Ġ embeds in a certain way.
Two disjoint sets of natural numbers S1 , S2 ⊂ N are called recursively inseparable if there is no recursive set T ⊂ N such that S1 ⊆ T and S2 ⊆ N \ T . The
set T is called separating set. Clearly, if two disjoint sets are recursively inseparable, then they cannot be recursive. Indeed, if, say, S1 and S2 are disjoint and,
say, S1 is recursive, then as a recursive separating set one could simply take
S1 . Nevertheless, it is a well-known fact that there exist disjoint recursively
enumerable and recursively inseparable sets. See, for example, [22].
Let us fix two disjoint recursively enumerable and recursively inseparable
sets N = {n1 , n2 , . . .} ⊂ N and M = {m1 , m2 , . . .} ⊂ N such that the maps
i 7→ ni and i 7→ mi are computable.
Now, for all n ∈ N, define An as a torsion-free abelian additive group of rank
two with basis {an,0 , an,1 }, i.e.
An = han,0 i ⊕ han,1 i
and such that the groups A1 , A2 , . . . are disjoint.
For all n ∈ N, define the groups Ȧn as follows.
if n ∈
/ N ∪M ,
An
Ȧn =
An / ≪ an,1 = 2i an,0 ≫ if n = ni ∈ N ,
An / ≪ an,1 = 3i an,0 ≫ if n = mi ∈ M.
For all n ∈ N and m ∈ {0, 1}, let us denote the images of an,m under the
natural homomorphisms An → Ȧn by ȧn,m .
Convention. In this text, whenever we deal with an additive group, say, A,
∗
with finite generating set, say, {a1 , . . . , ak }, by
P {±a1 , . . . , ±ak } we denote the
set of formal finite sums of the form w =
λi aji , where λi ∈ Z and aji ∈
5
{a1 , . . . , ak }, and we say that w is a word formed by letters {±a1 , . . . , ±ak }.
Note that this is the additive analogue of the central in combinatorial group
theory concept of words, where the alphabet composing P
the words is a set of
group generators. This is why the finite formal sums w = λi aji we call words
from {±a1 , . . . , ±ak }∗ .
Before moving forward, we prove the following important lemma.
Lemma 4. There exists an algorithm such that for each input n ∈ N and
w ∈ {±ȧn,0 , ±ȧn,1 }∗ , it decides whether or not w represents the trivial element
in the group Ȧn .
Proof. Indeed, since Ȧn is abelian with generating set {ȧn,0 , ȧn,1 }, each word w
from {±ȧn,0 , ±ȧn,1 }∗ can be effectively transformed to a word of the form
w′ = λ0 ȧn,0 + λ1 ȧn,1
which represents the same element in Ȧn as the initial word w, where λ0 , λ1 ∈ Z.
Now, assuming that λ0 6= 0, λ1 6= 0, in order w′ to represent the trivial
element in Ȧn it must be that n ∈ N ∪ M, because otherwise, by definition, the
group Ȧn is torsion-free abelian of rank 2 with basis {ȧn,0 , ȧn,1 }.
In case n ∈ N , by definition we have that ȧn,1 = 2x ȧn,0 , where x is the index
of n in N , i.e. n = nx .
Similarly, in case n ∈ M, by definition we have that ȧn,1 = 3x ȧn,0 , where x
is the index of n in M, i.e. n = mx .
Now, if λ0 = 0 and λ1 = 0, then clearly w′ (hence also w) represents the
trivial element in Ȧn . Therefore, without loss of generality we can assume that
at least one of λ0 and λ1 is not 0. Then, if we treat x as an unknown variable,
depending on whether n = nx or n = mx , the equality w′ = 0 would imply one
of the following equations:
λ0 + λ1 2x = 0
(2)
λ0 + λ1 3x = 0,
(3)
or
respectively.
This observation suggests that in case λ0 6= 0 or λ1 6= 0, in order to verify
whether or not w′ = 0 in Ȧn , we can first try to find x satisfying (2) or (3),
and in case such an x does not exist, conclude that w′ (hence, also w) does not
represent the trivial element in Ȧn . Otherwise, if x is the root of the equation
(2), we can check whether or not n = nx (since N is recursively enumerable,
this checking can be done algorithmically). Similarly, if x is the root of the
equation (3), we can check whether or not n = mx .
If as a result of this checking, we get n = nx (respectively, n = mx ), then
the conclusion will be that w′ (hence, also w) represents the trivial element in
6
Ȧn , otherwise, if n 6= nx (respectively, n 6= mx ), then the conclusion will be
that w′ (hence, also w) does not represent the trivial element in Ȧn .
Now, for all n ∈ N, define the group Bn as a torsion-free additive abelian
group of rank 2, that is
Bn = hbn,0 i ⊕ hbn,1 i
such that B1 , B2 , . . . are disjoint.
Now, for all n ∈ N, define the groups Ḃn as follows.
Bn
if n ∈
/ N ∪M ,
Ḃn =
Bn / ≪ bn,1 = 2i bn,0 ≫ if n = ni ∈ N or n = mi ∈ M.
For all n ∈ N, m ∈ {0, 1}, let us denote the images of bn,m under the natural
homomorphism Bn → Ḃn by ḃn,m .
It follows from the definitions of Ȧn and Ḃn that for all n ∈ N, these groups
are infinite and torsion free.
Lemma 5. There exists an algorithm such that for each input n ∈ N and
w ∈ {±ḃn,0 , ±ḃn,1 }∗ , it decides whether or not w represents the trivial element
in the group Ḃn .
Proof. Follows from the repetition of arguments of the proof of Lemma 4.
Lemma 6. The map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism
between the groups hȧn,0 , ȧn,1 i = Ȧn and hḃn,0 , ḃn,1 i = Ḃn if and only if n ∈
N \ M.
Proof. Indeed, in case n ∈ N , by definition, hȧn,0 , ȧn,1 i = hȧn,0 i and ȧn,1 =
2i ȧn,0 , where i is the index of n in N . Also hḃn,0 , ḃn,1 i = hḃn,0 i and ḃn,1 = 2i ḃn,0 .
Therefore, in case n ∈ N , the map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group
isomorphism between the groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i.
In case n ∈ N \ (N ∪ M), the groups Ȧn and Ḃn are torsion-free and abelian
of rank 2 with generating sets {ȧn,0 , ȧn,1 } and {ḃn,0 , ḃn,1 }, respectively. Therefore, if n ∈ N \ (N ∪ M), the map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group
isomorphism between the groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i as well.
Now suppose that n ∈ M. Then, hȧn,0 , ȧn,1 i = hȧn,0 i and hḃn,0 , ḃn,1 i =
hḃn,0 i, however, by definition, ȧn,1 = 3i ȧn,0 while ḃn,1 = 2i ḃn,0 . Therefore, the
map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 does not induce a group isomorphism between the
groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i when n ∈ M.
7
Now, let T = F (t1 , t2 , . . .) be a free group with countable free basis {t1 , t2 , . . .}.
Denote the infinite free products Ȧ1 ∗ Ȧ1 ∗ . . . and Ḃ1 ∗ Ḃ1 ∗ . . . by ∗∞
n=1 Ȧn
and ∗∞
Ḃ
,
respectively.
Then
define
n=1 n
∞
Ġ = (∗∞
n=1 Ȧn ) ∗ (∗n=1 Ḃn ) ∗ T / ≪ R ≫,
(4)
where the set of defining relators R is defined as
R = t−1
i ȧni ,0 ti = ḃni ,0 | i ∈ N .
Define
∞
Ġ0 = (∗∞
n=1 Ȧn ) ∗ (∗n=1 Ḃn ),
and for all k ∈ N, define Ġk as
∞
Ġk = (∗∞
n=1 Ȧn ) ∗ (∗n=1 Ḃn ) ∗ F (t1 , . . . , tk )/ ≪ Rk ≫,
where the set of defining relators Rk is defined as
Rk = t−1
i ȧni ,0 ti = ḃni ,0 | 1 ≤ i ≤ k .
Then, clearly the group Ġ is the direct limit of the sequence of group {Ġk }∞
k=0
connected by homomorphisms ǫk : Ġk → Ġk+1
such that ǫk are the homomorphisms induced by the identity maps from ȧn,0 , ȧn,1 , ḃn,0 , ḃn,1 , ti | n ∈ N, i ∈
{1, 2, . . . , k} to themselfs for all k ∈ N.
Let us denote
S0 = ± ȧn,m , ± ḃn,m | n ∈ N, m ∈ {0, 1}
and for k ∈ N,
Sk =
±1
± ȧn,m , ± ḃn,m , t±1
| n ∈ N, m ∈ {0, 1} .
1 , . . . , tk
Note that since the sets N and M are recursively enumerable, the groups
Ġ and Ġk have recursive presentations with respect to the generating sets
S0 ∪ {t1 , t2 , . . .} and Sk , k ∈ N ∪ {0}, respectively.
Lemma 7. There exists an algorithm such that for each input w ∈ S0∗ it decides
whether or not w = 1 in Ġ0 .
Moreover, there exists an algorithm such that for each input (w, i), w ∈ S0∗ ,
i ∈ N, it decides whether or not w represents an element from hȧni ,0 i, and in
case it represents such an element, the algorithm returns λȧni ,0 , λ ∈ Z, such
that w = λȧni ,0 in Ġ0 . Analogous statement remains true when we replace ȧni ,0
with ḃni ,0 .
8
Proof. Indeed, these properties immediately follow from the basic properties of
the direct products of groups combined with Lemmas 4 and 5.
Lemma 8. For all k ∈ N ∪ {0} and n ∈ N, the following holds.
(i). The groups Ȧn and Ḃn embed into Ġk under the maps induced by ȧn,m 7→
ȧn,m and ḃn,m 7→ ḃn,m for m ∈ {0, 1}, respectivley;
(ii). The group Ġk+1 is an HNN-extension of the group Ġk . More precisely,
Ġk+1 = hĠk , tk+1 | t−1
k+1 ȧnk+1 ,0 tk+1 = ḃnk+1 ,0 i.
Proof. Indeed, if k = 0, then (i) and (ii) are obvious. Now, let us apply induction with respect to k.
Suppose that for all 0 ≤ l < k, the statements of (i) and (ii) are true.
Then, since by the inductive assumption, Ġk is obtained from Ġk−1 as an HNNextension with respect to the isomorphic subgroups hȧnk ,0 i ⋍ hḃnk ,0 i, by the
basic properties of HNN-extensions (see Lemma 2), we get that the statement of
(i) holds for Ġk . Therefore, since the subgroups hȧnk+1 ,0 i ≤ Ġk and hḃnk+1 ,0 i ≤
Ġk are isomorphic, and in the definition of Ġk+1 the only defining relation which
−1
involves the letters t±1
k+1 is the relation tk+1 ȧnk+1 ,0 tk+1 = ḃnk+1 ,0 , we get that
the statement of (ii) holds as well.
Corollary 1. If k < l, then the group Ġk embeds into the group Ġl under the
map induced by
ȧn,m 7→ ȧn,m , ḃn,m 7→ ḃn,m for n ∈ N and m ∈ {0, 1}
and
t1 7→ t1 , . . . , tk 7→ tk .
Proof. Indeed, by Lemma 8, the group Ġl is obtained from the group Ġk by
(multiple) HNN-extensions. Therefore, the statement follows from the basic
properties of HNN-extensions, namely, by Lemma 2.
Corollary 2. The map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism
between the subgroups hȧn,0 , ȧn,1 i = Ȧn and hḃn,0 , ḃn,1 i = Ḃn of Ġ if and only
if n ∈ N \ M.
Proof. By Corollary 1, Ġ0 embeds in Ġ by the map induced by ȧn,0 7→ ȧn,0 ,
ȧn,1 7→ ȧn,1 , ḃn,0 7→ ḃn,0 , ḃn,1 7→ ḃn,1 for n ∈ N. Therefore, the statement of the
corollary follows from Lemma 6.
Definition 1 (Reduced words over Sk∗ ). Let k ∈ N. Then, for a given word
w ∈ Sk∗ , we say that w is a reduced word over Sk∗ if the following properties hold.
9
(0). w is freely reduced, i.e. w does not contain subwords of the form xx−1 ,
x ∈ Sk ;
(1). For all 1 ≤ i ≤ k, w does not contain subwords of the form t−1
i uti , where
u ∈ S0∗ is such that u = λȧni ,0 in Ġ0 for some λ ∈ Z;
(2). For all 1 ≤ i ≤ k, w does not contain subwords of the form ti vt−1
i , where
v ∈ S0∗ is such that v = λḃni ,0 in Ġ0 for some λ ∈ Z.
∗
Lemma 9. For all k ∈ N, if w ∈ Sk∗ \ Sk−1
is a reduced word over Sk∗ , then
∗
w 6= 1 in Ġk . Moreover, w 6= u in Ġk for any word u ∈ Sk−1
.
Proof. Let us prove by induction on k. If k = 1, then the group Ġ1 = hĠ0 , t1 |
t−1
1 ȧn1 ,0 t1 = ḃn1 ,0 i is an HNN-extension of Ġ0 with respect to the isomorphic
subgroups hȧn1 ,0 i ≤ Ġ0 and hḃn1 ,0 i ≤ Ġ0 . Therefore, by Britton’s Lemma (see
Lemma 1), w 6= 1 in Ġ1 provided that it is a reduced word over S1∗ .
Also for any u ∈ S0∗ , the word wu−1 is a reduced word with respect to
the HNN-extension Ġ1 = hĠ0 , t1 | t−1
1 ȧn1 ,0 t1 = ḃn1 ,0 i. Therefore, by Britton’s
Lemma (see Lemma 1), wu−1 6= 1 in Ġ1 or, in other words, w 6= u in Ġ1 .
∗
Now assume that k > 1 and w ∈ Sk∗ \ Sk−1
is a reduced word over Sk∗ . Also,
suppose that the statement of the lemma is true for all l < k. Then, first of
all, note that from the definition of the reduced words over Sk∗ it follows that if
∗
∗
.
, then v is a reduced word over Sk−1
v is a subword of w such that v ∈ Sk−1
−1
−1
Consequently, by the inductive hypothesis, if tk utk (or tk utk ) is a subword of
∗
w such that u ∈ Sk−1
and u represents an element from the image of Ȧnk (or
Ḃnk ) in Ġk , then u ∈ S0∗ . However, this contradicts the assumption that w is a
reduced word over Sk∗ . Therefore, since Ġk = hĠk−1 , tk | t−1
k ȧnk ,0 tk = ḃnk ,0 i is
an HNN-extension of Ġk−1 with respect to the isomorphic subgroups hank ,0 i =
Ȧnk ≤ Ġk−1 and hbnk ,0 i = Ḃnk ≤ Ġk−1 , we get that if w is a reduced word over
Sk∗ , then w is a reduced word over this HNN-extension. Hence, by Britton’s
Lemma, we get that w 6= 1 in Ġk . Similarly, for any u ∈ S0∗ , again by Britton’s
Lemma, we get that wu−1 6= 1 in Ġk or, in other words, w 6= u in Ġk .
Lemma 10. There exists an algorithm such that for each input (k, w), k ∈
N ∪ {0}, w ∈ Sk∗ , it decides whether or not w = 1 in Ġk .
Proof. Let (k, w) be a fixed input. Without loss of generality assume that w is
a freely reduced word in Sk∗ .
If k = 0, then one can apply the word problem algorithm for the group
Ġ0 = hS0∗ i. See Lemma 7.
Otherwise, if k ≥ 1, for each k1 ≤ k such that w contains a letter from
{tk1 , t−1
k1 }, do the following: Find all subwords of w which are of one of the forms
−1
∗
ut
t−1
k1 or tk1 vtk1 , where u, v ∈ S0 and u = λȧnk1 ,0 , v = λḃnk1 ,0 in Ġ0 for some
k1
λ ∈ Z. (By Lemma 7, subwords of these form can be found algorithmically.)
Then, if, say, a subword of the form t−1
k1 utk1 is found, replace it with λḃnk1 ,0 .
10
Thanks to the identity t−1
k1 λȧnk1 ,0 tk1 = λḃnk1 ,0 , the newly obtained word is
equal to w in Ġk . Then repeat this procedure on the newly obtained word until
there is no more subwords of the mentioned forms. Let w1 be the word obtained
as a result of this procedure. Then, by Lemma 9, either w1 ∈ S0∗ or for some
k0 ≥ 1, w1 ∈ Sk∗0 \ Sk∗0 −1 . Then, in the last case, by Lemma 9, w1 is a reduced
word over Sk∗0 . Also in the first case (i.e. when w1 ∈ S0∗ ), w1 = 1 in Ġk if and
only if w1 = 1 in Ġ0 , hence by Lemma 7, in this case, the identity w1 = 1 can
be checked algorithmically. In the second case, by Lemma 9, w1 6= 1 in Ġk .
Lemma 11. The word problem in Ġ is decidable with respect to the presentation
(4).
Proof. Suppose that w is a finite word with letters from
±1
Sk = ± ȧn,m , ± ḃn,m , t±1
| n ∈ N, m ∈ {0, 1} ,
1 , . . . , tk
where k is some natural number. Also suppose that w represents the trivial
element in Ġ. Then, since Ġ is a direct limit of the groups {Ġi }∞
i=1 , there exists
a minimal integer N ≥ 0 such that w represents the trivial element in ĠN .
We claim that N ≤ k. Indeed, if N > k, then since N was chosen as the
minimal index such that w = 1 in ĠN , we get w 6= 1 in Ġk . However, by
Corollary 1, Ġk embeds into ĠN under the map induces by
ȧn,m 7→ ȧn,m and t1 7→ t1 , . . . , tk 7→ tk , for n ∈ N, m ∈ {0, 1},
which implies that if w 6= 1 in Ġk , then w 6= 1 in ĠN . A contradiction.
Thus, if w ∈ Sk∗ represents the trivial element in Ġ, then it represents the
trivial element in Ġk as well. In other words, in order to check whether or not
w represents the trivial element in Ġ it is enough to check its triviality in Ġk .
Therefore, since for each w ∈ S ∗ one can algorithmically find (the minimal)
k ∈ N such that w ∈ Sk∗ , the decidability of the word problem in Ġ follows from
Lemma 10.
Lemma 12. The group Ġ is torsion-free.
Proof. First of all, notice that by the properties of the groups Ȧk , Ḃk , k ∈ N,
and by the basic properties of direct products, the group Ġ0 is torsion free.
Now, suppose that u ∈ S ∗ is such that it represents a torsion element of
Ġ. Then, since Ġ is a direct limit of the groups {Ġi }∞
i=1 , there exists k ∈ N
such that u ∈ Sk∗ and u represents a torsion element in Ġk as well. Since Ġk
is obtained from Ġ0 by multiple HNN-extensions, then, by Lemma 3, Ġk is a
torsion free group. Therefore, u represents the trivial element in Ġk as well as
in Ġ.
11
Now suppose that Φ : Ġ ֒→ G̈ is an embedding of the group Ġ into a finitely
generated torsion-free group G̈ such that the maps
φ1 : (n, m) 7→ Φ(ȧn,m ), φ2 : (n, m) 7→ Φ(ḃn,m ), and φ3 : n 7→ Φ(tn ),
where n ∈ N, m ∈ {0, 1},
are computable, and G̈ has decidable word problem. Then the next lemma
shows that the group G̈ has the desirable properties we were looking for.
Lemma 13. The group G̈ cannot be embedded in a group with decidable conjugacy problem.
Proof. By contradiction, let us assume that G̈ embeds in a group Ḡ which has
decidable conjugacy problem. Then, for the purpose of convenience, without
loss of generality let us assume that G̈ is a subgroup of the group Ḡ.
Below we show that the decidability of the conjugacy problem in Ḡ contradicts the assumption that N and M are disjoint and recursively inseparable.
Let us define C ⊆ N as
C = n ∈ N | Φ(ȧn,0 ) is conjugate to Φ(ḃn,0 ) in Ḡ .
Then, the decidability of the conjugacy problem in Ḡ implies that the set C is
recursive, because, since the group Ḡ has decidable conjugacy problem, and since
by our assumptions, the above mentioned maps φ1 , φ2 and φ3 are computable,
for any input n ∈ N one can algorithmically verify whether or not Φ(ȧn,0 ) is
conjugate to Φ(ḃn,0 ) in Ḡ.
Therefore, since for groups with decidable conjugacy problem one can algorithmically find conjugator element for each pair of conjugate elements of the
group, we also get that there exists a computable map
f : C → Ḡ
such that for all n ∈ C we have
f (n)−1 Φ(ȧn,0 )f (n) = Φ(ḃn,0 ).
For n ∈ C, let us denote
f (n) = gn ∈ Ḡ.
Now let us define
A = n ∈ C | gn−1 Φ(ȧn,1 )gn = Φ(ḃn,1 ) ⊆ N.
12
Since the word problem in Ḡ is decidable, the sets C is recursive and the maps
Φ and f are computable, we get that the set A is a recursive subset of N. Also
since the following identities
ȧni ,1 = 2i ȧni ,0 , ḃni ,1 = 2i ḃni ,0 and ti−1 ȧni ,0 ti = ḃni ,0 , for i ∈ N,
hold in Ġ, we get that in Ḡ the following identities hold
i
i
Φ(ȧni ,1 ) = Φ(ȧni ,0 )2 , Φ(ḃni ,1 ) = Φ(ḃni ,0 )2
and
Φ(ti )−1 Φ(ȧni ,0 )Φ(ti ) = Φ(ḃni ,0 ) for all ni ∈ N .
Therefore, we get that
N ⊆ A.
On the other hand, Corollary 2 implies that for any n ∈ M, the pairs of elements
Φ(ȧn,0 ), Φ(ḃn,0 ) and Φ(ȧn,1 ), Φ(ḃn,1 )
cannot be conjugate in Ḡ by the same conjugator. Therefore, we get that
A ∩ M = ∅.
Thus we get that N ⊆ A and A ∩ M = ∅, which implies that A ⊂ N is a
recursive separating set for N and M, which contradicts the assumption that
N and M are recursively inseparable.
Finally, the embedding Φ : Ġ ֒→ G̈ with the prescribed properties exists,
thanks to Theorem 4. Therefore, the group G̈ with the above mentioned properties exists. Also by a version of Higman’s embedding theorem described by
Aanderaa and Cohen in [1], the group G̈ can be embedded into a finitely presented group G with decidable word problem. By a recent result of Chiodo
and Vyas, [4], the group G defined this way will also inherit the property of
torsion-freeness from the group G̈.
Clearly, since G̈ cannot be embedded into a group with decidable conjugacy
problem, this property will be inherited by G. Thus Theorem 2 is proved.
13
References
[1] S. Aanderaa, D. E. Cohen, Modular machines I, II, in [2], Stud. Logic
Found. Math. 95 (1980), 1-18, 19-28.
[2] G. Baumslag, A. Myasnikov, V. Shpilrain et al., Open problems in Combinatorial and Geometric Group Theory, http://www.grouptheory.info/.
[3] W.W. Boone, F.B. Cannonito, and R.C. Lyndon, Word Problems: Decision
Problems and the Burnside Problem in Group Theory, Studies in Logic
and the Foundations of Mathematics, vol. 71, North-Holland, Amsterdam,
1973.
[4] M. Chiodo, R. Vyas, Torsion length and finitely presented groups,
arXiv:1604.03788, 2016.
[5] Clapham C.R.J., An embedding theorem for finitely generated groups.
Proc. London Math. Soc. (1967) 17:419-430.
[6] D.J. Collins, Representation of Turing reducibility by word and conjugacy
problems in finitely presented groups, Acta Math. 128 , (1972) no. 1-2,7390.
[7] D.J. Collins, C.F. Miller III , The conjugacy problem and subgroups of
finite index, Proc. London Math. Soc. (3) 34 (1977), no. 3, 535-556.
[8] A. Darbinyan, Word and conjugacy problems in lacunary hyperbolic
groups, arXiv:1708.04591.
[9] A. Darbinyan, Group embeddings with algorithmic properties, Communications in Algebra, 43:11 (2015), 4923-4935.
[10] A.V. Gorjaga, A.S. Kirkinski , The decidability of the conjugacy problem
cannot be transferred to finite extensions of groups. (Russian) Algebra i
Logika 14 (1975), no. 4, 393-406.
[11] G. Higman. Subgroups of finitely presented groups. Proc. Roy. Soc. Ser. A,
262:455-475, 1961.
[12] Kourovka Notebook. Unsolved Problems in Group Theory. 5th edition,
Novosibirsk, 1976.
[13] R.C. Lyndon, P.E. Schupp, Combinatorial group theory, Springer, Berlin,
1977.
[14] A. Mal’cev, Constructive algebras. I, Uspehi Mat. Nauk, vol. 16 (1961), no.
3 (99), pp. 3-60.
[15] A. Miasnikov, P. Schupp, Computational complexity and the conjugacy
problem. Computability, 2016.
14
[16] C. F. Miller III. On group-theoretic decision problems and their classification, volume 68 of Annals of Mathematics Studies. Princeton University
Press, 1971.
[17] C.F. Miller III, Decision Problems for Groups Survey and Reflections.
Algorithms and Classification in Combinatorial Group Theory (1992) 23:159.
[18] A.Yu. Olshanskii, M. Sapir, The conjugacy problem and Higman embeddings. Mem. Amer. Math. Soc. 170 (2004), no. 804
[19] A. Yu. Olshanskii, M.V. Sapir, Subgroups of finitely presented groups with
solvable conjugacy problem. Internat. J. Algebra Comput., 15(5-6):10751084, 2005.
[20] D. Osin, Small cancellations over relatively hyperbolic groups and embedding theorems, Ann. Math. 172 (2010), no. 1, 1-39.
[21] M. Rabin, Computable algebra, general theory and theory of computable
fields., Trans. Amer. Math. Soc., vol. 95 (1960), pp. 341-360.
[22] J. R. Shoenfield, Mathematical logic. Addison Wesley, 1967.
A. Darbinyan, Department of Mathematics, Vanderbilt University,
1326 Stevenson Center, Nashville, TN 37240
E-mail address: [email protected]
15
| 4math.GR
|
Unsupervised Pretraining for Sequence to Sequence Learning
Prajit Ramachandran and Peter J. Liu and Quoc V. Le
Google Brain
{prajit, peterjliu, qvl}@google.com
Abstract
arXiv:1611.02683v2 [cs.CL] 22 Feb 2018
This work presents a general unsupervised learning method to improve the accuracy of sequence to sequence (seq2seq)
models. In our method, the weights of
the encoder and decoder of a seq2seq
model are initialized with the pretrained
weights of two language models and then
fine-tuned with labeled data. We apply this method to challenging benchmarks in machine translation and abstractive summarization and find that it significantly improves the subsequent supervised models. Our main result is that
pretraining improves the generalization of
seq2seq models. We achieve state-of-theart results on the WMT English→German
task, surpassing a range of methods using both phrase-based machine translation and neural machine translation. Our
method achieves a significant improvement of 1.3 BLEU from the previous best
models on both WMT’14 and WMT’15
English→German. We also conduct human evaluations on abstractive summarization and find that our method outperforms a purely supervised learning baseline in a statistically significant manner.
1
Introduction
Sequence to sequence (seq2seq) models
(Sutskever et al., 2014; Cho et al., 2014;
Kalchbrenner and Blunsom, 2013; Allen, 1987;
Ñeco and Forcada, 1997) are extremely effective
on a variety of tasks that require a mapping
between a variable-length input sequence to
a variable-length output sequence. The main
weakness of sequence to sequence models, and
deep networks in general, lies in the fact that they
can easily overfit when the amount of supervised
training data is small.
In this work, we propose a simple and effective technique for using unsupervised pretraining
to improve seq2seq models. Our proposal is to
initialize both encoder and decoder networks with
pretrained weights of two language models. These
pretrained weights are then fine-tuned with the labeled corpus. During the fine-tuning phase, we
jointly train the seq2seq objective with the language modeling objectives to prevent overfitting.
We benchmark this method on machine translation for English→German and abstractive summarization on CNN and Daily Mail articles. Our
main result is that a seq2seq model, with pretraining, exceeds the strongest possible baseline
in both neural machine translation and phrasebased machine translation. Our model obtains
an improvement of 1.3 BLEU from the previous best models on both WMT’14 and WMT’15
English→German. On human evaluations for abstractive summarization, we find that our model
outperforms a purely supervised baseline, both in
terms of correctness and in avoiding unwanted
repetition.
We also perform ablation studies to understand
the behaviors of the pretraining method. Our study
confirms that among many other possible choices
of using a language model in seq2seq with attention, the above proposal works best. Our study
also shows that, for translation, the main gains
come from the improved generalization due to the
pretrained features. For summarization, pretraining the encoder gives large improvements, suggesting that the gains come from the improved optimization of the encoder that has been unrolled
for hundreds of timesteps. On both tasks, our proposed method always improves generalization on
the test sets.
W
X
Y
Z
<EOS>
Softmax
Second RNN Layer
First RNN Layer
Embedding
A
B
C
<EOS>
W
X
Y
Z
Figure 1: Pretrained sequence to sequence model. The red parameters are the encoder and the blue
parameters are the decoder. All parameters in a shaded box are pretrained, either from the source side
(light red) or target side (light blue) language model. Otherwise, they are randomly initialized.
2
Methods
In the following section, we will describe our basic
unsupervised pretraining procedure for sequence
to sequence learning and how to modify sequence
to sequence learning to effectively make use of the
pretrained weights. We then show several extensions to improve the basic model.
2.1
Basic Procedure
Given an input sequence x1 , x2 , ..., xm and an
output sequence yn , yn−1 , ..., y1 , the objective of
sequence to sequence learning is to maximize the
likelihood
p(yn , yn−1 , ..., y1 |x1 , x2 , ..., xm ).
Common sequence to sequence learning methods decompose this objective
as
p(yn , yn−1 , ..., y1 |x1 , x2 , ..., xm )
=
Q
n
p(y
|y
,
...,
y
;
x
,
x
,
...,
x
).
t t−1
1 1 2
m
t=1
In sequence to sequence learning, an RNN encoder is used to represent x1 , ..., xm as a hidden
vector, which is given to an RNN decoder to produce the output sequence. Our method is based
on the observation that without the encoder, the
decoder essentially acts like a language model on
y’s. Similarly, the encoder with an additional output layer also acts like a language model. Thus it
is natural to use trained languages models to initialize the encoder and decoder.
Therefore, the basic procedure of our approach
is to pretrain both the seq2seq encoder and decoder networks with language models, which can
be trained on large amounts of unlabeled text data.
This can be seen in Figure 1, where the parameters in the shaded boxes are pretrained. In the following we will describe the method in detail using
machine translation as an example application.
First, two monolingual datasets are collected,
one for the source side language, and one for the
target side language. A language model (LM) is
trained on each dataset independently, giving an
LM trained on the source side corpus and an LM
trained on the target side corpus.
After two language models are trained, a multilayer seq2seq model M is constructed. The embedding and first LSTM layers of the encoder and
decoder are initialized with the pretrained weights.
To be even more efficient, the softmax of the decoder is initialized with the softmax of the pretrained target side LM.
2.2
Monolingual language modeling losses
After the seq2seq model M is initialized with the
two LMs, it is fine-tuned with a labeled dataset.
However, this procedure may lead to catastrophic
forgetting, where the model’s performance on the
language modeling tasks falls dramatically after
fine-tuning (Goodfellow et al., 2013). This may
hamper the model’s ability to generalize, especially when trained on small labeled datasets.
To ensure that the model does not overfit the labeled data, we regularize the parameters that were
pretrained by continuing to train with the monolingual language modeling losses. The seq2seq and
language modeling losses are weighted equally.
In our ablation study, we find that this technique
is complementary to pretraining and is important
in achieving high performance.
2.3
Other improvements to the model
Pretraining and the monolingual language modeling losses provide the vast majority of improvements to the model. However in early experimentation, we found minor but consistent improvements with two additional techniques: a) residual
connections and b) multi-layer attention (see Figure 2).
X
W
Attention
+
A
W
B
(a)
<EOS>
C
(b)
Figure 2: Two small improvements to the baseline
model: (a) residual connection, and (b) multi-layer
attention.
Residual connections: As described, the input
vector to the decoder softmax layer is a random
vector because the high level (non-first) layers of
the LSTM are randomly initialized. This introduces random gradients to the pretrained parameters. To avoid this, we use a residual connection
from the output of the first LSTM layer directly to
the input of the softmax (see Figure 2-a).
Multi-layer attention: In all our models, we use
an attention mechanism (Bahdanau et al., 2015),
where the model attends over both top and first
layer (see Figure 2-b). More concretely, given a
query vector qt from the decoder, encoder states
from the first layer h11 , . . . , h1T , and encoder states
L
from the last layer hL
1 , . . . , hT , we compute the attention context vector ct as follows:
exp(qt · hN
i )
αi = PT
N
j=1 exp(qt · hj )
cN
t =
T
X
αi hN
i
c1t =
T
X
αi h1i
i=1
ct = [c1t ; cN
t ]
i=1
3
Experiments
In the following section, we apply our approach
to two important tasks in seq2seq learning: ma-
chine translation and abstractive summarization.
On each task, we compare against the previous
best systems. We also perform ablation experiments to understand the behavior of each component of our method.
3.1
Machine Translation
Dataset and Evaluation: For machine translation, we evaluate our method on the WMT
English→German task (Bojar et al., 2015). We
used the WMT 14 training dataset, which is
slightly smaller than the WMT 15 dataset. Because the dataset has some noisy examples, we
used a language detection system to filter the
training examples. Sentences pairs where either the source was not English or the target
was not German were thrown away. This resulted in around 4 million training examples.
Following Sennrich et al. (2015a), we use subword units (Sennrich et al., 2015b) with 89500
merge operations, giving a vocabulary size around
90000. The validation set is the concatenated newstest2012 and newstest2013, and our test sets are
newstest2014 and newstest2015. Evaluation on
the validation set was with case-sensitive BLEU
(Papineni et al., 2002) on tokenized text using
multi-bleu.perl. Evaluation on the test
sets was with case-sensitive BLEU on detokenized
text using mteval-v13a.pl. The monolingual
training datasets are the News Crawl English and
German corpora, each of which has more than a
billion tokens.
Experimental settings: The language models
were trained in the same fashion as (Jozefowicz
et al., 2016) We used a 1 layer 4096 dimensional
LSTM with the hidden state projected down to
1024 units (Sak et al., 2014) and trained for one
week on 32 Tesla K40 GPUs. Our seq2seq model
was a 3 layer model, where the second and third
layers each have 1000 hidden units. The monolingual objectives, residual connection, and the modified attention were all used. We used the Adam
optimizer (Kingma and Ba, 2015) and train with
asynchronous SGD on 16 GPUs for speed. We
used a learning rate of 5e-5 which is multiplied
by 0.8 every 50K steps after an initial 400K steps,
gradient clipping with norm 5.0 (Pascanu et al.,
2013), and dropout of 0.2 on non-recurrent connections (Zaremba et al., 2014). We used early
stopping on validation set perplexity. A beam size
of 10 was used for decoding. Our ensemble is con-
System
Phrase Based MT (Williams et al., 2016)
Supervised NMT (Jean et al., 2015)
Edit Distance Transducer NMT (Stahlberg et al., 2016)
Edit Distance Transducer NMT (Stahlberg et al., 2016)
Backtranslation (Sennrich et al., 2015a)
Backtranslation (Sennrich et al., 2015a)
Backtranslation (Sennrich et al., 2015a)
No pretraining
Pretrained seq2seq
Pretrained seq2seq
ensemble?
single
single
ensemble 8
single
ensemble 4
ensemble 12
single
single
ensemble 5
BLEU
newstest2014 newstest2015
21.9
23.7
22.4
21.7
24.1
22.9
25.7
22.7
25.7
23.8
26.5
24.7
27.6
21.3
24.3
24.0
27.0
24.7
28.1
Table 1: English→German performance on WMT test sets. Our pretrained model outperforms all other
models. Note that the model without pretraining uses the LM objective.
M
LST
&
s
s
pus
r
ding
ding
er
ia
cor
d
d
l
d
e
e
e
o
ode
l
b
b
l
c
c
iped
a
e
m
n
m
e
k
r
v
i
e
e
e
d
i
a
g
t
p
W
n
n
n
n
n
c
ai
ai
ai
ai
n
n
ini
obje
retr
retr
retr
retr
in o
in o
etra
p
p
p
p
a
a
r
M
r
r
t
p
y
L
y
y
y
t
l
l
l
l
Pre
No
On
No
On
On
On
Pre
0.0
0.3
Difference in BLEU
0.5
1.0
1.0
1.5
1.6
2.0
2.1
2.0
2.0
1.6
2.0
Figure 3: English→German ablation study measuring the difference in validation BLEU between various
ablations and the full model. More negative is worse. The full model uses LMs trained with monolingual
data to initialize the encoder and decoder, plus the language modeling objective.
structed with the 5 best performing models on the
validation set, which are trained with different hyperparameters.
Results: Table 1 shows the results of our
method in comparison with other baselines. Our
method achieves a new state-of-the-art for single model performance on both newstest2014
and newstest2015, significantly outperforming the
competitive semi-supervised backtranslation technique (Sennrich et al., 2015a). Equally impressive
is the fact that our best single model outperforms
the previous state of the art ensemble of 4 models.
Our ensemble of 5 models matches or exceeds the
previous best ensemble of 12 models.
Ablation study: In order to better understand
the effects of pretraining, we conducted an ablation study by modifying the pretraining scheme.
We were primarily interested in varying the pretraining scheme and the monolingual language
modeling objectives because these two techniques
produce the largest gains in the model. Figure
3 shows the drop in validation BLEU of various
ablations compared with the full model. The full
model uses LMs trained with monolingual data to
initialize the encoder and decoder, in addition to
the language modeling objective. In the follow-
• Only pretraining the decoder is better than
only pretraining the encoder: Only pretraining the encoder leads to a 1.6 BLEU point
drop while only pretraining the decoder leads
to a 1.0 BLEU point drop.
the pretrained models degrade less as the labeled
dataset becomes smaller.
BLEU
ing, we interpret the findings of the study. Note
that some findings are specific to the translation
task.
Given the results from the ablation study, we
can make the following observations:
Pretrain
No pretrain
22
21
20
19
18
17
16
15
• Pretrain as much as possible because the benefits compound: given the drops of no pretraining at all (−2.0) and only pretraining the
encoder (−1.6), the additive estimate of the
drop of only pretraining the decoder side is
−2.0 − (−1.6) = −0.4; however the actual
drop is −1.0 which is a much larger drop than
the additive estimate.
Figure 4: Validation performance of pretraining
vs. no pretraining when trained on a subset of the
entire labeled dataset for English→German translation.
• Pretraining the softmax is important: Pretraining only the embeddings and first LSTM
layer gives a large drop of 1.6 BLEU points.
3.2
• The language modeling objective is a strong
regularizer: The drop in BLEU points of pretraining the entire model and not using the
LM objective is as bad as using the LM objective without pretraining.
• Pretraining on a lot of unlabeled data is essential for learning to extract powerful features: If the model is initialized with LMs
that are pretrained on the source part and
target part of the parallel corpus, the drop
in performance is as large as not pretraining at all. However, performance remains
strong when pretrained on the large, nonnews Wikipedia corpus.
To understand the contributions of unsupervised pretraining vs. supervised training, we track
the performance of pretraining as a function of
dataset size. For this, we trained a a model with
and without pretraining on random subsets of the
English→German corpus. Both models use the
additional LM objective. The results are summarized in Figure 4. When a 100% of the labeled
data is used, the gap between the pretrained and
no pretrain model is 2.0 BLEU points. However,
that gap grows when less data is available. When
trained on 20% of the labeled data, the gap becomes 3.8 BLEU points. This demonstrates that
20
40
60
80
100
Percent of entire labeled dataset used for training
Abstractive Summarization
Dataset and Evaluation: For a low-resource
abstractive summarization task, we use the
CNN/Daily Mail corpus from (Hermann et al.,
2015). Following Nallapati et al. (2016), we modify the data collection scripts to restore the bullet
point summaries. The task is to predict the bullet
point summaries from a news article. The dataset
has fewer than 300K document-summary pairs. To
compare against Nallapati et al. (2016), we used
the anonymized corpus. However, for our ablation study, we used the non-anonymized corpus.1
We evaluate our system using full length ROUGE
(Lin, 2004). For the anonymized corpus in particular, we considered each highlight as a separate sentence following Nallapati et al. (2016). In
this setting, we used the English Gigaword corpus (Napoles et al., 2012) as our larger, unlabeled
“monolingual” corpus, although all data used in
this task is in English.
Experimental settings: We use subword units
(Sennrich et al., 2015b) with 31500 merges, resulting in a vocabulary size of about 32000. We
use up to the first 600 tokens of the document and
1
We encourage future researchers to use the nonanonymized version because it is a more realistic summarization setting with a larger vocabulary. Our numbers on
the non-anonymized test set are 35.56 ROUGE-1, 14.60
ROUGE-2, and 25.08 ROUGE-L. We did not consider highlights as separate sentences.
System
Seq2seq + pretrained embeddings (Nallapati et al., 2016)
+ temporal attention (Nallapati et al., 2016)
Pretrained seq2seq
ROUGE-1
32.49
35.46
32.56
ROUGE-2
11.84
13.30
11.89
ROUGE-L
29.47
32.65
29.44
Table 2: Results on the anonymized CNN/Daily Mail dataset.
STM
gs
pus
er
er
cor
ddin
ddin
d
d
l
e
e
o
o
e
l
b
b
al
ve
dec
em
em
enc
par
ecti
ning
ain
ain
ain
ain
j
n
i
r
r
r
r
t
b
t
t
t
o
a
r
e
o
e
e
e
t
n
pre
y pr
LM
y pr
y pr
y pr
trai
No
Onl
No
Onl
Onl
Onl
Pre
L
gs &
0
Difference in ROUGE
1
2
3
4
5
ROUGE1
ROUGE2
ROUGEL
Figure 5: Summarization ablation study measuring the difference in validation ROUGE between various
ablations and the full model. More negative is worse. The full model uses LMs trained with unlabeled
data to initialize the encoder and decoder, plus the language modeling objective.
predict the entire summary. Only one language
model is trained and it is used to initialize both
the encoder and decoder, since the source and target languages are the same. However, the encoder
and decoder are not tied. The LM is a one-layer
LSTM of size 1024 trained in a similar fashion to
Jozefowicz et al. (2016). For the seq2seq model,
we use the same settings as the machine translation experiments. The only differences are that
we use a 2 layer model with the second layer having 1024 hidden units, and that the learning rate is
multiplied by 0.8 every 30K steps after an initial
100K steps.
Results: Table 2 summarizes our results on the
anonymized version of the corpus. Our pretrained
model is only able to match the previous baseline seq2seq of Nallapati et al. (2016). Interestingly, they use pretrained word2vec (Mikolov
et al., 2013) vectors to initialize their word em-
beddings. As we show in our ablation study, just
pretraining the embeddings itself gives a large improvement. Furthermore, our model is a unidirectional LSTM while they use a bidirectional LSTM.
They also use a longer context of 800 tokens,
whereas we used a context of 600 tokens due to
GPU memory issues.
Ablation study: We performed an ablation
study similar to the one performed on the machine translation model. The results are reported in Figure 5. Here we report the drops on
ROUGE-1, ROUGE-2, and ROUGE-L on the nonanonymized validation set.
Given the results from our ablation study, we
can make the following observations:
• Pretraining appears to improve optimization: in contrast with the machine translation
model, it is more beneficial to only pretrain
the encoder than only the decoder of the sum-
marization model. One interpretation is that
pretraining enables the gradient to flow much
further back in time than randomly initialized
weights. This may also explain why pretraining on the parallel corpus is no worse than
pretraining on a larger monolingual corpus.
• The language modeling objective is a strong
regularizer: A model without the LM objective has a significant drop in ROUGE scores.
Human evaluation: As ROUGE may not be
able to capture the quality of summarization, we
also performed a small qualitative study to understand the human impression of the summaries
produced by different models. We took 200 random documents and compared the performance of
a pretrained and non-pretrained system. The document, gold summary, and the two system outputs were presented to a human evaluator who was
asked to rate each system output on a scale of 1-5
with 5 being the best score. The system outputs
were presented in random order and the evaluator
did not know the identity of either output. The
evaluator noted if there were repetitive phrases or
sentences in either system outputs. Unwanted repetition was also noticed by Nallapati et al. (2016).
Table 3 and 4 show the results of the study. In
both cases, the pretrained system outperforms the
system without pretraining in a statistically significant manner. The better optimization enabled by
pretraining improves the generated summaries and
decreases unwanted repetition in the output.
NP > P
29
NP = P
88
NP < P
83
Table 3: The count of how often the no pretrain
system (NP) achieves a higher, equal, and lower
score than the pretrained system (P) in the side-byside study where the human evaluator gave each
system a score from 1-5. The sign statistical test
gives a p-value of < 0.0001 for rejecting the null
hypothesis that there is no difference in the score
obtained by either system.
4
Related Work
Unsupervised pretraining has been intensively
studied in the past years, most notably is the work
by Dahl et al. (2012) who found that pretraining
with deep belief networks improved feedforward
Pretrain
No repeats
Repeats
No pretrain
No repeats Repeats
67
65
24
44
Table 4: The count of how often the pretrain and
no pretrain systems contain repeated phrases or
sentences in their outputs in the side-by-side study.
McNemar’s test gives a p-value of < 0.0001 for
rejecting the null hypothesis that the two systems
repeat the same proportion of times. The pretrained system clearly repeats less than the system
without pretraining.
acoustic models. More recent acoustic models
have found pretraining unnecessary (Xiong et al.,
2016; Zhang et al., 2016; Chan et al., 2015), probably because the reconstruction objective of deep
belief networks is too easy. In contrast, we find
that pretraining language models by next step prediction significantly improves seq2seq on challenging real world datasets.
Despite its appeal, unsupervised learning has
not been widely used to improve supervised training. Dai and Le (2015); Radford et al. (2017) are
amongst the rare studies which showed the benefits of pretraining in a semi-supervised learning
setting. Their methods are similar to ours except
that they did not have a decoder network and thus
could not apply to seq2seq learning. Similarly,
Zhang and Zong (2016) found it useful to add an
additional task of sentence reordering of sourceside monolingual data for neural machine translation. Various forms of transfer or multitask learning with seq2seq framework also have the flavors
of our algorithm (Zoph et al., 2016; Luong et al.,
2015; Firat et al., 2016).
Perhaps most closely related to our method is
the work by Gulcehre et al. (2015), who combined
a language model with an already trained seq2seq
model by fine-tuning additional deep output layers. Empirically, their method produces small improvements over the supervised baseline. We suspect that their method does not produce significant
gains because (i) the models are trained independently of each other and are not fine-tuned (ii) the
LM is combined with the seq2seq model after the
last layer, wasting the benefit of the low level LM
features, and (iii) only using the LM on the decoder side. Venugopalan et al. (2016) addressed (i)
but still experienced minor improvements. Using
pretrained GloVe embedding vectors (Pennington
et al., 2014) had more impact.
Related to our approach in principle is the work
by Chen et al. (2016) who proposed a two-term,
theoretically motivated unsupervised objective for
unpaired input-output samples. Though they did
not apply their method to seq2seq learning, their
framework can be modified to do so. In that case,
the first term pushes the output to be highly probable under some scoring model, and the second
term ensures that the output depends on the input.
In the seq2seq setting, we interpret the first term
as a pretrained language model scoring the output sequence. In our work, we fold the pretrained
language model into the decoder. We believe that
using the pretrained language model only for scoring is less efficient that using all the pretrained
weights. Our use of labeled examples satisfies the
second term. These connections provide a theoretical grounding for our work.
In our experiments, we benchmark our method
on machine translation, where other unsupervised
methods are shown to give promising results (Sennrich et al., 2015a; Cheng et al., 2016). In backtranslation (Sennrich et al., 2015a), the trained
model is used to decode unlabeled data to yield
extra labeled data. One can argue that this method
may not have a natural analogue to other tasks
such as summarization. We note that their technique is complementary to ours, and may lead
to additional gains in machine translation. The
method of using autoencoders in Cheng et al.
(2016) is promising, though it can be argued that
autoencoding is an easy objective and language
modeling may force the unsupervised models to
learn better features.
5
Conclusion
We presented a novel unsupervised pretraining
method to improve sequence to sequence learning.
The method can aid in both generalization and optimization. Our scheme involves pretraining two
language models in the source and target domain,
and initializing the embeddings, first LSTM layers, and softmax of a sequence to sequence model
with the weights of the language models. Using
our method, we achieved state-of-the-art machine
translation results on both WMT’14 and WMT’15
English to German. A key advantage of this technique is that it is flexible and can be applied to a
large variety of tasks.
References
Robert B. Allen. 1987. Several studies on natural language and back-propagation. IEEE First International Conference on Neural Networks.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly
learning to align and translate. In ICLR.
Ondřej Bojar, Rajen Chatterjee, Christian Federmann,
Barry Haddow, Matthias Huck, Chris Hokamp,
Philipp Koehn, Varvara Logacheva, Christof Monz,
Matteo Negri, Matt Post, Carolina Scarton, Lucia
Specia, and Marco Turchi. 2015. Findings of the
2015 workshop on statistical machine translation.
In Proceedings of the Tenth Workshop on Statistical
Machine Translation.
William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol
Vinyals. 2015. Listen, attend and spell. arXiv
preprint arXiv:1508.01211.
Jianshu Chen, Po-Sen Huang, Xiaodong He, Jianfeng
Gao, and Li Deng. 2016. Unsupervised learning
of predictors from unpaired input-output samples.
abs/1606.04646.
Yong Cheng, Wei Xu, Zhongjun He, Wei He, Hua
Wu, Maosong Sun, and Yang Liu. 2016. Semisupervised learning for neural machine translation.
arXiv preprint arXiv:1606.04596.
Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger
Schwenk, and Yoshua Bengio. 2014. Learning
phrase representations using RNN encoder-decoder
for statistical machine translation. In EMNLP.
G. E. Dahl, D. Yu, L. Deng, and A. Acero. 2012.
Context-dependent pre-trained deep neural networks
for large-vocabulary speech recognition. IEEE
Transactions on Audio, Speech, and Language Processing, 20(1):30–42.
Andrew M. Dai and Quoc V. Le. 2015.
supervised sequence learning. In NIPS.
Semi-
Orhan Firat, Baskaran Sankaran, Yaser Al-Onaizan,
Fatos T. Yarman-Vural, and Kyunghyun Cho.
2016.
Zero-resource translation with multilingual neural machine translation. arXiv preprint
arXiv:1606.04164.
Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron
Courville, and Yoshua Bengio. 2013. An empirical investigation of catastrophic forgetting in
gradient-based neural networks. arXiv preprint
arXiv:1312.6211.
Caglar Gulcehre, Orhan Firat, Kelvin Xu, Kyunghyun
Cho, Loic Barrault, Huei-Chi Lin, Fethi Bougares,
Holger Schwenk, and Yoshua Bengio. 2015. On using monolingual corpora in neural machine translation. arXiv preprint arXiv:1503.03535.
Karl Moritz Hermann, Tomas Kocisky, Edward
Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In NIPS.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2015a.
Improving neural machine translation
models with monolingual data. arXiv preprint
arXiv:1511.06709.
Sébastien Jean, Orhan Firat, Kyunghyun Cho, Roland
Memisevic, and Yoshua Bengio. 2015. Montreal
neural machine translation systems for WMT’15. In
Proceedings of the Tenth Workshop on Statistical
Machine Translation.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2015b.
Neural machine translation of rare
words with subword units.
arXiv preprint
arXiv:1508.07909.
Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam
Shazeer, and Yonghui Wu. 2016.
Exploring
the limits of language modeling. arXiv preprint
arXiv:1602.02410.
Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent
continuous translation models. In EMNLP.
Diederik Kingma and Jimmy Ba. 2015. Adam: A
method for stochastic optimization. In ICLR.
Chin-Yew Lin. 2004. ROUGE: a package for automatic evaluation of summaries. In Proceedings of
the Workshop on Text Summarization Branches Out
(WAS 2004).
Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol
Vinyals, and Lukasz Kaiser. 2015. Multi-task sequence to sequence learning. In ICLR.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In NIPS.
Ramesh Nallapati, Bing Xiang, and Bowen Zhou.
2016. Sequence-to-sequence RNNs for text summarization. arXiv preprint arXiv:1602.06023.
Courtney Napoles, Matthew Gormley, and Benjamin
Van Durme. 2012. Annotated gigaword. In Proceedings of the Joint Workshop on Automatic Knowledge Base Construction and Web-scale Knowledge
Extraction. ACL.
Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: A method for automatic
evaluation of machine translation. In ACL.
Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio.
2013. On the difficulty of training recurrent neural
networks. ICML.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for
word representation. In EMNLP.
Alec Radford, Rafal Jozefowicz, and Ilya Sutskever.
2017. Learning to generate reviews and discovering
sentiment. arXiv preprint arXiv:1704.01444.
Hasim Sak, Andrew W. Senior, and Françoise Beaufays. 2014. Long short-term memory recurrent
neural network architectures for large scale acoustic
modeling. In INTERSPEECH.
Felix Stahlberg, Eva Hasler, and Bill Byrne. 2016.
The edit distance transducer in action: The university of cambridge english-german system at wmt16.
In Proceedings of the First Conference on Machine
Translation, pages 377–384, Berlin, Germany. Association for Computational Linguistics.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
Sequence to sequence learning with neural networks. In NIPS.
Subhashini Venugopalan, Lisa Anne Hendricks, Raymond Mooney, and Kate Saenko. 2016. Improving LSTM-based video description with linguistic knowledge mined from text. arXiv preprint
arXiv:1604.01729.
Philip Williams, Rico Sennrich, Maria Nadejde,
Matthias Huck, Barry Haddow, and Ondřej Bojar.
2016. Edinburgh’s statistical machine translation
systems for wmt16. In Proceedings of the First
Conference on Machine Translation, pages 399–
410, Berlin, Germany. Association for Computational Linguistics.
Wayne Xiong, Jasha Droppo, Xuedong Huang, Frank
Seide, Mike Seltzer, Andreas Stolcke, Dong Yu, and
Geoffrey Zweig. 2016. Achieving human parity in
conversational speech recognition. abs/1610.05256.
Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals.
2014. Recurrent neural network regularization.
arXiv preprint arXiv:1409.2329.
Jiajun Zhang and Chengqing Zong. 2016. Exploiting source-side monolingual data in neural machine
translation. In EMNLP.
Yu Zhang, William Chan, and Navdeep Jaitly. 2016.
Very deep convolutional networks for end-to-end
speech recognition. abs/1610.03022.
Barret Zoph, Deniz Yuret, Jonathan May, and Kevin
Knight. 2016. Transfer learning for low-resource
neural machine translation. In EMNLP.
Ramon P. Ñeco and Mikel L. Forcada. 1997. Asynchronous translations with recurrent neural nets.
Neural Networks.
A
Example outputs
Source Document
( cnn ) like phone booths and typewriters , record stores are a vanishing breed –
another victim of the digital age . camelot music . virgin megastores . wherehouse
music . tower records . all of them gone . corporate america has largely abandoned
brick - and - mortar music retailing to a scattering of independent stores , many of
them in scruffy urban neighborhoods . and that s not necessarily a bad thing . yes , it
s harder in the spotify era to find a place to go buy physical music . but many of the
remaining record stores are succeeding – even thriving – by catering to a passionate
core of customers and collectors . on saturday , hundreds of music retailers will
hold events to commemorate record store day , an annual celebration of , well , your
neighborhood record store . many stores will host live performances , drawings ,
book signings , special sales of rare or autographed vinyl and other happenings .
some will even serve beer . to their diehard customers , these places are more than
mere stores : they are cultural institutions that celebrate music history ( the entire
duran duran oeuvre , all in one place ! ) , display artifacts ( aretha franklin on vinyl !
) , and nurture the local music scene ( hey , here s a cd by your brother s metal band
! ) . they also employ knowledgeable clerks who will be happy to debate the relative
merits of blood on the tracks and blonde on blonde . or maybe , like jack black in
high fidelity , just mock your lousy taste in music . so if you re a music geek , drop
by . but you might think twice before asking if they stock i just called to say i love
you .
Ground Truth summary
saturday is record store day , celebrated at music stores around the world . many
stores will host live performances , drawings and special sales of rare vinyl .
No pretrain
corporate america has largely abandoned brick - brick - mortar music . many of the
remaining record stores are succeeding – even thriving – by catering to a passionate
core of customers .
Pretrained
hundreds of music retailers will hold events to commemorate record store day . many
stores will host live performances , drawings , book signings , special sales of rare or
autographed vinyl .
Table 5: The pretrained model outputs a highly informative summary, while the no pretrain model outputs
irrelevant details.
Source Document
( cnn ) hey , look what i did . that small boast on social media can trigger a whirlwind
that spins into real - life grief , as a texas veterinarian found out after shooting a cat
. dr. kristen lindsey allegedly shot an arrow into the back of an orange tabby s head
and posted a proud photo this week on facebook of herself smiling , as she dangled its
limp body by the arrow s shaft . lindsey added a comment , cnn affiliate kbtx reported
. my first bow kill , lol . the only good feral tomcat is one with an arrow through it
s head ! vet of the year award ... gladly accepted . callers rang the phones hot at
washington county s animal clinic , where lindsey worked , to vent their outrage .
web traffic crashed its website . high price of public shaming on the internet then
an animal rescuer said that lindsey s prey was probably not a feral cat but the pet of
an elderly couple , who called him tiger . he had gone missing on wednesday , the
same day that lindsey posted the photo of the slain cat . cnn has not been able to
confirm the claim . as the firestorm grew , lindsey wrote in the comments underneath
her post : no i did not lose my job . lol . psshh . like someone would get rid of me
. i m awesome ! that prediction was wrong . the clinic fired lindsey , covered her
name on its marquee with duct tape , and publicly distanced itself from her actions
. our goal now is to go on and try to fix our black eye and hope that people are
reasonable and understand that those actions do nt anyway portray what we re for
here at washington animal clinic , said dr. bruce buenger . we put our heart and soul
into this place . the clinic told wbtx that lindsey was not available for comment .
cnn is reaching out to her . she removed her controversial post then eventually shut
down her facebook page . callers also complained to the brenham police department
and washington county animal control , as her facebook post went viral . the sheriff
s office in austin county , where the cat was apparently shot , is investigating , and
lindsey could face charges . its dispatchers were overloaded with calls , the sheriff
posted on facebook . we are asking you to please take it easy on our dispatchers .
as soon as the investigation is complete , we will post the relevant information here
on this page , the post read . animal rights activists are pushing for charges . animal
cruelty must be taken seriously , and the guilty parties should be punished to the
fullest extent of the law , said cat advocacy activist becky robinson . her organization
, alley cat allies , is offering a $ 7,500 reward for evidence leading to the arrest and
conviction of the person who shot the cat . but others stood up for lindsey . she s
amazing . she s caring , said customer shannon stoddard . she s a good vet , so maybe
her bad choice of posting something on facebook was not good . but i do nt think she
should be judged for it . she dropped off balloons at the animal clinic for lindsey with
a thank you note . cnn s jeremy grisham contributed to this report .
Ground Truth summary
dr. kristen lindsey has since removed the post of her holding the dead cat by an arrow
. her employer fired her ; the sheriff s office is investigating . activist offers $ 7,500
reward .
No pretrain
dr. kristen lindsey allegedly shot an arrow into the back of an orange orange tabby s
head . it s the only good good tomcat is one with an arrow through it s head ! vet vet
of the year award .
Pretrained
lindsey lindsey , a texas veterinarian , shot an arrow into the back of an orange tabby
s head . she posted a photo of herself smiling , as she dangled its limp body by the
arrow s shaft . lindsey could face charges , the sheriff s department says .
Table 6: The pretrained model outputs a highly relevant summary but makes a mistake on the feline
executioner’s name. The no pretrain model degenerates into irrelevant details and repeats itself.
Source Document
eugenie bouchard s run of poor form continued as the top seed was beaten 6 - 3 , 6 - 1
by american lauren davis in the second round at the family circle cup in charleston on
wednesday . davis , 21 , had lost her only career meeting with bouchard , but was in
control this time against the world no 7 . davis won nine of the final 11 games of the
match and broke bouchard s serve twice in the final set to pull off the upset . eugenie
bouchard fires down a serve during her second - round match at the family circle cup
bouchard shows her frustrations during her straight - sets defeat by lauren davis on
wednesday i ve never beaten here before , so i came out knowing i had nothing to
lose , said davis , ranked 66th in the world . bouchard was a semi-finalist last year
but had struggled in the lead - up to the charlston green - clay event , losing three of
her last five matches to lower - ranked opponents . davis used precise groundstrokes
to keep bouchard on her heels throughout the second set . davis broke bouchard s
serve to love to take a 3 - 1 lead as the 21 - year - old canadian had a double fault ,
then sailed a forehand long to fall behind . two games later , davis hit a backhand to
send back bouchard s 102mph serve for a winner down the left sideline and a 5 - 1
lead . davis hits a forehand on her way to an impressive win against the canadian top
seed bouchard has struggled recently , this time slumping to her fourth defeat in six
matches the match ended on bouchard s forehand into the net , davis waving to the
cheering crowd . bouchard said : as soon as it was over , definitely a bit of anger , but
also kind of this confusion , slash , like quest to find what s wrong , like this kind of
searching feeling that , ok , like i know something s off . i know something s not right
. so i want to find it . i was definitely a little bit slow today , overpowered . usually ,
i m the one dominating . so it was definitely , just , i do nt know , just not good .
Ground Truth summary
eugenie bouchard suffered her fourth defeat in six matches . the canadian top seed
lost to lauren davis at the family circle cup . world no 66 davis won 6 - 3 , 6 - 1 in
the second round in charleston . davis won nine of the final 11 games of the match to
seal victory . click here for all the latest news from charleston .
No pretrain
bouchard beat american lauren davis 6 - 3 , 6 - 1 in the second round of the family
circle cup in charleston on wednesday . bouchard had lost her only career meeting
with bouchard but was in control this time against the world no 7 . bouchard won
nine of the final 11 games of the match and broke bouchard s serve twice in the final
set to pull off the upset .
Pretrained
eugenie bouchard was beaten 6 - 3 , 6 - 1 by american lauren davis in the second
round . davis had lost her only career meeting with bouchard , but was in control this
time against the world no 7 . davis hit a backhand to send back bouchard s 102mph
serve for a winner down the left sideline .
Table 7: Both models output a relevant summary, but the no pretrain model uses the same name to refer
to both players.
Source Document
( cnn ) mike rowe is coming to a river near you . sometimes , you hear about a person
who makes you feel good about humanity , but bad about yourself , rowe says . on
thursday s episode of somebody s got ta do it , rowe meets up with chad pregracke ,
the founder of living lands & waters , who does just that . pregracke wants to clean
up the nation s rivers one piece of detritus at a time . his quota ? always more . read
mike rowe s facebook post on how to break our litter habit . since he founded the
nonprofit in 1998 at the ripe age of 23 , pregracke and more than 87,000 volunteers
have collected 8.4 million pounds of trash from u.s. waterways . those efforts helped
him earn the 2013 cnn hero of the year award , along with numerous other honors
. wherever you are , no matter if there s a stream , a creek , a lake , whatever , that
needs to be cleaned up , you can do it . just organize it and do it , he told cnn s
anderson cooper after his win . pregracke also gives rowe a tour of the 150 - foot ,
solar - powered barge that the living lands & waters staff calls home during lengthy
cleanups . the part - home , part - office , part - dumpster has seven bedrooms , two
bathrooms , a classroom and a kitchen – and just happens to be made from a recycled
strip club . according to the organization s latest annual report , pregracke has made
it his mission in 2015 to remove 500,000 more pounds of trash . if you d like to help
achieve this goal , visit his website to learn how to help : livinglandsandwaters.org /
get - involved / .
Ground Truth summary
chad pregracke was the 2013 cnn hero of the year . mike rowe visited pregracke for
an episode of somebody s got ta do it .
No pretrain
rowe meets up with chad pregracke , founder of living lands & waters . pregracke
and more than 87,000 volunteers collected 8.4 million pounds of trash from u.s. waterways .
Pretrained
rowe is the founder of living lands & waters , who does just that . pregracke also
gives rowe a tour of the 150 - foot barge that the living lands & waters gets .
Table 8: A failure case. The pretrained model outputs irrelevant details while the no pretrain model
successfully summarizes the document.
Source
Mayor Bloomberg told reporters that, because of that court order, the city had suspended the reopening of the public space and protesters were informed, however, that
local laws do not allow them to re-install with camping shops and sleeping bags.
Ground Truth
Bürgermeister Bloomberg stellt vor der Presse klar , das aufgrund dieser richterlichen Anordnung die erneute Öffnung des Platzes für den Publikumsverkehr und
die Demonstranten aufgehoben worden sei . Die Demonstranten wies er darauf hin ,
dass die Stadtgesetze ihnen nicht erlaubten , sich erneut mit Zelten und Schlafsäcken
an diesem Ort einzurichten .
No pretrain
Der Bürgermeister Bloomberg sagte den Reportern , dass die Stadt aufgrund dieser
Gerichtsentscheidung die Wiedereröffnung des öffentlichen Raumes und die Information der Demonstranten ausgesetzt habe , dass die lokalen Gesetze ihnen nicht
erlauben , mit den Campingplätzen und Schlafsäcken neu zu installieren .
Pretrained
Bürgermeister Bloomberg erklärte gegenüber Journalisten , dass die Stadt aufgrund
dieser Gerichtsentscheidung die Wiedereröffnung des öffentlichen Raums ausgesetzt
habe und dass die Demonstranten darüber informiert wurden , dass die örtlichen
Gesetze es ihnen nicht erlauben würden , sich mit Campingplätzen und Schlafsälen
neu zu installieren .
Table 9: The no pretrain model makes a complete mistranslation when outputting ”und die Information
der Demonstranten ausgesetzt habe”. That translates to ”the reopening of the public space and the information [noun] of the protesters were suspended”, instead of informing the protesters. Furthermore,
it wrongly separated the two sentences, so the first sentence has extra words and the second sentence
is left without a subject. The pretrained model does not make any of these mistakes. However, both
models make a vocabulary mistake of ”zu installieren”, which is typically only used to refer to installing
software. A human evaluator fluent in both German and English said that the pretrained version was
better.
Source
The low February temperatures, not only did they cause losses of millions for the
agricultural sector, but they limited the possibilities of the state economy to grow,
causing a contraction of the economic activity in general of 3.6 percent in the first
half of the year, mainly supported by the historic fall of 31.16 per cent in agriculture,
which affected the dynamics of other economic sectors.
Ground Truth
Die niedrigen Temperaturen im Februar verursachten nicht nur Verluste in Millionenhöhe in der Landwirtschaft , sondern steckten darüber hinaus dem Wachstum der
Staatswirtschaft enge Grenzen und verursachten im ersten Vierteljahr einen allgemeinen Rückgang der Wirtschaftstätigkeit um 3,6 Prozent Dieser geht hauptsächlich
auf den historischen Abbau der landwirtschaftlichen Entwicklung um 31,16 Prozent
zurück , der sich bremsend auf weitere Wirtschaftssektoren auswirkte .
No pretrain
Die niedrigen Temperaturen im Februar führten nicht nur zu Verlusten für die Landwirtschaft , sondern sie beschränkten die Möglichkeiten der staatlichen Wirtschaft ,
wachsen zu wachsen , wodurch die Wirtschaftstätigkeit insgesamt von 3,6 Prozent
in der ersten Hälfte des Jahres , hauptsächlich durch den historischen Rückgang
von 31.16 % in der Landwirtschaft , beeinflusst wurde , was die Dynamik anderer
Wirtschaftssektoren betraf .
Pretrained
Die niedrigen Temperaturen im Februar führten nicht nur zu Verlusten von Millionen
für den Agrarsektor , sondern beschränkten die Möglichkeiten der Staatswirtschaft
, zu wachsen , was zu einer Schrumpfung der Wirtschaftstätigkeit im Allgemeinen
von 3,6 Prozent in der ersten Hälfte des Jahres führte , hauptsächlich durch den historischen Einbruch von 316 Prozent in der Landwirtschaft , der die Dynamik anderer
Wirtschaftsbereiche beeinflusst hatte .
Table 10: The human evaluator noted that the pretrained version is better, as it correctly captures the
meaning and sentence structure of the middle. The no pretrain model does not misses translating the
word ”million”, repeats itself in ”wachsen zu wachsen”, and puts the verb ”beeinflusst wurde” is an
unnatural position. However, the pretrained model makes a mistake in the percentage (316% instead of
31.16%).
Source
To facilitate the inception of the Second World War, they allowed bankers and politicians to create a latent conflict situation by saddling Germany with huge war reparations, thereby making a radicalist example of the impoverished masses, it remained
only to introduce a sufficiently convincing culprit and a leader with a simple solution,
while also creating a multi-racial Czechoslovakia with a strong German minority to
play, and indeed did, the role of a fifth colony, once the war had been ignited.
Ground Truth
Um den Zweiten Weltkrieg einfacher entfachen zu können , ließen die Banker durch
die Politik eine latente Konfliktsituation schaffen , indem sie Deutschland mit gigantischen Kriegsreparationen belegten ; dadurch schufen sie die Voraussetzung verarmter Massen , so dass sie den Deutschen nur noch einen ausreichend starken Führer
unterjubeln mussten , der die Schuldigen benannte und einfache Lösungen anbot ; ein
weiterer Faktor war die Schaffung des Vielvölkerstaates Tschechoslowakei mit einer
starken deutschen Minderheit , die die Funktion einer fünften Kolonne einnehmen
sollte und auch einnahm , um den Kriegsbrand zu entfachen .
No pretrain
Um die Gründung des Zweiten Weltkriegs zu erleichtern , ermöglichte es den
Bankern und Politikern , eine latente Konfliktlage zu schaffen , indem sie Deutschland mit enormen Reparationsforderungen konfrontierte , wodurch ein radikalislamistisches Beispiel der verarmten Massen entstand , es blieb nur , einen ausreichend aussagekräftigen Schuldigen und einen Führer mit einer einfachen Lösung zu
etablieren , während gleichzeitig eine multi-ethnische Tschechoslowakei mit einer
starken deutschen Minderheit zu spielen war und tatsächlich die Rolle einer fünften
Kolonie war .
Pretrained
Um die Einführung des Zweiten Weltkrieges zu erleichtern , ließen sie Banker
und Politiker eine latente Konfliktlage schaffen , indem sie Deutschland mit riesigen Reparationszahlungen belieferten , wodurch ein radikalislamistisches Beispiel
der verarmten Massen entstand , es blieb nur , einen ausreichend überzeugenden
Schuldigen und einen Führer mit einer einfachen Lösung zu präsentieren , während
gleichzeitig eine multiethnische Tschechoslowakei mit einer starken deutschen Minderheit geschaffen wurde , um zu spielen , und tatsächlich , die Rolle einer fünften
Kolonie , sobald der Krieg entfacht worden war .
Table 11: An example where the English source is poorly worded. Both models output poor translations,
but the evaluator noted that the pretrained version is still better than the no pretrain version. Interestingly,
both models mistranslate ”radical” as ”radikalislamistisches”, which means ”radical Islam”, which is
probably a bias in the training data.
Source
The total vote count will also be done if at the end of the ordinary calculation is
established that the difference between the winner and the candidate placed on second
position is equal to or less than one percentage point, as long as there is a request of
the representative of the political party whose candidate came on the second position,
case in which there will be excluded the electoral boxes that have been considered
during the partial recount.
Ground Truth
Die Stimmenauszählung kann auch in ihrer Gesamtheit erfolgen , wenn nach Abschluss der ordentlichen Berechnung festgestellt wird , dass der Unterschied zwischen dem mutmaßlichen Gewinner und dem Kandidaten auf dem zweiten Platz gleich
oder geringer als ein Prozent ist , vorausgesetzt es liegt ein ausdrücklicher Antrag von
einem Vertreter der Partei , deren Kandidat Zweiter geworden ist , vor . In diesem
Fall würden die Wahlpakete , die einer teilweisen Auszählung ausgesetzt wurden ,
ausgeschlossen .
No pretrain
Die gesamte Stimmenanzahl wird auch dann erreicht , wenn am Ende der ordentlichen Berechnung festgestellt wird , dass der Unterschied zwischen dem Sieger
und dem Kandidaten , der auf der zweiten Position liegt , gleich oder weniger als
einen Prozentpunkt beträgt , vorausgesetzt , dass der Vertreter der Partei , deren Kandidat auf der zweiten Position ist , der Fall ist , in dem die Wahlunterlagen , die
während der teilweisen Rückzählung berücksichtigt wurden , ausgeschlossen werden
.
Pretrained
Die Gesamtzahl der Stimmzettel wird auch dann durchgeführt , wenn am Ende der ordentlichen Berechnung festgestellt wird , dass der Unterschied zwischen dem Gewinner und dem auf den zweiten Platz platzierten Kandidaten gleich oder weniger als
einen Prozentpunkt beträgt , solange es einen Antrag des Vertreters der politischen
Partei gibt , dessen Kandidat auf die zweite Position kam , in dem es die Wahlzettel
ausklammert , die während der Teilzählung berücksichtigt wurden .
Table 12: Another example where the English source is poorly worded. Both models get the structure
right, but have a variety of problematic translations. Both models miss the meaning of ”total vote count”.
They both also translate ”electoral boxes” poorly - the no pretrain model calls it ”electoral paperwork”
while the pretrained model calls it ”ballots”. These failures may be because of the poorly worded English
source. The human evaluator found them both equally poor.
| 9cs.NE
|
On the ERM Principle with Networked Data
Yuanhong Wang1,2 , Yuyi Wang3 , Xingwu Liu4,5 , Juhua Pu1,2∗
arXiv:1711.04297v2 [cs.LG] 22 Nov 2017
1
State Key Laboratory of Software Development Environment, Beihang University, Beijing, China
2
Research Institute of Beihang University in Shenzhen, Shenzhen, China
3
Disco Group, ETH Zurich, Switzerland
4
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
5
University of Chinese Academy of Sciences, Beijing, China
Abstract
Networked data, in which every training example involves two
objects and may share some common objects with others, is
used in many machine learning tasks such as learning to rank
and link prediction. A challenge of learning from networked
examples is that target values are not known for some pairs of
objects. In this case, neither the classical i.i.d. assumption nor
techniques based on complete U-statistics can be used. Most
existing theoretical results of this problem only deal with the
classical empirical risk minimization (ERM) principle that
always weights every example equally, but this strategy leads
to unsatisfactory bounds. We consider general weighted ERM
and show new universal risk bounds for this problem. These
new bounds naturally define an optimization problem which
leads to appropriate weights for networked examples. Though
this optimization problem is not convex in general, we devise
a new fully polynomial-time approximation scheme (FPTAS)
to solve it.
1
Introduction
“No man is an island, entire of itself ...”, the beginning of a wellknown poem by the 17th century English poet John Donne,
might be able to explain why social networking websites are
so popular. These social media not only make communications convenient and enrich our lives but also bring us data,
of an unimaginable amount, that is intrinsically networked.
Social network data nowadays is widely used in research on
social science, network dynamics, and as an inevitable fate,
data mining and machine learning (Scott 2017). Similar examples of networked data such as traffic networks (Min and
Wynter 2011), chemical interaction networks (Szklarczyk et
al. 2014), citation networks (Dawson et al. 2014) also abound
throughout the machine learning world.
Admittedly, many efforts have been made to design practical algorithms for learning from networked data, e.g., (LibenNowell and Kleinberg 2007, Macskassy and Provost 2007, Li
et al. 2016, Garcia-Duran et al. 2016). However, not many theoretical guarantees of these methods have been established,
which is the main concern of this paper. More specifically,
this paper deals with risk bounds of classifiers trained with
networked data (CLANET) whose goal is to train a classifier
∗
Corresponding author
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
with examples in a data graph G. Every vertex of G is an
object and described by a feature vector X ∈ X that is drawn
independently and identically (i.i.d.) from an unknown distribution, while every edge corresponds to a training example
whose input is a pair of feature vectors (X, X 0 ) of the two
ends of this edge and whose target value Y is in {0, 1}.
A widely used principle to select a proper model from a
hypothesis set is Empirical Risk Minimization (ERM). Papa,
Bellet, and Clémençon (2016) establish risk bounds for ERM
on complete data graphs, and the bounds are independent
of the distribution of the data. These bounds are of the order O(log(n)/n), where n is the number of vertices in the
complete graph. However, in practice it is very likely that
one cannot collect examples for all pairs of vertices and then
G is usually incomplete, thus techniques based on complete
U -processes in (Papa, Bellet, and Clémençon 2016) cannot
be applied and the risk bounds of the order O(log(n)/n) are
no longer valid in this setting. By generalizing the moment
inequality for U -processes to the case of incomplete graphs,
we prove novel risk bounds for the incomplete graph.
Usually, every training example is equally weighted (or unweighted) in ERM, which seems much less persuasive when
the examples are networked, in particular when the graph is
incomplete. But, most existing theoretical results of learning
from networked examples are based on the unweighted ERM
(Usunier, Amini, and Gallinari 2006, Ralaivola, Szafranski,pand Stempfel 2009), and their bounds are of the order
O( χ∗ (DG )/m) where DG is the line graph of G and χ∗
is the fractional chromatic number of DG (see Section A
in the online appendix1 ) and m is the number of training
examples. In order to improve this bound, Wang, Guo, and
Ramon (2017) propose weighted ERM which adds weights
to training examples according to the data graph, and show
that the
p risk bound for weighted ERM can be of the order
O(1/ ν ∗ (G)) where ν ∗ (G) is the fractional matching number of G, so using weighted ERM networked data can be
more effectively exploited than the equal weighting method,
∗
as basic graph theory tells us νG
≥ m/χ∗ (DG ). However,
Wang, Guo, and Ramon (2017) (in fact, Usunier, Amini, and
Gallinari (2006) and Ralaivola, Szafranski, and Stempfel
(2009) also) assume that any two examples can be arbitrarily
correlated if they share a vertex, which cannot lead to an
1
https://arxiv.org/abs/1711.04297
O(log(n)/n) bound when the graph is complete.
We show that the “low-noise” condition, also called the
Mammen-Tsybakov noise condition (Mammen and Tsybakov
1998), which is commonly assumed in many typical learning
problems with networked data, e.g., ranking (Clémençon,
Lugosi, and Vayatis 2008) and graph reconstruction (Papa,
Bellet, and Clémençon 2016), can be used to reasonably
bound the dependencies of two examples that share a common vertex and then leads to tighter risk bounds.
In summary, in this paper we mainly
Though Wang, Guo, and Ramon (2017) show improved
risk bounds, as far as we know, there is no known tight risk
bound on incomplete graphs for tasks such as pairwise ranking and graph reconstruction that satisfy the “low-noise” condition. Under this condition, the method proposed in (Wang,
Guo, and Ramon 2017) does not work (see Section 6.1).
Before we show new risk bounds and new weighting methods, we present the following three aspects to convey some
intuitions.
• prove new universal risk bounds for CLANET which
Line Graphs Compared to Janson’s decomposition which
is based on line graphs, our method utilizes the additional dependency information in the data graph G. For example, the
complete line graph with three vertices (i.e., triangle) corresponds to two different data graphs, as illuminated in Figure 1.
Hence, line graph based methods ignore some important information in the data graph. This negligence makes it unable
to improve bounds, no matter whether considering weighted
ERM or not (see Section A.1 in the online appendix). In
Section 6.2, we show that our bounds are tighter than that of
line graph based methods.
– can be applied to learning from networked data even if
the data graph is incomplete;
– exploit the property of the “low-noise” condition, and
then become tighter than previous results;
– allow non-identical weights on different examples, so
it is possible to achieve better learning guarantee by
choosing these weights.
• formulate a non-convex optimization problem inspired by
our new risk bounds (because our risk bounds depend
on the weights added to every training example, and a
better weighting scheme leads to a tighter bound), and
then we also design a new efficient algorithm to obtain
an approximate optimal weighting vector and show that
this algorithm is a fully polynomial-time approximation
scheme for this non-convex program.
2
Intuitions
We now have a look at previous works that are closely related
to our work, as shown in Table 1, and present the merits
of our method. Biau and Bleakley (2006), Clémençon, Lugosi, and Vayatis (2008) and Papa, Bellet, and Clémençon
(2016) deal with the case when the graph is complete, i.e.,
the target value of every pair of vertices is known. In this
case, Clémençon, Lugosi, and Vayatis (2008) formulate the
“low-noise” condition for the ranking problem and demonstrate that this condition can lead to tighter risk bounds by
the moment inequality for U -processes. Papa, Bellet, and
Clémençon (2016) further consider the graph reconstruction
problem introduced by Biau and Bleakley (2006) and show
this problem always satisfies the “low-noise” condition.
If the graph is incomplete, one can use either Janson’s
decomposition (Janson 2004, Usunier, Amini, and Gallinari
2006, Ralaivola, Szafranski, and Stempfel 2009, Ralaivola
and Amini 2015) or the fractional matching approach by
Wang, Guo, and Ramon (2017) to derive risk bounds. The
main differences between these two approaches are:
• Wang, Guo, and Ramon (2017) consider the data graph
G while Janson’s decomposition uses only the line graph
DG .
• The fractional matching approach considers weighted
ERM while Janson (2004), Usunier, Amini, and Gallinari (2006), Ralaivola, Szafranski, and Stempfel (2009)
and Ralaivola and Amini (2015) only prove bounds for
unweighted ERM.
Asymptotic Risk As mentioned by Wang, Guo, and Ramon (2017), if several examples share a vertex, then we are
likely to put less weight on them because the influence of
this vertex to the empirical risk should be bounded. Otherwise, if we treat every example equally, then these dependent
examples may dominate the training process and lead to the
risk bounds that do not converge to 0 (see the example in
Section 6.2).
Uniform Bounds Ralaivola and Amini (2015) prove an
entropy-base concentration inequality for networked data
using Janson’s decomposition, but the assumption there is
usually too restrictive to be satisfied (see Section A.2 in the
online appendix). To circumvent this problem, our method
uses the “low-noise” condition (also used in (Papa, Bellet, and
Clémençon 2016)) to establish uniform bounds, in absence
of any restrictive condition imposed on the data distribution.
3
Preliminaries
In this section, we begin with the detailed probabilistic framework for CLANET, and then give the definition of weighted
ERM on networked examples.
3.1
Problem Statement
Consider a graph G = (V, E) with a vertex set V =
{1, . . . , n} and a set of edges E ⊆ {{i, j} : 1 ≤ i 6= j ≤ n}.
For each i ∈ V , a continuous random variable (r.v.) Xi , taking its values in a measurable space X , describes features of
vertex i. The Xi ’s are i.i.d. r.v.’s following some unknown
distribution PX . Each pair of vertices (i, j) ∈ E corresponds
to a networked example whose input is a pair (Xi , Xj ) and
target value is Yi,j ∈ Y. We focus on binary classification
in this paper, i.e., Y = {0, 1}. Moreover, the distribution of
target values only depends on the features of the vertices it
contains but does not depend on features of other vertices,
Principles
Unweighted ERM
(equally weighted)
Weighted ERM
Table 1: Summary of methods for CLANET.
With “low-noise”
Graph type
condition
Complete
graphs
Clémençon, Lugosi, and Vayatis (Ann. Stat. 2008),
Papa, Bellet, and Clémençon
(NIPS 2016)
General
graphs
Ralaivola and Amini (ICML
2015)
General
graphs
This paper
Without “low-noise”
condition
Biau and Bleakley
(Statistics and Decisions
2006)
Usunier, Amini, and Gallinari (NIPS 2006),
Ralaivola, Szafranski, and
Stempfel (AISTATS 2009)
Wang, Guo, and Ramon
(ALT 2017)
that is, there is a probability distribution PY|X 2 such that for
every pair (i, j) ∈ E, the conditional probability
(Hoeffding 1948) of its excess risk can be always obtained if
the problem is symmetric (see Lemma 2).
P [Yi,j = y | x1 , . . . , xn ] = PY|X 2 [y, xi , xj ] .
Definition 2 (symmetry). A learning problem is symmetric if for every xi , xj ∈ X , yi,j ∈ Y and r ∈ R,
`(r, (xi , xj , yi,j )) = `(r, (xj , xi , yj,i )).
Example 1 (pairwise ranking). (Liu 2009) categorize ranking problems into three groups by their input representations
and loss functions. One of these categories is pairwise ranking that learns a binary classifier telling which document
is better in a given pair of documents. A document can be
described by a feature vector from the X describing title,
volume, . . . The target value (rank) between two documents,
that only depends on features of these two documents, is 1 if
the first document is considered better than the second, and
0 otherwise.
The training set S := {(Xi , Xj , Yi,j )}(i,j)∈E is dependent
copies of a generic random vector (X1 , X2 , Y1,2 ) whose distribution P = PX ⊗ PX ⊗ PY|X 2 is fully determined by
the pair (PX , PY|X 2 ). Let R be the set of all measurable
functions from X 2 to Y and for all r ∈ R, the loss function
`(r, (x1 , x2 , y1,2 )) = 1y1,2 6=r(x1 ,x2 ) .
Given a graph G with training examples S and a hypothesis
set R ⊆ R, the CLANET problem is to find a function r ∈ R,
with risk
L(r) := E[`(r, (X1 , X2 , Y1,2 ))]
(1)
that achieves a comparable performance to the Bayes rule
r∗ = arg inf r∈R L(r) = 1η(x1 ,x2 )≥1/2 , whose risk is denoted by L∗ , where η(x1 , x2 ) = PY|X 2 [1, x1 , x2 ] is the regression function.
The main purpose of this paper is to devise a principle to
select a classifier r̂ from the hypothesis set R and establish
bounds for its excess risk L(r̂) − L∗ .
Definition 1 (“low-noise” condition). Let us consider a
learning problem, in which the hypothesis set is F and
the Bayes rule is f ∗ . With slightly abusing the notation,
this problem satisfies the “low-noise” condition if ∀f ∈
F, L(f ) − L∗ ≥ C θ (E[|f − f ∗ |])θ where C is a positive
constant.
As mentioned, the “low-noise” condition can lead to tighter
risk bounds. For this problem, we show that the “low-noise”
condition for the i.i.d. part of the Hoeffding decomposition
Many typical learning problems are symmetric. For example, pairwise ranking problem with symmetric functions r
in the sense that r(X1 , X2 ) = 1 − r(X2 , X1 ) satisfies the
symmetric condition.
3.2
Weighted ERM
ERM aims to find the function from a hypothesis set that minimizes the empirical estimator of (1) on the training examples
S = {(Xi , Xj , Yi,j )}(i,j)∈E :
Lm (r) :=
1 X
`(r, (Xi , Xj , Yi,j )).
m
(2)
(i,j)∈E
where m is the number of training examples. In this paper,
we consider its weighted version, in which we put weights
on the examples and select the minimizer rw of the weighted
empirical risk
Lw (r) :=
1
kwk1
X
wi,j `(r, (Xi , Xj , Yi,j ))
(3)
(i,,j)∈E
where w is a fractional matching of G and kwk1 > 0.
Definition 3 (fractional matching). Given a graph G =
(V, E), a fractional matching w is a non-negative
vector
P
(wi,j )(i,j)∈E that for every vertex i ∈ V, j:(i,j)∈E wi,j ≤
1.
4
Universal Risk Bounds
In this section, we use covering numbers as the complexity
measurement of hypothesis sets to prove that tighter universal
risk bounds are always attained by the minimizers of the
weighted empirical risk (3).
4.1
Covering Numbers
∗
The excess risk L(rw ) − L depends on the hypothesis set
R whose complexity can be measured by covering number
(Cucker and Zhou 2007). A similar but looser result using VCdimension (Vapnik and Chervonenkis 1971) can be obtained
as well.
Definition 4 (covering numbers). Let (F, Lp ) be a metric
space with Lp -pseudometric. We define the covering number
N (F, Lp , ) be the minimal l ∈ N such that there exist l
disks in F with radius covering F. If the context is clear,
we simply denote N (F, Lp , ) by Np (F, ).
In this paper, we focus on the L∞ covering number
N∞ (F, ) and suppose that it satisfies the following assumption.
Assumption 1. There exists a nonnegative number β < 1
and a constant K such that log N∞ (F, ) ≤ K−β for all
∈ (0, 1].
Similar to (Massart and Nédélec 2006) and (Rejchel 2012),
we restrict to β < 1, whereas in the empirical process theory
this exponent usually belongs to [0, 2). This restriction is
needed to prove Lemma 1, which involves the integral of
log N∞ (F, ) through 0. Dudley (1974), Korostelev and Tsybakov (1993) and Mammen and Tsybakov (1995) presented
various examples of classes F satisfying Assumption 1. We
also refer interested readers to (Mammen and Tsybakov 1998,
p. 1813) for more concrete examples of hypothesis classes
with smooth boundaries satisfying Assumption 1.
4.2
order O((1/n)1/(1+β) ) achieve the same results as in (Papa,
Bellet, and Clémençon 2016)2
Remark. Theorem 1 provides universal risk bounds no matter what the distribution of the data is. The factor of 2 in front
of the approximation error inf r∈R L(r) − L∗ has no special
meaning and can be replaced by any constant larger than
1 with a cost of increasing the constant C. Wang, Guo, and
Ramon (2017) obtain risk bounds that has a factor 1 in front
of the approximation
error part, but in their result the bound
p
is O(1/ kwk1 ). Hence, Theorem 1 improves their results if
the approximation error does not dominate the other terms
in the bounds.
In the rest of this section, we outline the main ideas to
obtain this result. We first define
qr (x1 , x2 , y1,2 ) := `(r, x1 , x2 , y1,2 ) − `(r∗ , x1 , x2 , y1,2 )
for every (x1 , x2 , y1,2 ) ∈ X ×X ×Y and let Λ(r) := L(r)−
L∗ = E[qr (X1 , X2 , Y1,2 )] be the excess risk with respect to
the Bayes rule. Its empirical estimate by weighted ERM is
Λw (r) = Lw (r) − Lw (r∗ )
X
1
=
wi,j qr (Xi , Xj , Yi,j ).
kwk1
(i,j)∈E
By Hoeffding’s decomposition (Hoeffding 1948), for all
r ∈ R, one can write
ew (r),
Λw (r) = Tw (r) + Uw (r) + U
where
Risk Bounds
Now we are ready to show the tighter risk bounds for
weighted empirical risk by the following theorem.
Theorem 1 (risk bounds). Let rw be a minimizer of the
weighted empirical risk Lw over a class R that satisfies
Assumption 1. There exists a constant C > 0 such that for all
δ ∈ (0, 1], with probability at least 1 − δ, the excess risk of
rw satisfies
K 0 C log(1/δ)
L(rw ) − L∗ ≤ 2( inf L(r) − L∗ ) +
r∈R
(1 − β)2/(β+1) kwk1
β/(1+β)
kwk1
+ max kwk2 , kwkmax (log(1/δ))1/2 ,
kwk∞ (log(1/δ))
where kwkmax = maxi
√
max(K, K, K 1/(1+β) ).
(5)
qP
j:(i,j)∈E
2
wi,j
and K 0
(4)
=
According to Theorem 1, if the parameter δ is greater than
the value exp − min(kwk2 /kwk∞ , kwk22 /kwk2max ) ,
then the risk bounds above are of the order
O (1/kwk1 )1/(1+β) + kwk2 /kwk1 . In this case,
p
our bounds arep tighter than O(1/ kwk1 ) as
kwk2 /kwk1 ≤ 1/ kwk1 (recall that w must be a
fractional matching and 0 < β < 1). If G is complete
and every example is equally weighted, the bounds of the
Tw (r) = Λ(r) +
n
2 X
kwk1 i=1
X
wi,j hr (Xi )
j:(i,j)∈E
is a weighted average of i.i.d. random variables with
hr (Xi ) = E[qr (Xi , Xj , Yi,j ) | Xi ] − Λ(r),
X
1
Uw (r) =
wi,j (ĥr (Xi , Xj )
kwk1
(i,j)∈E
is a weighted degenerated (i.e., the symmetric kernel
ĥr (x1 , x2 ) such that E[ĥr (X1 , X2 ) | X1 = x1 ] = 0 for
all x1 ∈ X ) U -statistic ĥr (Xi , Xj ) = E[qr (Xi , Xj , Yi,j ) |
Xi , Xj ] − Λ(r) − hr (Xi ) − hr (Xj ) and
X
ew (r) = 1
U
wi,j h̃r (Xi , Xj , Yi,j )
kwk1
(i,j)∈E
with a degenerated kernel h̃r (Xi , Xj , Yi,j ) = qr (Xi , Xj
, Yi,j ) − E[qr (Xi , Xj , Yi,j ) | Xi , Xj ]. In the following, we
ew in (5) respectively.
bound the three terms Tw , Uw and U
Lemma 1 (uniform approximation). Under the same assumptions as in Theorem 1, for any δ ∈ (0, 1/e), we have with
probability at least 1 − δ,
√
kwk log(1/δ)
max(K, K)C1
2
sup|Uw (r)|≤
max
,
1
−
β
kwk
1
r∈R
2
They consider the same range of δ.
kwkmax (log(1/δ))3/2 kwk∞ (log(1/δ))2
,
kwk1
kwk1
Let A be the event that
sup|Uw (r)|≤ κ1 ,
r∈R
and
√
max(K, K)C2 kwk2
e
sup|Uw (r)|≤
1−β
kwk1
r∈R
kwk
3/2
kwk∞ (log(1/δ))2
max (log(1/δ))
,
+ max
kwk1
kwk1
where
κ1 =
where C1 , C2 < +∞ are constants.
ew (r) can be
To prove Lemma 1, we show that Uw (r) and U
bounded by Rademacher chaos using classical symmetrization and randomization tricks combined with the decoupling
method. We handle these Rademacher chaos by generalizing
the moment inequality for U -statistics in (Clémençon, Lugosi, and Vayatis 2008). Specifically, we utilize the moment
inequalities from (Boucheron et al. 2005) to convert them
into a sum of simpler processes, which can be bounded by the
metric entropy inequality for Khinchine-type processes (see
Arcones and Gine 1993, Proposition 2.6) and Assumption 1.
The detailed proofs can be found in Section C in the online
appendix.
Lemma 1 shows that the contribution of the degenerated
ew (r) to the excess risk can be bounded.
parts Uw (r) and U
This implies that minimizing Λw (r) is approximately equivalent to minimizing Tw (r) and thus rw is a ρ-minimizer of
Tw (r) in the sense that Tw (rw ) ≤ ρ + inf r∈R Tw (r). In
order to analyze Tw (r), which can be treated as a weighted
empirical risk on i.i.d. examples, we generalize the results in
(Massart and Nédélec 2006) (see Section B in the online appendix). Based on this result, tight bounds for the excess risk
with respect to Tw (r) can be obtained if the variance of the
excess risk is controlled by its expected value. By Lemma 2,
Tw (r) fulfills this condition, which leads to Lemma 3.
Lemma 2 (condition leads to “low-noise”, (Papa, Bellet,
and Clémençon 2016, Lemma 2)). If the learning problem
CLANET is symmetric, then
Var [E[qr (X1 , X2 , Y1,2 ) | X1 ]] ≤ Λ(r)
(6)
holds for any distribution P and any function r ∈ R.
Lemma 3 (risk bounds for i.i.d. examples). Suppose that r0
is a ρ-minimizer of Tw (r) in the sense that Tw (r0 ) ≤ ρ +
inf r∈R Tw (r) and R satisfies Assumption 1, then there exists
a constant C such that for all δ ∈ (0, 1], with probability at
least 1 − δ, the risk of r0 satisfies
kwk log(1/δ) kwk
3/2
C1
2
max (log(1/δ))
max
,
,
1−β
kwk1
kwk1
kwk∞ (log(1/δ))2
kwk1
for an appropriate constant C1 . Then by Lemma 1, P [A] ≥
1 − δ/4. Similarly, let B be the event that
ew (r)|≤ κ2 .
sup|U
r∈R
where
C2
κ2 =
1−β
kwk
3/2
kwk2
max (log(1/δ))
+ max
,
kwk1
kwk1
kwk∞ (log(1/δ))2
kwk1
for an appropriate constant C2 . Then P [B] ≥ 1 − δ/4.
By (5), it is clear that, if both A and B happen, rw is
a ρ-minimizer of Tw (r) over r ∈ R in the sense that the
difference between the value of this latter quantity at its
minimum and rw is at most (κ1 + κ2 ). Then, from Lemma 3,
with probability at least 1 − δ/2, rw is a (κ1 + κ2 )-minimizer
of Tw (r), which the result follows.
An intuition obtained from our result is how to choose
weights for networked data. By Theorem 1, to obtain tight
risk bounds, we need to maximize kwk1 (under the constraint that this weight vector is a fractional matching),
which resembles the result of (Wang, Guo, and Ramon
2017) (but they only needpto maximize kwk1 and this is
why they end in the O(1/ ν ∗ (G)) bound), while making
kwk2 , kwkmax , kwk∞ as small as possible, which appears
to suggest putting nearly average weights on examples and
vertices respectively. These two objectives, maximizing kwk1
and minimizing kwk2 , kwkmax , kwk∞ , seem to contradict
each other. In the next section, we discuss how to solve this
problem.
5
Weighting Vector Optimization
With Lemma 1 Lemma 3, now we are ready to prove
Theorem 1.
In this section, we first formulate the optimization problem
that minimizes the risk bounds in Theorem 1. Although this
optimization problem is not convex unless β = 0, which
usually means that there is no general efficient way to solve
it, we devise a fully polynomial-time approximation scheme
(FPTAS) to solve it.
Proof of Theorem 1. Let us consider the Hoeffding decomposition (5) of Λw (r) that is minimized over r ∈ R. The
idea of this proof is that the degenerate parts Uw (r) and
ew (r) can be bounded by Lemma 1. Therefore, rw is an
U
approximate minimizer of Tw (r), which can be handled by
Lemma 3.
Definition 5 (FPTAS). An algorithm A is a FPTAS for a
minimization problem Π, if for any input I of Π and > 0,
A finds a solution s in time polynomial in both the size of I
and 1/ that satisfies fΠ (s) ≤ (1 + ) · fΠ (s∗ ), where fΠ is
the (positive) objective function of Π and s∗ is an optimal
solution for I.
CK 1/(1+β) log(1/δ)
Λ(r ) ≤ 2 inf Λ(r) + 2ρ +
.
r∈R
(kwk1 (1 − β)2 )1/(1+β)
0
5.1
Optimization Problem
According to Theorem 1, given a graph G, β ∈ (0, 1) and
δ ∈ (0, 1], one can find a good weighting vector with tight
risk bounds by solving the following program:
1
β/(1+β)
min
kwk1
+ max kwk2 ,
w
kwk1
1/2
kwkmax (log(1/δ)) , kwk∞ (log(1/δ))
X
s.t. ∀(i, j) ∈ E, wi,j ≥ 0 and ∀i,
wi,j ≤ 1
Algorithm 1 FPTAS for weighting vector optimization.
Input: , β, δ and a graph G that contains n vertices and m
edges.
Output: An approximate optimal weighting vector p̄ for the
program (8).
1: Solve the following linear program (LP) efficiently3 , and
obtain an -approximation amin ;
min a
a,p
s.t. ∀(i, j) ∈ E, pi,j ≥ 0
X
∀i,
pi,j − a ≤ 0
j:(i,j)∈E
(7)
To get rid of the fraction of norms in the program above,
we consider a distribution p on P
edges pi,j := wi,j /kwk1
and then kwk1 ≤ 1/maxi=1,...,n j:(i,j)∈E pi,j . Every distribution p corresponds to a valid weighting vector w. By
introducing two auxiliary variables a and b, solving the original program (7) is equivalent to solving
min
a,b,p
s.t.
∀(i, j) ∈ E, pi,j ≥ 0
∀(i, j) ∈ E, pi,j log(1/δ) − b ≤ 0
X
∀i,
pi,j − a ≤ 0
j:(i,j)∈E
(8)
p2i,j log(1/δ)
−b≤0
∀(i, j) ∈ E, pi,j ≥ 0
∀(i, j) ∈ E, pi,j log(1/δ) − b ≤ 0
X
∀i,
pi,j − a ≤ 0
j:(i,j)∈E
(10)
1/2
and
X
pi,j = 1
Note that the constraints are all convex. If β = 0, e.g., the
hypothesis set is finite, then the objective function becomes
linear and thus (8) is a convex optimization problem that
can be solved by some convex optimization method (see
e.g., (Boyd and Vandenberghe 2004)) such as interior-point
method.
If β > 0, the objective function is not convex any more. In
fact, the program (8) becomes a concave problem that may
be optimized globally by some complex algorithms (Benson
1995, Hoffman 1981) that often need tremendous computation. Instead, one may only need to approximate it using
some efficient methods, e.g., Concave-Convex Procedure
(Yuille 2001) and Coordinate Descent (Wright 2015). However, these methods lack in complexity analysis and may lead
to a local optimum.
A Fully Polynomial-time Approximation
Scheme
To solve the program (8) efficiently, we propose Algorithm
1 and show that it is a fully polynomial-time approximation
scheme for (8).
Theorem 2. Algorithm 1 is a FPTAS for the program (8).
3
n(1 − amin )/(1 + β)} and Solutions := ∅.
3: for a ∈ Grid do
Use some efficient interior-point method to obtain an
4:
s.t.
(i,j)∈E
5.2
2: Let Grid := {amin + i · (1 + β)/n | i ∈ N and i ≤
b,p
j:(i,j)∈E
kpk2 −b ≤ 0
pi,j = 1
(i,j)∈E
min b
1/2
X
X
-approximation of the following program and add the
solution (a, b, p) into Solutions.
a1/(1+β) + b
∀i,
(9)
j:(i,j)∈E
For example, some interior-point method.
∀i,
X
p2i,j log(1/δ)
−b≤0
j:(i,j)∈E
kpk2 −b ≤ 0
and
X
pi,j = 1
(i,j)∈E
5: return the vector p̄ which makes a1/(1+β) + b smallest
from Solutions.
Proof. We first analyze the running time of this algorithm.
Note that 1/n ≤ a ≤ 1 if the graph is not empty. In
Algorithm 1, we first divide the problem into at most
1 − 1/n
n−1
=
(1 + β)/n
(1 + β)
convex programs, each of which produces an -approximate
solution by some interior-point method. Since interior-point
method is FPTAS for convex problems (Boyd and Vandenberghe 2004), solving each of these programs needs polynomial time in the problem size m + n and 1/. Thus, the
complexity of Algorithm 1 is also polynomial in m + n and
1/.
Now we show that this algorithm indeed results in an approximation of this optimal solution.
For any optimal solution (a∗ , b∗ , p∗ ), if a∗ achieves minimum for the program (9), we can find a0 in Grid (actually
amin ) such that
X1
0 1/(1+β)
(a )
1/(1+β)
≤ (1 + )
(a )
∗ 1/(1+β)
≤ (1 + )(a )
X2
X1
X2
X1
X4
X3
X2
∗ 1/(1+β)
.
0
X3
(11)
X1
∗
0
Otherwise, we can also find a in Grid such that a ≤ a <
a∗ + (1 + β)/n and thus
X2
X3
(a)
X4
(b)
X3
(c)
X4
(d)
(a0 )1/(1+β) ≤ (a∗ + (1 + β)/n)1/(1+β)
≤ (a∗ )1/(1+β) + (1 + β)/n
1
(a∗ )−β/(1+β)
1+β
(12)
≤ (1 + )(a∗ )1/(1+β)
The third inequality follows from the fact that 1/n ≤ a∗ .
We assume that the optimal solution for the program (10) is
b = b0 when we fix a = a0 . Because (a∗ , b∗ ) is feasible and
a0 > a∗ , (a0 , b∗ ) is always a feasible solution for the program
(10), which leads to b0 ≤ b∗ . Besides, interior-point method
can produce an -approximate solution b00 such that
b00 ≤ (1 + )b0 ≤ (1 + )b∗ .
(13)
Finally, we select the best approximate weighting vector p̄
from all solutions in Solutions. Combining (11), (12) and
(13), we have the objective value for p̄
(ap̄ )1/(1+β) + bp̄ ≤ (a0 )1/(1+β) + b00
≤ (1 + )((a∗ )1/(1+β) + b∗ ).
6
Discussion
In this section, we first show that, according to our bounds,
equal weighting is indeed the best weighting scheme for
complete graphs. Then, we discuss the performance of this
equal weighting scheme when the graph is incomplete.
6.1
Complete Graphs
When graph G is complete, weighting all examples
equally gives
P the best risk bound, as all the terms
maxi=1,...,n j:(i,j)∈E pi,j , kpk2 , kpkmax and kpk∞
achieve minimum. Compared to the results in (Wang, Guo,
and Ramon 2017), our theory puts additional constrains on
kpk2 , kpkmax and kpk∞ which encourages weighting examples fairly in this case, as illustrated in Figure 1. Besides,
this scheme, which coincides with U -statistics that average
the basic estimator applied to all sub-samples, produces the
smallest variance among all unbiased estimators (Hoeffding
1948).
6.2
Figure 1: (a) and (b) are two different data graphs, but both of
them correspond to the same line graph (a triangle). (c) and
(d) are two weighting schemes for a complete graph formed
by points X1 , X2 , X3 , X4 . Solid line means its weight p > 0
while dash line means p = 0. (c): Weight every example
equally. (d): Only the two examples in an independent subset get equally non-zero weights and
P other weights are 0
(dashed line). Note that maxi=1,...,n j:(i,j)∈E pi,j of these
two weighting schemes are the same, but (c) has the tighter
risk bounds, as kpk2 , kpkmax and kpk∞ of (c) are smaller
than that of (d) respectively.
Equal Weighting
Let us discuss further the equal weighting scheme that gives
every example the same weight. Denote by ∆(G) the maximum degree of G (note that this is not the maximum degree
of DG ) and let pi,j = 1/m (recall that m is the number of
examples) for all (i, j) ∈ E. According to program (8), using
equal weighting scheme, the risk bounds are of the order
∆(G) 1/(1+β)
1
O (
)
+√
,
(14)
m
m
if δ ∈ (exp(−m/∆(G)), 1]. In some cases, such as bounded
degree graphs and complete graphs, this scheme provides
reasonable risk bounds. Note that ∆(G) is smaller than the
maximum size of cliques in its corresponding line graph DG
and χ∗ (DG ) is larger than the maximum size of cliques in
DG , these bounds
p above are always better than the bounds of
the order O( χ∗ (DG )/m) built by Janson’s decomposition.
However, as argued in Section 2, one can construct examples to illustrate that if we use the equal weighting strategy
when ∆(G) is large (e.g., if it is linear to m), the risk bounds
(14) are very large and do not converge to 0, while this problem can be solved by simply using a better weighting strategy.
Example 2. Consider a data graph with |E|= m 1 and E
consists of m/2 disjoint edges and m/2 edges sharing a common vertex, then ∆(G) = m/2. Using the equal weighting
scheme, the risk bounds are of the order O(1) that is meaningless. A much better weighting scheme of this case is to weight
the examples of disjoint edges with 2/(m + 2) while weight
the examples of adjacent edges with
pro 4/m(m + 2), which
p
1/(1+β)
vides risk bounds of the order O (1/m)
+ 1/m .
7
Conclusion
In this paper, we consider weighted ERM of the symmetric
CLANET problem and establish new universal risk bounds under the “low-noise” condition. These new bounds are tighter
in the case of incomplete graphs and can be degenerate to
the known tightest bound when graphs are complete. Based
on this result, one can train a classifier with a better risk
bound by putting proper weights on training examples. We
propose an efficient algorithm to obtain the approximate optimal weighting vector and prove that the algorithm is a FPTAS
for the weighting vector optimization problem. Finally, we
discuss two cases to show the merits of our new risk bounds.
Acknowledgments
The authors would like to thank the anonymous reviewers
for their helpful comments. This work was supported by
Guangdong Shannon Intelligent Tech. co., Ltd., National Key
R&D Program of China (2017YFC0803700), National Natural Science Foundation of China (61502320, 61420106013,
61521092) and Science Foundation of Shenzhen City in
China (JCYJ20160419152942010).
References
[Arcones and Gine 1993] Arcones, M. A., and Gine, E. 1993.
Limit theorems for u-processes. The Annals of Probability
1494–1542.
[Benson 1995] Benson, H. P. 1995. Concave Minimization:
Theory, Applications and Algorithms. Springer US.
[Biau and Bleakley 2006] Biau, G., and Bleakley, K. 2006.
Statistical inference on graphs. Statistics & Decisions
24(2):209–232.
[Boucheron et al. 2005] Boucheron, S.; Bousquet, O.; Lugosi,
G.; Massart, P.; et al. 2005. Moment inequalities for functions
of independent random variables. The Annals of Probability
33(2):514–560.
[Bousquet 2002] Bousquet, O. 2002. A bennett concentration inequality and its application to suprema of empirical
processes. Comptes Rendus Mathematique 334(6):495–500.
[Boyd and Vandenberghe 2004] Boyd, S., and Vandenberghe,
L. 2004. Convex optimization. Cambridge university press.
[Clémençon, Lugosi, and Vayatis 2008] Clémençon, S.; Lugosi, G.; and Vayatis, N. 2008. Ranking and empirical
minimization of u-statistics. Annals of Statistics 36(2):844–
874.
[Cucker and Zhou 2007] Cucker, F., and Zhou, D. X. 2007.
Learning theory: an approximation theory viewpoint, volume 24. Cambridge University Press.
[Dawson et al. 2014] Dawson, S.; Gašević, D.; Siemens, G.;
and Joksimovic, S. 2014. Current state and future trends:
A citation network analysis of the learning analytics field.
In Proceedings of the Fourth International Conference on
Learning Analytics and Knowledge, 231–240.
[De la Pena and Giné 2012] De la Pena, V., and Giné, E.
2012. Decoupling: from dependence to independence.
Springer Science & Business Media.
[Dudley 1974] Dudley, R. M. 1974. Metric entropy of some
classes of sets with differentiable boundaries. Journal of
Approximation Theory 10(3):227–236.
[Garcia-Duran et al. 2016] Garcia-Duran, A.; Bordes, A.;
Usunier, N.; and Grandvalet, Y. 2016. Combining two
and three-way embedding models for link prediction in
knowledge bases. Journal of Artificial Intelligence Research
55:715–742.
[Hoeffding 1948] Hoeffding, W. 1948. A class of statistics
with asymptotically normal distribution. The Annals of Mathematical Statistics 293–325.
[Hoffman 1981] Hoffman, K. L. 1981. A method for globally
minimizing concave functions over convex sets. Mathematical Programming 20(1):22–32.
[Janson 2004] Janson, S. 2004. Large deviations for sums of
partly dependent random variables. Random Structures &
Algorithms 24(3):234–248.
[Korostelev and Tsybakov 1993] Korostelev, A. P., and Tsybakov, A. B. 1993. Minimax Theory of Image Reconstruction.
Springer-Verlag.
[Li et al. 2016] Li, J.; Hu, X.; Wu, L.; and Liu, H. 2016.
Robust unsupervised feature selection on networked data. In
Proceedings of the 2016 SIAM International Conference on
Data Mining, 387–395.
[Liben-Nowell and Kleinberg 2007] Liben-Nowell, D., and
Kleinberg, J. 2007. The link-prediction problem for social networks. Journal of the Association for Information
Science and Technology 58(7):1019–1031.
[Liu 2009] Liu, T.-Y. 2009. Learning to rank for information
retrieval. Foundations and Trends R in Information Retrieval
3(3):225–331.
[Macskassy and Provost 2007] Macskassy, S. A., and
Provost, F. 2007. Classification in networked data: A toolkit
and a univariate case study. Journal of Machine Learning
Research 8(May):935–983.
[Mammen and Tsybakov 1995] Mammen, E., and Tsybakov,
A. B. 1995. Asymptotical minimax recovery of sets with
smooth boundaries. Annals of Statistics 23(2):502–524.
[Mammen and Tsybakov 1998] Mammen, E., and Tsybakov,
A. B. 1998. Smooth discrimination analysis. Annals of
Statistics 27(6):1808–1829.
[Massart and Nédélec 2006] Massart, P., and Nédélec, É.
2006. Risk bounds for statistical learning. Annals of Statistics
2326–2366.
[Min and Wynter 2011] Min, W., and Wynter, L. 2011. Realtime road traffic prediction with spatio-temporal correlations.
Transportation Research Part C: Emerging Technologies
19(4):606–616.
[Papa, Bellet, and Clémençon 2016] Papa, G.; Bellet, A.; and
Clémençon, S. 2016. On graph reconstruction via empirical
risk minimization: Fast learning rates and scalability. In
Advances in Neural Information Processing Systems, 694–
702.
[Ralaivola and Amini 2015] Ralaivola, L., and Amini, M.-R.
2015. Entropy-based concentration inequalities for dependent
variables. In International Conference on Machine Learning,
2436–2444.
[Ralaivola, Szafranski, and Stempfel 2009] Ralaivola,
L.;
Szafranski, M.; and Stempfel, G. 2009. Chromatic pac-bayes
bounds for non-iid data. In Artificial Intelligence and
Statistics, 416–423.
[Rejchel 2012] Rejchel, W. 2012. On ranking and generalization bounds. Journal of Machine Learning Research
13(May):1373–1392.
[Scott 2017] Scott, J. 2017. Social network analysis. Sage.
[Szklarczyk et al. 2014] Szklarczyk, D.; Franceschini, A.;
Wyder, S.; Forslund, K.; Heller, D.; Huerta-Cepas, J.; Simonovic, M.; Roth, A.; Santos, A.; Tsafou, K. P.; et al. 2014.
String v10: protein–protein interaction networks, integrated
over the tree of life. Nucleic Acids Research 43(D1):D447–
D452.
[Tsybakov 2004] Tsybakov, A. B. 2004. Optimal aggregation
of classifiers in statistical learning. Annals of Statistics 135–
166.
[Usunier, Amini, and Gallinari 2006] Usunier, N.; Amini,
M.-R.; and Gallinari, P. 2006. Generalization error bounds
for classifiers trained with interdependent data. In Advances
in neural information processing systems, 1369–1376.
[Vapnik and Chervonenkis 1971] Vapnik, V., and Chervonenkis, A. Y. 1971. On the uniform convergence of relative frequencies of events to their probabilities. Theory of
Probability & Its Applications 16(2):264–280.
[Wang, Guo, and Ramon 2017] Wang, Y.; Guo, Z.-C.; and
Ramon, J. 2017. Learning from networked examples. In Proceedings of the 28th International Conference on Algorithmic
Learning Theory, to appear.
[Wright 2015] Wright, S. J. 2015. Coordinate descent algorithms. Mathematical Programming 151(1):3–34.
[Yuille 2001] Yuille, A. L. 2001. The concave-convex procedure (cccp). In Advances in Neural Information Processing
Systems 14, 915 – 936.
Appendix
This appendix is organized as follows. Section A provides
new risk bounds for weighted ERM following the Janson’s
decomposition and show that it cannot improve prior results.
Section B establishes universal risk bounds for weighted
ERM on i.i.d. examples under the ”low-noise” condition
and proves upper bounds involved in covering number for
weighted empirical processes. Section C presents the classical
symmetrization and randomization tricks and decoupling inequality for degenerated weighted U -processes and the degenew (r). Then, some useful inequalities including
erated part U
the moment inequality are proved for weighted Rademacher
chaos. We mainly use these inequalities to bound Uw (r) and
ew (r). Section D provides technical proofs omitted from the
U
main track and the appendix. For the sake of completeness,
we present the Khinchine inequality and the metric entropy
inequality for Rademacher chaos in Section E.
A
Janson’s Decomposition
There are literatures that use Janson’s decomposition derived from (Janson 2004) to study the problem learning
from networked examples (Usunier, Amini, and Gallinari
2006, Biau and Bleakley 2006, Ralaivola, Szafranski, and
Stempfel 2009, Ralaivola and Amini 2015). They usually
model the networked data with the line graph of G.
Definition 6 (line graph). Let G = (V, E) be a data graph
we consider in the main track. We define the line graph of
G as a graph DG = (DV , DE ), in which DV = E and
{i, j} ∈ DE if and only if ei ∩ ej 6= ∅ in G.
This framework differs from our setting and detains less
information from the data graph, as argued in Section 2. One
can analyze this framework by the fractional coloring and the
Janson’s decomposition.
Definition 7 (fractional coloring). Let DG = (DV , DE ) be
a graph. C = {(Cj , qj )}j∈(1,...,J) , for some positive integer
J, with Cj ⊆ DV and qj ∈ [0, 1] is an fractional coloring of
DG , if
• ∀j, Cj is an independent set, i.e., there is no connection
between vertices in Cj .
P
• it is an exact cover of G: ∀v ∈ DV , j:v∈Cj qj = 1.
PJ
The weight W (C) of C is given by W (C) = j=1 qj and
the minimum weight χ∗ (DG ) = minC W (C) over the set of
all fractional colorings is the fractional chromatic number
of DG . By the Janson’s decomposition, one splits all examples into several independent sets according to a fractional
coloring of DG and then analyze each set using the standard
method for i.i.d. examples.
In particular, Theorem 3 (also see Usunier, Amini, and Gallinari 2006, Theorem 2) provides a variation of McDiarmid’s
theorem using the Janson’s decomposition. For simplicity, let
m
S = {zi }m
and
i=1 be the training examples drawn from Z
i
SCj be the examples included in Cj . Also, let kj be the index
of the i-th example of Cj in the training set S.
Theorem 3. Using the notations defined above, let C =
{(Cj , qj )}Jj=1 be a fractional coloring of DG . Let f : Z m →
R such that
• there exist J functions Z |Cj | → R which satisfy ∀S ∈ Z m ,
PJ
f (S) = j=1 qj fj (SCj ).
• there exist ci , . . . , cm ∈ R+ such that ∀j, ∀SCj , SCkj such
that SCj and SkCj differ only in the k-th example, |fj (SCj )−
fj (SCkj )|≤ ckji .
Then, we have
P [f (S) − E[f (S)] ≥ ] ≤ exp −
A.1
22
Pm
χ∗ (DG ) i=1 c2i
.
Weighted ERM
Using the above theorem, we show that the risk bounds for
weighted ERM derived from the Janson’s decomposition
cannot improve the results of (Usunier, Amini, and Gallinari
2006), as shown in the following theorem.
Theorem 4. Consider a minimizer rw of the weighted empirical risk Lw over a class R. For all δ ∈ (0, 1], with probability
at least 1 − δ, we have
p
kwk2
L(r) − Lw (r) ≤ R∗w (R, S) + χ∗ (DG ) log(1/δ)
kwk1
where
J
X
X
2
R∗w (R, S) = Eσ
qj sup
wkji σ i r(zkji ) .
N
r∈R
j=1
i∈Cj
is the weighted empirical fractional Rademacher complexity
of R with respect to DG .
In this setting, although the weights w are no limit to
the fractional matching, the risk bounds cannot improve the
results of p
(Usunier, Amini, and Gallinari 2006),
√ which is of
the order χ∗ (DG )/m, as kwk2 /kwk1 ≤ 1/ m.
A.2
“Low-noise” Condition
If we considering the complete graph with equal weighting
scheme, by the Janson’s decomposition, the empirical risk
Lm (L) can be represented as an average of sums of i.i.d. r.v.’s
bm/2c
X
1
1 X
`(r, (Xπ(i) , Xπ(i)+bm/2c ,
m!
bm/2c i=1
π∈Gn
(15)
Yπ(i),π(i)+bm/2c ))
where the sum is taken over all permutations of Gm , the
symmetric group of order m, and buc denotes the integer part
of any u ∈ R. From (Biau and Bleakley 2006),
√ the bounds for
excess risk of (15) are of the order O(1/ n). Moreover, by
the result in (Ralaivola and Amini 2015), tighter risk bounds
may be obtained under the following assumption which can
lead to “low-noise” condition (Tsybakov 2004, Massart and
Nédélec 2006).
Assumption 2. There exists C > 0 and θ ∈ [0, 1] such that
for all > 0,
1
P |η(X1 , X2 ) − |≤ ≤ Cθ/(1−θ) .
2
The risk bounds of the order O(log(n)/n) may be
achieved if θ = 1 (Massart and Nédélec 2006), which however is very restrictive. We can use the example in Papa,
Bellet, and Clémençon (2016) to show this.
Example 3. Let N be a positive integer. For each vertex
i ∈ {1, . . . , n}, we observe Xi = (Xi1 , Xi2 ) where Xi1 and
Xi2 are two distinct elements drawn from {1, . . . , N }. This
may, for instance, correspond to the two preferred items of a
user i among a list of N items. Consider now the case that
two nodes are likely to be connected if they share common
preference, e.g., Yi,j ∼ Ber(#(Xi ∩ Xj )/2). One can easily check that P [|η(X1 , X2 ) − 1/2|= 0] > 0, so tight risk
bounds cannot be obtained for minimizers of (15).
B
Universal Risk Bounds for Weighted ERM
on i.i.d. Examples
In section 3, the excess risk is split into two types of processes:
the weighted empirical process of i.i.d. examples and two
degenerated processes. In this section, we prove general risk
bounds for the weighted empirical process of i.i.d. examples.
The main idea is similar to (Massart and Nédélec 2006) that
tighter bounds for the excess risk can be obtained if the
variance of the excess risk is controlled by its expected value.
B.1
Bennett Concentration Inequality
First, we prove a concentration inequality for the supremum
of weighted empirical processes derived from (Bousquet
2002).
Theorem 5. Assume the (X1 , . . . , Xi ) are i.i.d. random variable according to P . Let F be a countable set of functions
from X to R and assume that all functions f in F are P measurable and square-integrable. If supf ∈F |f |≤ b, we denote
n
1 X
wi (f (Xi ) − E[f (Xi )]).
Z = sup
f ∈F kwk1 i=1
which {wi }ni=1 are bounded weights such that 0 ≤ wi ≤ 1
for i = 1, . . . , n and kwk1 > 0. Let σ be a positive real
number such that σ 2 ≥ supf ∈F Var[f (X)] almost surely,
then for all x ≥ 0, we have
v
u kwk2
u 2(
2
2 + 4bE[Z])x
σ
2bx
t kwk1
−x
P Z − E[Z] ≥
+
≤e .
kwk1
3kwk1
(16)
It is a variant of Theorem 2.3 of (Bousquet 2002) by just
applying Theorem 2.1 of (Bousquet 2002) with the weighted
empirical process. Then, by Theorem 5, we can generalize
the results of (Massart and Nédélec 2006) to weighted ERM.
We start by describing the probabilistic framework adapts to
our problem.
B.2
General Upper Bounds
Suppose that one observes independent variable ξ1 , . . . , ξn
taking their values in some measurable space Z with common
distribution P . For every i, the variable ξi = (Xi , Yi ) is a
copy of a pair of random variables (X, Y ) where X take its
values in measurable space X . Think of R as being the set of
all measurable functions from X to {0, 1}. Then we consider
some loss function
γ : R × Z → [0, 1].
(17)
Basically one can consider some set R, which is known to
contain the Bayes classifier r∗ that achieves the best (smallest) expected loss P [γ(r, ·)] when r varies in R. The relative
expected loss `¯ is defined by
¯ ∗ , r) = P [γ(r, ·) − γ(r∗ , ·)], ∀r ∈ R
`(r
(18)
Since the empirical process on i.i.d. examples split from the
excess risk is with non-negative weights on all examples, we
define the weighted loss as
n
1 X
γw (r) =
wi γ(r, ξi )
(19)
kwk1 i=1
where
Pn 0 ≤ wi ≤ 1 for all i = 1, . . . , n and kwk1 =
i=1 wi > 0. Weighted ERM approach aims to find a minimizer of the weighted empirical loss r̂ in the hypothesis set
R ⊂ R to approximate r∗ .
We introduce the weighted centered empirical process γw
defined by
γw (r) = γw (r) − P [γ(r, ·)].
(20)
¯ we shall
In addition to the relative expected loss function `,
need another way to measure the closeness between the elements of R. Let d be some pseudo-distance on R × R such
that
Var[γ(r, ·) − γ(r∗ , ·)] ≤ d2 (r, r∗ ), ∀r ∈ R.
(21)
A tighter risk bound for weighted ERM is derived from
Theorem 5 which combines two different moduli of uniform
continuity: the stochastic modulus of uniform continuity of
γw over R with respect to d and the modulus of uniform
¯
continuity of d with respect to `.
Next, we need to specify some mild regularity conditions
functions that we shall assume to be verified by the moduli
of continuity involved in the following result.
Definition 8. We denote by D the class of nondecreasing
and continuous functions ψ from R+ to R+ such that x →
ψ(x)/x is nonincreasing on (0, +∞) and ψ(1) ≥ 1.
In order to avoid measurability problems, we need to consider some separability condition on R. The following one
will be convenient.
Assumption 3. There exists some countable subset R0 of R
such that, for every r ∈ R, there exists some sequence {rk }
of elements of R0 such that, for every ξ ∈ Z, γ(rk , ξ) tends
to γ(r, ξ) as k tends to infinity.
The upper bound for the relative expected loss of any empirical risk minimizer on some given model R will depend
¯ ∗ , R) = inf r∈R `(r
¯ ∗ , r) and the flucon the bias term `(r
tuations of the empirical process γw on R. As a matter of
fact, we shall consider some slightly more general estimators. Namely, given some nonnegative number ρ, we consider
some ρ-empirical risk minimizer, that is, any estimator r
taking its values in R such that γw (r̂) ≤ ρ + inf r∈R γw (r).
Theorem 6 (risk bound for weighted ERM). Let γ be a loss
function such r∗ minimizes P [γ(r, ·)] when r varies in R. Let
φ and ψ belong to the class of functions D defined above and
let R be a subset of R satisfying the separability Assumption
3. Assume that, on the one hand,
p
q
kwk1
∗
¯ ∗ , r)), ∀r ∈ R,
(22)
d(r , r) ≤
ψ( `(r
kwk2
and that, on the other hand, one has, for every r ∈ R0 ,
p
sup
kwk1 E
[γw (r0 ) − γw (r) ≤ φ(σ)
kwk2
kwk1
r 0 ∈R0 , √
d(r 0 ,r)≤σ
Lemma 4. Let F be a countable collection of measurable
functions such that f ∈ [0, 1] for every f ∈ F, and let f0
be a measurable function such that f0 ∈ [0, 1]. Let σ be a
positive number such that P [|f − f0 |] ≤ σ 2 for every f ∈ F.
Then, setting
Z σ
ϕ(σ) =
(log N∞ (F, 2 ))1/2 d,
0
the following inequality is available:
p
kwk1 max(E[sup vw (f0 − f )], E[sup vw (f − f0 )])
f ∈F
≤ 12ϕ(σ).
p
provided that 4ϕ(σ) ≤ σ 2 kwk1 .
C
ew (r)
Inequalities for Uw (r) and U
In this section, we first show the classical symmetrization and randomization tricks for the degenerated weighted
ew (r).
U -statistics Uw (r) and the degenerated part U
Then we establish general exponential inequalities for
weighted Rademacher chaos. This result is generalized from
(Clémençon, Lugosi, and Vayatis 2008) based on moment inequalities obtained for empirical processes and Rademacher
chaos in (Boucheron et al. 2005). With this moment inequality, we prove the inequality for weighted Rademacher chaos,
which involves the L∞ covering number of the hypothesis
set.
Lemma 5 (decoupling and undecoupling). Let (Xi0 )ni=1 be
an independent copy of the sequence (Xi )ni=1 . Then, for all
q ≥ 1, we have:
X
E[ sup |
wi,j fi,j (Xi , Xj )|q ]
fi,j ∈F
(23)
for every positive σ such that φ(σ) ≤ kwk1 σ 2 , where
R0 is given by Assumption 3. Let ∗ be the unique positive
solution of the equation
p
kwk1 2∗ = φ(ψ(∗ )).
(24)
fi,j ∈F
Maximal Inequality for Weighted Empirical
Processes
Next, we present the maximal inequality involved in covering
number for weighted empirical processes. Let us fix some
notation. We consider i.i.d. random variables ξ1 , . . . , ξn with
values in some measurable space Z and common distribution
P . For any
-integrable function f on Z, we define Pw (f ) =
PP
n
1
w
i
i=1 f (ξi ) and vw (f ) = Pw (f ) − P (f ) where
kwk1
Pn
0 ≤ wi ≤ 1 for all i = 1, . . . , n and kwk1 = i=1 wi >
0. Given a collection F of P -integrable functions f , our
purpose is to control the expectation of supf ∈F vw (f ) or
supf ∈F −vw (f ).
X
wi,j fi,j (Xi , Xj0 )|q ]
(26)
(i,j)∈E
If the functions fi,j are symmetric in the sense that for all
Xi , Xj ,
fi,j (Xi , Xj ) = fj,i (Xj , Xi )
and (wi,j )(i,j)∈E is symmetric, then the inequality can be
reversed, that is,
X
E[ sup |
wi,j fi,j (Xi , Xj0 )|q ]
fi,j ∈F
B.3
(i,j)∈E
≤ 4q E[ sup |
p
Then there exists an absolute constant K such that, for every
y ≥ 1, the following inequality holds:
∗
¯ , r̂) > 2ρ + 2`(r
¯ ∗ , R) + Ky2 ≤ e−y .
P `(r
(25)
∗
f ∈F
(i,j)∈E
≤ 4q E[ sup |
fi,j ∈F
X
wi,j fi,j (Xi , Xj )|q ]
(27)
(i,j)∈E
Lemma 6 (randomization). Let (σi )ni=1 and (σi0 )ni=1 be two
independent sequences of i.i.d. Radermacher variables, independent from the (Xi , Xi0 )’s. If f is degenerated, we have
for all q ≥ 1,
X
E[sup |
wi,j f (Xi , Xj0 )|q ]
f ∈F
(i,j)∈E
≤ 4q E[sup |
f ∈F
X
(i,j)∈E
σi σj0 wi,j f (Xi , Xj0 )|q ]
Lemma 7. Let (Xi0 )ni=1 be an independent copy of the sequence (Xi )ni=1 . Consider random variables valued in {0, 1},
(Ỹi,j )(i,j)∈E , conditionally independent given the Xi0 ’s and
the Xi ’s and such that P [Ỹi,j = 1 | Xi , Xj0 ] = η(Xi , Xj0 ).
We have for all q ≥ 1,
X
E[sup|
wi,j h̃r (Xi , Xj , Yi,j )|q ]
r∈R
(i,j)∈E
X
≤ 4q E[sup|
r∈R
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
(28)
(i,j)∈E
and the inequality can be reversed,
X
E[sup|
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
r∈R
(i,j)∈E
X
≤ 4q E[sup|
r∈R
wi,j h̃r (Xi , Xj , Yi,j )|q ]
(29)
(i,j)∈E
Lemma 8. Let (σi )ni=1 and (σi0 )ni=1 be two independent sequences of i.i.d. Radermacher variables, independent from
the (Xi , Xi0 , Yi,j , Ỹi,j )’s. Then, we have for all q ≥ 1,
X
E[sup|
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
r∈R
1
t
t
min (
)2 ,
,
C
E[Uσ ]
E[M ] + F kwk2
s
t
t
(
)2/3 ,
kwkmax F
kwk∞ F
qP
2
where kwkmax = maxi
j:(i,j)∈E wi,j .
≤ exp
If the hypothesis set F is a subset of L ∞ (X 2 ) (upper
bounds on the uniform covering number with L∞ metric
can be calculated (Cucker and Zhou 2007)), we show E[Zσ ],
E[Uσ ] and E[M ] can be bounded by N∞ (F, ) since all these
Rademacher random variables satisfy the Khinchine inequality (see Section E). Following the metric entropy inequality
for Khinchine-type processes (see Section E), it is easy to get
the following Corollary.
Corollary 1. With the same setting of Theorem 7, if F ⊂
L ∞ (X 2 ), we have for any δ < 1/e,
P [Z ≤ κ] ≥ 1 − δ
where
Z
κ ≤C kwk2
X
≤ 4 E[sup|
r∈R
with a universal constant C.
D
D.1
Technical Proofs
Proofs Omitted in Section 4
Proof of Lemma 1. Since R ⊂ L ∞ (X 2 ) (Assumption 1),
by Corollary 1, the weighted degenerated U -process
supr∈R |Uw (r)| can be bounded by the L∞ covering number
of R, that is, for any δ ∈ (0, 1/e), we have
P [sup|Uw (r)|≤ κ] ≥ 1 − δ
r∈R
(i,j)∈E
where E[f (X, x)] = 0 for all x ∈ X , f ∈ F. Assume also
for all x, x0 ∈ X , f (x, x0 ) = f (x0 , x) (symmetric) and
supf ∈F kf k∞ = F . Let (σ i )ni=1 be i.i.d. Rademacher random variables and introduce the random variables
X
Zσ = sup |
wi,j σi σj f (Xi , Xj )|
f ∈F
where
κ≤
Z 1
C1
kwk2
log N∞ (R, )d
kwk1
0
Z 1p
log N∞ (R, )d,
+ max kwk2 log(1/δ)
0
(i,j)∈E
(log(1/δ))
Uσ = sup
X
sup
f ∈F α:kαk2 ≤1
M=
sup
f ∈F ,k=1,...,n
|
3/2
kwkmax , (log(1/δ))2 kwk∞
.
wi,j σi αj f (Xi , Xj )
(i,j)∈E
X
wi,k σi f (Xi , Xk )|
i:(i,k)∈E
Then exists a universal constant C such that for all n and
t > 0,
P [Z ≥ CE[Zσ ] + t]
2F
p
log N∞ (F, )d,
0
(log(1/δ))3/2 kwkmax , (log(1/δ))2 kwk∞ .
(i,j)∈E
f ∈F
log N∞ (F, )d
Z
+ max kwk2 log(1/δ)
σi σj0 wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
Lemma 5 and 6 are applications of Theorem 3.1.1 and
Theorem 3.5.3 of (De la Pena and Giné 2012) respectively,
providing decoupling and randomization inequalities for degenerated weighted U -statistics of order 2. Lemma 7 and
8 are the variants of Lemma 5 and 6 respectively, which is
ew (r).
suitable for the degenerated part U
Theorem 7 (moment inequality). Let X, X1 , . . . , Xn be i.i.d.
random variables and let F be a class of kernels. Consider
a weighted Rademacher chaos Zσ of order 2 on the graph
G = (V, E) indexed by F,
X
Z = sup |
wi,j f (Xi , Xj )|
2F
0
(i,j)∈E
q
−
with a universal constant C1 < ∞. Then the first inequality
for supr∈R |Uw (r)| follows the fact that R satisfies Assumption 1. Similarly, by Lemma 7 and Lemma 8, we can convert
ew (r)| to the moment of Rademacher
the moment of supr∈R |U
chaos
X
4q E[sup|
σi σj0 wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
r∈R
(i,j)∈E
which can be handled by the by-products of Theorem 7. More
specifically, using (46) and (47) combined with the arguments
in Corollary 1 and Assumption 1 will gives the second inew (r)|.
equality for supr∈R |U
Proof of Lemma 2. For any function r ∈ R, observe first
that
E[qr (X1 , X2 , Y1,2 ) | X1 ]
= E[E[qr (X1 , X2 , Y1,2 ) | X1 , X2 ] | X1 ]
= E[|1 − 2η(X1 , X2 )|1r(X1 ,X2 )6=r∗ (X1 ,X2 ) | X1 ]
Then observing that
|1 − 2η(X1 , X2 )|2 ≤ |1 − 2η(X1 , X2 )|
which implies that the modulus of continuity ψ can be taken
as
√
ψ() = 5.
(31)
Then from Lemma 4, we have
p
sup
kw̄k1 E
|γw̄ (r) − γw̄ (r0 )| ≤ φ(σ).
kw̄k2
r 0 ∈R, √
kw̄k1
d(r,r 0 )≤σ
p
provided that φ(σ)/3 ≤ kw̄k1 σ 2 . It remains to bound the
excess risk of rw by the tight bounds for weighted ERM on
i.i.d. examples by Theorem 6. For any δ ∈ (0, 1), we have
with probability at least 1 − δ,
L(rw ) − L∗ ≤ 2( inf L(r) − L∗ ) + 2ρ
r∈R
almost sure, and combining with Jensen inequality, we have
+ K log(1/δ)2∗ )
Var[E[qr (X1 , X2 , Y1,2 ) | X1 ]]
where C is a universal constant and ∗ is the unique positive
solution of the equation
p
kw̄k1 2∗ = φ(ψ(∗ )).
≤ E[(E[qr (X1 , X2 , Y1,2 ) | X1 ])2 ]
≤ E[|1 − 2η(X1 , X2 )|1r(X1 ,X2 )6=r∗ (X1 ,X2 ) ]
= Λ(r).
Proof of Lemma 3. First, we introduce some notations of
weighted
ERM of the i.i.d. case. We denote {w̄i =
P
j:(i,j)∈E wi,j : i = 1, . . . , n} the weights on vertices and
introduce the “loss function”
γ(r, X) = 2hr (X) + Λ(r)
n
1 X
w̄i γ(r, Xi ) = Tw (r).
kw̄k1 i=1
n
1 X
w̄i (γ(r, Xi ) − Λ(r))
kw̄k1 i=1
Proof Omitted in Section A
i∈Cj
(33)
Now, consider, for each j,
X wkji
(E[`(r, z)] − `(r, zkji )).
fj (SCj ) = sup
kwk1
r∈R
and the pseudo-distance
p
1/2
kw̄k1
0
d(r, r ) =
E[(γ(r, X) − γ(r0 , X))2 ]
kw̄k2
for every r, r0 ∈ R. Let φ be
Z σ
φ(σ) = 12
(log N∞ (R, 2 ))1/2 d.
which completes the proof.
Proof of Theorem 4. We write, for all r ∈ R,
"
#
m
1 X
L(r) − Lw (r) ≤ sup E[`(r, z)] −
wi `(r, zi )]
kwk1 i=1
r∈R
J
X
X wkji
≤
pj sup
(E[`(r, z)] − `(r, zkji )) .
kwk
1
r∈R
j=1
Define centered empirical process
γw̄ (r) =
When R satisfies Assumption 1, there exists a universal constant C 0 such that
1
2∗ ≤ C 0 K 1/(1+β) (
)1/(1+β)
(1 − β)2 kw̄k1
D.2
and the weighted empirical loss of vertices
γw̄ (r) =
(32)
i∈Cj
(30)
0
From the definition of “loss function” γ, we have the excess
risk of r is
¯ r∗ ) = Λ(r) − Λ(r∗ ) = Λ(r).
`(r,
According to Lemma 1, as Λ2 (r) ≤ Λ(r), we have for every
r ∈ R,
p
q
kw̄k1
∗
¯ r∗ )
d(r, r ) ≤
5`(r,
kw̄k2
Let f is defined by, for all training set S, f (S) =
PJ
j=1 pj fj (SCj ), then f satisfies the conditions of Theorem 3 with, for any i ∈ {1, . . . , m}, βi ≤ wi /kwk1 . Therefore, we can claim that, with probability at least 1 − δ,
"
"
##
m
1 X
wi `(r, zi )]
L(r) − Lw (r) ≤ E sup E[`(r, z)] −
kwk1 i=1
r∈R
+
p
χ∗ (DG ) log(1/δ)
kwk2
kwk1
Then, using the standard symmetrization technique (see
Usunier, Amini, and Gallinari 2006, Theorem 4), one can
bound the first item in the right hand side by R∗w (R, S) which
completes the proof.
D.3
Proofs Omitted in Section B
Proof of Theorem 5. We use an auxiliary random variable
n
X
e = kwk1 Z = sup 1
Z
wi (f (Xi ) − E[f (Xi )]).
2b
f ∈F 2b i=1
We denote by fk a function such that
X
1
fk =
sup
wi (f (Xi ) − E[f (Xi )]).
2b f ∈F
i6=k
We introduce following auxiliary random variables for k =
1, . . . , n,
X
1
Zk =
sup
wi (f (Xi − E[f (Xi )]))
2b f ∈F
i6=k
and
1
wi (f (Xk ) − E[f (Xk )]).
2b
Denoting by f0 the function achieving the maximum in Z,
we have
1
Ze − Zk ≤ wi (f0 (Xk ) − E[f0 (Xk )]) ≤ 1 a.s.,
2b
e − Zk − Zk0 ≥ 0
Z
Zk0 =
and
E[Zk0 ] = 0.
The first inequality is derived from wi ≤ 1 and
supf ∈F ,X∈X f (X) − E[f (X)] ≤ 2b. Also, we have
n
X
1 X
e
(n − 1)Z =
wi (f0 (Xi ) − E[f0 (Xi )])
2b
≤
k=1
n
X
i6=k
Zk ,
k=1
and
n
X
n
Ekn [Zk02 ] =
k=1
1 X
E[wi2 (fk (Xk ) − E[fk (Xk )])2 ]
2b
k=1
1
kwk22 sup Var[f (X)]
4b2
f ∈F
1
≤ 2 kwk22 σ 2 .
4b
where σ 2 ≥ supf ∈F Var[f (X)]. Notice that we use the fact
the Xi have identical distribution. Applying Theorem 1 of
2
e + kwk22 σ 2 will give
(Bousquet 2002) with v = 2E[Z]
4b
√
e ≥ 2vx + x ] ≤ e−x ,
P [Ze − E[Z]
3
and then
r
kwk1
kwk1
kwk22 2
P[
(Z − E[Z]) ≥ 2x(
E[Z] +
σ )
2b
b
4b2
x
+ ] ≤ e−x
3
which proves the inequality.
≤
Proof of Theorem 6. Since R satisfies Condition 3, we notice
that, by dominated convergence, for every r ∈ R, considering the sequence {rk } provided by Condition 3, one has
P [γ(·, rk )] that tends to P [γ(·, r)] as k tends to infinity. De¯ ∗ , R) = inf r∈R `(r
¯ ∗ , r). Hence,
note the bias term of loss `(r
¯ ∗ , R) = `(r
¯ ∗ , R0 ), which implies that there exists some
`(r
point π(r∗ ) (which may depend on ∗ ) such that π(r∗ ) ∈ R0
and
¯ ∗ , π(r∗ )) ≤ `(r
¯ ∗ , R) + 2 .
`(r
(34)
∗
We start from the identity
¯ ∗ , r̂) = `(r
¯ ∗ , π(r∗ )) + γw (r̂) − γw (π(r∗ ))
`(r
+ γw (π(r∗ )) − γw (r̂)
which, by definition of r̂, implies that
¯ ∗ , r̂) ≤ ρ + `(r,
¯ π(r∗ )) + γw (π(r∗ )) − γw (r̂).
`(r
√
Let x = K 0 y∗ , where K 0 is a constant to be chosen later
such that K 0 ≥ 1 and
γw (π(r∗ )) − γw (r)
.
Vx = sup ¯ ∗
∗
2
r∈R `(r , r) + + x
Then,
¯ ∗ , r̂) ≤ ρ + `(r
¯ ∗ , π(r∗ )) + Vx (`(r
¯ ∗ , r̂) + x2 + 2 )
`(r
∗
and therefore, on the event Vx < 1/2, one has
¯ ∗ , r̂) ≤ 2(ρ + `(r
¯ ∗ , π(r∗ ))) + 2 + x2 ,
`(r
∗
yielding
¯ ∗ , r̂) ≤ 2(ρ + `(r
¯ ∗ , π(r∗ ))) + 32 + x2 ]
P [`(r
∗
(35)
1
≤ P [Vx ≥ ].
2
¯
Since ` is bounded by 1, we may always assume x (and thus
∗ ) to be not larger than 1. Assuming that x ≤ 1, it remains
to control the variable Vx via Theorem 5. In order to use
Theorem 5, we first remark that, by Condition 3,
γw (π(r∗ )) − γw (r)
Vx = sup ¯ ∗
∗
2
r∈R0 `(r , r) + + x
which means that we indeed have to deal with a countably
indexed empirical process. Note that the triangle inequality
implies via (21), (34) and (22) that
1
(Var[γ(r, ·) − γ(π(r∗ ), ·)]) 2 ≤ d(r∗ , r) + d(r∗ , π(r∗ ))
p
q
kwk1
¯ ∗ , r) + 2 )
≤2
ψ( `(r
∗
kwk2
(36)
Since γ takes its values in [0, 1], introducing the functions
ψ1 = min(1, 2ψ) and, we derive from (36) that
√
kwk
( kwk2 1 ψ1 ())2
γw (π(r∗ )) − γw (t)
sup Var[ ¯ ∗
] ≤ sup
(2 + x2 )2
`(r , t) + ∗ + x2
r∈R
≥0
≤
kwk1
ψ1 () 2
sup(
) .
2
2
kwk2 x ≥0 max(, x)
Now the monotonicity assumptions on ψ imply that either
ψ() ≤ ψ(x) if x ≥ or ψ()/ ≤ ψ(x)/x if x ≤ . Hence,
one has in any case ψ()/(max(, x)) ≤ ψ(x)/x, which
finally yields
kwk1 ψ12 (x)
γw (π(r∗ )) − γw (r)
.
]
≤
sup Var[ ¯ ∗
kwk22 x4
`(r , r) + ∗ + x2
r∈R
On the other hand, since γ takes its values in [0, 1], we have
supk
r∈R
γ(r, ·) − γ(π(r∗ ), ·)
1
k∞ ≤ 2 .
∗
2
¯
x
`(r , r) + x
We can therefore apply Theorem 5 with v = ψ12 (x)x−4 and
b = x−2 , which gives that, on a set Ωy with probability larger
than 1 − exp(−y), the inequality
s
2y(ψ12 (x)x−2 + 4E[Vx ])
2y
+
.
Vx < E[Vx ] +
2
kwk1 x
3kwk1 x2
(37)
Now since ∗ is assumed to be not larger than 1, one has
ψ(∗ ) ≥ ∗ and therefore, for every σ ≥ ψ(∗ ), the following
inequality derives from the definition of ∗ by monotonicity:
φ(ψ(∗ ))
φ(ψ(∗ )) p
φ(σ)
≤
≤
= kwk1 .
2
2
σ
w (∗ )
2∗
Thus, (23) holds for every σ ≥ ψ(∗ ). In order to
control E[Vx ], we intend to use Lemma A.5 of (Massart and Nédélec 2006). For every r ∈ R0 , we intro¯ ∗ , π(r∗ )), `(r
¯ ∗ , r)). Then by (34),
duce a2 (r) = max(`(r
∗
2
∗
2
¯
¯
`(r , r) ≤ a (r) ≤ `(r , r) + ∗ . Hence, we have, on the
one hand, that
E[Vx ] ≤ E[ sup
r∈R0
γw (π(r∗ )) − γw (r)
].
a2 (r) + x2
and, on the other hand, that, for every ≥ ∗ ,
E[
sup
γw (π(r∗ )) − γw (r)]
r∈R0 ,a(r)≤
≤ E[
sup
¯ ∗ ,r)≤2
r∈R0 ,`(r
γw (π(r∗ )) − γw (r)].
¯ ∗ , r) ≤
Now by (34) if there exists some r ∈ R0 such that `(r
¯ ∗ , π(r∗ )) ≤ 2 + 2 ≤ 22 and therefore,
2 , then `(r
∗
by assumption (22)
of θ → ψ(θ)/θ,
√ and monotonicity
√
√ √kwk1
kwk1
∗
d(π(r ), r) ≤ 2 kwk2 ψ( 2) ≤ 2 2 kwk2 ψ(), then
√
√kwk2 d(π(r∗ ), r) ≤ 2 2ψ(). Thus, we derive from (23)
kwk1
that, for every ≥ ∗ ,
E[
sup
¯ ∗ ,r)≤2
r∈R0 ,`(r
√
γw (π(r∗ )) − γw (r)] ≤ φ(2 2ψ())
√
and since θ → φ(2 2ψ(θ))/θ is nonincreasing, we can use
Lemma A.5 of (Massart and Nédélec 2006) to get
p
√
E[Vx ] ≤ 4φ(2 2ψ(x))/( kwk1 x2 ),
and by monotonicity of θ → φ(θ)/θ,
p
√
E[Vx ] ≤ 8 2φ(ψ(x))/( kwk1 x2 ).
Thus, using the monotonicity of θ → φ(ψ(θ))/θ, and the
definition of ∗ , we derive that
√
√
√
√
8 2φ(ψ(∗ ))
8 2
8 2∗
8 2
E[Vx ] ≤ p
=
≤√ 0 ≤√ ,
x
Ky
K0
kwk1 x∗
(38)
provided that x ≥ ∗ , which holds since K 0 ≥ 1. Now, the
monotonicity of θ → ψ1 (θ)/θ implies that x−2 ψ12 (x) ≤
2
−2
∗ ψ1 (∗ ), but since φ(θ)/θ ≥ φ(1) ≥ 1 for every θ ∈
[0, 1], we derive from (24) and the monotonicity of φ and
θ → φ(θ)/θ that
ψ12 (∗ )
φ2 (ψ1 (∗ ))
φ2 (2ψ(∗ ))
φ2 (ψ(∗ ))
≤
≤
≤
4
2∗
2∗
2∗
2∗
and, therefore, x−2 ψ12 (x) ≤ 4kwk1 2∗ . Plugging this inequality together with (38) into (37) implies that, on the set Ωy ,
s
√
√
2y
8 2
2y(4kwk1 2∗ + 32/ K 0 )
+
+
.
Vx < √
2
0
kwk1 x
3kwk1 x2
K
It remains to replace x2 by its value K 0 y2∗ to derive that, on
the set Ωy , the following inequality holds:
s
√
√
8 2
8(1 + 4(kwk1 2∗ K 0 )−1 )
2
Vx < √ +
+
.
0
0
K
3kwk1 K 0 2∗
K
Taking into account that φ(ψ(θ)) ≥ φ(min(1, ψ(θ))) ≥ θ
for every θ ∈ [0, 1], we deduce from the definition of ∗
that kwk1 2∗ ≥ 1 and, therefore, the preceding inequality
becomes, on Ωy ,
s
√
√
2
8(1 + 4/ K 0 )
8 2
+
Vx < √
+
.
0
0
0
K
3K
K
Hence, choosing K 0 as a large enough numerical constant
guarantee that Vx < 1/2 on Ωy and, therefore, (35) yields
¯ ∗ , r̂) ≤ 2(ρ + `(r
¯ ∗ , π(r∗ ))) + 32 + x2 ]
P [`(r
∗
≤ P [Ωcy ]
≤ e−y .
We get the required probability bound (6) by setting K =
K 0 + 3.
Proof of Lemma 4. We first perform the control of
E[supf ∈F vw (f − f0 )]. For simplicity, we denote
H∞ (F, ) = log N∞ (F, ). For any integer j, we set
σj = σ2−j and Hj = H∞ (F, σj2 ). By definition of
Hj = H∞ (F, σj2 ), for any integer j ≥ 1, we can define a
mapping Πj from F to some finite collection of functions
such that
log #{Πj F} ≤ Hj
(39)
and
Πj f ≤ f with P (f − Πj f ) ≤ σj2 , ∀f ∈ F.
(40)
For j = 0, we choose Π0 to be identically equal to f0 . For
this choice of Π0 , we still have
P (|f − Π0 f |) = P [|f − f0 |] ≤ σ02 = σ
(41)
for every f ∈ F. Furthermore, since we may always assume
that the extremities of the balls used to cover F take their
values in [0, 1], we also have for every integer j that
0 ≤ Πj f ≤ 1.
H1 ≤ σ1−2 ϕ2 (σ),
p
and under the condition 4ϕ(σ) ≤ σ 2 kwk1 , one has H1 ≤
σ12 kwk1 . Thus, since j → Hj σj−2 increases to infinity, the
set {j ≥ 0 : Hj ≤ σj2 kwk1 } is a nonvoid interval of the
form
{j ≥ 0 : Hj ≤ σj2 kwk1 } = [0, J],
with J ≥ 1. For every f ∈ F, starting from the decomposition
−vw (f ) =
vw (Πj f ) − vw (Πj+1 f ) + σJ2
j=0
and, therefore,
E[ sup [−vw (f )]]
f ∈F
J−1
X
j=0
and, on the other hand,
p
p
2
4 kwk1 σJ+1
≤ 4σJ+1 Hj+1 .
Hence, plugging these inequalities in (43) yields
J+1
X p
p
kwk1 E[sup −vw (f )] ≤ 6
σj Hj ,
E[sup vw (Πj f ) − vw (Πj+1 f )] + σJ2 .
(42)
D.4
Now, it follows from (40) and (41) that, for every integer j
and every f ∈ F, one has
2
2
P [|Πj f − Πj+1 f |] ≤ σj2 + σj+1
= 5σj+1
Zi = Xi if σi = 1, and Xi0 otherwise,
Zi0 = Xi0 if σi = 1, and Xi0 otherwise.
Conditionally upon the Xi and Xi0 , the random vector
(Zi , Zi0 ) takes the values (Xi , Xi0 ) or (Xi0 , Xi ), each with
probability 1/2. In particular, we have:
L(X1 , . . . , Xn , X10 , . . . , Xn0 ) =
and, therefore, since |Πj f − Πj+1 f |≤ 1,
L(Z1 , . . . , Zn , Z10 , . . . , Zn0 )
2
P [|Πj f − Πj+1 f |2 ] ≤ 5σj+1
.
Moreover, (39) ensures that the number of functions of the
form Πj f − Πj+1 f when varies in F is not larger than
exp(Hj + Hj+1 ) ≤ exp(2Hj+1 ), Hence, we derive from
the maximal inequality for random vectors (see Massart and
Nédélec 2006, Lemma A.1) and the by-product of the proof
of Bernstein’s inequality for the weighted sum of networked
random variables (see Wang, Guo, and Ramon 2017, Lemma
16) that
p
kwk1 E[sup [vw (Πj f ) − vw (Πj+1 f )]]
≤ 2[σj+1
p
1
5Hj+1 + p
Hj+1 ]
3 kwk1
Proofs Omitted in Section C
Proof of Lemma 7. This Lemma is derived from Lemma 5,
thus we can follow the similar arguments that can be found
in (De la Pena and Giné 2012).
For any random variable X, we denote by L(X) its distribution. We denote by Σ (respectively Σ0 ) the sigma-filed
generated by {X1 , . . . , Xn } (respectively {X10 , . . . , Xn0 }).
0
Let (Yi,j
)(i,j)∈E be Bernoulli random variables such that
0
P [Yi,j = 1 | Σ, Σ0 ] = η(Xi0 , Xj0 ). Let (σi )ni=1 be independent Rademacher variables and define:
f ∈F
f ∈F
j=1
and the result follows. The control of E[supf ∈F vw (f − f0 )]
can be performed analogously.
we derive, since ΠJ (f ) ≤ f and P (f − Πj (f )) ≤ σJ2 , that
≤
[σj
f ∈F
+vw (ΠJ f ) − vw (f ),
J−1
X
J
X
p
p
1
2
Hj ] + 4 kwk1 σJ+1
5Hj + p
.
3
kwk
1
j=1
(43)
It follows from the definition of J that, on the one hand, for
every j ≤ J,
1
1p
p
Hj ≤
Hj
3
3 kwk1
vw (Πj f ) − vw (Πj+1 f )
j=0
−vw (f ) =
f ∈F
≤2
Noticing that since u → H∞ (F, u2 ) is nonincreasing,
J−1
X
because wi ≤ 1, ∀i ∈ 1, . . . , n, and (42) becomes
p
kwk1 E[sup −vw (f )]
(44)
and
L(X1 , . . . , Xn ) = L(Z1 , . . . , Zn ).
(45)
Let {Ỹi,j }(i,j)∈E be Bernoulli random variables such that
P [Ỹi,j = 1 | Σ, Σ0 ] = η(Xi , Xj0 ) and define for (i, j) ∈ E,
Yi,j if σi = 1 and σj = −1
0
Yi,j if σi = −1 and σj = 1
Ŷi,j =
Ỹ
if σi = 1 and σj = 1
i,j
Ỹi,j if σi = −1 and σj = −1
Notice that for all f ,
Eσ [h̃r (Zi , Zj0 , Ỹi,j )]
1
= (h̃r (Xi , Xj , Yi,j ) + h̃r (Xi0 , Xj , Ỹi,j )
4
0
+ h̃r (Xi , Xj0 , Ỹi,j ) + h̃r (Xi0 , Xj0 , Yi,j
))
≤
where Eσ denotes the expectation taken with respect to
{σi }ni=1 . Moreover, using
0
E[h̃r (Xi0 , Xj0 , Yi,j
) | Σ] = 0
E[h̃r (Xi , Xj0 , Ỹi,j ) | Σ] = 0
(i,j)∈E
| Σ] = 0
(i,j)∈E
X
1
wi,j Eσ [h̃r (Zi , Zj , Ỹi,j )]|q ]
= E[sup|4
2 r∈R
h̃r (Xi , Xj , Yi,j ) = 4E[h̃r (Zi , Zj0 , Ŷi,j ) | Σ]
(i,j)∈E
For all q ≥ 1, we therefore have
X
E[ sup|
wi,j h̃r (Xi , Xj , Yi,j )|q ]
X
1
+ E[sup|2
wi,j h̃r (Xi , Xj , Yi,j )|q ]
2 r∈R
(i,j)∈E
(i,j)∈E
= E[sup|
r∈R
X
(Jansen’s Inequality and the convexity of sup|·|q )
4wi,j E[h̃r (Zi , Zj0 , Ỹi,j ) | Σ]|q ]
r∈R
X
1
wi,j h̃r (Zi , Zj , Ỹi,j )|q ]
≤ E[sup|4
2 r∈R
(i,j)∈E
≤ 4q E[sup|
r∈R
X
wi,j h̃r (Zi , Zj0 , Ỹi,j )|q ]
(i,j)∈E
(i,j)∈E
p
derived from the facts that the supreme and |x| (p ≥ 1) are
convex functions and the Jansen inequality. According to
(44) and the fact that the distribution of Ŷi,j only depends on
the realization Zi , Zj0 , i.e. P [Ŷi,j | Zi , Zj0 ] = η(Zi , Zj0 ), we
obtain
X
4q E[ sup|
wi,j ĥr (Zi , Zj0 , Ỹi,j )|q ]
r∈R
(i,j)∈E
q
= 4 E[sup|
r∈R
X
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
(i,j)∈E
that
r∈R
(i,j)∈E
(According to (45))
X
1
= E[sup|4
wi,j h̃r (Xi , Xj , Yi,j )|q ]
2 r∈R
(i,j)∈E
X
1
wi,j h̃r (Xi , Xj , Yi,j )|q ]
+ E[sup|2
2 r∈R
(i,j)∈E
X
= 4q E[sup|
wi,j h̃r (Xi , Xj , Yi,j )|q ]
(i,j)∈E
which (29) follows.
Proof of Lemma 8. Re-using the notations used in the proof
of Lemma 5, we further introduce (Xi00 )ni=1 , a copy of
(Xi0 )ni=1 , independent from Σ, Σ0 , and denote by Σ00 its sigma00
field. Let (Ỹi,j
)(i,j)∈E Bernoulli random variables such that
00
P [Ỹi,j = 1 | Σ, Σ0 , Σ00 ] = η(Xi , Xj00 ). We now use classic
randomization techniques and introduce our ”ghost” sample:
X
00 q
E[sup|
wi,j h̃r (Xi , Xj0 , Ỹi,j
)| ]
(i,j)∈E
= E[sup|
X
1
+ E[sup|2
wi,j h̃r (Xi , Xj , Yi,j )|q ]
2 r∈R
r∈R
which concludes the proof of (28).
By the symmetry of h̃r in the sense
h̃r (Xi , Xj , Yi,j ) = h̃r (Xj , Xi , Yj,i ), we have
X
E[ sup|
wi,j h̃r (Xi , Xj0 )|q ]
r∈R
+ h̃r (Xi0 , Xj , Ỹi,j ) + h̃r (Xi , Xj , Yi,j )
X
1
0
+ E[sup|2
wi,j h̃r (Xi0 , Xj0 , Yi,j
)|q ]
4 r∈R
we easily get
r∈R
(i,j)∈E
0
+ h̃r (Xi0 , Xj0 , Yi,j
)|q ]
X
1
wi,j h̃r (Xi , Xj , Yi,j )|q ]
+ E[sup|2
4 r∈R
and (degenerated)
E[h̃r (Xi0 , Xj , Ỹi,j )
X
1
wi,j (h̃r (Xi , Xj0 , Ỹi,j )
E[sup|
2 r∈R
1 X
wi,j (h̃r (Xi , Xj0 , Ỹi,j )
2
(i,j)∈E
+ h̃r (Xi0 , Xj , Ỹi,j ))|q ]
1 X
= E[sup|
wi,j (h̃r (Xi , Xj0 , Ỹi,j )
r∈R 2
r∈R
(i,j)∈E
(i,j)∈E
(h̃r is degenerated)
X
00
= E[sup|
wi,j (h̃r (Xi , Xj0 , Ỹi,j
)
+ h̃r (Xi0 , Xj , Ỹi,j ) + h̃r (Xi , Xj , Yi,j )
0
+ h̃r (Xi0 , Xj0 , Yi,j
)
1 X
−
wi,j h̃r (Xi , Xj , Yi,j )
2
r∈R
(i,j)∈E
00
− EΣ00 [h̃r (Xi , Xj00 , Ỹi,j
)])|q ]
(i,j)∈E
(Jansen’s Inequality)
X
00
≤ E[sup|
wi,j (h̃r (Xi , Xj0 , Ỹi,j
)
1 X
0
−
wi,j h̃r (Xi0 , Xj0 , Yi,j
)|q ]
2
r∈R
(i,j)∈E
q
(Triangle’s Inequality and the convexity of sup|·| )
r∈R
(i,j)∈E
00
− h̃r (Xi , Xj00 , Ỹi,j
))|q ]
= E[sup|
n
X
X
r∈R j=1
i:(i,j)∈E
−
√
≤ Eσ [Zσ ] + 3 qEσ Uσ + 4qB
00
wi,j (h̃r (Xi , Xj0 , Ỹi,j
)
where B is defined below
00
h̃r (Xi , Xj00 , Ỹi,j
))|q ]
B = sup
f
Let (σi )ni=1 be independent Rademacher variables, independent of Σ, Σ0 and Σ00 , then we have:
n
X
E[ sup|
X
r∈R j=1
i:(i,j)∈E
wi,j (h̃r (Xi , Xj0 , Ỹi,j )
i:(i,j)∈E
00
− h̃r (Xi , Xj00 , Ỹi,j
))|q | Σ]
(Triangle’s Inequality and the convexity of sup|·|q )
r∈R
n
X
≤ 2q E[sup|
σj
r∈R j=1
X
X
r∈R j=1
00
wi,j h̃r (Xi , Xj00 , Ỹi,j
))|q | Σ]
i:(i,j)∈E
n
X
≤ 2q E[sup|
σj
r∈R j=1
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q | Σ]
i:(i,j)∈E
n
X
q
+ 2 E[sup|
σj
X
α,α0 :kαk2 ,kα0 k2 ≤1
Σ]
i:(i,j)∈E
and get
It remains to derive suitable upper bounds for the expectation
of the three terms on the right hand side.
First term: E[Eσ [Zσ ]q ]. Using the symmetrization trick,
we have
E[Eσ [Zσ ]q ] ≤ 4q E[Eσ [Zσ0 ]q ]
P
0
which Zσ = supf | (i,j)∈E σi σj0 f (Xi , Xj )|. Note that Eσ
now denotes expectation taken with respect to both the σ and
the σ 0 . For simplicity, we denote by A = Eσ [Zσ0 ]. In order
to apply Corollary 3 of (Boucheron et al. 2005), define, for
k = 1, . . . , n, the random variables
X
Ak = Eσ [sup|
wi,j σi σj0 f (Xi , Xj )|].
(i,j)∈E,i,j6=k
It is easy to see that Ak ≤ A.
On the other hand, defining
X
Rk = sup|
wi,k σi f (Xi , Xk )|,
f
E[sup|
r∈R
X
i:(i,k)∈E
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
M = max Rk
(i,j)∈E
k
≤ 2q E[sup|
n
X
σj
r∈R j=1
X
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q ]
Proof of Theorem 7. By the decoupling, undecoupling and
randomization techniques (see Lemma 5, Lemma 6), the
symmetry and the degeneration of f and the symmetry of
(wi,j )(i,j)∈E , we have
X
E[sup|
wi,j f (Xi , Xj )|q ]
(i,j)∈E
≤ 16q E[sup|
f
≤ 64q E[sup|
f
and denoting by f ∗ the function achieving the maximum in
the definition of Z, we clearly have
i:(i,j)∈E
Then repeating the same argument but for the (Xi )ni=1 will
give the similar inequality. The desired inequality will follows
putting these two inequalities together.
f
(i,j)∈E
The second inequality above follows by Theorem 14 of
(Boucheron et al. 2005).
Using the inequality (a + b + c)q ≤ 3(q−1) (aq + bq + cq )
valid for q ≥ 2, a, b, c > 0, we have
f
wi,j h̃r (Xi , Xj0 , Ỹi,j )|q |
wi,j αi αj0 f (Xi , Xj )|.
Eσ [Zσq ] ≤ 3q−1 (Eσ [Zσ ]q + 3q q q/2 Eσ [Uσ ]q + 4q q q B q ).
00
− h̃r (Xi , Xj00 , Ỹi,j
))|q | Σ]
n
X
X
= E[sup|
σj
wi,j (h̃r (Xi , Xj0 , Ỹi,j )
r∈R j=1
X
|
sup
X
and
n
X
(A − Ak ) ≤ 2A.
k=1
Therefore,
n
X
(A − Ak )2 ≤ 4AEσ M.
k=1
Then by Corollary 3 of (Boucheron et al. 2005), we obtain
wi,j σi σj0 f (Xi , Xj0 )|q ]
E[Eσ [Zσ0 ]q ] ≤ 2q−1 (2q (E[Zσ0 ])q + 5q q q E[Eσ [M ]q ]) (46)
wi,j σi σj f (Xi , Xj )|q ]
To bound E[Eσ [M ]q ], observe that Eσ [M ] is a conditional Rademacher average, for which Theorem 13
of (Boucheron et al. 2005) could be applied. Since
maxk supf,i wi,k f (Xi , Xk ) ≤ kwk∞ F , we have
(i,j)∈E
X
A − Ak ≤ 2Eσ [M ]
(i,j)∈E
It means we can convert the moment of the original U -process
to the moment of Rademacher chaos which can be handled
by moment inequalities of (Boucheron et al. 2005).
In particular, for any q ≥ 2,
(Eσ [Zσq ])1/q ≤ Eσ [Zσ ] + (Eσ [(Zσ − Eσ [Zσ ])q+ ])1/q
E[Eσ [M ]q ] ≤ 2q−1 (2q E[M ]q + 5q q q kwkq∞ F q ).
E[Zσ0 ]
(47)
E[Eσ,σ0 [Zσ0 ]]
By undecoupling, we have
=
E[4Eσ [Zσ ]] = 4E[Zσ ]. Collecting all terms, we have
E[Eσ [Zσ ]q ] ≤ 64q E[Zσ ]q + 160q q q E[M ]q
≤
+ 400q kwkq∞ F q q 2q .
Second term: E[Eσ [Uσ ]q ].
By the Cauchy-Schwartz inequality we can observe that
X
sup sup
wi,j αj f (Xi , Xj ) ≤ kwkmax F.
f,i α:kαk2 ≤1
the L∞ metric. For all f, s, we have L2 (zσ0 (f ), zσ0 (g)) ≤
L∞ (f, g). The fact that if ∀f, s ∈ F, L(f, s) ≤ L0 (f, s) then
N (F, L, ) ≤ N (F, L0 , ) combined with (49) will give
E[Zσ ] = 2E[Eσ [sup|zσ (f )|]]
f
j:(i,j)∈E
q
Then similar to the bound of E[Eσ [M ] ], we have
Third term:q
E[B q ]. By the Cauchy-Schwartz inequality,
P
2
we have B ≤
(i,j)∈E wi,j F = kwk2 F so
E[B q ] ≤ kwkq2 F q .
Now it remains to simply put the pieces together to obtain
X
E[sup|
wi,j f (Xi , Xj )|q ]
≤ 2Kkwk2
+
+
+
kwkqmax F q q 3q/2
for an appropriate constant C. In order to derive the exponential inequality, we use Markov inequality P [X ≥ t] ≤
t−q E[Z q ] and choose
t
t
t
t
q = C min (
)2 ,
,
)2/3 ,
,(
E[Uσ ]
E[M ] F kwk2 kwkmax F
s
t
kwk∞ F
for an appropriate constant C.
Proof of Corollary 1. From Theorem 3, it is easy to know
p
κ =C(E[Zσ ] + max(E[Uσ ] log(1/δ), E[M ] log(1/δ),
where K is a universal constant. Similarly,
we
can
also
bound
E[M ]
by
Kkwkmax
R 2F p
∗
log N∞ (F, )d. For Uσ , let α be the (random)
0
vector that maximizes Uσ and define
{uσ (f ) =
n
X
E
Theorem 8 (Khinchine inequality for Rademacher chaos,
De la Pena and Giné (2012, Theorem 3.2.1)). Let F be a
normed vector space and let {σ i }∞
i=1 be a Rademacher sequence. Denote by
X =x+
n
X
X
xi σi +
xi1 ,i2 σ i1 σ i2 + . . .
i1 <i2 ≤n
i=1
X
+
xi1 ...id σ i1 . . . σ id
i1 <···<id ≤n
the Rademacher chaos of order d. Let 1 < p ≤ q < ∞ and
let
p − 1 1/2
γ=(
) .
q−1
Then, for all d ≥ 1,
(i,j)∈E<
(E[|x +
n
X
γxi σi +
X
+
γ 2 xi1 ,i2 σ i1 σ i2 + . . .
γ d xi1 ...id σ i1 . . . σ id |q ])1/q
i1 <···<id ≤n
n
X
X
i=1
i1 <i2 ≤n
≤ (E[|x +
+
X
i1 <i2 ≤n
i=1
0
Recall that supf kf k∞ = F , we have D = 2F . Since this
metric function d is intractable, we need to convert it to
(i,j)∈E
Metric Entropy Inequality
(log(1/δ)) kwk∞ )).
f,g
wi,j αj∗ f (Xi , Xj ), f ∈ F}.
The following theorems are more or less classical and well
known. We present them here for the sake of completeness.
2
where E< = {(i, j) : {i, j} ∈ E, i < j}. Let zσ0 =
zσ /kwk2 , following Theorem 8, we can easily have that
zσ0 satisfies (50) with degree 2. Thus from Theorem 9, we
have
Z D
0
0
Eσ [sup|zσ (f )−zσ (s)|] ≤ K
log N (F, L2 , )d. (49)
X
σi
Clearly, uσ satisfies the Khintchine inequality with degree 1. Also, we need to convert its metric distance
and L2 (uσ (f ), uσ (g)) ≤ L∞ (f, g). Thus, E[Uσ ] ≤
R 2F p
log N∞ (F, )d.
Kkwk2 0
Plugging all these part into (48) will complete the corollary.
log(1/δ)kwk2 , (log(1/δ))3/2 kwkmax ,
(48)
An important character of these Rademacher processes in
(48) is that they all satisfy the Khinchine inequality (50). For
simplicity, we denote the weighted Rademacher processes of
Zσ by
X
{zσ (f ) =
wi,j σi σj f (Xi , Xj ), f ∈ F}.
log N∞ (F, )d
0
i=1
kwkq∞ F q q 2q
F q kwkq2 q q )
2F
Z
(i,j)∈E
≤ C(E[Zσ ]q + q q/2 E[Uσ ]q + q q E[M ]q
log N (F, L2 , )d
0
E[Eσ [Uσ ]q ] ≤ 2q−1 (2q E[Uσ ]q + 5q q q kwkqmax F q ).
f
2F
Z
≤ 2Kkwk2
xi σi +
X
i1 <···<id ≤n
xi1 ,i2 σ i1 σ i2 + . . .
xi1 ...id σ i1 . . . σ id |p ])1/p
Theorem 9 (metric entropy inequality, Arcones and Gine
(1993, Proposition 2.6)). If a process {Yf : f ∈ F} satisfies
(E[|Yf − Yg |p ])1/p ≤ (
p − 1 m/2
)
(E[|Yf − Yg |q ])1/q , (50)
q−1
for 1 < q < p < ∞ and some m ≥ 1, and if
d(f, g) = (E[|Yf − Yg |2 ])1/2 ,
there is a constant K < ∞ such that
Z D
E[sup|Yf − Yg |] ≤ K
(log N (F, d, ))m/2 d.
f,g
0
where D is the d-diameter of F.
(51)
(52)
| 8cs.DS
|
Can the early human visual system compete with Deep Neural Networks?
Samuel Dodge and Lina Karam
Arizona State University
arXiv:1710.04744v1 [cs.CV] 12 Oct 2017
{sfdodge,karam}@asu.edu
Abstract
we wish to test the early human vision system by limiting
the stimuli display time to 100ms. Within 100ms, there is
no time for eye movements [1], thus the human visual system is limited to more global “gist” representations [14].
Can the human visual system still recognize distorted images only by the gist? By contrast, the experiments in [7]
allow the subject to view the image for unlimited duration.
This allows the subject to analyze more local information
that can help classify the stimuli.
We study and compare the human visual system and
state-of-the-art deep neural networks on classification of
distorted images. Different from previous works, we limit
the display time to 100ms to test only the early mechanisms
of the human visual system, without allowing time for any
eye movements or other higher level processes. Our findings
show that the human visual system still outperforms modern
deep neural networks under blurry and noisy images. These
findings motivate future research into developing more robust deep networks.
1.1. Related Works
Human performance on distorted stimuli has been extensively studied. Torralba et al. [22] showed that humans
are able to recognize very low resolution images. Similarly,
studies on face images show that the human visual system
can perform well in the presence of blur [2] and noise [5].
There are also several works that study deep neural network performance on distorted data. Dodge and Karam [6]
studied several different types of common image distortions
and found that noise and blur have the largest effect on
the performance of DNNs on the ImageNet dataset. Rodner et al. show similar findings on smaller fine grained
datasets [17].
Comparing human and machine vision performance has
also been studied in the past. Borji and Itti [3] compare
14 different computer vision models on several datasets and
compare with human performance. However the study does
not consider modern neural networks, which greatly outperform older vision models.
Fleuret et al. compare human and machine vision performance on synthetic visual reasoning tasks [9]. These tasks
are designed such that there must be some sort of reasoning,
instead of pure pattern recognition. For many of these tasks
humans outperform artificial vision systems. A followup
study showed that, for some of these synthetic problems,
state-of-the-art neural networks surprisingly achieve accuracy equivalent with random chance [20].
Parikh [15] studies human and artificial vision system
performance on jumbled images. These jumbled images
are formed by randomly permuting blocks of an image. On
these jumbled images human classification performance is
1. Introduction
It has been shown that state-of-the-art deep neural networks (DNNs) often achieve better performance compared
with human subjects on large scale classification tasks [18].
Given this, we might conclude that the DNNs do a better
job of representing and organizing visual data. But does the
success of DNNs carry over to other more difficult tasks?
Recently it has been discovered that deep neural networks perform poorly in the presence of distortions such as
blur and noise [6, 17]. Blur removes high frequency information and likewise, noise injects high frequency information. Current deep networks seem to experience difficulty
reasoning in the presence of high levels of such distortions.
Do humans have a similar trouble with distorted images?
By studying the human visual system (HVS), we can perhaps gain insight into how to build DNN models that are
more robust to distortions. If human performance on distorted images is better than DNNs, then this exposes a vulnerability in the DNN’s representation of visual data. But
if human performance is also poor, then recognition under
distortions may be inherently difficult.
Previous studies have tested human capability for recognition under noise and blur [22, 5], finding that humans have
some robustness with respect to these distortions. In this
work, we wish to compare human ability and DNN ability
on a common task. Recent works perform similar experiments with distorted images [7, 10]. However, in this work
1
degraded to near the performance of a bag-of-words based
classifier. While the jumbled images may give some insight
into the human visual system, jumbled images are not typical visual stimuli.
Kheradpisheh et al. [13] compare human and DNN performance on images of objects with arbitrary backgrounds
and rotations. The highest performing DNNs match human performance. This is consistent with other studies
(e.g., [18]) that show that DNN classification performance
is at-par with or superior to human performance.
To further evaluate human vs. DNNs for classification
we design our experiments to test highly distorted images.
We extend on the work in [7], where Amazon Mechanical
Turk testing is used to compare human and machine performance on a subset of the ImageNet dataset with added
distortions. In this work there are two primary differences.
First we use human subjects in a controlled lab setting. This
is in contrast to Amazon Mechanical Turk studies where
there is no mechanism to control viewing distance, screen
brightness, etc. Secondly, we use a fixed display time instead of free viewing, which lets us analyze the accuracy
of the HVS without allowing for higher level processes like
eye movements.
A concurrent independent study [10] compares human
and deep learning performance on distorted images. However, the study does not fine tune networks on distorted images. This gives an unfair advantage to the human subjects,
which may have previously seen distorted images. Additionally, we chose a 100ms display time which is less than
the 200ms display time in [10]. This ensures that we are
testing early “gist”-based processes of the visual system.
2. Methods
In this section we first introduce our dataset and image
distortions. Then, we describe our experimental setup to
test classification under distortion by both human subjects
and deep neural networks.
2.1. Dataset
In this paper we wish to test the early vision system.
Thus we chose a simple coarse-grained dataset, instead
of a more difficult fine-grained dataset (as in [7]). We
select 8 classes from the Caltech101 dataset [8]: Butterfly, Crocodile, Dolphin, Elephant, Flamingo, Leopard, and
Llama. This dataset was deliberately chosen to be “easy”
such that under no distortion, neither the human subjects
nor the deep neural networks would have any difficulty in
classification.
We consider two types of distortion: random additive
Gaussian noise and Gaussian blur. For Gaussian noise we
consider noise with a standard deviation ranging from 0
(clean images) to 200 (highly corrupted images). Similarly,
for blur we consider Gaussian blur with a kernel of standard
deviation that ranges from 0 to 10. Figure 1 shows example stimuli of the 8 classes for 3 difference levels of each
distortion.
Our dataset consists of 200 clean training images (25 per
class), and 40 clean validation images (5 per class). We use
80 unique images for testing blur at 5 possible levels for a
total of 400 images. Similarly there are 80 unique images
for testing noise at 5 levels for a total of 400 images. We additionally randomly distort the training images as explained
in Section 2.3. The same training, testing, and validation
split was used for each human trial, as well as for the deep
neural networks.
2.2. Human Experiments
The goal of our human experiments is to test the ability
of the human visual system to identify images using 100ms
“gist”-level information. Other studies have shown that humans can accurately recognize images at display times below 100ms [12, 16]. We choose 100ms because this is the
same display time in the experiments in [4] which are used
to correlate human and DNN neural responses.
We design our experiment to mimic the training, validation, and testing stages used for training and evaluating
DNNs.
As in [7] we first allow the subjects to freely view the
training images. The subject must view images in all of the
categories before allowed to continue with the experiment.
This allows the subjects to familiarize themselves with the
image categories. This training stage is analogous to the
training stage used for deep neural networks.
Before we test the subject on distorted images, we first
test clean images in a validation stage. This stage is to ensure that the subjects are able to correctly classify clean images. As in the rest of our experiments a central fixation
cross is first shown for 500 milliseconds, followed by the
image stimuli for 100 milliseconds, and finally a second fixation cross for 500 milliseconds (Figure 2). Next, a choice
screen allows the subject to choose the most appropriate of
the 8 categories and continue to the next image. This is a
forced-choice response, so the subject must choose a class,
even if the subject is not perfectly clear as to the correct
class.
In the next stage of the experiment the subject is asked to
classify distorted images. The experiment proceeds exactly
as in the validation stage, except that the images are now
distorted. For each image we wish to test the maximum
distortion level at which the subject can make a correct decision. Testing every possible distortion level for every image can accomplish this but has practical limitations. First,
with many distortion levels, the total duration of the experiment becomes prohibitive. Secondly, multiple exposure to
the same image can induce a memory effect that could help
subjects identify highly distorted images that they would
Original
2
Blur σ
6
10
40
Noise σ
120
200
Figure 1. Example image stimuli. Our image categories consist of “coarse” classes that are easily recognizable under no distortion. In
this figure we show three levels of distortions among the 5 levels used in the experiments.
Fixation
Stimuli
500ms
Ti
m
e
Fixation
100ms
Choice
500ms
Figure 2. Timing of validation and testing stages. The subject
is first shown a central fixation marker for 500ms followed by the
image stimuli for 100ms, another fixation marker for 500ms and
finally a choice screen for the subject to input the class estimate.
otherwise not be able to identify.
Instead we follow the procedure in [7], which makes
the assumption that subjects can correctly classify an image with a certain distortion level if the subject has already
correctly classified an image with a higher distortion level.
Thus we begin with each stimuli image at the maximum
distortion level. If the subject incorrectly identifies the image, then the distortion level is reduced and the image is
randomly shown later. If the subject correctly classifies an
image at a particular distortion level, then all of the lower
distortion levels are assumed to be correctly classified and
the lower distortion levels are not tested.
We recruit 8 subjects to participate in our experiment.
All of the subjects were tested for vision and color blindness
before beginning the experiment.
2.3. Deep Neural Networks
Deep networks are the current state-of-the-art approaches for many image recognition tasks. These net-
Figure 3. Human experiment training screen. During the training stage, the subject must view example images from each category
before continuing with the experiment.
works consist of layers of convolutional filtering, pooling,
and nonlinear operations. The parameters of the layers can
be learned by fitting to a training set using gradient descent
based optimization algorithms.
We consider three classification models: VGG16 [19],
Googlenet [21], and ResNet50 [11]. The VGG16 network is
a popular architecture which consists of 16 layers and uses
convolution layers with small 3x3 filters. The Googlenet
architecture introduces blocks that perform parallel convolution with filters of varying size. ResNets employ skip
connections, which aids in training and can yield a more
accurate network.
Each network has been pre-trained on the ImageNet
dataset [18]. We perform fine-tuning to adapt the model to
our new dataset. Specifically, we replace the last fully connected layer with a new fully connected layer with 8 units
(corresponding to the 8 classes in the dataset). The learning
rate of the new layer is 10x the learning rate of the pretrained layers. The network is fine-tuned using stochastic
gradient descent with momentum. We stop learning when
the performance on the validation set plateaus.
We test two scenarios for training. In the first scenario,
the networks are fine-tuned on clean images as previously
described. In the second scenario, the networks are finetuned on a mixture of clean images and distorted images.
We refer to these networks as “distortion-tuned”. During
each mini-batch, half of the training images remain undistorted and the other half are distorted with a random level
uniformly chosen from the minimum distortion level to the
maximum distortion level. This procedure is identical to
that in [7], and ensures that the distortion-tuned networks
can perform well at all levels of distortions, as well as for
clean images.
We test using the same procedure as in the human experiments. The network is first tested on the highest distortion
level, and if the prediction is correct the network is assumed
to correctly classify the same image at all lower levels of
distortions. If the prediction is incorrect, then the network
is tested on a lower level of distortion. This process continues until the network predicts the correct class, or the level
of distortion becomes 0 (i.e., a clean image).
3. Results
On the validation portion of the human test, subjects
scored an average accuracy of 99.3%. This shows that the
subjects are diligently performing the experiment, and that
the subjects are able to classify the images correctly, even
at the display time of 100ms.
Figure 4 shows the comparison between human accuracy
and machine accuracy for our experiments. For clean images human accuracy and DNN accuracy are nearly identical. When distortion is added, human accuracy greatly
exceeds that of neural networks trained on clean images.
When networks are fine-tuned on distorted data, the network performance approaches human performance for most
levels of distortion. However at the largest levels of distortion, the human performance still exceeds that of the
distortion-tuned deep networks.
The network architecture can affect how resilient the network is to distortion, even when the network is distortiontuned. For our dataset, the disortion-tuned ResNet50 model
performs the best on noisy images, but is not the best performing on blurry images. For blurry images the distortiontuned VGG16 model acheives closest to human perfor-
100
80
80
Accuracy (%)
Accuracy (%)
100
60
40
20
0
0
50
100
σ
150
200
(a) Noise
60
40
20
0
0
2
4
σ
6
8
10
Human
VGG16
VGG16 Distortion-tuned
GoogleNet
GoogleNet Distortion-tuned
ResNet50
ResNet50 Distortion-tuned
(b) Blur
ResNet50
Distortion-tuned
ResNet50
Figure 4. Accuracy comparison between DNNs and human subjects for distorted stimuli. Human subjects greatly outperform DNNs
trained on clean data. When networks are trained on data from the respective distortion, the performance gap decreases.
σ=0
σ = 40
σ = 80
σ = 120
σ = 160
σ = 200
σ=0
σ = 40
σ = 80
σ = 120
σ = 160
σ = 200
1.0 Human predicts more often
0.5
0.0
−0.5
−1.0 DNN predicts more often
1.0 Human predicts more often
0.5
0.0
−0.5
−1.0 DNN predicts more often
Figure 5. Confusion matrix difference between humans and ResNet50 for noise. Columns of a confusion matrix correspond to the
actual class, and rows correspond to the predicted class. Rows are normalized to sum to one. This figure shows the difference between
the human confusion matrix and that of the ResNet50 model. Red regions denote pairs that humans predict more often, and blue regions
denote pairs that networks predict more often.
σ=2
σ=4
σ=6
σ=8
σ = 10
σ=0
σ=2
σ=4
σ=6
σ=8
σ = 10
VGG16
Distortion-tuned
VGG16
σ=0
1.0 Human predicts more often
0.5
0.0
−0.5
DNN predicts more often
−1.0
1.0 Human predicts more often
0.5
0.0
−0.5
−1.0 DNN predicts more often
Figure 6. Confusion matrix difference between humans and ResNet50 for blur. Columns of a confusion matrix correspond to the actual
class, and rows correspond to the predicted class. Rows are normalized to sum to one. This figure shows the difference between the human
confusion matrix and that of the VGG16 model. Red regions denote pairs that humans predict more often, and blue regions denote pairs
that networks predict more often.
mance.
An analysis of confusion patterns can yield insight into
the behavior of the DNNs. We compute the difference between the confusion matrices of human subjects and the best
performing DNN model. This tells us which categories are
confused by the DNN relative to the human and vice-versa.
Figure 5 shows this for noise using ResNet50, and Figure
6 shows this for blur using VGG16. For a DNN that has
not seen distorted data during training, at high distortions
the network tends to predict a single class regardless of the
input. The network fine-tuned on distortions shows a confusion matrix that is more similar to humans.
We also show examples of images that are “difficult” or
“easy” for humans or the distortion-tuned ResNet50 model
(Figure 7). We consider an image to be correctly classified by the human subjects if 90% or greater of the subjects selected the correct class. Some images are easy for
both the neural network and human subjects. Others are
consistently missclassified by both. Finally there are some
images correctly classified by humans, but missclassified by
DNNs. Likewise some images classified correctly by DNNs
are classified incorrectly by the human subjects.
Human Incorrect
ResNet50 Incorrect
ResNet50 Correct
Human Correct
Figure 7. Examples of correct and incorrect classifications of
human subjects and distortion-tuned ResNet50 models.
4. Discussion and Conclusion
We performed a set of human experiments for classifying distorted images under a 100ms time constraint. We
find that human performance still exceeds the performance
of state-of-the-art neural networks on distorted images, even
if the display time is very short (100ms). Fine-tuning with
distorted images reduces this gap, but nevertheless a gap
still exists. The 100ms display time does not allow for eye
movements or other higher level visual processes. This tells
us that the human visual system is more efficient at processing “gist” level information than DNNs.
Note that our dataset is noticeably easier than the dataset
in [7]. This was to ensure that the human subjects could
still perform recognition at 100ms display times. A sideeffect of this choice of dataset is that it is much easier for the
deep neural networks to perform recognition. For example,
for such simple classes it may be sufficient to use simple
color information to recognize the images. A more difficult
dataset may expose a larger difference between human and
DNN performance.
These experiments provide a look into the bias and shortcomings of DNNs by comparing them with the human visual system. Future work could analyze other scenarios
where humans still outperform DNNs. These studies can
help motivate future work into more robust deep neural networks.
Acknowledgment
The authors would like to thank NVIDIA Corporation
for the donation of a TITAN X GPU used in these experi-
ments.
References
[1] T. J. Andrews and D. M. Coppola. Idiosyncratic characteristics of saccadic eye movements when viewing different visual environments. Vision research, 39(17):2947–2953,
1999. 1
[2] T. Bachmann. Identification of spatially quantised tachistoscopic images of faces: How many pixels does it take to
carry identity? European Journal of Cognitive Psychology,
3(1):87–103, 1991. 1
[3] A. Borji and L. Itti. Human vs. computer in scene and object
recognition. In IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pages 113–120, 2014. 1
[4] C. F. Cadieu, H. Hong, D. L. Yamins, N. Pinto, D. Ardila,
E. A. Solomon, N. J. Majaj, and J. J. DiCarlo. Deep neural networks rival the representation of primate IT cortex for
core visual object recognition. PLoS computational biology,
10(12):e1003963, 2014. 2
[5] Y. Chen, R. McBain, and D. Norton. Specific vulnerability
of face perception to noise: A similar effect in schizophrenia patients and healthy individuals. Psychiatry research,
225(3):619–624, 2015. 1
[6] S. Dodge and L. Karam. Understanding how image quality
affects deep neural networks. International Conference on
Quality of Multimedia Experience (QoMEX), 2016. 1
[7] S. Dodge and L. Karam. A study and comparison of human
and deep learning recognition performance under visual distortions. International Conference on Computer Communications and Networks, 2017. 1, 2, 3, 4, 6
[8] L. Fei-Fei, R. Fergus, and P. Perona. Learning generative
visual models from few training examples: An incremental
bayesian approach tested on 101 object categories. Computer
Vision and Image Understanding, 2007. 2
[9] F. Fleuret, T. Li, C. Dubout, E. K. Wampler, S. Yantis, and
D. Geman. Comparing machines and humans on a visual
categorization test. Proceedings of the National Academy of
Sciences, 108(43):17621–17625, 2011. 1
[10] R. Geirhos, D. H. J. Janssen, H. H. Schtt, J. R. M. Bethge,
and F. A. Wichmann. Comparing deep neural networks
against humans: object recognition when the signal gets
weaker. arXiv preprint arXiv:1706.06969, 2017. 1, 2
[11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016.
4
[12] C. Keysers, D.-K. Xiao, P. Földiák, and D. I. Perrett. The
speed of sight. Journal of cognitive neuroscience, 13(1):90–
101, 2001. 2
[13] S. R. Kheradpisheh, M. Ghodrati, M. Ganjtabesh, and
T. Masquelier. Deep networks can resemble human feedforward vision in invariant object recognition. Scientific reports, 6:32672, 2016. 2
[14] A. Oliva. Gist of the scene. Neurobiology of attention, 2005.
1
[15] D. Parikh. Recognizing jumbled images: The role of local
and global information in image classification. In IEEE International Conference on Computer Vision (ICCV), pages
519–526, 2011. 1
[16] M. C. Potter, B. Wyble, C. E. Hagmann, and E. S. McCourt.
Detecting meaning in rsvp at 13 ms per picture. Attention,
Perception, & Psychophysics, 76(2):270–279, 2014. 2
[17] E. Rodner, M. Simon, R. B. Fisher, and J. Denzler. Finegrained recognition in the noisy wild: Sensitivity analysis of
convolutional neural networks approaches. British Machine
Vision Conference (BMVC), 2016. 1
[18] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
Recognition Challenge. International Journal of Computer
Vision (IJCV), 115(3):211–252, 2015. 1, 2, 4
[19] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. International
Conference on Learning Representations, 2014. 4
[20] S. Stabinger, A. Rodrı́guez-Sánchez, and J. Piater. 25 years
of cnns: Can we compare to human abstraction capabilities?
In International Conference on Artificial Neural Networks,
pages 380–387. Springer, 2016. 1
[21] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9,
2015. 4
[22] A. Torralba, R. Fergus, and W. T. Freeman. 80 million tiny
images: A large data set for nonparametric object and scene
recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(11):1958–1970, 2008. 1
[23] F. Wang, M. Jiang, C. Qian, S. Yang, C. Li, H. Zhang,
X. Wang, and X. Tang. Residual attention network for image classification. IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2017.
[24] A. R. Zamir, T.-L. Wu, L. Sun, W. Shen, J. Malik, and
S. Savarese. Feedback networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
| 1cs.CV
|
Thread-Modular Static Analysis for Relaxed Memory Models
Markus Kusano
Chao Wang
Virginia Tech
Blacksburg, VA, USA
University of Southern California
Los Angeles, CA, USA
arXiv:1709.10077v1 [cs.PL] 28 Sep 2017
ABSTRACT
We propose a memory-model-aware static program analysis method
for accurately analyzing the behavior of concurrent software running on processors with weak consistency models such as x86-TSO,
SPARC-PSO, and SPARC-RMO. At the center of our method is a
unified framework for deciding the feasibility of inter-thread interferences to avoid propagating spurious data flows during static
analysis and thus boost the performance of the static analyzer. We
formulate the checking of interference feasibility as a set of Datalog
rules which are both efficiently solvable and general enough to
capture a range of hardware-level memory models. Compared to
existing techniques, our method can significantly reduce the number of bogus alarms as well as unsound proofs. We implemented
the method and evaluated it on a large set of multithreaded C programs. Our experiments show the method significantly outperforms
state-of-the-art techniques in terms of accuracy with only moderate
runtime overhead.
CCS CONCEPTS
• Software and its engineering → Automated static analysis;
Formal software verification;
KEYWORDS
Concurrency, Abstract interpretation, Thread-modular reasoning,
Datalog, Relaxed memory model, TSO, PSO, RMO
1
INTRODUCTION
Concurrent software written for modern computer architectures,
though ubiquitous, remains challenging for static program analysis.
Although abstract interpretation [13] is a powerful static analysis
technique and prior thread-modular methods [19, 40, 47–49] mitigated interleaving explosion, none was specifically designed for
software running on weakly consistent memory. This is a serious
deficiency since weakly consistent memory may exhibit behaviors
not permitted by uniprocessors. For example, slow memory accesses
may be delayed, increasing performance, but also introducing additional inter-thread non-determinism. Thus, multithreaded software
running on such processors may exhibit erroneous behaviors not
manifesting on sequentially consistent (SC) memory.
Consider x86-TSO (total store order) as an example. Under TSO,
each processor has a store buffer caching memory write operations so they do not block the execution of subsequent instructions [4]. Conceptually, each processor has a queue of pending
writes to be flushed to memory at a later time. The flush occurs
non-deterministically at any time during the program’s execution.
This delay between the time a write instruction executes and the
time it takes effect may cause the write to appear reordered with
subsequent instructions within the same thread. Figure 1 shows an
example where the assertion holds under SC but not TSO. Since x
and y are initialized to 0 and they are not defined as atomic variables, the write operations (x=1 and y=1) may be stored in buffers,
one for each thread, and thus delayed after the read operations.
void thread1 () {
x = 1;
a = y;
}
void thread2 () {
y = 1;
b = x;
}
assert( !(a == 0 && b == 0) );
Figure 1: The assertion holds under SC, but not under x86TSO, SPARC-PSO, and SPARC-RMO memory models.
SPARC-PSO (partial store order) permits even more non-SC behaviors: it uses a separate store buffer for each memory address.
That is, x=1 and y=1 within the same thread may be cached in different store buffers and flushed to memory independently. This permits
the reordering of a write to x with a subsequent read from y, but
also with a subsequent write (e.g., to variable z) in the same thread.
The situation is similar under SPARC-RMO (relaxed-memory order).
We detail how such relaxation leads to errors in Section 2.
Broadly speaking, existing thread-modular abstract interpreters
fall into two categories, neither modeling weak-memory related
behaviors. The first are SC-specific [17, 18, 40]: they are designed
to be flow-sensitive in terms of modeling thread interactions but
consider only behaviors compatible with the SC memory. The second [47–49] are oblivious to memory models (MM-oblivious): they
permit all orderings of memory-writes across threads. Therefore,
MM-oblivious methods may report spurious errors (bogus alarms)
whereas SC-specific methods, although more accurate for SC memory, may miss real errors on weaker memory (bogus proofs). This
flaw is not easy to fix using conventional approaches [49]. For example, maintaining relational invariants at all program points makes
the analysis prohibitively expensive. In Section 2, we use examples
to illustrate issues related to these techniques.
We propose the first thread-modular abstract interpreter for
analyzing concurrent programs under weakly consistent memory.
Our method models thread interactions with flow-sensitivity, and is
memory-model specific: it models memory operations assuming a
processor-level memory model, as shown in Figure 2. In this figure,
the boxes with bold text highlight our main contributions.
Our method builds on a unified framework for modeling the
memory consistency semantics. Specifically, the feasibility of thread
interactions is formulated as a constraint problem via Datalog: it
is efficiently solvable in polynomial time, and adaptable to various
hardware-level memory models. Additionally, our method handles
thread interactions in a flow-sensitive fashion while being threadmodular. Analyzing one thread at a time, as opposed to the entire
program, increases efficiency, especially for large programs. However, unlike prior MM-oblivious methods we do not join all the
effects of remote stores before propagating them to a thread, thus
preserving accuracy. Overall, our method differs from the state-ofthe-art, which either are non-thread-modular [14, 17, 39, 45] or not
specifically targeting weak memory [40, 47–49].
Our method also differs significantly from techniques designed
for bug hunting as opposed to obtaining correctness proofs. For
example, in concurrency testing, stateless dynamic model checking [22, 25] was extended from SC to weaker memory models [1,
void thread1 () {
x = 5;
fence ;
y = 10;
}
Thread-Modular
Analysis
Sequential Abstract
Interpretation
Feasibility of
Thread Interference
Parameterized by
Memory Models
(SC, TSO, PSO,
RMO, etc.)
Figure 2: FruitTree: Our memory-model-aware, threadmodular, static program analysis procedure.
2, 16, 32, 51, 52, 67]. In bounded model checking, Alglave et al. [8]
modeled weak memory through code transformation or direct symbolic encoding [7, 9]. However, these methods cannot be used to
verify properties: if they do not find bugs, it does not mean the
program is correct. In contrast, our method, like other abstract
interpreters, is geared toward obtaining correctness proofs.
We implemented our new method in a tool named FruitTree,
using Clang/LLVM [5] as the C front-end, Apron [37] for abstract
domains, and the µZ [30] Datalog engine in Z3 [15]. We evaluated FruitTree on 209 litmus tests, and 52 larger multithreaded
programs totaling of 61,981 lines of C code. Reachability properties were expressed as embedded assertions. Our results show that
FruitTree is significantly more accurate than state-of-the-art techniques with moderate runtime overhead.
Specifically, we compared FruitTree against the MM-oblivious
analyzer of Miné [49], the SC-specific thread-modular analyzer
Watts [40], and a non thread-modular analyzer named Duet [17,
18]. On the litmus tests, FruitTree is more accurate than the other
three methods. On the larger benchmarks, including Linux device
drivers, FruitTree proved 4,577 properties, compared to 1,752
proved by Miné’s method.
To summarize, we make the following contributions:
• We propose a memory-model aware static analysis method
based on thread-modular abstract interpretation.
• We introduce a declarative analysis framework for deducing the feasibility of thread-interferences on weak memory.
• We implement and evaluate our method on a set of benchmarks to demonstrate its high accuracy and moderate runtime overhead.
The remainder of this paper is organized as follows. First, we
motivate our technique via examples in Section 2. Then, we provide background on memory models and abstract interpretation in
Section 3. We present our new declarative analysis for checking the
feasibility of thread inferences in Section 4, followed by the main
algorithm for thread-modular abstract interpretation in Section 5.
We present our experimental results in Section 6, review related
work in Section 7, and conclude in Section 8.
2
MOTIVATION
Consider the program in Figure 3. The assertion holds under SC,
TSO, PSO, and RMO. But, removing the fence causes it to fail under
PSO and RMO. In this section, we show why MM-oblivious methods
may generate bogus errors, why SC-specific ones may generate
bogus proofs, and how our new method fixes both issues.
2.1
Behaviors under SC, TSO, PSO, and RMO
First, note that the assertion in Figure 3 holds under SC since each
thread executes its instructions in program order, i.e., x = 5 takes
void thread2 () {
if ( y == 10) {
assert ( x == 5);
}
}
Figure 3: The assertion always holds, but if the fence is removed, the assertion may fail under PSO and RMO.
effect before y = 10. So, thread two observing y to be 10 implies x
must have been set to 5.
Next, we explain why the assertion holds under TSO [4]. TSO
permits the delay of a store after a subsequent load to a disjoint
memory address (as in Figure 1). This program-order relaxation is
a performance optimization, e.g., buffering slow stores to speed
up subsequent loads. However, since all stores in a thread go into
the same buffer, TSO does not allow the reordering of two stores
(thread 1, Figure 3). Thus, even without the fence, x = 5 always
takes effect before y = 10, meaning the assertion holds.
Next, we show why removing the fence causes the assertion to
fail under PSO and RMO. Both permit store–store reordering by
allowing each processor to have a separate store buffer for each
memory address. Thus x and y are in separate buffers. Since buffers
are flushed to memory independently, with the fence removed,
y = 10 may take effect before x = 5, as if the two instructions
were reordered in this thread. Thus, the second thread may read 10
from y before 5 is written to x in global memory, thus causing the
assertion to fail.
The fence is important because it forces all stores issued before
the fence to be visible to all loads issued after the fence, i.e., x = 5
takes effect before y = 10, even under PSO and RMO. Thus, the
assertion holds again.
2.2
Ineffectiveness of Existing Methods
MM-oblivious methods [47–49] report bogus alarms because they
were not designed for weak memory, and they ignore the causality of inter-thread data flows. Thus, they tend to drastically overapproximate the interferences between threads.
For example, an MM-oblivious static analysis may work as follows. First, it analyzes each thread as if it were a sequential program.
Then, it joins the effects of all stores on global memory—known as
the thread interferences. Next, it individually analyzes each thread
again, this time in the presence of the thread interferences computed
from the previous iteration: when a thread performs a memory read,
the value may come from any one of these thread interferences.
This iterative process repeats until a fixed point is reached.
Next, we demonstrate how the MM-oblivious analyzer works
on Figure 3. Consider the thread interferences to be a map from
variables to abstract values in the interval domain [13]. Thread
1 generates interferences x 7→ [5, 5] and y 7→ [10, 10]. Within
thread 2, the load of y may read from local memory, [0, 0], or the
interference [10, 10]. Thus, y = [0, 0] ⊔ [10, 10] = [0, 10], where ⊔
is the join operator in the interval domain. Similarly, the load of x
may read from local memory, [0, 0], or the interference [5, 5], i.e.,
x = [0, 5]. Thus, the assertion is incorrectly reported as violated.
While our previous example used the non-relational interval
domain the bogus alarm remains when using a relational abstract
domain: the propagation of interferences in MM-oblivious methods
is inherently non-relational. Inferences map variables to a single values, causing all relations to be forgotten. Conventional approaches
Method
MM-oblivious (e.g. [47–49])
SC
(bogus)
alarm
SC-specific (e.g. [17, 18, 40])
proof
Our method
proof
Program in Figure 1
with fences
without fences
TSO
PSO/RMO
SC
TSO
PSO/RMO
(bogus)
(bogus)
(bogus)
alarm
alarm
alarm
alarm
alarm
(bogus)
(bogus)
(bogus)
(bogus)
proof
proof
proof
proof
proof
proof
proof
proof
alarm
alarm
SC
(bogus)
alarm
proof
proof
Program in Figure 3
with fences
without fences
TSO
PSO/RMO
SC
TSO
PSO/RMO
(bogus)
(bogus)
(bogus)
(bogus)
alarm
alarm
alarm
alarm
alarm
(bogus)
(bogus)
(bogus)
(bogus)
proof
proof
proof
proof
proof
proof
proof
proof
proof
alarm
Figure 4: Comparing the effectiveness of various methods in handling the example programs in Figure 1 and Figure 3.
cannot easily fix this since maintaining relational invariants at all
global program points is prohibitively expensive.
In contrast, prior SC-specific methods [17, 18, 40] do not report
bogus alarms: they assume x = 5 takes effect before y = 10. This
leads to more accurate analysis results for SC, but is unsound under
weak memory, e.g., they miss the assertion failure in Figure 3 under
PSO or RMO when the fence is removed.
Figure 4 summarizes the ineffectiveness of prior techniques on
the programs in Figures 1 and 3 with and without fences. Note
that in Figure 1 the fence instruction may be added between the
write and read instructions of both threads. The table in Figure 4
shows how prior MM-oblivious methods report bogus alarms, prior
SC-specific methods report bogus proofs, while our new method
eliminates both.
2.3
How Our Method Handles Memory Models
Some prior techniques lead to bogus alarms because they overapproximate thread interferences, i.e., they allow a load to read
from any remote store regardless of whether such a data flow, or
combination of flows, is feasible, while others lead to missed bugs
because they under-approximate thread interferences, i.e., they do
not allow any non-SC data flow. Consider Figure 3: the load of
x may read 0 or 5, and the load of y may read 0 or 10, but the
combination of x reading 0 and y reading 10 is infeasible. Realizing
this, our method checks the feasibility of interference combinations
under weak-memory semantics before propagating them.
Toward this end, we propose two new techniques. The first is
the flow-sensitive propagation of thread interferences. Instead of
eagerly joining all interfering stores, we handle each combination
separately. The second is a declarative modeling of the memory
consistency semantics general enough to capture SC, TSO, PSO,
and RMO [4, 58, 65]. Together, these techniques prune infeasible
combinations of thread interferences such as x and y reading 0 and
10, respectively, in Figure 3.
Our new method analyzes thread 2 in Figure 3 by considering
four different interference combinations, ρ 1 –ρ 4 , separately.
•
•
•
•
ρ1
ρ2
ρ3
ρ4
= y 7→ [0, 0] and x 7→ [0, 0],
= y 7→ [10, 10] and x 7→ [5, 5],
= y 7→ [0, 0] and x 7→ [5, 5],
= y 7→ [10, 10] and x 7→ [0, 0].
We gain accuracy in two ways. First, we remove spurious values
caused by an eager join (e.g., we no longer have y = [0, 10]). Second, we query a lightweight constraint system to quickly deduce
infeasibility of an interference combination on demand. ρ 1 , ρ 2 , and
ρ 3 are all feasible but they do not cause assertion failures.
Our check for infeasibility of an interference combination is
implemented using Datalog (Horn clauses within finite domains),
solvable in polynomial time. We will provides details of this constraint system in Section 4. For now, consider ρ 4 in Figure 3: it is
infeasible (unless we assume the program runs under PSO or RMO
with the fence removed). We deduce infeasibility as follows:
• y = 10 has executed (it is being read from),
• thus x = 5 has executed (due to the program-order requirement on SC and TSO, and the fence on PSO and RMO),
• so the load of x must not read from its initial value [0, 0].
This deduction leads to a formal proof that ρ 4 can not exist in any
concrete execution. Since the combinations ρ 1−3 do not violate the
assertion, and ρ 4 is proved to be infeasible, the property is verified.
3
PRELIMINARIES
In this section, we review weak memory models at the processor
level (as opposed to the programming language level) and static
program analysis based on abstract interpretation.
3.1
Concurrent Programs
We are concerned with a program consisting of a finite set of threads.
Each thread assesses a set {a, b, c, . . .} of local variables. All threads
access a set {x, y, z, . . .} of global variables via load and store instructions. A thread creates a child thread with ThreadCreate, and
waits it to terminate with ThreadJoin.
We represent a program using a set G = {G 1 , . . . , G k } of flow
graphs. Each flow graph G ∈ G, where G = ⟨N , n 0 , δ ⟩, is a thread:
N ⊆ N is the set of program locations of the thread, n 0 ∈ N is the
entry point, and δ is the transition relation. That is, (n ′, n) ∈ δ iff
there exists an edge from n ′ to n.
Each program location n ∈ N is associated with an atomic instruction that may be a load, store, or fence. Non-atomic statements
such as y = x+1, where both x and y are global variables, can be
transformed to a sequence of atomic instructions, e.g., the load a =
x followed by the store y = a+1, where a is a local variable in both
cases. When accessing variables on the global memory, threads
may use a special fence instruction to impose a strict program order
between memory operations issued before and after the fence.
3.2
Memory Consistency Models
The simplest memory model is sequential consistency (SC) [43]. SC
corresponds to a system running on a single coherent memory timeshared by operations executed from different threads. There are two
important characteristics of SC: the program-order requirement and
the write-atomicity requirement. The program-order requirement
says that the processor must ensure that instructions within a
thread take effect in the order they appear in the program. The
write-atomicity requirement says that the processor must maintain
the illusion of a single sequential order among operations from all
Memory Model
SC [43]
TSO [55, 65]
PSO [65]
RMO [58, 65]
R(v 1 )→R(v 1 )
R(v 1 )→W(v 1 )
no
no
no
no
no
no
no
no
Which Program-Order Relaxation Is Allowed?
W(v 1 )→R(v 1 ) W(v 1 )→W(v 1 ) R(v 1 )→R(v 2 ) R(v 1 )→W(v 2 )
no
no*
no*
no*
no
no
no
no
no
no
no
yes
no
no
no
yes
W(v 1 )→R(v 2 )
W(v 1 )→W(v 2 )
no
yes
yes
yes
no
no
yes
yes
Write-Atomicity
read own
write early
no
yes
yes
yes
Figure 5: Allowed relaxations of various processor-level memory models (cf. [4]). v 1 and v 2 are distinct variables, and * indicates
rule needs to be relaxed to allow read-own-write-early behaviors (see Section 4.6 for explanation).
threads. That is, the effect of any store operation must take effect
and become visible either to all threads or to none of the threads.
SC is an ideal memory model: In real CPUs, the hardware-level
memory models are often weaker than SC, and can be characterized
by their corresponding relaxations of the program-order and writeatomicity requirements as shown in Figure 5. Here, R(v 1 ) → W (v 2 )
is a read of v 1 followed by a write of v 2 in the same thread.
Specifically, TSO allows x=1;a=y to be reordered as a=y;x=1,
according to W(v 1 )→R(v 2 ) in Column 8, where v 1 is x and v 2
is y. PSO further allows x=1;y=2 reordered to y=2;x=1, according to W(v 1 )→W(v 2 ) in Column 9. As shown in Section 2, these
program-order relaxations, conceptually, are the effect of store
buffering, which delay the stores past subsequent stores/loads
within a thread. Neither TSO nor PSO permits the delay of a load.
Weaker still is RMO, which permits the relaxations of R(v 1 )→R(v 2 )
and R(v 1 )→W(v 2 ), as shown in Columns 6 and 7 of the table in
Figure 5.
By relaxing the write-atomicity requirement, all three weaker
memory models allow a thread to read its own write early. That is,
the thread can read a value it has written before the value reaches
the global memory and hence becomes visible to other threads.
3.3
Abstract Interpretation
Abstract interpretation is a popular technique for conducting static
program analysis [13]. In this context, a numerical abstract domain
defines, for every n ∈ N of the program, a memory environment
s. It is a map from each program variable to its abstract value1 .
Consider intervals, which map each variable to a region defined
by the lower and upper bounds. For a program with two integer
variables x and y where both may have any value initially, the
memory environment associated with the entry point n 0 ∈ N is
s 0 = {x 7→ ⊤, y 7→ ⊤}, where ⊤ = (−∞, +∞). After executing x =
1, the memory environment becomes s 1 = {x 7→ [1, 1], y 7→ ⊤}.
The process of computing s 1 based on s 0 is represented by the
transfer function of x = 1. Additionally, the join is defined as
[l 1 , u 1 ] ⊔ [l 2 , u 2 ] = [min(l 1 , l 2 ), max(u 1 , u 2 )]. The partial-order relation is defined as [l 1 , u 1 ] ⊑ [l 2 , u 2 ] if and only if l 1 ≥ l 2 and u 1 ≤ u 2 .
For example [1, 3] ⊔ [7, 10] = [1, 10] and [4, 6] ⊑ [1, 10].
We use S to denote the set of all memory environments. S is
a lattice with properly defined top (⊤) and bottom (⊥) elements,
join (⊔), partial-order (⊑), and a widening operator [13]. Each node
n ∈ N has a transfer function t ∈ S → S, taking an environment
s ′ ∈ S as input (before executing the atomic operation in n) and
returns a new environment s ∈ S as output.
Let TFunc ∈ N → (S → S) be a map from each node to its
transfer function. For example, given a node n ∈ N whose operation
1 For
ease of presentation we assume a variable maps to a single value. Our analysis
can trivially use relational domains.
is x = a+1, if s ′ = {x 7→ [1, 3], a 7→ [2, 5]}, the new environment
is s = (TFunc(n))(s ′ ) = {x 7→ [3, 6], a 7→ [2, 5]}.
The goal of an abstract interpreter is to compute an environment
map M ∈ (N → S) over-approximating the memory state at every
program location. M, typically, initially maps all variables in the
entry node to ⊤, and all variables in other nodes to ⊥. Then, it
iteratively applies the transfer function TFunc(n) and joins the
resulting environments for all n, until they reach a fixed point.
Without getting into more details (refer to the literature [13]),
we define the sequential analyzer as a fixed-point computation with
respect to the function AnalyzeSeq ∈ (N → S) → (N → S):
Ä
AnalyzeSeq(M) = n 7→ (TFunc(n))(
M(n ′ ))
(n ′,n)∈δ
Here, M(n ′ ) is the
à environment produced by a predecessor node
n ′ of n, and s ′ = (n ′,n)∈δ M(n ′ ) is the join of these environments.
(TFunc(n))(s ′ ) is the new memory environment produced by executing the operation in n. Applying this function to all nodes of a sequential program until a fixed point leads to an over-approximated
memory state for each program location.
However, directly applying the sequential analyzer to each execution of a multithreaded program is not practical because it
leads to an exponential complexity. Instead, thread-modular techniques [40, 47–49] iteratively apply AnalyzeSeq to each thread, as if
it were a sequential program, and then merge/propagate the global
memory effects across threads. The iterative process continues until
memory environments in all threads stabilize.
Since each thread is analyzed in isolation, this approach is more
scalable than non-thread modular techniques. However, it may
result in accuracy loss because the analyzer for each thread relies
on a coarse-grained abstraction of interferences from other threads.
When analyzing a thread t in the presence of a set of threads T , for
example, the interferences are the effects of global memory stores
from all t ′ ∈ T . The interferences are a map (V → 2S ) from each
variable v ∈ V read by thread t to the set of memory environments
produced by interfering stores, where V is the set of all program
variables, and 2S is the power set of S.
Prior thread-modular techniques [47–49] eagerly join all interfering memory states from the other threads in T before propagating
them to the current thread t. As such, they often introduce bogus
store-to-load data flows into the static analysis or miss valid storeto-load data flows. In the remainder of this paper, we present our
method for mitigating this problem.
4
DECIDING INTERFERENCE FEASIBILITY
In this section, we describe our new method for quickly deciding the
feasibility of a combination of store-to-load data-flows under a given
memory model. An interference combination is a set ic = {(l, s), . . .}
where each (l, s) ∈ ic is a load l and an interfering store s.
Checking the feasibility of ic is formulated as a deductive analysis with inputs: (1) the flow graph of the current thread, (2) the
flow graphs of all interfering threads, and (3) the existing set of
store-to-load data flows represented by {(l, s)|(l, s) ∈ ReadsFrom}.
The output of this deductive analysis is the relation MustNotReadFrom. (l, s) ∈ MustNotReadFrom means the load l must not read
from the store s since our analysis proved the data flow from s to l
is infeasible given the input ReadsFrom relation in ic.
Consider the program in Figure 3 as an example. One thread
interference combination we want to check is the load of y from
y=10 and the load of x from the initial value 0. Let these load
and store instructions be denoted ly , sy10 , l x , and s x0 , respectively.
Then, the feasibility problem is stated as follows: given (ly , sy10 ) ∈
ReadsFrom, check if (l x , s x0 ) ∈ MustNotReadFrom.
4.1
Notations
Before presenting the details of our feasibility checking procedure,
we define a set of unary and binary relations over instructions and
program variables. Specifically, (s 1 , v 1 ) ∈ IsLoad denotes s 1 is a load
of variable v 1 , and (s 2 , v 2 ) ∈ IsStore denotes s 2 is a store to variable
v 2 . We use (s 1 , _) ∈ IsLoad if we do not care about the variable.
Similarly, we use f ∈ IsFence to denote that f is a fence. We
also use IsLLMembar, IsLSMembar, IsSLMembar, IsSSMembar to
denote load–load, load–store, store–load, and store–store memory
barriers as defined in the SPARC architecture [65]; for example,
a load–store membar prevents loads before the barrier from being
reordered with subsequent stores.
We define binary relations over instructions s 1 and s 2 : the first
four relations (Dominates, NotReachableFrom, ThreadCreates,
ThreadJoins) are determined by the program’s flow graphs. Based
on them, we deduce the MHB relation, which must be satisfied by
all program executions. The ReadsFrom relation comes from the
given ic, from which we deduce the MustNotReadFrom relation.
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
(s 1, s 2 ) ∈
Dominates means that s 1 dominates s 2 in the control flow
graph of a thread.
NotReachableFrom means that s 1 cannot be reached from s 2
in the control flow graph of a thread.
ThreadCreates means s 1 is the thread creation and s 2 is the
first operation of the child thread.
ThreadJoins means s 1 is the thread join and s 2 is the last
operation of the child thread.
MHB means that s 1 must happen before s 2 in all executions of
the program.
ReadsFrom means that s 1 is a load that reads the value written
by the store s 2 .
MustNotReadFrom means that s 1 must not read from the
value written by s 2 .
Consider Figure 3 again, where we want to check if the load
of y in the second thread reads from y=10, then is it possible for
the load of x to read from the initial value 0? In this case, we
encode the assumption as (ly , sy10 ) ∈ ReadsFrom. Next, we deduce
the MustNotReadFrom relation. Finally, we check if (l x , s x0 ) ∈
MustNotReadFrom.
4.2
Relaxing the Program-Order Requirement
To model the program order imposed by different memory models,
we define a new relation NoReorder such that (s 1 , s 2 ) ∈ NoReorder
if the reordering of s 1 and s 2 within the same thread is not allowed.
We define the rules for NoReorder based on the allowed programorder relaxations for different memory models (Figure 5).
For SC, NoReorder is defined as:
⊤
(s 1, s 2 ) ∈ NoReorder
(under SC)
That is, no reordering is ever allowed under SC (row SC Figure 5).
For TSO, NoReorder is defined as:
(s 1, _) ∈ IsLoad
(s 1, s 2 ) ∈ NoReorder
(s 2, _) ∈ IsStore
(s 1, s 2 ) ∈ NoReorder
(under TSO)
(under TSO)
Under TSO, two operations (s 1 , s 2 ) can not reorder in six of the
eight cases. The first rule above disallows Columns 2, 3, 6, and 7
(Figure 5), while the second disallows Columns 3, 5, 7, and 9. Thus,
reordering is permitted in two cases: Columns 4 and 8.
Although this is counter-intuitive, note that W (v 1 ) → R(v 1 )
(Column 4) may be reordered in our analysis under TSO (and PSO
and RMO) for soundness: it permits read-own-write-early behaviors.
We detail this shortly in Section 4.6.
For PSO, NoReorder is defined as:
(s 1, _) ∈ IsLoad
(s 1, s 2 ) ∈ NoReorder
(under PSO)
(s 1, v 1 ) ∈ IsStore ∧ (s 2, v 1 ) ∈ IsStore
(s 1, s 2 ) ∈ NoReorder
(under PSO)
Under PSO, two operations (s 1 , s 2 ) can not reorder in five of the
eight cases. The first rule above disallows Columns 2, 3, 6, and 7,
while the second disallows Column 5. Thus, reordering is permitted
only in the remaining three cases (Columns 4, 8, and 9).
For RMO, the inference rules are defined as:
(s 1, v 1 ) ∈ IsLoad ∧ (s 2, v 1 ) ∈ IsLoad
(s 1, s 2 ) ∈ NoReorder
(s 1, v 1 ) ∈ IsLoad ∧ (s 2, v 1 ) ∈ IsStore
(s 1, s 2 ) ∈ NoReorder
(s 1, v 1 ) ∈ IsStore ∧ (s 2, v 1 ) ∈ IsStore
(s 1, s 2 ) ∈ NoReorder
(under RMO)
(under RMO)
(under RMO)
Similarly, the above inference rules can be directly translated from
Columns 2, 3, and 6 of the table in Figure 5.
4.3
Handling Fences and Memory Barriers
Next, we present the ordering constraints imposed by fences and
memory barriers. We consider four variants of the membar instruction, which prevents loads and/or stores before the membar from
being reordered with subsequent loads and/or stores [65].
m ∈ IsLLMembar ∧ (s 1, _) ∈ IsLoad ∧ (s 2, _) ∈ IsLoad
∧(s 1, m) ∈ Dominates ∧ (m, s 2 ) ∈ Dominates
(s 1, s 2 ) ∈ NoReorder
m ∈ IsLSMembar ∧ (s 1, _) ∈ IsLoad ∧ (s 2, _) ∈ IsStore
∧(s 1, m) ∈ Dominates ∧ (m, s 2 ) ∈ Dominates
(s 1, s 2 ) ∈ NoReorder
m ∈ IsSLMembar ∧ (s 1, _) ∈ IsStore ∧ (s 2, _) ∈ IsLoad
∧(s 1, m) ∈ Dominates ∧ (m, s 2 ) ∈ Dominates
(s 1, s 2 ) ∈ NoReorder
m ∈ IsSSMembar ∧ (s 1, _) ∈ IsStore ∧ (s 2, _) ∈ IsStore
∧(s 1, m) ∈ Dominates ∧ (m, s 2 ) ∈ Dominates
(s 1, s 2 ) ∈ NoReorder
We also model fences in terms of membars since they prevent
loads and stores from being reordered with subsequent loads and
stores as well.
(s, s sta ) ∈ ThreadCreates
(s, s end ) ∈ ThreadJoins
(s, s sta ) ∈ MHB
(s end , s) ∈ MHB
(s 1, s 2 ) ∈ Dominates ∧ (s 2, s 1 ) ∈ NotReachableFrom
∧ (s 1, s 2 ) ∈ NoReorder
(s 1, s 2 ) ∈ MHB
(s 1, s 2 ) ∈ MHB ∧ (s 2, s 3 ) ∈ MHB
(s 1, s 3 ) ∈ MHB
(l, s 1 ) ∈ ReadsFrom ∧ (s 1, s 2 ) ∈ MHB
∧ (l, v) ∈ IsLoad ∧ (s 1, v) ∈ IsStore ∧ (s 2, v) ∈ IsStore
s1
(1)
s1
DOMINATES
DOMINATES
(2)
s2
s2
(3)
(4)
Figure 8: Example illustrating Rule (2)
(l, s 2 ) ∈ MHB
Store(v)
s1
Figure 6: Deduction rules for MHB (must-happen-before).
(l, s) ∈ MHB
(l, s) ∈ MustNotReadFrom
(l 1, s 1 ) ∈ ReadsFrom ∧ (l 1, s 2 ) ∈ MHB ∧ (s 2, l 2 ) ∈ MHB
∧(l 1, v) ∈ IsLoad ∧ (l 2, v) ∈ IsLoad ∧ (s 2, v) ∈ IsStore
READSFROM
(5)
MHB
l
(6)
(l 2, s 1 ) ∈ MustNotReadFrom
Store(v)
MHB
Load(v)
s2
Figure 7: Deduction rules for the MustNotReadFrom.
Figure 9: Example illustrating Rule (4)
f ∈ IsFence
f ∈ IsFence
f ∈ IsLLMembar
f ∈ IsFence
f ∈ IsLSMembar
f ∈ IsFence
f ∈ IsSLMembar
f ∈ IsSSMembar
Store(v)
Rules for Deducing MustNotReadFrom
We divide our inference rules into two groups. The first (Figure 6)
use the relations ThreadCreates, ThreadJoins, Dominates, and
NoReorder to generate the must-happen-before (MHB) relation.
Rule (1) states that if the instruction s creates a thread with
entry instruction s sta , then s must happen before s sta . Similarly,
if instruction s joins a thread with exit instruction s end , then s end
must happen before s.
Rule (2) states that if s 1 dominates s 2 within a thread’s CFG,
and s 1 is not reachable from s 2 , (i.e., no loop encompasses both s 1
and s 2 ), then, if permitted by the memory model, s 1 must happen
before s 2 . Figure 8 exemplifies this rule: the loop in the left CFG is
outside the Dominates edge, thus (s 1 , s 2 ) ∈ NotReachableFrom.
The loop in the right CFG encompasses the Dominates edge, thus
(s 1 , s 2 ) < NotReachableFrom.
Rule (3) states that the MHB relation is transitive: if s 1 must
happen before s 2 , and s 2 must happen before s 3 , then s 1 must happen
before s 3 . Correctness follows from the definition of MHB.
Rule (4) states that if a load l reads from the value written by the
store s 1 , then l must happen before some second store to the same
variable s 2 takes effect. This is intuitive because, if s 2 takes effect
READSFROM
l1
In addition to fences explicitly added to the program, there are
fences implicitly added to thread routines such as lock/unlock and
signal/wait. For example, in the code snippet x = 1; lock(lk);
a = y; unlock(lk), there is a fence inside lock(lk), to ensure x = 1 always takes effect before a = y. This is how most
modern programming systems guarantee data-race-freedom [3]
to application-level code (i.e., programs without data races have
only SC behaviors). Thus, we model every call sc to a POSIX thread
routine using sc ∈ IsFence.
4.4
s1
Load(v)
MHB
MustNotReadFrom
s2
Store(v)
MHB
Load(v)
l2
Figure 10: Example illustrating Rule (6).
before l (but after the first store s 1 ), then l can no longer read from
s 1 . Figure 9 exemplifies this rule. Its correctness is obvious.
The second group of inference rules (Figure 7) takes the relations
MHB and ReadsFrom and generates the MustNotReadFrom relation. Recall that if a load-store pair (l, s) ∈ MustNotReadFrom,
the value stored by s can never flow to l. Thus, MustNotReadFrom
may be used to eliminate infeasible data flows.
Rule (5) states that if a load l must happen before a store s, then l
cannot read from s. This follows from the definition of MHB. Note
that a store s “happens” when it propagates to main memory.
Rule (6) states that if a load l 1 reads from a store s 1 , and l 1 must
happen before some other store s 2 , and s 2 must happen before a
second load l 2 , then l 2 cannot read from s 1 . Figure 10 exemplifies
this rule. This is correct because l 2 reading from s 1 would mean s 1
takes effect after s 2 thus preventing l 1 from reading s 1 .
4.5
Soundness and Incompleteness
When deciding the feasibility of an interference combination our
analysis is designed to be sound but incomplete. By sound we mean
it permits all possible program behaviors allowed by a memory
model. Therefore, if it says a certain interference combination is
infeasible it must be infeasible. However, there is no guarantee
every infeasible interference combination will be found.
Incompleteness is expected: the intent is a quick pruning of infeasible combinations before the computationally expensive threadlevel analysis. The overhead of insisting on completeness would
outweigh its benefit: the feasibility checking problem, in the worst
case, is as hard as program verification itself, which is undecidable.
Now, we formally state the soundness of our deductive procedure. First, our deduction of the NoReorder relation relaxing the
program order requirement, from Figure 5, is sound.
this trace, we must allow the following interference combination:
s 2 reads from s 1 , s 3 reads from the initial value 0, and s 6 reads from
the initial value 0. This combination is feasible only when we avoid
enforcing the program order between s 1 and s 2 . Specifically, the
statements in thread 2 follow program order (s 4 , s 5 , s 6 ) from the
fence. In thread 1, s 2 and s 3 are ordered since they are added to
NoReorder under TSO. But, statements s 1 and s 2 are not added to
NoReorder, thus preventing the assertion from being (incorrectly)
verified.
Theorem 4.1. Let s 1 and s 2 be two instructions in the same thread.
If our rules deduce (s 1 , s 2 ) ∈ NoReorder, then the reordering of s 1
and s 2 is not allowed by the corresponding memory model.
5
The proof of this theorem is straightforward, since our inference
rules for deducing NoReorder directly follow the memory model
semantics provided by Adve and Gharachorloo [4] in Figure 5.
Next, we note that, given the ReadsFrom relation, the deduction
of the MustNotReadFrom relation is also sound.
Theorem 4.2. Let l and s be two instructions. If our rules deduce
to (l, s) ∈ MustNotReadFrom, then l cannot read from s.
The proof of this theorem is straightforward: it amounts to proofs
of Rules (1)–(6). During the previous presentation, we have argued
why each rule is correct. More formal proofs can be obtained via
proof-by-contradiction, which is straightforward. We omit the details
for brevity.
4.6
Relaxing the Write-Atomicity Requirement
Our method soundly models buffer forwarding, which corresponds
to the write-atomicity requirement (Column 10 Figure 5). This
allows a thread to read its own write before the written value is
flushed to the memory, thus becoming visible to other threads. This
is modeled in both the thread-level analyzer (AnalyzeSeq) and the
deduction rules.
AnalyzeSeq captures the relaxation for free. During this analysis
each thread is treated as a sequential program: all loads read their
values from the preceding writes within the same thread.
The deduction rules for NoReorder (Section 4.2) always permit the reordering of a store with a subsequent load of the same
variable (Column 4, Figure 5). That is, if (s 1 , v 1 ) ∈ IsStore and
(s 2 , v 1 ) ∈ IsLoad, we do not deduce (s 1 , s 2 ) ∈ NoReorder due to
buffer forwarding (even though it is counter-intuitive). Within a
thread t, it may appear to be the case that the store and load are
reordered from the perspective of all threads t ′ , t. Forbidding
this reordering would be equivalent to forcing a full flush of the
store-buffers before every load, thus prohibiting any thread from
reading its own store earlier than other threads.
void thread1 () {
x = 1; // s1
a = x; // s2
b = y; // s3
}
void thread2 () {
y = 1; // s4
fence ; // s5
c = x ; // s6
}
assert( !(a ==1 && b == 0 && c == 0) );
Figure 11: Write atomicity example under TSO.
Figure 11 exemplifies the requirement of this relaxation. First,
the assertion may be violated under TSO. An error trace is: x = 1;
a = 1; b = 0; y = 1; flush y; c = 0; flush x. To permit
THE THREAD-MODULAR ANALYSIS
Next, we present the integration of our interference analysis (Section 4) with a thread-modular analyzer. The thread-modular analyzer itself is standard, whose full details may be found in several
prior works including [48, 49] and [40]. Thus, our presentation
of the analyzer itself will be terse. Instead, we shall focus on our
main contribution, which is adding the capability of deducing infeasible interference combinations for weak-memory models: our
method is sound for not only SC but also TSO, PSO, and RMO. Prior
techniques were either MM-oblivious, or sound only for SC.
Given a load l of v, the interferences on l, within the threadmodular analysis, are the environments after all stores to v from
other threads. The function n(G) takes a graph as input and returns
the nodes of the graph. The interferences on the loads in a thread
G is the least fixed point of the function Interfs ′ .
Interfs ′ (G, M, I ) = l 7→ {e} ∪ I (l)
where e is the environment after a remote store s < n(G)
to the same variable as loaded by load l ∈ n(G)
Interfs(G, M) = lfp(Interfs ′ (G, M), I ⊥ )
We use Interfs ′ (G, M, I ) as shorthand for Interfs ′ (G, M)(I ), where
Interfs ′ (G, M) is a partially-applied function, and use I ⊥ as the initial map from loads to interfering-environments, i.e., one mapping
all nodes to {⊥}. lfp computes the least fixed point. Interfs ′ depends
on the existence of M, a map from each program location, in all
threads, to an environment. We show shortly that the computation
of M and the interferences is done in a nested fixed point.
We refer to an interference combination, ic ∈ (N 7→ S), as a map
from a load l to the memory environment after a store instruction from which l reads. This differs slightly from the definition
of Section 4 where it is defined as a set of load-stores pairs. The
two can be easily converted as the analysis keeps track of all the
environments associated with each store. Given the set of interferences I from Interfs, the set of all interference-combinations are
all permutations of selecting a single environment from I for each
load. The iterative thread-modular analysis separately considers
each interference-combination thus increasing accuracy.
The thread analyzer adapts the sequential analyzer (AnalyzeSeq,
Section 3) to use interference-combinations. AnalyzeTM ′ takes a
thread G and an interference combination ic and computes the input
environment e for some node n in G by joining the environment
after the predecessors of n with n’s environment in ic, denoted ic(n).
Then, e is passed to n’s transfer function to update M(n).
AnalyzeTM ′ (G, ic, M) = n 7→ TFunc(n)(e)
Ä
where e =
M(n ′ ) ⊔ ic(n)
(n ′,n)∈t(G)
AnalyzeTM(G, ic) = lfp(AnalyzeTM ′ (G, ic), M ⊥ )
AnalyzeTM is the least fixed point of AnalyzeTM ′ . t(G) returns
the transition-relation of a graph. M ⊥ is the initial memory map
mapping the entry nodes of each thread to ⊤ and all others to ⊥.
Given a set of threads Gs and a set of interference-combinations I ,
applying AnalyzeTM to each G ∈ Gs and each ic ∈ I computes the
analysis over all threads.
What remains is to show how the thread analyzer and the calculation of interferences can be done simultaneously since they are
dependent: the interference computation depends on the analysis
result, M, and the analysis result depends on the set of interferences, I . The solution is a nested fixed point: the outer computation
produces M, and the inner computation produces I . The process
iterates until M (and thus I ) reach a fixed point.
Analyze(G, M) = M ′
where I = Interfs(G, M)
I ′ = FilterFeasible(I )
M ′ = JoinMM(map(AnalyzeTM(G), I ′ ))
AnalyzeAll ′ (Gs, M) = JoinMM(map(Analyze(M), Gs))
AnalyzeAll(Gs) = lfp(AnalyzeAll ′ (Gs), M ⊥ )
Analyze operates as follows: first, it takes M, the current analysis
results over all threads, and computes the interferences, I , wrt
the thread under test, G. The function FilterFeasible integrates the
thread-level analyzer with the feasibility analysis of Section 4. It
expands the interferences I into a set of interference combinations
I ′ , and filters any infeasible combination.
Specifically, given the interferences on a thread, I = {(l 1 7→
{e 1 , e 2 , . . .}), (l 2 7→ {e 3 , e 4 , . . .}) . . .}, FilterFeasible creates all combinations of pairing each load to a single interfering environment,
e.g., Ie = {{⟨l 1 , e 1 ⟩, ⟨l 2 , e 3 ⟩, . . .}, {⟨l 1 , e 2 ⟩, ⟨l 2 , e 3 ⟩, . . .}, . . .}. Then, it
maps each environment in Ie to the associated store generating the
environment, e.g., Is = {{⟨l 1 , s 1 ⟩, ⟨l 2 , s 3 ⟩}, . . .}. Each set of pairs
of load and store statements in Is is then passed to the deduction
analysis of Section 4. If it is infeasible, it is discarded, otherwise it
is added to the set I ′ returned by FilterFeasible.
map(AnalyzeTM(G), I ′ ) ∈ 2(N7→S) is the set of the results of
applying AnalyzeTM(G, i) for each i ∈ I ′ . Specifically, map ∈ (A →
B) → 2A → 2B takes a function f and a set S, and returns a set
containing the application of f on each element of S.
JoinMM ∈ (2(N7→S) → (N 7→ S)) takes the join of memory
environments on matching nodes across a set of maps to join them
into a single map. Similarly, AnalyzeAll ′ joins the results of applying
Analyze to the set Gs of threads. AnalyzeAll computes the fixed
point of AnalyzeAll ′ starting with the initial map M ⊥ .
The following is a high-level example. Initially, each thread G is
analyzed in the presence of M ⊥ resulting in the set of interferences,
I , being empty (all stores map to ⊥). The results of analyzing each
thread are merged into a new map M. Each thread is then analyzed
using M, resulting in the sets I and I ′ to be (potentially) non-empty,
causing AnalyzeTM to be called once per-combination. Within a
thread, the results of AnalyzeTM are joined, then, across threads, the
results of Analyze are joined, creating M ′ . The procedure repeats,
thus growing the size of M, I , and I ′ until M = M ′ .
We handle loops the same way as in prior techniques (e.g., [40]).
Given a load l within a loop the previously described analysis can
generate an infinite number of interference combinations for l,
e.g., when l is within an infinite loop. Loops are unrolled when
possible, and, when not, we join all the feasible interfering memory
environments into a single value. An interfering environment e is
infeasible to interfere on l if the store generating e must-happen
after l; otherwise, it is feasible. This is sound for verifying assertions
embedded in a concurrent program [40].
6
EXPERIMENTS
We implemented our weak-memory-aware abstract interpreter in a
tool named FruitTree, building upon open-source platforms such
as LLVM [5], Apron [37], and µZ [30]. Specifically, we use LLVM
to translate C programs into LLVM bit-code, based on which we
perform static analysis. We use the Apron library to manipulate
abstract domains in the thread analyzer. We use the µZ fixed-point
engine in Z3 [15] to solve Datalog constraints that encode the
feasibility of interference combinations.
We implemented the state-of-the-art MM-oblivious abstract interpretation method of Miné [49], and the SC-specific method,
Watts [40], on the same platform to facilitate experimental evaluation. We also compared against a previously implemented version
of Duet [17, 18]. While Duet may be unsound, and Watts is unsound, we include their results because they are closely related to
our new technique.
All methods implemented in FruitTree use the clustering and
property-directed optimizations [40], where clustering considers
interferences only within sets of loads, similar to the packing of
relational domains, and property-direction filters interference combinations unrelated to properties under test. These optimizations
reduce the number of interference combinations, which is crucial
since it grows exponentially with respect to program’s size.
We evaluated FruitTree on a large set of programs written
using the POSIX threads. These benchmarks fall into two categories. The first are 209 litmus tests exposing non-SC behaviors
under various processor-level memory models [8]. The second are
52 larger applications [17, 60, 61], including several Linux device
drivers. The benchmarks total 61,981 lines of code. The properties
under verification are assertions embedded in the program’s source
code: a property is valid if and only if the assertion holds over all
executions under a given memory model.
Our experiments were designed to answer the following research
questions: (1) Is our new method more effective than prior techniques in obtaining correctness proofs on relaxed memory? (2) Is
our new method more accurate than prior techniques in detecting
potential violations on relaxed memory? (3) Is our new method
reasonably efficient when used as a static program analysis technique? We conducted all experiments on a Linux computer with 8
GB RAM, and a 2.60 GHz CPU.
6.1
Litmus Test Results
First, we present the litmus test results. Since these programs are
small in terms of code size, all methods under evaluation (Miné,
Watts, Duet, and FruitTree) finished quickly. Thus, our focus
is not on comparing the runtime performance but comparing the
accuracy of their results. Specifically, we compare our method to
these state-of-the-art techniques in terms of the number of true
proofs, bogus proofs, true alarms, and bogus alarms.
Here, a bogus alarm is a valid property which cannot be proved.
A bogus proof is a property which may be violated yet is unsoundly
and incorrectly proved. The litmus tests are particularly useful not
only because they cover corner cases, but also because we know a
priori if a property holds or not.
Table 1: Results on the litmus test programs under TSO.
Method
Miné [49]
Duet [17]
Watts [40]
FruitTree
True Alarm
Bogus Alarm
True Proof
Bogus Proof
77
77
63
77
207
181
13
72
8
34*
0
143
0
0
216⋆
0
Time (s)
12.9
473.1
71.0
89.2
Table 1 summarizes the litmus test results under TSO. The first
column shows the name of each method, and the next four show the
number of true alarms, bogus alarms, true proofs, and bogus proofs
generated by each method, respectively. Since Watts [40] was
designed to be SC-specific, it ignores non-SC behaviors, meaning
its proofs are unsound under weaker memory (marked by ⋆ ). The
last column is the total analysis time over all tests.
Overall, the results show the prior thread-modular technique of
Miné admits many infeasible executions thus leading to 207 bogus
alarms. Duet reported 181 bogus alarms. In contrast, our method
(FruitTree) reported only 72 bogus alarms, together with 143 true
proofs. Therefore, it is more accurate than these prior techniques.
Although Watts reported only 13 bogus alarms, it is unsound
for TSO: it only considers SC behaviors and cannot be trusted.
Furthermore, the soundness of Duet under TSO or any other nonSC memory model was not clear (since Duet was only designed
for SC). Thus, in the result table, its 34 proofs are marked with *.
Table 2: Results on the litmus test programs under PSO.
Method
Miné [49]
Duet [17]
Watts [40]
FruitTree
True Alarm
Bogus Alarm
True Proof
Bogus Proof
81
81
64
81
203
177
12
68
8
34*
0
143
0
0
216⋆
0
Time (s)
12.9
473.1
71.0
281.4
Table 2 summarizes the results under PSO. Again, Watts may be
unsound for weak memory. The same litmus programs were used
under PSO as in TSO but the properties changed, i.e., whether an
alarm is true or bogus. Note that Miné only verified 8 properties,
Duet verified 34, whereas our method verified 143.
abstracting all other threads into a set of interferences—it cannot
capture ordering constraints involving more than two threads. In
principle, this limitation can be lifted by extending our interference
feasibility analysis: we leave this as future work.
6.2
Results on Larger Applications
Next, we present our results on the larger benchmark programs.
Since execution time is no longer negligible, we compare, across
methods, both the run time and accuracy. However, since the programs are larger (60K lines of code) and there are far too many
properties to manually inspecting each case, we do not report the
number of bogus alarms and bogus proofs due to lack of the ground
truth. Instead, we compare the total number of proofs reported by
each method, to show our method is more accurate even though
all methods are approximate.
Table 4 shows our results under TSO, where ⋆ and * mark the
unsoundly verified properties. Since the results for PSO and RMO
are similar to Table 4, we omit them for brevity. Column 1 of this
table shows the name of the benchmark program. Columns 2–3,
4–5, 6–7, and 8–9 show the run time and number of properties
verified by Miné, Duet, Watts, and FruitTree, respectively.
Again, while the proofs reported by FruitTree and Miné’s
method are sound, the proofs reported by Watts are not, and the
soundness of Duet on weak memory is unclear.
Overall, FruitTree proved 4,577 properties compared to only
1,712 proved by Miné, an increase of 2.7x more properties relative to
prior state-of-the-art. Additionally, though Duet may be unsound,
it proved only 2,432 properties. The definitely-unsound Watts
“proved” 4,583 properties, possibly including bogus proofs.
In terms of the run time, FruitTree took 5,387 seconds, which
is similar to Watts, and slower than Duet and Miné. However,
the additional time is well justified due to the significant increase
in the number of proofs. Furthermore, the runtime performance
– proving 1 property per second – remains competitive as a static
analysis technique.
To summarize, our new method has modest runtime overhead
compared to prior techniques, but vastly improved accuracy in
terms of the analysis results, and is provably sound in handling not
only SC but also three other processor-level memory models.
Table 3: Results on the litmus test programs under RMO.
7
Method
Miné [49]
Duet [17]
Watts [40]
FruitTree
True Alarm
Bogus Alarm
True Proof
Bogus Proof
28
11
0
28
67
58
0⋆
13
8
34
75
62
0
0
28⋆
0
Time (s)
4.9
187.8
33.9
46.9
Table 3 summarizes the results under RMO. Under RMO, a different set of litmus programs were used since the instruction set
for processors using RMO differs from TSO and PSO. Nevertheless,
we observed similar results: FruitTree obtained significantly more
true proofs and fewer bogus alarms than the other methods.
In general, our method was more accurate than prior techniques.
However, since the analysis is over-approximated, it does not eliminate all bogus alarms. Currently, most bogus alarms reported by
FruitTree require reasoning across more than two threads, e.g.,
the correctness of a property may require reasoning that thread T1
reading x = 1 from thread T2 implies y = 1 in thread T3 . Since our
method is thread-modular—threads are analyzed individually by
RELATED WORK
We reviewed prior work on thread-modular abstract interpretation,
which are either MM-oblivious [47–49] or SC-specific [40] in processor memory-models. There are also techniques [17, 18, 35, 53]
that are not thread-modular.
There are code-transformation techniques [14, 39, 45] that transform a non-SC program into an SC program and then apply abstract
interpretation. They generally follow the sequentialization approach
pioneered by Lal and Reps [41], with a focus on code transformation as opposed to abstract interpretation. To ensure termination,
they make various assumptions to bound the program’s behavior.
Furthermore, they are not thread-modular, and often do not directly
handle C code. Instead, they admit only models of concurrent programs written in artificial languages; because of this, we were not
able to perform a direct experimental comparison.
In the context of bounded model checking, Alglave et al. proposed several methods for concurrent software on relaxed memory.
They are based on either sequentializing concurrent programs [8]
or encoding weak memory semantics using SAT/SMT solvers [7, 9].
Table 4: Results on larger applications (total of 61,981 LOC).
Minè [49]
Duet [17]
Watts [40]
FruitTree
Name
Time
Verif
Time
Verif
Time
Verif
Time
Verif
thread00
threadcreate01
threadcreate02
sync_01_true
sync_02_true
intra01
dekker1
fk2012_v2
keybISR
ib700wdt_01
ib700wdt_02
ib700wdt_03
i8xx_tco_01
i8xx_tco_02
i8xx_tco_03
machzwd_01
machzwd_02
machzwd_03
mixcomwd_01
mixcomwd_02
mixcomwd_03
pcwd_01
pcwd_02
pcwd_03
pcwd_04
sbc60xxwdt_01
sbc60xxwdt_02
sbc60xxwdt_03
sc1200wdt_01
sc1200wdt_02
sc1200wdt_03
smsc37b787wdt_01
smsc37b787wdt_02
smsc37b787wdt_03
sc520wdt_01
sc520wdt_02
sc520wdt_03
w83877fwdt_01
w83877fwdt_02
w83877fwdt_03
wdt01
wdt02
wdt03
wdt977_01
wdt977_02
wdt977_03
wdt_pci01
wdt_pci02
wdt_pci03
pcwd_pci_01
pcwd_pci_02
pcwd_pci_03
0.01
0.02
0.02
0.03
0.03
0.02
0.10
0.04
0.04
1.7
13.2
23.1
1.0
8.5
10.7
0.6
1.6
4.1
0.8
2.2
3.5
0.7
4.4
10.2
25.4
1.1
3.3
7.0
1.0
6.9
26.2
1.0
9.2
26.4
1.5
12.6
27.8
12.5
29.0
54.2
0.2
0.3
0.5
1.3
2.4
8.2
1.2
8.9
23.9
4.7
9.8
20.3
0
1
1
1
1
1
3
3
4
23
63
81
14
34
37
14
24
39
12
32
64
10
27
40
60
21
40
60
22
58
102
22
76
130
15
41
58
34
50
66
3
5
6
9
13
25
11
31
51
56
70
88
1.1
0.8
0.7
1.3
1.2
1.2
1.3
1.4
1.2
1.8
2.7
2.8
2.8
4.6
4.7
1.3
1.4
1.8
1.9
4.4
2.9
0.9
1.3
1.7
2.1
2.0
2.3
3.2
1.3
2.2
3.3
1.2
2.3
3.4
1.1
1.7
2.2
1.8
2.3
2.7
1.4
1.5
1.5
1.9
2.3
2.9
1.1
1.8
3.0
1.3
1.4
2.0
0
1
2
1
1
0
2
3
4
35
95
122
28
68
74
14
24
39
21
41
65
22
56
82
122
0
0
0
10
28
50
23
77
131
16
42
59
34
50
66
13
21
25
35
51
99
31
91
151
89
132
186
0.03
0.05
0.03
0.06
0.07
0.04
6.1
0.3
2.1
14.5
108.4
178.4
97.4
1288.3
1677.0
107.5
240.5
488.7
169.3
768.0
88.3
1.8
8.5
18.5
46.1
4.9
13.7
33.4
15.1
64.5
197.0
16.7
91.4
286.1
15.6
74.1
155.6
83.9
189.6
357.9
65.9
600.2
1479.1
83.0
115.6
264.9
5.9
33.2
93.5
27.2
52.6
97.7
0
2
2
1
1
2
4
4
6
46
126
162
39
99
108
35
65
110
34
64
100
31
82
121
181
43
81
121
33
87
153
46
154
262
45
123
174
137
201
265
14
22
26
43
63
123
52
152
252
116
158
212
0.03
0.04
0.03
0.06
0.07
0.04
6.6
0.2
2.3
11.4
89.4
156.3
53.3
757.3
952.3
42.1
75.3
128.9
15.3
61.7
84.0
1.4
6.8
15.4
39.8
3.0
8.0
17.9
10.6
47.6
146.7
12.0
67.3
197.5
11.0
56.0
116.9
71.0
159.6
301.9
27.6
306.5
766.2
49.7
76.0
193.9
4.6
26.3
72.7
21.1
40.6
72.7
0
2
2
1
1
2
4
4
5
46
126
162
39
99
108
34
64
109
33
63
100
31
82
121
181
43
81
121
33
87
153
46
154
262
45
123
174
137
201
265
14
22
26
43
63
123
52
152
252
116
158
212
Total
415 s
1752
106 s
2432*
9830 s
4583⋆
5387 s
4577
interpretation, which covers all possible program behaviors and
therefore is geared toward obtaining correctness proofs.
Thread-modular analysis was also used in model checking [23,
24], where it was combined with predicate abstraction [29] to help
mitigate state explosion and thus increase the scalability. However,
model checking is significantly different from abstract interpretation in that each thread must be first abstracted into a finite-state
model. Thread-modular analysis was also used to conduct shape
analysis [26] and prove thread termination [12]. Hoenicke et al. [31]
introduced a hierarchy of proof systems that leverage thread modularity in compositional verification on SC memory.
Similar to the interference analysis in Watts [40], we check
the feasibility of thread interactions using Datalog. Datalog-based
declarative program analysis was a framework introduced by Whaley and Lam [66]. Previously, it has been used to implement pointsto [11, 42], dependency [28, 59] and change-impact analyses [27],
uncover security bugs [44] and detect data races [50].
In abstract interpretation of sequential programs, Miné [46] proposed a technique for abstracting the global memory into a set
of byte-level cells to support a variety of casts and union types.
Ferrara et al. [20, 21] integrated heap abstraction and numerical
abstraction during static analysis, where the heap is represented
as disjunctions of points-to constraints based on values. Jeannet
and Serwe [38] also proposed a method for abstracting the data and
control portions of a call-stack for analyzing sequential programs
with potentially infinite recursion. Subsequently, Jeannet [36] extended the work to handle concurrent programs as well. However,
none of these methods was designed specifically for handling weak
memory models.
8
CONCLUSIONS
We have presented a thread-modular static analysis method for
concurrent programs under weak memory models, building upon a
lightweight constraint system for quickly identifying the infeasibility of thread interference combinations, so they are skipped during
the expensive abstract-interpretation based analysis. The constraint
system is also general enough to handle a range of processor-level
memory models. We have implemented the method and conducted
experiments on a large number of benchmark programs. We showed
the new method significantly outperformed three state-of-the-art
techniques in terms of accuracy while maintaining only a moderate
runtime overhead.
ACKNOWLEDGMENTS
Alglave et al. also developed techniques for modeling and testing
weak-memory semantics of real processors [10], and characterized
the memory models of some GPUs [6]. However, these techniques
are primarily for detecting buggy behaviors as opposed to proving
that such behaviors do not exist.
In the context of systematic testing, often based on stateless
model checking [22, 25, 64] or predictive analysis [33, 54, 56, 57, 62,
63], a number of methods have been proposed to handle weak memory such as TSO/PSO [1, 16, 34, 67], PowerPC [2], and C++11 [51,
52]. However, since they rely on concretely executing the program,
and require the user to provide test inputs, they can only be used
to detect bugs. That is, since testing does not cover all program
behaviors, if no bug is detected, these methods cannot obtain a
correctness proof. In contrast, our method is based on abstract
This material is based upon research supported in part by the U.S.
National Science Foundation under grants CNS-1405697 and CCF1722710 and the U.S. Office of Naval Research under award number
N00014-13-1-0527.
REFERENCES
[1] Parosh Aziz Abdulla, Stavros Aronis, Mohamed Faouzi Atig, Bengt Jonsson,
Carl Leonardsson, and Konstantinos F. Sagonas. Stateless model checking for
TSO and PSO. In International Conference on Tools and Algorithms for
Construction and Analysis of Systems, pages 353–367, 2015.
[2] Parosh Aziz Abdulla, Mohamed Faouzi Atig, Bengt Jonsson, and Carl
Leonardsson. Stateless model checking for POWER. In International Conference
on Computer Aided Verification, pages 134–156, 2016.
[3] Sarita V. Adve and Hans-Juergen Boehm. Memory models: a case for rethinking
parallel languages and hardware. Commun. ACM, 53(8):90–101, 2010.
[4] Sarita V. Adve and Kourosh Gharachorloo. Shared memory consistency models:
A tutorial. Computer, 29(12):66–76, 1996.
[5] Vikram Adve, Chris Lattner, Michael Brukman, Anand Shukla, and Brian Gaeke.
LLVM: A low-level virtual instruction set architecture. In ACM/IEEE
international symposium on Microarchitecture, Dec 2003.
[6] Jade Alglave, Mark Batty, Alastair F. Donaldson, Ganesh Gopalakrishnan, Jeroen
Ketema, Daniel Poetzl, Tyler Sorensen, and John Wickerson. GPU concurrency:
Weak behaviours and programming assumptions. In International Conference on
Architectural Support for Programming Languages and Operating Systems, pages
577–591, 2015.
[7] Jade Alglave, Daniel Kroening, Vincent Nimal, and Daniel Poetzl. Don’t sit on
the fence. In International Conference on Computer Aided Verification, pages
508–524, 2014.
[8] Jade Alglave, Daniel Kroening, Vincent Nimal, and Michael Tautschnig.
Software verification for weak memory via program transformation. In
European Symposium on Programming, pages 512–532, 2013.
[9] Jade Alglave, Daniel Kroening, and Michael Tautschnig. Partial orders for
efficient bounded model checking ofÂăconcurrentÂăsoftware. In International
Conference on Computer Aided Verification, pages 141–157, 2013.
[10] Jade Alglave, Luc Maranget, and Michael Tautschnig. Herding cats: modelling,
simulation, testing, and data-mining for weak memory. In ACM SIGPLAN
Conference on Programming Language Design and Implementation, page 7, 2014.
[11] Martin Bravenboer and Yannis Smaragdakis. Strictly declarative specification of
sophisticated points-to analyses. In ACM SIGPLAN Conference on Object
Oriented Programming, Systems, Languages, and Applications, pages 243–262,
2009.
[12] Byron Cook, Andreas Podelski, and Andrey Rybalchenko. Proving thread
termination. ACM SIGPLAN Conference on Programming Language Design and
Implementation, pages 320–330, 2007.
[13] P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for
static analysis of programs by construction or approximation of fixpoints. In
ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages,
pages 238–252, 1977.
[14] Andrei Dan, Yuri Meshman, Martin Vechev, and Eran Yahav. Effective
abstractions for verification under relaxed memory models. In International
Conference on Verification, Model Checking, and Abstract Interpretation, pages
449–466, 2015.
[15] Leonardo De Moura and Nikolaj Bjørner. Z3: An efficient SMT solver. In
International Conference on Tools and Algorithms for the Construction and
Analysis of Systems, pages 337–340, 2008.
[16] Brian Demsky and Patrick Lam. SATCheck: SAT-directed stateless model
checking for SC and TSO. In ACM SIGPLAN Conference on Object Oriented
Programming, Systems, Languages, and Applications, pages 20–36, 2015.
[17] Azadeh Farzan and Zachary Kincaid. Verification of parameterized concurrent
programs by modular reasoning about data and control. In ACM
SIGACT-SIGPLAN Symposium on Principles of Programming Languages, pages
297–308, 2012.
[18] Azadeh Farzan and Zachary Kincaid. Duet: Static analysis for unbounded
parallelism. In International Conference on Computer Aided Verification, pages
191–196, 2013.
[19] Pietro Ferrara. Static analysis via abstract interpretation of the happens-before
memory model. In International Conference on Tests and Proofs, pages 116–133.
2008.
[20] Pietro Ferrara. Generic combination of heap and value analyses in abstract
interpretation. In International Conference on Verification, Model Checking, and
Abstract Interpretation, pages 302–321, 2014.
[21] Pietro Ferrara, Peter Müller, and Milos Novacek. Automatic inference of heap
properties exploiting value domains. In International Conference on Verification,
Model Checking, and Abstract Interpretation, pages 393–411, 2015.
[22] C. Flanagan and P. Godefroid. Dynamic partial-order reduction for model
checking software. In ACM SIGACT-SIGPLAN Symposium on Principles of
Programming Languages, pages 110–121, 2005.
[23] Cormac Flanagan, Stephen N. Freund, and Shaz Qadeer. Thread-modular
verification for shared-memory programs. In European Symposium on
Programming, pages 262–277, 2002.
[24] Cormac Flanagan and Shaz Qadeer. Thread-modular model checking. In
International SPIN Workshop on Model Checking Software, pages 213–224, 2003.
[25] Patrice Godefroid. VeriSoft: A tool for the automatic analysis of concurrent
reactive software. In International Conference on Computer Aided Verification,
pages 476–479, 1997.
[26] Alexey Gotsman, Josh Berdine, Byron Cook, and Mooly Sagiv. Thread-modular
shape analysis. ACM SIGPLAN Conference on Programming Language Design
and Implementation, pages 266–277, 2007.
[27] Shengjian Guo, Markus Kusano, and Chao Wang. Conc-iSE: Incremental
symbolic execution of concurrent software. In IEEE/ACM International
Conference On Automated Software Engineering, 2016.
[28] Shengjian Guo, Markus Kusano, Chao Wang, Zijiang Yang, and Aarti Gupta.
Assertion guided symbolic execution of multithreaded programs. In ACM
SIGSOFT Symposium on Foundations of Software Engineering, pages 854–865,
2015.
[29] Thomas A. Henzinger, Ranjit Jhala, Rupak Majumdar, and Shaz Qadeer.
Thread-modular abstraction refinement. In International Conference on
Computer Aided Verification, pages 262–274, 2003.
[30] Krystof Hoder, Nikolaj Bjørner, and Leonardo de Moura. muZ - an efficient
engine for fixed points with constraints. In International Conference on
Computer Aided Verification, pages 457–462, 2011.
[31] Jochen Hoenicke, Rupak Majumdar, and Andreas Podelski. Thread modularity
at many levels: A pearl in compositional verification. In ACM SIGACT-SIGPLAN
Symposium on Principles of Programming Languages, pages 473–485, 2017.
[32] Alan Huang. Maximally stateless model checking for concurrent bugs under
relaxed memory models. In International Conference on Software Engineering,
pages 686–688, 2016.
[33] Jeff Huang, Patrick O’Neil Meredith, and Grigore Rosu. Maximal sound
predictive race detection with control flow abstraction. In ACM SIGPLAN
Conference on Programming Language Design and Implementation, pages
337–348, 2014.
[34] Shiyou Huang and Jeff Huang. Maximal causality reduction for TSO and PSO.
In ACM SIGPLAN Conference on Object Oriented Programming, Systems,
Languages, and Applications, pages 447–461, 2016.
[35] Thuan Quang Huynh and Abhik Roychoudhury. A memory model sensitive
checker for c#. In International Symposium on Formal Methods, pages 476–491,
2006.
[36] Bertrand Jeannet. Relational interprocedural verification of concurrent
programs. Software & Systems Modeling, 12(2):285–306, 2012.
[37] Bertrand Jeannet and Antoine Miné. Apron: A library of numerical abstract
domains for static analysis. In Ahmed Bouajjani and Oded Maler, editors,
International Conference on Computer Aided Verification, pages 661–667. 2009.
[38] Bertrand Jeannet and Wendelin Serwe. Abstracting call-stacks for
interprocedural verification of imperative programs. In International Conference
on Algebraic Methodology and Software Technology, pages 258–273, 2004.
[39] Michael Kuperstein, Martin T. Vechev, and Eran Yahav. Partial-coherence
abstractions for relaxed memory models. In ACM SIGPLAN Conference on
Programming Language Design and Implementation, pages 187–198, 2011.
[40] Markus Kusano and Chao Wang. Flow-sensitive composition of thread-modular
abstract interpretation. In ACM SIGSOFT Symposium on Foundations of Software
Engineering, 2016.
[41] Akash Lal and Thomas W. Reps. Reducing concurrent analysis under a context
bound to sequential analysis. Formal Methods in System Design, 35(1):73–97,
2009.
[42] Monica S. Lam, John Whaley, V. Benjamin Livshits, Michael C. Martin, Dzintars
Avots, Michael Carbin, and Christopher Unkel. Context-sensitive program
analysis as database queries. In ACM SIGACT-SIGMOD-SIGART Symposium on
Principles of Database Systems, pages 1–12, 2005.
[43] Leslie Lamport. How to make a multiprocessor computer that correctly
executes multiprocess programs. IEEE Transactions on Computers, 28(9), 1979.
[44] V. Benjamin Livshits and Monica S. Lam. Finding security vulnerabilities in Java
applications with static analysis. In USENIX Security Symposium, 2005.
[45] Yuri Meshman, Andrei Dan, Martin Vechev, and Eran Yahav. Synthesis of
memory fences via refinement propagation. In International Symposium on
Static Analysis, pages 237–252, 2014.
[46] Antoine Miné. Field-sensitive value analysis of embedded C programs with
union types and pointer arithmetics. In ACM SIGPLAN/SIGBED Conference on
Language, Compilers, and Tool Support for Embedded Systems, pages 54–63, 2006.
[47] Antoine Miné. Static analysis of run-time errors in embedded critical parallel C
programs. In Programming Languages and Systems, pages 398–418. 2011.
[48] Antoine Miné. Static analysis by abstract interpretation of sequential and
multi-thread programs. In Proc. of the 10th School of Modelling and Verifying
Parallel Processes, pages 35–48, 2012.
[49] Antoine Miné. Relational thread-modular static value analysis by abstract
interpretation. In International Conference on Verification, Model Checking, and
Abstract Interpretation, pages 39–58, 2014.
[50] Mayur Naik, Alex Aiken, and John Whaley. Effective static race detection for
Java. In ACM SIGPLAN Conference on Programming Language Design and
Implementation, pages 308–319, 2006.
[51] Brian Norris and Brian Demsky. CDSchecker: checking concurrent data
structures written with C/C++ atomics. In ACM SIGPLAN Conference on Object
Oriented Programming, Systems, Languages, and Applications, pages 131–150,
2013.
[52] Peizhao Ou and Brian Demsky. Checking concurrent data structures under the
C/C++11 memory model. In ACM SIGPLAN Symposium on Principles and
Practice of Parallel Programming, pages 45–59, 2017.
[53] Abhik Roychoudhury and Tulika Mitra. Specifying multithreaded java
semantics for program verification. In International Conference on Software
Engineering, pages 489–499, 2002.
[54] Mahmoud Said, Chao Wang, Zijiang Yang, and Karem Sakallah. Generating data
race witnesses by an SMT-based analysis. In NASA Formal Methods, pages
313–327, 2011.
[55] Peter Sewell, Susmit Sarkar, Scott Owens, Francesco Zappa Nardelli, and
Magnus O. Myreen. X86-TSO: A rigorous and usable programmer’s model for
x86 multiprocessors. Commun. ACM, 53(7):89–97, July 2010.
[56] Arnab Sinha, Sharad Malik, Chao Wang, and Aarti Gupta. Predicting
serializability violations: SMT-based search vs. DPOR-based search. In Haifa
Verification Conference, pages 95–114, 2011.
[57] Arnab Sinha, Sharad Malik, Chao Wang, and Aarti Gupta. Predictive analysis for
detecting serializability violations through trace segmentation. In International
Conference on Formal Methods and Models for Co-Design, pages 99–108, 2011.
[58] Richard Sites. Alpha Architecture Reference Manual. Digital Press, 1992.
[59] Chungha Sung, Markus Kusano, Nishant Sinha, and Chao Wang. Static DOM
event dependency analysis for testing web applications. In ACM SIGSOFT
Symposium on Foundations of Software Engineering, 2016.
[60] SVCOMP. International competition on software verification.
http://sv-comp.sosy-lab.org/2015/benchmarks.php, Accessed: 2015-05-06.
[61] TLDP. Interrupt handlers: Linux kernel module programming guide.
http://www.tldp.org/LDP/lkmpg/2.6/html/x1256.html, Accessed: 2015-05-06.
[62] Chao Wang and Malay Ganai. Predicting concurrency failures in generalized
traces of x86 executables. In International Conference on Runtime Verification,
pages 4–18, September 2011.
[63] Chao Wang, Sudipta Kundu, Malay Ganai, and Aarti Gupta. Symbolic predictive
analysis for concurrent programs. In International Symposium on Formal
Methods, pages 256–272, 2009.
[64] Chao Wang, Mahmoud Said, and Aarti Gupta. Coverage guided systematic
concurrency testing. In International Conference on Software Engineering, pages
221–230, 2011.
[65] David L Weaver and Tom Gremond. The SPARC architecture manual. PTR
Prentice Hall Englewood Cliffs, NJ 07632, 1994.
[66] John Whaley and Monica S. Lam. Cloning-based context-sensitive pointer alias
analysis using binary decision diagrams. In ACM SIGPLAN Conference on
Programming Language Design and Implementation, pages 131–144, 2004.
[67] Naling Zhang, Markus Kusano, and Chao Wang. Dynamic partial order
reduction for relaxed memory models. In ACM SIGPLAN Conference on
Programming Language Design and Implementation, pages 250–259, 2015.
| 6cs.PL
|
LQG Control and Sensing Co-design
arXiv:1802.08376v2 [math.OC] 21 Mar 2018
Vasileios Tzoumas,1,2 Luca Carlone,2 George J. Pappas,1 Ali Jadbabaie2
Abstract—Linear-Quadratic-Gaussian (LQG) control is concerned with the design of an optimal controller and estimator
for linear Gaussian systems with imperfect state information.
Standard LQG control assumes the set of sensor measurements
to be fed to the estimator to be given. However, in many
problems arising in networked systems and robotics, one may be
interested in designing a suitable set of sensors for LQG control.
In this paper, we introduce the LQG control and sensing co-design
problem, where one has to jointly design a suitable sensing,
estimation, and control policy. In particular, we consider two
dual instances of the co-design problem: the sensing-constrained
LQG control problem, where the design maximizes the control
performance subject to sensing constraints, and the minimumsensing LQG control, where the design minimizes the amount
of sensing subject to performance constraints. We focus on the
realistic case in which the sensing design has to be selected
among a finite set of possible sensing modalities, where each
modality is associated with a (possibly) different cost. While we
observe that the computation of the optimal sensing design is
intractable in general, we present the first scalable LQG codesign algorithms to compute near-optimal policies with provable
sub-optimality guarantees. To this end, (i) we show that a
separation principle holds, which partially decouples the design
of sensing, estimation, and control; (ii) we frame LQG codesign as the optimization of (approximately) supermodular
set functions; (iii) we develop novel algorithms to solve the
resulting optimization problems; (iv) we prove original results
on the performance of these algorithms and establish connections
between their suboptimality gap and control-theoretic quantities.
We conclude the paper by discussing two practical applications
of the co-design problem, namely, sensing-constrained formation
control and resource-constrained robot navigation.
I. I NTRODUCTION
Traditional approaches to the control of dynamical systems
with partially observable state assume the choice of sensors
used to observe the system to be given [1]. The choice of
sensors usually results from a preliminary design phase in
which an expert designer selects a suitable sensor suite that
accommodates estimation requirements (e.g., observability,
desired estimation error) and system constraints (e.g., size,
cost). However, modern control applications, from large networked systems to miniaturized robotics systems, pose serious
limitations to the applicability of this traditional paradigm: in
particular, in large-scale networked systems (e.g., smart grids,
or robot swarms), in which new nodes are continuously added
1 The authors are with the Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA (email:
{pappagsg, vtzoumas}@seas.upenn.edu).
2 The authors are with the Institute for Data, Systems and Society, and
the Laboratory for Information & Decision Systems, Massachusetts Institute of Technology, Cambridge, MA 02139, USA (email: {lcarlone,
jadbabai, vtzoumas}@mit.edu).
This work was partially supported by the AFOSR Complex Networks
Program, and by the Army Research Laboratory’s DCIST program.
and removed from the network, a manual re-design of the
sensors becomes cumbersome and expensive, and it is not scalable; in miniaturized robot systems, while the set of onboard
sensors is fixed, it may be desirable to selectively activate only
a subset of the sensors during different phases of operation, to
minimize power consumption. Overall, in both applications,
although a designer has access to a (possibly large) list of
potential sensors, due to resource constraints (size, weight,
power, cost) the designer can utilize only a subset of them.
Thus, the need for online and large-scale sensor selection
demands for automated approaches that efficiently select a
subset of sensors to maximize system performance.
Motivated by the aforementioned need, in this paper we
consider the problem of jointly designing control, estimation,
and sensing for systems with partially observable state.
Related work in control theory. Related work in control
theory focuses on either the co-design of estimation and
control in presence of communication constraints [1]–[7], or
on the design of the system’s sensing and actuation [8]–[19]
(sensor and actuator selection). In more detail:
a) LQG control design: The line of work [1]–[7] assumes the set of sensors and actuators to be given, and
either focuses on the co-design of estimation and control
over band-limited communication channels, or investigates the
trade-offs between communication constraints (e.g., data rate,
quantization, delays) and control performance (e.g., stability)
in networked control systems. These works provide results on
the impact of quantization [2], finite data rates [3], [4], as
well as, separation principles for LQG design with communication constraints [5]. More recent work focuses on privacy
constraints [6]. In addition, [7] studies rationally inattentive
control laws for LQG control and discusses their effectiveness
in stabilizing the system. We refer the reader to the surveys [1],
[20]–[22] for a comprehensive review on LQG control.
b) Sensor and actuator selection: The line of work [8]–
[19] focuses on selecting the system’s active sensors and actuators, independently of the control task at hand. In particular, [8]
proposes a sensor placement algorithm to maximize the accuracy of maximum likelihood estimation over static parameters,
whereas [9]–[13] focus on maximizing the estimation accuracy
for non-static parameters; [9]–[12] present sensor scheduling
algorithms for optimal Kalman filtering, while [13] presents
sensor scheduling algorithms for optimal batch state estimation
(smoothing); [14] considers fixed-lag smoothing and investigates sensor scheduling and feature selection for vision-based
agile navigation of autonomous robots. Finally, [15]–[19]
present sensor and actuator selection algorithms to optimize
the average observability and controllability of systems.
Extending the focus of the aforementioned works, more
recent work focuses on the co-design of control and estimation,
as well as, of sensing [23], [24], by augmenting the standard
LQG cost with an information-theoretic regularizer, and by
optimizing the sensing capabilities of each of the system’s
sensors using semi-definite programming. The main difference
between [23], [24], and our proposal in this paper is that
in [23], [24] the choice of sensors is arbitrary, rather than
being restricted to a finite set of available sensors.
Related work on set function optimization. The algorithms for sensor and actuator selection discussed above
employ either convex relaxation techniques [8]–[10], [12], [17]
or combinatorial optimization techniques [11], [13]–[16], [18].
The advantage of the combinatorial optimization techniques
is that they lead to algorithms with provable suboptimality
guarantees and low running time. The literature on combinatorial optimization, which is more relevant for the discussion in
this paper, includes investigation into (i) submodular optimization subject to cardinality constraints [25]; (ii) submodular
optimization subject to heterogeneous-cost constraints [26]–
[28]; and (iii) approximately submodular optimization subject
to cardinality constraints [29]. We note that the related literature does not cover the case of approximately submodular
optimization subject to heterogeneous-cost constraints, which
is indeed the setup of interest for our LQG control and
sensing co-design problems, hence requiring us to develop
novel algorithms and results for this case.
Contributions to control theory. We introduce the LQG
control and sensing co-design problem, that involves the
joint design of sensing, control, and estimation, by extending
Linear-Quadratic-Gaussian (LQG) control to the case where,
besides designing an optimal controller and estimator, one
has to choose a set of sensors to observe the system state.
We consider the realistic case in which the choice of sensors,
rather than being arbitrary (see, e.g., [23]), is restricted to a
finite selection from a set of available sensors. In particular,
in our formulation each available sensor is associated with a
cost that quantifies the penalty incurred when using that sensor
(trivially, if there is no cost associated to using a sensor, one
would always prefer to select and use all available sensors).
In more detail, we consider the general case in which each
sensor has a potentially different cost, hence capturing the
practical scenarios where each sensor may have a different
monetary cost, power consumption, or bandwidth utilization.
We formulate two dual instances of the LQG co-design
problem. The first instance, named sensing-constrained LQG
control, involves the joint design of the controller, estimator,
and sensing policies that minimize the LQG objective (quantifying tracking performance and control effort) while satisfying
a given constraint on the maximum cost of the selected
sensors. The second instance, named minimum-sensing LQG
control, involves the joint design of the controller, estimator,
and sensing that minimizes the cost of the selected sensors
while satisfying a given bound on the LQG performance.
We then leverage a separation principle1 to partially decouple the design of control, estimation, and sensing, and
we frame the sensor design subproblem as the optimization
of (approximately) supermodular set functions. While the
computation of the optimal sensing strategies is combinatorial
in nature, we provide the first scalable co-design algorithms,
which retrieve a near-optimal choice of sensors, as well as
the corresponding control and estimation policies. We show
that the suboptimality gaps of these algorithms depend on the
supermodularity ratio γf of the set function f appearing in
our problem, and we establish connections between the supermodularity ratio γf and control-theoretic quantities, providing
also a computable lower bound for γf .
Contributions to set function optimization. In proving the
aforementioned results, we extend the literature on supermodular optimization. In particular, (i) we provide the first efficient algorithms for the optimization of approximately supermodular functions subject to heterogeneous-cost constraints;
and (ii) we improve known suboptimality bounds that also
apply to the optimization of (exactly) supermodular functions:
specifically, the proposed algorithm for approximate supermodular optimization with heterogeneous-cost constraints can
achieve in the exactly supermodular case the approximation
bound (1−1/e), which is superior to the previously established
bound 1/2(1 − 1/e) in the literature [28].
Application examples. We motivate the importance of the
LQG control and sensing co-design problem, and demonstrate the effectiveness of the proposed algorithms in numerical experiments, by considering two application scenarios,
namely, a sensing-constrained formation control scenario and
a resource-constrained robot navigation scenario. In particular, we present a Monte Carlo analysis for both scenarios,
which demonstrates that (i) the proposed sensor selection
strategy is near-optimal, and in particular, the resulting LQGcost (tracking performance) matches the optimal selection in
all tested instances for which the optimal selection could
be computed via a brute-force approach; (ii) a more naive
selection which attempts to minimize the state estimation
covariance [11] (rather than the LQG cost) has degraded LQG
tracking performance, often comparable to a random selection;
and (iii) the selection of a small subset of sensors using the
proposed algorithms ensures an LQG cost that is close to the
one obtained by using all available sensors, hence providing an
effective alternative for control under sensing constraints [14].
Comparison with the preliminary results in [30]. This paper extends our preliminary results [30], and provides a more
comprehensive presentation of the LQG co-design problem,
including both sensing-constrained LQG control (introduced
in [30]) and the minimum-sensing LQG control problem (not
previously published). Moreover, we generalize the original
setup in [30] to account for heterogeneous sensor costs (in [30]
each sensor has unit cost) and extend the numerical analysis
accordingly. Most of the technical results, including Theo1 The separation principle leverages standard results in LQG control and
follows the line of [23], hence we do not claim it to be an original contribution.
rems 1–3, Proposition 1, as well as Algorithms 2–4 are novel,
and have not been previously published.
Organization of the rest of the paper. Section II formulates the LQG control and sensing co-design problems
considered in this paper. Section III presents a separation
principle and provides scalable, near-optimal algorithms for
the co-design problems. Section IV characterizes the running
time and approximation performance of the proposed algorithms, and establishes connections between their suboptimality bounds and control-theoretic quantities. Section V presents
two practical examples of co-design problems and provides
a numerical analysis of the proposed algorithms. Section VI
concludes the paper. All proofs are given in the Appendix.
Notation. Lowercase letters denote vectors and scalars (e.g.,
v), and uppercase letters denote matrices (e.g., M ). We use
calligraphic fonts to denote sets (e.g., S). The identity matrix
of size n is denoted with In (the dimension index is omitted
when it is clear from the context). For a matrix M and a vector
v of appropriate dimension, we define kvk2M , v T M v. For matrices M1 , M2 , . . . , Mk , we let diag (M1 , M2 , . . . , Mk ) be the
block diagonal matrix with diagonal blocks M1 , M2 , . . . , Mk .
II. LQG C ONTROL AND S ENSING C O - DESIGN :
P ROBLEM S TATEMENT
In this section we formalize the LQG control and sensing
co-design problem considered in this paper. In particular, we
present two “dual” statements of the co-design problem: the
sensing-constrained LQG control, where the design maximizes
the control performance subject to sensing constraints, and the
minimum-sensing LQG control, where the design minimizes
sensing subject to performance constraints.
A. System, sensors, and control policies
We start by introducing the notions of system, sensors, and
control policies. These notions are standard, except that only a
subset of sensors is actually used to observe the system’s state
(these are referred to as “active” sensors in Definition 1), and
that we associate a cost to each sensor (Definition 2).
a) System: We consider a standard discrete-time (possibly time-varying) linear system with additive Gaussian noise:
xt+1 = At xt + Bt ut + wt ,
t = 1, 2, . . . , T,
(1)
where xt ∈ Rn represents the state of the system at time t,
ut ∈ Rmt represents the control action, wt represents the
process noise, At and Bt are matrices of suitable dimensions,
and T is a finite horizon. In addition, we consider the system’s
initial condition x1 to be a Gaussian random variable with
covariance Σ1|0 , and wt to be a Gaussian random variable with
mean zero and covariance Wt , such that wt is independent of
x1 and wt′ for all t′ = 1, 2, . . . , T , t′ 6= t.
b) Sensors: We consider the case in which we have a
(potentially large) set of available sensors, which can take
noisy linear observations of the system’s state. In particular,
let V be a set of indices such that each index i ∈ V uniquely
identifies a sensor that can be used to observe the state of the
system. We consider sensors of the form
yi,t = Ci,t xt + vi,t ,
i ∈ V,
(2)
where yi,t ∈ Rpi,t represents the measurement of sensor i at
time t, Ci,t is a sensing matrix of suitable dimension, and
vi,t represents the measurement noise of sensor i. We assume
vi,t to be a Gaussian random variable with mean zero and
positive definite covariance Vi,t , such that vi,t is independent
of x1 , and of wt′ for any t′ 6= t, and independent of vi′ ,t′ for
all t′ 6= t, and any i′ ∈ V, i′ 6= i.
In this paper we are interested in the case in which we
cannot use all the available sensors and, as a result, we need
to select a convenient subset of sensors in the set V to meet
given specifications on the control performance (formalized in
Problem 1 and Problem 2 below).
Definition 1 (Active sensor set and measurement model).
Given a set of available sensors V, we say that S ⊆ V is an
active sensor set if we can observe the measurements from each
sensor i ∈ S for all t = 1, 2, . . . , T . Given an active sensor
set S = {i1 , i2 . . . , i|S| }, we define the following quantities
yt (S)
Ct (S)
Vt (S)
,
,
,
[yiT1 ,t , yiT2 ,t , . . . , yiT|S| ,t ]T,
[CiT1 ,t , CiT2 ,t , . . . , CiT|S| ,t ]T,
diag Vi1 ,t , Vi2 ,t , . . . , Vi|S| ,t ,
(3)
which lead to the definition of the measurement model:
yt (S) = Ct (S)xt + vt (S),
(4)
where vt (S) is a zero-mean Gaussian noise with covariance Vt (S). Despite the availability of a possibly large set
of sensors V, our observer will only have access to the
measurements produced by the active sensors.
In this paper we focus on the case where each sensor in the
set of available sensors V is associated with a (possibly different) cost, which captures, for instance, the sensor’s monetary
cost, its power consumption, or its bandwidth utilization.
Definition 2 (Cost of sensor and cost of active sensor set).
Given a set of available sensors V, we denote the cost of
sensor i ∈ V by c(i)≥ 0. Moreover, we denote the cost of an
active sensor set S ⊆ V by c(S) and set it equal to the sum
of the sensor costs c(i) for all active sensors i ∈ S:
X
c(S),
c(i).
(5)
i∈S
The following paragraph formalizes how the choice of the
active sensors affects the control policies.
c) Control policies: We consider control policies ut for
all t = 1, 2, . . . , T that are only informed by the measurements
collected by the active sensors:
ut = ut (S) = ut (y1 (S), y2 (S), . . . , yt (S)),
Such policies are called admissible.
t = 1, 2, . . . , T.
B. LQG co-design problems
The LQG co-design problem considered in this paper consists in the joint design of sensing, estimation, and control
strategies to meet given design specifications. We consider
two different types of specifications that lead to two codesign problems, named sensing-constrained LQG control
(Problem 1) and minimum-sensing LQG control (Problem 2).
Problem 1 (Sensing-constrained LQG control). Given a
system, a set of available sensors V, and a sensor budget b ≥
0, find a sensor set S ⊆ V to be active across all times
t = 1, 2, . . . , T , with cost c(S) at most b, and an admissible
control policy u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)} to
minimize the LQG cost function, that is:
min
S ⊆ V,
u1:T (S)
T
X
t=1
E kxt+1 (S)k2Qt +kut (S)k2Rt , s.t. c(S) ≤ b,
(6)
where the state-cost matrices Q1 , Q2 , . . . , QT are positive
semi-definite, the control-cost matrices R1 , R2 , . . . , RT are
positive definite, and the expectation is taken with respect to
the initial condition x1 , the process noises w1 , w2 , . . . , wT ,
and the measurement noises v1 (S), v2 (S), . . . , vT (S).
The sensing-constrained LQG control Problem 1 models
the practical case in which one cannot use all the available
sensors due to power, cost, or bandwidth constraints, and needs
to compute a suitable set of active sensors and controls that
maximize LQG performance. Note that if the budget constraint
is relaxed, all sensors are active (no penalty is incurred in using
all sensors) and Problem 1 reduces to standard LQG control.
While Problem 1 imposes constraints on the maximum
amount of sensing, the following “dual” problem formulation
imposes a constraint on the desired LQG performance.
Problem 2 (Minimum-sensing LQG control). Given a system, a set of available sensors V, and an upper bound κ > 0
for the LQG cost, find a minimum-cost sensor set S ⊆ V to
be active across all times t = 1, 2, . . . , T and an admissible
control policy u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)} such
that the LQG cost is at most κ:
min
S ⊆ V,
u1:T (S)
c(S), s.t.
T
X
t=1
E kxt+1 (S)k2Qt +kut (S)k2Rt ≤ κ,
(7)
where the state-cost matrices Q1 , Q2 , . . . , QT are positive
semi-definite, the control-cost matrices R1 , R2 , . . . , RT are
positive definite, and the expectation is taken with respect to
the initial condition x1 , the process noises w1 , w2 , . . . , wT ,
and the measurement noises v1 (S), v2 (S), . . . , vT (S).
The minimum-sensing LQG control Problem 2 models the
practical case in which one wants to design a system that
guarantees a desired level of performance, while incurring in
the smallest sensing cost (again the cost can be monetary or
connected to the use of limited resources).
Remark 1 (Case of uniform-cost sensors). When all sensors
i ∈ V have the same cost , say c(i) = c̄ > 0, the sensor budget
constraint can be rewritten as a cardinality constraint, since:
X
b
c(S) ≤ b ⇔
c(i) ≤ b ⇔ |S|c̄ ≤ b ⇔ |S|≤ , (8)
c̄
i∈S
which bounds the cardinality of the set of active sensors.
Similarly, under the uniform-cost assumption, the objective of
Problem 2 becomes the minimal cardinality objective |S|.
Problem 1 and Problem 2 generalize the imperfect stateinformation LQG control problem from the case where all
sensors in the set of available sensors V are active, and only
optimal control policies are to be found [1, Chapter 5], to the
case where only a few sensors in V can be active, and both
optimal sensors and control policies are to be found, jointly.
While we noticed that admissible control policies depend
on the active sensor set S, it is worth noticing that this in turn
implies that the state evolution will also depend on S, per the
system’s dynamics eq. (1); for this reason we write xt+1 (S) in
eqs. (6) and (7). Thereby, the intertwining between control and
sensing calls for a joint design strategy and, as a result, in the
following section we focus on the design of a jointly optimal
control and sensing solution to Problem 1 and Problem 2.
III. C O - DESIGN P RINCIPLES
AND
E FFICIENT A LGORITHMS
In this section we first present a separation principle that
decouples sensing, estimation, and control, and allows designing them in cascade (Section III-A). We then present
scalable algorithms for the sensing and control design in both
Problem 1 (Section III-B) and Problem 2 (Section III-C).
Theoretical guarantees bounding the suboptimality gap of the
proposed algorithms are given in Section IV.
A. Separability of optimal sensing and control design
We characterize the jointly optimal control and sensing
solutions for Problem 1 and Problem 2, and prove they can be
found in two separate steps, where first the sensing design is
computed, and second the control design is found (Theorem 1).
Theorem 1 (Separability of optimal sensing and control design). For any active sensor set S ⊆ V, let x̂t (S)
be the Kalman estimator of the state xt , i.e., x̂t (S) ,
E[xt |y1 (S), y2 (S), . . . , yt (S)], and Σt|t (S) be x̂t (S)’s error
covariance, i.e., Σt|t (S) , E[(x̂t (S) − xt )(x̂t (S) − xt )T ] [1,
Appendix E]. In addition, let the matrices Θt and Kt be the
solution of the following backward Riccati recursion
St
Nt
Mt
Kt
Θt
= Qt + Nt+1 ,
−1
−1 T −1
= AT
At ,
t (St + Bt Rt Bt )
T
= Bt St Bt + Rt ,
= −Mt−1 BtT St At ,
= KtT Mt Kt ,
(9)
with boundary condition NT +1 = 0 (notably, all matrices in
eq. (9) are independent of the active sensor set S).
1) (Separability in Problem 1) Let the sensor set S ⋆
and the controllers u⋆1 , u⋆2 , . . . , u⋆T be a solution to the
Algorithm 1 Joint Sensing and Control design for Problem 1.
Input: Time horizon T , available sensor set V, sensor selection budget b, covariance Σ1|0 of initial condition x1 ; for
all t = 1, 2, . . . , T , system matrix At , input matrix Bt ,
process noise covariance Wt , and LQG cost matrices Qt
and Rt ; for all sensors i ∈ V, measurement matrix Ci,t ,
measurement noise covariance Vi,t , and sensor cost c(i).
b and controls û1 , û2 , . . . , ûT .
Output: Active sensors S,
1: Compute the matrices Θ1 , Θ2 , . . . , ΘT using the backward
Riccati recursion in eq. (9).
b as the sensor set returned by Al2: Return the sensor set S
gorithm 2, which finds a (possibly approximate) solution
to the optimization problem in eq. (10);
3: Compute the matrices K1 , K2 , . . . , KT using the backward Riccati recursion in eq. (9).
4: At time t = 1, 2, . . . , T , compute the Kalman estimate of
the state xt , i.e., the estimate:
5:
b y2 (S),
b . . . , yt (S)];
b
x̂t , E[xt |y1 (S),
At time t = 1, 2, . . . , T , return the control ût = Kt x̂t .
sensing-constrained LQG Problem 1. Then, S ⋆ and u⋆1 ,
u⋆2 , . . . , u⋆T can be computed in cascade as follows:
S ⋆ ∈ arg min
S⊆V
T
X
t=1
⋆
u⋆t = Kt x̂t (S ),
tr[Θt Σt|t (S)], s.t. c(S) ≤ b, (10)
t = 1, . . . , T.
(11)
2) (Separability in Problem 2) Let the sensor set S ⋆ and the
controllers u⋆1 , u⋆2 , . . . , u⋆T be a solution to the minimumsensing LQG Problem
define the constant
2.PMoreover,
T
κ̄ , κ − tr Σ1|0 N1 − t=1 tr (Wt St ). Then, S ⋆ and
u⋆1 , u⋆2 , . . . , u⋆T can be computed in cascade as follows:
S ⋆ ∈ arg min c(S), s.t.
S⊆V
u⋆t = Kt x̂t (S ⋆ ),
T
X
tr[Θt Σt|t (S)] ≤ κ̄, (12)
t=1
t = 1, . . . , T.
(13)
Remark 2 (Certainty equivalence principle). The control
gain matrices K1 , K2 , . . . , KT are the same as the ones that
make the controllers (K1 x1 , K1 x2 , . . . , KT xT ) optimal for
the perfect state-information version of Problem 1, where the
state xt is known to the controllers [1, Chapter 4].
Theorem 1 decouples the design of the sensing from the
controller design. In particular, it suggests that once an optimal
sensor set S ⋆ is found, then the optimal controllers are equal
to Kt x̂t (S ⋆ ), which correspond to the standard LQG control
policy. This should not come as a surprise, since for a given
sensing strategy, Problem 1 reduces to standard LQG control.
Moreover, for a given sensor set, Problem 2 becomes a
feasibility problem and, as a result, admits multiple controls
that satisfy the LQG cost bound; one such set of controls are
the control actions computed in eq. (13), since they minimize
the LQG cost and, hence, they also belong to Problem 2’s
feasible set whenever Problem 2 admits a solution.
Algorithm 2 Sensing design for Problem 1.
Input: Time horizon T , available sensor set V, sensor selection budget b, covariance Σ1|0 of initial condition x1 , and
for all t = 1, 2, . . . , T and any sensor i ∈ V, matrix Θt ,
process noise covariance Wt , measurement matrix Ci,t ,
measurement noise covariance Vi,t , and sensor cost c(i).
b
Output: Sensor set S.
PT
b
1: S1 ← arg mini∈V,c(i)≤b
t=1 tr[Θt Σt|t ({i})];
b2 ← ∅; V ′ ← V;
2: S
b2 ) ≤ b do
3: while V ′ 6= ∅ and c(S
′
4:
for all a ∈ V do
5:
Sb2,α ← Sb2 ∪ {a}; Σ1|0 (Sb2,α ) ← Σ1|0 ;
6:
for all t = 1, . . . , T do
7:
Σt|t (Sb2,α ) ←
8:
[Σt|t−1 (Sb2,α )−1 +Ct (Sb2,α )T Vt (Sb2,α )−1 Ct (Sb2,α )]−1 ;
9:
Σt+1|t (Sb2,α ) ← At Σt|t (Sb2,α )AT
t + Wt ;
10:
end for P
T
11:
gaina ← t=1 tr{Θt [Σt|t (Sb2 ) − Σt|t (Sb2,α )]};
12:
end for
13:
s ← arg maxa∈V ′ [gaina /c(a)];
14:
Sb2 ← Sb2 ∪ {s};
15:
V ′ ← V ′ \ {s};
16: end while
b2 ) > b then
17: if c(S
b
18:
S2 ← Sb2 \ {s};
19: end if
PT
b ← arg min
20: S
t=1 tr[Θt Σt|t (S)].
S∈{Sb1 ,Sb2 }
We conclude the section with a remark providing an intuitive interpretation of the sensor design steps in eqs. (10)
and (12) for Problem 1 and Problem 2, respectively.
Remark 3 (Control-aware P
sensor design). In order to proT
vide insight on the function t=1 tr[Θt Σt|t (S)] appearing in
in eqs. (10) and (12), we rewrite it as:
T
X
tr[Θt Σt|t (S)] =
t=1
T
X
E tr{[xt − x̂t (S)]T Θt [xt − x̂t (S)]}
t=1
=
T
X
t=1
E kKt xt − Kt x̂t (S)k2Mt ,
(14)
where
in the first line we used
the fact that Σt|t (S) =
E (xt − x̂t (S))(xt − x̂t (S))T , and in the second line we
substituted the definition of Θt = KtT Mt Kt from eq. (9).
From eq. (14), it is clear that each term tr[Θt Σt|t (S)]
captures the expected mismatch between the imperfect stateinformation controller ut (S) = Kt x̂t (S) (which is only
aware of the measurements from the active sensors) and the
perfect state-information controller Kt xt . This is an important
distinction from the existing sensor selection literature. In particular, while standard sensor selection attempts to minimize
the estimation covariance, for instance by minimizing
T
X
t=1
tr[Σt|t (S)] ,
T
X
t=1
E kxt − x̂t (S)k22 ,
(15)
the proposed LQG cost formulation selectively minimizes the
estimation error focusing on the states that are most informative for control purposes. For instance, the contribution to the
total control mismatch in eq. (14) of all xt − x̂t (S) in the null
space of Kt is zero; accordingly, the proposed sensor design
approach has no incentive in activating sensors to observe
states which are irrelevant for control purposes. Overall, the
importance of a state for control purposes is indeed captured
by the weighting matrix Θt . Hence, in contrast to minimizing
the cost function in eq. (15), minimizing the cost function in
eq. (14) results in a control-aware sensing design.
B. Scalable near-optimal co-design algorithms
for sensing-constrained LQG control (Problem 1)
This section proposes a practical algorithm for the sensingconstrained LQG control Problem 1. The pseudo-code of the
algorithm is presented in Algorithm 1. Algorithm 1 follows the
result of Theorem 1 and jointly designs sensing and control by
first computing an active sensor set (Algorithm 1’s lines 1-2)
and then computing a control policy (Algorithm 1’s lines 3-5).
We discuss each step of the design below.
Near-optimal sensing design for Problem 1. Theorem 1
implies that an optimal sensor design for Problem 1 can be
computed by solving the optimization problem in eq. (10).
To this end, Algorithm 1 (line 1) first computes the matrices
Θ1 , Θ2 , . . . , ΘT , which appear in the objective function of
the optimization problem in eq. (10) and, as result, they
are necessary for its evaluation. Next, since the optimization
problem in eq. (10) is combinatorial in nature, because it
requires to select a subset of sensors out of all the available
sensors in V that has sensor cost at most b and induces the
smallest LQG cost, Algorithm 1’s line 2 proposes a greedy
algorithm, whose pseudo-code is given in Algorithm 2, to
compute a (possibly approximate) solution to the problem in
eq. (10). Our interest towards Algorithm 2 is motivated by that
it is scalable and provably close to the solution of the problem
in eq. (10) (in Section IV we quantify its running time and
provide sub-optimality bounds for its performance).
The steps that Algorithm 2 follows to compute a (possibly
approximate) solution to the problem in eq. (10) are as follows:
first, Algorithm 2 creates two candidate active sensor sets Sb1
and Sb2 (lines 1-2), of which only one will be selected as the
(possibly approximate) solution to the problem in eq. (10)
(line 20). In more detail, Algorithm 2’s line 1 lets the set Sb1
be composed of a single sensor, namely the sensor i ∈ V
that achieves the smallest value of the objective function in
eq. (10) and having cost not exceeding the sensor selection
budget (c(i) ≤ b). Then, Algorithm 2’s line 2 initializes the
candidate active sensor set Sb2 with the empty set, and after the
construction of the set Sb2 in Algorithm 2’s lines 3–19 (which
are explained below), Algorithm 2’s line 20 computes which of
Algorithm 3 Joint Sensing and Control design for Problem 2.
Input: Time horizon T , available sensor set V, LQG-cost
bound κ, covariance Σ1|0 of initial condition x1 ; for all
t = 1, 2, . . . , T , system matrix At , input matrix Bt , LQG
cost matrices Qt and Rt , process noise covariance Wt ;
and for all sensors i ∈ V, measurement matrix Ci,t ,
measurement noise covariance Vi,t , and sensor cost c(i).
b and controls û1 , û2 , . . . , ûT .
Output: Active sensors S,
1: Compute the matrices Θ1 , Θ2 , . . . , ΘT using the backward
Riccati recursion in eq. (9).
b as the sensor set returned by Al2: Return the sensor set S
gorithm 4, which finds a (possibly approximate) solution
to the optimization problem in eq. (7) ;
3: Compute the matrices K1 , K2 , . . . , KT using the backward Riccati recursion in eq. (9).
4: At time t = 1, 2, . . . , T , compute the Kalman estimate of
the state xt , i.e., the estimate:
5:
b y2 (S),
b . . . , yt (S)];
b
x̂t , E[xt |y1 (S),
At time t = 1, 2, . . . , T , return the control ût = Kt x̂t .
the two sets Sb1 and Sb2 achieves the smallest value for the objective function in eq. (10), and returns this set as the (possibly
approximate) solution to the optimization problem in eq. (10).
Lines 3–19 in Algorithm 2 populate the set Sb2 as follows:
at each iteration of the “while loop” (lines 3-16) a sensor
is greedily added to the set Sb2 , as long as Sb2 ’s sensor cost
does not exceed the sensor selection budget b. In particular,
for each available sensor (the set V ′ contains the available
sensors, excluding the ones already included in Sb2 ), the “for
loop” (lines 4-12) computes first the estimation covariance
resulting by adding the sensor to Sb2 , and second the corresponding marginal gain in the objective function in eq. (10)
(line 11). Then, the sensor that induces the largest sensorcost-normalized marginal gain is selected (line 13), and it is
added to the current set Sb2 (line 14). Finally, the “if” statement
(lines 17-19) ensures that the constructed set Sb2 has sensor
cost at most b, by possibly removing the sensor that was added
in Sb2 during the last iteration of the “while” loop in lines 3-16.
Control design for Problem 1. Theorem 1 implies that
given an active sensor set, the controls for Problem 1 can be
computed according to the eq. (11). To this end, Algorithm 1
first computes the matrices K1 , K2 , . . . , KT (line 3), and then,
at each time t = 1, 2, . . . , T , the Kalman estimate of the
current state xt (line 4), and the corresponding control (line 5).
C. Scalable near-optimal co-design algorithms
for minimum-sensing LQG control (Problem 2)
This section proposes a practical algorithm for the
minimum-sensing LQG control Problem 2. The pseudo-code
of the algorithm is presented in Algorithm 3. Algorithm 3
follows the result of Theorem 1 and jointly designs sensing and
control by first computing an active sensor set (Algorithm 3’s
lines 1-2) and then computing a control policy (Algorithm 3’s
lines 2-5). We discuss the first step (sensor design) in the rest
of this section, while the second step (control design) is as in
Algorithm 1’s line 2, and is explained in Section III-B.
Near-optimal sensing design for Problem 2. Theorem 1
implies that an optimal sensor design for Problem 2 can be
computed by solving the optimization problem in eq. (12).
To this end, similarly to Algorithm 1’s line 1, Algorithm 3’s
line 1 computes the matrices Θ1 , Θ2 , . . . , ΘT , which are
necessary for the evaluation of the cost function appearing in
eq. (12). Next, Algorithm 3’s line 2 calls Algorithm 4 to find a
(possibly approximate) solution to the optimization problem in
eq. (12). Analogously to the previous section, Algorithm 4 is
a greedy algorithm that returns a near-optimal solution for the
problem in eq. (12). The running time and the sub-optimality
bounds of the algorithm are analyzed in Section IV.
Algorithm 4 finds a (possibly approximate) solution to the
optimization problem in eq. (12) as follows: first, Algorithm 4
defines the constant κ̄ (line 1), appearing in the definition
of the optimization problem in eq. (12), and then initializes
the sensor set Sb with the empty set (line 2). Afterwards,
Algorithm 4 populates Sb in lines 3–16 using the following
steps: at each iteration of the “while loop” (lines 3-16) a sensor
b as long as Problem 2’s LQGis greedily added to the set S,
cost bound κ has not been met, which eq. (12) guarantees
to be equivalent to checking whether the second condition
in Algorithm 4’s line 3 holds. In particular, for each sensor
in V ′ (set of available sensors, excluding the ones already
b the “for loop” (lines 4-12) computes first the
included in S)
b and
estimation covariance resulting by adding the sensor to S,
then the corresponding marginal gain in the objective function
in eq. (10) (line 11). Then, the sensor that induces the largest
sensor-cost-normalized marginal gain is selected (line 13), and
added to the current candidate active set Sb (line 14). Finally,
the added sensor s is removed from V ′ (line 15).
In the following section we characterize the approximation
and running-time performance of Algorithm 1 and Algorithm 3
for Problem 1 and Problem 2, respectively.
IV. P ERFORMANCE
GUARANTEES FOR
LQG
CO - DESIGN
We prove that Algorithm 1 and Algorithm 3 are the first
scalable algorithms for the joint sensing and control design
Problem 1 and Problem 2, respectively, and that they achieve
an objective value that is close to the optimal. We start by
introducing the notion of supermodularity ratio (Section IV-A),
which will enable to bound the sub-optimality gap of Algorithm 1 (Section IV-B) and Algorithm 3 (Section III-C).
We then establish connections between the supermodularity
ratio and control-theoretic quantities (Section IV-D).
A. Supermodularity ratio of monotone functions
This section introduces the notion of supermodularity ratio
of a monotone set function (Definition 5). We start by defining
the notions of monotonicity (Definition 3) and of supermodularity (Definition 4).
Definition 3 (Monotonicity). Consider any finite set V. The
set function f : 2V 7→ R is non-increasing if and only if for
any sets A ⊆ B ⊆ V, it holds f (A) ≥ f (B).
Algorithm 4 Sensing design for Problem 2.
Input: Time horizon T , available sensor set V, LQG performance bound κ, covariance Σ1|0 of initial condition x1 ,
and for all t = 1, 2, . . . , T and any sensor i ∈ V, matrix
Θt , process noise covariance Wt , measurement matrix
Ci,t , measurement noise covariance Vi,t , and sensor cost
c(i).
b
Output: Sensor set S.
PT
1: κ̄ ← κ − tr Σ1|0 N1 −
t=1 tr (Wt St )
′
b
2: S ← ∅;
V ← V;P
T
b
3: while V ′ 6= ∅ and
t=1 tr[Θt Σt|t (S)] > κ̄ do
′
4:
for all a ∈ V do
5:
Sbα ← Sb ∪ {a}; Σ1|0 (Sbα ) ← Σ1|0 ;
6:
for all t = 1, . . . , T do
7:
Σt|t (Sbα ) ←
8:
[Σt|t−1 (Sbα )−1 + Ct (Sbα )T Vt (Sbα )−1 Ct (Sbα )]−1 ;
9:
Σt+1|t (Sbα ) ← At Σt|t (Sbα )AT
t + Wt ;
10:
end for P
b − Σt|t (Sbα )]};
11:
gaina ← Tt=1 tr{Θt [Σt|t (S)
12:
end for
13:
s ← arg maxa∈V ′ [gaina /c(a)];
14:
Sb ← Sb ∪ {s};
15:
V ′ ← V ′ \ {s};
16: end while
Definition 4 (Supermodularity [31, Proposition 2.1]). Consider any finite set V. The set function f : 2V 7→ R is
supermodular if and only if for any sets A ⊆ B ⊆ V, and any
element v ∈ V, it holds f (A)−f (A∪{v}) ≥ f (B)−f (B∪{v}).
In words, a set function f is supermodular if and only
if it satisfies the following diminishing returns property: for
any element v ∈ V, the marginal drop f (A) − f (A ∪ {v})
diminishes as the set A grows; equivalently, for any A ⊆ V
and v ∈ V, the drop f (A) − f (A ∪ {v}) is non-increasing.
Definition 5 (Supermodularity ratio). Consider any finite
set V, and a non-increasing set function f : 2V 7→ R. We
define the supermodularity ratio of f as
γf ,
min
A⊆B⊆V,v∈V\B
f (A) − f (A ∪ {v})
.
f (B) − f (B ∪ {v})
In words, the supermodularity ratio of a monotone set
function f measures how far f is from being supermodular.
In particular, as per Definition 5 of the supermodularity ratio,
the supermodularity ratio γf takes values in [0, 1], and
• γf = 1 if and only if f is supermodular, since if γf = 1,
then Definition 5 implies f (A) − f (A ∪ {v}) ≥ f (B) −
f (B ∪ {v}), i.e., the drop f (A) − f (A ∪ {v}) is nonincreasing as new elements are added in the set A.
• 0 <γf < 1 if and only if f is approximately supermodular, in the sense that if γf < 1, then Definition 5 implies
f (A) − f (A ∪ {v}) ≥ γf [f (B) − f (B ∪ {v})], i.e., the
drop f (A) − f (A ∪ {v}) is approximately non-increasing
as new elements are added in A; specifically, the ratio
B. Performance analysis for Algorithm 1
In this section we quantify Algorithm 1’s running time
and approximation performance (Theorem 2 below), using the
notion of supermodularity ratio introduced in Section IV-A.
Theorem 2 (Performance of Algorithm 1). For any active
sensor set S ⊆ V, and any admissible control policies
u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)}, let h[S, u1:T (S)] be
Problem 1’s cost function, i.e.,
PT
h[S, u1:T (S)] , t=1 E(kxt+1 (S)k2Qt +kut (S)k2Rt ).
Further define the following set-valued function and scalar:
g(S) , minu1:T (S) h[S, u1:T (S)],
(16)
g ⋆ , minS⊆V,u1:T (S) h[S, u1:T (S)], s.t. c(S) ≤ b;
that is, given a sensor set S ⊆ V, g(S) is the optimal value of
h[S, u1:T (S)] across all admissible control policies u1:T (S),
and g ⋆ is the optimal objective value of Problem 1.
The following results hold true:
1) (Approximation quality) Algorithm 1 returns an active
b at most b, and the
sensor set Sb ⊆ V having cost c(S)
b The
corresponding admissible control policies u1:T (S).
b
b
active sensors S and controls u1:T (S) are such that:
b u1:T (S)]
b
h[∅, u1:T (∅)] − h[S,
≥
⋆
h[∅, u1:T (∅)] − g
i (17)
hγ
b
g
1 − e−γg , 1 − e−γg c(S)/b ,
max
2
where γg is the supermodularity ratio of g(S) in eq. (16).
2) (Running time) Algorithm 1 runs in O(|V|2 T n2.4 ) time,
where n is the system size in eq. (1).
b u1:T (S)]
b quantifies
Note that the term h[∅, u1:T (∅)] − h[S,
b
the marginal gain of selecting the set S, and ineq. (17)
guarantees that the marginal gain is sufficiently large compared
to the optimal marginal gain h[∅, u1:T (∅)] − g ⋆, in the sense
that their ratio is lower bounded by the maximum between
b
γg
−γg
) and 1 − e−γg c(S)/b. We further comment on the
2 (1 − e
bound in ineq. (17) in the following proposition and remarks.
Proposition 1 (Extension of the bound in ineq. (17) to
sensor sets of any cost). Consider the modified version of
Algorithm 1 where Algorithm 2’s “if” statement (lines 17-19)
is removed. Then, Algorithm 1’s approximation performance
bound remains as in ineq. (17), even when Algorithm 1 returns
b that exceeds Problem 1’s budget b.
a set Sb of cost c(S)
Remark 4 (Comparison of bounds in ineq. (17)). In Fig. 1
we plot Algorithm 1’s approximation performance bounds in
ineq. (17), namely the bound γg /2 (1 − e−γg ) (function f1 (γg )
b
in Fig. 1) and the bound 1 − e−γg c(S)/b (functions f2 (γg ),
1
0.8
fi (γg ), i = 1, 2, 3, 4
γf captures how much ones needs to discount the drop
f (B) − f (B ∪ {v}), such that f (A) − f (A ∪ {v}) remains
greater then, or equal to, f (B) − f (B ∪ {v}).
We next use the above supermodularity ratio notion to quantify the sub-optimality gap of Algorithm 1 and Algorithm 3.
f1 (γg ) = γg /2 (1 − e−γg )
f2 (γg ) = 1 − e−2γg /5
f3 (γg ) = 1 − e−γg
f4 (γg ) = 1 − e−2γg
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
γg
Fig. 1. Plot of fi (γg ) (i = 1, 2, 3, 4) versus supermodularity ratio γg of
a monotone supermodular function g. By Definition 5 of supermodularity
ratio, γg takes values between 0 and 1. As γg increases from 0 to 1 then:
f1 (γg ) increases from 0 to 1/2(1 − e−1 ) ≃ 0.32; f3 (γg ) increases from 0
to 1 − e−2/5 ≃ 0.32; f2 (γg ) increases from 0 to 1 − e−1 ≃ 0.64; f4 (γg )
increases from 0 to 1 − e−2 ≃ 0.87.
b
f3 (γg ), and f4 (γg ) in Fig. 1, which correspond to c(S)/b
equal to 2/5, 1, and 2, respectively; we note that for the latter
b
case where c(S)/b
is equal to 2, we consider that Algorithm 1
has been modified per Proposition 1 to allow for active sensor
sets with costs that exceed the selection budget b). We make
two observations from Fig. 1: first, we observe that for ratio
b
b
values c(S)/b
> 2/5, the bound 1 − e−γg c(S)/b in ineq. (17)
dominates (i.e., is always larger —for all values of γg — than)
the bound γg /2 (1 − e−γg ) (compare plot of f2 (γg ) against
that of f1 (γg )). Also, we observe from Fig. 1 that as the cost
b
ratio c(S)/b
and the supermodularity ratio γg increase, the
b
bound 1 − e−γg c(S)/b tends to 1 (see plot of f4 (γg )).
Remark 5 (Novelty of Algorithm 2 and of bound in
ineq. (17)). Algorithm 2 (used as a subroutine in Algorithm 1)
is the first scalable algorithm with provable suboptimality
guarantees for the minimization of a (possibly) approximately
supermodular set function g, that is, a function g with supermodularity ratio γg (possibly) less than 1, subject to a
heterogeneous-cost constraint. This generalizes existing algorithms for optimization with heterogeneous-cost constraints,
which only focus on the special case of (exactly) supermodular
functions (see, e.g., [28]), that is, functions g with supermodularity ratio γg (exactly) equal to 1.
In addition, Algorithm 2 offers a tighter approximation performance bound for the optimization of (exactly) supermodular
functions. Specifically, although the previous algorithms for
the optimization of supermodular functions (see, e.g., [28])
have the same running time as Algorithm 2 and
achieve the
approximation performance bound 1/2 1 − e−1 , which is the
same as Algorithm 2’s performance bound γg /2 (1 − e−γg )
for γg = 1 (that is, for supermodular functions), Algorithm 2
b
also achieves the cost-dependent bound 1 − e−γg c(S)/b, which
for γg = 1 is superior to 1/2 1 − e−1 when the cost ratio
b
c(S)/b
is more than 2/5 (Remark 4).
Theorem 2 guarantees that Algorithm 1 achieves an objective value for Problem 1 that is finitely close to optimal,
whenever the supermodularity ratio γg is non-zero. In more
detail, the extreme values of the bound in ineq. (17), as
well as their interpretation with respect to Algorithm 1’s
approximation performance are as follows: the maximum
value of the bounds in ineq. (17) is 1, which is achieved for
b
supermodularity ratio γg = 1 and ratio c(S)/b
−→ +∞;
as discussed in Proposition 1, the latter is possible when
Algorithm 1 is modified to return an active sensor set with
cost larger than b. On the other hand, when Algorithm 1 is not
modified to return an active sensor set with cost larger than the
budget b, and it always returns a sensor set with cost at most
b, then the maximum value the bound in ineq. (17) can take
b = b); notably, this is the best
is 1 − 1/e (for γg = 1 and c(S)
bound one can achieve in the worst-case in polynomial time
even for supermodular objective functions [32]. The minimum
value the bound in ineq. (17) is 0, which occurs for γg = 0.
The interpretation of the extreme values 0 and 1 of the
bound in ineq. (17) is as follows: when the bound in ineq. (17)
takes the value 1, then ineq. (17) implies that the approximate
b u1:T (S)]
b to Problem 1 is equal to the (optimal)
value h[S,
value g ⋆ of Problem 1, that is, Algorithm 1 is exact. Finally,
when the bound in ineq. (17) is 0, ineq. (17) implies that
b u1:T (S)]
b ≤ h[∅, u1:T (∅)], which is trivially satisfied2
h[S,
and, as a result, it is uninformative on the approximation
performance of Algorithm 1. In Section IV-D we present
conditions under which the supermodularity ratio in ineq. (17)
is guaranteed to be non-zero, in which case Algorithm 1
achieves near-optimal approximation performance.
Theorem 2 also ensures that Algorithm 1 is the first scalable
algorithm for Problem 1. In particular, Algorithm 1’s running
time O(|V|2 T n2.4 ) is in the worst-case quadratic in the
number of the available sensors |V| (in the case where all the
sensors in V are chosen as active) and linear in the Kalman
filter’s running time across the time horizon {1, 2, . . . , T };
specifically, the contribution n2.4 T in Algorithm 1’s running
time comes from the computational complexity of using the
Kalman filter to compute the state estimation error covariances
Σt|t for each t = 1, 2, . . . , T [1, Appendix E].
C. Performance analysis for Algorithm 3
We quantify Algorithm 3’s running time and approximation
performance (Theorem 3 below), using the notion of supermodularity ratio introduced in Section IV-A.
Theorem 3 (Performance of Algorithm 3). Consider the
notation introduced in the statement of Theorem 2 (Section IV-B): for any active sensor set S ⊆ V, and any admisb u1:T (S)]
b ≤ h[∅, u1:T (∅)] simply states that a control
inequality h[S,
policy that is informed by the active sensor set Sb has better performance than
a policy that does not use any sensor; for a more formal proof we refer the
reader to Appendix B.
2 The
sible control policies u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)},
let h[S, u1:T (S)] be the LQG cost function in Problem 2’s
constraint, i.e.,
P
h[S, u1:T (S)] , Tt=1 E(kxt+1 (S)k2Qt +kut(S)k2Rt );
Further define the following set-valued function:
g(S) , minu1:T (S) h[S, u1:T (S)];
(18)
that is, given a sensor set S ⊆ V, g(S) is the optimal value of
h[S, u1:T (S)] across all admissible control policies u1:T (S).
Finally, let b⋆ be the optimal value of Problem 2, namely:
b⋆ ,
min
S⊆V,u1:T (S)
c(S), s.t. h[S, u1:T (S)] ≤ κ.
The following results hold true:
1) (Approximation quality) Algorithm 3 returns an active
sensor set Sb ⊂ V and admissible control policies
b Let sl denote the last sensor added to Sb by
u1:T (S).
b
Algorithm 3: the active sensors Sb and controls u1:T (S)
are such that:
b u1:T (S)]
b ≤ κ;
h[S,
b ≤ c(sl ) +
c(S)
1
log
γg
h[∅, u1:T (∅)] − κ
h[Sbl−1 , u1:T (Sbl−1 )] − κ
!
(19)
b⋆,
(20)
where Sbl−1 is the subset of Sb that results by removing
from Sb the last sensor added to it by Algorithm 3, i.e.,
Sbl−1 , Sb \ {sl }; γg is the supermodularity ratio of g(S).
2) (Running time) Algorithm 3 runs in O(|V|2 T n2.4 ) time,
where n is the maximum system size in eq. (1).
Remark 6 (Novelty of Algorithm 4 and of bound in
ineq. (20)). Algorithm 4 (used as a subroutine in Algorithm 3)
is the first scalable algorithm with provable suboptimality
guarantees for the minimum heterogeneous-cost set selection
subject to a constraint on a (possibly) approximately supermodular function g, that is, a function g with supermodularity
ratio γg less than 1. In particular, it generalizes previous
algorithms that only focus on the special case of (exactly)
supermodular functions (see, e.g., [25]), that is, functions g
with supermodularity ratio γg equal to 1. Notably, for the case
where the supermodularity ratio γg is equal to 1 (that is, the
set function g is supermodular), and the sensor cost of the last
sensor added to the returned set Sb by Algorithm 3 is equal to 1
(c(sl ) = 1), then the bound in ineq. (20) becomes the same as
the known bound in the supermodular function optimization
literature for minimum cost set-selection [25, Theorem 1].
Theorem 3, with ineq. (19), implies that Algorithm 3 returns
a (possibly approximate) solution to Problem 2 that guarantees
that the LQG-cost constraint in Problem 2 is satisfied.
Ineq. (20) also guarantees that for non-zero supermodularity
ratio γg Algorithm 3 achieves an objective value for Problem 2
that is finitely close to the optimal, since for non-zero γg the
sensor cost of the set returned by Algorithm 3 is up to a finite
multiplicative factor away from the optimal sensor set cost b⋆.
In addition, ineq. (20) suggests that the approximation bound
increases as the LQG-cost performance bound parameter κ decreases, that is, as we require from Algorithm 3 to find a sensor
set that achieves a better (lower) LQG-cost performance.
Theorem 3 also ensures that Algorithm 3 is the first scalable
algorithm for Problem 2. Notably, Algorithm 3’s running time
is equal in the worst-case to the running time of Algorithm 1
(which we discussed in Section IV-B).
In the following section we present control-theoretic conditions under which the supermodularity ratio γg in both
Algorithm 1’s and Algorithm 3’s approximation bound in
ineqs. (17) and (19) is non-zero, in which case Algorithm 1 and
Algorithm 3 achieve near-optimal approximation performance.
D. Conditions for non-zero supermodularity ratio
In this section we provide conditions such that the supermodularity ratio γg in ineqs. (17) and (19) is non-zero, in
which case both Algorithm 1 and Algorithm 3 guarantee a
close to optimal approximate performance (per Theorem 2
and Theorem 3, respectively).
PT In particular, we first prove
that if the strict inequality t=1 Θt ≻ 0 holds, where each
Θt is defined as in eq. (9), then the supermodularity ratio γg
is
PTnon-zero (Theorem 4). Then, we prove that the condition
t=1 Θt ≻ 0 holds true in all LQG control problem instances
where a zero controller would result in
PaT suboptimal behavior
for the system; that is, we prove that t=1 Θt ≻ 0 holds true
in all system instances where LQG control design is necessary
to achieve a desired system performance (Theorem 5).
The next theorem provides a non-zero computable bound for
the supermodularity ratio γg in Theorem 2 and in Theorem 3.
in ineq. (21) increases when on average all the directions in
the state space become equally important for control purposes.
Indeed, in the extreme case where λmax (Θt ) = λmin (Θt ) = λ,
the cost function in eq. (10) that Algorithm 1 minimizes to
select the active sensor set becomes
T
X
t=1
tr[Θt Σt|t (S)] = λ
T
X
tr[Σt|t (S)],
t=1
which matches the cost function in the classical sensor selection where all states are equally important (per eq. (15)).
b) Case 2 where the bound of γg in ineq. (21) increases:
When either the numerators of the last two fractions in the
right-hand-side of ineq. (21) increase or the denominators of
the last two fractions in the right-hand-side of ineq. (21) decrease, then the right-hand-side in ineq. (21) increases. In particular, the numerators of the last two fractions in the righthand-side of ineq. (21) capture the (best) estimation quality
when all available sensors in V are used, via the terms of the
T
form λmin [Σt|t (V)] and λmin [C̄i,t Σt|t (V)C̄i,t
]. Interestingly,
this suggests that the right-hand-side of ineq. (21) increases
when the available sensors in V are inefficient in achieving
low estimation error, that is, when the terms of the form
T
λmin [Σt|t (V)] and λmin [C̄i,t Σt|t (V)C̄i,t
] increase. Similarly,
the denominators of the last two fractions in the right-handside of ineq. (21) capture the (worst) estimation quality when
no sensor is active, via the terms of the form λmax [Σt|t (∅)] and
T
λmax [C̄i,t Σt|t (∅)C̄i,t
]. This suggests that the right-hand-side of
ineq. (21) increases when the measurement noise increases.
Theorem 4 states that the supermodularity ratio γg is nonPT
zero whenever t=1 Θt ≻ 0. To provide insight on the type of
control problems for which this result holds, inP
the following
T
theorem we translate the technical condition t=1 Θt ≻ 0
into an equivalent control-theoretic condition.
Theorem 4 (Non-zero computable bound for the supermodularity ratio γg ). Let the matrices Θt for all t = 1, 2, . . . , T
be defined as in eq. (9), the set function g(S) be defined
as in eq. (16), and for any sensor i ∈ V, the matrix
−1/2
Theorem 5 (Control-theoretic condition for near-optimal
C̄i,t , Vi,t Ci,t be the whitened measurement matrix.
PT
co-design). Consider the (noiseless, perfect state-information)
If the strict inequality
t=1 Θt ≻ 0 holds, then the
LQG problem where at any time t = 1, 2, . . . , T , the state xt
supermodularity ratio γg is non-zero. In addition, if we assume
is known to each controller ut and the process noise wt is
(for simplicity in presentation) that the Frobenius norm of
zero, i.e., the optimal control problem
T
= 1, and that tr[Σt|t (∅)] ≤
each C̄i,t is 1, i.e., tr C̄i,t C̄i,t
P
λ2max [Σt|t (∅)], then γg ’s lower bound is
minu1:T Tt=1 [kxt+1 k2Qt +kut (xt )k2Rt ] Σ =W =0 .
(22)
t
t|t
PT
λmin ( t=1 Θt ) mint∈{1,2,...,T } λ2min [Σt|t (V)]
Let At be
γg ≥
PT
PTinvertible for all t = 1, 2, . . . , T ; the strict
λmax ( t=1 Θt ) maxt∈{1,2,...,T } λ2max [Σt|t (∅)]
inequality
t=1 Θt ≻ 0 holds if and only if for all non(21)
T
zero
initial
conditions
x1 , the all-zeroes control policy u◦1:T ,
1 + mini∈V,t∈{1,2...,T } λmin [C̄i Σt|t (V)C̄i ]
.
(0, 0, . . . , 0) is not an optimal solution to the optimal control
2 + maxi∈V,t∈{1,2...,T } λmax [C̄i Σt|t (∅)C̄iT ]
problem in eq. (22):
Ineq. (21) suggests two cases under which γg can increase,
PT
◦
/ arg minu1:T t=1 [kxt+1 k2Qt +kut (xt )k2Rt ] Σ =W =0 .
and, correspondingly, the approximation performance bounds u1:T ∈
t
t|t
of Algorithm 1 and of Algorithm 3 in ineqs. (17) and (19),
PT
respectively, can improve; in more detail:
Theorem 5 suggests that the condition
t=1 Θt ≻ 0
a) Case 1 where the bound of γg in ineq. (21) increases: (which ensures a non-zero supermodularity ratio γg per ThePT
PT
When the fraction λmin ( t=1 Θt )/λmax ( t=1 Θt ) increases orem 4) holds if and only if for any non-zero initial conto 1, then the right-hand-side in ineq. (21) increases. Therefore, dition x1 the all-zeroes control policy u◦1:T = (0, 0, . . . , 0)
since the matrices Θt weight the states depending on their is suboptimal for the noiseless, perfect state-information
relevance for control purposes (Remark 3), the right-hand-side LQG problem in eq. (22); intuitively, this encompasses
9
8
7
6
5
4
3
y [meters]
2
1
0
-1
-2
-3
-4
-5
-6
-7
-8
-9
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
x [meters]
(a) formation control
(b) unmanned aerial robot
Fig. 2. Examples of applications of the proposed sensing-constrained LQG
control framework: (a) sensing-constrained formation control and (b) resourceconstrained robot navigation.
most practical control design problems where a zero controller would result in a suboptimal behavior of the system
(LQG control design itself would be unnecessary in the
case where a zero controller, i.e., no control action, can
already attain the desired system performance).
Overall, Algorithm 1 and Algorithm 3 are the first scalable
algorithms for Problem 1 and for Problem 2, respectively, and
(for the LQG control problem instances where a zero controller
would result in a suboptimal behavior for the system and, as
a result, for the system instances where LQG control design
is necessary to achieve a desired system performance) they
achieve close to optimal approximate performance.
V. N UMERICAL E XPERIMENTS
We consider two application scenarios for the proposed sensing-constrained LQG control framework: sensingconstrained formation control and resource-constrained robot
navigation. We present a Monte Carlo analysis for both scenarios, which demonstrates that (i) the proposed sensor selection
strategy is near-optimal, and in particular, the resulting LQGcost (tracking performance) matches the optimal selection in
all tested instances for which the optimal selection could
be computed via a brute-force approach, (ii) a more naive
selection which attempts to minimize the state estimation
covariance [11] (rather than the LQG cost) has degraded LQG
tracking performance, often comparable to a random selection,
(iii) in the considered instances, a clever selection of a small
subset of sensors can ensure an LQG cost that is close to the
one obtained by using all available sensors, hence providing an
effective alternative for control under sensing constraints [14].
A. Sensing-constrained formation control
Simulation setup. The first application scenario is illustrated in Fig. 2(a). A team of n agents (blue triangles) moves
in a 2D scenario. At time t = 1, the agents are randomly
deployed in a 10m × 10m square and their objective is to
reach a target formation shape (red stars); in the example of
Fig. 2(a) the desired formation has an hexagonal shape, while
in general for a formation of n, the desired formation is an
equilateral polygon with n vertices. Each robot is modeled as
a double-integrator, with state xi = [pi vi ]T ∈ R4 (pi is the 2D
position of agent i, while vi is its velocity), and can control
its own acceleration ui ∈ R2 ; the process noise is chosen
as a diagonal matrix W = diag [1e−2 , 1e−2 , 1e−4 , 1e−4 ] .
Each robot i is equipped with a GPS receiver, which can
measure the agent position pi with a covariance Vgps,i = 2·I2 .
Moreover, the agents are equipped with lidar sensors allowing
each agent i to measure the relative position of another agent j
with covariance Vlidar,ij = 0.1 · I2 . The agents have very
limited on-board resources, hence they can only activate a
subset of k sensors. Hence, the goal is to select the subset of
k sensors, as well as to compute the control policy that ensure
best tracking performance, as measured by the LQG objective.
For our tests, we consider two problem setups. In the first
setup, named homogeneous formation control, the LQG weigh
matrix Q is a block diagonal matrix with 4 × 4 blocks, with
each block i chosen as Qi = 0.1 · I4 ; since each 4 × 4 block
of Q weights the tracking error of a robot, in the homogeneous
case the tracking error of all agents is equally important.
In the second setup, named heterogeneous formation control,
the matrix Q is chose as above, except for one of the agents,
say robot 1, for which we choose Q1 = 10 · I4 ; this setup
models the case in which each agent has a different role or
importance, hence one weights differently the tracking error
of the agents. In both cases the matrix R is chosen to be the
identity matrix. The simulation is carried on over T time steps,
and T is also chosen as LQG horizon. Results are averaged
over 100 Monte Carlo runs: at each run we randomize the
initial estimation covariance Σ1|0 .
Compared techniques. We compare five techniques. All
techniques use an LQG-based estimator and controller, and
they only differ by the selections of the sensors used.
The first approach is the optimal sensor selection, denoted as
optimal, which attains the minimum of the cost function in
eq. (10), and that we compute by enumerating all possible
subsets; this brute-force approach is only viable when the
number of available sensors is small. The second approach
is a pseudo-random sensor selection, denoted as random∗ ,
which selects all the GPS measurements and a random subset
of the lidar measurements; note that we do not consider a
fully random selection since in practice this often leads to an
unobservable system, hence causing divergence of the LQG
cost. The third approach, denoted as logdet, selects sensors
so to minimize the average log det of the estimation covariance
over the horizon; this approach resembles [11] and is agnostic
to the control task. The fourth approach is the proposed sensor
selection strategy, described in Algorithm 2, and is denoted as
s-LQG. Finally, we also report the LQG performance when all
sensors are selected; this is clearly infeasible in practice, due to
the sensing constraints, and it is only reported for comparison
purposes. This approach is denoted as allSensors.
Results. The results of our numerical analysis are reported
in Fig. 3. When not specified otherwise, we consider a
formation of n = 4 agents, which can only use a total of
k = 6 sensors, and a control horizon T = 20. Fig. 3(a)
shows the LQG cost attained by the compared techniques for
increasing control horizon and for the homogeneous case. We
B. Resource-constrained robot navigation
Simulation setup. The second application scenario is illustrated in Fig. 2(b). An unmanned aerial robot (UAV) moves
in a 3D scenario, starting from a randomly selected initial
12
250
random*
optimal
logdet
s-LQG
allSensors
8
random*
optimal
logdet
s-LQG
allSensors
200
LQG cost
LQG cost
10
6
150
100
4
2
50
10
15
20
25
30
10
15
20
30
(b) heterogeneous
(a) homogeneous
16
300
12
10
random*
optimal
logdet
s-LQG
allSensors
250
LQG cost
random*
optimal
logdet
s-LQG
allSensors
14
LQG cost
25
horizon
horizon
200
150
8
100
6
4
50
4
5
6
7
8
9
10
4
5
6
7
8
9
10
maxNrUsedSensors
maxNrUsedSensors
(d) heterogeneous
(c) homogeneous
25
180
random*
optimal
logdet
s-LQG
allSensors
15
random*
optimal
logdet
s-LQG
allSensors
160
140
LQG cost
20
LQG cost
note that, in all tested instance, the proposed approach s-LQG
matches the optimal selection optimal, and both approaches
are relatively close to allSensors, which selects all the
2
available sensors ( n+n
2 ). On the other hand logdet leads
to worse tracking performance, and it is often close to the
pseudo-random selection random∗ . These considerations are
confirmed by the heterogeneous setup, shown in Fig. 3(b).
In this case the separation between the proposed approach
and logdet becomes even larger; the intuition here is that
the heterogeneous case rewards differently the tracking errors
at different agents, hence while logdet attempts to equally
reduce the estimation error across the formation, the proposed
approach s-LQG selects sensors in a task-oriented fashion,
since the matrices Θt for all t = 1, 2, . . . , T in the cost
function in eq. (10) incorporate the LQG weight matrices.
Fig. 3(c) shows the LQG cost attained by the compared
techniques for increasing number of selected sensors k and for
the homogeneous case. We note that for increasing number of
sensors all techniques converge to allSensors (the entire
ground set is selected). As in the previous case, the proposed
approach s-LQG matches the optimal selection optimal.
Interestingly, while the performance of logdet is in general
inferior with respect to s-LQG, when the number of selected
sensors k decreases (for k < n the problem becomes unobservable) the approach logdet performs similarly to s-LQG.
Fig. 3(d) shows the same statistics for the heterogeneous case.
We note that in this case logdet is inferior to s-LQG even
in the case with small k. Moreover, an interesting fact is that
s-LQG matches allSensors already for k = 7, meaning
that the LQG performance of the sensing-constraint setup is
indistinguishable from the one using all sensors; intuitively,
in the heterogeneous case, adding more sensors may have
marginal impact on the LQG cost (e.g., if the cost rewards
a small tracking error for robot 1, it may be of little value
to take a lidar measurement between robot 3 and 4). This
further stresses the importance of the proposed framework as a
parsimonious way to control a system with minimal resources.
Fig. 3(e) and Fig. 3(f) show the LQG cost attained by the
compared techniques for increasing number of agents, in the
homogeneous and heterogeneous case, respectively. To ensure
observability, we consider k = round (1.5n), i.e., we select a
number of sensors 50% larger than the smallest set of sensors
that can make the system observable. We note that optimal
quickly becomes intractable to compute, hence we omit values
beyond n = 4. In both figures, the main observation is that
the separation among the techniques increases with the number
of agents, since the set of available sensors quickly increases
with n. Interestingly, in the heterogeneous case s-LQG remains relatively close to allSensors, implying that for the
purpose of LQG control, using a cleverly selected small subset
of sensors still ensures excellent tracking performance.
10
120
100
80
5
0
60
40
3
5
7
nrRobots
(e) homogeneous
9
11
3
5
7
9
11
nrRobots
(f) heterogeneous
Fig. 3. LQG cost for increasing (a)-(b) control horizon T , (c)-(d) number of
selected sensors k, and (e)-(f) number of agents n. Statistics are reported for
the homogeneous formation control setup (left column), and the heterogeneous
setup (right column). Results are averaged over 100 Monte Carlo runs.
location. The objective of the UAV is to land, and more
specifically, it has to reach the position [0, 0, 0] with zero
velocity. The UAV is modeled as a double-integrator, with
state xi = [pi vi ]T ∈ R6 (pi is the 3D position of agent i,
while vi is its velocity), and can control its own acceleration
ui ∈ R3 ; the process noise is chosen as W = I6 . The UAV
is equipped with multiple sensors. It has an on-board GPS
receiver, measuring the UAV position pi with a covariance
2 · I3 , and an altimeter, measuring only the last component
of pi (altitude) with standard deviation 0.5m. Moreover, the
UAV can use a stereo camera to measure the relative position
of ℓ landmarks on the ground; for the sake of the numerical
example, we assume the location of each landmark to be
known only approximately, and we associate to each landmark
an uncertainty covariance (red ellipsoids in Fig. 2(b)), which is
randomly generated at the beginning of each run. The UAV has
limited on-board resources, hence it can only activate a subset
of k sensors. For instance, the resource-constraints may be due
to the power consumption of the GPS and the altimeter, or
may be due to computational constraints that prevent to run
multiple object-detection algorithms to detect all landmarks
2500
120
LQG cost
LQG cost / T
2450
random*
optimal
logdet
s-LQG
allSensors
110
100
2400
2350
2300
90
10
20
30
40
horizon
(a) heterogeneous
50
2250
intractable in general, we provide the first scalable algorithms
that can compute a design that is provably close to the optimal
one. While developing these algorithms, we also extend the
literature on supermodular optimization, by providing the
first efficient algorithms for the optimization of (approximately) supermodular functions subject to heterogeneous-cost
constraints, and by improving existing performance bounds.
12 Notably, our performance bounds have a clear connection to
control-theoretic quantities and are proven to be non-vanishing
under very general conditions (we prove that the suboptimality
gap is non-vanishing whenever the open loop behavior of the
system deviates from the desired closed-loop behavior, hence
encompassing most real-world control problems). Finally, we
provided illustrative examples and a numerical analysis considering problems in robotics and networked control.
This paper opens a number of avenues for future research.
First, while this paper provides an introduction to LQG sensing
and control co-design, other interesting co-design problems
exist; for instance, one may consider actuation-and-control codesign problems, or even sensing-actuation-control co-design.
Second, one may extend the LQG co-design problem to account for potential sensor failures, where some of the selected
sensors do not work as expected; to this end, one could leverage recent results on resilient submodular optimization [33].
Finally, while we currently provide bounds between our sensor
design and the optimal sensor design, we find interesting to
provide bounds that compare the LQG performance attained
when an optimal subset of sensors is used with the LQG
performance attained when all available sensors is used.
random*
optimal
logdet
s-LQG
allSensors
130
4
6
8
10
maxNrUsedSensors
(b) heterogeneous
Fig. 4. LQG cost for increasing (a) control horizon T , and (b) number of
selected sensors k. Statistics are reported for the heterogeneous setup. Results
are averaged over 100 Monte Carlo runs.
on the ground. Similarly to the previous case, we phrase the
problem as a sensing-constraint LQG problem,
and we use
Q = diag [1e−3 , 1e−3 , 10, 1e−3 , 1e−3 , 10] and R = I3 .
Note that the structure of Q reflects the fact that during
landing we are particularly interested in controlling the vertical
direction and the vertical velocity (entries with larger weight
in Q), while we are less interested in controlling accurately the
horizontal position and velocity (assuming a sufficiently large
landing site). In the following, we present results averaged
over 100 Monte Carlo runs: in each run, we randomize the
covariances describing the landmark position uncertainty.
Compared techniques. We consider the five techniques
discussed in the previous section. As in the formation control
case, the pseudo-random selection random∗ always includes
the GPS measurement (which alone ensures observability) and
a random selection of the other available sensors.
Results. The results of our numerical analysis are reported
in Fig. 4. When not specified otherwise, we consider a total of
k = 3 sensors to be selected, and a control horizon T = 20.
Fig. 4(a) shows the LQG cost attained by the compared
techniques for increasing control horizon. For visualization
purposes we plot the cost normalized by the horizon, which
makes more visible the differences among the techniques. Similarly to the formation control example, s-LQG matches the
optimal selection optimal, while logdet and random∗
have suboptimal performance.
Fig. 4(b) shows the LQG cost attained by the compared
techniques for increasing number of selected sensors k.
Clearly, all techniques converge to allSensors for increasing k, but in the regime in which few sensors are used s-LQG
still outperforms alternative sensor selection schemes, and
matches in all cases the optimal selection optimal.
VI. C ONCLUDING R EMARKS
In this paper, we introduced the LQG control and sensing co-design problem, where one has to jointly design a
suitable sensing, estimation, and control policy to steer the
behavior of a linear Gaussian systems under resource constraints. We discussed two variants of the problem, named
sensing-constrained LQG control and minimum-sensing LQG
control, which are central in modern control applications
ranging from large-scale networked systems to miniaturized
robotics networks. While the resulting co-design problems are
R EFERENCES
[1] D. P. Bertsekas, Dynamic programming and optimal control, Vol. I.
Athena Scientific, 2005.
[2] N. Elia and S. Mitter, “Stabilization of linear systems with limited
information,” IEEE Trans. on Automatic Control, vol. 46, no. 9, pp.
1384–1400, 2001.
[3] G. Nair and R. Evans, “Stabilizability of stochastic linear systems with
finite feedback data rates,” SIAM Journal on Control and Optimization,
vol. 43, no. 2, pp. 413–436, 2004.
[4] S. Tatikonda and S. Mitter, “Control under communication constraints,”
IEEE Trans. on Automatic Control, vol. 49, no. 7, pp. 1056–1068, 2004.
[5] V. Borkar and S. Mitter, “LQG control with communication constraints,”
Comm., Comp., Control, and Signal Processing, pp. 365–373, 1997.
[6] J. L. Ny and G. Pappas, “Differentially private filtering,” IEEE Trans.
on Automatic Control, vol. 59, no. 2, pp. 341–354, 2014.
[7] E. Shafieepoorfard and M. Raginsky, “Rational inattention in scalar LQG
control,” in Proceedings of the 52th IEEE Conference in Decision and
Control, 2013, pp. 5733–5739.
[8] S. Joshi and S. Boyd, “Sensor selection via convex optimization,” IEEE
Transactions on Signal Processing, vol. 57, no. 2, pp. 451–462, 2009.
[9] V. Gupta, T. H. Chung, B. Hassibi, and R. M. Murray, “On a stochastic
sensor selection algorithm with applications in sensor scheduling and
sensor coverage,” Automatica, vol. 42, no. 2, pp. 251–260, 2006.
[10] J. L. Ny, E. Feron, and M. A. Dahleh, “Scheduling continuous-time
kalman filters,” IEEE Trans. on Aut. Control, vol. 56, no. 6, pp. 1381–
1394, 2011.
[11] S. T. Jawaid and S. L. Smith, “Submodularity and greedy algorithms in
sensor scheduling for linear dynamical systems,” Automatica, vol. 61,
pp. 282–288, 2015.
[12] Y. Zhao, F. Pasqualetti, and J. Cortés, “Scheduling of control nodes
for improved network controllability,” in Proceedings of the 55th IEEE
Conference on Decision and Control, 2016, pp. 1859–1864.
[13] V. Tzoumas, A. Jadbabaie, and G. J. Pappas, “Sensor placement for
optimal Kalman filtering,” in Amer. Contr. Conf., 2016, pp. 191–196.
[14] L. Carlone and S. Karaman, “Attention and anticipation in fast visualinertial navigation,” in Proceeding of the IEEE International Conference
on Robotics and Automation, 2017, pp. 3886–3893.
[15] T. H. Summers, F. L. Cortesi, and J. Lygeros, “On submodularity and
controllability in complex dynamical networks,” IEEE Transactions on
Control of Network Systems, vol. 3, no. 1, pp. 91–101, 2016.
[16] E. Nozari, F. Pasqualetti, and J. Cortés, “Time-invariant versus timevarying actuator scheduling in complex networks,” in American Control
Conference, 2017, pp. 4995–5000.
[17] T. Summers and J. Ruths, “Performance bounds for optimal feedback
control in networks,” arXiv e-prints:1707.04528, 2017.
[18] T. Summers and M. Kamgarpour, “Performance guarantees for greedy
maximization of non-submodular set functions in systems and control,”
arXiv e-prints:1712.04122, 2017.
[19] A. F. Taha, N. Gatsis, T. Summers, and S. Nugroho, “Time-varying
sensor and actuator selection for uncertain cyber-physical systems,”
ArXiv e-prints:1708.07912, 2017.
[20] G. Nair, F. Fagnani, S. Zampieri, and R. Evans, “Feedback control under
data rate constraints: An overview,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 108–137, 2007.
[21] J. Hespanha, P. Naghshtabrizi, and Y.Xu, “A survey of results in networked control systems,” Pr. of the IEEE, vol. 95, no. 1, p. 138, 2007.
[22] J. Baillieul and P. Antsaklis, “Control and communication challenges in
networked real-time systems,” Proceedings of the IEEE, vol. 95, no. 1,
pp. 9–28, 2007.
[23] T. Tanaka and H. Sandberg, “SDP-based joint sensor and controller
design for information-regularized optimal LQG control,” in IEEE 54th
Annual Conference on Decision and Control, 2015, pp. 4486–4491.
[24] T. Tanaka, P. M. Esfahani, and S. K. Mitter, “LQG control with
minimum directed information: Semidefinite programming approach,”
IEEE Trans. on Automatic Control, vol. 63, no. 1, pp. 37–52, 2018.
[25] L. A. Wolsey, “An analysis of the greedy algorithm for the submodular
set covering problem,” Combinatorica, vol. 2, no. 4, pp. 385–393, 1982.
[26] S. Khuller, A. Moss, and J. S. Naor, “The budgeted maximum coverage
problem,” Info. Processing Letters, vol. 70, no. 1, pp. 39–45, 1999.
[27] M. Sviridenko, “A note on maximizing a submodular set function subject
to a knapsack constraint,” Operations Research Letters, vol. 32, no. 1,
pp. 41–43, 2004.
[28] A. Krause and C. Guestrin, “A note on the budgeted maximization of
submodular functions,” 2005.
[29] A. A. Bian, J. M. Buhmann, A. Krause, and S. Tschiatschek, “Guarantees for greedy maximization of non-submodular functions,” in Proceedings of the 34th International Confer. on Machine Learning, 2017.
[30] V. Tzoumas, L. Carlone, G. J. Pappas, and A. Jadbabaie, “SensingConstrained LQG Control,” arXiv e-prints: 1709.08826, 2017.
[31] G. Nemhauser, L. Wolsey, and M. Fisher, “An analysis of approximations
for maximizing submodular set functions – I,” Mathematical Programming, vol. 14, no. 1, pp. 265–294, 1978.
[32] U. Feige, “A threshold of ln(n) for approximating set cover,” Journal
of the ACM, vol. 45, no. 4, pp. 634–652, 1998.
[33] V. Tzoumas, K. Gatsis, A. Jadbabaie, and G. J. Pappas, “Resilient
monotone submodular maximization,” in Proceedings of the 56th IEEE
Conference on Decision and Control, 2017.
[34] D. S. Bernstein, Matrix mathematics. Princeton University Press, 2005.
[35] D. Coppersmith and S. Winograd, “Matrix multiplication via arithmetic
progressions,” J. of Symbolic Comp., vol. 9, no. 3, pp. 251–280, 1990.
A PPENDIX A
P RELIMINARY FACTS
This appendix contains a set of lemmata that will be used
to support the proofs in this paper (Appendices B–F).
Lemma 1 ([34, Proposition 8.5.5]). Consider two positive
definite matrices A1 and A2 . If A1 A2 then A2−1 A−1
1 .
Lemma 2 (Trace inequality [34, Proposition 8.4.13]). Consider a symmetric matrix A, and a positive semi-definite matrix
B of appropriate dimension. Then,
λmin (A)tr (B) ≤ tr (AB) ≤ λmax (A)tr (B) .
Lemma 3 (Woodbury identity [34, Corollary 2.8.8]). Consider matrices A, C, U and V of appropriate dimensions, such
that A, C, and A + U CV are invertible. Then,
(A + U CV )−1 = A−1 − A−1 U (C −1 + V A−1 U )−1 V A−1.
Lemma 4 ([34, Proposition 8.5.12]). Consider two symmetric
matrices A1 and A2 , and a positive semi-definite matrix B.
If A1 A2 , then tr (A1 B) ≤ tr (A2 B).
Lemma 5 ([1, Appendix E]). For any sensor set S ⊆ V, and
for all t = 1, 2, . . . , T , let x̂t (S) be the Kalman estimator
of the state xt , i.e., x̂t (S) , E[xt |y1 (S), y2 (S), . . . , yt (S)],
and Σt|t (S) be x̂t (S)’s error covariance, i.e., Σt|t (S) ,
E[(x̂t (S) − xt )(x̂t (S) − xt )T ]. Then, Σt|t (S) is the solution
of the Kalman filtering recursion
Σt|t (S) = [Σt|t−1 (S)−1 + Ct (S)T Vt (S)−1 Ct (S)]−1,
Σt+1|t (S) = At Σt|t (S)AT
t + Wt ,
(23)
with boundary condition Σ1|0 (S) = Σ1|0 .
Lemma 6. For any sensor set S ⊆ V, let Σ1|1 (S) be defined as
in eq. (23), and consider two sensor sets S1 , S2 ⊆ V. If S1 ⊆
S2 , then Σ1|1 (S1 ) Σ1|1 (S2 ).
Proof of Lemma 6: Let D = S2 \ S1 , and observe that
for all t = 1, 2, . . . , T , the notation in Definition 1 implies
X
T
Ct (S2 )T Vt (S2 )−1 Ct (S2 ) =
Ci,t
Vi,t Ci,t
i∈S2
=
X
T
Ci,t
Vi,t Ci,t +
i∈S1
X
X
T
Ci,t
Vi,t Ci,t
i∈D
T
Ci,t
Vi,t Ci,t
i∈S1
= Ct (S1 )T Vt (S1 )−1 Ct (S1 ).
(24)
Therefore, Lemma 1 and ineq. (24) imply
T
−1
Ct (S2 )]−1
Σ1|1 (S2 ) = [Σ−1
1|0 + C1 (S2 ) Vt (S2 )
T
−1
[Σ−1
Ct (S1 )]−1 = Σ1|1 (S1 ).
1|0 + C1 (S1 ) Vt (S1 )
Lemma 7. Let Σt|t be defined as in eq. (23) with boundary
condition Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (23) with
boundary condition Σ̄1|0 . If Σt|t Σ̄t|t , then Σt+1|t Σ̄t+1|t .
Proof of Lemma 7: We complete the proof in two steps:
T
first, from Σt|t Σ̄t|t , it follows At Σt|t AT
t At Σ̄t|t At .
T
Then, from eq. (23), it its Σt+1|t = At Σt|t At + Wt
At Σ̄t|t AT
t + Wt = Σ̄t+1|t .
Lemma 8. Let Σt|t−1 be defined as in eq. (23) with boundary
condition Σ1|0 ; and, let Σ̄t|t−1 be defined as in eq. (23) with
boundary condition Σ̄1|0 . If Σt|t−1 Σ̄t|t−1 , then Σt|t Σ̄t|t .
Proof of Lemma 8: From eq. (23), it is Σt|t =
T −1
−1
T −1
−1
(Σ−1
(Σ̄−1
= Σ̄t|t ,
t|t−1 + Ct Vt Ct )
t|t−1 + Ct Vt Ct )
since Lemma 1 and the condition Σt|t−1 Σ̄t|t−1 imply
−1
T −1
T −1
Σ−1
t|t−1 + Ct Vt Ct Σ̄t|t−1 + Ct Vt Ct , which in turn
−1
T −1
−1
T −1
−1
implies (Σt|t−1 + Ct Vt Ct )
(Σ̄−1
t|t−1 + Ct Vt Ct ) .
Corollary 1. Let Σt|t be defined as in eq. (23) with boundary
condition Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (23)
with boundary condition Σ̄1|0 . If Σt|t Σ̄t|t , then Σt+i|t+i
Σ̄t+i|t+i for any positive integer i.
Proof of Corollary 1: If Σt|t Σ̄t|t , from Lemma 7
we get Σt+1|t Σ̄t+1|t , which, from Lemma 8, implies
Σt+1|t+1 Σ̄t+1|t+1 . By repeating the previous argument
another (i − 1) times, the proof is complete.
Corollary 2. Let Σt|t be defined as in eq. (23) with boundary
condition Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (23) with
boundary condition Σ̄1|0 . If Σt|t Σ̄t|t , then Σt+i|t+i−1
Σ̄t+i|t+i−1 for any positive integer i.
Proof of Corollary 2: If Σt|t Σ̄t|t , from Corollary 1,
we get Σt+i−1|t+i−1 Σ̄t+i−1|t+i−1 , which, from Lemma 7,
implies Σt+i|t+i−1 Σ̄t+i|t+i−1 .
Lemma 9. Consider positive
real numbers a, b, γ,
Pn
a1 , a2 , . . . , an such that i=1 ai = a. The function
f (a1 , a2 , . . . , an ) = 1 −
n
Y
i=1
ai
1−γ
b
achieves its minimum at a1 = a2 = . . . = an = a/n.
In particular,
aγ n
f (a/n, a/n, . . . , a/n) = 1 − 1 −
≥ 1 − e−aγ/b.
bn
Proof of Lemma 9: To find f ’s minimum we use
the method of Lagrange multipliers. In particular, the partialPderivative of φ(a1 , a2 , . . . , an ) , f (a1 , a2 , . . . , an ) +
n
λ ( i=1 ai − a), where λ is the Lagrangian multiplier, with
respect to aj is as follows:
γ Y
ai
∂φ
+ λ.
(25)
=
1−γ
∂aj
b
b
i6=j
At an f ’s minimum, the partial derivative in eq. (25) is zero
for all j, which implies that for all j:
γ Y
ai
λ=−
;
(26)
1−γ
b
b
i6=j
Since λ is constant, eq. (26) implies for any j1 and j2 that
γ Y
ai γ Y
ai
=
,
1−γ
1−γ
b
b
b
b
i6=j1
i6=j2
which in turn implies that
aj
aj
1−γ 1 =1−γ 2,
b
b
from where we conclude aj1 = aj2 (for any j1 and j2 ).
The lower bound for the minimum value of f follows from
the fact that 1 − x ≤ e−x for all real x.
Proposition 2 (Monotonicity of cost function in eq. (10)).
Consider the cost function in eq. P
(10), namely, for any
sensor
T
set S ⊆ V the set function
tr
Θ
Σ
(S)
. Then,
t
t|t
t=1
for
any
sensor
sets
such
that
S
⊆
S
⊆
V,
it
holds
1
2
PT
PT
tr
Θ
Σ
(S
)
≥
tr
Θ
Σ
(S
)
.
t
1
t
2
t|t
t|t
t=1
t=1
Proof: Lemma 6 implies Σ1|1 (S1 ) Σ1|1 (S2 ), and
then, Corollary 1 implies Σt|t (S1 ) Σt|t (S2 ). Finally, for
any t = 1, 2,. . . , T , Lemma 4 implies tr Θt Σt|t (S1 ) ≥
tr Θt Σt|t (S2 ) , since each Θt is symmetric.
A PPENDIX B
P ROOF OF T HEOREM 1
We first prove part (1) of Theorem 1 (Appendix B.1), and
then we prove part (2) of Theorem 1 (Appendix B.2).
B.1. Proof of part (1) of Theorem 1
We use the following lemma to prove Theorem 1’s part (1).
Lemma 10. For any active sensor set S ⊆ V, and any admissible control policy u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)},
let h[S, u1:T (S)] be Problem 1’s cost function, i.e.,
PT
h[S, u1:T (S)] , t=1 E(kxt+1 (S)k2Qt +kut(S)k2Rt ).
Further define the following set-valued function:
g(S) , minu1:T (S) h[S, u1:T (S)].
Consider any sensor set S ⊆ V, and let u⋆1:T (S) be the vector of control policies (K1 x̂1 (S), K2 x̂2 (S), . . . , KT x̂T (S)).
Then u⋆1:T (S) is an optimal control policy:
u⋆1:T (S) ∈ arg min h[S, u1:T (S)],
(27)
u1:T (S)
i.e., g(S) = h[S, u⋆1:T (S)], and in particular, u⋆1:T (S) attains
a (sensor-dependent) LQG cost equal to:
g(S) = E(kx1 kN1 )+
T
X
tr[Θt Σt|t (S)] + tr (Wt St ) . (28)
t=1
Proof of Lemma 10: Let ht [S, ut:T (S)] be the LQG cost
in Problem 1 from time t up to time T , i.e.,
ht [S, ut:T (S)] ,
T
X
E(kxk+1 (S)k2Qt +kuk (S)k2Rt ).
k=t
and define gt (S) , minut:T (S) ht [S, ut:T (S)]. Clearly, g1 (S)
matches the LQG cost in eq. (28).
We complete the proof inductively. In particular, we first
prove Lemma 10 for t = T , and then for any other t ∈
{1, 2, . . . , T − 1}. To this end, we use the following observation: given any sensor set S, and any time t ∈ {1, 2, . . . , T },
gt (S) = min E(kxt+1 (S)k2Qt +kut (S)k2Rt ) + gt+1 (S) ,
ut (S)
(29)
with boundary condition gT +1 (S) = 0. In particular, eq. (29)
holds since
gt (S) = min E kxt+1 (S)k2Qt +kut (S)k2Rt )+
ut (S)
min
ut+1:T (S)
ht+1 [S, ut+1:T (S)]} ,
where one can easily recognize the second summand to match
the definition of gt+1 (S).
We prove Lemma 10 for t = T . From eq. (29), for t = T ,
gT (S) = minuT (S) E(kxT +1 (S)k2QT +kuT (S)k2RT )
2
= minuT (S) E(kA
T xT + BT uT (S) + wT kQT +
kuT (S)k2RT ) ,
(30)
since xT +1 (S) = AT xT + BT uT (S) + wT , as per eq. (1);
we note that for notational simplicity we drop henceforth the
dependency of xT on S since xT is independent of uT (S),
which is the variable under optimization in the optimization
problem (30). Developing eq. (30) we get:
gT (S)
= minuT (S) E(uT (S)T BTT QT BT uT (S) + wTT QT wT +
T T
T
xT
T AT QT AT xT + 2xT AT QT BT uT (S)+
T T
T
Q
w
+
2u
(S)
BT QT wT + kuT (S)k2RT )
A
2xT
T
T T T T
= minuT (S) E(uT (S)T BTT QT BT uT (S) + kwT k2QT +
T
T T
2
xT
T AT QT AT xT + 2xT AT QT BT uT (S) + kuT kRT ) ,
(31)
where the latter equality holds since wT has zero mean
and wT , xT , and uT (S) are independent. From eq. (31),
rearranging the terms, and using the notation in eq. (9),
=
=
=
=
(i)
=
=
=
(ii)
We now prove that if Lemma 10 holds for t = l + 1, it
also holds for t = l. To this end, assume eq. (29) holds for
t = l + 1. Using the notation in eq. (9),
gl (S) = minul (S) E(kxl+1 (S)k2Ql +kul (S)k2Rl ) + gl+1 (S)
= minul (S) E(kxl+1 (S)k2Ql +kul (S)k2Rl )+
PT
E(kxl+1 (S)k2Nl+1 ) + k=l+1 tr Θk Σk|k (S) +
tr (Wk Sk )]}
= minul (S) E(kxl+1 (S)k2Sl +kul (S)k2Rl )+
o
PT
k=l+1 [tr Θk Σk|k (S) + tr (Wk Sk )]
PT
= k=l+1 [tr Θk Σk|k (S) + tr (Wk Sk )]+
minul (S) E(kxl+1 (S)k2Sl +kul (S)k2Rl ).
(34)
In eq. (34), the minimization in the last summand can
be solved by following the same steps as for the proof of
Lemma 10 for t = T , leading to:
2
minul (S) E(kxl+1 (S)k2Sl +ku
l (S)kRl ) =
2
E(kxl kNl ) + tr Θl Σl|l (S) + tr (Wl Ql ) ,
(35)
and ul (S) = Kl x̂l (S). Therefore, by substituting eq. (35) back
into eq. (34), we get:
P
gT (S)
gl (S) = E(kxl k2Nl ) + Tk=l [tr Θk Σk|k (S) + tr (Wk Sk )].
T
T
minuT (S) E(uT (S) (BT QT BT + RT )uT (S)+
(36)
T
T T
kwT k2QT +xT
T AT QT AT xT + 2xT AT QT BT uT (S)
which proves that if Lemma 10 holds for t = l + 1, it also
T
minuT (S) E(kuT (S)k2MT +kwT k2QT +xT
T AT QT AT xT + holds for t = l. By induction, this also proves that Lemma 10
T T
2xT AT QT BT uT (S)
holds for l = 1, and we already observed that g1 (S) matches
T
Q
A
x
−
A
minuT (S) E(kuT (S)k2MT +kwT k2QT +xT
T
T
T
the
original LQG cost in eq. (28), hence concluding the proof.
T T
−1
T
Q
B
M
)M
u
(S)
(−A
2xT
T T
T
T T T T 2
2
T T
minuT (S) E(kuT (S)k
Proof of Theorem 1’s part (1): The proof follows
MT +kwT kQT +xT AT QT AT xT −
T
from
Lemma 10. In particular, eq. (10) is a direct conseK
M
u
(S)
2xT
T T
T T
quence
of eq. (28), since the value of Problem 1 is equal
minuT (S) E(kuT (S) − KT xT k2MT +kwT k2QT +
to
min
g(S), and both E(kx1 kN1 ) = tr Σ1|0 N1
T
T
xT
PTS⊆V,c(S)≤b
T (AT QT AT − KT MT KT )xT
and t=1 tr (Wt St ) in eq. (28) are independent of the choice
minuT (S) E(kuT (S) − KT xT k2MT +kwT k2QT +
of
the sensor set S. Finally, eq. (11) directly follows from
T
T
xT (AT QT AT − ΘT )xT
eq. (27).
2
2
2
min
E(ku (S) − K x k +kw k +kx k
T
uT (S)
= minuT (S) E(kuT (S) −
E(kxT k2NT ),
T
T MT
KT xT k2MT )
T QT
T NT
+ tr (WT QT ) +
B.1. Proof of part (2) of Theorem 1
(32)
where equality (i) follows from completion
of
squares,
and
2
T
=
equality (ii) holds
since
E(kw
k
)
=
E
tr
w
Q
w
T
T
T
Q
T
T
tr E(wTT wT )QT = tr (WT QT ). Now we note that
We use the following lemma to prove Theorem 1’s part (2).
Lemma 11. The sensor set S ⋆ and the controllers u⋆1:T are a
solution to Problem 2 if and only if they are a solution to the
optimization problem
min E(kuT (S) − KT xT k2MT )
min
uT (S)
= E(kKT x̂T (S) − KT xT k2MT )
= tr ΘT ΣT |T (S) ,
S⊆V,u1:T (S)
(37)
where
(33)
since x̂T (S) is the Kalman estimator of the state xT , i.e.,
the minimum mean square estimator of xT , which implies
that KT x̂T (S) is the minimum mean square estimator of
KT xT (S) [1, Appendix E]. Substituting (33) back into
eq. (32), we get:
gT (S) = E(kxT k2NT ) + tr ΘT ΣT |T (S) + tr (WT QT ) ,
which proves that Lemma 10 holds for t = T .
c(S), s.t. g(S) ≤ κ,
g(S) , min h [S, u1:T (S)] .
u1:T (S)
Proof of Lemma 11: We prove the lemma by contradiction. In particular, first we prove that if the sensor set S ⋆ and
the controllers u⋆1:T are a solution to Problem 2, then they are
also a solution to the problem in eq. (37); and second, we
prove that if S ⋆ and u⋆1:T are a solution to the problem in
eq. (37), then they are also a solution to Problem 2.
We prove that if the sensor set S ⋆ and the controllers u⋆1:T
are a solution to Problem 2, then they are also a solution
to the problem in eq. (37). To this end, let the sensor set
S ⋆ and the controllers u⋆1:T be a solution to Problem 2, and
assume by contradiction that they are not a solution to the
problem in eq. (37), which instead has solution Sb and u
b1:T . By
optimality of Sb and u
b1:T (and suboptimality of S ⋆ and u⋆1:T )
b < c(S ⋆ ). In addition, it also holds
in eq. (37), it follows c(S)
b
b
g(S) ≤ κ, since (S, u
b1:T ) must be feasible for
the problem
bu
in eq. (37). However, the latter implies that h S,
b1:T ≤ κ.
bu
Therefore, (S,
b1:T ) is feasible for Problem 2 and has a better
objective value with respect to the optimal solution (S ⋆ , u⋆1:T )
b < c(S ⋆ )), leading to contradiction.
(we already observed c(S)
Hence, if the sensor set S ⋆ and the controllers u⋆1:T are a
solution to Problem 2, then they are also a solution to (37).
We now prove that if the sensor set S ⋆ and the controllers
u⋆1:T are a solution to the problem in eq. (37), then they are also
a solution to Problem 2. To this end, let the sensor set S ⋆ and
the controllers u⋆1:T be a solution to the problem in eq. (37),
and assume that they are not a solution to Problem 2, which
bu
bu
instead has solution (S,
b1:T ). By optimality of (S,
b1:T ) (and
suboptimality of S ⋆ and u⋆1:T ) in Problem
2
,
it follows
⋆
b
b
c(S) < c(S ). In addition, it is h S, u
b1:T ≤ κ, since
b
(S, u
b1:T ) must be feasible for Problem 2, and, as a result,
b ≤ κ. Therefore, (S,
bu
it also holds g(S)
b1:T ) is feasible for
the problem in eq. (37) and has a better objective value with
respect to the optimal solution (S ⋆ , u⋆1:T ) (we already observed
b < c(S ⋆ )), leading to contradiction. Hence, if the sensor
c(S)
set S ⋆ and the controllers u⋆1:T are a solution to the problem
in eq. (37), then they are also a solution to Problem 2.
Proof of Theorem 1’s part (2): The proof follows from
Lemma 10 and Lemma 11. In particular, similarly to the proof
of Theorem 1’s part (1), Lemma 10, along
with eq. (28) and
the fact that E(kx1 kN1 ) = tr Σ1|0 N1 , implies that if the
sensor set S ⋆ and the controllers u⋆1:T are a solution to the
optimization problem in eq. (37), then S ⋆ and the controllers
u⋆1:T can be computed in cascade as follows:
S ⋆ ∈ arg min c(S), s.t. tr[Θt Σt|t (S)] ≤
S⊆V
T
X
tr (Wt St ) , (38)
κ − tr Σ1|0 N1 −
t=1
u⋆t = Kt x̂t (S ⋆ ),
t = 1, . . . , T.
(39)
In addition, Lemma 11 implies that (S ⋆ , u⋆1:T ) is a solution to
Problem 2. As a result, eqs. (12)-(13) hold true.
P ROOF
OF
A PPENDIX C
T HEOREM 2 AND P ROPOSITION 1
We first prove Theorem 2 (Appendix C.1), and then prove
Proposition 1 (Appendix C.2).
C.1. Proof of Theorem 2
We consider theP
following notation: for any sensor set S ⊆
V, we let f (S) , Tt=1 tr[Θt Σt|t (S)] be the cost function in
eq. (10), S ⋆ be a solution in eq. (10), and b⋆ , c(S ⋆ ), that
is, b⋆ is the cost of the sensor set S ⋆. In addition, consider
the computation of the set Sb2 in Algorithm 2 (lines 3-19),
where Sb2 refers to the set that Algorithm 2 has constructed
by the end of the line 19; we let G , Sb2 . We also let si
be the i-th element added in G during the i-th iteration of
Algorithm 2’s “while loop” (lines 3-16). Moreover, we let
Gi , {s1 , s2 , . . . , si }, that is, Gi is the subset of G constructed
during the first i iterations of Algorithm 2’s “while loop”
(lines 3-16). Finally, we consider that Algorithm 2’s “while
loop” (lines 3-16) terminates after l + 1 iterations.
There are two scenarios under which Algorithm 2’s “while
loop” (lines 3-16) terminates: (i) the trivial scenario where
V ′ = ∅, that is, where all available sensors in V can
been chosen by Algorithm 2 as active while satisfying the
budget constraint b; and (ii) the non-trivial scenario where
Algorithm 2’s “while” loop (lines 3–16) terminates because
c(Gl+1 ) > b, that is, where the addition of the sensor sl+1 in Gl
makes the sensor cost of Gl+1 violate the budget constraint b.
Henceforth, we focus on the second, non-trivial scenario,
which implies that sl+1 will be removed by the “if” statement
in Algorithm 2’s lines 17–19 and, as a result, Gl = Sb2 .
We prove Theorem 2 using the following two lemmas.
Lemma 12 (Generalization of [28, Lemma 2]). For i =
1, 2, . . . , l + 1, it holds
γf c(si )
(f (Gi−1 ) − f (S ⋆ )).
b⋆
Proof of Lemma 12: Due to the monotonicity of the cost
function f in eq. (10) (Proposition 2), it holds
f (Gi−1 ) − f (Gi ) ≥
f (Gi−1 ) − f (S ⋆ ) ≤ f (Gi−1 ) − f (S ⋆ ∪ Gi−1 )
= f (Gi−1 ) − f [(S ⋆ \ Gi−1 ) ∪ Gi−1 ].
Let {z1 , z2 , . . . , zm } , S ⋆ \ Gi−1 , and also let
dj , f (Gi−1 ∪{z1 , z2 , . . . , zj−1 })−f (Gi−1 ∪{z1 , z2 , . . . , zj }),
Pm
for j = 1, 2, . . . , m. Then, f (Gi−1 ) − f (S ⋆ ) ≤ j=1 dj .
Notice that
dj
f (Gi−1 ) − f (Gi−1 ∪ {zj })
f (Gi−1 ) − f (Gi )
≤
≤
,
c(zj )
γf c(zj )
γf c(si )
where the first inequality holds due to the Definition 5 of the
supermodularity ratio γf , and the second inequality holds due
to the greedy rule (Algorithm
2’s line 13) and the definitions
Pm
of Gi , and si . Since j=1 c(zj ) ≤ b⋆ , it holds that
f (Gi−1 ) − f (S ⋆ ) ≤
m
X
j=1
dj ≤ b⋆
f (Gi−1 ) − f (Gi )
.
γf c(si )
Lemma 13 (Adapation of [28, Lemma 3]). For i =
1, 2, . . . , l + 1, it holds
i
Y
γ
c(s
)
f
j
[f (∅) − f (S ⋆ )].
f (∅) − f (Gi ) ≥ 1 −
1−
⋆
b
j=1
Proof of Lemma 13: We complete the proof inductively.
In particular, for i = 1, we need to prove f (∅) − f (G1 ) ≥
γf c(s1 )/b⋆ [f (∅) − f (S ⋆ )], which follows from Lemma 12 for
i = 1. Then, we have for i > 1:
f (∅) − f (Gi ) = f (∅) − f (Gi−1 ) + [f (Gi−1 ) − f (Gi )]
≥ f (∅) − f (Gi−1 )+
γf c(si )
(f (Gi−1 ) − f (S ⋆ ))
⋆
b
γf c(si )
= 1−
[f (∅) − f (Gi−1 ])+
b⋆
γf c(si )
[f (∅) − f (S ⋆ )]
b⋆
i−1
Y
γf c(si )
γf c(sj )
≥ 1−
1−
1−
b⋆
b⋆
j=1
γf c(si )
[f (∅) − f (S ⋆ )] +
[f (∅) − f (S ⋆ )]
b⋆
i
Y
γf c(sj )
[f (∅) − f (S ⋆ )],
1−
= 1 −
⋆
b
j=1
where we used Lemma 12 for the first inequality and the
induction hypothesis for the second inequality.
Proof of Theorem 2’s part (1) (Algorithm 1’s approximation quality): We first prove the approximation bound
b
γg
−γg
) in ineq. (17) and, then, the bound 1−e−γg c(S)/b.
2 (1 − e
γg
−γg
)
To prove Algorithm 1’s approximation bound 2 (1 − e
Pl+1
in ineq. (17), we let b′ , j=1 c(sj ). It holds
γf c(sj )
1−
f (∅) − f (Gl+1 ) ≥ 1 −
[f (∅) − f (S ⋆ )]
⋆
b
j=1
′
⋆
≥ 1 − e−γf b /b [f (∅) − f (S ⋆ )],
≥ 1 − e−γf [f (∅) − f (S ⋆ )],
(40)
l+1
Y
where the first inequality follows from Lemma 13, the second
inequality from Lemma 9, and ineq. (40) from the fact that
′
⋆
b′ /b⋆ ≥ 1 and, as a result, e−γf b /b ≤ e−γf, that is, 1 −
′
⋆
e−γf b /b ≥ 1 − e−γf.
In addition, it is f (∅)− f (Sb1 ) ≥ γf [f (Gl )− f (Gl+1 )] due to
the Definition 5 of the supermodularity ratio and, as a result,
f (∅) − f (Sb1 ) ≥ γf [f (Gl ) − f (∅) + f (∅) − f (Gl+1 )], which
after rearranging its terms gives
γf [f (∅) − f (Gl+1 )]
≤ f (∅) − f (Sb1 ) + γf [f (∅) − f (Gl )]
n
o
≤ 2 max f (∅) − f (Sb1 ), γf [f (∅) − f (Gl )] .
(41)
By substituting ineq. (40) in ineq. (41) and rearranging the
terms we have
n
o
max f (∅) − f (Sb1 ), γf [f (∅) − f (Gl )]
γf
1 − e−γf [f (∅) − f (S ⋆ )],
≥
2
which implies the inequality
h
i
max f (∅) − f (Sb1 ), f (∅) − f (Gl )
γf
≥
1 − e−γf [f (∅) − f (S ⋆ )],
2
(42)
since γf takes values in [0, 1] by the Definition 5 of the
supermodularity ratio.
γ
Algorithm 1’s approximation bound 2g (1 − e−γg ) in
ineq. (17) follows from ineq. (42) as the combination of the
following three observations:
b2 , due to the definition of Gl , and, as a result,
• it is Gl = S
f (∅) − f (Gl ) = f (∅) − f (Sb2 ).
b such at Sb ∈
• Algorithm 2 returns the set S
arg maxS∈{Sb1 ,Sb2 } [f (∅) − f (S)] (per Algorithm 2’s
line 20) and, as a result, the previous observation, along
with ineq. (42), gives:
b ≥ γf 1 − e−γf [f (∅) − f (S ⋆ )]. (43)
f (∅) − f (S)
2
•
Finally, Lemma 10 implies that for
sets S, S ′ ⊆ V it
Pany
T
is g(S) = f (S) + E(kx1 kN1 ) + t=1 tr (Wt St ), where
PT
E(kx1 kN1 ) + t=1 tr (Wt St ) is independent of S. As a
result, for any sets S, S ′ ⊆ V it is f (S)−f (S ′ ) = g(S)−
g(S ′ ), which implies γf = γg due to the Definition 5 of
the supermodularity ratio. In addition, Lemma 10 implies
that for any S ⊆ V it is g(S) = h[S, u1:T (S)] and g ⋆ =
g(S ⋆ ). Thereby, for any S ⊆ V it is f (∅)−f (S) = g(∅)−
g(S) = h[∅, u1:T (∅)]−h[S, u1:T (S)] and f (∅)−f (S ⋆ ) =
g(∅) − g(S ⋆ ) = h[∅, u1:T (∅)] − g ⋆. Overall, ineq. (43) is
written as
b u1:T (S)]
b ≥
h[∅, u1:T (∅)] − h[S,
γf
1 − e−γf {h[∅, u1:T (∅)] − g ⋆ } ,
2
which implies the bound
γg
2
(1 − e−γg ) in ineq. (17).
b
It remains to prove the bound 1 − e−γg c(S)/b in ineq. (17).
To this end, we first have:
l
Y
γf c(sj )
[f (∅) − f (Gl )]
1−
f (∅) − f (Gl ) ≥ 1 −
b⋆
j=1
⋆
≥ 1 − e−γf c(Gl )/b [f (∅) − f (S ⋆ )],
(44)
≥ 1 − e−γf c(Gl )/b [f (∅) − f (S ⋆ )],
where the first inequality follows from Lemma 13, the second
inequality from Lemma 9, and ineq. (44) from the fact
that c(Gl )/b⋆ ≥ c(Gl )/b, since b⋆ ≤ b, which implies
⋆
′
⋆
e−γf c(Gl )/b ≤ e−γf c(Gl )/b, i.e., 1−e−γf b /b ≥ 1−e−γf c(Gl )/b.
The rest of the proof is completed using the combination of
the three observations in the previous paragraph, that we used
γ
to prove Algorithm 1’s approximation bound 2g (1 − e−γg ) in
ineq. (17).
Proof of Theorem 2’s part (2) (Algorithm 1’s running
time): We compute Algorithm 1’s running time by adding
the running times of Algorithm 1’s lines 1-5:
a) Running time of Algorithm 1’s line 1: Algorithm 1’s
line 1 needs O(T n2.4 ) time, using the Coppersmith algorithm
for both matrix inversion and multiplication [35].
b) Running time of Algorithm 1’s line 2: Algorithm 1’s
line 2 running time is the running time of Algorithm 2, whose
running time we show next to be O(|V|2 T n2.4 ). To this end,
we first compute the running time of Algorithm 2’s line 1,
and finally the running time of Algorithm 2’s lines 3–16.
Algorithm 2’s lines 3–16 are repeated at most |V|2 times, since
before the end of each iteration of the “while loop” in line 3
the added element in Sb2 (line 14) is removed from V ′ (line 15).
We now need to find the running time of Algorithm 2’s lines 4–
16; to this end, we first find the running time of Algorithm 2’s
lines 4–12, and then the running time of Algorithm 2’s line 13.
In more detail, the running time of Algorithm 2’s lines 4–12
is O(|V|T n2.4 ), since Algorithm 2’s lines 4–12 are repeated
at most |V| times and Algorithm 2’s lines 5–10, as well as,
line 11 need O(T n2.4 ) time, using the Coppersmith-Winograd
algorithm for both matrix inversion and multiplication [35].
Moreover, Algorithm 2’s line 13 needs O[|V|log(|V|)] time,
since it asks for the maximum among at most |V| values of
the gain(·) , which takes O[|V|log(|V|)] time to be found, using,
e.g., the merge sort algorithm. In sum, Algorithm 2’s running
time is upper bounded by O[|V|2 T n2.4 + |V|2 log(|V|)], which
is equal to O(|V|2 T n2.4 ).
c) Running time of Algorithm 1’s lines 3-5: Algorithm 1’s lines 3-5 need O(T n2.4 ) time, using the Coppersmith
algorithm for both matrix inversion and multiplication [35].
In sum, Algorithm 1’s running time is upper bounded by
O(|V|2 T n2.4 + 2T n2.4 ) which is equal to O(|V|2 T n2.4 ).
C.2. Proof of Proposition 1
The proof of Proposition 1 follows the same steps as the
proof of Theorem 2 and for this reason we omit it.
A PPENDIX D
P ROOF OF T HEOREM 3
We consider the following
notation: for any sensor set S ⊆
PT
V, we let f (S) ,
tr[Θ
t Σt|t (S)] be the cost function
t=1
in eq. (12), the sensor set S ⋆ be a solution to Problem 2,
and b⋆ be equal to c(S ⋆ ), that is, b⋆ is the optimal value of
Problem 2. In addition, consider the computation of the set
b We also let
Sb in Algorithm 4 (lines 3-16): we let G , S.
si be the i-th element added in G during the i-th iteration
of Algorithm 4’s “while loop” (lines 3-16). Finally, we let
Gi , {s1 , s2 , . . . , si }, that is, Gi is the subset of Sb constructed
during the first i iterations of Algorithm 4’s “while loop”
(lines 3-16).
We prove Theorem 3 using the following two lemmas.
Lemma 14 (Adaptation of Lemma 12). For i = 1, 2, . . . , |G|,
it holds
f (Gi−1 ) − f (Gi ) ≥
γf c(si )
(f (Gi−1 ) − f (S ⋆ )).
b⋆
Proof of Lemma 14: Due to the monotonicity of the cost
function f in eq. (10) (Proposition 2), it holds
f (Gi−1 ) − f (S ⋆ ) ≤ f (Gi−1 ) − f (S ⋆ ∪ Gi−1 )
= f (Gi−1 ) − f [(S ⋆ \ Gi−1 ) ∪ Gi−1 ].
Let {z1 , z2 , . . . , zm } , S ⋆ \ Gi−1 , and also let
dj , f (Gi−1 ∪{z1 , z2 , . . . , zj−1 })−f (Gi−1 ∪{z1 , z2 , . . . , zj }),
Pm
for j = 1, 2, . . . , m. Then, f (Gi−1 ) − f (S ⋆ ) ≤ j=1 dj .
Notice that
f (Gi−1 ) − f (Gi−1 ∪ {zj })
f (Gi−1 ) − f (Gi )
dj
≤
≤
,
c(zj )
γf c(zj )
γf c(si )
where the first inequality holds due to the Definition 5 of the
supermodularity ratio γf , and the second inequality holds due
to the greedy rule (Algorithm
4’s line 13) and the definitions
Pm
of Gi and si . Since j=1 c(zj ) ≤ b⋆ , it holds that
f (Gi−1 ) − f (S ⋆ ) ≤
m
X
j=1
dj ≤ b⋆
f (Gi−1 ) − f (Gi )
.
γf c(si )
Lemma 15 (Adaptation of Lemma 13). For i = 1, 2, . . . , |G|,
it holds
i
Y
γ
c(s
)
f
j
[f (∅) − f (S ⋆ )].
1−
f (∅) − f (Gi ) ≥ 1 −
⋆
b
j=1
Proof of Lemma 15: We complete the proof inductively.
In particular, for i = 1, we need to prove f (∅) − f (G1 ) ≥
γf c(s1 )/b⋆ [f (∅) − f (S ⋆ )], which follows from Lemma 14 for
i = 1. Then, we have for i > 1:
f (∅) − f (Gi ) = f (∅) − f (Gi−1 ) + [f (Gi−1 ) − f (Gi )]
≥ f (∅) − f (Gi−1 )+
γf c(si )
(f (Gi−1 ) − f (S ⋆ ))
⋆
b
γf c(si )
= 1−
[f (∅) − f (Gi−1 ])+
b⋆
γf c(si )
[f (∅) − f (S ⋆ )]
b⋆
i−1
Y
γf c(si )
γf c(sj )
≥ 1−
1−
1−
b⋆
b⋆
j=1
γf c(si )
[f (∅) − f (S ⋆ )]
[f (∅) − f (S ⋆ )] +
b⋆
i
Y
γf c(sj )
1−
= 1 −
[f (∅) − f (S ⋆ )],
⋆
b
j=1
using Lemma 14 for the first inequality and the induction
hypothesis for the second inequality.
Proof of Theorem 3’s part (1) (Algorithm 3’s approximation quality): We first observe that ineq. (19) holds since Algob u1:T (S)]
b ≤κ
rithm 3 returns the set Sb once the condition h[S,
is satisfied (Algorithm 4’s line 3).
It remains to prove ineq. (20). Let l , |G|; then, Gl = G,
by the definition of Gi , and from Lemma 13 for i = l − 1 it
holds
l−1
Y
γ
c(s
)
f
j
[f (∅) − f (S ⋆ )]
1−
f (∅) − f (Gl−1 ) ≥ 1 −
⋆
b
j=1
⋆
≥ 1 − e−γf c(Gl−1 )/b [f (∅) − f (S ⋆ )],
(45)
where ineq. (45) follows from Lemma 9. Moreover, Lemma 10
implies that forPany sensor sets S, S ′ ⊆ V it is g(S) = f (S)+
T
E(kx k ) + t=1 tr (Wt St ), where the term E(kx1 kN1 ) +
PT 1 N1
t=1 tr (Wt St ) is independent of S, and, as a result, it is
f (S) − f (S ′ ) = g(S) − g(S ′ ), which implies γf = γg .
Moreover, Lemma 10 implies for any S ⊆ V that g(S) =
h[S, u1:T (S)] and, as a result, it is f (∅) − f (Gl−1 ) =
h[∅, u1:T (∅)] − h[Gl−1 , u1:T (Gl−1 )] and f (∅) − f (S ⋆ ) =
h[∅, u1:T (∅)] − h[S ⋆, u1:T (S ⋆ )]. In sum, ineq. (45) is the same
as the inequality
h[∅, u1:T (∅)] − h[Gl−1 , u1:T (Gl−1 )] ≥
⋆
1 − e−γg c(Gl−1 )/b {h[∅, u1:T (∅)] − h[S ⋆, u1:T (S ⋆ )]} ,
⋆
which, by letting β , 1 − e−γg c(Gl−1 )/b and rearranging its
terms, is simplified to the inequality
(Algorithm 1’s running time) and for this reason we omit it.
A PPENDIX E
P ROOF OF T HEOREM 4
Proof of Theorem 4: We complete the proof by first
deriving a lower bound for the numerator of the supermodularity ratio γg , and then, by deriving an upper bound for the
denominator of the supermodularity ratio γg .
use the following notation: c , E(xT
1 N1 x1 ) +
PWe
T
tr
(W
S
),
and
for
any
sensor
set
S
⊆
V,
and time
t
t
t=1
t = 1, 2, . . . , T , ft (S) , tr Θt Σt|t (S) . Then, the cost
P
function g(S) in eq. (16) is written as g(S) = c+ Tt=1 ft (S),
due to eq. (28) in Lemma 10.
a) Lower bound for the numerator of the supermodularity ratio γg : Per the supermodularity ratio Definition 5, the
numerator of the submodularity ratio γg is of the form
T
X
[ft (S) − ft (S ∪ {v})],
(47)
t=1
for some sensor set S ⊆ V, and sensor v ∈ V; to lower bound
the sum in (47), we lower bound each ft (S) − ft (S ∪ {v}).
To this end, from eq. (23) in Lemma 5, observe
X
T
C̄i,t
C̄i,t ]−1.
Σt|t (S ∪ {v}) = [Σ−1
(S
∪
{v})
+
t|t−1
i∈S∪{v}
h[Gl−1 , u1:T (Gl−1 )] ≤ (1 − β)h[∅, u1:T (∅)] + βh[S ⋆, u1:T (S ⋆ )] Define Ω = Σ−1 (S) + PT C̄ T C̄ , and Ω̄ =
t
t
i∈S i,t i,t
t|t−1
≤ (1 − β)h[∅, u1:T (∅)] + βκ,
(46) Σ−1 (S ∪{v})+ PT C̄ T C̄ ; using the Woodbury identity
i∈S i,t i,t
t|t−1
where the second inequality holds because S ⋆ is a solution to in Lemma 3,
Problem 2 and, as result, h[S ⋆, u1:T (S ⋆ )] ≤ κ. To complete the
−
ft (S ∪ {v}) = tr Θt Ω̄−1
t
proof, we recall that Algorithm 4 returns the set G = Gl when
−1 T
T −1
.
C̄v,t Ω̄−1
tr Θt Ω̄t C̄v,t (I + C̄v,t Ω̄−1
t
t C̄v,t )
for i = l it is the first time that h[Gi , u1:T (Gi )] ≤ κ. Therefore,
h[Gl−1 , u1:T (Gl−1 )] > κ and, as a result, there exists a real Therefore, for any time t ∈ {1, 2 . . . , T },
number ǫ > 0 such that h[Gl−1 , u1:T (Gl−1 )] = (1 + ǫ)κ, and
ft (S) − ft (S ∪ {v}) =
ineq. (46) gives
+
− tr Θt Ω̄−1
tr Θt Ω−1
t
t
(1 + ǫ)κ ≤ (1 − β)h[∅, u1:T (∅)] + βκ ⇒
−1 T −1
−1 T
≥
tr Θt Ω̄t C̄v,t (I + C̄v,t Ω̄t C̄v,t ) C̄v,t Ω̄−1
t
ǫκ ≤ (1 − β)h[∅, u1:T (∅)] − (1 − β)κ ⇒
−1
−1 T −1
−1 T
(48)
tr Θt Ω̄t C̄v,t (I + C̄v,t Ω̄t C̄v,t ) C̄v,t Ω̄t ,
ǫκ ≤ (1 − β){h[∅, u1:T (∅)] − κ} ⇒
⋆
where ineq. (48) holds because tr Θt Ω−1
≥ tr Θt Ω̄−1
ǫκ ≤ e−γg c(Gl−1 )/b {h[∅, u1:T (∅)] − κ} ⇒
t
t .
In particular, the inequality tr Θt Ω−1
≥ tr Θt Ω̄−1
t
t
ǫκ
log
≤ −γg c(Gl−1 )/b⋆ ⇒
is implied as follows: Lemma 6 implies Σ1|1 (S)
h[∅, u1:T (∅)] − κ
Σ1|1 (S ∪ {v}). Then, Corollary 2 implies Σt|t−1 (S)
1
h[∅, u1:T (∅)] − κ ⋆
Σt|t−1 (S ∪ {v}), and as a result, Lemma 1 implies
c(Gl−1 ) ≤
b ⇒
log
γg
ǫκ
Σt|t−1 (S)−1 Σt|t−1 (S ∪ {v})−1. Now, Σt|t−1 (S)−1
1
h[∅, u1:T (∅)] − κ ⋆
Σt|t−1 (S ∪ {v})−1 and the definition of Ωt and of Ω̄t
c(G) ≤ c(sl ) +
b,
log
γg
ǫκ
imply Ωt Ω̄t . Next, Lemma 1 implies Ω−1
Ω̄t−1.
t
As a result, since also Θt is a symmetric
matrix, Lem
where the latter holds since G = Gl−1 ∪ {sl }, due to
ma 4 gives the desired inequality tr Θt Ω−1
≥ tr Θt Ω̄−1
.
t
t
the definitions of G, Gl−1 , and sl , and since c(G) =
Continuing from the ineq. (48),
c(Gl−1 ) + c(sl ). Finally, since the definition of ǫ implies
b
ǫκ = h[Gl−1 , u1:T (Gl−1 )]−κ, and the definition of G is G = S,
ft (S) − ft (S ∪ {v}) ≥
the proof of ineq. (19) is complete.
−1 T −1
−1 T
≥
tr C̄v,t Ω̄−1
t Θt Ω̄t C̄v,t (I + C̄v,t Ω̄t C̄v,t )
Proof of Theorem 3’s part (2) (Algorithm 3’s running
−1 T
−1
−1 T −1
λmin ((I + C̄v,t Ω̄t C̄v,t ) )tr C̄v,t Ω̄t Θt Ω̄t C̄v,t , (49)
time): The proof is similar to the proof of Theorem 2’s part (2)
where ineq. (49) holds due to Lemma 2. From ineq. (49),
ft (S) − ft (S ∪ {v}) ≥
−1 T
−1
−1 T
= λ−1
max (I + C̄v,t Ω̄t C̄v,t )tr C̄v,t Ω̄t Θt Ω̄t C̄v,t
≥
=
λ−1
max (I
−1
λmax (I
+
+
T
C̄v,t Σt|t (∅)C̄v,t
)tr
T
C̄v,t Σt|t (∅)C̄v,t )tr
T
C̄v,t Ω̄t−1 Θt Ω̄t−1 C̄v,t
T
C̄v,t Ω̄t−1 ,
Θt Ω̄t−1 C̄v,t
(50)
where we used Ω̄−1
Σt|t (∅), which holds because of the
t
−1
following: the definition of Ω̄t implies Ω̄t Σt|t−1
(S ∪ {v}),
−1
and as a result, from Lemma 1 we get Ω̄t Σt|t−1 (S ∪{v}).
In addition, Corollary 2 and the fact that Σ1|1 (S ∪ {v})
Σ1|1 (∅), which holds due to Lemma 6, imply Σt|t−1 (S ∪
{v}) Σt|t−1 (∅). Finally, from eq. (23) in Lemma 5 it is
Σt|t−1 (∅) = Σt|t (∅). Overall, the desired inequality Ω̄−1
t
Σt|t (∅) holds.
Consider a time t′ ∈ {1, 2 . . . , T } such that for any time
−1
−1
−1 T
T
t ∈ {1, 2, . . . , T } it is Ω̄−1
t′ C̄v,t′ C̄v,t′ Ω̄t′ Ω̄t C̄v,t C̄v,t Ω̄t ,
−1 T
−1
and let Φ be the matrix Ω̄t′ C̄v,t′ C̄v,t′ Ω̄t′ ; similarly, let l be
T
the mint∈{1,2...,T },v∈V λ−1
max (I + C̄v,t Σt|t (∅)C̄v,t ). Summing
ineq. (50) across all times t ∈ {1, 2 . . . , T }, and using
Lemmata 4 and 2,
g(S) − g(S ∪ {v}) ≥ l
T
X
−1
T
tr Θt Ω̄−1
t C̄v,t C̄v,t Ω̄t
t=1
≥l
T
X
= ltr Φ
≥ lλmin
T
X
Θt
t=1
T
X
Θt
tr (Φ)
PT
which is non-zero because t=1 Θt ≻ 0 and Φ is a non-zero
positive semi-definite matrix.
Finally, we lower bound tr (Φ), using Lemma 2:
−1
T
tr (Φ) = tr Ω̄−1
t′ C̄v,t′ C̄v,t′ Ω̄t′
T
= tr Ω̄−2
t′ C̄v,t′ C̄v,t′
≥
=
≥
t=1
for some sensor set S ′ ⊆ V, and sensor v ∈ V; to upper bound
it, from eq. (23) in Lemma 5 of Appendix A, observe
X
T
′
C̄i,t
C̄i,t ]−1,
Σt|t (S ′ ∪ {v}) = [Σ−1
(S
∪
{v})
+
t|t−1
i∈S ′ ∪{v}
PT
′
T
and let Ht = Σ−1
i∈S ′ C̄i,t C̄i,t , and H̄t =
t|t−1 (S ) +
P
T
−1
T
C̄i,t ; using the Woodbury idenΣt|t−1 (S ′ ∪ {v}) + i∈S ′ C̄i,t
tity in Lemma 3,
ft (S ′ ∪ {v}) = tr Θt H̄t−1 −
T −1
T
) C̄v,t H̄t−1 .
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
Therefore,
T
X
T
λmin (Ω̄−2
t′ )tr C̄v,t′ C̄v,t′
T
λ2min (Ω̄−1
t′ )tr C̄v,t′ C̄v,t′
T
λ2min (Σt′ |t′ (V))tr C̄v,t
′ C̄v,t′ ,
[ft (S ′ ) − ft (S ′ ∪ {v})] =
t=1
T
X
T
X
!
[ft (S ′ ) − ft (S ′ ∪ {v})],
[tr Θt Ht−1 − tr Θt H̄t−1 +
T −1
T
) C̄v,t H̄t−1 ] ≤
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
!
t=1
T
X
t=1
tr (Θt Φ)
t=1
> 0,
b) Upper bound for the denominator of the supermodularity ratio γg : The denominator of the submodularity ratio γg
is of the form
(51)
where ineq. (51) holds because Ω̄−1
t′ Σt′ |t′ (V). In particular,
the inequality Ω̄−1
t′ Σt′ |t′ (S ∪ {v}) is derived by applying
T
T
C̄v,t
= Σt−1
Lemma 1 to the inequality Ω̄t′ Ω̄t′ + C̄v,t
′ |t′ (S ∪
{v}), where the equality holds by the definition of Ω̄t′ . In addition, due to Lemma 6 it is Σ1|1 (S ∪ {v}) Σ1|1 (V), and as
a result, from Corollary 1 it also is Σt′ |t′ (S ∪{v}) Σt′ |t′ (V).
Overall, the desired inequality Ω̄t−1
Σt′ |t′ (V) holds.
′
[tr Θt Ht−1 +
t=1
T −1
T
(52)
) C̄v,t H̄t−1 ],
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
where ineq. (52) holds since tr Θt H̄t−1 is non-negative. In
eq. (52), the second term in the sum is upper bounded as
follows, using Lemma 2:
T −1
T
) C̄v,t H̄t−1 =
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
T −1
T
≤
)
(I + C̄v,t H̄t−1 C̄v,t
tr C̄v,t H̄t−1 Θt H̄t−1 C̄v,t
−1 T −1
−1 T
−1
tr C̄v,t H̄t Θt H̄t C̄v,t λmax [(I + C̄v,t H̄t C̄v,t ) ] =
−1
T
T
)≤
λmin (I + C̄v,t H̄t−1 C̄v,t
tr C̄v,t H̄t−1 Θt H̄t−1 C̄v,t
−1
−1 T
−1
T
), (53)
tr C̄v,t H̄t Θt H̄t C̄v,t λmin (I + C̄v,t Σt|t (V)C̄v,t
T
T
) ≥ λmin (I + C̄v,t Σt|t (V)C̄v,t
),
since λmin (I + C̄v,t H̄t−1 C̄v,t
−1
−1
because H̄t Σt|t (V). In particular, the inequality H̄t
T
Σt|t (V) is derived as follows: first, it is H̄t H̄t + C̄v,t
C̄v,t =
′
−1
Σt|t (S ∪ {v}) , where the equality holds by the definition of
H̄t , and now Lemma 1 implies H̄t−1 Σt|t (S ′ ∪ {v}). In addition, Σt|t (S ′ ∪ {v}) Σt|t (V) is implied from Corollary 1,
since Lemma 6 implies Σ1|1 (S ′ ∪ {v}) Σ1|1 (V). Overall,
the desired inequality H̄t−1 Σt|t (V) holds.
T
Let l′ = maxt∈{1,2...,T },v∈V λ−1
min (I + C̄v,t Σt|t (V)C̄v,t ).
From ineqs. (52) and (53),
PT
[ft (S ′ ) − ft (S ′ ∪ {v})] ≤
Pt=1
(54)
T
−1
T
C̄v,t H̄t−1 ].
+ l′ tr Θt H̄t−1 C̄v,t
t=1 [tr Θt Ht
Consider times t′ ∈ {1, 2 . . . , T } and t′′ ∈ {1, 2 . . . , T } such
that for any time t ∈ {1, 2, . . . , T }, it is Ht−1
Ht−1 and
′
−1 T
−1
−1
−1 T
H̄t′′ C̄v,t′′ C̄v,t′′ H̄t′′ H̄t C̄v,t C̄v,t H̄t , and let Ξ = Ht−1
′
−1
T
and Φ′ = H̄t−1
′ C̄v,t′ C̄v,t′ H̄t′ . From ineq. (54), and Lemma 4,
T
X
t=1
T
X
[ft (S ′ ) − ft (S ′ ∪ {v})] ≤
since xt+1 = At xt = At At−1 xt−1 = . . . = At At−1 · · · A1 x1
when all u1 , u2 , . . . , uT are zero.
From eqs. (59) and (60), the inequality
[tr (Θt Ξ) + l′ tr (Θt Φ′ )] ≤
t=1
tr Ξ
T
X
Θt
t=1
!
′
+ l tr Φ
′
T
X
Θt
t=1
T
X
(tr (Ξ) + l′ tr (Φ′ ))λmax (
!
≤
T
xT
1 N1 x1 < x1
(55)
Finally, we upper bound tr (Ξ)+l tr (Φ′ ) in ineq. (55), using
Lemma 2:
T
≤
l′ λ2max (H̄t−1
′′ )tr C̄v,t′′ C̄v,t′′
T
tr Σt′ |t′ (∅) + l′ λ2max (Σt′′ |t′′ (∅))tr C̄v,t
′′ C̄v,t′′ ,
holds for any non-zero x1 if and only if
N1 ≺
(57)
A PPENDIX F
P ROOF OF T HEOREM 5
For the proof of Theorem 5, we use Lemmata 16-19 below.
Lemma 16 (System-level condition for near-optimal co-design). Let N1 be defined as in eq. (9). The control policy
u◦1:T , (0, 0, . . . , 0) is suboptimal for the LQG problem in
eq. (22) for all non-zero initial conditions x1 if and only if
PT
T
T
(58)
t=1 A1 · · · At Qt At · · · A1 ≻ N1 .
Proof of Lemma 16: For any initial condition x1 , eq. (28)
in Lemma 10 implies for the noiseless perfect state information
LQG problem in eq. (22):
t=1
T
AT
1 · · · At Qt At At−1 · · · A1 .
Θt = AT
t St At + Qt−1 − St−1 .
(56)
where ineq. (57) holds because
Σt′ |t′ (∅), and similarly, H̄t−1
Σt′′ |t′′ (∅). In particular, the inequality Ht−1
′′
′
Σt′ |t′ (∅) is implied as follows: first, by the definition of Ht′ ,
it is Ht−1
= Σt′ |t′ (S ′ ); and finally, Corollary 1 and the fact
′
that Σ1|1 (S ′ ) Σ1|1 (∅), which holds due to Lemma 6, imply
Σt′ |t′ (S ′ ) Σt′ |t′ (∅). In addition, the inequality H̄t−1
′′
Σt′′ |t′′ (∅) is implied as follows: first, by the definition of H̄t′′ ,
′
it is H̄t′′ Σ−1
t′′ |t′′ −1 (S ∪ {v}), and as a result, Lemma 1
−1
implies H̄t′′ Σt′′ |t′′ −1 (S ′ ∪ {v}). Moreover, Corollary 2
and the fact that Σ1|1 (S ∪ {v}) Σ1|1 (∅), which holds due to
Lemma 6, imply Σt′′ |t′′ −1 (S ′ ∪ {v}) Σt′′ |t′′ −1 (∅). Finally,
from eq. (23) in Lemma 5 it is Σt′′ |t′′ −1 (∅) = Σt′′ |t′′ (∅).
Overall, the desired inequality H̄t−1
′′ Σt′′ |t′′ (∅) holds.
u1:T
T
X
t=1
Proof of Lemma 17: Using the Woobury identity in
Lemma 3, and the notation in eq. (9),
−1
Nt = AT
+ Bt Rt−1 BtT )−1 At
t (St
Ht−1
′
T
X
T
T
AT
1 A2 · · · At Qt At At−1 · · · A1 x1
Lemma 17. For any t = 1, 2, . . . , T ,
tr (Ξ) + l′ tr (Φ′ ) ≤
tr Ht−1
+
′
min
T
X
t=1
Θt ).
t=1
′
since E(kx1 k2N1 ) = xT
1 N1 x1 , because x1 is known (Σ1|1 = 0),
and Σt|t and Wt are zero. In addition, for u1:T = (0, 0, . . . , 0),
the objective function in the noiseless perfect state information
LQG problem in eq. (22) is
PT
2
2
t=1 [kxt+1 kQt +kut (xt )kRt ] Σt|t =Wt =0
PT
(60)
= t=1 xT
Qx
PT t+1 T t Tt+1 T
= xT
Q
A
A
·
·
·
A
x
,
·
·
·
A
A
A
t t t−1
1 1
t
1
t=1 1 2
[kxt+1 k2Qt +kut (xt )k2Rt ]
Σt|t =Wt =0
= xT
1 N1 x1 ,
(59)
−1 T
= AT
t (St − St Bt Mt Bt St )At
= AT
t St At − Θt .
The latter, gives Θt = AT
t St At −Nt . In addition, from eq. (9),
−Nt = Qt−1 − St−1 , since St = Qt + Nt+1 .
P
T
T
Lemma 18. Tt=1 AT
1 A2 · · · At Qt At At−1 · · · A1 ≻ N1 if
and only if
T
X
T
T
AT
1 A2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0.
t=1
Proof of Lemma 18: For i = t − 1, t − 2, . . . , 1, we preand post-multiply the identity in Lemma 17 with AT
i and Ai ,
respectively:
Θt = AT
t St At + Qt−1 − St−1 ⇒
T
T
T
AT
t−1 Θt At−1 = At−1 At St At At−1 + At−1 Qt−1 At−1 −
T
At−1 St−1 At−1 ⇒
T
T
T
AT
t−1 Θt At−1 = At−1 At St At At−1 + At−1 Qt−1 At−1 −
Θt−1 + Qt−2 − St−2 ⇒
T
T
Θt−1 + AT
t−1 Θt At−1 = At−1 At St At At−1 +
T
At−1 Qt−1 At−1 + Qt−2 − St−2 ⇒
... ⇒
T
T
Θ2 + AT
2 Θ3 A2 + . . . + A2 · · · At−1 Θt At−1 · · · A2 =
T
T
T
A2 · · · At St At · · · A2 + A2 · · · AT
t−1 Qt−1 At−1 · · · A2 +
. . . + AT
2 Q2 A2 + Q1 − S1 ⇒
T
T
Θ1 + AT
1 Θ2 A1 + . . . + A1 · · · At−1 Θt At−1 · · · A1 =
T
T
T
A1 · · · At St At · · · A1 + A1 · · · AT
t−1 Qt−1 At−1 · · · A1 +
. . . + AT
Q1 A1 − N1 ⇒
1
PT
AT · · · AT
t−1 Θt At−1 · · · A1 =
t=1
PT1
T
T
A
·
t=1 1 · · At Qt At · · · A1 − N1 .
(61)
The last equality in eq. (61) implies Lemma 18.
Lemma 19.
for any t = 1, 2, . . . , T that At is
PConsider
T
T T
T
A
A
invertible.
t=1 1 2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0 if
and only if
T
X
Θt ≻ 0.
t=1
Proof of Lemma 19: Let Ut = At−1 At−2 · · · A1 .
We Pfirst prove that for any non-zero vector z, if
T
T T
T
it is
t=1 A1 A2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0, then
PT
T
t=1 z Θt z > 0. In particular, since Ut is invertible, —
because for any t ∈ {1, 2, . . . , T }, At is,—
PT
PT
T
T −⊤ T
Ut Θt Ut Ut−1 z
t=1 z Θt z = Pt=1 z Ut
(62)
T
T T
= t=1 tr φt φt Ut Θt Ut ,
where we let φt = Ut−1 z. Consider a time t′ such that for any
T
time t ∈ {1, 2 . . . , T }, φt′ φT
t′ φt φt . From eq. (62), using
Lemmata 4 and 2,
T
X
z T Θt z ≥
t=1
T
X
T
tr φt′ φT
t′ Ut Θt Ut
t=1
≥ tr
φt′ φT
t′
T
X
UtT Θt Ut
t=1
!
T
X
λ
(
UtT Θt Ut )
≥ tr φt′ φT
′
min
t
t=1
T
X
UtT Θt Ut )
= kφt′ k22 λmin (
t=1
> 0.
finally provePthat for any non-zero vector z, if
PWe
T
T
T
T
Θ
t=1 zA1 · · · At−1 Θt At−1 · · · A1 z ≻ 0.
t=1 t ≻ 0, then
In particular,
T
X
t=1
z T UtT Θt Ut z =
T
X
t=1
tr ξtT Θt ξt ,
(63)
where we let ξt = Ut z. Consider time t′ such that for any
time t ∈ {1, 2 . . . , T }, ξt′ ξtT′ ξt ξtT . From eq. (62), using
Lemmata 4 and 2,
!
T
T
X
X
T
T
Θt
tr ξt Θt ξt ≥ tr ξt′ ξt′
t=1
t=1
T
X
Θt )
≥ tr ξt′ ξtT′ λmin (
t=1
T
X
Θt )
= kξt′ k22 λmin (
t=1
> 0.
Proof of Theorem 5: Theorem 5 follows from the
sequential application of Lemmata 16, 18, and 19.
| 3cs.SY
|
arXiv:1711.10761v1 [cs.NE] 29 Nov 2017
Transfer Learning with Binary Neural Networks
Sam Leroux • Steven Bohez • Tim Verbelen • Bert Vankeirsbilck
Pieter Simoens • Bart Dhoedt
Ghent University - imec, IDLab, Department of Information Technology
Ghent University
Ghent, Belgium
[email protected]
Abstract
Previous work has shown that it is possible to train deep neural networks with
low precision weights and activations. In the extreme case it is even possible to
constrain the network to binary values. The costly floating point multiplications are
then reduced to fast logical operations. High end smart phones such as Google’s
Pixel 2 and Apple’s iPhone X are already equipped with specialised hardware for
image processing and it is very likely that other future consumer hardware will also
have dedicated accelerators for deep neural networks. Binary neural networks are
attractive in this case because the logical operations are very fast and efficient when
implemented in hardware. We propose a transfer learning based architecture where
we first train a binary network on Imagenet and then retrain part of the network
for different tasks while keeping most of the network fixed. The fixed binary part
could be implemented in a hardware accelerator while the last layers of the network
are evaluated in software. We show that a single binary neural network trained on
the Imagenet dataset can indeed be used as a feature extractor for other datasets.
1
Introduction
Deep learning really took off in 2012 when Krizhevsky et al. showed record breaking results on the
Imagenet dataset [1]. They demonstrated that deep convolutional neural networks trained end to
end on large labelled datasets can beat most other techniques for image recognition. Deep learning
quickly became the default algorithm for image classification and now even achieves super-human
level performance [2]. Deep learning has also revolutionized other fields like speech recognition and
natural language processing [3].
The two key ingredients needed to successfully apply deep neural networks are large amounts of
labelled training data and powerful computing systems such as GPUs. Mobile devices including
smartphones, Internet-of-Things (IoT) devices or smart home assistants have very limited processing
power because of their intrinsic limitations on size and energy consumption. One possible solution is
to offload all computations to the cloud but this introduces a latency and potentially even a privacy
risk when sensitive data is processed remotely.
There is a considerable amount of active research on techniques to reduce the computational cost of deep learning models. One approach is to prune the network by removing redundant
weights. The idea of pruning already goes back to the eighties when LeCun et al. used second order
derivatives to calculate the impact of each weight on the loss of the network [4]. Weights with a small
impact are then removed from the network. More recently pruning was used on modern deep neural
networks. Song et al. proposed a pruning pipeline where first weights with a small magnitude where
removed after which the network was fine-tuned to recover the lost accuracy [5]. Network pruning is
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. Machine
Learning on the Phone and other Consumer Devices workshop
especially effective when a network is used for transfer learning. In transfer learning the model is first
trained on a large dataset such as Imagenet and is then fine-tuned on a small domain specific dataset.
Because the network was pretrained on a general dataset it will contain convolutional kernels that
are not useful for the domain specific dataset. Molchanov et al. introduced a criterion based on a
first-order Taylor expansion to decide which feature maps to remove and demonstrated impressive
results when used together with transfer learning [6].
Instead of compressing a trained model it is also possible to train efficient models from
scratch. The recently introduced MobileNets [7] use depthwise separable convolutions to reduce
the computational cost. Depthwise separable convolutions factorize a standard convolution into a
depthwise convolution and a 1x1 pointwise convolution. The depthwise convolution applies a single
convolution to each input channel while the pointwise convolution combines the information in the
different channels. Factorizing a traditional convolution into these two convolutions dramatically
reduces the computational cost and size of the network with only a minimal reduction in accuracy.
Most implementations of deep neural networks use 32 bit floating point numbers for weights and
activations. Various works have shown that this is not necessary and that it is possible to use 16 bit
[8] or 8 bit [9] numbers. In the extreme case it is even possible to use binary weights and activations.
Courbariaux et al. successfully trained convolutional neural networks for image recognition with
binary weights and activations [10]. This works surprisingly well for small scale datasets such as
CIFAR10 but there is still a large drop in accuracy for larger datasets such as Imagenet. Neural
networks with binary weights and activations are attractive because they replace the costly floating
point multiplications and additions with bitwise XNORs and left and right bit shifts. These operations
are very efficient to implement in hardware.
Another problem with deep learning is the need for large labelled datasets. Training a new
model from scratch requires a large amount of training data. A well known technique is to use
transfer learning where a model is first trained on a large dataset like Imagenet and afterwards the last
layer is removed and retrained using a small amount of new domain specific training data. Transfer
learning works because the first layers in the network learn to detect features such as color transitions
and basic shapes that are present in images from different domains [11].
2
A hybrid hardware-software architecture
We propose a hybrid hardware-software architecture based on this idea of transfer learning. We train
a neural network with binary weights and activations on the Imagenet dataset and use this network as
a fixed feature extractor that could be optimised on hardware level. The last layer is implemented
in software and is evaluated on the CPU (or GPU) of the device. Most of the computations are
offloaded to the custom circuit. Since we only need to retrain the last layer of the network it even
becomes feasible to train on the device itself instead of offloading the training to the cloud. This is
very attractive from a privacy point of view because the training data never leaves the mobile device.
Our architecture also allows app developers to embed custom neural networks into their apps. Right
now this is often impossible because a typical neural network quickly requires hundreds of megabytes
of storage just for the weights. If the device however is equipped with the fixed feature extractor we
only need to ship the last (domain specific) layer of the network with the app. A schematic overview
of our architecture is shown in Figure 1.
0
0.5
1
0
1
-0.1
0
Software
0.2
1
1
Hardware
0.9
1
Figure 1: Schematic overview of our proposed architecture. Most layers of a binary neural network
trained on Imagenet are embedded in a specialised circuit in hardware. This circuit is used as a fixed
feature extractor. The last layer is implemented in software and retrained for different applications.
2
3
Experiments
In this section we show that it is possible to use a binary neural network trained on Imagenet as a
feature extractor for other datasets. We trained a binary version of the Alexnet architecture [1] on
the ILSVRC2012 training set following the binarization technique of Courbariaux et al. [10]. Our
binary network obtains a top 5 accuracy of 67% (42% top 1) while the floating point Alexnet obtains
a top 5 accuracy of 80% (57% top 1). We then retrained the last layer on the domain specific datasets
without changing the weights of the other layers. We report results for the three fine-grained image
datasets summarized in Table 1. For all our models we resized the input images such that the longest
side was of length 256. During training we took random 224 by 224 pixel crops. For the test set we
used center crops. The floating point networks where trained with stochastic gradient descent with
momentum. For the binary networks we found that Adam [12] gives better results which is consistent
with Courbariaux et al. [10].
Table 1: The different datasets used in our experiments. We pretrained our network on the Imagenet
dataset and then retrained the last layer on the three smaller domain specific datasets.
Dataset
Number of classes
Training images
Testing images
ILSVRC2012
1000
1,200,000
50,000
Flowers [13]
UCSD Birds [14]
MIT Indoor scenes [15]
102
200
67
6,149
5,994
5.360
1,020
5,794
1,340
The results are summarized in Table 2. The first part of the table shows the accuracies for the baseline
floating point networks, either trained from scratch on the domain specific datasets or fine-tuned from
the Imagenet model. As expected the fine-tuned models consistently outperform the models trained
from scratch. The second part of the table shows the results for our binary models. The fine-tuned
models again outperform the models trained from scratch which shows that transfer learning also
works when the network uses binary weights and activations. For the last layer we can use binary
weights and activations (a) but because this layer is evaluated in software in our proposed architecture
we also experimented with floating point weights and activations (b). This consistently increases the
accuracy on all datasets and pushes the accuracy closer to the accuracy of the floating point models.
Table 2: The top 1 and top 5 accuracies for each domain specific dataset. We report results for the
binary network trained from scratch and for the fine-tuned networks, both with binary weights in the
last layer (a) and floating point weights (b). We compare to a floating point baseline.
4
Flowers
Top 1
Top 5
Birds
Top 1
Top 5
Indoor scenes
Top 1
Top 5
FP
trained from scratch
fine-tuned
71.7%
85.4%
90.5%
97.2%
32.4%
50.4%
57.2%
79.6%
29.7%
57.4%
56.4%
86.8%
Binary
trained from scratch
(a) fine-tuned binary last layer
(b) fine-tuned FP last layer
63.2%
80.6%
84.0%
85.1%
94.5%
95.5%
13.8%
33.9%
36.7%
32.4%
62.9%
63.4%
29.0%
46.7%
48.3%
54.9%
77.4%
78.6%
Conclusion and future work
We introduced a hybrid hardware-software architecture where a binary neural network trained on
Imagenet can be embedded in a dedicated circuit. The last layer is implemented in software and is
retrained for each specific task. We showed that transfer learning works very well for binary neural
networks and experimented with a hybrid binary-floating point network where only the last layer uses
floating point operations. This is a good trade-off between accuracy and computational cost since
most of the computations can be offloaded to the fixed hardware accelerator. In future work we will
explore other hybrid architectures. We now only considered finetuning the last layer of the network
but we can also retrain more than one layer which should give a higher accuracy but will also incur a
higher computational cost since a larger part of the network is evaluated in software.
3
Acknowledgments
We gratefully acknowledge the support of NVIDIA Corporation with the donation of GPU hardware
used for this research. Steven Bohez is funded by a Ph.D. grant of the Agency for Innovation by
Science and Technology in Flanders (IWT).
References
[1] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Advances in neural information processing systems, pages
1097–1105, 2012.
[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers:
Surpassing human-level performance on imagenet classification. CoRR, abs/1502.01852, 2015.
[3] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444,
2015.
[4] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In D. S. Touretzky, editor, Advances in Neural Information Processing Systems 2, pages 598–605. Morgan-Kaufmann,
1990.
[5] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections
for efficient neural network. In Advances in Neural Information Processing Systems, pages
1135–1143, 2015.
[6] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional
neural networks for resource efficient transfer learning. arXiv preprint arXiv:1611.06440, 2016.
[7] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias
Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural
networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
[8] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia,
Boris Ginsburg, Michael Houston, Oleksii Kuchaev, Ganesh Venkatesh, et al. Mixed precision
training. arXiv preprint arXiv:1710.03740, 2017.
[9] Vincent Vanhoucke, Andrew Senior, and Mark Z Mao. Improving the speed of neural networks
on cpus. In Proc. Deep Learning and Unsupervised Feature Learning NIPS Workshop, volume 1,
page 4, 2011.
[10] Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks: Training deep neural networks with weights and activations constrained to+ 1
or-1. arXiv preprint arXiv:1602.02830, 2016.
[11] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in
deep neural networks? In Advances in neural information processing systems, pages 3320–3328,
2014.
[12] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[13] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of
classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image
Processing, Dec 2008.
[14] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD
Birds 200. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010.
[15] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In Computer Vision and
Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 413–420. IEEE, 2009.
4
| 1cs.CV
|
QUANTILE GRAPHICAL MODELS: PREDICTION AND CONDITIONAL
INDEPENDENCE WITH APPLICATIONS TO SYSTEMIC RISK
arXiv:1607.00286v2 [math.ST] 5 Dec 2017
ALEXANDRE BELLONI∗ , MINGLI CHEN‡ , AND VICTOR CHERNOZHUKOV†
Abstract. The understanding of co-movements, dependence, and influence between variables of interest is
key in many applications. Broadly speaking such understanding can lead to better predictions and decision
making in many settings. We propose Quantile Graphical Models (QGMs) to characterize prediction and
conditional independence relationships within a set of random variables of interest. Although those models
are of interest in a variety of applications, we draw our motivation and contribute to the financial risk
management literature. Importantly, the proposed framework is intended to be applied to non-Gaussian
settings, which are ubiquitous in many real applications, and to handle a large number of variables and
conditioning events.
We propose two distinct QGMs. First, Condition Independence Quantile Graphical Models (CIQGMs)
characterize conditional independence at each quantile index revealing the distributional dependence structure. Second, Prediction Quantile Graphical Models (PQGMs) characterize the best linear predictor under
asymmetric loss functions. A key difference between those models is the (non-vanishing) misspecification
between the best linear predictor and the conditional quantile functions.
We also propose estimators for those QGMs. Due to high-dimensionality, the two distinct QGMs require
different estimators. The estimators are based on high-dimensional techniques including (a continuum of)
`1 -penalized quantile regressions (and low biased equations), which allow us to handle the potential large
number of variables. We build upon a recent literature to obtain new results for valid choice of the penalty
parameters, rates of convergence, and confidence regions that are simultaneously valid.
We illustrate how to use QGMs to quantify tail interdependence (instead of mean dependence) between
a large set of variables which is relevant in applications concerning with extreme events. We show that
the associated tail risk network can be used for measuring systemic risk contributions. We also apply the
framework to study international financial contagion and the impact of market downside movement on the
dependence structure of assets’ returns.
Key words: High-dimensional approximately sparse model, tail risk network, conditional independence,
nonlinear correlation, penalized quantile regression, systemic risk, financial contagion, downside movement
Date: First version: November 2012, this version December 6, 2017.
We would like to thank Don Andrews, Debopam Bhattacharya, Peter Bickel, Marianne Bitler, Peter Bühlmann, Colin
Cameron, Xu Cheng, Valentina Corradi, Alan Crawford, Francis Diebold, Iván Fernández-Val, Bulat Gafarov, Kenji Fukumizu,
Dalia Ghanem, Bryan Graham, Jiaying Gu, Ran Gu, Jungbin Hwang, Cheng Hsiao, Hidehiko Ichimura, Michael Jansson,
Oscar Jorda, Chihwa Kao, Kengo Kato, Shakeeb Khan, Roger Koenker, Chenlei Leng, Arthur Lewbel, Michael Leung, Song Liu,
Whitney Newey, Dong Hwan Oh, David Pacini, Andrew Patton, Aureo de Paula, Victor de la Pena, Elisabeth Perlman, Stephen
Portnoy, Demian Pouzo, James Powell, Geert Ridder, Stephen Ross, Shu Shen, Senay Sokullu, Sami Stouli, Aleksey Tetenov,
Takuya Ura, Tiemen Woutersen, Zhijie Xiao, Wenkai Xu, and Yanos Zylberberg for comments and discussions. We would
also like to thank the seminar and workshop participants from Boston College, 2015 Warwick Summer Workshop, 11th World
Congress of the Econometric Society, 2017 UCL Workshop on the Theory of Big Data, 2017 California Econometrics Conference,
University of Bristol, University of Connecticut, Humboldt University Berlin, the Institute of Statistical Mathematics, USC,
UC Berkeley, UC Davis, Warwick Statistics Department and University of Tokyo.
∗ Duke University, e-mail:[email protected].
‡ University of Warwick, e-mail:[email protected].
† Massachusetts Institute of Technology, e-mail:[email protected].
1
2
1. Introduction
The understanding of co-movements, dependence and influence between variables of interest is key in
many applications. Such understanding can lead to better predictions and decision making in many settings.
This is clearly of interest in financial management settings where quantifying risk is crucial. For example,
the dependence between stock returns plays a key role in hedging strategies. Those hedging decisions are
typically focused on the tail of the distribution of returns rather than the mean. Moreover, such strategies
aiming to reduce risk are critical precisely during market downside movement. Therefore, it is also instructive
to understand how the dependence (and policies) would change as the downside movement of the market
becomes more extreme. Moreover, recent empirical evidence [7, 6, 66] points to non-Gaussianity of the
distribution of stock returns, especially during market downturns. These issues require models that allow for
non-Gaussian settings and can also accommodate various conditioning events (e.g., downside movements).
Motivated by those features, this work proposes Quantile Graphical Models (QGMs) to characterize (and
visualize) the dependence structure of a set of random variables. The proposed framework allows us to
understand prediction and conditional independence between those variables. Moreover, it also enables us
to focus on specific parts of the distributions of those variables such as tail events. Such understanding
plays an important role in applications like financial contagion and measuring systemic risk contributions
where extreme events are the main interests for regulators. Our techniques are intended to be applicable to
high-dimensional settings where the total number of variables (or additional conditioning variables) is large
– possibly larger than the sample size.
QGMs provide an alternative route to learn conditional independence and prediction under asymmetric
loss functions which is appealing in non-Gaussian settings. In the Gaussian setting, those notions essentially
coincide; however, in non-Gaussian settings different estimation approaches are needed. Conditional independence hinges on the equivalence between conditional probabilities and conditional quantiles. Prediction
under asymmetric loss function hinges on the solution of a quantile regression with non-vanishing misspecification (M -estimation problem). Although we build upon the quantile regression literature ([50, 18]), we
derive new results for penalized quantile regression in high dimensional settings to handle misspecification,
many controls and a continuum of additional conditioning events.
Conditional independence has a long history in statistical models with consequences towards parameter
identification, causal inference, prediction sufficiency, and many others, see [32]. Conditional Independence
Quantile Graphical Models (CIQGMs) aim to characterize conditional independence via the conditional
quantile functions. In such models we consider a flexible specification that can approximate well the conditional quantile functions (up to a vanishing approximation error). In turn, this allows to detect which
variables have a strong or near zero impact on others which can then be used to provide guidance on
conditional independence.
Prediction Quantile Graphical Models (PQGMs) focus on prediction of a variable based on the other
variables (a reduced form relation). An important motivation for proposing PQGMs is to allow for misspecification as the conditional quantile function is typically non-linear in non-Gaussian settings but a linear
specification is widely used in practice. Misspecification in quantile regression models was first properly
justified by [8] in which it was shown that we can recover a suitable “best approximation” for the conditional
quantile function. We directly characterize the good prediction properties under asymmetric loss functions
QUANTILE GRAPHICAL MODELS
3
which is appealing in empirical applications. Other papers investigated the impact of misspecification on
the specification of the quantile function are [8], [49], [55] and [1]. Nonetheless, this work seems to be the
first to accommodate non-vanishing misspecification for high-dimensional quantile regression.
Broadly speaking QGMs enhance our understanding of statistical dependence among the components of a
d-dimensional random vector XV , where the set V contains the labels of the components. QGMs will provide
a way to visualize such dependence via graphs where nodes represent components of XV and edges represent
conditional relationships (as in Gaussian Graphical Models, see below). Given that for each specific quantile
index τ we will obtain one such graph, we could have a graph process indexed by τ ∈ (0, 1). The structure
represented by the τ -quantile graph represents a local relation and can be valuable in applications where
the tail interdependence (corresponding to high or low quantile index) might be of special interest. This is
akin to the contrast between quantile regression and linear regression, where the latter provides information
only on the conditional mean, while the former can provide a more complete description of the distribution
of the outcome.
The graph process induced by QGMs has several important features. First, a τ -quantile graph enables
different values of edge strength in different directions. This is important because for undirected networks,
the distinction between exposure and contribution is unclear. Second, QGMs are able to capture the tail interdependence through estimating at a high or low quantile index. Third, QGMs can capture the asymmetric
dependence structure at different quantiles, which can be particularly useful in empirical applications (e.g.,
stock market dependence, exchange rate dependence). By considering all the quantiles at once we can characterize conditional independence structure for a set of variables which are not jointly Gaussian distributed,
i.e. the case where the covariance matrix cannot characterize conditional independence completely.
A key feature of our work is it provides estimation procedures to learn QGMs from the data observed. The
estimators are geared to cover high-dimensional settings where the size of the model is large relative to the
sample size. Those estimators are based on `1 -penalized quantile regression and low biased equations. For
CIQGMs, under mild regularity conditions, we provide rates of convergence and edge properties of the estimated graph that hold uniformly over a large class of data generating processes. We provide simultaneously
valid confidence regions (post-selection) for the coefficients of the CIQGM that are uniformly valid, despite of
possible model selection mistakes. Furthermore, based on proper thresholding, recovery of the QGMs pattern
is possible when coefficients are well separated from zero which parallel the results for graph recovery in the
Gaussian case based on the estimation of the precision matrix. 1 For PQGMs, we provide an estimator that
achieves an adaptive rate of convergence which might differ under different conditioning events. Therefore
we contribute to the recent literature on simultaneous valid confidence regions post-model selection that has
been an active research area in econometrics [10, 17, 16, 41, 23, 28] and statistics [71, 80, 19, 18, 45, 64, 74];
in particular, the penalty choices and theoretical results are uniformly valid and adaptive to the relevant
conditioning events.
QGMs can play important roles in applications where tail events are relevant. For example, with certain
rescaling of the edge weights, we are able to capture the importance of each node or measure its systemic risk
contribution. In parallel with [5], many approaches for measuring systemic risk fit naturally into QGMs. For
1Similar to graph recovery in the Gaussian case such exact recovery is subject to the lack of uniformity validity critiques of
Leeb and Pötscher [56].
4
BELLONI, CHEN, AND CHERNOZHUKOV
b|a
example, one can view ∆CoV aRτ , a measure of the impact of institution a on institution b, as the weight
of an edge of a τ -quantile graph. Then, the systemic risk contribution of institution a, equals to the sum
P
sys|a
b|a
of coefficients over b ∈ V , i.e. ∆CoV aRτ
= b∈V ∆CoV aRτ . Similarly, the summation over a ∈ V
measures exposures of individual institution to systemic shocks from the network. QGMs can also be used
to study contagion and network spillover effects since it is useful for characterizing tail risk spillovers. We
apply QGM to the study of international financial contagion in volatilities, specializing in learning the risk
transmission channels, see [30] for an overview of international financial contagions. After learning the risk
transmission channels, we can use our new network-cooperated ∆CoV aR to measure the contribution and
exposure of each country to the whole market.
Our work is complementary to a large body of works that study a set of jointly Gaussian distributed
random variables. Gaussian Graphical Model (GGM) provides a graphical model representation of those
variables. It is well known that conditional independence structure is completely characterized by the support
of the precision matrix (i.e. the inverse of the covariance matrix) of the random variables of interest, hence
recovering the structure of an undirected Gaussian graph is equivalent to recovering the support of the
precision matrix, [33, 53, 31, 38]. Methods work with the precision matrix (or covariance matrix) using
hypothesis testing can be find in [38, 35, 36, 37]. Recently, due to the advance of regularization techniques,
high-dimensional GGMs have been extensively studied in the statistical and machine learning literature:
[62] propose neighborhood selection, i.e. using LASSO for each node in the graph and combine the results
column-by-column; [79, 9, 43] penalize the log-likelihood function hence work with the precision matrix
directly; other refined estimators including [78, 22, 60, 69, 59, 29]. [58] extended the result to a more
general class of models called nonparanormal models or semiparametric Gaussian copula models, i.e., the
variables have a multivariate Gaussian distribution after a set of unknown monotone transformations (see
also [57, 76, 77]). However, those methods assume the (transformed) random vectors follows a joint Gaussian
distribution. In addition, they characterize the conditional mean predictability by linear combinations of the
other variables.
Our work also contributes to a growing literature that rely on quantile based models to characterize
the data generating process. [81] considers a globally adaptive quantile regression model, establishes oracle
properties and improved rates of convergence for the high-dimensional case. Screening procedures based
on moment conditions motivated by the quantile models have been proposed and analyzed in [44] and [75]
in the high-dimensional case. [46] considers tail dependence defined via conditional probabilities in a low
dimensional setting. Several other quantile based models have been proposed, see e.g. [50]. Among the
contributions of this work is to consider a high-dimensional setting and propose techniques that can be
robust to small coefficients (i.e. allowing for model selection mistakes), non-vanishing misspecification in the
conditional quantile function, and uniformly valid over additional conditioning events.
The rest of the paper is organized as follows. Section 2 lays out the foundation of QGMs. Section 3
contains estimators for QGMs while Section 4 contains the theoretical guarantees of the estimators. Section
5 provides empirical applications of QGMs to measure systemic risk contribution and to hedging conditional
on the downside movements of the US stock market. Finally, the appendix contains proofs, simulations and
implementation details of the estimators.
QUANTILE GRAPHICAL MODELS
5
Notation. For an integer k, we let [k] := {1, . . . , k} denote the set of integers from 1 to k. For a random
variable X we denote by X its support. We use the notation a ∨ b = max{a, b} and a ∧ b = min{a, b}. We
use kvkp to denote the p-norm of a vector v. We denote the `0 -“norm” by k · k0 (i.e., the number of nonzero
components). Given a vector δ ∈ Rp , and a set of indices T ⊂ {1, ..., d}, we denote by δT the vector in
Pn
which δT j = δj if j ∈ T , δT j = 0 if j ∈
/ T . We use En to abbreviate the notation n−1 i=1 ; for example,
Pn
En [f ] := En [f (ωi )] := n−1 i=1 f (ωi ).
2. Quantile Graphical Models
In this section we describe quantile graphical models associated with a d-dimensional random vector XV
where the set V = [d] = {1, . . . , d} denotes the labels of the components. These models aim to provide
a description of the dependence between the random variables in XV . In particular, these models induce
graphs that allow for visualizing dependence structures. Nonetheless, because of the non-Gaussianity, we
consider two fundamentally distinct models (one geared towards conditional independence and one geared
towards prediction).
2.1. Conditional Independence Quantile Graphical Models. Conditional independence graphs have
been used to provide visualization and insight on the dependence structure between random variables. Each
node of the graph is associated with a component of XV . We denote the conditional independence graph as
GI = (V, E I ) where GI is an undirected graph with vertex set V and edge set E I which is represented by an
I
I
adjacency matrix (Ea,b
= 1 if the edge (a, b) ∈ GI , and Ea,b
= 0 otherwise). An edge (a, b) is not contained
in the graph if and only if
Xa ⊥ Xb | XV \{a,b} ,
(2.1)
namely Xb and Xa are independent conditional on all remaining variables XV \{a,b} = {Xk ; k ∈ V \{a, b}}.
Comment 2.1 (Conditional Independence Under Gaussianity). In the case that XV is jointly Gaussian
distributed, XV ∼ N (0, Σ) with Σ as the covariance matrix of XV , the conditional independence structure
between two components is determined by the inverse of the covariance matrix, i.e. the precision matrix
Θ = Σ−1 . It follows that the nonzero elements in the precision matrix corresponds to the nonzero coefficients
of the associated (high dimensional) mean regression. The family of Gaussian distributions with this property
is known as a Gauss-Markov random field with respect to the graph G. This observation has motivated a
large literature [53] and interesting extensions that allow for transformations of Gaussian variables [58, 57].
In order to achieve a tractable concept for non-Gaussian settings, we use that (2.1) occurs if and only if
FXa (·|XV \{a} ) = FXa (·|XV \{a,b} ) for all XV \{a} ∈ XV \{a} .
(2.2)
In turn, by the equivalence between conditional probabilities and conditional quantiles to characterize a
random variable, we have that (2.1) occurs if and only if
QXa (τ |XV \{a} ) = QXa (τ |XV \{a,b} )
for all τ ∈ (0, 1), and XV \{a} ∈ XV \{a} .
(2.3)
For a quantile index τ ∈ (0, 1), the τ -quantile conditional independence graph is a directed graph GI (τ ) =
(V, E I (τ )) with vertex set V and edge set E I (τ ). An edge (a, b) is not contained in the edge set E I (τ ) if
6
BELLONI, CHEN, AND CHERNOZHUKOV
and only if
QXa (τ |XV \{a} ) = QXa (τ |XV \{a,b} ) for all XV \{a} ∈ XV \{a} .
(2.4)
By the equivalence between (2.2) and (2.3), the union of τ -quantile graphs over τ ∈ (0, 1) represents the
conditional independence structure of X, namely E I = ∪τ ∈(0,1) E I (τ ). We also consider a relaxation of (2.1).
For a set of quantile indices T ⊂ (0, 1), we say that
Xa ⊥T Xb | XV \{a,b} ,
(2.5)
Xa and Xb are T -conditionally independent given XV \{a,b} , if (2.4) holds for all τ ∈ T . Thus, we have that
(2.1) implies (2.5).We define the T -quantile graph as GI (T ) = (V, E I (T )) where
E I (T ) = ∪τ ∈T E I (τ ).
Although the conditional independence concept relates to all quantile indices, the quantile characterization
described above also lends itself to quantile specific impacts which can be of independent interest.2
Comment 2.2 (Simulation and Conditional Independence). Although not pursued in this work, the tools
developed here can also be used to develop simulation tools for high-dimensional random vectors. Specifically,
we can simulate a random vector X as follows
X1 ∼ QX1 (U1 ), X2 ∼ QX2 (U2 | X1 ), X3 ∼ QX3 (U3 | X1 , X2 ), . . . , Xd ∼ QXd (Ud | X1 , . . . , Xd−1 )
where Uj ∼ Uniform(0, 1), for all j ∈ [d] and estimates of the conditional quantiles can be obtained based on
a sample (Xi ∈ Rd )ni=1 and the tools discussed here. It is clear that the order of the procedure can impact
the estimation. In particular, if most variables are independent of (say) Xd , skipping them from the process
are likely to increase the accuracy of the simulation procedure.
2.2. Prediction Quantile Graphical Models. Prediction Quantile Graphical Models (PQGMs) are motivated by prediction accuracy under an asymmetric loss function (instead of conditional independence as in
Section 2.1). More precisely, for each a ∈ V , we are interested in predicting Xa based on linear combinations
of the remaining variables, XV \{a} , where accuracy is measured with respect to an asymmetric loss function.
Formally, PQGMs measure accuracy as
0
La (τ | V \{a}) = min E[ρτ (Xa − X−a
β)]
(2.6)
β
where X−a = (1, XV0 \{a} )0 , and the asymmetric loss function ρτ (t) = (τ − 1{t 6 0})t is the check function
used in [51].
Importantly, PQGMs are concerned with the best linear predictor under the asymmetric loss function
ρτ which is a specification that is widely used in practice. This is a fundamental distinction with respect
to CIQGMs discussed in Section 2.1 where the specification of the conditional quantile was approximately
a linear function of transformations of XV \{a} .3 Indeed, we note that under suitable conditions the linear
predictor that solves the minimization problem in (2.6) approximates the conditional quantile regression as
shown in [15]. (In fact, the conditional quantile function would be linear if XV was jointly Gaussian distributed.) However, PQGMs do not assume that the conditional quantile function of Xa is well approximated
by a linear function and instead it focuses on the best linear predictor.
2For example, we might be interested in some extreme events which typically correspond to crises in financial systems.
3In Section 2.1 the vector Z a in equation (3.10) collects the functions of the vector X
.
V \{a}
QUANTILE GRAPHICAL MODELS
7
We define that Xb is predictively uninformative for Xa given XV \{a,b} if
La (τ | V \{a}) = La (τ | V \{a, b})
for all τ ∈ (0, 1),
i.e., considering a linear function of Xb will not improve our performance of predicting Xa with respect to
the asymmetric loss function ρτ .
Again we can visualize the predictive relationship through a graph process indexed by τ ∈ (0, 1). That is,
for each τ ∈ (0, 1) we have a directed graph GP (τ ) = (V, E P (τ )), where an edge (a, b) ∈ GP (τ ) only if Xb is
predictively informative for Xa given XV \{a,b} at the quantile τ . Finally, it is also convenient to define the
PQGM associated with a subset T ⊂ (0, 1) as GP (T ) = (V, E P (T )) where
E P (T ) = ∪τ ∈T E P (τ ).
2.3. W-Conditional Quantile Graphical Models. In what follows, we discuss an extension of the QGMs
discussed in Sections 2.1 and 2.2 to allows for conditioning on a (possible infinity) family of events $ ∈ W.4
Such extension is motivated by several applications in which the interdependence between the random
variables in XV maybe substantially impacted by additional observable events (e.g. downside movements of
the market). This general framework allows different forms of conditioning. The main implication of this
extension is that QGMs are now graph processes indexed by τ ∈ T ⊂ (0, 1) and $ ∈ W.
We define Xa and Xb are (T , $)-conditionally independent,
Xa ⊥T Xb | XV \{a,b} , $
(2.7)
QXa (τ |XV \{a} , $) = QXa (τ |XV \{a,b} , $).
(2.8)
if for all τ ∈ T we have
The conditional independence edge set associated with (τ, $) is defined analogously as before. We denote
them by E I (τ, $) and E I (T , $) = ∪τ ∈T E I (τ, $) for each $ ∈ W.
The extension of PQGMs proceeds by defining the accuracy under the asymmetric loss function conditionally on $. More precisely, we define
0
La (τ |V \{a}, $) = min E[ρτ (Xa − X−a
β) | $].
β
(2.9)
The prediction edge set associated with (τ, $) is also defined analogously as before. We denote them by
E P (τ, $) and E P (T , $) = ∪τ ∈T E P (τ, $), for each $ ∈ W.
Example 1 (PQGMs for Stock Returns Under Market Downside Movements). Hedging decisions rely on the
dependence of various stocks returns. Moreover, hedging is even more relevant during market downside movements, which motivates us to understand interdependence conditional on those events. We can parameterize
the downside movements by using a random variable M , which could be the market index, and conditional
on the event Ω$ = {M 6 $}. This allows us to define a $-conditional-CIQGM as GI (τ, $) = (V, E I (τ, $))
and a $-conditional-PQGM as GP (τ, $) = (V, E P (τ, $)), for each $ ∈ W.
4With a slight abuse of notation, we let $ to denote the event and also the index of such event. For example, we write P($)
as a shorthand for P(W ∈ Ω$ ).
8
BELLONI, CHEN, AND CHERNOZHUKOV
3. Estimators for High-Dimensional Quantile Graphical Models
In this section, we propose and discuss estimators for QGMs introduced in Section 2. Throughout it is
assumed that we observe a d-dimensional i.i.d. random vector XV , namely {XiV : i = 1, . . . , n}. Based on
the data observed, unless additional assumptions are imposed we cannot estimate the quantities of interest
for all τ ∈ (0, 1). Instead, in what follows we will consider a (compact) set of quantile index T ⊂ (0, 1). The
estimators are intended to handle high dimensional models and a continuum of conditioning events in W.
3.1. Estimators for CIQGMs. We discuss the specification and propose an estimator for CIQGMs. Although in general it is potentially hard to correctly specify coherent models, the following are simple examples.
Example 2 (Gaussian). Consider the Gaussian case, XV ∼ N (µ, Σ). It follows that for each a ∈ V , the
conditional distribution Xa | XV \{a} satisfies
X (Σ−1 )aj
1
Xa | XV \{a} ∼ N µa −
(Xj − µj ), −1 .
(Σ−1 )aa
(Σ )aa
j∈V \{a}
Therefore the conditional quantile function of Xa is linear in XV \{a} and is given by
QXa (τ |XV \{a} ) =
Φ−1 (τ )
1/2
(Σ−1 )aa
X
+ µa −
j∈V \{a}
(Σ−1 )aj
(Xj − µj ).
(Σ−1 )aa
Example 3 (Mixture of Gaussians). Similar to the prior example, consider the case XV | $ ∼ N (µ$ , Σ$ )
for each $ ∈ W. It follows that for a ∈ V , the conditional distribution satisfies
X (Σ−1 )$aj
1
.
Xa | XV \{a} , $ ∼ N µ$a −
(Xj − µ$j ), −1
(Σ−1 )$aa
(Σ )$aa
j∈V \{a}
Again the conditional quantile function of Xa is linear in XV \{a} and is given by
QXa (τ |XV \{a} , $) =
Φ−1 (τ )
1/2
(Σ−1 )$aa
X
+ µ$a −
j∈V \{a}
(Σ−1 )$aj
(Xj − µ$j ).
(Σ−1 )$aa
Example 4 (Monotone Transformations). Consider the Gaussian case, for each a ∈ V , Xa = ha (Ya ) and
YV ∼ N (µ, Σ). It follows that for each a ∈ V , the conditional quantile function satisfies
−1
X (Σ−1 )aj
Φ
(τ
)
QXa (τ |XV \{a} ) = ha
+ µa −
(h−1 (Xj ) − µj ) .
1/2
(Σ−1 )aa j
(Σ−1 )aa
j∈V \{a}
In particular, if (ha : a ∈ V ) are monotone polynomials, the expression above is a sum of monomials with
fractional and integer exponents.
Example 5 (Multiplicative Error Model). Consider d = 2 so that V = {1, 2}. Assume that X2 and ε are
independent positive random variables. Assume further that they relate to X1 as
X1 = α + εX2 .
In this case, we have that the conditional quantile functions are linear and given by
QX1 (τ |X2 ) = α + Fε−1 (τ )X2
and QX2 (τ |X1 ) = (X1 − α)/Fε−1 (1 − τ ).
QUANTILE GRAPHICAL MODELS
9
Example 6 (Additive Error Model). Consider d = 2 so that V = {1, 2}. Let X2 ∼ U (0, 1) and ε ∼ U (0, 1)
be independent random variables. Also define the random variable X1 as
X1 = α + βX2 + ε.
It follows that QX1 (τ |X2 ) = α + βX2 + τ . However, if β = 0, we have QX2 (τ |X1 ) = τ , and for β > 0, direct
calculations yield that
(
τ
β (X1 − α), if X1 6 α + β
QX2 (τ |X1 ) =
τ + (1 − τ )(X1 − α − β), if X1 > α + β
where we note that X1 ∈ [α, 1 + α + β].
Although a linear specification is correct for Examples 2 and 5, Example 6 illustrates that we need
to consider a more general transformation of the covariates XV in the specification for each conditional
quantile function. Nonetheless, specifications with additional non-linear terms can approximate non-drastic
departures from normality.
We will consider a conditional quantile representation for each a ∈ V . It is based on transformations of
the original covariates XV \{a} that create a p-dimensional random vector Z a = Z a (XV \{a} ) such that
QXa (τ |XV \{a} ) = Z a βaτ + raτ , βaτ ∈ Rp , for all τ ∈ T ,
(3.10)
where raτ denotes a small approximation error. For b ∈ V \{a} we let Ia (b) := {j : Zja depends on Xb }.
That is, Ia (b) contains the components of Z a that are functions of Xb . Under correct specification, if Xa
and Xb are conditionally independent, we have βaτ j = 0 for all j ∈ Ia (b), τ ∈ (0, 1).
This allows us to connect the conditional independence quantile graph estimation problem with model
selection with quantile regression. Indeed, the representation (3.10) has been used in several quantile regression models, see [50]. Under mild conditions this model allows us to identify the process (βaτ )τ ∈T as the
solution of the following moment equation
E[(τ − 1{Xa 6 Z a βaτ + raτ })Z a ] = 0.
(3.11)
In order to allow for a flexible specification, so that the approximation errors are negligible, it is attractive
to consider a high-dimensional Z a where its dimension p is possibly larger than the sample size n. In turn,
having a large number of technical controls creates an estimation challenge if the number of coefficients p
is not negligible with respect to the sample size n. In such a high dimensional setting, a widely applicable
condition that makes estimation possible is approximate sparsity [40, 10, 17]. Formally we require
p
2
max sup kβaτ k0 6 s, max sup {E[raτ
]}1/2 . s/n, and max sup |E[faτ raτ Z a ]| = o(n−1/2 ), (3.12)
a∈V τ ∈T
a∈V τ ∈T
a∈V τ ∈T
where the sparsity parameter s of the model is allowed to grow (at a slower rate) as n grows, and faτ =
fXa |XV \{a} (QXa (τ |XV \{a} )|XV \{a} ) denotes the conditional density function evaluated at the corresponding
conditional quantile value. This sparsity also has implications on the maximum degree of the associated
quantile graph.
Algorithm 3.1 below contains our proposal to estimate βaτ , a ∈ V , τ ∈ T . It is based on three procedures
in order to overcome high-dimensionality. In the first step, we apply a (post-)`1 -penalized quantile regression.
The second step applies (post-)Lasso where the data is weighted by the conditional density function at the
10
BELLONI, CHEN, AND CHERNOZHUKOV
conditional quantile.5 Finally, the third step relies on constructing (orthogonal) score function that provides
immunity to (unavoidable) model selection mistakes.
There are several parameters that need to be specified for Algorithm 3.1. The penalty parameter λV T is
chosen to be larger than the `∞ -norm of the (rescaled) score at the true quantile function. The work in [11]
exploits the fact that this quantity is pivotal in their setting. Here, additional correlation structure would
have an impact and the distribution is pivotal only for each a ∈ V . The penalty is based on the maximum
of the quantiles of the following random variables (each with pivotal distribution), for a ∈ V
ΛaT = sup max
τ ∈T j∈[p]
|En [(1{U 6 τ } − τ )Zja ]|
p
Z
τ (1 − τ )b
σaj
(3.13)
Z
where {Ui : i = 1, . . . , n} are i.i.d. uniform (0, 1) random variables, and σ
baj
= {En [(Zja )2 ]}1/2 for j ∈ [p].
The penalty parameter λV T is defined as
λV T := max ΛaT (1 − ξ/|V | | Z a ),
a∈V
that is, the maximum of the 1−ξ/|V | conditional quantile of ΛaT given in (3.13). Regarding the penalty term
for the weighted Lasso in Step 2, we recommend a (theoretically valid) iterative choice. We refer to Appendix
P Z
A for the implementation details of the algorithm. We denote kβk1,bσZ := j σ
baj |βj | the standardized version
of the `1 -norm.
Algorithm 3.1. (CIQGM Estimator.) For each a ∈ V , τ ∈ T , and j ∈ [p]
p
Step 1. Compute βbaτ from k · k1,bσZ -penalized τ -quantile regression of Xa on Z a with penalty λV T τ (1 − τ ).
p
Z
σak
}.
Compute βeaτ from τ -quantile regression of Xa on {Zka : |βbaτ k | > λV T τ (1 − τ )/b
j
a
Step 2. Compute γ
eaτ
from the post-Lasso estimator of faτ Zja on faτ Z−j
.
a
a
a
a
j
) and for
Step 3. Construct the score function ψbi (α) = (τ − 1{Xia 6 Zij
α + Zi,−j
βeaτ,−j })fiaτ (Zij
− Zi,−j
γ
eaτ
2
2
b
b
Laτ j (α) = |En [ψi (α)]| /En [ψi (α)], set β̌aτ j ∈ arg minα∈Aaτ j Laτ j (α).
Algorithm 3.1 above has been studied in [18] where it is applied to a single triple (a, τ, j), and we have used
Z
the following parameter space for α, Aaτ j = {α ∈ R : |α − βeaτ j | 6 10/{b
σaj
log n}}. Under similar conditions,
results that hold uniformly over (a, τ, j) ∈ V × T × [p] are achievable (as shown in the next sections) building
upon the tools developed in [11] and [25]. Algorithm 3.1 is tailored to achieve good rates of convergence in
the `∞ -norm. In particular, under standard regularity conditions, with probability approaching to 1 we have
r
log(p|V |n)
sup kβaτ − β̌aτ k∞ .
.
n
τ ∈T
In order to create an estimate of E I (τ ) = {(a, b) ∈ V × V : maxj∈Ia (b) |βaτ j | > 0}, we define
b I (τ ) = (a, b) ∈ V × V : max |β̌aτ j | > cv
E
j∈Ia (b) se(β̌aτ j )
2
−1 1/2
a
a
j
where se(β̌aτ j ) = {τ (1 − τ )En [e
viaτ
}
with ṽiaτ j = fbiaτ {Zij
− Zi,−j
γ̃aτ
}, is an estimate of the standard
j]
deviation of the estimator, and the critical value cv is set to account for the uniformity over a ∈ V , τ ∈ T ,
and j ∈ [p]. We discuss in the following sections a data driven procedure based on multiplier bootstrap that
is theoretically valid in this high dimensional setting.
5We note that an estimate for f
aτ is available from `1 -penalized quantile regression estimators for τ + h and τ − h where h
is a bandwidth parameter, see [50, 18] and Comment 3.2.
QUANTILE GRAPHICAL MODELS
11
Comment 3.1 (Stepdown Procedure for cv). Setting a critical value cv that accounts for the multiple
b I (τ ). Further improvements can be
hypotheses being tested plays an important role to estimate the graph E
obtained by considering the stepdown procedure of [68] for multiple hypothesis testing that was studied for
the high-dimensional case in [24]. The procedure iteratively creates a suitable sequence of decreasing critical
values. In each step only null hypotheses that were not rejected are considered to determine the critical
value. Thus, as long as any hypothesis is rejected at a step, the critical value decreases and we continue to
the next iteration. The procedure stops when no hypothesis in the current active set is rejected.
Comment 3.2 (Estimation of Conditional Density Function). The algorithm above requires the conditional
density function faτ which typically needs to be estimated in practice. It turns out that estimation of
conditional quantiles yields a natural estimator for the conditional density function as
faτ =
1
.
∂QXa (τ |Z a )/∂τ
Therefore, based on `1 -penalized quantile regression estimates at the τ + hn and τ − hn quantile, where
h = hn → 0 denotes a bandwidth parameter, we have
fbaτ =
2h
a
b
b X (τ − h|Z a )
QXa (τ + h|Z ) − Q
a
(3.14)
as an estimator of faτ . Under smoothness conditions, it has an bias of order h2 . See [18] and the references
therein for additional comments and estimators.
3.2. Estimators for PQGMs. In this section we propose an estimator for PQGMs in which case we are
interested in the prediction of Xa , a ∈ V , using a linear combination of XV \{a} under the asymmetric
loss discussed in (2.6). We will add an intercept as one of the variables for the sake of notation so that
X−a = (1, XV0 \{a} )0 . Given the loss function ρτ , the target d-dimensional vector of parameters βaτ is defined
as (part of) the solution of the following optimization problem
0
βaτ ∈ arg min E[ρτ (Xa − X−a
β)].
β
(3.15)
As we are interested in the case that d is large, the use of high-dimensional tools to achieve consistent
estimators is needed. The estimation procedure we proposed is based on `1 -penalized quantile regression but
additional issues need to be considered to cope with the (non-vanishing) difference between the best linear
predictor and the conditional quantile function. Again we consider models that satisfy an approximately
sparse condition. Formally, we require the existence of sparse coefficients {β̄aτ : a ∈ V, τ ∈ T } such that
p
0
max sup kβ̄aτ k0 6 s and max sup {E[{X−a
(βaτ − β̄aτ )}2 ]}1/2 . s/n,
(3.16)
a∈V τ ∈T
a∈V τ ∈T
where (again) the sparsity parameter s of the model is allowed to grow as n grows. The high-dimensionality
prevents us from using (standard) quantile regression methods and regularization methods are needed to
achieve good prediction properties.
A key issue is to set the penalty parameter properly so that it bounds from above
0
max sup max |En [(1{Xa 6 X−a
βaτ } − τ )X−a,j ]|.
a∈V τ ∈T j∈[d]
(3.17)
However, it is important to note that we do not assume that the conditional quantile of Xa is a linear
function of X−a . Under correct linear specification of the conditional quantile function, `1 -penalized quantile
12
BELLONI, CHEN, AND CHERNOZHUKOV
regression estimator has been studied in [11]. The case that the conditional quantile function differs from
a linear specification by vanishing approximation errors has been considered in [47] and [18]. The analysis
proposed here aims to allow for non-vanishing misspecification of the quantile function relative to a linear
specification while still guarantees good rates of convergence in the `2 -norm to the best linear specification.
Thus the penalty parameter in the penalized quantile regression needs to account for such misspecification
and is no longer pivotal as in [11].
In order to handle this issue we propose a two step estimation procedure. In the first step, the penalty
parameter λ0 is conservative and is set via bounds constructed based on symmetrization arguments, similar
p
in spirit to [70, 12]. This leads to λ0 = 2(1 + 1/16) 2 log(8|V |2 /ξ)/n. Although this is conservative, under
mild conditions this would lead to estimates that can be leverage to fine tune the penalty choice. The second
step uses the preliminary estimator to bootstrap (3.17) based on the tools in [24] as follows. Specifically, for
0
estimates εbiaτ of the “noise” εiaτ = 1{Xia 6 Xi,−a
βaτ } − τ for i ∈ [n], for a ∈ V define
Λ̄aT := 1.1 sup max
τ ∈T j∈[d]
|En [gi εbiaτ Xi,−aj ]|
2
{En [b
ε2iaτ Xi,−aj
]}1/2
(3.18)
where (gi )ni=1 is a sequence of i.i.d. standard Gaussian random variables. The new penalty parameter λ̄V T
is defined as
λ̄V T := max Λ̄aT (1 − ξ|X−a )
a∈V
(3.19)
that is, the maximum of the (1 − ξ) conditional quantile of Λ̄aT . The penalty choice above adapts to
the unknown correlation structure across components and quantile indices. The following algorithm states
P X
X
the procedure where we denote weighted `1 -norms by kβk1,bσX := j σ
baj |βj | with σ
baj
= {En [Xj2 ]}1/2 the
P εX
2
εX
standardized version of the `1 -norm and kβk1,bε := j σ
]}1/2 a norm based
baτ j |βj | with σ
baτ
ε2aτ X−a,j
j = {En [b
on the estimated residuals.
Algorithm 3.2. (PQGM Estimator.) For each a ∈ V , and τ ∈ T
Step 1. Compute βbaτ from k · k1,bσX -penalized τ -quantile regression of Xa on X−a with penalty λ0 .
X
Compute βeaτ from τ -quantile regression of Xa on {Xk : |βbaτ k | > λ0 /b
σak
}.
0
e
Step 2. For εbiaτ = 1{Xia 6 Xi,−a βaτ } − τ for i ∈ [n], and ξ = 1/n, compute λ̄V T via (3.19).
Step 3. Recompute βbaτ from k · k1,bε -penalized τ -quantile regression of Xa on X−a with penalty λ̄V T .
εX
Compute β̌aτ from τ -quantile regression of Xa on {Xk : |βbaτ k | > λ̄V T /b
σaτ
k }.
Under regularity conditions stated in Section 4, with probability approaching 1, we have
r
s log(|V |n)
max sup kβaτ − β̌aτ k .
.
a∈V τ ∈T
n
The estimate of the prediction quantile graph is given by the support of (β̌aτ )a∈V,τ ∈T , namely
n
o
εX
b P (τ ) = (a, b) ∈ V × V : |βbaτ b | > λ̄V T /b
E
σaτ
b .
That is, it is induced by covariates selected by the `1 -penalized estimator. Those thresholded estimators not
only have the same rates of convergence as of the original penalized estimators but also possess additional
sparsity guarantees.
QUANTILE GRAPHICAL MODELS
13
3.3. Estimators for W-Conditional Quantile Graphical Models. In order to handle the additional
conditioning events Ω$ , $ ∈ W, we propose to modify Algorithms 3.1 and 3.2 based on kernel smoothing.
To that extent, we assume the observed data is of the form {(XiV , Wi ) : i = 1, . . . , n}, where Wi might be
defined through additional variables. Furthermore, we assume for each conditioning event $ ∈ W we have
access to a kernel function K$ that is applied to W , to represent the relevant observations associated with
$ (recall that we denote P(W ∈ Ω$ ) as P($)). We assume that K$ (W ) = 1{W ∈ Ω$ }.
Example 7 (PQGMs for Stock Returns Under Market Downside Movements, continued). In Example 1,
we have W as the market return and the conditioning event as Ω$ = {W 6 $} which is parameterized by
$ ∈ W, a closed interval in R. We might be interest in a fixed $ or on a family of values $ ∈ (−$̄, 0]. The
latter induces W = {Ω$ = {W 6 $} : $ ∈ (−$̄, 0]}. The kernel function is simply K$ (t) = 1{t 6 $}.
This framework encompasses the previous framework by having K$ (W ) = 1 for all W . Moreover, it
allows for a richer class of estimands which require estimators whose properties should hold uniformly over
$ ∈ W as well. Next we propose estimators for this setting, i.e. we generalize the previous methods
to account for the additional conditioning on $ ∈ W. In what follows, with a slight abuse of notation
we use $ to denote not only the index but also the event Ω$ . For further notational convenience, we
denote u = (a, τ, $) ∈ U := V × T × W so that the set U collects all the three relevant indices. With
P
Z
Z
σ
ba$j
= {En [K$ (W )(Zja )2 ]}1/2 , we define the following weighted `1 -norm kβk1,$ = j∈[p] σ
ba$j
|βj |. This
norm is $ dependent and provides the proper adjustments as we condition on different events associated
with different $’s.
We first consider estimators of CIGMs conditional on the events in W. In this setting, the model is
correctly specified up to small approximation errors. The definition of the penalty parameter will be based
on the random variable
ΛaT W =
sup
max
τ ∈T ,$∈W j∈[p]
En [K$ (W )(1{U 6 τ } − τ )Zja ]
p
Z
τ (1 − τ )b
σa$j
where Ui are independent uniform (0, 1) random variables, and set the penalty
λV T W = max ΛaT W (1 − ξ/{|V |n1+2dW }|Z a , W ),
a∈V
that is, the maximum of the (1 − ξ/{|V |n1+2dW }) conditional quantile of ΛaT W . Algorithm 3.3 provides
Z
the definition of the estimator. Here Auj = {α ∈ R : |α − βeuj | 6 10/{b
σa$j
log n}}, and denote λu :=
p
λV T W τ (1 − τ ).
Algorithm 3.3. (W-Conditional CIQGM Estimator.) For (a, τ, $) ∈ V × T × W and j ∈ [p]
Step 1. Compute βbu from k · k1,$ -penalized τ -quantile regression of K$ (W )(Xa ; Z a ) with penalty λu .
Z
Compute βeu from τ -quantile regression of K$ (W )(Xa ; {Zka : |βbuk | > λu /b
σa$j
}).
j
a
a
Step 2. Compute γ
eu from the post-Lasso estimator of K$ (W )fu Zj on K$ (W )fu Z−j .
a
a
a
a
Step 3. Construct the score function ψbi (α) = K$ (Wi )(τ − 1{Xia 6 Zij
α + Zi,−j
βeu,−j })fiu (Zij
− Zi,−j
γ
euj )
and for Luj (α) = |En [ψbi (α)]|2 /En [ψbi2 (α)], set β̌uj ∈ arg minα∈Auj Luj (α) .
Next we consider estimators of PQGMs conditional on the events in W. Similar to the previous case, for
a ∈ V define
|En [K$ (W )gb
εaτ $ X−a,j ]|
(3.20)
Λ̄aT W := 1.1 sup max
2
2
εaτ $ X−a,j
]}1/2
τ ∈T ,$∈W j∈[d] {En [K$ (W )b
14
BELLONI, CHEN, AND CHERNOZHUKOV
where (gi )ni=1 is a sequence of i.i.d. standard Gaussian random variables. The new penalty parameter λ̄V T
is defined as
λ̄V T W := max Λ̄aT W (1 − ξ|X−a )
a∈V
(3.21)
that is, the maximum of the (1 − ξ) conditional quantile of Λ̄aT W . It will also be useful to define another
P εX
εX
2
weighted `1 -norm, kβk1,$bε :=
baτ $j |βj | with σ
baτ
ε2aτ $ X−a,j
]}1/2 . We also denote
$j = {En [K$ (W )b
jσ
X
2
1/2
σ
ba$j = {En [K$ (W )X−a,j ]} . The penalty choice and weighted `1 -norm adapt to the unknown correlation
structure across components and quantile indices. The following algorithm states the procedure, with λ0W =
p
2(1 + 1/16) 2 log(8|V |2 {ne/dW }2dW /ξ)/n.
Algorithm 3.4. (W-Conditional PQGM Estimator.) For (a, τ, $) ∈ V × T × W
Step 1. Compute βbu from k · k1,$ -penalized τ -quantile regression of Xa on X−a with penalty λ0W .
X
Compute βeu from τ -quantile regression of K$ (W )(Xa ; {X−a,k : |βbuk | > λ0W /b
σa$k
}).
0
Step 2. For εbiu = 1{Xia 6 Xi,−a
βeu } − τ for i ∈ [n], and ξ = 1/n, compute λ̄V T W via (3.21).
Step 3. Recompute βbu from k·k1,$bε -penalized τ -quantile regression of K$ (W )(Xa ; X−a ) with penalty λ̄V T W .
εX
Compute β̌u from τ -quantile regression of K$ (W )(Xa ; {X−a,k : |βbuk | > λ̄V T W /b
σuk
}).
Comment 3.3 (Computation of Penalty Parameter over W). The penalty choices require one to maximize
over a ∈ V , τ ∈ T and $ ∈ W. The set V is discrete and does not pose a significant challenge. However
both other sets are continuous and additional care is needed. In most applications we are concerned with
the case that W is a low dimensional VC class of sets and it impacts the calculation only through indicator
functions, which is precisely the case of T . It follows that only a polynomial number (in n) of different values
of τ and $ would need to be considered.
4. Main Theoretical Results
This section is devoted to theoretical guarantees associated with the proposed estimators. We will establish
rates of convergence results for the proposed estimators as well as the (uniform) validity of confidence regions.
These results build upon and contribute to an increasing literature on the estimation of many processes of
interest with (high-dimensional) nuisance parameters.
Throughout, we will provide results for the estimators of the W-conditional quantile graphical models
as those can be generalized the other models by setting K$ (W ) = 1. Although some of the tools are
similar, CIQGMs and PQGMs require different estimators and are subject to different assumptions. Thus,
substantial different analyses are required.
4.1. W-Conditional CIQGM. For u = (a, τ, $) ∈ U, define the τ -conditional quantile function of Xa
given XV \{a} and $ as
QXa (τ |XV \{a} , $) = Z a βu + ru ,
(4.22)
a
where Z is a p-dimensional vector of (known) transformations of XV \{a} , and ru is an approximation error.
The event $ ∈ W will be used for further conditioning through the function K$ (W ) = 1{W ∈ $}.
We let fXa |XV \{a} ,$ (·|XV \{a} , $) denote the conditional density function of Xa given XV \{a} and $ ∈ W.
We define fu := fXa |XV \{a} ,$ (QXa (τ |XV \{a} , $)|XV \{a} , $) as the value of the conditional density function
QUANTILE GRAPHICAL MODELS
15
evaluated at the τ -conditional quantile. In our analysis we will consider for u ∈ U
E[fu {Z a δ}2 |$]
and f U = min f u .
u∈U
kδk=1 E[{Z a δ}2 |$]
f u = inf
(4.23)
Moreover, for each u ∈ U and j ∈ [p] we define
a
γuj = arg min E[fu2 K$ (W )(Zja − Z−j
γ)2 ].
(4.24)
γ
This provides a weighted projection to construct the residuals
a
vuj = fu (Zja − Z−j
γuj )
a
that satisfy E[fu Z−j
vuj |$] = 0 for each (u, j) ∈ U × [p].
The estimands of interest are βu ∈ Rp , u ∈ U, and can be written as the solution of (a continuum of)
moment equations. Letting βuj denote the jth component of βu so that βuj ∈ R solves
E[ψuj (X, W, β, ηuj )] = 0,
where the function ψuj is given by
(1)
(3)
(2)
a
a
ψuj (X, W, β, ηuj ) = K$ (W )(τ − 1{Xa 6 Zja β + Z−j
ηuj + ηuj })fu (Zja − Z−j
ηuj ),
(1)
(2)
(3)
(1)
(2)
and the true value of the nuisance parameter is given by ηuj = (ηuj , ηuj , ηuj ) with ηuj = βu,−j , ηuj = γuj ,
(3)
and ηuj = ru . In what follows c, C denote some fixed constant, δn and ∆n denote sequences go to zero with
δn = n−µ for some sufficiently small µ. Denote µW = inf $∈W P($).
Condition CI. Let u = (a, τ, $) ∈ U := V × T × W and (Xi , Wi )ni=1 denote a sequence of independent
and identically distributed random vectors generated accordingly to models (4.22) and (4.24):
(i) Suppose supu∈U ,j∈[p] {kβu k + kγuj k} 6 C and T is a fixed compact set: (a) there exists s = sn such
that supu∈U ,j∈[p] {kβu k0 + kγ̄uj k0 } 6 s, supu∈U ,j∈[p] kγ̄uj − γuj k + s−1/2 kγ̄uj − γuj k1 6 C{n−1 s log(|V |pn)}1/2 ,
where γ̄uj is approximately sparse; (b) the conditional distribution function of Xa given XV \{a} and $ is
absolutely continuous with continuously differentiable density fXa |XV \{a} ,$ (t|XV \{a} , $) bounded by f¯ and
its derivative bounded by f¯0 uniformly over u ∈ U; (c) |fu − fu0 | 6 Lf ku − u0 k, kβu − βu0 k 6 Lβ ku − u0 kκ with
κ ∈ [1/2, 1], and E[|K$ (W ) − K$0 (W )|] 6 LK k$ − $0 k; (d) the VC dimension dW of the set W is fixed,
{QXa (τ |XV \{a} , $) : (τ, $) ∈ T × W} is a VC-subgraph with VC-dimension 1 + CdW for every a ∈ V ;
(ii) The following moment conditions hold uniformly over u ∈ U and j ∈ [p]: E[|fu vuj Zka |2 |$]1/2 6
Cf u , mina∈V inf kδk=1 E[{(Xa , Z a )δ}2 |$] > c, maxa∈V supkδk=1 E[{(Xa , Z a )δ}4 |$] 6 C, E[fu2 (Z a δ)2 |$] 6
Cf 2u E[(Z a δ)2 |$], maxj,k
E[|fu vuj Zka |3 |$]1/3
E[|fu vuj Zka |2 |$]1/2
log1/2 (pn|V |) 6 δn {nP($)}1/6 ;
(iii) Furthermore, for some fixed q > 4∨(1+2dW ), supu∈ U ,kδk=1 E[|(Xa , Z a )δ|2 ru2 |$] 6 CE[ru2 |$] 6 Cs/n,
maxu∈U ,j∈[p] |E[fu ru vuj |$]| 6 δn n−1/2 , E[maxi6n supu∈U |K$ (W )riu |q ] 6 C, and with probability 1 − ∆n ,
2
uniformly over u ∈ U, j ∈ [p]: En [ru2 vuj
|$] + En [ru2 |$] . n−1 s log(p|V |n), En [K$ (W ){|ru | + ru2 }(Z a δ)2 ] 6
δn En [K$ (W )fu (Z a δ)2 ];
(iv) For a fixed q > 4 ∨ (1 + dW ), diam(W) 6 n1/2q , E[maxi6n kXiV kq∞ ∨ maxa∈V kZia kq∞ ]1/q /µW 6
Mn , E[maxi6n supu∈U ,j∈[p] |viuj |q ]1/q 6 Ln , (Lf + LK )2 Mn2 log2 (p|V |n) 6 δn nµ3W f 6U , Mn4 log(p|V |n) log n 6
δn2 nµ2W f 2U , s2 log2 (p|V |n) 6 δn2 nf 4U µ6W , s3 log3 (p|V |n) 6 δn4 nf 2U µ3W , L2n s log3/2 (p|V |n) 6 δn f U (nµW )1/2 ,
p
Mn s log(p|V |n) 6 δn n1/2 µW f U .
16
BELLONI, CHEN, AND CHERNOZHUKOV
Condition CI assumes various conditional moment conditions to allow for the estimation to be conditional
on $ ∈ W. Those are analogous to the (unconditional) conditions in the high-dimensional literature in
quantile regression models, [18]. In particular, condition CI(i) assumes smoothness of the density function,
and of coefficients. Condition CI(ii) assumes conditions on the (conditional) population design matrices such
as the ratio between eigenvalues. Condition CI(iii) pertains to the approximations errors and assumes mild
moment conditions. Finally Condition CI(iv) provides sufficient conditions on the allowed growth of the
model via p and |V | relative to the available sample size n. Based on Condition CI, we derive our main
results regarding the proposed estimator. Moreover, we also establish new results for `1 -penalized quantile
regression methods that hold uniformly over the indices u ∈ U. The following theorems summarize these
results. Note, Condition CI(iii) also assume dW is bounded by fixed q, and the proof can easily be extended
to other cases.
Theorem 1 (Uniform Rates of Convergence for W-Conditional Penalized Quantile Regression). Under
Condition CI, we have that with probability at least 1 − o(1)
s
s(1 + dW ) log(p|V |n)
b
, uniformly over u = (a, τ, $) ∈ U
k βu − βu k .
nf u P($)
p
λ̄
Moreover, the thresholded estimator βbλ̄ , with λ̄ = (1 + dW ) log(p|V |n)/n and βbuj
= βbuj 1{|βbuj | >
Z
λ̄b
σa$j
}, satisfies the same rate and kβbλ̄ k0 . s.
Theorem 1 builds upon ideas in [11] however the proof strategy is designed to derive rates that are adaptive
to each u ∈ U. Indeed the rates of convergence are u-dependent and they show a slower rate for rare events
$ ∈ W.
Theorem 2 (Uniform Rates of Convergence for W-Conditional Weighted Lasso). Under Condition CI, we
have that with probability at least 1 − o(1)
s
1
s(1 + dW ) log(p|V |n)
kb
γuj − γuj k .
and kb
γuj k0 . s, uniformly over u = (a, τ, $) ∈ U, j ∈ [p].
fu
nP($)
The following result establishes a uniform Bahadur representation for the final estimators.
Theorem 3 (Uniform Bahadur Representation for W-Conditional CIQGM). Under Condition CI, the estimator (β̌uj )u∈U ,j∈[p] satisfies
−1 √
σuj
n(β̌uj − βuj ) = Un (u, j) + OP (δn ) in `∞ (U × [p]),
2
2 −1
where σuj
= τ (1 − τ )E[K$ (W )vuj
] and
Un (u, j) :=
n
2
{τ (1 − τ )E[K$ (W )vuj
]}−1/2 X
√
(τ − 1{Ui (a, $) 6 τ })K$ (Wi )vi,uj ,
n
i=1
where U1 (a, $), . . . , Un (a, $) are i.i.d. uniform (0, 1) random variables, independent of v1,uj , . . . , vn,uj .
Theorem 3 plays a key role. However, it is important to note that the marginal distribution of Un (u, j)
is pivotal. Nonetheless, there is a non-trivial correlation structure between U (a, $) and U (ã, $̃). In order to construct confidence regions with non-conservative guarantees, we rely on a multiplier bootstrap
method. We will approximate the process N = (Nuj )u∈U ,j∈[p] by the Gaussian multiplier bootstrap
QUANTILE GRAPHICAL MODELS
17
2
based on estimates ψbuj := {τ (1 − τ )En [K$ (W )b
vuj
]}−1/2 (τ − 1{Xa 6 Z a βbu })K$ (Wi )b
vuj of ψ̄uj (U, W ) =
2
−1/2
{τ (1 − τ )En [K$ (W )vuj ]}
(τ − 1{U (a, $) 6 τ })K$ (W )vuj , namely
(
)
n
X
1
Gb = (Gbuj )u∈U ,j∈[p] = √
gi ψbuj (Xi , Wi )
n i=1
u∈U ,j∈[p]
(gi )ni=1
are independent standard normal random variables which are independent from the data
where
n
(Wi )i=1 . Based on Theorem 5.2 of [24], the following result shows that the multiplier bootstrap provides a
√
valid approximation to the large sample probability law of n(β̌uj − βuj )u∈U ,j∈[p] which is suitable for the
construction of uniform confidence bands over the set of indices associated with Ia (b) for all a, b ∈ V .
Corollary 1 (Gaussian Multiplier Bootstrap for W-Conditional CIQGM). Under Condition CI with δn =
o({(1 + dW ) log(p|V |n)}−1/2 ), and (1 + dW ) log(p|V |n) = o({(n/L2n )1/7 ∧ (n1−2/q /L2n )1/3 }), we have that
|β̌uj − βuj |
0
0
n
b
PP max
sup
sup
∈ [t, t ] − PP max |Guj | ∈ [t, t ] | (Xi , Wi )i=1 = o(1)
j∈Ia (b) n−1/2 σuj
j∈Ia (b)
P ∈Pn t,t0 ∈R,u∈U ,b∈V
Corollary 1 allows the construction of simultaneous confidence regions for the coefficients that are uniformly valid over the set of data generating processes induced by Condition CI. Based on the coefficients whose intervals do not overlap zero, we can construct a conditional independence graph process
b I (τ, $), τ ∈ T , $ ∈ W that contains the true conditional independence quantile graph with a specified
E
probability.
4.2. W-Conditional PQGM. In this section, we derive theoretical guarantees for the W-conditional predictive quantile estimators uniformly over u = (a, τ, $) ∈ U. For each u ∈ U the estimand of interest is
βu ∈ Rp that corresponds to the best linear predictor under asymmetric loss function, namely
0
βu ∈ arg min E[ρτ (Xa − X−a
β) | $]
β
(4.25)
where the event $ ∈ W is used for further conditioning. In the analysis below, the conditioning is implemented through the function K$ (W ) = 1{W ∈ $}.
In the analysis of this case, the main issue is to handle the inherent misspecification of the linear form
0
X−a
βu with respect to the true conditional quantile. The first consequence is to handle the identification con0
dition. Given X−a and $ ∈ W, we let fu := fXa |X−a ,$ (X−a
βu |X−a , $) denote the value of the conditional
0
density function evaluated at X−a βu . In our analysis, we will consider
0
E[fu {X−a
δ}2 | $]
and f U = min f u .
0
u∈U
kδk=1 E[(X−a δ)2 | $]
f u = inf
(4.26)
We remark that f u defined in (4.26) differs from (4.23) which is the standard conditional density at the
true quantile value. It turns out that Knight’s identity can be used by exploiting the first order condition
associated with the optimization problem (4.25) which yields zero mean condition similar to the conditional
quantile condition.
A second consequence of the misspecification is the lack of pivotality of the score. Such pivotal property
was convenient in the previous section to define penalty parameters and to conduct inference. We will exploit
bounds on the VC-dimension of the relevant classes of sets formally stated below.
18
BELLONI, CHEN, AND CHERNOZHUKOV
Condition P. Let U = V × T × W and (Xi , Wi )ni=1 denote a sequence of independent and identically
distributed random vectors generated accordingly to models (4.25):
(i) Suppose that supu∈U kβu k 6 C and T is a fixed compact set: (a) there exists s = sn and β̄u such that
p
supu∈U kβ̄u k0 6 s, supu∈U kβ̄u − βu k + s−1/2 kβ̄u − βu k1 6 s/n; (b) the conditional distribution function of
Xa given X−a and $ is absolutely continuous with continuously differentiable density fXa |X−a ,$ (t | X−a , $)
such that its values are bounded by f¯ and its derivative is bounded by f¯0 uniformly over u ∈ U; (c) |fu −fu0 | 6
Lf ku − u0 k, kβu − βu0 k 6 Lβ ku − u0 kκ with κ ∈ [1/2, 1], and E[|K$ (W ) − K$0 (W )|] 6 LK k$ − $0 k; (d)
0
the VC dimension dW of the set W is fixed, {1{Xa 6 X−a
βu } : (τ, $) ∈ T × W} is a VC-class with
VC-dimension 1 + dW for every a ∈ V ;
0
(ii) The following moment conditions hold uniformly over u ∈ U: mina∈V inf kδk=1 E[{X−a
δ}2 |$] > c,
0
maxa∈V supkδk=1 E[{X−a
δ}4 |$] 6 C;
0
0
((iii) With probability 1 − ∆n , uniformly over u ∈ U and a ∈ V : En [K$ (W ){|X−a
(β̄u − βu )| + |X−a
(β̄u −
2
2
a 2
0
βu )| }(Z δ) ] 6 δn En [K$ (W )fu (X−a δ) ];
(iv) For a fixed q > 4 ∨ (1 + dW ), we have that: diam(W) 6 n1/2q , E[maxi6n kXiV kq∞ ]1/q /µW 6
Mn , Mn2 log7 (n|V |) 6 δn nf 2U µ2W , Mn4 log(n|V |) log n 6 δn nµW , (Lf + LK )2 Mn2 log2 (|V |n) 6 δn nµ3W f 6U ,
p
Mn2 s log3/2 (n|V |) 6 δn f U (nµW )1/2 , Mn s log(n|V |) 6 δn (nµW )1/2 , and s3 log5 (n|V |) 6 δn nf 2U µ2W .
Condition P is a high-level condition. It allows to cover conditioning events $ ∈ W whose probability
can decrease to zero (although slower than n−1/4 ).
Next we derive our main results regarding the proposed estimator for the best linear predictor. These
results are also new `1 -penalized quantile regression methods as it holds under possible misspecification of the
conditional quantile function and hold uniformly over the indices u ∈ U. The following theorem summarizes
the result.
Theorem 4 (Uniform Rates of Convergence for W-Conditional Penalized Quantile Regression under Misspecification). Under Condition P, we have that with probability at least 1 − o(1), uniformly over u =
(a, τ, $) ∈ U,
s
s(1 + dW ) log(|V |n)
kβbu − βu k .
.
nf u P($)
The data-driven choice of penalty parameter helps diminish the regularization bias and also allow to
obtain sparse estimators with provably rates of convergence (through thresholding). Moreover, the u specific
penalty parameter combined with the new analysis yields an adaptive rate of convergence to each u ∈ U
unlike previous works.
Comment 4.1 (Simultaneous Confidence Bands for Coefficients in PQGMs). We note that in some applications we might be interested in constructing (simultaneous) confidence bands for the coefficients in PQGMs.
In particular, this would include the cases practitioners are using a misspecified linear specification in a
0
quantile regression model. Provided the conditional density function at X−a
βu can be estimated, a version
of Algorithm 3.3 using the penalty parameters in Algorithm 3.4 for the initial step can deliver such confidence
regions via a multiplier bootstrap.
QUANTILE GRAPHICAL MODELS
19
5. Empirical Applications of QGMs
5.1. CoVaR, Network Spillover Effects, and Systemic Risk. Traditional risk measures, such as Value
of Risk (VaR), focus on the loss of an individual institution only. CoVaR proposed by [4] is to measure the
VaR of the whole financial system or a particular financial institution by conditioning on another institution
being in distress.
[4] define institution b’s CoVaR at level τ conditional on a particular outcome of institution a, as the value
b|a
of CoV aRτ that solves
P(Xb 6 CoV aRτb|a |C(Xa )) = τ.
(5.27)
A special case is C(Xa ) = {Xa = V aRτa } which, as interpreted by [4], means with probability τ institution
b is in trouble given that institution a is in trouble. The estimation procedure is defined as: firstly, quantile
regression Xb on Xa gives the value at risk of institution b conditional on institution a,
V aRτb |Xa = β0b (τ ) + βab (τ )Xa ,
(5.28)
then replacing variable Xa by its τ -th quantile, i.e. V aRτa , yields
b|Xa =V aRτa
CoV aRτ
= β0b (τ ) + βab (τ )V aRτa ,
(5.29)
and
a
∆CoV aRτb|a = βab (τ )(V aRτa − V aR50%
).
(5.30)
Below, we show with QGM we can incorporate (tail) network spillover effects into risk measuring. (Note
the identified risk connections between all financial institutions constitute a systemic risk network.) Define
P(Xb 6 CoV aRτb|a,V \{a,b} |C(Xa , XV \{a,b} )) = τ,
(5.31)
we then have
b|Xa =V aRτa ,XV \{a,b} =V aRτV \{a,b}
CoV aRτ
= β̌0b (τ ) + β̌ab (τ )V aRτa + β̌Vb \{a,b} (τ )V aRτV \{a,b} ,
(5.32)
and
a
∆CoV aRτb|a,V \{a,b} = β̌ab (τ )(V aRτa − V aR50%
),
(5.33)
where β̌ b (τ ) = {β̌0b (τ ), β̌Vb \{b} (τ )} is estimated via Algorithm 3.1 or 3.2.
b|a,V \{a,b}
We stack ∆CoV aRτ
as the (a, b)-th element of an d×d matrix E β (τ ) which represents a weighted
and directed network of institutions. Following [5], the systemic risk contribution of institution a is called
P
k|a,V \{a,k}
the network to-degree, defined as δato = ∆CoV aRsys|a =
. To-degrees measure
k ∆CoV aR
contributions of individual institutions to the overall risk of systemic network events. Similarly, from-degree of
P
a|k,V \{a,k}
institution a is defined as δaf rom = ∆CoV aRa|sys = k ∆CoV aRτ
. From-degrees measure exposure
P
of individual institutions to systemic shocks from the network. The total degree δ := a ∆CoV aRsys|a ,
aggregates institution-specific systemic risk across institutions hence provides a measure of total systemic risk
in the whole financial system. Finally, for institution a, we define its net contribution as net-∆CoV aRa =
f rom
δ to
. For more about network analysis, see [52].
a − δa
20
BELLONI, CHEN, AND CHERNOZHUKOV
5.2. International Financial Contagion. In this section we apply PQGMs to the study of international
financial contagion and then with the estimated network structure we can measure the systemic risk contributions of each country, as mentioned in Section 5.1. We focus on examining financial contagion through
the volatility spillovers perspective. [39] reported that international stock markets are related through their
volatilities instead of returns. [34] studied the return and volatility spillovers of 19 countries and found
differences in return and volatility spillovers. For a survey of financial contagion see [30]. We also illustrate
how PQGMs can highlight asymmetric dependence between the random variables.
We use daily equity index returns, September 2009 to September 2013 (1044 observations), from Morgan
Stanley Capital International (MSCI). The returns are all translated into dollar-equivalents as of September
6th 2013. We use absolute returns as a proxy for volatility. We have a total of 45 countries in our sample, there
are 21 developed markets (Australia, Austria, Belgium, Canada, Denmark, France, Germany, Hong Kong,
Ireland, Italy, Japan, Netherlands, New Zealand, Norway, Portugal, Singapore, Spain, Sweden, Switzerland,
the United Kingdom, the United States), 21 emerging markets (Brazil, Chile, Mexico, Greece, Israel, China,
Colombia, Czech Republic, Egypt, Hungary, India, Indonesia, Korea, Malaysia, Peru, Philippines, Poland,
Russia, Taiwan, Thailand, Turkey), and 3 frontier markets (Argentina, Morocco, Jordan).
Below in Figure 1 we provide a full-sample analysis of global volatility spillovers at different tails. We
denote 20% quantile as Low Tail, 50% quantile as Median, 80% quantile as Up Tail. Both PQGMs and GGM
are presented. Our purpose is to show the usefulness of PQGM in representing nonlinear tail interdependence
allowing for heteroscedasticity and to show that PQGM can measure correlation asymmetry through looking
at the tails of the distribution (not specific to any model).
There are significant differences in the network structure in terms of volatility spillovers when using
PQGM and GGM. PQGM permits asymmetries in correlation dynamics, suited to investigate the presence
of asymmetric responses. We find significant increase interdependence at the up tail between the volatility
series, i.e. we find downside correlations (high volatility) are much larger than upside correlations (low
volatility). This confirms findings in the finance literature that financial markets become more interdependent
during high volatility periods.
We also find if two countries locate in the same geographic region, with many similarities in terms of
market structure and history, they tend to be more closely connected (homophily effect as stated in network
terminology); while two economies locate in separate geographic regions are less likely directly connected.
In addition, we find among European Union member countries, Germany appears to play a major role in the
transmission of shocks to others; while in Asia, Hong Kong, Thailand, and Singapore appear to play major
roles; and among all the north and south American countries, Canada and US play major roles.
In addition, we present net-∆CoV aR discussed in Section 5.1 with τ = 0.8, i.e. the Up Tail, in Figure 2
which shows that: globally, total volatility spillovers from Germany and France to the others are much larger
than total volatility spillovers from the others to them, and their net-∆CoV aR are positive. Both Greece and
Spain have negative net-∆CoV aR. The estimated network structure is important here as it demonstrates
that shocks originated in some markets may be amplified during their transmission throughout the system,
posing greater risks to the whole market than other shocks’ origination.
QUANTILE GRAPHICAL MODELS
21
Gaussian Graph
Median
TWN
TWN
HKG
HKG
MYS
MYS
CHN
EGY
SGP
KOR
CHN
EGY
SGP
PHL
JPN
KOR
IDN
IDN
NZL
NZL
IND
IND
JOR
ARG
PHL
JPN
JOR
THA
ISR
GRC
AUS
ARG
THA
ISR
ESP
GRC
AUS
COL
IRL
CAN
GBR
AUT
CAN
ITA
BEL
USA
NLD
SWE
NLD
FRA
DEU
PRT
CZE
SWE
FRA
DEU
MAR
ITA
BEL
TUR
PRT
CZE
AUT
CHE
MEX
TUR
USA
POL
GBR
CHE
DNK
NOR
RUS
POL
MEX
IRL
HUN
BRA
DNK
NOR
RUS
COL
PER
HUN
BRA
MAR
Up Tail
Low Tail
TWN
TWN
HKG
HKG
MYS
MYS
CHN
EGY
SGP
KOR
CHN
EGY
SGP
PHL
JPN
KOR
IDN
ARG
PHL
JPN
IDN
NZL
NZL
IND
JOR
ESP
CHL
CHL
PER
IND
JOR
THA
ISR
GRC
AUS
ARG
THA
ISR
ESP
GRC
AUS
CHL
ESP
CHL
COL
PER
IRL
HUN
BRA
CAN
GBR
POL
AUT
BEL
ITA
GBR
POL
MAR
USA
FRA
SWE
NLD
DEU
FRA
MAR
Figure 1. International Financial Contagion. Note: we show the volatility transmission channel
is asymmetric (at different tails).
ITA
PRT
CZE
NLD
DEU
BEL
TUR
PRT
CZE
SWE
AUT
CHE
MEX
TUR
USA
CAN
DNK
NOR
RUS
CHE
MEX
IRL
HUN
BRA
DNK
NOR
RUS
COL
PER
BELLONI, CHEN, AND CHERNOZHUKOV
2.0
22
1.5
DEU
FRA
1.0
CHE
HKG
0.5
ITA
BRA
SWE
GBR
NLD
MYS
JOR
PRTSGP
ISR
MAR
NOR
CHN
USA
CZE
0.0
BEL
COL
JPN
IDN
NZL
MEX
CHL
KOR
PHL
THA
TWN
AUS
PER
DNK
−0.5
EGY
ESP
IND
RUS
IRL
ARG
HUN
TUR
POL
AUT
−1.0
Net ∆CoVaR
CAN
GRC
Countries
Figure 2. Volatility Spillovers Net Contribution of Each Country
QUANTILE GRAPHICAL MODELS
23
5.3. Stock Returns Under Market Downside Movements. Stock returns are in general non-Gaussian.
[6] find correlation asymmetries in the data and reject the null hypothesis of multivariate Gaussian distributions at daily, weekly, and monthly frequencies, conditional on market “downside” movements. See also
[61, 67], among other studies, in the empirical finance literature for evidence of the non-Gaussian feature of
financial markets. Hence, generally, in the financial market context, linear correlation measures only convey
partial and often misleading information on the actual underlying conditional dependencies.
In fact, when we are particularly interested in the conditional dependencies of stocks returns conditioning on market “downside” movements, this can be modeled by the W-Conditional CIQGM
with W = {Market return 6 $}, $ as the τm -th quantile of the market index returns, and τm =
{0.15, 0.5, 0.75, 0.9, 1}. In the following example, we obtain daily stock returns from CRSP and use S&P
500 as the market index. The full sample consists of 2769 observations for 86 stocks from Jan 2, 2003 to
December 31, 2013. The total number of stocks is 86 due to data availability at CRSP. In this case, we
define market downside movement as when the market index returns are below a pre-specified level (e.g.
τm -th quantile), hence conditioning on a particular $ simply corresponds to consider the subsample based
on whether the corresponding date’s market return is less equal to the τm -th quantile of the market index
returns. We reported the number of edges (there is no linkage between two stocks if they are conditional
independent), under different market conditions in Table 1 below.
Table 1. Edges Produced by Different Graph Estimators
τm
$
obs
PQGM
CIQGM
CIQGM(0.1)
CIQGM(0.3)
0.15
416
82
7302
4086
388
0.5
1385
196
7254
1308
74
0.75
2077
238
7226
908
50
0.9
2492
272
7202
740
46
1
2769
304
7274
784
54
>0
1246
186
7262
1516
80
<0
1522
180
7254
1308
64
Note: The results are computed by Algorithm 3.3 and 3.4. CIQGM(0.1)
means additional thresholding at 0.1, i.e. we keep the edges that have
correlation stronger than 0.1. Similarly, CIQGM(0.3) means additional
thresholding at 0.3.
As shown in Table 1, there are significant differences between PQGM and CIQGMs, hence assuming
Gaussianity for the distribution of stock returns could result in false correlation conclusions (e.g., estimation
bias due to asymmetry in correlations conditional on market upside or downside moves). When using PQGM,
the number of edges increases with τm . The number of edges in CIQGM is significantly higher when no
thresholding is applied. The results also show stronger correlation (and more connections) under market
downside moments. All those empirical findings support evidence from the empirical finance literature.
24
BELLONI, CHEN, AND CHERNOZHUKOV
References
[1] Alberto Abadie, Guido Imbens, and Fanyin Zheng. Inference for misspecified models with fixed regressors. Journal of the
American Statistical Association, 109(508):1601–1614, 2014.
[2] Daron Acemoglu, Asuman Ozdaglar, and Alireza Tahbaz-Salehi. Cascades in networks and aggregate volatility. Technical
report, National Bureau of Economic Research, 2010.
[3] Daron Acemoglu, Asuman Ozdaglar, and Alireza Tahbaz-Salehi. Systemic risk and stability in financial networks. The
American Economic Review, 105(2):564–608, 2015.
[4] Tobias Adrian and Markus Brunnermeier. Covar. The American Economic Review, 106(7):1705–1741, 2016.
[5] Torben Andersen, Tim Bollerslev, Peter Christoffersen, and Francis Diebold. Financial risk measurement for financial risk
management. Handbook of the Economics of Finance, 2:1127–1220, 2013.
[6] Andrew Ang and Joseph Chen. Asymmetric correlations of equity portfolios. Journal of Financial Economics, 63(3):443–
494, 2002.
[7] Andrew Ang, Joseph Chen, and Yuhang Xing. Downside risk. Review of Financial Studies, 19(4):1191–1239, 2006.
[8] Joshua Angrist, Victor Chernozhukov, and Iván Fernández-Val. Quantile regression under misspecification, with an application to the us wage structure. Econometrica, 74(2):539–563, 2006.
[9] Onureena Banerjee, Laurent El Ghaoui, and Alexandre d’Aspremont. Model selection through sparse maximum likelihood
estimation for multivariate gaussian or binary data. The Journal of Machine Learning Research, 9:485–516, 2008.
[10] Alexandre Belloni, Daniel Chen, Victor Chernozhukov, and Christian Hansen. Sparse models and methods for optimal
instruments with an application to eminent domain. Econometrica, 80(6):2369–2429, 2012.
[11] Alexandre Belloni and Victor Chernozhukov. `1 -penalized quantile regression for high dimensional sparse models. Annals
of Statistics, 39(1):82–130, 2011.
[12] Alexandre Belloni and Victor Chernozhukov. Inference methods for high-dimensional sparse econometric models. Advances
in Economics and Econometrics, 10th World Congress of Econometric Society, III:245–295, 2013.
[13] Alexandre Belloni and Victor Chernozhukov. Least squares after model selection in high-dimensional sparse models.
Bernoulli, 19(2):521–547, 2013. ArXiv, 2009.
[14] Alexandre Belloni, Victor Chernozhukov, Denis Chetverikov, and Ying Wei. Uniformly valid post-regularization confidence
regions for many functional parameters in z-estimation framework. arXiv:1512.07619, 2015.
[15] Alexandre Belloni, Victor Chernozhukov, and Iván Fernández-Val. Conditional quantile processes based on series or many
regressors. arXiv, 2011.
[16] Alexandre Belloni, Victor Chernozhukov, Ivan Fernández-Val, and Christian Hansen. Program evaluation and causal
inference with high-dimensional data. Econometrica, 85(1):233–298, 2017.
[17] Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. Inference on treatment effects after selection among highdimensional controls. The Review of Economic Studies, 81(2):608–650, 2014.
[18] Alexandre Belloni, Victor Chernozhukov, and Kengo Kato. Robust inference in high-dimensional approximately sparse
quantile regression models. arXiv preprint arXiv:1312.7186, 2013.
[19] Alexandre Belloni, Victor Chernozhukov, and Kengo Kato. Uniform post-selection inference for least absolute deviation
regression and other z-estimation problems. Biometrika, 102(1):77–94, 2015.
[20] Alexandre Belloni, Victor Chernozhukov, and Lie Wang. Square-root-lasso: Pivotal recovery of sparse signals via conic
programming. Biometrika, 98(4):791–806, 2011. Arxiv, 2010.
[21] Peter Bickel, Yaacov Ritov, and Alexandre Tsybakov. Simultaneous analysis of Lasso and Dantzig selector. Annals of
Statistics, 37(4):1705–1732, 2009.
[22] Tony Cai, Weidong Liu, and Xi Luo. A constrained `1 minimization approach to sparse precision matrix estimation. Journal
of the American Statistical Association, 106(494), 2011.
[23] Mehmet Caner and Anders Kock. Asymptotically honest confidence regions for high dimensional parameters by the desparsified conservative lasso. Journal of Econometrics, forthcoming.
[24] Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Gaussian approximations and multiplier bootstrap for maxima
of sums of high-dimensional random vectors. The Annals of Statistics, 41(6):2786–2819, 2013.
[25] Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Gaussian approximation of suprema of empirical processes.
The Annals of Statistics, 42(4):1564–1597, 2014.
QUANTILE GRAPHICAL MODELS
25
[26] Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Comparison and anti-concentration bounds for maxima of
gaussian random vectors. Probability Theory and Related Fields, 162:47–70, 2015.
[27] Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Empirical and multiplier bootstraps for supreme of empirical processes of increasing complexity, and related gaussian couplings. Stochastic Processes and their Applications,
126(12):3632–3651, 2016.
[28] Victor Chernozhukov, Christian Hansen, and Martin Spindler. Post-selection and post-regularization inference in linear
models with very many controls and instruments. American Economic Review: Papers and Proceedings, 105(5):486–490,
2015.
[29] Khai Chiong and Roger Moon. Estimation of graphical lasso using the l1, 2 norm. The Econometrics Journal, forthcoming.
[30] Stijn Claessens and Kristin Forbes. International Financial Contagion. Springer, 2001.
[31] David Cox and Nanny Wermuth. Multivariate dependencies: Models, analysis and interpretation, volume 67. CRC Press,
1996.
[32] Philip Dawid. Conditional independence in statistical theory. Journal of the Royal Statistical Society. Series B (Methodological), pages 1–31, 1979.
[33] Arthur Dempster. Covariance selection. Biometrics, pages 157–175, 1972.
[34] Francis Diebold and Kamil Yilmaz. Measuring financial asset return and volatility spillovers, with application to global
equity markets. The Economic Journal, 119(534):158–171, 2009.
[35] Mathias Drton and Michael Perlman. Model selection for gaussian concentration graphs. Biometrika, 91(3):591–602, 2004.
[36] Mathias Drton and Michael Perlman. Multiple testing and error control in gaussian graphical model selection. Statistical
Science, 22(3):430–449, 2007.
[37] Mathias Drton and Michael Perlman. A sinful approach to gaussian graphical model selection. Journal of Statistical
Planning and Inference, 138(4):1179–1200, 2008.
[38] David Edwards. Introduction to graphical modelling. Springer, 2000.
[39] Robert Engle and Raul Susmel. Common volatility in international equity markets. Journal of Business & Economic
Statistics, 11(2):167–176, 1993.
[40] Jianqing Fan, Jinchi Lv, and Lei Qi. Sparse high dimensional models in economics. Annual review of economics, 3:291,
2011.
[41] Max Farrell. Robust inference on average treatment effects with possibly more covariates than observations. Journal of
Econometrics, 174(2):1–23, 2015.
[42] Rina Foygel and Mathias Drton. Extended bayesian information criteria for gaussian graphical models. In Advances in
Neural Information Processing Systems, pages 604–612, 2010.
[43] Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Sparse inverse covariance estimation with the graphical lasso.
Biostatistics, 9(3):432–441, 2008.
[44] Xuming He, Lan Wang, and Hyokyoung Grace Hong. Quantile-adaptive model-free variable screening for high-dimensional
heterogeneous data. The Annals of Statistics, 41(1):342–369, 2013.
[45] Jana Jankova and Sara van de Geer. Confidence intervals for high-dimensional inverse covariance estimation. Electronic
Journal of Statistics, 9(1):1205–1229, 2015.
[46] Harry Joe. Multivariate models and multivariate dependence concepts. CRC Press, 1997.
[47] Kengo Kato. Group lasso for high dimensional sparse quantile regression models. Preprint, ArXiv, 2011.
[48] Keith Knight. Limiting distributions for L1 regression estimators under general conditions. The Annals of Statistics,
26:755–770, 1998.
[49] Keith Knight. Asymptotics of the regression quantile basic solution under misspecification. Applications of Mathematics,
[50]
[51]
[52]
[53]
[54]
53(3):223–234, 2008.
Roger Koenker. Quantile regression. Cambridge University Press, New York, 2005.
Roger Koenker and Gilbert Bassett. Regression quantiles. Econometrica, 46(1):33–50, 1978.
Eric Kolaczyk. Statistical analysis of network data. Springer, 2009.
Steffen Lauritzen. Graphical models. Oxford University Press, 1996.
Michel Ledoux and Michel Talagrand. Probability in Banach Spaces (Isoperimetry and processes). Ergebnisse der Mathematik undihrer Grenzgebiete, Springer-Verlag, 1991.
[55] Ying-Ying Lee. Efficiency bounds for semiparametric estimation of quantile regression under misspecification. University
of Wisconsin-Madison, 2009.
26
BELLONI, CHEN, AND CHERNOZHUKOV
[56] Hannes Leeb and Benedikt Pötscher. Can one estimate the unconditional distribution of post-model-selection estimators?
Econometric Theory, 24(02):338–376, 2008.
[57] Han Liu, Fang Han, Ming Yuan, John Lafferty, and Larry Wasserman. High-dimensional semiparametric gaussian copula
graphical models. The Annals of Statistics, 40(4):2293–2326, 2012.
[58] Han Liu, John Lafferty, and Larry Wasserman. The nonparanormal: Semiparametric estimation of high dimensional
undirected graphs. The Journal of Machine Learning Research, 10:2295–2328, 2009.
[59] Han Liu and Lie Wang. Tiger: A tuning-insensitive approach for optimally estimating gaussian graphical models. Electronic
Journal of Statistics, 11(1):241–294, 2017.
[60] Weidong Liu and Xi Luo. High-dimensional sparse precision matrix estimation via sparse column inverse operator. arXiv
preprint arXiv:1203.3896, 2012.
[61] Francois Longin and Bruno Solnik. Extreme correlation of international equity markets. The Journal of Finance, 56(2):649–
676, 2001.
[62] Nicolai Meinshausen and Peter Bühlmann. High-dimensional graphs and variable selection with the lasso. The Annals of
Statistics, pages 1436–1462, 2006.
[63] Sahand Negahban, Pradeep Ravikumar, Martin Wainwright, and Bin Yu. A unified framework for high-dimensional analysis
of m-estimators with decomposable regularizers. Statistical Science, 27(4):538–557, 2012.
[64] Yang Ning and Han Liu. A general theory of hypothesis tests and confidence regions for sparse high dimensional models.
The Annals of Statistics, 45(1):158–195, 2017.
[65] Roberto Imbuzeiro Oliveira. The lower tail of random quadratic forms, with applications to ordinary least squares and
restricted eigenvalue properties. Probability Theory and Related Fields, 166(3-4):1175–1194, 2016.
[66] Andrew Patton. On the out-of-sample importance of skewness and asymmetric dependence for asset allocation. Journal of
Financial Econometrics, 2(1):130–168, 2004.
[67] Andrew Patton. Modelling asymmetric exchange rate dependence. International economic review, 47(2):527–556, 2006.
[68] Joseph Romano and Michael Wolf. Exact and approximate stepdown methods for multiple hypothesis testing. Journal of
the American Statistical Association, 100(469):94–108, 2005.
[69] Tingni Sun and Cun-Hui Zhang. Sparse matrix inversion with scaled lasso. The Journal of Machine Learning Research,
14(1):3385–3418, 2013.
[70] Sara van de Geer. High-dimensional generalized linear models and the lasso. Annals of Statistics, 36(2):614–645, 2008.
[71] Sara Van de Geer, Peter Bühlmann, Yaacov Ritov, and Ruben Dezeure. On asymptotically optimal confidence regions and
tests for high-dimensional models. The Annals of Statistics, 42(3):1166–1202, 2014.
[72] Aad van der Vaart and Jon Wellner. Weak Convergence and Empirical Processes. Springer Series in Statistics, 1996.
[73] Aad van der Vaart and Jon Wellner. Empirical process indexed by estimated functions. IMS Lecture Notes-Monograph
Series, 55:234–252, 2007.
[74] Jialei Wang and Mladen Kolar. Inference for high-dimensional exponential family graphical models. In Proc. of AISTATS,
volume 51, pages 751–760, 2016.
[75] Yuanshan Wu and Guosheng Yin. Conditional quantile screening in ultrahigh-dimensional heterogeneous data. Biometrika,
102(1):65–76, 2015.
[76] Lingzhou Xue and Hui Zou. Regularized rank-based estimation of high-dimensional nonparanormal graphical models. The
Annals of Statistics, 40(5):2541–2571, 2012.
[77] Lingzhou Xue, Hui Zou, and Tianxi Cai. Nonconcave penalized composite conditional likelihood estimation of sparse ising
models. The Annals of Statistics, 40(3):1403–1429, 2012.
[78] Ming Yuan. High dimensional inverse covariance matrix estimation via linear programming. The Journal of Machine
Learning Research, 99:2261–2286, 2010.
[79] Ming Yuan and Yi Lin. Model selection and estimation in the gaussian graphical model. Biometrika, 94(1):19–35, 2007.
[80] Cun-Hui Zhang and Stephanie Zhang. Confidence intervals for low dimensional parameters in high dimensional linear
models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(1):217–242, 2014.
[81] Qi Zheng, Limin Peng, and Xuming He. Globally adaptive quantile regression with ultra-high dimensional data. Annals of
statistics, 43(5):2225, 2015.
QUANTILE GRAPHICAL MODELS
27
Appendix A. Implementation Details of Algorithms
This section provides details of the algorithms mentioned in Section 3. Note, for the weighted-Lasso
b τ = diag[Γ
b τ kk , k ∈
estimator, the choice of penalty level λ := 1.1n−1/2 2Φ−1 (1 − ξ/Nn ) and penalty loading Γ
[p]\{j}] is a diagonal matrix defined by the following procedure: (1) Compute the post-Lasso estimator
j
b τ kk = maxkfiaτ Z a k∞ {En [|faτ Z a |2 ]}1/2 . (2) Compute the residuals
γ
eaτ
based on λ and initial values Γ
i
k
i6n
a
a
j
vbiaτ j = fiaτ (Zij
− Zi,−j
γ
eaτ
) and update the loadings
b τ kk =
Γ
q
2 |Z a v
2
En [faτ
k baτ j | ], k ∈ [p]\{j}
(A.34)
j
and use them to recompute the post-Lasso estimator γ
eaτ
. In the case of Algorithm 3.1 we can take Nn =
3 3
Z
|V |p n , in the case of Algorithm 3.3 we take Nn = |V |p2 {pn3 }1+dW . Denote σ
baj
= {En [(Zja )2 ]}1/2 ,
X
2
Z
X
2
σ
baj
= {En [X−a,j
]}1/2 , σ
ba$j
= {En [K$ (W )(Zja )2 ]}1/2 , and σ
ba$j
= {En [K$ (W )X−a,j
]}1/2 .
Detailed version of Algorithm 3.1 (CIQGM)
For each a ∈ V , τ ∈ T , and j ∈ [p], perform the following:
a e
(1) Run Post-`1 -quantile regression of Xa on Z a ; keep fitted value Z−j
βaτ,−j ,
p
P
p
Z
βbaτ ∈ arg minβ En [ρτ (Xa − Z a β)] + λV T τ (1 − τ ) j=1 σ
baj
|β |
pj
Z
a
e
b
σaj
.
βaτ ∈ arg minβ En [ρτ (Xa − Z β)] : βj = 0 if |βaτ j | 6 λV T τ (1 − τ )/b
a
a
a
j
},
(2) Run Post-Lasso of faτ Zja on faτ Z−j
; keep the residual vei := fiaτ {Zij
− Zi,−j
γ
eaτ
a
2
j
b τ γk1
(Zja − Z−j
γ)2 ] + λkΓ
∈ arg minγ En [faτ
γ
baτ
a
a
2
j
2
j
).
γaτ
γ
eaτ ∈ arg minγ En [faτ (Zj − Z−j γ) ] : support(γ) ⊆ support(b
a e
(3) Run Instrumental Quantile Regression of Xa − Z−j
βaτ,−j on Zja using ve as the instrument for Zja ,
β̌aτ,j ∈ arg min
α∈Aaτ j
with Aaτ j
a e
{En [(1{Xa 6 Zja α + Z−j
βaτ,−j } − τ )e
v ]}2
,
En [(1{Xa 6 Z a α + Z a βeaτ,−j } − τ )2 ve2 ]
j
−j
Z
= {α ∈ R : |α − βeaτ j | 6 10/{b
σaj
log n}}.
Detailed version of Algorithm 3.2 (PQGM)
For each a ∈ V , and τ ∈ T , perform the following:
(1) Run Post-`1 -quantile regression of Xa on X−a ,
P
0
X
βbaτ ∈ arg minβ En [ρτ (Xa − X−a
β)] + λ0 j∈[d] σ
baj
|βj |
0
X
e
b
βaτ ∈ arg minβ En [ρτ (Xa − X−a β)] : βj = 0 if |βaτ j | 6 λ0 /b
σaj
.
0
β̃aτ } − τ for i ∈ [n]. Compute the penalty level λ̄V T via (3.19).
(2) Set εbiaτ = 1{Xia 6 Xi,−a
(3) Run Post-`1 -quantile regression of Xa on X−a ,
P
0
2
βbaτ ∈ arg minβ En [ρτ (Xa − X−a
β)] + λ̄V T j∈[d] {En [b
ε2aτ X−a,j
]}1/2 |βj |
0
2
β̌aτ ∈ arg minβ En [ρτ (Xa − X−a β)] : βj = 0 if |βbaτ j | 6 λ̄V T /{En [b
ε2aτ X−a,j
]}1/2 .
28
BELLONI, CHEN, AND CHERNOZHUKOV
Detailed version of Algorithm 3.3 (W-Conditional CIQGM)
For each u = (a, τ, $) ∈ U = V × T × W, and j ∈ [p], perform the following:
a e
(1) Run Post-`1 -quantile regression of Xa on Z a ; keep fitted value Z−j
βu,−j ,
P
p
Z
|βj |
βbu ∈ arg minβ En [K$ (W )ρτ (Xa − Z a β)] + λu j=1 σ
ba$j
Z
βeu ∈ arg minβ En [K$ (W )ρτ (Xa − Z a β)] : βj = 0 if |βbuj | 6 λu /b
σa$j
.
a
a
(2) Run Post-Lasso of fu Zja on fu Z−j
; keep the residual ve := fu (Zja − Z−j
γ
euj ),
a
b u γk1
γ
buj ∈ arg minθ En [K$ (W )fu2 (Zja − Z−j
γ)2 ] + λkΓ
a
γ
euj ∈ arg minγ En [K$ (W )fu2 (Zja − Z−j
γ)2 ] : support(γ) ⊆ support(b
γuj ).
a e
(3) Run Instrumental Quantile Regression of Xa − Z−j
βu,−j on Zja using ve as the instrument,
β̌uj ∈ arg min
α∈Auj
a e
{En [K$ (W )(1{Xa 6 Zja α + Z−j
βu,−j } − τ )e
v ]}2
En [K$ (W )(1{Xa 6 Z a α + Z a βeu,−j } − τ )2 ve2 ]
j
−j
Z
where Auj := {α ∈ R : |α − βeuj | 6 10/{b
σa$j
log n}}.
Detailed version of Algorithm 3.4 (W-Conditional PQGM)
For each u = (a, τ, $) ∈ U = V × T × W perform the following:
(1) Run Post-`1 -quantile regression of Xa on X−a ,
P
0
X
β)] + λ0W j∈[d] σ
βbu ∈ arg minβ En [K$ (W )ρτ (Xa − X−a
ba$j
|βj |
X
0
b
e
σa$j
.
βu ∈ arg minβ En [K$ (W )ρτ (Xa − X−a β)] : βj = 0 if |βuj | 6 λ0W /b
0
(2) Set εbiu = 1{Xia 6 Xi,−a
βeu } − τ for i ∈ [n], compute λ̄V T W via (3.21).
(3) Run Post-`1 -quantile regression of Xa on X−a ,
P
2
0
β)] + λ̄V T W j∈[d] {En [K$ (W )b
]}1/2 |βj |.
βbu ∈ arg minβ En [K$ (W )ρτ (Xa − X−a
ε2u X−a,j
0
2
β̌u ∈ arg minβ En [K$ (W )ρτ (Xa − X−a
β)] : βj = 0 if |βbuj | 6 λ̄V T W /{En [K$ (W )b
ε2u X−a,j
]}1/2 .
Appendix B. Simulations of Quantile Graphical Models
In this section, we perform numerical examples to illustrate the performance of the estimators proposed
for QGMs. We will consider several different designs. In order to compare with other proposals we will
consider both Gaussian and non-Gaussian examples.
B.1. Isotropic Non-Gaussian Example. In general, the equivalence between a zero in the inverse covariance matrix and a pair of conditional independent variables will break down for non-gaussian distributions.
The nonparanormal graphical models extends Gaussian Graphical Models to Semiparametric Gaussian Copula models by transforming the variables with smooth functions. We illustrate the applicability of CIQGM
in representing the conditional independence structure of a set of variables when the random variables are
not even jointly nonparanormal.
Consider i.i.d. copies of an d-dimensional random vector X̃V = (X̃1 , . . . , X̃d−1 , X̃d ) from the following
multivariate normal distribution, X̃V ∼ N (0, Id×d ), where Id×d is the identity matrix. Further, we generate
q
q
2
2
π
Xd = − 3π−2
+ 3π−2
X̃d−1
|X̃d |.
(B.35)
QUANTILE GRAPHICAL MODELS
It follows that E[Xd ] =
q
π
3π−2 (E[|X̃d |]
−
p
2/π) = 0 and Var(Xd ) =
29
π
2
3π−2 (E[X̃d
4
· X̃d−1
]−
2
π)
= 1. In
addition, equation (B.35) is a location-scale-shift model in which the conditional median of the response is
zero while quantile functions other than the median are nonzero. We define vector XV as
XV = (Xd , X̃1 , ..., X̃d−1 )0 .
In this new set of variables, only Xd and X̃d−1 (i.e., node 1 and 15, when d = 15) are not conditionally
independent. Nonetheless, the covariance matrix of XV is still Id×d .
Next we consider an example with n = 300 and d = 15. We show graphs, in Figure 3 and 4, estimated by
both CIQGM(s) and GGMs in this non-Gaussian setting.
(b) CIQGM(0.5)
(a) CIQGM(0.2)
11
(c) CIQGM(0.8)
11
11
13
13
9
13
9
9
12
12
5
12
5
15
7
5
15
7
15
7
2
2
14
2
14
3
14
3
4
3
4
8
10
8
10
10
1
6
4
8
1
6
(d) CIQGM-Union
(e) Gaussian
11
(f) Nonparanormal
11
11
13
13
9
13
9
9
12
12
5
12
5
15
7
5
15
7
15
7
2
2
14
2
14
3
14
3
4
3
4
8
10
1
4
8
10
6
1
6
8
10
6
1
6
1
Figure 3. QGM and GGM
In Figure 3, Gaussian means the graph is estimated by using graphical lasso without any transformation
of XV , and the final graph is chosen by Extended Bayesian Information Criterion (ebic), see [42]. Nonparanormal means the graph is estimated using graphical lasso (likelihood based approach) with nonparanormal
30
BELLONI, CHEN, AND CHERNOZHUKOV
transformation of XV , see [58], and again the final graph is chosen by ebic. Both graphs are estimated using
R-package huge.
In Figure 4, as a robustness check, we also compare results produced by CIQGM with those produced by
neighborhood selection methods (pseudo-likelihood approach),
e.g. TIGER of [59] in R-package flare the
q
left graph is when choosing the turning parameter to be logn d while the right graph is when choosing the
q
q
tuning parameter to be 2 logn d . Throughout, we use Tiger2 represent TIGER with penalty level 2 logn d . As
expected, GGM cannot detect the correct dependence structure when the joint distribution is non-Gaussian
while CIQGM can still represent the right conditional independence structure.
(b) Tiger2
(a) Tiger1
11
11
13
13
9
9
12
12
5
5
15
7
15
7
2
2
14
14
3
3
4
4
8
8
10
10
6
1
6
1
Figure 4. TIGER
B.2. Gaussian Examples.
B.2.1. Graph Recovery. In this subsection, we start with comparing the numerical performance of QGM and
other methods, e.g. TIGER of [59] and graphical lasso algorithm (Glasso) of [43], in graph recovery using
simulated datasets with different pairs of (n, d). We start with one simulation for illustration purpose (the
results are summarized in Figure 5), and then we show the performance of QGM through estimated degree
distribution with 100 simulations (the results are summarized in Figure 6).
We mainly consider the Hub graph, as mentioned in [59], which also corresponds to the star network
mentioned in [2, 3]. In line with [59], we generate a d-dimensional sparse graph GI = (V, E I ) represents
the conditional independence structure between the variables. In our simulations, we consider 12 settings
to compare these methods: (A) n = 200, d = 10; (B) n = 200, d = 20; (C) n = 200, d = 40; (D) n = 400,
d = 10; (E) n = 400, d = 20; (F) n = 400, d = 40; (G) n = 200, d = 100; (H) n = 200, d = 200; (I) n = 200,
d = 400; (J) n = 400, d = 100; (K) n = 400, d = 200; (L) n = 400, d = 400. We adopt the following model
for generating undirected graphs and precision matrices.
QUANTILE GRAPHICAL MODELS
31
Hub graph. The d nodes are evenly partitioned into d/20 (or d/10 when d < 20) disjoint groups with
each group contains 20 (or 10) nodes. Within each group, one node is selected as the hub and we add edges
between the hub and the other 19 (or 9) nodes in that group. For example, the resulting graph has 190 edges
when d = 200 and 380 edges when d = 400. Once the graph is obtained, we generate an adjacency matrix
E I by setting the nonzero off-diagonal elements to be 0.3 and the diagonal elements to be 0. We calculate
its smallest eigenvalue Λmin (E I ). The precision matrix is constructed as
Θ = D[E I + (|Λmin (E I )| + 0.2) · Id×d ]D
(B.36)
where D ∈ Rd×d is a diagonal matrix with Djj = 1 for j = 1, ..., d/2 and Djj = 1.5 for j = d/2 + 1, ..., d.
The covariance matrix Σ := Θ−1 is then computed to generate the multivariate normal data: X1 , ...., Xd ∼
N (0, Σ). Below we provide simulation results using different estimators: PQGM6, TIGER and Glasso. We
start with one simualtion as an illustration:
(a) n = 200, d = 10
(b) n = 200, d = 20
(c) n = 200, d = 40
QR−Union
TIGER
QR−Union
TIGER
QR−Union
TIGER
Glasso
True
Glasso
True
Glasso
True
(d) n = 400, d = 10
(e) n = 400, d = 20
(f) n = 400, d = 40
QR−Union
TIGER
QR−Union
TIGER
QR−Union
TIGER
Glasso
True
Glasso
True
Glasso
True
Figure 5. One simulation
6Given the graphs are generated from multivariate Gaussian distribution we can use PQGM to simplify the computation.
32
BELLONI, CHEN, AND CHERNOZHUKOV
(g) n = 200, d = 100
(h) n = 200, d = 200
(i) n = 200, d = 400
QR−Union
TIGER
QR−Union
TIGER
QR−Union
TIGER
Glasso
True
Glasso
True
Glasso
True
(j) n = 400, d = 100
(k) n = 400, d = 200
(l) n = 400, d = 400
QR−Union
TIGER
QR−Union
TIGER
QR−Union
TIGER
Glasso
True
Glasso
True
Glasso
True
Figure 5. One simulation (Cont.)
Figure 5 shows that: for the low dimensional cases, d = 10, 20, 40, n is large compared to d, CIQGM is
comparable to TIGER and both are better than Glasso in terms of false positives; for the high dimensional
cases, d = {100, 200, 300}, we can compare the performance of different graph estimators through looking at
the denseness of the estimated graph (e.g., whether it is even or not), and again, both CIQGM and TIGER
perform well in terms of graph recovery as compared to Glasso, and their performance are getting better
when n is increasing.
In what follows, Figure 6 shows the degree distribution of true graph, the estimated ones, and the standard
deviations of the degree difference (between the true graph and the estimated ones). It is based on simulations
of Hub graph with n = 500 and d = 40. Simulated 100 times.
QUANTILE GRAPHICAL MODELS
33
40
20
0
Number of vertices
True Degree Sequence
5
10
15
Degree
40
20
0
Number of vertices
Estimated Degree of Sequence
5
10
15
Degree
40
20
0
Number of vertices
Standard Deviations of Degree Differences
0.5
1.0
1.5
SD
Figure 6. Upper panel shows the degree distribution of the true graphs. Middle panel
shows the degree distribution of the estimated graphs. Bottom panel shows the standard
deviations of the degree difference (between the true graph and the estimated ones). Hub
graph with n = 500 and d = 40. Simulated 100 times.
B.2.2. Inference. In this subsection Table 2 shows the numerical performance of CIQGM, based on Algorithm
3.1, on estimating Erdős-Rényi random graphs. More precisely, we construct approximate 90% confidence
intervals for βab with τ = 0.5, and we report the coverage probabilities. Note, in the jointly Gaussian
distributed case, we have closed form solution of βab as shown in Example 2.
p
Erdős-Rényi random graph. We add an edge between each pair of nodes with probability log d/n/d
independently. Once the graph is obtained, we construct the adjacency matrix E I and generate the precision
matrix Θ using (B.36) but setting Djj = 1 for j = 1, ..., d/2 and Djj = 1.5 for j = d/2 + 1, ..., d. We then
invert Θ to get the covariance matrices Σ := Θ−1 and generate the multivariate Gaussian data: X1 , ...., Xd ∼
N (0, Σ).
34
BELLONI, CHEN, AND CHERNOZHUKOV
Table 2. Erdős-Rényi Random Graph
d =20
d=50
d=100
(a, b)
n = 200
n = 500
n = 1000
(1,20)
84.0
87.5
91.5
(10,11)
84.0
88.0
92.5
(19, 20)
86.5
86.0
90.0
ACP
86.3
89.4
89.8
(1,50)
86.5
93.0
90.5
(25,26)
88.0
87.0
91.0
(49, 50)
87.5
90.5
87.5
ACP
86.7
89.4
90.1
(1,100)
82.5
81
89.0
(50,51)
85.0
86
92.0
(99, 100)
78.5
84
87.0
ACP
86.5
86.8
90
(a, b), coverage probability for βab ; ACP, average coverage
probability for βab , with a ∈ V , b ∈ V \{a}. Simulated 200
times.
Appendix C. Proofs of Section 4
Proof of Theorem 1. By Lemma 6, under Condition CI, for any θ such that kθk0 6 Cs`n , `n → ∞ slowly,
we have that
p
k fu Z a θkn,$ /{E[K$ (W )fu (Z a θ)2 ]}1/2 = 1 + oP (1).
Moreover, E[K$ (W )fu (Z a θ)2 ] > f u E[K$ (W )(Z a θ)2 ], E[K$ (W )(Z a θ)2 ] = E[(Z a θ)2 |$]P($), and
E[(Z a θ)2 | $] > ckθk2 by Condition CI. Lemma 6 further implies that the ratio of the minimal and maximal
eigenvalues of order s`n are bounded away from zero and from above uniformly over $ ∈ W and a ∈ V
with probability 1 − o(1). Therefore, since c{P($)}1/2 kδk1 6 kδk1,$ 6 C{P($)}1/2 kδk1 , we have κu,2c > c
uniformly over u ∈ U with the same probability for n large enough, see for instance [21].
To establish rates of convergence of the estimator obtained in Step 1 we will apply Lemma 1. Consider
the events Ω1 , Ω2 , and Ω3 as defined in (D.49), (D.50) and (D.51). By the choice of λu we have P(Ω1 ) >
1 − o(1). By Condition CI with R̄uξ 6 Cs log(p|V |n)/n and Lemma 2 we have P(Ω2 ) > 1 − o(1). Moreover,
p
P(Ω3 ) > 1 − o(1) by Lemma 3 with t3 6 Cn−1/2 (1 + dW ) log(p|V |nLf ).
QUANTILE GRAPHICAL MODELS
35
Using the same argument (with Z a replacing X−a ) as in (C.46), (C.47), and (C.48), for
p
p
δ ∈ Au := ∆$,2c ∪ {v : kvk1,$ 6 2cR̄uξ /λu , k fu Z a vkn,$ > C s(1 + dW ) log(p|V |n)/n/κu,2c },
with the restricted set defined as ∆u,2c̃ = {δ : kδTuc k1 6 2c̃kδTu k1 } for u ∈ U. we have q̄Au >
1/2 √
c(f 3/2
/f¯0 )µW /{ s maxa∈V,i6n kZia k∞ } where maxa∈V,i6n kZia k∞ .P Mn . Thus the conditions on q̄Au
U
are satisfied since Condition CI assumes Mn2 s2 log(p|V |n) 6 nµW f 3U . The conditions on the approximation
error are assumed in Condition CI.
Therefore, setting ξ = 1/ log n, by Lemma 1 we have uniformly over u = (a, τ, $) ∈ U
q
p
√
√
W ) log(p|V |n)
k fu Z a (βbu − βu )kn,$ . (1 + (t3 /λu )R̄uξ + (λu + t3 ) s . s(1+dnτ
(1−τ )
q
(1+dW ) log(p|V |n)
b
kβu − βu k1,$ . s
n
q
(1+dW ) log(p|V |n)
. Indeed by Lemma
n
Λaτ $ (1 − ξ/{|V |n1+2dW }|X−a , W ) under
here we used that λu 6 C
1/2
(C.37)
a
15 with x̃ij = K$ (Wi )Zij
and σ
bj >
cP($) , we can bound
Mn2 log(p|V |n/{τ (1 − τ )}) = o(nτ (1 −
τ )µW ) for all τ ∈ T and $ ∈ W, and the bound on λu follows from the union bound.
Let δu = βbu − βu . By triangle inequality it follows that
p
{E[K$ (W )fu (Z a δu )2 ]}1/2 6 k fu Z a δu kn,$ + kδu k1 {|(En − E)[K$ (W )fu (Z a δu )2 ]/kδu k21 ]|}1/2
(C.38)
and the last term can be bounded by
sup |(En − E)[K$ (W )fu (Z a δ)2 /kδk21 ]|
kδk1 61
6 maxk,j |(En − E)[K$ (W )fu Zka Zja ]|
q
|n)
. (1+dW ) log(p|V
n
with probability 1 − o(1) by Lemma 19 under our conditions.
Combining the relations above with (C.38), under (1 + dW )s2 log(p|V |n) = o(n) we have uniformly over
u∈U
kδu k . {E[(Z a δu )2 | $]}1/2 . {P($)}−1/2 {E[K$ (W )(Z a δu )2 ]}1/2
. {P($)f u }−1/2 {E[K$ (W )fu (Z a δu )2 ]}1/2
q
√
|n)
6 {P($)f u }−1/2 k fu Z a δu kn,$ + {P($)f u }−1/2 4 (1+dW ) log(p|V
kδu k1
n
r
s(1+dW ) log(p|V |n)
6C
.
nf P($)
u
given kδu k1 6 kδu k1,$ /P($)1/2 , (C.37), and s2 (1 + dW ) log(p|V |n) = o(nµ4W f 2U ) assumed in Condition CI.
p
Finally, let βbuλ̄ obtained by thresholding the estimator βbu with λ̄ := (1 + dW ) log(p|V |n)/n (note that
each component is weighted by En [K$ (W )(Zja )2 ]1/2 ). By Lemma 17, we have with probability 1 − o(1)
p
kZ a (βbuλ̄ − βu )kn,$ . s(1 + dW ) log(p|V |n)/n
p
kβbuλ̄ − βu k1,$ . s (1 + dW ) log(p|V |n)/n
|support(βbuλ̄ )| . s
by the choice of λ̄ and the rates in (C.37)
36
BELLONI, CHEN, AND CHERNOZHUKOV
Proof of Theorem 2. We verify Assumption C4 and Condition WL for the weighted Lasso model with index
a
a
set U × [p] where Yu = K$ (W )Zja , Xu = K$ (W )Z−j
, θu = γ̄uj , au = (fu , r̄uj ), r̄uj = K$ (W )Z−j
(γuj −
a
a
a
γ̄uj ), Suj = K$ (W )fu2 (Zja − Z−j
γuj )Z−j
= K$ (W )fu vuj Z−j
, and wu = K$ (W )fu2 . We will take Nn =
2
3 1+dW
|V |p {pn }
in the definition of λ.
2
a
a
We first verify Condition WL. We have E[Sujk
] 6 f¯2 E[|vuj Z−j,k
|2 ] 6 f¯2 {E[|vuj |4 |Z−j,k
|4 ]}1/2 6 C by the
bounded fourth moment condition. We have that
a
E[|fu vuj Z−jk
|3 | $]1/3
E[|Sujk |3 | $]1/3
E[|Sujk |3 ]1/3
−1/6
=
{P($)}
=
{P($)}−1/6 =: Muk
a |2 | $]1/2
E[|Sujk |2 ]1/2
E[|Sujk |2 | $]1/2
E[|fu vuj Z−jk
p
By the choice of Nn and Φ−1 (1 − t) 6 C log(1/t), we have Muk Φ−1 (1 − ξ/{2pNn }) 6 Muk C(1 +
dW ) log1/2 (pn|V |) 6 Cδn n1/6 where the last inequality holds by Condition CI so Condition WL(i) holds.
To verify Condition WL(ii) we will establish the validity of the choice of Nn . We will consider u =
(a, τ, $) ∈ U and u0 = (a, τ 0 , $0 ) ∈ U. By Condition CI we have that
|fu − fu0 | 6 Lf ku − u0 k and E[|K$ (W ) − K$0 (W )|] 6 LK k$ − $0 k.
(C.39)
Further, by Lemma 5 we have
kγuj − γuj 0 k 6 Lγ {ku − u0 k + k$ − $0 k1/2 }.
(C.40)
By definition we have
a
a
Sujk − Su0 jk = {K$ (W )fu2 − K$0 (W )fu20 }{Zja − Z−j
γuj }Zka − K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}Zka
a
a
γuj |2 . Moreover,
and note that fu + fu0 6 2fu + Lku − u0 k, |Zj − Z−j
γuj | · |Zka | 6 |Zj |2 + 2|Zka |2 + |Z−j
|K$ (W )fu2 − K$0 (W )fu20 | 6 K$ (W )K$0 (W )|fu2 − fu20 | + (fu + fu0 )2 |K$ (W ) − K$0 (W )|
6 2f¯K$ (W )K$0 (W )|fu − fu0 | + 4f¯2 |K$ (W ) − K$0 (W )|
Using these relations we have |En [Sujk − Su0 jk ]| 6 (I) + (II) where
(I)
a
γuj }Zka |]
= En [|K$ (W )fu2 − K$0 (W )fu20 | · |{Zja − Z−j
2
6 maxi6n kZia k2∞ (1 + kγuj k1 )En [2f¯K$ (W )K$0 (W )|fu − fu0 | + 4f |K$ (W ) − K$0 (W )|]
√
6 (f¯ + f¯2 )C s maxi6n kZ a k2 {Lf ku − u0 k + En [|K$ (W ) − K$0 (W )|]}
i
(II)
∞
a
= En [K$0 (W )fu20 |Z−j
(γuj − γuj 0 )Zka |]
6 f¯2 En [kZ a k2∞ ]kγuj − γuj 0 k1
√
6 f¯2 En [kZ a k2∞ ] pLγ {ku − u0 k + k$ − $0 k1/2 }
Moreover, we have that maxi6n kZia k2∞ .P Mn2 . For dU = k · k, an uniform -cover of U satisfies
(6diam(U)/)1+dW > N (, U, k · k). We will set 1/ = (1 + f¯2 ){Lγ + Lf }2 pnMn2 log2 (p|V |n)/{µW f 2U } 6 pn3
so that with probability 1 − o(1), for any pair u, u0 ∈ U, ku − u0 k 6 , we have
√
|En [Sujk − Su0 jk ]| . (f¯ + f¯2 ) s maxi6n kZia k2∞ {Lf + En [|K$ (W ) − K$0 (W )|]}
√
+f¯2 En [kZia k2∞ ] pLγ { + 1/2 }
√
. δn n−1/2 {µW f 2U }1/2 + sMn log(n)En [|K$ (W ) − K$0 (W )|]
QUANTILE GRAPHICAL MODELS
37
by the choice of . To control the last term, note that W is a VC-class of events with VC dimension dW .
Thus by Lemma 19, with probability 1 − o(1)
En [|K$ (W ) − K$0 (W )|] 6 |(En − E)[|K$ (W ) − K$0 (W )|]| + E[|K$ (W ) − K$0 (W )|]
6
sup
|(En − E)[|K$ (W ) − K$0 (W )|]| + LK
$,$ 0 ∈W,k$−$ 0 k6
q
1/2 + dW log(n/)
+ LK
. dW log(n/)
n
n
which yields uniformly over u ∈ U and j ∈ [p]
1/2
|En [Sujk − Su0 jk ]| . δn n−1/2 µW f U
(C.41)
1/2
1/2
p
under dW log(n/)Mn log n = o(µW f U ) and dW log(n/)Mn log n = o(n1/2 µW f U ) assumed in Condition
CI. In turn this implies
|En [Sujk − Su0 jk ]|
6 δn n−1/2
sup
max
2 ]1/2
E[Sujk
|u−u0 |6 j,k∈[p],j6=k
2
since E[Sujk
] > cµW f 2U . Using the same choice of , similar arguments also imply
2
|E[Sujk
− Su20 jk ]|
6 δn
2 ]
E[Sujk
|u−u0 |6 j,k∈[p],j6=k
sup
max
(C.42)
To establish the last requirement of Condition WL(ii), note that
sup
max
u∈U j,k∈[p],j6=k
2
|(En − E)[Sujk
]| 6 sup
max
u∈U j,k∈[p],j6=k
2
|(En − E)[Sujk
]| + ∆n
(C.43)
2
where ∆n := supu,u0 ∈U ,ku−u0 k6 maxj,k∈[p],j6=k |(En − E)[Sujk
] − (En − E)[Su20 jk ]|.
To bound the first term, we will apply Corollary 2 with k = 1, Ub := U × [p] and the vector {(X̄)uj =
b In this case note that
Suj , (u, j) ∈ U}.
K 2 = E[max sup
max
i6n u∈U j,k∈[p],j6=k
2
Sujk
] 6 E[max
sup
i6n u∈U ,j∈[p]
|viuj |2 kfiu Zia k2∞ ] 6 f¯2 Mn2 L2n .
Therefore, by Corollary 2 and Markov inequality, we have with probability 1 − o(1) that
sup
max
u∈U j,k∈[p],j6=k
2
|(En − E)[Sujk
]| 6 Cn−1/2 Mn Ln log1/2 (p|V |n) 6 Cδn µW f U + ∆n
under Mn2 L2n log(p|V |n) 6 δn nµ2W f 2U .
To control ∆n , note that
2
2
2
|(En − E)[Sujk
] − (En − E)[Su20 jk ]| 6 |En [Sujk
− Su20 jk ]| + |E[Sujk
− Su20 jk ]|
2
6 En [|Sujk − Su0 jk |] supu∈U maxi6n |2Siu0 jk | + |E[Sujk
− Su20 jk ]|
−1/2 1/2
a
. δn n
µW f U f¯ supu∈U maxi6n |viuj |kZi k∞ + δn µW f 2U
1/2
. δn n−1/2 µW f U Mn Ln log n + δn µW f 2U
with probability 1−o(1) where we used (C.41) and (C.42). Therefore ∆n . δn µW f 2U with probability 1−o(1)
as required.
a
To verify Assumption C4(a), note that [∂θ Mu (Yu , X, θu ) − ∂θ Mu (Yu , X, θu , au )]0 δ = −fu2 K$ (W )r̄uj Z−j
δ,
so that by Cauchy-Schwartz, we have
En [∂θ Mu (Yu , X, θu ) − ∂θ Mu (Yu , X, θu , au )]0 δ
a
a
6 kfu r̄uj kn,$ kfu Z−j
δkn,$ 6 Cun kfu Z−j
δkn,$
38
BELLONI, CHEN, AND CHERNOZHUKOV
where we choose Cun so that {Cun > maxj∈[p] kfu r̄uj kn,$ : u ∈ U} with probability 1 − o(1). To bound Cun ,
by Lemma 4, uniformly over u ∈ U, j ∈ [p] we have with probability 1 − o(1)
a
kfu r̄uj kn,$ = kfu Z−j
(γuj − γ̄uj )kn,$ . f u {P($)}1/2 {n−1 s log(p|V |n)}1/2
so that setting Cun = f u {P($)}1/2 {n−1 s log(p|V |n)}1/2 suffices.
Next we show that Assumption C4(b) holds. First, by (C.43) and the corresponding bounds, note the
uniform convergence of the loadings
sup
u∈U ,j,k∈[p],j6=k
2
2
a
(|En [Sujk
] − E[Sujk
]| + |(En − E)[K$ (W )fu2 |Zja Z−jk
|2 ]|) 6 δn µW f 2U
2
2
En [Sujk
]/E[Sujk
]
so that
= 1 + oP (1). It follows that c̃ is bounded above by a constant for n large enough.
Indeed, uniformly over u ∈ U, j ∈ [p], since cf u 6 E[|fu vuj Zka |2 | $]1/2 6 Cf u , with probability 1 − o(1) we
b u0jj 6 Cf P($)1/2 so that c/C 6 kΨ
b u0 k∞ kΨ
b −1 k∞ 6 C/c.
have cf P($)1/2 6 Ψ
u
u
u0
a
θ)2 with
Assumption C4(c) follows directly from the choice of Mu (Yu , Xu , θ) = K$ (W )fu2 (Zja − Z−j
q̄Au = ∞.
The result for the rate of convergence then follows from Lemma 22, namely
r
r
b u0 k∞ s log(p|V |n)
f u P($)1/2 s log(p|V |n)
kΨ
0
j
j
+ Cun .
kfu Xu (b
γu − γu )kn,$ .
κ̄u,2c
n
κ̄u,2c
n
(C.44)
By Lemma 6 we have that for sparse vectors, kθk0 6 `n s satisfies
a
a
kfu Z−j
θk2n,$ /E[K$ (W )fu2 (Z−j
θ)2 ] = 1 + oP (1)
so that φmax (`n s, uj) 6 Cf 2u P($) and sbuj 6 minm∈Mu φmax (m, uj)L2u 6 Cs provided L2u . s{f 2u P($)}−1 .
b −1 k∞ 6 Cf −1 P($)−1/2 , so that Lu . f −1 P($)−1/2 n {Cun +
Indeed, with probability 1 − o(1), we have kΨ
u0
λ
u
u
Lun }. Moreover, we can take Cun . f u {P($)n−1 s log(p|V |n)}1/2 , and Lun . {n−1 s log(p|V |n)}1/2 in
Assumption C4 because
|{En [∂γ Mu (Yu , Xu , γ
buj ) − ∂γ Mu (Yu , Xu , γuj )]}0 δ|
γuj − γuj )}Xu0 δ]|
= 2|En [K$ (W )fu2 {Xu0 (b
0
j
j
6 2kfu Xu (b
γu − γu )kn,$ kfu Xu0 δkn,$ =: Lun kfu Xu0 δkn,$ ,
where the last inequality hold by (C.44) since κ̄u,2c > cf u {P($)}1/2 . The bound on the restricted eigenvalue
√
κ̄u,2c holds7 by arguments similar to (C.46) and using that kδk1 6 C skδk for any δ ∈ ∆u,2c , and since for
any kδk = 1, we have
cf u P($) 6 E[K$ (W )fu (Z a δ)2 ]
6 {E[K$ (W )fu2 (Z a δ)2 ]}1/2 {E[K$ (W )(Z a δ)2 ]}1/2
6 {E[K$ (W )fu2 (Z a δ)2 ]}1/2 C{P($)}1/2
where the first inequality follows from the definition of f u , kδk = 1, and Condition CI, so that we have
{E[K$ (W )fu2 (Z a δ)2 ]}1/2 > c0 f u {P($)}1/2 .
7Note that there are two restricted eigenvalues definitions, one used for the quantile regression (κ
u,2c ), and another used
here for the weighted lasso (κ̄u,2c ). It is a consequence of the use of different norms.
QUANTILE GRAPHICAL MODELS
Return to the rate of convergence we have by (C.44) and κ̄u,2c > cf u {P($)}1/2 that
r
r
f u P($)1/2 s log(p|V |n)
s log(p|V |n)
0
j
j
.
kfu Xu (b
γu − γu )kn,$ .
κ̄u,2c
n
n
39
(C.45)
and the result follows by noting that kfu Xu0 (b
γuj − γuj )kn,$ > cf u P($)1/2 kb
γuj − γuj k with probability 1 − o(1)
by arguments similar to (C.46) under Condition CI.
The sparsity result follows from Lemma 21. The result for Post Lasso follows from Lemma 20 under the
growth requirements in Condition CI.
Proof of Theorem 3. We will verify Assumptions C1 and C2, and the result follows from Theorem 5. The
estimate of the nuisance parameter is constructed from the estimators in Steps 1 and 2 of the Algorithm.
a
For each u = (a, τ, $) ∈ U and j ∈ [p], let Wuj = (W, Xa , Z a , vuj , ru ), where vuj = fu (Zja − Z−j
γuj ) and
let θuj ∈ Θuj = {θ ∈ R : |θ − βuj | 6 c/ log n} (Assumption C1(i) holds). The score function is
a
a
ψuj (Wuj , θ, ηuj ) = K$ (W ){τ − 1{Xa 6 Zja θ + Z−j
βu,−j + ru }}fu (Zja − Z−j
γuj )
where the nuisance parameter is ηuj = (βu,−j , γuj , ru ) and the last component is a function ru = ru (X).
Recall that K$ (W ) ∈ {0, 1} and let an = max(n, p, |V |). Define the nuisance parameter set Huj = {η =
(3)
(2)
(1)
(1) (2) (3)
(η (1) , η (2) , η (3) ) : kη − ηuj ke 6 τn } where kη − ηuj ke = k(δη , δη , δη )ke = max{kδη k, kδη k, E[|δη |2 ]1/2 },
and
s
τn := C sup
u∈U
1
1 ∧ fU
s log an
nµW
The differentiability of the mapping (θ, η) ∈ Θuj ×Huj 7→ Eψuj (Wuj , θ, η) follows from the differentiability
of the conditional probability distribution of Xa given XV \{a} and $. Let η = (η (1) , η (2) , η (3) ), δη =
(1)
(2)
(3)
(δη , δη , δη ), and θr̄ = θ + r̄δθ , ηr̄ = η + r̄δη .
To verify Assumption C1(v)(a) with α = 2, for any (θ, η), (θ̄, η̄) ∈ Θuj × Huj note that fXa |X−a ,$ is
uniformly bounded from above by f¯, therefore
6
a
f¯E[|Z−j
(η (2)
(2)
E[{ψuj (Wuj , θ, η) − ψuj (Wuj , θ̄, η̄)}2 ]1/2
a
a
(2) 2
(3)
(3)
a
+ f E[(Zj − Z−j η̄ ) {|η − η̄ | + |Z−j
(η (1) − η̄ (1) )| + |Zja (θ − θ̄)|}]1/2
a
+ f¯E[(Zja − Z−j
η̄ (2) )4 ]1/4 {E[|η (3) − η̄ (3) |2 ]1/4 + Ckη (1) − η̄ (1) k + |θ − θ̄|}1/2
1/2
6 C 0 |θ − θ̄|1/2 ∨ kη − η̄ke
2 1/2
− η̄ )| ]
6 Ckη (2) − η̄ (2) k
¯2
¯ 2 ]1/2 6 Ckξk
¯ for all vectors ξ,
¯ and the
for some constance C 0 < ∞ since by Condition CI we have E[|Z a ξ|
p
√
j
conditions supu∈U ,j∈[p] kγu k 6 C, supθ∈Θuj |θ| 6 C, and s log(an ) 6 δn n. This implies that kη (2) −
(2)
(2)
η̄ (2) k 6 kη (2) − ηuj k + kηuj − η̄ (2) k 6 1 so that kη (2) − η̄ (2) k 6 kη (2) − η̄ (2) k1/2 .
(1)
(3)
a
To verify Assumption C1(v)(b), let tr̄ = Zja θr̄ + Z−j
ηr̄ + ηr̄ . We have
∂r E(ψuj (Wuj , θ + rδθ , η + rδη ))|r=r̄ =
(2)
(3)
a
a (1)
−E[K$ (W )fXa |X−a ,$ (tr̄ )(Zja − Z−j
ηr̄ ){Zja δθ + Z−j
δη + δη }]
a (2)
−E[K$ (W ){τ − FXa |X−a ,$ (tr̄ )}Z−j
δη ]
40
BELLONI, CHEN, AND CHERNOZHUKOV
Applying Cauchy-Schwartz we have that
∂r E(ψuj (Wuj , θ + rδθ , η + rδη ))|r=r̄
(2)
(3)
a
a (1) 2 1/2
a (2) 2 1/2
6 f¯E[(Zja − Z−j
ηr̄ )2 ]1/2 {E[(Zja )2 ]1/2 |δθ | + E[(Z−j
δη ) ] + E[|δη |2 ]1/2 } + f¯E[(Z−j
δη ) ]
6 B̄1n (|δθ | ∨ kη − ηuj ke )
where B̄1n 6 C by the same arguments of bounded (second) moments of linear combinations.
Assumption C1(v)(c) follows similarly as
∂r2 E(ψuj (Wuj , θ + rδθ , η + rδη )) r=r̄ =
(2)
(3)
0
a
a (1)
−E[K$ (W )fX
(tr̄ )(Zja − Z−j
ηr̄ ){Zja δθ + Z−j
δη + δη }2 ]
a |X−a ,$
(2)
(1)
(3)
a
a
+2E[K$ (W )fXa |X−a ,$ (tr̄ )(Z−j
δη ){Zja δθ + Z−j
δη + δη }]
0
and under |fX
| 6 f¯0 , from Cauchy-Schwartz inequality we have
a |X−a ,$
∂r2 E(ψuj (Wuj , θ + rδθ , η + rδη )) r=r̄
(3)
(2)
a
a (1) 4 1/2
ηr̄ )2 ]1/2 {E[(Zja )4 ]|δθ |2 + E[(Z−j
δη ) ] } + CE[{δη }2 ]
6 |f¯n0 E[(Zja − Z−j
(3)
a (1) 2 1/2
a (2) 2 1/2
δη ) ] + E[{δη }2 ]1/2 }
+2f¯E[(Z−j
δη ) ] {E[(Zja )2 ]1/2 |δθ | + E[(Z−j
6 B̄2n (δθ2 ∨ kη − ηuj k2e )
where B̄2n 6 C(1 + f¯n0 ) by the same arguments of bounded (fourth) moments as before and using that
(3)
(3)
(2)
(3)
(2)
(3)
a
a
ηr̄ )2 (δη )2 ]}1/2 E[(δη )2 ]1/2 6 CE[(δη )2 ].
|E[(Zja − Z−j
ηr̄ )(δη )2 ]| 6 {E[(Zja − Z−j
To verify the near orthogonality condition, note that for all u ∈ U and j ∈ [p], since by definition
fu = fXa |X−a ,$ (Z a βu + ru ) we have
(2)
a
|Du,j,0 [η̃uj − ηuj ]| = | − E[K$ (W )fu {Z−j
(η̃ (2) − ηuj ) + ru }vuj ]| 6 δn n−1/2
a
a
by the relations E[K$ (W )(τ − FXa |X−a ,$ (Z a βu + ru ))Z−j
] = 0 and E[K$ (W )fu Z−j
vuj ] = 0 implied by the
−1/2
model, and |E[K$ (W )fu ru vuj ]| 6 δn n
by Condition CI. Thus, condition (H.79) holds.
Furthermore, since Θuj ⊂ θuj ± C/ log n, for Juj = ∂θ E[ψuj (Wuj , θuj , ηuj )] = E[K$ (W )fu Zja vuj ] =
2
2
a
E[K$ (W )vuj
] = E[vuj
|$]P($) as E[K$ (W )fu Z−j
vuj ] = 0, we have that for all θ ∈ Θuj
1
E[ψuj (Wuj , θ, ηuj )] = Juj (θ − θuj ) + ∂θ2 E[ψuj (Wuj , θ̄, ηuj )](θ − θuj )2
2
2
0
a 2
¯
where |∂θ E[ψuj (Wuj , θ̄, ηuj )]| 6 f E[|Zj | |vuj | | $]P($) 6 f¯0 E[|Zja |4 |$]1/2 E[|vuj |2 |$]1/2 P($) 6 C f¯0 P($)
so that for all θ ∈ Θuj
2
|E[ψuj (Wuj , θ, ηuj )]| > {|E[vuj
| $]| − (C 2 f¯0 )/ log n}P($)|θ − θuj |
and we can take jn > c inf $∈W P($) = cµW .
Next we verify Assumption C2 with Hujn = {η = (β, γ, 0) : kβk0 6 Cs, kγk0 6 Cs, kβ − βu,−j k 6
√
Cτn , kγ − γuj k 6 Cτn , kγ − γuj k1 6 C sτn }. We will show that ηbuj = (βeu,−j , γ
euj , 0) ∈ Hujn with probability
1 − o(1), uniformly over u ∈ U and j ∈ [p].
Under Condition CI and the choice of penalty parameters, by Theorems 1 and 2, with probability 1−o(1),
uniformly over u ∈ U we have
kβeu − βu k 6 Cτn , max sup ke
γuj − γuj k 6 Cτn , and max sup ke
γuj k0 6 C̄s,
j∈[p] u∈U
j∈[p] u∈U
QUANTILE GRAPHICAL MODELS
41
further by thresholding we can achieve supu∈U kβeu k0 6 C̄s using Lemma 17.
Next we establish the entropy bounds. For η ∈ Hujn we have that
a
a
ψuj (Wuj , θ, η) = K$ (W )(τ − 1{Xa 6 Zja θ + Z−j
β−j })fu {Zja − Z−j
γ}
It follows that F1 ⊂ WG1 G2 G3 ∪ F¯0 where F¯0 = {ψuj (Wuj , θ, ηuj ) : u ∈ U, j ∈ [p], θ ∈ Θuj }, G1 = {τ −1{Xa 6
Z a β} : kβk0 6 Cs, τ ∈ T , a ∈ V }, G2 = {Z a → Z a (1, −γ), kγk0 6 Cs, kγk 6 C, a ∈ V }, G3 = {fu : u ∈ U }.
Under Condition CI, W is a VC class of sets with VC index dW (fixed). It follows that G1 and G2 are p
choose O(s) VC-subgraph classes with VC indices at most O(s). Therefore, ent(G1 ) ∨ ent(G2 ) ∨ ent(W) 6
Cs log(an /) + CdW log(e/) by Theorem 2.6.7 in [72] and by standard arguments. Also, since fu is Lipschitz
in u by Condition CI, we have ent(G3 ) 6 (1 + dW ) log(an Lf /). Moreover, an envelope FG for F1 satisfies
E[FGq ]
a
(γ − γuj )|q ]
= E[supu∈U ,j∈[p],kγ−γuj k1 6C √sτn |vuj − fu Z−j
√
6 2q−1 E[supu∈U ,j∈[p] |vuj |q ] + 2q−1 f¯E[maxa∈V kZ a kq∞ ]{C sτn }q
√
6 2q−1 Lq + 2q−1 f¯{Mn C sτn }q 6 2q Lq
n
n
√
since Mn C sτn 6 δn Ln /f¯ and δn 6 1 for n large.
Next we bound the entropy in F¯0 . Note that for any ψuj (Wuj , θ, ηuj ) ∈ F¯0 , there is some δ ∈ [−C, C]
such that
ψuj (Wuj , θ, ηuj ) = K$ (W ){τ − 1{Xa 6 Zja δ + QXa (τ | X−a , $)}}vuj
and therefore F¯0 ⊂ W{T − φ(V)}L where φ(t) = 1{t 6 0}, V = ∪a∈V,j∈[p] Vaj with
Vaj := {Xa − Zja δ − QXa (τ | X−a , $) : τ ∈ T , $ ∈ W, |δ| 6 C},
a
and L = ∪a∈V,j∈[p] (Laj + {vūj }) where Laj = {(X, W ) 7→ vuj − vūj = fu Z−j
(γuj − γūj ) : u ∈ U }. Note that
each Vaj is a VC subgraph class of functions with index 1 + CdW as {QXa (τ |X−a , $) : (τ, $) ∈ W × T } is
a VC-subgraph with VC-dimension CdW for every a ∈ V . Since φ is monotone, φ(V) is also the union of
VC-dimension of order 1 + CdW .
Letting F1 = 1 be an envelope for W and T −φ(V). By Lemma 5, it follows that kγuj −γuj 0 k 6 Lγ {ku−u0 k+
k$ − $0 k1/2 } for some Lγ satisfying log(Lγ ) 6 C log(p|V |n) under Condition CI. Therefore, |vuj − vūj | =
√
a
|Z−j
(γuj − γūj )| 6 kZ a k∞ pkγuj − γūj k. For a choice of envelope Fa = Mn−1 kZ a k∞ + 2 supu∈U |vuj | which
satisfies kFa kP,q . Ln , we have
log N (kFa kQ,2 , Laj , k · kQ,2 ) 6 log N ( Mn k kZ a k∞ kQ,2 , Laj , k · kQ,2 )
√
6 log N (/{Mn pLγ }, U, | · |) 6 Cdu log(Mn pLγ /)
Since L = ∪a∈V,j∈[p] (Laj + {vūj }), taking FL = maxa∈V Fa , we have that
log N (kFL F1 kQ,2 , F̄0 , k · kQ,2 ) 6 log N ( 4 kF1 kQ,2 , W, k · kQ,2 ) + log N ( 4 kF1 kQ,2 , T − φ(V), k · kQ,2 )
P
+ log a∈V,j∈[p] N ( 2 kFa kQ,2 , Laj , k · kQ,2 )
6 log(p|V |) + 1 + C 0 {dW + du } log(4eMn |V |pLγ /)
where the last line follows from the previous bounds.
Next we verify the growth conditions in Assumption C2 with the proposed F1 and Kn . CLn . We
take sn(U ,p) = (1 + dW )s and an = max{n, p, |V |}. Recall that B̄1n 6 C, B̄2n 6 C, jn > cµW .
√
√ s log(p|V |n)
Thus, we have n(τn /jn )2 . n n(1∧f
6 δn under s2 log2 (p|V |n) 6 n(1 ∧ f 4U )µ6W . Moreover,
2 )µ3
U
W
42
BELLONI, CHEN, AND CHERNOZHUKOV
α/2
(τn /jn )
p
sn(U ,p) log(an ) .
r
4
(1+dW )3 s3 log3 (p|V |n)
n(1∧f 2 )µ3W
U
. δn under dW fixed and s3 log3 (p|V |n) 6 δn4 n(1 ∧
1
1
1
f 2U )µ3W and sn(U ,p) n− 2 Kn log(an ) log n . (1 + dW )sn q − 2 Mn log(p|V |n) log n 6 δn under our conditions.
Finally, the conditions of Corollary 4 hold with ρn = (1 + dW ) since the score is the product of VC-subgraph
classes of function with VC index bounded by C(1 + dW ).
0
Proof of Theorem 4. We will invoke Lemma 7 with β̄u as the estimand and riu = Xi,−a
(βu − β̄u ), therefore
0
E[K$ (W )(τ − 1{Xa 6 X−a β̄u + ru })X−a ] = 0. To invoke the lemma we verify that the events Ω1 , Ω2 , Ω3
and Ω4 hold with probability 1 − o(1)
Ω1
Ω2
Ω3
Ω4
X
:= {λu > c|Suj |/b
σa$j
, for all u ∈ U, j ∈ V },
bu (β̄u ) 6 R̄uξ : u ∈ U}
:= {
R
n
:= supu∈U ,1/√n6kδk1,$ 6√n |En [gu (δ, X, W ) − E[gu (δ, X, W ) | X−a , W ]]|/kδk1,$ 6 t3
o
X
:= {Ku σ
ba$j
> |En [huj (X−a , W )]|, for all u ∈ U, j ∈ V \{a}}
0
0
(β̄u + δ)) − ρτ (Xa − X−a
β̄u )}, huj (X−a , W ) = E[K$ (W ){τ −
where gu (δ, X, W ) = K$ (W ){ρτ (Xa − X−a
0
FXa |X−a ,W (X−a β̄u + ru )}Xj | X−a , W ].
p
By Lemma 8 with ξ = 1/n, by setting λu = λ0 = c2(1 + 1/16) 2 log(8|V |2 {ne/dW }2dW n)/n, we have
P(Ω1 ) = 1 − o(1). By Lemma 9, setting R̄uξ = Cs(1 + dW ) log(|V |n)/n we have P(Ω2 ) = 1 − o(1) for some
p
ξ = o(1). By Lemma 10 weqhave P(Ω3 ) = 1 − o(1) by setting t3 := C (1 + dW ) log (|V |nMn /ξ). Finally,
(1+dW ) log(|V |n)
n
we have P(Ω4 ) = 1 − o(1).
√
√
√
√
1
R̄uξ = o( n) for all u ∈ U.
Moreover, we have that kβ̄u k1,$ 6 skβ̄u k2,$ 6 C s = o( n) and λu (1−1/c)
Finally, we verify condition (F.64) holds for all
p
p
0
vkn,$ > C s(1 + dW ) log(n|V |)/n/κu,2c },
δ ∈ Au := ∆u,2c ∪ {v : kvk1,$ 6 2cR̄uξ /λu , k fu X−a
√
p
u
q̄Au /4 > ( f¯ + 1)kru kn,$ + [λu + t3 + Ku ] κ3cu,2cs and q̄Au > {2c 1 + t3 +K
R̄uγ }1/2 .
λu
by Lemma 11 with Ku = C
0
0
]. By Lemma 6, with probability
] and E[K$ (W )fu X−a X−a
Consider the matrices En [K$ (W )fu X−a X−a
p
√
1 − o(1), it follows that we can take η = ηn = CMn s(1 + dW ) log(|V |n) log(1 + s){log n}/ n and Dkk = 2η
√
in Lemma 16. (Note that we increase δn by a factor of log n.) Therefore, with at least the same probability
we have (taking s > 2)
0
0
δ 0 En [K$ (W )fu X−a X−a
]δ > δ 0 E[K$ (W )fu X−a X−a
]δ − 4ηkδk21 /s
(C.46)
and by definition of f u we have
0
0
En [K$ (W )fu |X−a
δ|2 ] > f u E[K$ (W )|X−a
δ|2 ] − 4ηkδk21 /s > cf u P($)kδk2 − 4ηkδk21 /s.
√
For δ ∈ ∆u,2c we have kδk1 6 Ckδk1,$ /{P($)}1/2 6 C 0 kδTu k1,$ /{P($)}1/2 6 C 0 skδTu k2 . Note that we
p
can assume kδk > c s(1 + dW ) log(n|V |)/n otherwise we are done. So that for δ ∈ Au \∆u,2c we have that
p
p
√
kδk1 /kδk2 6 Cs log(|V |n)/n/ s(1 + dW ) log(n|V |)/n 6 C 0 s.
Similarly we have
0
0
En [K$ (W )|X−a
δ|2 ] 6 E[K$ (W )|X−a
δ|2 ] + 4ηkδk21 /s 6 CP($)kδk2 − 4ηkδk21 /s.
(C.47)
QUANTILE GRAPHICAL MODELS
43
Under the condition that η = o(f U µW ), which holds by Condition P, for n sufficiently large we have with
probability 1 − o(1) that
0
0
En [K$ (W )fu |X−a
δ|2 ]3/2
En [K$ (W )fu |X−a
δ|2 ]3/2
c
c
> ¯0 inf
>
inf
0 δ|3 ]
0 δ|2 ] max
f δ∈Au En [K$ (W )|X−a
f¯0 δ∈Au En [K$ (W )|X−a
i6n kXi k∞ kδk1
0
2 3/2
0 3/2
{c f u P($)kδk }
c f u P($)1/2 kδk
c
c
> ¯0 inf
> ¯0 inf
f δ∈Au C 0 P($)kδk2 maxi6n kXi k∞ kδk1
f δ∈Au C 0 maxi6n kXi k∞ kδk1
3/2
1/2
f
µW
> C 00 U¯0 √
s maxi6n kXi k∞
f
q̄Au
(C.48)
where maxi6n kXi k∞ 6 `n Mn with probability 1 − o(1) for any `n → ∞. Therefore, under the condition
p
√
Mn s log(p|V |n) = o( nµW ) assumed in Condition P, the conditions on q̄Au are satisfied.
By Lemma 7, we have uniformly over all u = (a, τ, $) ∈ U := V × T × W
r
p
0
k fu X−a
(βbu − βu )kn,$ 6 C
r
√
(1 + dW ) log(n|V |)
(1 + dW ) log(n|V |) s
s
b
and kβu − βu k1,$ 6 C
n
κu,2c
n
κu,2c
where κu,2c is bounded away from zero with probability 1 − o(1) for n sufficiently large. Consider the
thresholded estimators βbuλ̄ for λ̄ = {(1 + dW ) log(n|V |)/n}1/2 . By Lemma 17 we have kβbuλ̄ k0 6 Cs and the
same rates of convergence as βbu . Therefore, by refitting over the support of βbuλ̄ we have by Lemma 14,
b u 6 λu kβbuλ̄ − βu k1,$ . Cs(1 +
the estimator βeu has the same rate of convergence where we used that Q
dW ) log(|V |n)/n (the other conditions of Lemma 14 hold as for the conditions in Lemma 7).
Next we will invoke Lemma 7 for the new penalty choice and penalty loadings. (We note that minor
modifications cover the new penalty loadings.)
Ω1
Ω2
2
:= {λu > c|Suj |/{En [Kw (W )ε2u X−a,j
]}1/2 , for all u ∈ U, j ∈ V },
bu (β̄u ) 6 R̄uξ : u ∈ U}
:= n
{R
Ω3
:= supu∈U ,1/√n6kδk1,$ 6√n |En [gu (δ, X, W ) − E[gu (δ, X, W )|X−a , W ]]|/{θu kδk1,$ } 6 t3
Ω4
Ω5
X
:= {Ku θu σ
ba$j
> |En [huj (X−a , W )]|, for all u ∈ U, j ∈ V \{a}}
X
:= {θu > maxj∈V σ
ba$j
/{En [K$ (W )ε2u Xj2 ]}1/2 }
o
where event Ω5 simply makes the relevant norms equivalent, k · k1,u 6 k · k1,$ 6 θu k · k1,u . Note that we can
always take θu 6 1/{τ (1 − τ )} 6 C since T is a fixed compact set.
Next we show that the bootstrap approximation of the score provides a valid choice of penalty parameter.
0 e
Let εbu := 1{Xa 6 X−a
βu } − τ . For notational convenience for u ∈ U, j ∈ V \{a} define
ψ̄iuj =
K$ (Wi )εiu Xij
K$ (Wi )b
εiu Xij
K$ (Wi )εiu Xij
, ψiuj =
, ψbiuj =
.
2
2
2
2
1/2
1/2
E[K$ (W )εu Xj ]
En [K$ (W )εu Xj ]
En [K$ (W )b
ε2u Xj2 ]1/2
We will consider the following processes:
n
S̄uj =
√1
n
Pn
i=1 ψ̄iuj , Suj =
√1
n
Pn
i=1 ψiuj , G uj =
√1
n
1 X b
b
gi ψiuj ,
i=1 gi ψ̄iuj , Guj = √
n i=1
Pn
and N is a tight zero-mean Gaussian process with covariance operator given by E[ψ̄uj ψ̄u0 j 0 ]. Their supremum
∗
b∗ :=
are denoted by Z̄S := supu∈U ,j∈V \{a} |S̄uj |, ZS := supu∈U ,j∈V \{a} |Suj |, Z̄G
:= supu∈U ,j∈V \{a} |G uj |, Z
G
supu∈U ,j∈V \{a} |Gbuj |, and ZN := supu∈U ,j∈V \{a} |Nuj |.
44
BELLONI, CHEN, AND CHERNOZHUKOV
b∗ . We have that
The penalty choice should majorate ZS and we simulate via Z
G
b∗ 6 t)| 6 |P(ZS 6 t) − P(Z̄S 6 t)| + |P(Z̄S 6 t) − P(ZN 6 t)|
|P(ZS 6 t) − P(Z
G
∗
∗
b∗ 6 t)|
+|P(ZN 6 t) − P(Z̄G
6 t)| + |P(Z̄G
6 t) − P(Z
G
We proceed to bound each term. We have that
|ZS − Z̄S | 6 Z̄S
sup
u∈U ,j∈V \{a}
6 Z̄S
sup
u∈U ,j∈V \{a}
E[K$ (W )ε2u Xj2 ]1/2
−1
En [K$ (W )ε2u Xj2 ]1/2
(En − E)[K$ (W )ε2u Xj2 ]
En [K$ (W )ε2u Xj2 ]1/2 {En [K$ (W )ε2u Xj2 ]1/2 + E[K$ (W )ε2u Xj2 ]1/2 }
0
Therefore, since {1{Xa 6 X−a
βu } : u ∈ U} is a VC-subgraph of VC dimension 1 + dW , and W is a VC class
of sets of dimension dW , we apply Lemma 19 with envelope F = kXk2∞ and σ 2 6 maxj∈V E[Xj4 ] 6 C to
obtain with probability 1 − o(1)
r
(1 + dW ) log(|V |n) Mn2 (1 + dW ) log(|V |n)
2 2
0
+
sup
|(En − E)[K$ (W )εu Xj ]| . δ1n :=
n
n
u∈U ,j∈V \{a}
where δ1n = o(µ2W ) under Condition P. Note that this implies that the denominator above is bounded away
from zero by cµW . Therefore,
0
|ZS − Z̄S | .P δ1n := Z̄S δ1n
/µW .
4
where Z̄S .P {(1 + dW ) log(n|V |)}1/2 . By Theorem 2.1 in [27], since E[ψ̄uj
] 6 C, there is a version of ZN
such that
!
1/3
Mn (1 + dW ) log(n|V |) Mn ((1 + dW ) log(n|V |))2/3
|Z̄S − ZN | .P δ2n :=
+
n1/2
n1/6
and by Theorem 2.2 in [27], there is also a version of
1/2
|ZN −
∗
Z̄G
|
.P
Mn (1 + dW ) log(n|V |) Mn ((1 + dW ) log(n|V |))3/4
+
n1/2
n1/4
!
Finally, we have that
n
X
∗
b∗ | 6 sup √1
|Z̄G
−Z
gi (ψbiuj − ψ̄iuj )
G
n i=1
u∈U ,j
Pn
where conditional on (Xi , Wi ), i = 1, . . . , n, √1n i=1 gi (ψbiuj − ψ̄iuj ) is a zero-mean Gaussian with variance
En [(ψbiuj − ψ̄iuj )2 ] 6 δ̄ 2 . Next we bound δ̄n . We have
n
δ̄n
6 En [(ψbuj − ψuj )2 ]1/2 + En [(ψuj − ψ̄uj )2 ]1/2 6 En [(ψbuj − ψuj )2 ]1/2 + δ1n /µW ,
and
En [(ψbuj − ψuj )2 ]1/2
En [(K$ (W )Xij |b
εu −εu |)2 ]1/2
2 ε
En [K$ (W )Xij
b2u ]1/2
2 2 1/2
En [K$ (W )Xij
εu ]
2 2 1/2
+
|En [K$ (W )Xij
εbu ]
cP($)
6
6
2
En [K$ (W )Xij
|b
εu
2 1/2
− εu | ]
2 2 1/2
− En [K$ (W )Xij
εu ] |
1
2 ε
En [K$ (W )Xij
b2u ]1/2
+
2 2 1/2
εu ]
En [K$ (W )Xij
cP($)
,
note that the term in the curly brackets is bounded by C/P($)1/2 with probability 1 − o(1). To bound the
0 e
0
0 e
other term note that |b
εu − εu |2 = |1{Xa 6 X−a
βu } − 1{Xa 6 X−a
βu }|. Note that εbu = 1{Xa 6 X−a
βu } − τ
QUANTILE GRAPHICAL MODELS
45
0 e
0
where kβeu k0 6 Cs. Therefore, we have {1{Xa 6 X−a
βu } : u ∈ U} ⊂ {1{Xa 6 X−a
β} : kβk0 6 Cs} which
|V |
0
is the union of Cs VC subgraph classes of functions with VC dimension C s. Moreover, we have
2
E[K$ (W )Xij
|b
εu − εu |2 ]
2
0 e
0
= E[K$ (W )Xij
|1{Xa 6 X−a
βu } − 1{Xa 6 X−a
βu }|]
0
2
e
¯
6 f E[K$ (W )Xij |X−a (βu − βu )|]
0
4 1/2
] E[K$ (W )|X−a
(βeu − βu )|2 ]1/2
6 f¯E[K$ (W )Xij
p
6 C(f¯/f 1/2
)P($)1/2 s(1 + dW ) log(n|V |)/n
U
Therefore, by Lemma 19, with probability 1 − o(1) we have
r
p
s(1 + dW ) log(n|V |)
2
2
(En − E)[K$ (W )Xij |b
εu − εu | ] .
C(f¯/f 1/2
) s(1 + dW ) log(n|V |)/n
U
n
p
Under s(1 + dW ) log(n|V |)/n = o(f U µW ) we have that with probability 1 − o(1) that
δ̄n 6 C{s(1 + dW ) log(n|V |)/n}1/4 .
Therefore, using again the sparsity of βeu in the definition of ψbiuj
n
sup
u∈U ,j∈V \{a}
1 X
√
gi (ψbiuj − ψ̄iuj )
n i=1
.P δ̄n
p
s(1 + dW ) log(|V |n)
.P δ3n := {s log(|V |n)/n}1/4
p
s(1 + dW ) log(|V |n)
The rest of the proof follows similarly to Corollary 2.2 in [14] since under Condition P (and the bounds
above) we have that rn := δ1n + δ2n + δ3n = o({E[ZN ]}−1 )) where E[ZN ] . {(1 + dW ) log(|V |n)}1/2 . Then
b∗ 6 t)| = oP (1) which in turn implies that
we have supt |P(ZS 6 t) − P(Z
G
P(Ω1 )
= P(ZS 6 b
c∗G (ξ))
∗
b∗ 6 b
> P(ZbG
6b
c∗G (ξ)) − |P(ZS 6 b
c∗G (ξ)) − P(Z
c∗G (ξ))|
G
> 1 − ξ + oP (1)
Note that the occurrence of the events Ω2 , Ω3 and Ω4 follows by similar arguments. The result follows by
Lemma 7, thresholding and applying Lemma 17 and Lemma 14 similarly to before.
Appendix D. Technical Lemmas for Conditional Independence Quantile Graphical Model
Let u = (a, τ, $) ∈ U := V × T × W, and Tu = support(βu ) where |Tu | 6 s for all u ∈ U.
Define the pseudo-norms
n
kvk2n,$
1X
:=
K$ (Wi )(vi )2 ,
n i=1
kδk2,$
1/2
p
X
Z
:=
{b
σa$j
}2 |δj |2
,
j=1
and kδk1,$ :=
p
X
Z
σ
ba$j
|δj |,
j=1
Z
where σ
ba$j
= {En [{K$ (W )Zja }2 ]}1/2 . These pseudo-norms induce the following restricted eigenvalue as
√
k fu Z a δkn,$
√ .
κu,c =
min
kδTuc k1,$ 6ckδTu k1,$ kδk1,$ / s
The restricted eigenvalue κu,c is an counterpart of the restricted eigenvalue proposed in [21] for our setting.
We note that κu,c typically will vary with the events $ ∈ W.
46
BELLONI, CHEN, AND CHERNOZHUKOV
We will consider three key events in our analysis. Let
Z
Ω1 := {λu > c|Suj |/b
σa$j
, for all u ∈ U, j ∈ [p]}
(D.49)
which occurs with probability at least 1−ξ by the choice of λu . For CIQGMs, we have Suj := En [K$ (W )(τ −
p
1{Xa 6 Z a βu +ru })Zja ], and λu = λV T W τ (1 − τ ). (In the case of PQGMs, we have Suj := En [K$ (W )(τ −
0
X
2
1{Xa 6 X−a
βu })X−a ], σ
ba$j
= {En [K$ (W )X−a,j
]}1/2 and λu = λ0 .)
To define the next event, for each u ∈ U, consider the function defined as
bu (βu ) = En [K$ (W ){ρu (Xa − Z a β) − ρu (Xa − Z a βu − ru ) − (τ − 1{Xa 6 Z a βu + ru })(Z a β − Z a βu − ru )}]
R
bu (βu ) > 0.
in the case of CIQGMs. (In the case of PQGMs, we replace Z a with X−a .) By convexity we have R
The event
bu (βu ) 6 R̄uξ : u ∈ U}
Ω2 := {R
(D.50)
where R̄uξ are chosen so that Ω2 occurs with probability at least 1 − ξ. Note that by Lemma 2,
bu (βu ) 6 R̄uξ :=
bu (βu )|X−a , W ]] 6 f¯kru k2n,$ /2 and with probability at least 1 − ξ, R
we have En [E[R
p
4 max{f¯kru k2n,$ , kru kn,$ C log(n1+dW p/ξ)/n} 6 C 0 s log(n1+dW p/ξ)/n.
Define gu (δ, X, W ) = K$ (W ){ρτ (Xa − Z a (βu + δ)) − ρτ (Xa − Z a βu )} so that event Ω3 is defined as
)
(
|En [gu (δ, X, W ) − E[gu (δ, X, W )|X−a , W ]]|
6 t3
(D.51)
Ω3 :=
sup
√
√
kδk1,$
u∈U ,1/ n6kδk1,$ 6 n
where t3 is given in Lemma 3 so that Ω3 holds with probability at least 1 − ξ.
√
1+1/c
1
kβu k1,$ + λu (1−1/c)
R̄uξ 6 n for all
Lemma 1. Suppose that Ω1 , Ω2 and Ω3 holds. Further assume 2 1−1/c
p
¯
u ∈ U, and (F.64) holds for all
δ ∈ Au := ∆u,2c ∪ {v : kvk1,$ 6 2cR̄uξ /λu }, q̄Au /4 > ( f + 1)kru kn,$ +
√
[λu + t3 ] κ3cu,2cs and q̄Au > {2c 1 + λt3u R̄uξ }1/2 . Then uniformly over all u ∈ U we have
√
k fu Z a (βbu − βu )kn,$
kβbu − βu k1,$
r
6 8c 1 + λt3u R̄uξ + (f¯1/2 + 1)kru kn,$ +
√ √
6 (1 + 2c) sk fu Z a δu kn,$ /κu,2c + λ2cu R̄uξ
√
3cλu s
κu,2c
√
s
+ t3 (1+c)
κu,2c
Proof of Lemma 1. Let u = (a, τ, $) ∈ U and δu = βbu − βu . By convexity and definition of βbu we have
bu (βbu ) − R
bu (βu ) + Su0 δu
R
= En [K$ (W )ρu (Xa − Z a βbu )] − En [K$ (W )ρu (Xa − Z a βu )]
6 λu kβu k1,$ − λu kβbu k1,$
(D.52)
Z
where Su is defined as in (D.49) so that under Ω1 we have λu > c|Suj |/b
σa$j
.
bu (βbu ) > 0, we have
Under Ω1 ∩ Ω2 , and since R
bu (βu ) −
−R
λu
c kδu k1,$
bu (βu + δu ) − R
bu (βu ) + En [K$ (W )(τ − 1{Xa 6 Z a βu + ru })Z a δu ]
6R
= En [K$ (W )ρu (Xa − Z a (δu + βu ))] − En [K$ (W )ρu (Xa − Z a βu )]
6 λu kβu k1,$ − λu kδu + βu k1,$
so that for c = (c + 1)/(c − 1)
kδTuc k1,$ 6 ckδTu k1,$ +
c
bu (βu ).
R
λu (c − 1)
(D.53)
QUANTILE GRAPHICAL MODELS
47
To establish that δu ∈ Au := ∆u,2c ∪ {v : kvk1,$ 6 2cR̄uξ /λu } we consider two cases. If kδTuc k1,$ >
2ckδTu k1,$ we have
1
c
bu (βu )
kδTuc k1,$ 6
R
2
λu (c − 1)
and consequentially
2c b
Ru (βu ).
kδu k1,$ 6 {1 + 1/(2c)}kδTuc k1,$ 6
λu
Otherwise, we have kδTuc k1,$ 6 2ckδTu k1,$ which implies
√ p
kδu k1,$ 6 (1 + 2c)kδTu k1,$ 6 (1 + 2c) sk fu Z a δu kn,$ /κu,2c
by definition of κu,2c . Thus we have δu ∈ Au under Ω1 ∩ Ω2 .
Furthermore, (D.53) also implies that
kδu + βu k1,$
bu (βu )/λu
6 kβu k1,$ + 1c kδu k1,$ + R
bu (βu )/λu .
6 (1 + 1/c)kβu k1,$ + (1/c)kδu + βu k1,$ + R
which in turn establishes
1
1
1 + 1/c
bu (βu ) 6 2 1 + 1/c kβu k1,$ +
kβu k1,$ +
R
R̄uξ
kδu k1,$ 6 2
1 − 1/c
λu (1 − 1/c)
1 − 1/c
λu (1 − 1/c)
√
where the last inequality holds under Ω2 . Thus, kδu k1,$ 6 n under our condition. In turn, δu is considered
in the supremum that defines Ω3 .
Under Ω1 ∩ Ω2 ∩ Ω3 we have
En [E[K$ (W ){ρu (Xa − Z a (βu + δu )) − ρu (Xa − Z a βu )} | X−a , W ]
6 En [K$ (W ){ρu (Xa − Z a (βu + δu )) − ρu (Xa − Z a βu )}] + t3 kδu k1,$
6 λu kδ
u k1,$ + t3 kδu k1,$√
√
6 2c 1 + λ1u t3 R̄uξ + k fu Z a δu kn,$ [λu + t3 ] κ3cu,2cs
√ √
here we used the bound kδu k1,$ 6 (1 + 2c) sk fu Z a δu kn,$ /κu,2c +
2c
λu R̄uξ
(D.54)
under Ω1 ∩ Ω2 .
Using Lemma 12, since (F.64) holds, we have for each u ∈ U
En [E[K$ (W ){ρu (Xa − Z a (βu + δu )) − ρu (Xa − Z a βu )} | X−a , W ]
p
√
Z
> −( f¯ + 1)kru kn,$ k fu Z a δu kn,$ − supu∈U ,j∈[p] |En [E[Suj |X−a , W ]/b
σa$j
]| kδu k1,$
√
a
2
√
k fu Z δu kn,$
+
∧ {q̄Au k fu Z a δu kn,$ }
4
here we have E[Siuj |Xi,−a , Wi ] = 0 since τ = P(Xa 6 Z a βu + ru |X−a , W ) by the definition of conditional
quantile.
Note that for positive numbers (t2 /4) ∧ qt 6 A + Bt implies t2 /4 6 A + Bt provided q/2 > B and 2q 2 > A.
2
2
(Indeed, otherwise (t2 /4) > qt so that
in turn implies that
h t > 4q which
2q + qt/2 6 (t /4) ∧ qt 6 A + Bt.)
√ i
p
Since q̄A /4 > ( f¯ + 1)kru kn,$ + {λu + t3 } 3c s and q̄A > {2c 1 + t3 R̄uξ }1/2 , the minimum on the
κu,2c
u
u
λu
right hand side is achieved by the quadratic part for all u ∈ U. Therefore we have uniformly over u ∈ U
√
q
√
p
k fu Z a δu k2n,$
t3
3c s
6 2c 1 +
R̄uξ + k fu Z a δu kn,$ ( f¯ + 1)kru kn,$ + {λu + t3 }
4
λu
κu,2c
which implies that
√
k fu Z a δu kn,$
6
r
8c 1 +
t3
λu
hp
√ i
R̄uξ + ( f¯ + 1)kru kn,$ + {λu + t3 } κ3cu,2cs .
48
BELLONI, CHEN, AND CHERNOZHUKOV
bu (βu )|X−a , $]] 6 f¯kru k2 /2,
Lemma 2 (CIQGM, Event Ω2 ). Under Condition CI we have En [E[R
n,$
b
Ru (βu ) > 0 and
−1
¯
b
P sup Ru (βu ) 6 C{1 + f }{n s(1 + dW ) log(p|V |n)} = 1 − o(1).
u∈U
bu (βu ) > 0 by convexity of ρτ . Let εiu = Xia −Z a βu −riu where kβu k0 6 s
Proof of Lemma 2. We have that R
i
a
and riu = QXa (τ |X−a , $) − Z βu .
R
bu (βu ) = −En [K$ (W )ru 1 1{εu 6 −tru } − 1{εu 6 0} dt] > 0.
By Knight’s identity (F.65), R
0
R1
bu (βu )|X−a , $] = En [K$ (W )ru FX |X ,$ (Z a βu + (1 − t)ru ) − FX |X ,$ (Z a βu + ru ) dt]
En [E[R
a
−a
a
−a
R01
6 En [K$ (W )ru 0 f¯tru dt] 6 f¯kru k2n,$ /2 6 C f¯s/n.
bu (βu ) 6 C f¯s/n) >
Since Condition CI assumes E[kru k2n,$ ] 6 P($)s/n, by Markov’s inequality we have P(R
R1
b
1/2. Define ziu := − 0 1{εiu 6 −triu } − 1{εiu 6 0} dt, so that Ru (βu ) = En [K$ (W )ru zu ] where |ziu | 6 1.
By Lemma 2.3.7 in [73] (note that the Lemma does not require zero mean stochastic processes), for t >
2C f¯s/n we have
1
P sup |En [K$ (W )ru zu ]| > t 6 2P sup |En [εK$ (W )ru zu ]| > t/4
2
u∈U
u∈U
where εi , i = 1, . . . , n are Rademacher random variables independent of the data.
Next consider the class of functions F = {−K$ (W )ru (1{εiu 6 −Bi riu } − 1{εiu 6 0}) : u ∈ U} where
Bi ∼ Uniform(0, 1) independent of (Xi , Wi )ni=1 . It follows that K$ (W )ru zu = E[−K$ (W )ru (1{εiu 6
−Bi riu } − 1{εiu 6 0})|Xi , Wi ] where the expectation is taken over Bi only. Thus we will bound the entropy
of F = {E[f |X, W ] : f ∈ F} via Lemma 25. Note that R := {ru = QXa (τ |X−a , $) − Z a βu : u ∈ U}
where G := {Z a βu : u ∈ U} is contained in the union of at most |V | ps VC-classes of dimension Cs and
H := {QXa (τ |X−a , $) : u ∈ U}} is the union of |V | VC-class of functions of dimension (1+dW ) by Condition
CI. Finally note that E := {εiu : u ∈ U} ⊂ {Xia : a ∈ V } − G − R.
Therefore, we have
supQ log N (kF̄ kQ,2 , F, k · kQ,2 ) 6 supQ log N ((/4)2 kF kQ,2 , F, k · kQ,2 )
6 supQ log N ( 18 (2 /16), W, k · kQ,2 )
+ supQ log N ( 81 (2 /16)kF kQ,2 , R, k · kQ,2 )
+ supQ log N ( 81 (2 /16), 1{E + {B}R 6 0} − 1{E 6 0}, k · kQ,2 )
We will apply Lemma 19 with envelope F̄ = supu∈U |K$ (W )ru |, so that E[maxi6n F̄i2 ] 6 C, and
supu∈U E[K$ (W )ru2 ] 6 Cs/n =: σ 2 by Condition CI. Thus, we have that with probability 1 − o(1)
r
r
s(1 + dW ) log(p|V |n) s
s(1 + dW ) log(p|V |n)
s(1 + dW ) log(p|V |n)
sup |En [εK$ (W )ru zu ]| .
+
.
n
n
n
n
u∈U
p
under Mn s2 /n 6 C.
Lemma 3 (CIQGM, Event Ω3 ). For u = (a, τ, $) ∈ U := V × T × W, define the function gu (δ, X, W ) =
K$ (W ){ρτ (Xa − Z a (βu + δ)) − ρτ (Xa − Z a βu )}, and the event
(
)
|En [gu (δ, X, W ) − E[gu (δ, X, W ) | X−a , W ]]|
Ω3 :=
sup
< t3 .
√
√
kδk1,$
u∈U ,1/ n6kδk1,$ 6 n
QUANTILE GRAPHICAL MODELS
49
Then, under Condition CI we have P (Ω3 ) > 1 − ξ for any t3 satisfying
q
√
1+d /κ
t3 n > 12 + 16 2 log(64|V |p2 n3+2dW log(n)Lβ W Mn /ξ)
√
Proof. We have that Ωc3 := {maxa∈V Aa > t3 n} for
Aa :=
√
sup
n
(τ,$)∈T ×W,N 6kδk1,$ 6N̄
En [gu (δ, X, W ) − E[gu (δ, X, W ) | X−a , W ]]
.
kδk1,$
√
√
Therefore, for N = 1/ n and N̄ = n we have by Lemma 13 with ρ = κ, Lη = Lβ , x̃ = Z a
√
P(Ωc3 ) = P(maxa∈V Aa > t3 n)
√
6 |V | maxa∈V P(Aa > t3 n)
√
= |V | maxa∈V EX−a ,W {P(A
a > t3 n | X−a , W )}
n
o
b | · |W|
c · |Tb | exp(−(t3 √n/4 − 3)2 /32)
6 |V | maxa∈V EX−a ,W 8p|N
1+d
/κ
√
maxi6n kZia k∞ W
2
1+dW
6 exp(−(t3 n/4 − 3) /32)|V |64pn
log(n)Lβ EX−a
N 1+dW
6ξ
(1+dW /κ)/q
by the choice of t3 and noting that Mn
1+dW /κ
> EX−a [maxi6n kZia k∞
], 1 + dW /κ 6 q and Mn > 1.
Lemma 4 (CIQGM, Uniform Control of Approximation Error in Auxiliary Equation). Under Condition
CI, with probability 1 − o(1) uniformly over u ∈ U and j ∈ [p] we have
a
En [K$ (W )fu2 {Z−j
(γuj − γ̄uj )}2 ] . f 2u P($){n−1 s log(p|V |n)}1/2 .
a
Proof. Define the class of functions G = ∪a∈V,j∈[p] Gaj with Gaj := {Z−j
(γuj − γ̄uj ) : τ ∈ T , $ ∈ W}. Under
Condition CI we have supu∈U kγ̄uj k0 6 Cs, supu∈U ,j∈[p] kγ̄uj −γuj k∨
j
γ̄uk
j
kγ̄u
−γ j k
√ u 1
s
6 {n−1 s log(p|V |n)}1/2 . Without
j
a
loss of generality we can set
) : τ ∈ T , $ ∈ W}
=
(γuj −γuT
for k ∈ support(γ̄uj ). Letting Gaj,T := {Z−j
for T ⊂ {1, . . . , p}, it follows that G ⊂ ∪a∈V,j∈[p] ∪|T |6Cs Gaj,T .
j
γuk
By Lemma 5, we have kγuj − γuj 0 k 6 Lγ (ku − u0 k + ku − u0 k1/2 ) for each a ∈ V , j ∈ [p]. (Note that although
j
γ̄uj might not be Lipschitz in u, however, for each T , γuT
satisfies the same Lipschitz relation as γuj , in fact
j
kγ̄uT
− γuj 0 T k 6 kγuj − γuj 0 k by construction.) Therefore, for each T we have
j
a
a
k{Z−j
(γ̄uT
− γuj )}2 − {Z−j
(γ̄uj 0 T − γuj 0 )}2 kQ,2
j
j
a
a
6 kZ−j
(γ̄uT
− γ̄uj 0 T + γuj 0 − γuj )Z−j
(γ̄uT
− γuj + γ̄uj 0 T − γuj 0 )kQ,2
j
j
a 2
6 kkZ−j
k∞ kQ,2 kγ̄uT
− γ̄uj 0 T + γuj 0 − γuj k1 kγ̄uT
− γuj + γ̄uj 0 T − γuj 0 k1
√
j
a 2
6 4kkZ−j
k∞ kQ,2 supu∈U kγ̄uT − γuj k1 2pkγuj − γuj 0 k
a 2
6 kkZ−j
k∞ kQ,2 L0γ (ku − u0 k + ku − u0 k1/2 ).
√
where L0γ = 4{n−1 s2 log(p|V |n)}1/2 2pLγ . Thus, for the envelope G = maxa∈V kZ a k2∞ supu∈U kγ̄uj − γuj k21
that
sup
kγ̄ j −γ j k2
log N (kGkQ,2 , G, k · kQ,2 ) 6 Cs log(|V |p) + log N u∈U L0 u u 1 , U, dU 6 Cs(1 + dW )2 log(L0γ n/).
γ
Next define the functions W0 =
: u ∈ U}, W1 = {P($)−1 : $ ∈ W} and W2 = {K$ (W ) : $ ∈
W}. We have that W2 is VC class with VC index CdW and W1 is bounded by µ−1
W and covering number
1+dW
2
2
bounded by (CdW /{µW })
. Finally, since |K$ (W )fu − K$0 (W )fu0 | 6 K$ (W )K$0 (W )|fu2 − fu20 | +
{K$ (W )fu2
50
BELLONI, CHEN, AND CHERNOZHUKOV
f¯2 |K$ (W ) − K$0 (W )| 6 2f¯Lf ku − u0 k + f¯2 |K$ (W ) − K$0 (W )|, we have N (, U, | · |) 6 (C(1 + dW )/)1+dW .
Therefore, using standard bounds we have
2
0
¯
log N (kµ−1
W Gf kQ,2 , W0 W1 W2 G, k · kQ,2 ) . s(1 + dW ) log(Lγ Lf n/)
By Lemma 19 we have that with probability 1 − o(1) that
a
supu∈U ,j∈[p] |(En − E)[fu2 {Z−j
(γuj − γ̄uj )}2 /P($)]|
q
a (γ j −γ̄ j )}4 ]/P($)2
2 −1
j
j 2
s(1+dW )2 log(p|V |n) supu∈U E[K$ (W )fu4 {Z−j
s(1+dW )2 Mn
µW supu∈U kγ̄u
−γu
k1 log(p|V |n)
u
u
+
.
n
n
q
2
2 2
2
(1+dW ) Mn s log(p|V |n) s log(p|V |n)
|n) s log(p|V |n)
+
. s(1+dW µ) Wlog(p|V
n
n
nµW
n
r
2 2
(1+dW )2 Mn
s log(p|V |n)
s log(p|V |n)
s(1+dW )2 log(p|V |n)
.
µ
f
+
W U
n
µ3 f 2 n
nµ2 f
W
.
1/2
s log(p|V |n)
µW f U {δn
n
W
U
U
+ δn2 }
here we used that E[fu4 {Z a δ}4 |$] 6 f¯4 E[{Z a δ}4 |$] 6 Ckδk4 , kγ̄uj − γuj k + s−1/2 kγ̄uj − γuj k1 6
{n−1 s log(p|V |n)}1/2 , s(1 + dW )2 log(p|V |n) 6 δn nf 2U µ3W and (1 + dW )Mn s log1/2 (p|V |n) 6 δn n1/2 µW f U
a
by Condition CI. Furthermore, by Condition CI, the result follows from E[fu2 {Z−j
(γ̄uj − γuj )}2 |$] 6
Cf 2u kγ̄uj − γuj k2 6 Cf 2u n−1 s log(p|V |n).
Lemma 5. Under Condition CI, for u = (a, τ, $) ∈ U and u0 = (a, τ 0 , $0 ) ∈ U we have that
kγuj − γuj 0 k 6
C0
¯2 E[{K$0 (W ) − K$ (W )}2 ]1/2 + E[K$ (W )K$0 (W ){f 20 − f 2 }2 ]1/2 }.
u
u
{f
f 2u0 P($0 )
In particular, we have kγuj − γuj 0 k 6 Lγ {k$ − $0 k1/2 + ku − u0 k} for Lγ = C{Lf + LK }/{f 2U µW } under
E[|K$ (W ) − K$0 (W )|] 6 LK k$ − $0 k, K$ (W )K$0 (W )|fu0 − fu | 6 Lf ku0 − uk, and fu 6 f¯ 6 C.
Proof. Let u = (a, τ, $) and u0 = (a, τ 0 , $0 ). By Condition CI we have
kγuj − γuj 0 k2
a
a
(γuj − γuj 0 )}2 ]
6 CE[{Z−j
(γuj − γuj 0 )}2 |$0 ] 6 {C/P($0 )}E[K$0 (W ){Z−j
To bound the last term of the right hand side above, by the definition of f u0 and Cauchy-Schwarz’s inequality
we have
a
a
f u0 E[K$0 (W ){Z−j
(γuj − γuj 0 )}2 ] 6 E[K$0 (W )fu0 {Z−j
(γuj − γuj 0 )}2 ]
a
a
6 {E[K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}2 ] E[K$0 (W ){Z−j
(γuj − γuj 0 )}2 ]}1/2
a
a
so that E[K$0 (W ){Z−j
(γuj − γuj 0 )}2 ]1/2 6 {E[K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}2 ]}1/2 /f u0 . Therefore
a
kγuj − γuj 0 k2 6 {1/f u0 }2 {C/P($)}E[K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}2 ].
(D.55)
We proceed to bound the last term. The optimality of γuj and γuj 0 yields
a
a
E[K$ (W )fu2 Z−j
(Zja − Z−j
γuj )] = 0
and
a
a
E[K$0 (W )fu20 Z−j
(Zja − Z−j
γuj 0 )] = 0
Therefore, we have
a
a
E[K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}Z−j
]
a
a
= −E[K$0 (W )fu20 {Zja − Z−j
γuj }Z−j
]
2
2
a
a
= −E[{K$0 (W )fu0 − K$ (W )fu }{Zja − Z−j
γuj }Z−j
]
(D.56)
QUANTILE GRAPHICAL MODELS
51
Multiplying by (γuj − γuj 0 ) both sides of (D.56), we have
a
E[K$0 (W )fu20 {Z−j
(γuj − γuj 0 )}2 ]
a
a
6 E[{K$0 (W )fu20 − K$ (W )fu2 }2 ]1/2 {E[{Zja − Z−j
γuj }2 {Z−j
(γuj − γuj 0 )}2 ]}1/2
j
6 E[{K$0 (W )fu20 − K$ (W )fu2 }2 ]1/2 Ckγuj − γu0 k
by the fourth moment assumption in Condition CI. By Condition CI, fu , fu0 6 f¯, and it follows that
|K$ (W )fu2 − K$0 (W )fu20 | 6 K$ (W )K$0 (W )|fu2 − fu20 | + f¯2 |K$ (W ) − K$0 (W )|
(D.57)
From (D.55) we obtain
kγuj − γuj 0 k 6
C0
¯2 E[{K$0 (W ) − K$ (W )}2 ]1/2 + E[K$ (W )K$0 (W ){f 20 − f 2 }2 ]1/2 }.
u
u
{f
f 2u0 P($0 )
Lemma 6. Let U = V × T × W. Under Condition CI, for m = 1, 2, we have
"
#
E
sup
|(En − E)[K$ (W )fum (Z a θ)2 ]| . Cδn
where δn = Mn
u∈U ,kθk0 6k,kθk=1
p
p
k(1 + dW )C log(p|V |n) log(1 + k) log n/n. Moreover, under Condition CI, δn = o(µW ).
Proof. By symmetrization we have
"
E
{E[K$ (W )fum (Z a θ)2 ]}1/2
sup
u∈U ,kθk0 6k,kθk=1
sup
#
"
#
|(En − E)[K$ (W )fum (Z a θ)2 ]| 6 2E
u∈U ,kθk0 6k,kθk=1
|En [εK$ (W )fum (Z a θ)2 ]|
sup
u∈U ,kθk0 6k,kθk=1
where εi are i.i.d.
Rademacher random variables.
We have that |K$ (W )fum − K$0 (W )fum0 | 6
K$ (W )K$0 (W )|fu − fu0 |(1 + 2f¯) + f¯m |K$ (W ) − K$0 (W )| for m = 1, 2 where u and u0 have the same
a ∈ V . However, conditional on {(Wi , Xi ), i = 1, . . . , n}, {K$ (Wi ) : i = 1, . . . , n, $ ∈ W} induces at most
ndW different sequences by Corollary 2.6.3 in [72]. This induces (at most) ndW partitions of W such that
K$ (W ) = K$0 (W ) for any $, $0 in the same partition given the conditioning. Thus, for such suitable $0 we
have |K$ (W )fum − K$0 (W )fum0 | 6 K$ (W )|fu − fu0 |(1 + 2f¯) for m = 1, 2. (Thus it suffices to create a net for
each partition.) We can take a cover Ub of V × T × W such that ku − u0 k 6 {Lf (1 + 2f¯)nk maxi6n kZia k2∞ }−1
so that |fu − fu0 |(Z a θ)2 6 |fu − fu0 |kZ a k2∞ kθk21 6 |fu − fu0 |kZ a k2∞ kkθk2 which implies
sup
|En [εK$ (W )fum (Z a θ)2 ]| −
u∈U ,kθk0 6k,kθk=1
Consequentially
"
E
sup
u∈U ,kθk0 6k,kθk=1
|En [εK$ (W )fum (Z a θ)2 ]| 6 n−1
sup
b,kθk0 6k,kθk=1
u∈U
#
|En [εK$ (W )fum (Z a θ)2 ]|
"
#
6E
sup
|En [εK$ (W )fum (Z a θ)2 ]|
b,kθk0 6k,kθk=1
u∈U
b 6 |V |ndW {Lf (1 + 2f¯)nk maxi6n kZ a k2∞ }(1+dW ) .
where |U|
i
By Lemma 18 with K = K(W, X) = (1 + f¯2 ) supa∈V maxi6n kZia k∞ and
p
√ q
√
√
b + 1 + log p + log k log(p ∨ n) log n /√n
δn (W, X) := C̄K(W, X) k
log |U|
p
√
√
. K(W, X) k(1 + dW )C log(p|V |nK(W, X)) log(1 + k) log n/ n
+
1
n
52
BELLONI, CHEN, AND CHERNOZHUKOV
so that conditional on (W, X) we have
"
E
#
|En [εK$ (W )fum (Z a θ)2 ]|
sup
. δn (W, X)
b,kθk0 6k,kθk=1
u∈U
p
sup
En [K$ (W )fum (Z a θ)2 ]
b,kθk0 6k,kθk=1
u∈U
Therefore,
"
E
#
|En [εK$ (W )fum (Z a θ)2 ]|
sup
"
u∈U ,kθk0 6k,kθk=1
#
6 EW,X δn (W, X)
sup
p
En [K$ (W )fum (Z a θ)2 ]
b,kθk0 6k,kθk=1
u∈U
+
1
n
1
n
b,kθk0 6k,kθk=1
u∈U
p
p
p
Note that for a random variable A > 1, we have that E[A2 log(CA)] 6 E[A2 ] log(C) + E[A2 log(A)] 6
p
E[A2 ] log(C) + E[A2+1/4 ]. Therefore, under Condition CI, since q > 2 + 1/4 in the definition of Mn , we
have
p
p
√
EW,X [δn2 (W, X)]1/2 . Mn k(1 + dW )C log(p|V |n) log(1 + k) log n/ n.
6
EW,X [δn2 (W, X)]
+ EW,X [δn2 (W, X)]1/2
sup
E[K$ (W )fum (Z a θ)2 ]1/2 +
The results follows by setting δn = EW,X [δn2 (W, X)]1/2 .
Appendix E. Results for Prediction Quantile Graphical Models
In the analysis of PQGM we also use the following event for some sequence (Ku )u∈U
X
0
βu + ru })X−a |Xj , W ]], u ∈ U, j ∈ V \{a}}.
Ω4 = {Ku σ
ba$j
> En [E[K$ (W )(τ − 1{Xa 6 X−a
(E.58)
1+1/c
Lemma 7 (Rate for PQGM). Suppose that Ω1 , Ω2 , Ω3 and Ω4 hold. Further assume 2 1−1/c
kβu k1,$ +
√
1
R̄
6
n
for
all
u
∈
U,
and
(F.64)
holds
for
all
δ
∈
A
:=
∆
∪
{v
:
kvk
6
2c
R̄uξ /λu },
u
$,2c
1,$
λu (1−1/c) uξ
√
p
u
R̄uξ }1/2 . Then uniformly over
q̄Au /4 > ( f¯ + 1)kru kn,$ + [λu + t3 + Ku ] κ3cu,2cs and q̄Au > {2c 1 + t3 +K
λu
all u = (a, τ, $) ∈ U := V × T × W, the k · k1,$ -penalized estimator βbu satisfies
√
0
k fu X−a
(βbu − βu )kn,$
kβbu − βu k1,$
r
√
8c 1 + λt3u R̄uξ + (f¯1/2 + 1)kru kn,$ + [λu + t3 + Ku ] κ3cu,2cs
√ √
0
6 (1 + 2c) sk fu X−a
δu kn,$ /κu,2c + λ2cu R̄uξ
6
Proof of Lemma 7. The proof proceeds similarly to the proof of Lemma 1 by defining
bu (β)
R
0
0
= En [K$ (W ){ρu (Xa − X−a
β) − ρu (Xa − X−a
βu − ru )}]
0
0
0
−En [K$ (W ){(τ − 1{Xa 6 X−a βu + ru })(X−a β − X−a
βu − ru )}].
The same argument yields δu = βbu − βu ∈ Au := ∆$,2c ∪ {v : kvk1,$ 6 2cR̄uξ /λu } under Ω1 ∩ Ω2 . (Similarly
√
we also have kδu k1,$ 6 n.) Furthermore, under Ω1 ∩ Ω2 ∩ Ω3 we have that (D.54) also holds which implies
0
0
En [E[K$ (W ){ρu (Xa − X−a
(βu + δu )) − ρu (Xa − X−a
βu )}|X−a , W ] 6 (λu + t3 )kδu k1,$
Since the conditions of Lemma 12 hold we have
0
0
En [E[K$ (W ){ρτ (Xa − X−a
(βu + δu )) − ρτ (Xa − X−a
βu )}|X−a , W ]
p
√
0
¯
> −( f + 1)kru kn,$ k fu X−a δkn,$ − Ku kδu k1,$
√
0
√
k fu X−a
δu k2n,$
0
+
∧ q̄Au k fu X−a
δu kn,$
4
QUANTILE GRAPHICAL MODELS
53
where Ku is given in Ω4 which accounts for the misspecification the conditional quantile condition. Therefore,
we have
√
0
√
√
δu k2n,$
k fu X−a
0
0
∧ q̄Au k fu X−a
δu kn,$ 6 (f¯1/2 + 1)kru kn,$ k fu X
δu kn,$ + (λu + t3 + Ku )kδu k1,$
4
√−a
√
3c s
1/2
0
¯
6 {(f
+ 1)kru kn,$ + κu,2c (λu + t3 + Ku )}k fu X−a
δu kn,$
+(λu + t3 + Ku ) λ2cu R̄uξ
The result then follows with the same argument under the current assumptions that account for Ku .
Lemma 8 (PQGM, Event Ω1 ). Under Condition P, we have
!
2 !
0
|En [K$ (W )(τ − 1{Xa 6 X−a
βu + ru })X−a,j ]|
ne 2dW
t/(1 + δ̄n )
> t 6 8|V |(
P
sup
)
exp −
X
dW
2(1 + 1/16)
σ
ba$j
u∈U ,j∈[d]
0
2
where t > 4 supu∈U {E[K$ (W )(τ − 1{Xa 6 X−a
βu + ru })2 X−a,j
]}1/2 and δ̄n = o(1). In particular, the RHS
p
is less than ξ if t > 2(1 + δ̄n )(1 + 1/16) log(8|V |{ne/dW }2dW /ξ).
X
2
Proof. Set σa$j
:= E[K$ (W )X−a,j
]1/2 . We have that for any δ̄n → 0
P(λ0 6
sup
0
X
|En [K$ (W )(τ − 1{Xa 6 X−a
βu })X−a,j ]|/b
σa$j
)
u∈U ,j∈[d]
6 P(λ0 6 (1 + δ̄n )
sup
0
X
|En [K$ (W )(τ − 1{Xa 6 X−a
βu })X−a,j ]|/σa$j
)
u∈U ,j∈[d]
+P(
sup
X
X
σa$j
/b
σa$j
(E.59)
> (1 + δ̄n ))
u∈U ,j∈[d]
X
To bound the last term in (E.59), note that under Condition P, cµW 6 (σa$j
)2 6 C and W is a VC class
of set with VC dimension dW . Therefore, by Lemma 19 we have that with probability 1 − o(1)
r
(1 + dW ) log(|V |Mn /σ1 ) (1 + dW )Mn2 log(|V |Mn /σ1 )
2
sup (En − E)[K$ (W )X−a,j ] .
+
(E.60)
n
n
u∈U ,j∈[d]
2
2
for σ12 = maxu∈U ,j∈[d] E[K$ (W )X−a,j
] 6 maxj∈V E[X−a,j
] 6 C and envelope F = kXk2∞ so that kF kP,2 6
2
k maxi6n Fi kP,2 6 Mn . Thus for δ̄n → 0, provided (1 + dW )Mn2 log(|V |n) = o(n1/2 ) and (1 + dW ) log(|V |n) =
o(nδ̄n2 µ2W ), so that the RHS of (E.60) is o(δ̄n µW ), we have
!
X
σa$j
1
P
(E.61)
6 X 6 (1 + δ̄n ), for all u ∈ U, j ∈ [d] = 1 − o(1).
σ
ba$j
1 + δ̄n
Now we bound the first term of the RHS of (E.59). and let σ22 = supu∈U ,j∈[d] Var(K$ (W )(τ − 1{Xa 6
0
X
X−a
βu })X−a,j /σa$j
) 6 1. By symmetrization (adapting Lemma 2.3.7 in [73] to replace the “arbitrary”
factor 2 with 1 + δ̄n ), for δ := 1/(2(1 + nσ22 /t2 )) < 1/2 we have
Pn
0
X
(∗) := P(supu∈U ,j∈[d] | i=1 K$ (Wi )(τ − 1{Xia 6 Xi,−a
βu })Xi,−aj /σa$j
| > t)
Pn
0
X
6 2P(supu∈U ,j∈[d] | i=1 εi K$ (Wi )(τ − 1{Xia 6 Xi,−a βu })Xi,−aj /σa$j
! | > tδ)
6 2P
sup
u∈U ,j∈[d]
|
Pn
i=1
0
εi K$ (Wi )(τ −1{Xia 6Xi,−a
βu })Xi,−aj |
X
σ
ba$j
> tδ/(1 + δ̄n )
+ o(1)
where εi , i = 1, . . . , n, are Rademacher random variables independent of the data, and the last inequality
follows from (E.61).
54
BELLONI, CHEN, AND CHERNOZHUKOV
Therefore, by the union bound and symmetry, and iterated expectations we have
!#
"
Pn
0
i=1 εi K$ (Wi )(τ − 1{Xia 6 Xi,−a βu })Xi,−aj
> tδ/(1 + δ̄n ) | W, X
(∗) 6 4|V | max EW,X Pε sup
X
j∈[d]
σ
ba$j
u∈U
Next we use that {$ ∈ W} is a VC class of sets with VC dimension bounded by dW and {1{Xa 6
0
X−a
βu } : (τ, $) ∈ T × W} is a VC class of sets with VC dimension bounded by 1 + dW . By Corollary 2.6.3
in [72], we have that conditionally on (Wi , Xi )ni=1 , the set of (binary) sequences {(K$ (Wi ))i=1,...,n : $ ∈ W}
PdW −1 n
0
has at most j=0
j different values. Similarly, {(1{Xia 6 Xi,−a βu })i=1,...,n : u ∈ U} assumes at most
PdW n
Pk
n
k
j=0 j different values. Assuming that n > dW , we have
j=0 j 6 {ne/k} and
P
0
βu })Xi,−aj |
| ni=1 εi K$ (Wi )(τ −1{Xia 6Xi,−a
> tδ/(1 + δ̄n ) | W, X
Pε supu∈U
X
σ
ba$j
Pn
0
βu })Xi,−aj |
| i=1 εi K$ (Wi )(τ̃ −1{Xia 6Xi,−a
ne
dW −1 ne dW
6 { dW −1 }
{ dW } supu∈U Pε supτ̃ ∈T
> tδ/(1 + δ̄n ) | W, X
X
σ
ba$j
Pn
0
| i=1 εi K$ (Wi )(τ̃ −1{Xia 6Xi,−a βu })Xi,−aj |
2dW
6 {ne/dW }
supu∈U ,τ̃ ∈[τ ,τ̄ ] Pε
> tδ/(1 + δ̄n ) | W, X
σ
bX
a$j
2dW
6 2{ne/dW }
2
exp(−{tδ/[1 + δ̄n ]} )
here we used that the expression is linear in τ and so it is maximized at the extremes. Combining the bounds
in the last two displayed equations we have
(∗) 6 8|V |{ne/dW }2dW exp(−{tδ/[1 + δ̄n ]}2 ).
p
√
Therefore, setting λ0 = ct/n where t > 4 nσ2 and t > 2(1 + δ̄n )(1 + 1/16) 2 log(8p|V |{ne/dW }2dW /ξ).
√
(Note that t > 4 nσ2 implies that δ > 1/{2(1 + 1/16)}.)
Lemma 9 (PQGM, Event Ω2 ). Under Condition P we have
−1
¯
b
P sup Ru (β̄u ) 6 C{1 + f }{n s(1 + dW ) log(|V |n)} = 1 − o(1).
u∈U
bu (β̄u ) > 0 by convexity of ρτ . Let εiu = Xia − Xi,−a β̄u − riu where
Proof of Lemma 9. We have that R
R
0
bu (β̄u ) = −En [K$ (W )ru 1 1{εu 6
kβ̄u k0 6 s and riu = X−a (βu − β̄u ). By Knight’s identity (F.65), R
0
−tru } − 1{εu 6 0} dt] > 0.
R
bu (β̄u )] = En [K$ (W )ru 1 FX |X ,$ (X 0 β̄u + (1 − t)ru ) − FX |X ,$ (X 0 β̄u + ru ) dt]
En [E[R
−a
−a
a
−a
a
−a
R01
6 En [K$ (W )ru 0 f¯tru dt] 6 f¯[kru k2n,$ ]/2 6 C f¯s/n.
bu (β̄u ) 6 C f¯s/n) > 1/2.
Thus, by Markov’s inequality we have inf u∈U P(R
R1
bu (β̄u ) = En [K$ (W )ru zu ] with |ziu | 6 1. By
Define ziu := − 0 1{εiu 6 −triu } − 1{εiu 6 0} dt, so that R
Lemma 2.3.7 in [73] (note that the Lemma does not require zero mean stochastic processes), for t > 2C f¯s/n
we have
1
P sup |En [K$ (W )ru zu ]| > t 6 2P sup |En [εK$ (W )ru zu ]| > t/4
2
u∈U
u∈U
where εi , i = 1, . . . , n, are Rademacher random variables independent of the data.
Consider the class of functions F = {−K$ (W )ru (1{εiu 6 −Bi riu } − 1{εiu 6 0}) : u ∈ U } where
Bi ∼ Uniform(0, 1) independent of (Xi , Wi )ni=1 . It follows that K$ (W )ru zu = E[−K$ (W )ru (1{εiu 6
−Bi riu } − 1{εiu 6 0})|Xi , Wi ] where the expectation is taken over Bi only. Thus we will bound the
QUANTILE GRAPHICAL MODELS
55
0
0
βu − X−a
β̄u : u ∈ U }
entropy of F = {E[f |X, W ] : f ∈ F } via Lemma 25. Note that R := {ru = X−a
p
0
where G := {X−a β̄u : u ∈ U } is contained in the union of at most |V | s VC-classes of dimension Cs and
0
H := {X−a
βu : u ∈ U}} is a VC-class of functions of dimension (1 + dW ) by Condition P. Finally note that
E := {εiu : u ∈ U} ⊂ {Xia : a ∈ V } − G − R.
Therefore, we have
supQ log N (kF̄ kQ,2 , F, k · kQ,2 ) 6 supQ log N ((/4)2 kF kQ,2 , F, k · kQ,2 )
6 supQ log N ( 18 (2 /16), W, k · kQ,2 )
+ supQ log N ( 81 (2 /16)kF kQ,2 , R, k · kQ,2 )
+ supQ log N ( 81 (2 /16), 1{E + {B}R 6 0} − 1{E 6 0}, k · kQ,2 )
r
By Lemma 19 with envelope F̄ = kXk∞ supu∈U kβu − β̄u k1 , and (σmax
)2 = supu∈U E[K$ (W )ru2 ] . s/n by
Condition P, we have that with probability 1 − o(1)
r
p
r
Mn s2 /n log(|V |n)
s(1 + dW ) log(|V |n) s
s(1 + dW ) log(|V |n)
+
.
sup |En [εK$ (W )ru zu ]| .
n
n
n
n
u∈U
p
under Mn s2 /n 6 C.
Lemma 10 (PQGM, Event Ω3 ). Under Condition P, for u = (a, τ, $) ∈ V × T × W, define gu (δ, X, W ) =
0
0
K$ (W ){ρτ (Xa − X−a
(βu + δ)) − ρτ (Xa − X−a
βu )}, and
)
(
|En [gu (δ, X, W ) − E[gu (δ, X, W )|X−a , W ]]|
< t3 .
Ω3 :=
sup
√
√
kδk1,$
u∈U ,1/ n6kδk1,$ 6 n
Then, under Condition CI we have P(Ω3 ) > 1 − ξ for any
q
√
√
t3 n > 12 + 16 2 log 64|V |2 n1+dW log(n)(Lβ Mn n)1+dW /κ /ξ
√
Proof. We have that Ωc3 := {maxa∈V Aa > t3 n} for
Aa :=
sup
√
(τ,$)∈T ×W,N 6kδk1,$ 6N̄
n
En [gu (δ, X, W ) − E[gu (δ, X, W )|X−a , W ]]
.
kδk1,$
√
√
We will apply Lemma 13 with ρ = κ, Lη = Lβ , x̃ = X−a (so we take p = |V |), N = 1/ n and N̄ = n.
Therefore, we have by Lemma 13 and the union bound
√
P(Ωc3 ) = P(maxa∈V Aa > t3 n)
√
6 |V | maxa∈V P(Aa > t3 n)
√
= |V | maxa∈V EX−a ,W n
{P(Aa > t3 n | X−a , W )}
o
b | · |W|
c · |Tb | exp(−(t3 √n/4 − 3)2 /32)
6 |V | maxa∈V EX−a ,W 8|V | |N
1+d
/κ
√
maxi6n kXi,−a k∞ W
1+d /κ
6 C|V |2 n1+dW log(n)Lβ W EX−a
exp(−(t3 n/4 − 3)2 /32)
N 1+dW /κ
6ξ
where the last step follows by the choice of t3 .
Lemma 11 (PQGM, Event Ω4 ). Under Condition P, and setting Ku = C
P(Ω4 ) = 1 − o(1).
q
(1+dW ) log(|V |n)
,
n
we have that
56
BELLONI, CHEN, AND CHERNOZHUKOV
Proof. First note that by Lemma 19 we have that with probability 1 − o(1)
r
(1 + dW ) log(|V |Mn /σ) (1 + dW )Mn2 log(|V |Mn /σ)
2
+
sup (En − E)[K$ (W )X−a,j ] .
n
n
$∈W,j∈V
X
and σa$j
> cP($). Under (1 + dW ) log(|V |Mn /σ) 6 δn2 µ2W and (1 + dW )Mn2 log(|V |Mn /σ) 6 δn nµW , we
X
2
) for all u ∈ U. Therefore, we have
]| = o(σa$j
have that |(En − E)[K$ (W )X−a,j
X
P(supu∈U |En [huj (X−a , W )]|/{Ku σ
ba$j
} > 1)
X
6 P(supu∈U |En [huj (X−a , W )]|/{Ku σa$j } > 1 + O(δn )) + o(1)
X
Applying Lemma 19 to F = {huj (X−a , W )/σa$j
: u ∈ U}. For convenience define H̄j = {huj (X−a , W ) : u ∈
2
U} and K̄j := {E[K$ (W )Xj ] : $ ∈ W}. Note that K̄j has covering numbers bounded by the covering number
of Kj := {K$ (W )Xj2 : $ ∈ W} hence supQ log N (kK̄j kQ,2 , K̄j , k · kQ,2 ) 6 log supQ̃ N ((/4)2 kF kQ̃,2 , Kj , k ·
kQ̃,2 ) by Lemma 25. Similarly, Lemma 25 also allows us to bound covering numbers of H̄j via covering
0
βu })Xj : u ∈ U}.
numbers of Hj = {K$ (W )(τ − 1{Xa 6 X−a
supQ log N (kF kQ,2 , F, k · kQ,2 ) 6 p maxj∈[p] supQ log N (kFj kQ,2 , Fj , k · kQ,2 )
1/2
1/2
6 p maxj∈[p] supQ {log N ((1/2)kH̄j kQ,2 , H̄j , k · kQ,2 ) + supQ log N ((1/2)cµW , 1/K̄j , k · kQ,2 )}
3/2
1/2
6 p maxj∈[p] supQ {log N ((1/2)kH̄j kQ,2 , H̄j , k · kQ,2 ) + supQ log N ((1/2)cµW , K̄j , k · kQ,2 )}
3/2
6 p maxj∈[p] supQ {log N ((1/2)kH̄j kQ,2 , H̄j , k · kQ,2 ) + supQ log N (C(1/2)cµW , K̄j , k · kQ,2 )}
3/2
6 p maxj∈[p] supQ {log N ((1/2)kH̄j kQ,2 , H̄j , k · kQ,2 ) + supQ log N (1/(2C)cµW , K̄j , k · kQ,2 )}
6 p maxj∈[p] supQ̃ log N ((1/4)2 kHj kQ,2 , Hj , k · kQ,2 )
+p maxj∈[p] supQ̃ log N ((1/(4C 2 )2 c2 µ3W , Kj , k · kQ̃,2 )
1/2
where Fj = ckXk∞ /µW , Hj = kXk∞ . Since Kj is the product of a VC subgraph of dimension dW with a
single function, and Hj is the product of two VC subgraph of dimension 1 + dW and a single function, by
Lemma 19 with σ 2 = 1, we have with probability 1 − o(1)
r
(En − E)[huj (X−a , W )]
(1 + dW ) log(|V |n)
Mn (1 + dW ) log(|V |n)
sup
6C
+C
.
2
1/2
1/2
n
E[K$ (W )Xj ]
u∈U
nµW
q
1/2
Thus, under Mn (1 + dW ) log(|V |n) 6 n1/2 µW we have that we can take Ku = C (1+dW ) nlog(|V |n) .
Appendix F. Technical Results for High-Dimensional Quantile Regression
In this section we provide technical results for high-dimensional quantile regression. It is based on a
sample (ỹi , x̃i , Wi )ni=1 , independent across i, ρτ (t) = (τ − 1{t 6 0})t, τ ∈ T ⊂ (0, 1) a compact interval,
and a family of indicator functions Kw (W ) = 1 if W ∈ Ω$ , Kw (W ) = 0 otherwise, here Ω$ ∈ W. For
convenience we index the sets Ω$ by $ ∈ BW ⊂ RdW where we normalize the diameter of BW to be less or
0
equal than 1/6. Let fỹ|x̃,ru ,$ (·) denote the conditional density function, fỹ|x̃,ru ,$ (·) 6 f¯, |fỹ|x̃,r
(·)| 6 f¯0
u ,$
and fu := fỹ|x̃,ru ,$ (x̃0 ηu ). Moreover, we assume that
kηu − ηũ k1 6 Lη {|τ − τ̃ | + k$ − $̃kρ }.
(F.62)
Although the results can be applied more generally, these results will be used for (ηu , ru ), u = (ỹ, τ, $) ∈
U := {ỹ} × T × W satisfying
E[K$ (W )(τ − 1{ỹ 6 x̃0 ηu + ru })x̃] = 0.
QUANTILE GRAPHICAL MODELS
57
Note that this generality is flexible enough to allow us to cover the case that the τ -conditional quantile
function Qỹ (τ |x̃, $) = x̃0 η̃u +r̃u by setting ηu = η̃u and ru = r̃u in which case E[(τ −1{ỹ 6 x̃0 ηu +ru })|x̃, $] =
0. It also covers the case that
η̃u ∈ arg min E[K$ (W )ρτ (ỹ − x̃0 β)]
β
0
so that E[K$ (W )(τ − 1{ỹ 6 x̃ η̃u })x̃] = 0 holds by the first order condition by setting ηu = η̃u and ru = 0.
Moreover, it also covers the case that we work with a sparse approximation η̄u of η̃u by setting ηu = η̄u and
ru = x̃0 (η̃u − η̄u ).
Lemma 12 (Identification Lemma). For u = (a, τ, $) ∈ U, and a subset Au ⊂ Rp let
3/2
q̄Au = 1/(2f¯0 ) · inf En K$ (W )fu |x̃0 δ|2
/En K$ (W )|x̃0 δ|3
(F.63)
and assume that for all δ ∈ Au
En K$ (W )|ru | · |x̃0 δ|2 + En K$ (W )ru2 · |x̃0 δ|2 6 1/(4f¯0 )En [K$ (W )fu |x̃0 δ|2 ].
(F.64)
δ∈Au
Then we have
En [E[K$ (W )ρτ (ỹ − x̃0 (ηu + δ)) | x̃, ru , W ]] − En [E[K$ (W )ρτ (ỹ − x̃0 ηu ) | x̃, ru , W ]]
√
√
√
k fu x̃0 δk2n,$
∧ q̄Au k fu x̃0 δkn,$ − Kn2 k fu x̃0 δkn,$ − Kn1 kδk1,$ .
>
4
where Kn2 := (f¯1/2 + 1)kru kn,$ and Kn1 := supu∈U ,j∈[p]
|En [E[K$ (W )(τ −1{ỹ6x̃0 ηu +ru })x̃j |x̃,W ]]|
.
{En [K$ (W )x̃2j ]}1/2
Proof of Lemma 12. Let Tu = support(ηu ), and Qu (η) := En E[K$ (W )ρτ (ỹ − x̃0 η) | x̃, ru , W ]. The proof
proceeds in steps.
Step 1. (Minoration) Define the maximal radius over which the criterion function can be minorated by a
quadratic function
)
(
√
√
r : Qu (ηu + δ) − Qu (ηu ) + Kn2 k fu x̃0 δkn,$ + Kn1 kδk1,$ > 41 k fu x̃0 δk2n,$ ,
.
rAu = sup
√
∀δ ∈ Au , k fu x̃0 δkn,$ 6 r
r
Step 2 below shows that rAu > q̄Au . By construction of rAu and the convexity of Qu (·), k · k1,$ and k · kn,$ ,
√
Qu (ηu + δ) − Qu(
(ηu ) + Kn2 k fu x̃0 δkn,$ + Kn1 kδk1,$ >
)
√
√
√
inf δ̃ Qu (ηu + δ̃) − Qu (ηu ) + Kn2 k fu x̃0 δ̃kn,$ + Kn1 kδ̃k1,$
k fu x̃0 δk2n,$
k fu x̃0 δkn,$
>
∧
·
√
4
rA
δ̃ ∈ Au , k fu x̃0 δ̃ku > rAu
√
n
o
√
2
k fu x̃0 δk2n,$
k fu x̃0 δkn,$ rAu
>
∧
4
rAu
4
√
√
k fu x̃0 δk2n,$
0
>
∧ q̄Au k fu x̃ δkn,$ .
4
Step 2. (rAu > q̄Au ) Let Fỹ|x̃,ru ,$ denote the conditional distribution of ỹ given x̃, ru , $. From [48], for
any two scalars w and v the Knight’s identity is
Z v
ρτ (w − v) − ρτ (w) = −v(τ − 1{w 6 0}) +
(1{w 6 z} − 1{w 6 0})dz.
(F.65)
0
Using (F.65) with w = ỹi − x̃0i ηu and v = x̃0i δ and taking expectations with respect to ỹ, we have
Qu (ηu + δ) − Qu (ηu ) = −En [E
− 1{ỹ 6 x̃0 ηu })x̃0i δ | x̃, ru , W ]]
hR[K$ (W )(τ
i
K$ (W )x̃0 δ
+En 0
Fỹ|x̃,ru ,$ (x̃0 ηu + t) − Fỹ|x̃,ru ,$ (x̃0 ηu )dt .
58
BELLONI, CHEN, AND CHERNOZHUKOV
Using the law of iterated expectations and mean value expansion, the relation
|En [E [K$ (W )(τ − 1{ỹ 6 x̃0 ηu })x̃0 δ | x̃, ru , W ]]|
= |En K$ (W ){Fỹ|x̃,ru ,$ (x̃0 ηu + ru ) − Fỹ|x̃,ru ,$ (x̃0 ηu )}x̃0 δ
+En K$ (W ){τ − Fỹ|x̃,ru ,$ (x̃0 ηu + ru )}x̃0 δ | x̃, ru , W |
6 En [K$ (W )fu |ru | |x̃0 δ|] + f¯0 En [K$ (W )|ru |2 |x̃0 δ|] + Kn1 kδk1,$
√
√
6 k fu ru kn,$ k fu x̃0 δkn,$ + f¯0 kru kn,$ kru x̃0 δkn,$ + Kn1 kδk1,$
√
6 (f¯1/2 + 1)kru kn,$ k fu x̃0 δkn,$ + Kn1 kδk1,$
where we used our assumption on the approximation error and we have Kn2 = (f¯1/2 + 1)kru kn,$ . With that
and similar arguments we obtain for t̃x̃i ,t ∈ [0, t]
√
Qu (ηu + δ) − Qu (ηu ) + Kn2 k fu x̃0 δkn,$ + Kn1 kδk1,$ >
Qu (ηuh + δ) − Qu (ηu ) + En [E [K$ (W )(τ − 1{ỹ 6 x̃0 ηu })x̃i0 δ | x̃, ru , W ]] =
R K (W )x̃0 δ
= En 0 $
Fỹ|x̃,ru ,$ (x̃0 ηu + t) − Fỹ|x̃,ru ,$ (x̃0 ηu )dt
i
hR
0
2
K (W )x̃ δ
0
0
(x̃
η
+
t̃
)dt
= En 0 $
tfỹ|x̃,ru ,$ (x̃0 ηu ) + t2 fỹ|x̃,r
u
x̃,t
u ,$
hR
i
√
K (W )x̃0i δ
> 12 k fu x̃0 δk2n,$ − 16 f¯0 En [K$ (W )|x̃0 δ|3 ] − En 0 $
t[fỹ|x̃,ru ,$ (x̃0 ηu ) − fỹ|x̃,ru ,$ (x̃0 ηu + ru )]dt
√
√
> 1 k fu x̃0 δk2 + 1 k fu x̃0 δk2 − 1 f¯0 En [K$ (W )|x̃0 δ|3 ] − (f¯0 /2)En K$ (W )|r̃u | · |x̃0 δ|2 .
4
n,$
n,$
4
6
(F.66)
Moreover, by assumption we have
En K$ (W )|ru | · |x̃0 δ|2 6
1
E [K$ (W )fu |x̃0 δ|2 ]
4f¯0 n
(F.67)
√
Note that for any δ such that k fu x̃0 δkn,$ 6 q̄Au we have
p
3/2
k fu x̃0 δkn,$ 6 q̄Au 6 1/(2f¯0 ) · En K$ (W )fu |x̃0 δ|2
/En K$ (W )|x̃0 δ|3 .
It follows that (1/6)f¯0 En [K$ (W )|x̃0 δ|3 ] 6 (1/8)En [K$ (W )fu |x̃0 δ|2 ]. Combining this with (F.67) we have
f¯0
f¯0
1
En [K$ (W )fu |x̃0 δ|2 ] − En [K$ (W )|x̃0 δ|3 ] − En K$ (W )|ru | · |x̃0 δ|2 > 0.
4
6
2
Combining (F.66) and (F.68) we have rAu > q̄Au .
(F.68)
Lemma 13. Let W be a VC-class of sets with VC-index dW . Conditional on {(Wi , x̃i ), i = 1, . . . , n} we
have
Pỹ
sup
τ ∈ T , $ ∈ W,
N 6 kδk1,$ 6 N̄
Gn
K$ (W )
ρτ (ỹ − x̃0 (ηu + δ)) − ρτ (ỹ − x̃0 ηu )
kδk1,$
!
n
2
> M | (Wi , x̃i )i=1 6 Sn exp(−(M/4 − 3) /32)
b | · |W|
c · |Tb |, with
where Sn 6 8p|N
dW /ρ
√
√
√
c 6 ndW + 2 n maxi6n kx̃i k∞ Lη
b | 6 1 + 3 n log(N̄ /N ) , |Tb | 6 2 n maxi6n kx̃i k∞ Lη , |W|
|N
.
N
N
Proof of Lemma 13. Let giτ $ (b) = K$ (Wi ){ρτ (ỹi − x̃0i ητ $ + b) − ρτ (ỹi − x̃0i ητ $ )} 6 K$ (Wi )|b| since
K$ (Wi ) ∈ {0, 1}. Note that |giτ $ (b) − giτ $ (a)| 6 K$ (Wi )|b − a|. To easy the notation we omit the
conditioning on (x̃i , Wi ) from the probabilities.
QUANTILE GRAPHICAL MODELS
59
For any δ ∈ Rp , since ρτ is 1-Lipschitz, we have
0
2
0 2
0
τ $ (x̃ δ)} ]
$ (W )x̃ δ| ]
τ $ (x̃ δ)
6 En [{g
6 En [|K
=1
Var Gn gkδk
kx̃0 δk2n,$
kx̃0 δk2n,$
1,$
P
P
since by definition kδk1,$ = j kδj k1,$ = j kx̃0j δj kn,$ > kx̃0 δkn,$ .
Since we are conditioning on (Wi , x̃i )ni=1 the process is independent across i. Then, by Lemma 2.3.7 in
[72] (Symmetrization for Probabilities) we have for any M > 1
!
gτ $ (x̃0 δ)
P
sup
Gn
>M
kδk1,$
τ ∈T ,$∈W,N 6kδk1,$ 6N̄
!
gτ $ (x̃0 δ)
2
o
6
> M/4
P
sup
Gn
1 − M −2
kδk1,$
τ ∈T ,$∈W,N 6kδk1,$ 6N̄
where Gon is the symmetrized process.
Consider Ft,τ,$ = {δ : kδk1,$ = t}. We will consider the families of Ft,τ,$ for t ∈ [N , N̄ ], τ ∈ T and
$ ∈ W.
c×N
b of T × W × [N , N̄ ] such that
We will construct a finite net Tb × W
gτ $ (x̃0 δ)
gτ $ (x̃0 δ)
o
o
Gn
=: 3 + Ao .
63+
sup
sup Gn
sup
kδk1,$
t
δ∈F
b
c
b
τ ∈T ,$∈W,t∈[N ,N̄ ],δ∈Ft,τ,$
t,τ,$
τ ∈T ,$∈W,t∈N
By triangle inequality we have
0
δ)
gτ̃ $̃ (x̃0 δ̃)
Gon gτ $ (x̃
−
t
t̃
gτ $ (x̃0 δ)
t
0
δ)
Gon gτ̃ $̃ (x̃
t
6 Gon
+
The first term in (F.69) is such that
0
0
δ)
δ)
Gon gτ $ (x̃
− gτ̃ $ (x̃
t
t
6
6
6
−
−
gτ̃ $ (x̃0 δ)
t
gτ̃ $̃ (x̃0 δ̃)
t̃
+ Gon
gτ̃ $ (x̃0 δ)
t
−
gτ̃ $̃ (x̃0 δ)
t
√
2 n
0
t En [K$ (W )|x̃ (ητ $ − ητ̃ $ )|]
√
2 n
N maxi6n kx̃i k∞ En [K$ (W )]kητ $ − ητ̃ $ k1
√
2 n
0
N maxi6n kx̃i k∞ En [K$ (W )]Lη |τ − τ |.
(F.69)
(F.70)
Define a net Tb = {τ1 , . . . , τT } such that
|τk+1 − τk | 6
√ maxi6n kx̃i k∞
Lη
2 n
N
−1
.
To bound the second term in (F.69), note that W is a VC-class. Therefore, by Corollary 2.6.3 in [72]
we have that conditional on (Wi )ni=1 , there are at most ndW different sets $ ∈ W that induce a different
c with at most ndW
sequence {K$ (W1 ), . . . , K$ (Wn )}. Thus we can choose a (data-dependent) cover W
values of $. Further, similarly to (F.71) we have kητ̃ $ − ητ̃ $̃ k1 6 Lη k$ − $̃kρ and
√
√
0
0
δ)
δ)
Gon gτ̃ $ (x̃
− gτ̃ $̃ (x̃
6 2 t n En [K$ (W )|x̃0 (ητ̃ $ − ητ̃ $̃ )|] + 2 t n En [|K$ (W ) − K$̃ (W )||x̃0 δ|]
t
t
6
6
√
2 n
N
√
2 n
N
maxi6n kx̃i k∞ En [K$ (W )]kητ̃ $ − ητ̃ $̃ k1
maxi6n kx̃i k∞ En [K$ (W )]Lη k$ − $̃kρ .
(F.71)
c such that |W|
c 6n
We define a net W
dW
n √
odW /ρ
maxi6n kx̃i k∞
+ 2 n
Lη
N
60
BELLONI, CHEN, AND CHERNOZHUKOV
To bound the third term in (F.69), note that for any δ ∈ Ft,τ,$ , t 6 t̃, by considering δ̃ := δ(t̃/t) ∈ Ft̃,τ,$
we have
0
0
0
0
0
0
δ)
δ)
Gon gτ $ (x̃
− gτ $ (x̃ t̃δ(t̃/t))
6 Gon gτ $ (x̃
− gτ $ (x̃ tδ(t̃/t)) + Gon gτ $ (x̃ tδ(t̃/t)) − gτ $ (x̃ t̃δ(t̃/t))
t
t
gτ $ (x̃0 δ[t̃/t]) + Gon gτ $ (x̃0 δ(t̃/t)) · 1t − 1t̃
= 1t Gon gτ $ (x̃0 δ) −
√
√
)x̃0 δ| |t−t̃|
nEn (|K$ (W )x̃0 δ|) tt̃ 1t − 1t̃
6 nEn |K$ (W
t
t +
0
√
√
)x̃ δ|
t−t̃
6 2 n t−t t̃ .
= 2 nEn |K$ (W
t
t
b be a ε-net {N =: t1 , t2 , . . . , tK := N̄ } of [N , N̄ ] such that |tk − tk+1 |/tk 6 1/(2√n). Note that we
We let N
b | 6 1 + 3√n log(N̄ /N ) .
can achieve that with |N
By Markov bound, we have
P(Ao > K) 6 minψ>0 exp(−ψK)E[exp(ψAo )]
c · |N
b | minψ>0 exp(−ψK) exp 8ψ 2
6 8p|Tb | · |W|
c · |N
b | exp(−K 2 /32)
6 8p|Tb | · |W|
here we set ψ = K/16 and bound E[exp(ψAo )] as follows
"
o
E [exp (ψA )]
6(1)
6(2)
6(3)
=(4)
6(5)
6(6)
!#
gτ $ (x̃0 δ)
sup
E exp ψ sup
t
b ×W×
c N
c
kδk1,$ =t
(τ,$,t)∈T
"
!#
K$ (W )x̃0 δ
o
b
c
b
2|T | · |W| · |N |
sup
E exp 2ψ sup Gn
t
b ×W×
c N
c
kδk1,$ =t
(τ,$,t)∈T
#!#
"
"
|Gon (K$ (W )x̃j )|
kδk1,$
c · |N
b|
max
2|Tb | · |W|
sup
E exp 2ψ
sup
1/2
j6p {En [K$ (W )x̃2
t
b ×W×
c N
c
j ]}
(τ,$,t)∈T
kδk1,$ =t o
|Gn (K$ (W )x̃j )|
c · |N
b|
2|Tb | · |W|
sup
E exp 2ψ max
2 1/2
j6p
{E
n [K$ (W )x̃j ]}
b ×W×
c N
(τ,$,t)∈T
c
o
c · |N
b | max sup E exp 4ψ Gn (K$ (W )x̃j )
4p|Tb | · |W|
j6p
{En [K$ (W )x̃2j ]}1/2
c
$∈W
2
b
c
b
8p|T | · |W| · |N | exp 8ψ
Gon
c · |N
b|
2|Tb | · |W|
here (1) follows by exp(maxi∈I |zi |) 6 2|I| maxi∈I exp(zi ), (2) by contraction principle (apply Theorem 4.12
[54] with ti = K$ (Wi )x̃0i δ, and φi (ti ) = ρτ (K$ (Wi )ỹi − K$ (Wi )x̃0i ητ + ti ) − ρτ (K$ (Wi )ỹi − K$ (Wi )x̃0i ητ )
so that |φi (s) − φi (t)| 6 |s − t| and φi (0) = 0, (3) follows by
|Gon (K$ (W )x̃0 δ)| 6 kδk1,$ max |Gon (K$ (W )x̃j )/{En [K$ (W )x̃2j ]}1/2 |,
j6p
(4) by the definition of suprema, (5) we again use exp(maxi∈I |zi |) 6 2|I| maxi∈I exp(zi ), and (6) exp(z) +
exp(−z) 6 2 exp(z 2 /2).
Lemma 14 (Estimation Error of Refitted Quantile Regression). Consider an arbitrary vector ηbu and suppose
b u for
kηu k0 6 s. Let kriu 6 r̄u kn,$ , |support(b
ηu )| 6 sbu and En [K$ (W ){ρτ (ỹi − x̃0i ηbu ) − ρτ (ỹi − x̃0i ηu )}] 6 Q
all u ∈ U hold. Furthermore, suppose that
sup
u=(τ,$)∈U
ρτ (ỹ − x̃0 ηeu ) − ρτ (ỹ − x̃0 ηu )
ρτ (ỹ − x̃0 ηeu ) − ρτ (ỹ − x̃0 ηu )
t3
En K$ (W )
− E K$ (W )
| W, x̃
6 √ .
ke
ηu − ηu k1,$
ke
ηu − ηu k1,$
n
Under these events, we have for n large enough,
s
p
√
(b
su + s)
0
e
b 1/2
k fu x̃i (e
ηu − ηu )kn,$ . Nu :=
(Kn1 + t3 / n) + Kn2 + f¯r̄u + Q
u
φmin (u, sbu + s)
QUANTILE GRAPHICAL MODELS
61
√
where φmin (u, k) = inf kδk0 =k k fu x̃0 δk2n,$ /kδk2 , provided that
f¯0 En [K$ (W )(|ru | + |ru |2 )|x̃0 δ̄|2 ]
eu /q̄A → 0.
+N
u
En [K$ (W )fu |x̃0 δ̄|2 ]
u∈U ,kδ̄k0 6b
su +s
sup
(F.72)
where Au = {δ ∈ Rp : kδk0 6 sbu + s}.
Proof of Lemma 14. Let δbu = ηbu − ηu which satisfies kδbu k0 6 sbu + s. By optimality of ηeu in the refitted
quantile regression we have
En [K$ (W )ρτ (ỹi − x̃0i ηeu )] − En [K$ (W )ρτ (ỹi − x̃0i ηu )]
bu
6 En [K$ (W )ρτ (ỹi − x̃0i ηbu )] − En [K$ (W )ρτ (ỹi − x̃0i ηu )] 6 Q
(F.73)
where the second inequality holds by assumption.
Moreover, by assumption, uniformly over u ∈ U, we have conditional on (Wi , x̃i , riu )ni=1 that
!
ρτ (ỹ − x̃0 (ηu + δeu )) − ρτ (ỹ − x̃0 ηu )
Gn K$ (W )
6 t3 .
kδeu k1,$
(F.74)
Thus combining relations (F.73) and (F.74), we have
√
bu .
En [E[K$ (W ){ρu (ỹ − x̃0 (ηu + δeu )) − ρu (ỹ − x̃0 ηu )}|x̃, r̃, $]] 6 kδeu k1,$ t3 / n + Q
Invoking the sparse identifiability relation of Lemma 12, since the required condition on the approximation
errors ru ’s holds by assumption (F.72), for n large enough
√
o
p
p
k fu x̃0 δeu k2n,$ n
√
bu ,
∧ q̄Au k fu x̃0 δeu kn,$ 6 Kn2 k fu x̃0 δeu kn,$ + kδeu k1,$ (Kn1 + t3 / n) + Q
4
where q̄Au is defined with Au := {δ : kδk0 6 sbu + s}. Moreover, by the sparsity of δ̃u we have kδeu k1,$ 6
p
√
√
(b
su + s)/φmin (u, sbu + s)k fu x̃0 δeu kn,$ so that we have for t = k fu x̃0 δeu kn,$ ,
t2
∧ {q̄Au t}
4
6 t(Kn2 +
p
√
bu .
(b
su + s)/φmin (u, sbu + s){Kn1 + t3 / n}) + Q
Note that for positive numbers (t2 /4) ∧ (q̄Au t) 6 A + Bt implies t2 /4 6 A + Bt provided q̄Au /2 > B and
2
2
2q̄A
> A. (Indeed, otherwise (t2 /4) > qt so that t > 4q, which in turn implies that 2q̄A
+ q̄Au t/2 6
u
u
2
2
(t /4) ∧ q̄Au t 6 A + Bt.) Note that q̄Au /2 > B and 2q̄Au > A is implied by condition (F.72) when we set
p
√
b u and B = (Kn2 + (b
su + s)/φmin (u, sbu + su ){K1n + t3 / n}). Thus the minimum is achieved in the
A=Q
quadratic part. Therefore, for n sufficiently large, we have
p
√ p
b 1/2
k fu x̃0 δeu kn,$ 6 Q
su + s)/φmin (u, sbu + su ).
u + Kn2 + (Kn1 + t3 / n) (b
Under the condition maxi6n kx̃i k2∞ log(n∨p) = o(n minτ ∈T τ (1−τ )), the next result provides new bounds
for the data driven penalty choice parameter when the quantile indices in T can approach the extremes.
Lemma 15 (Pivotal Penalty Parameter Bound). Let τ = minτ ∈T τ (1 − τ ) and Kn = maxi6n,j∈[p] |x̃ij /b
σj |,
σ
bj = En [x̃2j ]1/2 . Under Kn2 log(p/τ ) = o(nτ ), for n large enough we have that for some constant C̄
r
log(16p/(τ ξ))
Λ(1 − ξ|x̃1 , . . . , x̃n ) 6 C̄
n
62
BELLONI, CHEN, AND CHERNOZHUKOV
where Λ(1 − ξ|x̃1 , . . . , x̃n ) is the 1 − ξ quantile of maxj∈[p] supτ ∈T
Pn
i=1
x̃ij (τ −1{Ui 6τ })
σ
bj
√
τ (1−τ )
conditional on
x̃1 , . . . , x̃n , and Ui are independent uniform(0, 1) random variables.
Proof. Conditional on x̃1 , . . . , x̃n , letting σ
bj2 = En [x2j ], we have that
Pn
i=1 x̃j (τ − 1{U 6 τ })
p
nΛ = max sup
.
j∈[p] τ ∈T
σ
bj τ (1 − τ )
Step 1. (Entropy Calculation) Let F = {x̃ij (τ − 1{Ui 6 τ })/b
σj : τ ∈ T , j ∈ [p]}, hτ =
G = {fτ /hτ : τ ∈ T }. We have that
p
τ (1 − τ ), and
d(fτ /hτ , fτ̄ /hτ̄ ) 6 d(fτ , fτ̄ )/hτ + d(fτ̄ /hτ , fτ̄ /hτ̄ )
6 d(fτ , fτ̄ )/hτ + d(0, fτ̄ /hτ̄ )|hτ − hτ̄ |/hτ
Therefore, since kF kQ 6 kGkQ by hτ 6 1, and d(0, fτ̄ /hτ̄ ) 6 1/hτ̄ we have
N (kGkQ , G, Q) 6 N (kF kQ /{2 min hτ }, F, Q)N (/{2 min h2τ }, T , | · |).
τ ∈T
τ ∈T
Thus we have for some constants K and v that
N (kGkQ , G, Q) 6 p(K/{ min h2τ })v .
τ ∈T
Step 2.(Symmetrization) Since we have E[g 2 ] = 1 for all g ∈ G, by Lemma 2.3.7 in [72] we have
√
P(Λ > t n) 6 4P(max sup |Gon (g)| > t/4)
j6p τ ∈T
Gon
: G → R is the symmetrized process generated by Rademacher variables. Conditional on
here
(x1 , u1 ), . . . , (xn , un ), we have that {Gon (g) : g ∈ G} is sub-Gaussian with respect to the L2 (Pn )-norm
by the Hoeffding inequality. Thus, by Lemma 16 in [11], for δn2 = supg∈G En [g 2 ] and δ̄n = δn /kGkPn , we
have
Z δ̄n /2
p
2
P(sup |Gon (g)| > CKδn log(pK/τ ) | {x̃i , Ui }ni=1 ) 6
−1 {p(K/{ min h2τ })v }−C +1 d
g∈G
τ ∈T
0
for some universal constant K.
δn2
In order to control δn , note that δn2 = supg∈G
maxi6n G2i , we have
P(sup |Gon (g 2 )| > C K̄δn max Gi
i6n
g∈G
Thus with probability 1 −
R 1/2
0
√1 Gn (g 2 )
n
+ E[g 2 ]. In turn, since supg∈G En [g 4 ] 6
Z
p
log(pK/τ ) | {x̃i , Ui }ni=1 ) 6
δ̄n /2
−1 {p(K/{τ })v }−C
2
+1
d.
0
√
d, since E[g 2 ] = 1 and maxi6n Gi 6 Kn / τ , we have
p
C 0 Kn log(pK/τ )
√ √
δn 6 1 +
.
n τ
−1 {p(K/τ )v }−C
2
+1
p
√ √
Therefore, under Kn log(pK/τ ) = o( n τ ), conditionally on {x̃i }ni=1 and n sufficiently large, with
R 1/2 −1
2
probability 1 − 2 0 {p(K/{τ })v }−C +1 d we have that
p
sup |Gon (g)| 6 2CK log(pK/τ )
g∈G
QUANTILE GRAPHICAL MODELS
The stated bound follows since for C > 2
Z 1/2
Z
2
2
2
−1 {p(K/{τ })v }−C +1 d 6 {p/τ }−C +1 2
0
1/2
63
2
−2+C d 6 {p/τ }−C
2
+1
.
0
Appendix G. Inequalities
b and Σ be p × p matrices with non-negative diagonal entries,
Lemma 16 (Transfer principle, [65]). Let Σ
and assume that for some η ∈ (0, 1) and s 6 p we have
b > (1 − η)v 0 Σv
∀v ∈ Rp , kvk0 6 s, v 0 Σv
b kk − (1 − η)Σkk . Then for all δ ∈ Rp we have
Let D be a diagonal matrix such that Dkk > Σ
b > (1 − η)δ 0 Σδ − kD1/2 δk21 /(s − 1).
δ 0 Σδ
λ
Z
= βbuj 1{b
σa$j
|βbuj | > λ}
Lemma 17. Consider βbu and βu with kβu k0 6 s. Denote by βbuλ the vector with βbuj
Z
where σ
ba$j
= {En [K$ (W )(Zja )2 ]}1/2 . We have that
kβbuλ − βu k1,$ 6 kβbu − βu k1,$ + λs
|support(βbuλ )| 6 s + kβbu − βu k1,$ /λ
q
√
√
kZ a (βbuλ − βu )kn,$ 6 kZ a (βbu − βu )kn,$ + φ̃max (s, $){2 sλ + kβbu − βu k1,$ / s}
a
a
here φ̃max (m, $) = sup16kθk0 6m kZ̃ a θkn,$ /kθk and Z̃ij
= Zij
/{En [K$ (W )(Zja )2 ]}1/2 .
Proof. Let Tu = support(βu ). The first relation follows from the triangle inequality
kβbuλ − βu k1,$
= k(βbuλ − βu )Tu k1,$ + k(βbuλ )Tuc k1,$
6 k(βbuλ − βbu )Tu k1,$ + k(βbu − βu )Tu k1,$ + k(βbuλ )Tuc k1,$
6 λs + k(βbu − βu )Tu k1,$ + k(βbu )Tuc k1,$
= λs + kβbu − βu k1,$
To show the second result note that kβbu − βu k1,$ > {|support(βbuλ )| − s}λ. Therefore,
|support(βbuλ )| 6 s + kβbu − βu k1,$ /λ
which yields the result.
To show the third bound, we start using the triangle inequality
kZ a (βbuλ − βu )kn,$ 6 kZ a (βbuλ − βbu )kn,$ + kZ a (βbu − βu )kn,$ .
Without loss of generality, assume that the components are ordered so that |(βbuλ − βbu )j |b
σuj is decreasing.
Let T1 be the set of s indices corresponding to the largest values of |(βbuλ − βbu )j |b
σuj . Similarly define Tk
as the set of s indices corresponding to the largest values of |(βbuλ − βbu )j |b
σuj outside ∪k−1
m=1 Tm . Therefore,
64
BELLONI, CHEN, AND CHERNOZHUKOV
Pdp/se
βbuλ − βbu = k=1 (βbuλ − βbu )Tk . Moreover, given the monotonicity of the components, k(βbuλ − βbu )Tk k2,$ 6
√
k(βbuλ − βbu )Tk−1 k1 / s. Then, we have
Pdp/se
kZ a (βbuλ − βbu )kn,$ = kZ a k=1 (βbuλ − βbu )Tk kn,$
P
6 kZ a (βbuλ − βbu )T1 kn,$ + k>2 kZ a (βbuλ − βbu )Tk kn,$
q
q
P
6 φ̃max (s, $)k(βbuλ − βbu )T1 k2,$ + φ̃max (s, $) k>2 k(βbuλ − βbu )Tk k2,$
q
q
P
√
√
6 φ̃max (s, $)λ s + φ̃max (s, $) k>1 k(βbuλ − βbu )Tk k1,$ / s
q
q
√
√
= φ̃max (s, $)λ s + φ̃max (s, $)kβbuλ − βbu k1,$ / s
q
√
√
6 φ̃max (s, $){2λ s + kβbu − βu k1,$ / s}
here the last inequality follows from the first result and the triangle inequality.
Lemma 18 (Supremum of Sparse Vectors on Symmetrized Random Matrices). Let Ub denote a finite set
and (Xiu )u∈Ub , i = 1, . . . , n, be fixed vectors such that Xiu ∈ Rp and max16i6n maxu∈Ub kXiu k∞ 6 K.
Furthermore define
p
p
p
√ q
√
b
log |U| + 1 + log p + log k log(p ∨ n) log n / n,
δn := C̄K k
where C̄ is a universal constant. Then,
"
E
#
0
2
max En [ε(θ Xu ) ]
sup
b
kθk0 6k,kθk=1 u∈U
6 δn
p
sup
En [(θ0 Xu )2 ].
b
kθk0 6k,kθk=1,u∈U
Proof. See [14] for the proof.
Corollary 2 (Supremum of Sparse Vectors on Many Random Matrices). Let Ub denote a finite set and
(Xiu )u∈Ub , i = 1, . . . , n, be independent (across i) random vectors such that Xiu ∈ Rp and
r
E[ max max kXiu k2∞ ] 6 K.
16i6n u∈U
b
Furthermore define
p
p
p
√ q
√
b + 1 + log p + log k log(p ∨ n) log n / n,
δn := C̄K k
log |U|
here C̄ is a universal constant. Then,
"
#
0
2
0
2
E
sup
max En (θ Xu ) − E[(θ Xu ) ]
6 δn2 + δn
b
kθk0 6k,kθk=1 u∈U
sup
p
En [E[(θ0 Xu )2 ]].
b
kθk0 6k,kθk=1,u∈U
We will also use the following result of [25].
Lemma 19 (Maximal Inequality). Work with the setup above. Suppose that F > supf ∈F |f | is a measurable
envelope for F with kF kP,q < ∞ for some q > 2. Let M = maxi6n F (Wi ) and σ 2 > 0 be any positive
constant such that supf ∈F kf k2P,2 6 σ 2 6 kF k2P,2 . Suppose that there exist constants a > e and v > 1 such
that
log sup N (kF kQ,2 , F, k · kQ,2 ) 6 v log(a/), 0 < 6 1.
Q
QUANTILE GRAPHICAL MODELS
65
Then
s
EP [sup |Gn (f )|] 6 K
vσ 2
log
f ∈F
akF kP,2
σ
vkM kP,2
√
+
log
n
akF kP,2
σ
!
,
here K is an absolute constant. Moreover, for every t > 1, with probability > 1 − t−q/2 ,
h
i
√
sup |Gn (f )| 6 (1 + α)EP [sup |Gn (f )|] + K(q) (σ + n−1/2 kM kP,q ) t + α−1 n−1/2 kM kP,2 t ,
f ∈F
f ∈F
∀α > 0 where K(q) > 0 is a constant depends only on q. In particular, setting a > n and t = log n, with
probability > 1 − c(log n)−1 ,
s
!
akF kP,2
vkM kP,q
akF kP,2
√
sup |Gn (f )| 6 K(q, c) σ v log
+
log
,
(G.75)
σ
σ
n
f ∈F
here kM kP,q 6 n1/q kF kP,q and K(q, c) > 0 is a constant depending only on q and c.
Appendix H. Confidence Regions for Function-Valued Parameters Based on Moment
Conditions
For completeness, in this section we collect an adaptation of the results of [14] that are invoked in our
proofs. The main difference is the weakening of the identification condition (which is allowed to decrease to
zero, see the parameter jn in Condition C1 below). We are interested in function-valued target parameters
indexed by u ∈ U ⊂ Rdu . The true value of the target parameter is denoted by
θ0 = (θuj )u∈U ,j∈[p̃] , where θuj ∈ Θuj for each u ∈ U and j ∈ [p̃].
For each u ∈ U and j ∈ [p̃], the parameter θuj is characterized as the solution to the following moment
condition:
E[ψuj (Wuj , θuj , ηuj )] = 0,
(H.76)
where Wuj is a random vector that takes values in a Borel set Wuj ⊂ Rdw , η 0 = (ηuj )u∈U ,j∈[p̃] is a nuisance
parameter where ηuj ∈ Tuj a convex set, and the moment function
ψuj : Wuj × Θuj × Tujn 7→ R, (w, θ, t) 7→ ψuj (w, θ, t)
(H.77)
is a Borel measurable map.
We assume that the (continuum) nuisance parameter η 0 can be modelled and estimated by ηb =
(b
ηuj )u∈U ,j∈[p̃] . We will discuss examples where the corresponding η 0 can be estimated using modern regularization and post-selection methods such as Lasso and Post-Lasso (although other procedures can be applied).
The estimator θ̌uj of θuj is constructed as any approximate n -solution in Θuj to a sample analog of the
moment condition (H.76), i.e.,
max sup |En [ψuj (Wuj , θ̌uj , ηbuj )]| − inf |En [ψuj (Wuj , θ, ηbuj )]| 6 n = oP (n−1/2 δn ).
(H.78)
j∈[p̃] u∈U
θj ∈Θuj
As discussed before, we rely on an orthogonality condition for regular estimation of θuj , which we will
state next.
66
BELLONI, CHEN, AND CHERNOZHUKOV
Definition 1 (Near Orthogonality Condition). For each u ∈ U and j ∈ [p̃], we say that ψuj obeys a
general form of orthogonality with respect to Huj uniformly in u ∈ U, if the following conditions hold: the
Gâteaux derivative map
!
n
h
io
Du,j,r̄ [η̃uj − ηuj ] := ∂r E ψuj Wuj , θuj , ηuj + r η̃uj − ηuj
r=r̄
exists for all r ∈ [0, 1), η̃ ∈ Huj , j ∈ p̃, and u ∈ U and vanishes at r = 0, namely,
|Du,j,0 [η̃uj − ηuj ]| 6 δn n−1/2
for all η̃uj ∈ Huj .
(H.79)
In what follows, we shall denote by c0 , c, and C some positive constants.
Assumption C1 (Moment Condition). Consider a random element W , taking values in a measure space
(W, AW ), with law determined by a probability measure P ∈ Pn . The observed data ((Wiu )u∈U )ni=1 consist of
n i.i.d. copies of a random element (Wu )u∈U which is generated as a suitably measurable transformation with
respect to W and u. Uniformly for all n > n0 and P ∈ Pn , the following conditions hold: (i) The true parameter value θuj obeys (H.76) and is interior relative to Θuj , namely there is a ball of radius Cn−1/2 un log n
√
centered at θuj contained in Θuj for all u ∈ U, j ∈ [p̃] with un := E[supu∈U ,j∈[p̃] | nEn [ψuj (Wuj , θuj , ηuj )]|];
(ii) For each u ∈ U and j ∈ [p̃], the map (θ, η) ∈ Θuj × Huj 7→ E[ψuj (Wuj , θ, η)]| is twice continuously differentiable; (iii) For all u ∈ U and j ∈ [p̃], the moment function ψuj obeys the orthogonality condition given in
Definition 1 for the set Huj = Hujn specified in Assumption C2; (iv) The following identifiability condition
holds: |E[ψuj (Wuj , θ, ηuj )]| > 21 |Juj (θ − θuj )| ∧ c0 for all θ ∈ Θuj , with Juj := ∂θ E[ψuj (Wuj , θ, ηuj )]|θ=θuj
satisfies 0 < jn < |Juj | < C < ∞ for all u ∈ U and j ∈ [p̃]; (v) The following smoothness conditions holds
E[{ψ
(W
,θ,η)−ψ
(W
,θ̄,η̄)}2 ]
uj
uj
uj
uj
6 C,
(a) supu∈U ,j∈[p̃],(θ,θ̄)∈Θ2uj ,(η,η̄)∈H2ujn
{|θ−θ̄|∨kη−η̄ke }α
(b) supu∈U ,(θ,η)∈Θuj ×Hujn ,r∈[0,1)
|∂r E [ψuj (Wuj , θ, ηuj + r{η − ηuj })] |/kη − ηuj ke 6 B̄1n ,
(c) supu∈U ,j∈[p̃],(θ,η)∈Θuj ×Hujn ,r∈[0,1)
|∂r2 E[ψuj (Wuj ,θuj +r{θ−θuj },ηuj +r{η−ηuj })]|
{|θ−θuj |2 ∨kη−ηuj k2e }
6 B̄2n .
Next we state assumptions on the nuisance functions. In what follows, let ∆n & 0, δn & 0, and τn & 0
be sequences of constants approaching zero from above at a speed at most polynomial in n (for example,
δn > 1/nc for some c > 0).
Assumption C2 (Estimation of Nuisance Functions). The following conditions hold for each n > n0 and all
P ∈ Pn . The estimated functions ηbuj ∈ Hujn with probability at least 1 − ∆n , Hujn is the set of measurable
maps η̃uj such that
sup max kη̃uj − ηuj ke 6 τn ,
u∈U j∈[p̃]
here the e-norm is the same as in Assumption C1, and whose complexity does not grow too quickly in the
sense that F1 = {ψuj (Wuj , θ, η) : u ∈ U, j ∈ [p̃], θ ∈ Θuj , η ∈ Hujn ∪ {ηuj }} is suitably measurable and its
uniform covering entropy obeys:
sup log N (kF1 kQ,2 , F1 , k · kQ,2 ) 6 sn(U ,p̃) (log(an /)) ∨ 0,
Q
where F1 (W ) is an envelope for F1 which is measurable with respect to W and satisfies F1 (W ) >
supu∈U ,j∈[p̃],θ∈Θuj ,η∈Hujn |ψuj (Wuj , θ, η)| and kF1 kP,q 6 Kn for q > 2. The complexity characteristics
QUANTILE GRAPHICAL MODELS
67
an > max(n, Kn , e) and sn(U ,p̃) > 1 obey the growth conditions:
p
1
n−1/2 sn(U ,p̃) log(an ) + n−1 sn(U ,p̃) n q Kn log(an ) 6 τn
p
1
1
{(1 ∨ B̄1n )(τn /jn )}α/2 sn(U ,p̃) log(an ) + sn(U ,p̃) n q − 2 Kn log(an ) log n 6 δn ,
√
nB̄2n (1 ∨ B̄1n )(τn /jn )2 6 δn
and
here B̄1n , B̄2n , jn , q and α are defined in Assumption C1.
Theorem 5 (Uniform Bahadur representation for a Continuum of Target Parameters). Under Assumptions
C1 and C2, for an estimator (θ̌uj )u∈U ,j∈[p̃] that obeys equation (H.78),
√ −1
nσuj (θ̌uj − θuj ) = Gn ψ̄uj + OP (δn ) in `∞ (U × [p̃]), uniformly in P ∈ Pn ,
−1 −1
−2 2
2
here ψ̄uj (W ) := −σuj
Juj ψuj (Wuj , θuj , ηuj ) and σuj
= E[Juj
ψuj (Wuj , θuj , ηuj )].
The uniform Bahadur representation derived in Theorem 5 is useful in the construction of simultaneous
confidence bands for (θuj )u∈U ,j∈[p̃] . This is achieved by new high-dimensional central limit theorems that
have recently been developed in [24] and [25]. We will make use of the following regularity condition. In what
−1 b−1
Juj ψuj (Wuj , θ̌uj , ηbuj )
follows δ̄n and ∆n are fixed sequences going to zero, and we denote ψbuj (W ) := −b
σuj
b
be the estimators of ψ̄uj (W ), with Juj and σ
buj being suitable estimators of Juj and σuj . In what follows,
k · kPn ,2 denotes the empirical L2 (Pn )-norm with Pn as the empirical measure of the data.
Assumption C3 (Score Regularity). The following conditions hold for each n > n0 and all P ∈ Pn . (i)
The class of function induced by the score F0 = {ψ̄uj (W ) : u ∈ U, j ∈ [p̃]} is suitably measurable and its
uniform covering entropy obeys:
sup log N (kF0 kQ,2 , F0 , k · kQ,2 ) 6 %n (log(An /)) ∨ 0,
Q
here F0 (W ) is an envelope for F0 which is measurable with respect to W and satisfies F0 (W ) >
supu∈U ,j∈[p̃] |ψ̄uj (W )| and kF0 kP,q 6 Ln for q > 4. Furthermore, c 6 supu∈U ,j∈[p̃] E[|ψ̄uj (W )|k ] 6 CLk−2
n
for k = 2, 3, 4. (ii) The set Fb0 = {ψ̄uj (W ) − ψbuj (W ) : u ∈ U, j ∈ [p̃]} satisfies the conditions log N (, Fb0 , k ·
kP ,2 ) 6 %̄n (log(Ān /)) ∨ 0, and sup
En [{ψ̄uj (W ) − ψbuj (W )}2 ] 6 δ̄n {ρn ρ̄n log(An ∨ n) log(Ān ∨ n)}−1
u∈U ,j∈[p̃]
n
with probability 1 − ∆n .
Assumption C3 imposes condition on the class of functions induced by ψ̄uj and on its estimators ψbuj .
Typically the bound Ln on the moment of the envelope is smaller than Kn , and in many settings ρ̄n = ρn . dU
the dimension of U.
Next let N denote a mean zero Gaussian process indexed by U × [p̃] with covariance operator given by
E[ψ̄uj (W )ψ̄u0 j 0 (W )] for j, j 0 ∈ [p̃] and u, u0 ∈ U. Because of the high-dimensionality, indeed p̃ can be larger
than the sample size n, the central limit theorem will be uniformly valid over “rectangles”. This class of
sets are rich enough to construct many confidence regions of interest in applications accounting for multiple
testing. Let R denote the set of rectangles R = {z ∈ Rp̃ : maxj∈A zj 6 t, maxj∈B (−zj ) 6 t} for all A, B ⊂ [p̃]
and t ∈ R. The following result is a consequence of Theorem 5 above and Corollary 2.2 of [26].
Corollary 3. Under Assumptions C1, C2 and Assumption C3(i), with δn = o({ρn log(An ∨ n)}−1/2 ), and
ρn log(An ∨ n) = o({(n/L2n )1/7 ∧ (n1−2/q /L2n )1/3 }), we have that
−1
sup sup PP {sup n1/2 σuj
(θ̌uj − θuj )}p̃j=1 ∈ R − PP (N ∈ R) = o(1).
P ∈Pn R∈R
u∈U
68
BELLONI, CHEN, AND CHERNOZHUKOV
In order to derive a method to build confidence regions we approximate the process N by the Gaussian
multiplier bootstrap based on estimates ψbuj of ψ̄uj , namely
)
(
n
X
1
gi ψbuj (Wi )
Gb = (Gbuj )u∈U ,j∈[p̃] = √
n i=1
u∈U ,j∈[p̃]
are independent standard normal random variables which are independent from the data (Wi )ni=1 .
here
Based on Theorem 5.2 of [24], the following result shows that the multiplier bootstrap provides a valid
√
approximation to the large sample probability law of n(θ̌uj − θuj )u∈U ,j∈[p̃] over rectangles.
(gi )ni=1
Corollary 4 (Uniform Validity of Gaussian Multiplier Bootstrap). Under Assumptions C1, C2
and Assumption C3, with δn = o({(1 + dW )ρn log(An ∨ n)}−1/2 ) and ρn log(An ∨ n) = o({(n/L2n )1/7 ∧
(n1−2/q /L2n )1/3 }), we have that
p̃
1/2 −1
sup sup PP {sup n σuj (θ̌uj − θuj )}j=1 ∈ R − PP (Gb ∈ R | (Wi )ni=1 ) = o(1)
P ∈Pn R∈R
u∈U
Appendix I. Continuum of `1 -Penalized M-Estimators
For the reader’s convenience, this section collects results on the estimation of a continuum of estimation
of high-dimensional models via `1 -penalized estimators. We refer to [14] for the proofs.
Consider a data generating process with a response variable (Yu )u∈U and observable covariates (Xu )u∈U
satisfies for each u ∈ U,
θu ∈ arg minp E[Mu (Yu , Xu , θ, au )],
θ∈R
(I.80)
here θu is a p-dimensional vector, au is a nuisance function that capture the misspecification of the model,
Mu is a pre-specified function, and the pu -dimensional (pu 6 p) covariate Xu could have been constructed
based on transformations of other variables. This implies that
∂θ E[Mu (Yu , Xu , θu , au )] = 0 for all u ∈ U.
The solution θu is assumed to be sparse in the sense that for some process (θu )u∈U satisfies
kθu k0 6 s for all u ∈ U.
Because of the nuisance function, such sparsity assumption is very mild and formulation (I.80) encompasses
several cases of interest including approximate sparse models. We focus on the estimation of (θu )u∈U and
we assume that an estimate b
au of the nuisance function au is available and the criterion Mu (Yu , Xu , θu ) :=
Mu (Yu , Xu , θu , b
au ) is used as a proxy for Mu (Yu , X, θu , au ).
In the case of linear regression we have Mu (y, x, θ) = 21 (y − x0 θ)2 . In the logistic regression case, we
have Mu (y, x, θ) = −{1(y = 1) log G(x0 θ) + 1(y = 0) log(1 − G(x0 θ))} with G is the logistic link function
G(t) = exp(t)/{1 + exp(t)}. Additional examples include quantile regression models for u ∈ (0, 1).
0
Example 8 (Quantile Regression Model). Consider a data generating process Y = FY−1
|X (U ) = X θU +
rU (X), with U ∼ Unif(0, 1), and X is a p-dimensional vector of covariates. The criterion Mu (y, x, θ) =
(u − 1{y 6 x0 θ})(y − x0 θ) with the (trivial) estimate b
au = 0 for the nuisance parameter au = ru .
QUANTILE GRAPHICAL MODELS
69
Example 9 (Lasso with Estimated Weights). We consider a linear model defined as fu Y = fu X 0 θu +
r̄u + ζu , E[fu Xζu ] = 0, here X are p̄-dimensional covariates, θu is a s-sparse vector, and r̄u is an approximation error satisfies supu∈U En [r̄u2 ] .P s log p̄/n. In this setting, (Y, X) are observed and only an
estimator fbu of fu is available. This corresponds to nuisance parameter au = (fu , r̄u ) and b
au = (fbu , 0) so
2
0
2
2
0 2
b
that En [Mu (Y, X, θ, au )] = En [fu (Y − X θ − r̄u ) ] and En [Mu (Y, X, θ)] = En [fu (Y − X θ) ].
We assume that n i.i.d. observations from dgps with (I.80) holds, {(Yiu , Xiu )u∈U }ni=1 , are observed to
b u , we define
estimate (θu )u∈U . For each u ∈ U, a penalty level λ, and a diagonal matrix of penalty loadings Ψ
the `1 -penalized Mu -estimator (Weighed-Lasso) as
λ b
θbu ∈ arg min En [Mu (Yu , Xu , θ)] + kΨ
u θk1 .
θ
n
(I.81)
Furthermore, for each u ∈ U, the post-penalized estimator (Post-Lasso) based on a set of covariates Teu is
then defined as
θeu ∈ arg min En [Mu (Yu , Xu , θ)] : support(θ) ⊆ Teu .
(I.82)
θ
Potentially, the set Teu contains support(θbu ) and possibly additional variables deemed as important (although
in that case the total number of additional variables should also obey the same growth conditions that s
obeys). We will set Teu = support(θbu ) unless otherwise noted.
bu =
In order to handle the functional response data, the penalty level λ and penalty loading Ψ
diag({b
luk , k = 1, . . . , p}) need to be set to control selection errors uniformly over u ∈ U. The choice of
b u0 = diag({luk , k =
loading matrix is problem specific and we suggest to mimic the following “ideal” choice Ψ
1, . . . , p}) with
luk = {En {∂θk Mu (Yu , Xu , θu , au )}2 }1/2
(I.83)
which is motivated by the use of self-normalized moderate deviation theory. In that case, it is suitable to
set λ so that with high probability
λ
b −1 En [∂θ Mu (Yu , Xu , θu , au )]
> c sup Ψ
u0
n
u∈U
∞
,
(I.84)
here c > 1 is a fixed constant. Indeed, in the case that U is a singleton the choice above is similar to [21],
[13], and [20]. This approach was first employed for a continuum of indices U in the context of `1 -penalized
quantile regression processes by [11].
To implement (I.84), we propose setting the penalty level as
√
λ = c nΦ−1 (1 − ξ/{2pNn }),
(I.85)
here Nn is a measure of the class of functions indexed by U, 1 − ξ (with ξ = o(1)) is a confidence level
associated with the probability of event (I.84), and c > 1 is a slack constant. In many settings we can take
Nn = ndU . If the set U is a singleton, Nn = 1 suffices which corresponds to what is used in [17].
I.1. Generic Finite Sample Bounds. In this subsection we derive finite sample bounds based on Assumption C4 below. This assumption provides sufficient conditions that are implied by a variety of settings
including generalized linear models.
70
BELLONI, CHEN, AND CHERNOZHUKOV
Assumption C4 (M-Estimation Conditions). Let {(Yiu , Xiu , u ∈ U), i = 1, . . . , n} be n i.i.d. observations
of the model (I.80) and let Tu = support(θu ), here kTu k0 6 s, u ∈ U. With probability 1 − ∆n we have that
for all u ∈ U there are weights wu = wu (Yu , Xu ) and Cun such that:
√
(a) |En [∂θ Mu (Yu , Xu , θu ) − ∂θ Mu (Yu , Xu , θu , au )]0 δ| 6 Cun k wu Xu0 δkPn ,2 ;
b u0 6 Ψ
b u 6 LΨ
b u0 for ` > 1/c, and let c̃ = Lc+1 supu∈U kΨ
b u0 k∞ kΨ
b −1 k∞ ;
(b) `Ψ
u0
`c−1
(c) for all δ ∈ Au there is q̄Au > 0 such that
√
En [Mu (Yu , Xu , θu + δ)] − En [Mu (Yu , Xu , θu )] − En [∂θ Mu (Yu , Xu , θu )]0 δ + 2Cun k wu Xu0 δkPn ,2
√
√
> k wu Xu0 δk2Pn ,2 ∧ q̄Au k wu Xu0 δkPn ,2 .
In many applications we take the weights to be wu = wu (Xu ) = 1 but we allow for more general weights.
Assumption C4(a) bounds the impact of estimating the nuisance functions uniformly over u ∈ U. In the
b u are assumed
setting with s-sparse estimands, we typically have Cun . {n−1 s log(pn)}1/2 . The loadings Ψ
b
larger (but not too much larger) than the ideal choice Ψu0 defined in (I.83). This is formalized in Assumption
C4(b). Assumption C4(c) is an identification condition that will be imposed for specific choices of Au and
qAu . It relates to conditions in the literature derived for the case of a singleton U and no nuisance functions,
see the restricted strong convexity8 used in [63] and the non-linear impact coefficients used in [11] and [18].
The following results establish rates of convergence for the `1 -penalized solution with estimated nuisance
functions (I.81), sparsity bounds and rates of convergence for the post-selection refitted estimator (I.82).
They are based on restricted eigenvalue type conditions and sparse eigenvalue conditions. With the restricted
√
eigenvalue is defined as κ̄u,2c̃ = inf δ∈∆u,2c̃ k wu Xu0 δkPn ,2 /kδTu k In the results for sparsity and post-selection
refitted models, the minimum and maximum sparse eigenvalues,
√
√
k wu Xu0 δk2Pn ,2
k wu Xu0 δk2Pn ,2
φmin (m, u) = min
and φmax (m, u) = max
,
kδk2
kδk2
16kδk0 6m
16kδk0 6m
are also relevant quantities to characterize the behavior of the estimators.
Lemma 20. Suppose that Assumption C4 holds with δ ∈ Au = {δ : kδTuc k1 6 2c̃kδTu k1 } ∪ {δ : kδk1 6
n
o
√
b −1 k∞ n
√
6ckΨ
λ s
1
0
u0
b
w
X
δk
}
and
q̄
>
3
(L
+
C
k
)k
Ψ
k
+
9c̃C
. Suppose that λ satisfies conP
,2
A
u
un
u0
∞
un
u
n
u
`c−1
λ
c
nκ̄u,2c̃
dition (I.84) with probability 1 − ∆n . Then, with probability 1 − 2∆n we have uniformly over u ∈ U
n
o
√
√
b u0 k∞ λ s + 9c̃Cun
k wu Xu0 (θbu − θu )kPn ,2 6 3 (L + 1c )kΨ
nκ̄u,2c̃
n
on
o
√
√
b −1 k∞ n
6ckΨ
s
1
u0
b u0 k∞ λ s + 9c̃Cun
kθbu − θu k1 6 3 (1+2c̃)
+
C
(L
+
)k
Ψ
un
κ̄u,2c̃
`c−1
λ
c
nκ̄u,2c̃
Lemma 21 (M-Estimation Sparsity). In addition to conditions of Lemma 20, assume that with probability
1 − ∆n for all u ∈ U and δ ∈ Rp we have
√
|{En [∂θ Mu (Yu , Xu , θbu ) − ∂θ Mu (Yu , Xu , θu )]}0 δ| 6 Lun k wu Xu0 δkPn ,2 .
Let Mu = {m ∈ N : m > 2φmax (m, u)L2u } with Lu =
we have that
b −1 k∞ n
ckΨ
u0
c`−1
λ
sbu 6 min φmax (m, u)L2u
m∈Mu
{Cun + Lun }, then with probability 1 − 3∆n
for all u ∈ U.
8Assumption C4 (a) and (c) could have been stated with {C /√s}kδk instead of C k√w X 0 δk
un
un
u u Pn ,2 .
1
QUANTILE GRAPHICAL MODELS
71
Lemma 22. Let Teu , u ∈ U, be the support used for post penalized estimator (I.82) and s̃u = kTeu k0
its cardinality. In addition to conditions of Lemma 20, supposethat Assumption C4(c)
holds also for
√
Au = {δ : kδk0 6 s̃u + s} with probability 1 − ∆n , q̄Au > 2
2{En [Mu (Yu , Xu , θ̃u )] −
s̃u +su kEn [Su ]k∞
√
+ 3Cun
φmin (s̃u +su ,u)
and q̄Au >
1/2
En [Mu (Yu , Xu , θu )]}+ .
√
k wu Xu0 (θ̃u − θu )kPn ,2
Then, we have uniformly over u ∈ U
√
s̃u + su kEn [Su ]k∞
1/2
6 {En [Mu (Yu , Xu , θ̃u )] − En [Mu (Yu , Xu , θu )]}+ + p
+ 3Cun .
φmin (s̃u + su , u)
In Lemma 22, if Teu = support(θbu ), we have that
En [Mu (Yu , Xu , θ̃u )] − En [Mu (Yu , Xu , θu )] 6 En [Mu (Yu , Xu , θbu )] − En [Mu (Yu , Xu , θu )] 6 λC 0 kθbu − θu k1
b u0 k∞ .
and supu∈U kEn [Su ]k∞ 6 C 0 λ with high probability, C 0 6 L supu∈U kΨ
These results generalize important results of the `1 -penalized estimators to the case of functional response
data and estimated of nuisance functions. A key assumption in Lemmas 20-22 is that the choice of λ satisfies
(I.84). We next provide a set of simple generic conditions that will imply the validity of the proposed choice.
These generic conditions can be verified in many applications of interest.
Condition WL. For each u ∈ U, let Su = ∂θ Mu (Yu , Xu , θu , au ), suppose that:
(i) sup max{E[|Suk |3 ]}1/3 /{E[|Suk |2 ]}1/2 Φ−1 (1 − ξ/{2pNn }) 6 δn n1/6 , for all u ∈ U, k ∈ [p];
u∈U k6p
(ii) Nn > N (, U, dU ), here is such that with probability 1 − ∆n :
sup
n [Su −Su0 ]k∞
max kEE[|S
6 δn n− 2 , and
2 1/2
uk | ]
1
dU (u,u0 )6 k6p
2
2
2
|E[Suk
−Su
0 k ]|+|(En −E)[Suk ]|
2]
E[|S
|
uk
dU (u,u0 )6 k6p
sup
max
6 δn .
The following technical lemma justifies the choice of penalty level λ. It is based on self-normalized
moderate deviation theory.
Lemma 23 (Choice of λ). Suppose Condition WL holds, let c0 > c > 1 be constants, ξ ∈ [1/n, 1/ log n], and
√
λ = c0 nΦ−1 (1 − ξ/{2pNn }). Then for n > n0 large enough depends only on Condition WL,
−1
b
P λ/n > c sup kΨu0 En [∂θ Mu (Yu , Xu , θu , au )]k∞ > 1 − ξ − o(ξ) − ∆n .
u∈U
We note that Condition WL(ii) contains high level conditions. See [16] for examples that satisfy these
conditions. The following corollary summarizes these results for many applications of interest in well behaved
designs.
Corollary 5 (Rates under Simple Conditions). Suppose that with probability 1 − o(1) we have that Cun ∨
Lun 6 C{n−1 s log(pn)}1/2 , (Lc+1)/(`c−1) 6 C, wu = 1, and Condition WL holds with log Nn 6 C log(pn).
Further suppose that with probability 1 − o(1) the sparse minimal and maximal eigenvalues are well behaved,
c 6 φmin (s`n , u) 6 φmax (s`n , u) 6 C for some `n → ∞ uniformly over u ∈ U. Then with probability 1 − o(1)
we have
r
r
s log(pn)
s2 log(pn)
0 b
b
, sup kθu − θu k1 .
, and sup kθbu k0 . s.
sup kXu (θu − θu )kPn ,2 .
n
n
u∈U
u∈U
u∈U
Moreover, if Teu = support(θbu ), we have that
r
sup kXu0 (θ̃u
u∈U
− θu )kPn ,2 .
s log(pn)
n
72
BELLONI, CHEN, AND CHERNOZHUKOV
Appendix J. Bounds on Covering entropy
Let (Wi )ni=1 be a sequence of independent copies of a random element W taking values in a measurable
space (W, AW ) according to a probability law P . Let F be a set of suitably measurable functions f : W → R,
equipped with a measurable envelope F : W → R. The proofs for the following lemmas can be found in [16].
Lemma 24 (Algebra for Covering Entropies). Work with the setup above.
(1) Let F be a VC subgraph class with a finite VC index k or any other class whose entropy is bounded above
by that of such a VC subgraph class, then the uniform entropy numbers of F obey
sup log N (kF kQ,2 , F, k · kQ,2 ) . {1 + k log(1/)} ∨ 0
Q
(2) For any measurable classes of functions F and F 0 mapping W to R,
log N (kF + F 0 kQ,2 , F + F 0 , k · kQ,2 ) 6 log N 2 kF kQ,2 , F, k · kQ,2 + log N 2 kF 0 kQ,2 , F 0 , k · kQ,2 ,
log N (kF · F 0 kQ,2 , F · F 0 , k · kQ,2 ) 6 log N 2 kF kQ,2 , F, k · kQ,2 + log N 2 kF 0 kQ,2 , F 0 , k · kQ,2 ,
N (kF ∨ F 0 kQ,2 , F ∪ F 0 , k · kQ,2 ) 6 N (kF kQ,2 , F, k · kQ,2 ) + N (kF 0 kQ,2 , F 0 , k · kQ,2 ) .
(3) For any measurable class of functions F and a fixed function f mapping W to R,
log sup N (k|f | · F kQ,2 , f · F, k · kQ,2 ) 6 log sup N (/2kF kQ,2 , F, k · kQ,2 )
Q
Q
(4) Given measurable classes Fj and envelopes Fj , j = 1, . . . , k, mapping W to R, a function φ : Rk → R
Pk
such that for fj , gj ∈ Fj , |φ(f1 , . . . , fk ) − φ(g1 , . . . , gk )| 6 j=1 Lj (x)|fj (x) − gj (x)|, Lj (x) > 0, and fixed
functions f¯j ∈ Fj , the class of functions L = {φ(f1 , . . . , fk ) − φ(f¯1 , . . . , f¯k ) : fj ∈ Fj , j = 1, . . . , k} satisfies
k
k
X
X
log sup N
L j Fj
, L, k · kQ,2 6
log sup N k kFj kQ,2 , Fj , k · kQ,2 .
Q
Q,2
j=1
j=1
Q
Proof. See Lemma L.1 in [16].
Lemma 25 (Covering Entropy for Classes obtained as Conditional Expectations). Let F denote a class of
measurable functions f : W × Y → R with a measurable envelope F . For a given f ∈ F, let f¯: W → R be
R
the function f¯(w) := f (w, y)dµw (y) here µw is a regular conditional probability distribution over y ∈ Y
R
conditional on w ∈ W. Set F̄ = {f¯: f ∈ F} and let F̄ (w) := F (w, y)dµw (y) be an envelope for F̄. Then,
for r, s > 1,
log sup N (kF̄ kQ,r , F̄, k · kQ,r ) 6 log sup N ((/4)r kF kQ,s
e , F, k · kQ,s
e ),
Q
e
Q
here Q belongs to the set of finitely-discrete probability measures over W such that 0 < kF̄ kQ,r < ∞, and
e belongs to the set of finitely-discrete probability measures over W × Y such that 0 < kF k e < ∞. In
Q
Q,s
particular, for every > 0 and any k > 1,
log sup N (, F̄, k · kQ,k ) 6 log sup N (/2, F, k · kQ,k
e ).
Q
Proof. See Lemma L.2 in [16].
e
Q
| 10math.ST
|
A FAST ALGORITHM FOR THE GAS STATION PROBLEM
arXiv:1706.00195v1 [math.CO] 1 Jun 2017
KLEITOS PAPADOPOULOS AND DEMETRES CHRISTOFIDES
Abstract. In the gas station problem we want to find the cheapest path between two vertices of an n-vertex graph. Our
car has a specific fuel capacity and at each vertex we can fill our car with gas, with the fuel cost depending on the vertex.
Furthermore, we are allowed at most ∆ stops for refuelling.
In this short paper we provide an algorithm solving the problem in O(∆n2 + n2 log n) steps improving an earlier result
by Khuller, Malekian and Mestre.
c : V → R+ . Finally we are also given positive numbers U
and ∆.
1. Introduction
There are numerous problems in the literature in which the
task is to optimise the travel from one location to another
or to optimise a tour visiting a specific set of locations. The
problems usually differ in the restrictions that we may put
into the way we can travel as well as in the notion of what
an ‘optimal’ route means. Of course one could theoretically
check all possible ways to travel and pick out the optimal
one. However we care about finding the optimal route in
a much quicker way as usually checking all possibilities is
impractical.
Each vertex v of G corresponds to a gas station and the
number c(v) corresponds to the cost of the fuel at this station. Given an edge e = uv of G, the number d(e) = d(u, v)
corresponds to the distance between the vertices u and
v, or what is essentially equivalent, to the amount of gas
needed to travel between u and v. Finally, the number U
corresponds to the maximum gas capacity of our car.
Our task is to find the cheapest way possible to move from
vertex s to vertex t if we are allowed to make at most ∆
refill stops.
One of the most widely known abstractions of travel optimisation problems is that of the shortest paths which
although very general in their definition, fail to take into
consideration most of the aspects that arise in real world.
Perhaps one of their more practical generalisations is the
‘gas station problem’, introduced by Khuller, Malekian and
Mestre in [2]. Out of the infinitude of possible of possible parameters it includes one more central aspect of the
travelling agent, that of its limited fuel capacity and fuel
consumption during the travelling. As it is the case for
the shortest paths problem its also with the gas station
problem that it can be used in a variety of problems that
are not directly related with travelling optimisation
We make two further assumptions:
Our first assumption concerns the function d : E → R+ .
We will follow the natural assumption that it satisfies the
triangle inequalities. I.e.
d(u, v) + d(v, w) > d(u, w)
for every u, v, w ∈ V .
Our second assumption concerns the amount of fuel that
we have initially in our car. We will make the assumption
that we start with an empty fuel tank. We also consider
the filling in our tank at this vertex as one of the refill
stops. This does not make much difference. Indeed suppose that initially we have an amount g of gas. Instead of
solving the gas station problem for the graph G, we modify
this graph by adding a new vertex s0 . We define the new
distances by d(s0 , v) = U − g + d(s, v) for every v ∈ V (G).
I.e. the vertex s0 has distance U −g from s and furthermore
The setting of the gas station problem is as follows:
We are given a complete graph G = (V, E), two specific vertices s, t of G and functions d : E → R+ and
Date: June 2, 2017.
2010 Mathematics Subject Classification. 90B06,90C35,68Q25.
Key words and phrases. gas station problem, transportation, algorithm complexity.
Kleitos Papadopoulos, InSPIRE, Agamemnonos 20, Nicosia, 1041, Cyprus, [email protected].
Demetres Christofides, School of Sciences, UCLan, 7080 Pyla, Larnaka, Cyprus, [email protected].
1
2
KLEITOS PAPADOPOULOS AND DEMETRES CHRISTOFIDES
the shortest path from s0 to any other vertex v of G is via
s. We also define c(s0 ) = 0. It is then obvious that if we
start from s0 , we should completely fill our tank and then
move to vertex s. The only difference is that we are now
allowed one fewer refill stop than before. So solving the
gas station problem for G starting from s with g units of
gas is equivalent to solving the gas station problem for G0
starting from s0 with no gas.
An algorithm solving the gas station problem that runs
in O(∆n2 log n) was introduced by Khuller, Malekian and
Mestre in [2]. The main result of our article is the following:
Theorem 1. Given an n vertex graph G, there is an algorithm which solves the gas station problem with ∆ stops
in at most O(∆n2 + n2 log n) steps.
We should point out that the algorithm in [2] makes similar
assumptions to ours. It explicitly mentions the assumption
that the car starts with an empty fuel tank. It does not
mention explicitly the assumption that the distances in the
graph need to satisfy the triangle inequalities. However it
does use it implicitly in its Lemma 1. (See the proof of our
Lemma 3 which makes explicit why we do indeed need the
distances to satisfy the triangle inequalities.)
We will prove Theorem 1 in the next section. In a couple of
instances our algorithm will call some familiar algorithms
with known running time. The interested reader can find
more details about those algorithms in many algorithms
or combinatorial optimisation books, for example in [1].
2. Proof of Theorem 1
We start by ordering all edge distances. Since there are
O(n2 ) edges, this can be done in O(n2 log n) steps, using
e.g. heapsort. In fact we will not use the full ordering of
the edge distances. What we will need are the following
local orderings:
For each v ∈ V we create an ordering v1 , . . . , vn−1 of the
vertices of V \ {v} such that d(v, vi ) 6 d(v, vj ) for i 6 j.
We will call this the local edge ordering at v. (Note
that this definition might be a bit misleading as the local
edge ordering at v is an ordering of the vertices of V \ {v}.
Of course, this gives an ordering of the edges incident to v
and this is where it gets its name from.)
Of course all of these local orderings can also be computed
in O(n2 log n) steps.
So it is enough to show how to solve the gas station problem in O(∆n2 ) time assuming that the edges are already
ordered by distance.
The fact that the edge distances satisfy the triangle inequalities is needed to prove the following simple lemma:
Lemma 2. There is an optimal route during which we
fill our car with a positive amount of gas at every station
(apart from the last one).
Proof. Amongst all optimal routes, pick one passing
through the smallest number of vertices. Suppose that
it passes through vertices v1 , v2 , . . . , vk in that order. We
definitely need to fill our car with gas at v1 as we start
with an empty tank. Suppose now for contradiction that
in this optimal path we do not fill our car at station vi for
some 1 < i < k. Then, instead of moving from vi−1 to
vi+1 through vi , we could have moved to it directly. This
is indeed possible as
d(vi−1 , vi+1 ) 6 d(vi−1 , vi ) + d(vi , vi+1 )
and it is a contradiction as we assumed that our optimal
path is both optimal and minimal.
Lemma 2 is needed to prove the following slightly modified lemma from [2]. Even though it looks completely
obvious, we nevertheless provide a detailed proof in order
to make explicit the need for using Lemma 2 and thus to
require that the distances in the graph satisfy the triangle
inequalities. Lin [3] also makes explicit this requirement.
Essentially the same lemma also appears in [4].
Lemma 3. There is an optimal route, say passing through
vertices v1 , v2 , . . . , vk in that order, where v1 = s and
vk = t, for which an optimal way to refill the tank is as
follows:
(i) For 1 6 i 6 k − 2, if c(vi ) < c(vi+1 ), then at
station vi we completely fill our tank.
(ii) For i = k − 1, if c(vi ) < c(vi+1 ), then at station
vi we fill the tank with just enough gas in order to
reach vertex vi+1 with an empty tank.
(iii) For 1 6 i 6 k −1, if c(vi ) > c(vi+1 ) then at station
vi we fill the tank with just enough gas in order to
reach vertex vi+1 with an empty tank.
Proof. Pick an optimal route as given by Lemma 2.
Suppose that at some point during travelling through this
optimal route we reach vertex vi , with 1 6 i 6 k − 2 and
suppose that c(vi ) < c(vi+1 ). By Lemma 2 we have filled
some gas at station vi+1 . If we did not fully filled our gas
A FAST ALGORITHM FOR THE GAS STATION PROBLEM
at station vi then we could have reduced our cost by filling
more gas at vi and less at vi+1 , a contradiction. So at vi
we definitely must completely fill our tank.
If we reach vk−1 then of course we fill the car with just
enough gas in order to reach vk with our tank completely
empty.
Finally suppose that at some point during travelling
through this optimal route we reach vertex vi , with 1 6
i 6 k − 1 and c(vi ) > c(vi+1 ). Suppose we fill the car at
vi with a units of gas and that when reaching vi+1 we still
have b > 0 units of gas left. Suppose that at vi+1 we fill
our tank with c > 0 units of gas. Of course we could still
reach vi+1 with the same amount of gas but spend at most
as much as before if we filled our car with a − b units of
gas at vi and b + c units of gas at vi+1 .
So indeed the process described in the Lemma is an optimal refilling process.
An immediate and important Corollary of Lemma 3 is the
following:
Corollary 4. There is an optimal route such that for every vertex u that we reach, either at the previous step we
had an empty tank, or two steps before we completely filled
our tank.
Given a vertex u ∈ V and an integer r ∈ {0, 1, . . . , ∆} we
write C0 [u, r] for the minimal cost of starting from s with
an empty tank, and finishing at u with an empty tank going through exactly r edges. We also write CU [u, r] for
the minimal cost of starting from s with an empty tank,
and finishing at u with a full tank going through exactly r
edges.
3
∞. At the end of the algorithm the minimal cost computed
will be less than ∞, guaranteeing that the optimal route
produced does not travel through such edges. (Unless of
course it is impossible to make such a travel, whatever the
cost.)
Note that the ordering of the edges we have already computed using the function d is also an ordering of the edges
using the function d0 . On the other hand d0 does not satisfy the triangle inequalities but we need not worry about
this as we have already used them for Corollary 4. Of
course changing d to d0 does not alter the conclusion of
the Corollary even though d0 does not satisfy the triangle
inequalities.
Our task is to calculate all C0 [t, r] for 0 6 r 6 ∆. The
required quantity will then be the minimum of these. This
minimum can be calculated in ∆ steps.
We will proceed inductively as follows:
Step 2r + 1: We calculate all C0 [u, r] with u ∈ U in O(n2 )
time.
Step 2r + 2: We calculate all CU [u, r] with u ∈ U in O(n2 )
time.
We can stop at Step 2∆ + 1 and so the total running time
will be O(∆n2 ) as required.
The first two steps are easy to perform and can even be
calculated in O(n) time as we have
(
0 u=s
C0 [u, 0] =
∞ u 6= s
and
CU [u, 0] = C0 [u, 0] + U c(u).
If it is impossible to be on the vertex u going through
exactly r edges then we define C0 [u, r] = CU [u, r] = ∞.
The use of the symbol ∞ is purely for convenience of the
proof. We could have just said that C0 [u, r] and CU [u, r]
are undefined. We will in fact make further use of this symbol. Note that given two vertices u, v with d(u, v) > U we
can never travel between u and v. Rather than separating
the cases in which d(u, v) > U or not it will be simpler to
make all these distances equal to ∞. To be more explicit,
we define a new function d0 : E → R+ ∪ {∞} by
(
d(e) if d(e) 6 U
0
d (e) =
∞
otherwise.
In what follows, we might often travel through edges e with
d0 (e) = ∞. The total cost of such a travel will be equal to
So now we assume that r > 1. For ease of exposition only,
we will first explain how to perform Step 2r + 2 as is is
easier to explain. To perform Step 2r + 2 we will assume
that we already performed Step 2r + 1. Of course there is
no problem with this as long as when we explain how to
perform Step 2r + 1 we do not make any use of the Step
2r + 2.
For Step 2r + 2 fix a specific u ∈ U . As there are n such
u’s, it is enough to show how to compute CU [u, r] in O(n)
time.
Note that from Corollary 4, if u is the r-th vertex that we
reach, either we reach it with an empty tank, or we leave
the (r − 1)-th vertex with a full tank.
4
KLEITOS PAPADOPOULOS AND DEMETRES CHRISTOFIDES
If the first case happens then
CU [u, r] = C0 [u, r] + U c(u).
If the second case happens and v is the (r − 1)-th vertex,
then
CU [u, r] = CU [v, r − 1] + d0 (v, u)c(u).
For each fixed v we will compute C000 [u, r; v] for all u 6= v in
O(n) time. This will be enough for our purposes. Indeed
this means that we can compute C000 [u, r; v] for all u, v ∈ V
in O(n2 ) time. But then for each fixed u we can compute
C000 [u, r] in an additional O(n) time and thus for all u in
an additional O(n2 ) time.
So CU [u, r] is the minimum of n quantities and can therefore be calculated in O(n) time as required. [Note that
in the second case we should have considered only those
v for which d(v, u) 6 U . Recall however that for all the
other vertices we defined d0 (v, u) = ∞ and so the total
cost through such vertices is ∞. The optimal cost is finite
and therefore the optimal route does not pass through such
vertices unless of course there is no possible way to reach
vertex u through r edges.]
To do this observe that
We now explain how to perform Step 2r + 1. We will in
fact consider many vertices at once, but for the moment
fix a specific u ∈ U and suppose that it is the r-th vertex
that we reach and we reach it with an empty tank. From
Corollary 4 there are two possible ways this could happen:
The two main observations that will allow us to compute
all C000 [u, r; v] fast enough are the following:
Type I: We reach the (r − 1)-st vertex with an empty tank.
Type II: We leave the (r − 2)-nd vertex with a full tank.
The main difficulty is to treat the ‘Type II’ possibilities as
a naive way to do it needs Θ(n2 ) time for each u ∈ V and
thus Θ(n3 ) time for all u ∈ V .
C000 [u, r; v] =
min
w∈S(u,v)
[CU [w, r − 2] + (d0 (w, v) + d0 (v, u) − U )c(v)]
where S(u, v) is the set of all vertices w for which such
‘Type II’ ways are possible. I.e. all w ∈ V \ {v} such that
d0 (w, v) + d0 (v, u) > U or equivalently all w ∈ V \ {v} such
that d(w, v) + d(v, u) > U
Firstly, we can rewrite
C000 [u, r; v] = d0 (v, u)c(v) +
min
w∈S(u,v)
[CU [w, r − 2] − (U − d0 (w, v))c(v)]
(∗)
To see the importance of this observation, suppose for a
moment that S(u, v) = V \ {v} for each u, v ∈ V . (We will
deal with the general situation in our second observation.)
Then we could compute all
min
w∈S(u,v)
[CU [w, r − 2] − (U − d0 (w, v))c(v)]
We write C00 [u, r] for the minimum cost over all ‘Type I’
ways in which we can reach u with an empty tank, and
C000 [u, r] for the minimum over all ‘Type II’ ways.
in O(n2 ) time by taking O(n) time for each v. But then
we could compute all C000 [u, r; v] in O(n2 ) time as required.
Note that if we reach u with a ‘Type I’ way through a
vertex v then
In principle however, the sets S(u, v) can be a lot different. However, for u, u0 with d(u, v) 6 d(u0 , v) we have
S(u, v) ⊆ S(u0 , v). Indeed if w ∈ S(u, v), then
C0 [u, r] = C0 [v, r − 1] + d0 (u, v)c(v)
So we can compute C00 [u, r] in O(n) time. Thus we can
compute C00 [u, r] for all u ∈ V in O(n2 ) time.
It remains to compute C000 [u, r]. If we can compute it for
all u ∈ V in O(n2 ) time then we will be done. Indeed, then
for each u ∈ V we have
C0 [u, r] =
min{C00 [u, r], C000 [u, r]}
and we can compute this in O(n) time for all u ∈ V .
To compute C000 [u, r] we introduce yet another piece of notation and define C000 [u, r; v] as the minimum over all ‘Type
II’ ways such that the (r − 1)-st vertex is v. Evidently,
C000 [u, r] =
min C000 [u, r; v]
v∈V \{u}
d(w, v) + d(v, u0 ) > d(w, v) + d(v, u) > U
so we also have w ∈ S(u0 , v).
This second observation says that there is some structure
to the sets S(u, v). We will use it in order to still manage
to compute all
min
w∈S(u,v)
[CU [w, r − 2] − (U − d0 (w, v))c(v)]
in O(n2 ) time by taking O(n) time for each v.
To prepare for its use we need the following lemma:
Lemma 5. For each fixed v ∈ V and let v1 , v2 , . . . , vn−1
be the local edge ordering at v. Define T (v1 , v) = S(v1 , v)
and T (vi , v) = S(vi , v) \ S(vi−1 , v) for 2 6 i 6 n − 1. Then
we can determine all T (vi , v) in O(n) time.
A FAST ALGORITHM FOR THE GAS STATION PROBLEM
Proof. Starting from the last vertex, we go backwards
through all the vertices of the local edge ordering at v
until we find one which does not belong to S(v1 , v). Suppose vk1 is this vertex. From our second observation we
have that
T (v1 , v) = S(v1 , v) = {vk1 +1 , . . . , vn }.
In the case that vk1 = vn , the understanding is that
T (v1 , v) = ∅.
Now we start from vk1 and we again go backwards through
all the vertices of the local edge ordering until we find one
which does not belong to S(v2 , v). Suppose vk2 is this
vertex. So from our second observation we have that
S(v2 , v) = {vk2 +1 , . . . , vn }
and
T (v2 , v) = {vk2 +1 , . . . , vk1 }.
Continuing inductively we can find k1 , k2 , . . . , kn−1 where
ki is the first vertex in the local edge ordering, starting
from ki−1 and going backwards, which does not belong to
S(vi , v). Then
5
We will prove our claim by induction on m. This is immediate for m = 1 since from (∗) we just need to compute
the minimum over all elements of T (v1 , v) = S(v1 , v). Suppose now that it is true for m = k. Then it is also true for
m = k + 1. Indeed again from our second observation we
have S(vk+1 , v) = S(vk , v) ∪ T (vk+1 , v). Thus to compute
the minimum of
CU [w, r − 2] − (U − d0 (w, v))c(v)
over all w ∈ S(vk+1 , v) we can start with the minimum over
all w ∈ S(vk , v) which is already computed and additionally consider the minimum over all elements of T (vk+1 , v).
This can be done in another O(|T (vk+1 , v)|) steps and since
the sets S(vk , v) and T (vk+1 , v) are disjoint, the total time
taken so far is O(|S(vk+1 , v)|). So the claim is also true
for m = k + 1 and this completes the proof of the claim.
3. Generalisations
We finally conclude with a list of comments mostly concerned with possible generalisations of the algorithm.
S(vi , v) = {vki +1 , . . . , vn }
and
T (vi , v) = {vki +1 , . . . , vki−1 }.
Note that at each step of the process we ask a question of
the form ‘Does vertex vi belong to the set S(vj , v)?’. This
question is answered in O(1) time by checking whether a
specific inequality holds. If needed, we then add vi to the
set T (vj , v) in O(1) time.
Furthermore, we have at most n questions from which we
obtain a ‘Yes’ answer, as each time we obtain a ‘Yes’ we
move on to the next vertex in the local edge ordering. We
also have at most n questions from which we obtain a ‘No’,
as each time we obtain a ‘No’, we move on to examining
inclusion in the next set of the form S(vj , v). So all these
checks and additions to lists can be done in O(n) time as
required.
Now we can proceed calculating C000 [u, r; v] for all u ∈ V
as follows:
Consider the local edge ordering v1 , v2 , . . . , vn−1 at v. We
will show that for each 1 6 m 6 n we can compute
C000 [v1 , r; v], . . . , C000 [vm , r; v]
in O(|S(vm , v)|) time.
Note that this completes the proof of Theorem 1. Indeed
by taking m = n − 1 we get that we can for fixed v ∈ V
compute all C000 [u, r; v] in O(n) time for all u ∈ V as required.
(1) The algorithm can also treat non-complete graphs
as well. For every edge uv that does not belong to
G we define d(u, v) = U + 1. Thus we will never
pass through that edge and so this is the same as
when the edge does not exist. In that case though,
the edge distances might not satisfy the triangle
inequalities and so we would have to compute all
shortest paths as in the previous generalisation. (If
for each edge uv we replace its edge length by the
lengths of the shortest path between u and v then
the new graph satisfies the triangle inequalities.)
(2) Given a graph with n vertices and m edges, by using Johnson’s algorithm the shortest paths can be
computed in O(n2 log n + mn) time. So in particular the total running time becomes O(n2 log n +
∆n2 + mn).
(3) It would be interesting to solve to gas station problem for particular classes of graphs. For example [4] gives a linear time algorithm in the case
that the graph is a path.
(4) The algorithm also treats directed graphs as well.
Here, given two vertices u and v we allow d(u, v) 6=
d(v, u). This is relevant in various situations. For
example going up the mountains needs more fuel
consumption than returning back from the mountains via the same road.
(5) If ∆ = n our algorithm solves the gas station problem in O(n3 ) time. The same time is also achieved
in [2] and in [3]. (In fact in [3] this is even achieved
for all pairs of vertices at once.)
6
KLEITOS PAPADOPOULOS AND DEMETRES CHRISTOFIDES
(6) We have actually solved the gas station problem
not only for the travel from the fixed starting vertex s to the fixed terminal vertex t, but actually
for the travel from the fixed starting vertex s to
every other vertex u of V .
(7) By applying the algorithm separately for each possible starting vertex (note that the local edge orderings are only computed once) we get a O(∆n3 )
algorithm for the all-pairs version of the gas station problem. This is better than the O(∆2 n3 ) algorithm in [2] but not as good as the O(n3 log ∆)
algorithm by Lin in [3].
(8) In practice, we do not only care about the minimum cost, but also on the minimum time needed
for the travel. This was actually the logic behind
the restriction of having at most ∆ stops. This
can still be treated by our algorithm if we have
a ‘cost function’ for the travel between each pair
of vertices and the ‘total cost’ is the sum of the
individual costs.
(9) The space complexity of the algorithm is O(n2 ).
Indeed note that in performing Steps 2r + 1 and
2r +2 we only need the knowledge of the following:
The n local edge orderings, with each local edge
ordering having space complexity O(n), the n2
edge distances, O(n) values of the form C0 [u, r −
1], CU [u, r − 1], C0 [u, r] and CU [u, r], where some
of those values are known from before and some of
those are calculated during the algorithm. Finally
we also need O(n2 ) values of the form C000 [u, r; v].
Note in particular than once we are done with
Steps 2r+1 and 2r+2 we do not use the C000 [u, r; v]
anymore.
References
[1] T. H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein, Introduction to algorithms, third edition, MIT Press, Cambridge,
MA, 2009.
[2] S. Khuller, A. Malekian and J. Mestre, To fill or not to fill: the
gas station problem, ACM Trans. Algorithms 7 (2011), Art. 36,
16 pp.
[3] S. H. Lin, Finding optimal refueling policies in transportation networks, in Algorithmic Aspects in Information and Management,
Lecture Notes in Computer Sciences 5034 (2008), 280–291.
[4] S. H. Lin, N. Gertsch and J. R. Russell, A linear-time algorithm
for finding optimal vehicle refueling policies, Oper. Res. Lett. 35
(2007), 290–296.
| 8cs.DS
|
Graph-Theoretic Framework for Unified Analysis of Observability and Data
Injection Attacks in the Smart Grid
Anibal Sanjab1 , Walid Saad1 , and Tamer Başar2
1
arXiv:1801.08951v1 [cs.SY] 26 Jan 2018
2
Wireless@VT, Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA, USA,
Emails: {anibals,walids}@vt.edu
Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, IL, USA, Email: [email protected]
Abstract—State estimation is a fundamental process needed for
the effective operation of the smart grid. As such, cyber-physical
attacks such as denial-of-service and data injection attacks, which
often target the availability and the integrity of the collected state
estimation measurements, can have detrimental consequences on
the operation of the system. In this paper, a novel graph-theoretic
framework is proposed to generalize the analysis of a broad set of
security attacks, including observability and data injection attacks,
that target the state estimator of a smart grid. First, the notion of
observability attacks – denial–of–service attacks on measurement
units which render the system unobservable – is defined based on
a proposed graph-theoretic construct. In this respect, an algorithm
is proposed to characterize the critical set of measurements which
must be removed along with a certain measurement to make the
system unobservable. It is then shown that, for the system to
be observable, these critical sets must be part of a maximum
matching over a proposed bipartite graph. In addition, it is shown
that stealthy data injection attacks are a special case of these
observability attacks. Then, various attack strategies and defense
policies, for observability and data injection attacks, are shown
to be amenable to analysis using variations of the formulated
maximum-matching problem over a bipartite graph. The proposed
framework is then shown to provide a unified basis for exact
analysis of four key security problems (among others), pertaining
to the characterization of: 1) The sparsest stealthy attack, 2) The
sparsest stealthy attack including a certain specific measurement,
3) A set of measurements which must be defended to thwart all
potential stealthy attacks, and 4) The set of measurements, which
when protected, can thwart any attack whose cardinality is below
a certain threshold. A case study using the IEEE 14-bus system
containing a set of 17 distributed measurement units is used to
corroborate the theoretical findings. In this case analysis, stealthy
attacks of lowest cardinality are characterized and shown to have
a cardinality equal to 2. In addition, it is shown, for example, that
defending only 3 out of the 17 measurements is enough to thwart
any stealthy attack with cardinality lower than 3, while defending
a minimum of 13 measurements is needed to thwart all possible
stealthy attacks.
Such attacks can become more pronounced when they target
critical power system functions such as state estimation. In this
regard, the power system state estimation is an integral smart grid
process in which system-wide measurements are collected and
processed to estimate the global state of operation of a power system [15]. State estimation is the basis for various grid operational
decisions such as congestion management, economic dispatch,
contingency analysis, and electricity pricing [16]. As a result,
the critical importance of state estimation to the sustainable
operation of the grid makes it a primary target of possible cyberphysical attacks [1], [2]. Such attacks may target the availability
of the collected measurements as well as their integrity.
In this respect, intercepting a subset of the collected measurement data using availability attacks (such as denial-of-service
attacks) can render the power system unobservable (i.e. not fully
observable), a state in which the collected measurements do not
provide enough independent equations to estimate the states.
Such cyber-physical attacks, to which we refer as observability
attacks hereinafter, will make the operator partially oblivious
to the real state of operation of the system, leading to uninformed operational decisions. Beyond observability attacks, data
injection attacks (DIAs) have emerged as a malicious type of
integrity attacks which aim at manipulating the collected state
estimation data, leading to inaccurate state estimation outcomes
that result in misinformed operational decisions with potentially
detrimental consequences [1], [4], [6]. As shown in [4], such
DIAs can stealthily target the power system state estimation
process – manipulating the collected measurements and altering
the state estimation outcome – while being undetectable by the
system operator using traditional bad data detection mechanisms.
Hence, due to their potential danger to system operation, such
stealthy data injection attacks (SDIAs) and observability attacks
have been the focus of various recent research efforts [17]–[23].
I. I NTRODUCTION
A. Related Works
With the integration of information and communication technologies in power systems, new security concerns have emerged
due to the potential exploitation of this cyber layer to infiltrate
and compromise the underlying physical system [1], [2]. Indeed,
in recent years, various studies have focused on analyzing the
security of emerging cyber-physical power systems [1]–[14] and
the effect of potential cyber attacks on the various operational
components of the grid, ranging from power system state estimation [4], [5], to electricity markets [6]–[8], power system
protection [9], [10], and power system dynamics and control [9],
[11]–[14].
This research was supported by the U.S. National Science Foundation under
Grants ECCS-1549894 and CNS-1446621.
In this regard, the works in [17] and [18] focused on computing a security set which comprises the minimum set of
measurements which must be attacked in addition to a certain
specific measurement in order to make the system unobservable.
Moreover, the work in [19] focused on computing the cardinality
of the smallest set of meters which when attacked render the
system unobservable. The authors in [20]–[22] extended such
observability problems to studying SDIAs. In this regard, these
works focused on characterizing the sparsest stealthy attack
containing a certain specific measurement. In addition, the work
in [23] focused on characterizing a set of measurements to
defend so that no attack which concurrently manipulates a set
of meters whose cardinality is below a certain threshold can be
stealthy. Hence, this latter analysis focuses on the defense against
resource-limited attackers. As such, these works have focused on
formulating and studying mathematical problems whose solutions enable anticipating potential sophisticated attacks – which
constitutes a first step towards deriving corresponding defense
mechanisms – and designing optimal defense strategies to thwart
such attacks and mitigate their potential effect.
The computational complexity of these problems [17]–[23]
has led to limiting the analysis of their solutions to special,
often approximated, cases or required the use of heuristics and
relaxation techniques which led to suboptimal solutions. For
example, for characterizing the sparsest observability attacks
containing a specific measurement, the work in [17] focused
on the special case of measurement sets of low cardinality
while the work in [18] derived an approximate solution that
is based on the solution of a min-cut problem. In addition,
with regard to the analysis of the sparsest SDIAs containing
a certain measurement [20]–[22], the work in [20] focused on
deriving an upper-bound on this stealthy attack set while the
work in [21] used min-cut relaxation techniques to approximate
the sought solution. Moreover, the work in [22] proposed a
heuristic algorithm which can approximate the solution of the
studied problem while an exact solution was found for the special
case in which power flows over all the transmission lines and
power injections into and out of every bus are assumed to be
measured. To defend against a resource-limited data injection
attacker, the authors in [23] used an l1 relaxation method for
characterizing the set of meters to defend to thwart SDIAs
launched by attackers whose attack space is limited by a certain
cardinality threshold. Other related security works are also found
in [24]–[29].
Therefore, this rich body of literature [17]–[29] employs
heuristics and approximation techniques to numerically approximate the solutions to these fundamental observability attacks
and SDIA problems rather than analytically characterizing their
solutions. As such, there is a need for an analytical framework
which allows modeling and studying such data availability and
integrity attacks and enables an analytical characterization of
mathematical solutions to such widely-studied security problems.
In addition, the fact that these works [17]–[23] studied correlated
problems but from different perspectives highlights the need for
a unified framework using which solutions to such correlated
observability attacks and SDIA problems can be studied and
derived.
B. Contributions
The main contribution of this paper is a novel unified graphtheoretic framework that enables a global detailed modeling
and understanding of observability attacks and SDIAs. As a
result, this framework provides a unified tool for analyzing
various widely-studied observability attacks and SDIA problems
such as those studied in [17]–[23], among others. In addition,
the proposed framework enables the characterization of exact
analytical solutions to such security problems, instead of relying
on numerical approximations or heuristics. This will enable a
fundamental analysis and modeling of potential attack strategies
and the derivation of defense strategies which can thwart and
mitigate the effect of such observability attacks and SDIAs.
In this regard, our proposed framework is based on a shift
in the modeling of observability attacks and SDIAs from a
linear algebra frame of reference to a graph-theoretic perspective.
As a result, based on this proposed framework, such attacks
can be modeled and analyzed by requiring only power system
topological data, namely, the power system 1-line diagram and
the location of deployed measurement units without the need for
neither line parameters data nor the exact knowledge of power
flow levels throughout the system.
To build the proposed framework, we first begin by introducing a graph-theoretic basis of observability attacks and, then, we
prove that SDIAs are a special case of such observability attacks.
In this respect, we introduce an algorithm providing a step-bystep approach for building critical sets, a set of measurements
– containing a certain specific measurement – which, when
removed, render the system unobservable. We then prove that for
a DIA to be stealthy, the attacked measurements should strictly
result in leaving critical sets unmatched as part of a maximum
matching over an introduced bipartite graph. As such, a graphtheoretic model of SDIAs is then introduced based on which
the solutions to various well-studied SDIA problems are analytically characterized. In particular, we show that our developed
framework can be readily applied to derive analytical solutions
to various SDIA problems such as, but not limited to: 1) Finding
the stealthy attack of lowest cardinality, 2) Finding the stealthy
attack of lowest cardinality, including a specific measurement,
3) Finding a set of measurements which when defended can
thwart all possible stealthy attacks, and 4) Finding a set of
measurements to defend against a resource-limited attacker,
among others. A case study using the IEEE 14-bus system, with
17 distributed measurement units, is considered throughout the
paper to showcase the developed analytical concepts. In this case
study, we characterize the sparsest SDIAs which can successfully
target the system and show that the cardinality of such attacks
is equal to 2. In addition, the performed case analysis on the
IEEE 14-bus system shows that defending a characterized set
of 13 (out of 17) measurements is necessary to prevent any
successful SDIAs, while defending only 3 measurements is
enough to thwart any stealthy attack of cardinality lower than
3. This, hence, enables the defender to build on some acquired
knowledge regarding the resources of the adversaries to derive
a corresponding defense strategy.
In this respect, the derived analytical results and presented
case study showcase the importance of the proposed framework
for studying various correlated observability attacks and SDIA
problems and pave the way for further analyzing additional
emerging problems in that field.
The rest of the paper is organized as follows. Section II
introduces state estimation and power system observability. Section III introduces our proposed graph-theoretic foundation of
observability attacks and shows its impact on modeling and
analyzing such data availability attacks. Section IV introduces
the proposed graph-theoretic framework for modeling SDIAs, as
well as studies and solves various well-studied SDIA problems.
Section V concludes the paper and provides an outlook detail-
ing the impact of the proposed framework on studying future
observability and data injection attacks.
II. S TATE E STIMATION AND O BSERVABILITY
We next provide an overview of state estimation and of the
algebraic and topological concepts of observability in power
systems. This overview provides background material which is
useful for the analysis that follows.
A. State Estimation Process
Consider a power system state estimation process which
uses various measurements collected from across the system to
estimate the voltage magnitudes and phase angles at every bus
in the system, known as the system states [15]. Let z ∈ Rm (m
being the number of measurements) be the vector of collected
measurements, which includes power flow levels (real and reactive) over transmission lines, power (real and reactive) injected
in or withdrawn from certain buses, as well as bus voltage
magnitudes. In addition, let x ∈ Rn be the vector of system
states. The relationship between the measurements and the states
directly follows from the linearized power flow equations [15]:
z = Hx + e,
(1)
where H ∈ Rm×n is the measurement Jacobian matrix and
e ∈ Rm is the vector of random errors that typically follows
a Gaussian distribution, N (0, R), where R is positive definite.
Here m ≥ n, that is the dimension of x cannot be larger than the
dimension of the measurement vector, z. Further, we assume that
H is a full-rank matrix. Using a maximum-likelihood estimator
– a weighted least squares estimator (WLS) for a Gaussian error
vector e – an estimate of the states, x̂, will be:
x̂ = (H T R−1 H)−1 H T R−1 z.
(2)
This estimate of all the states provides visibility of the steadystate operating conditions of the system, based on which various
operational decisions are performed [15].
B. Power System Observability
The observability of the power system consists of the ability
to uniquely determine its states based on the collected set of
measurements [15]. Observability, hence, requires the collected
measurements to provide enough independent equations to allow
the estimation of the state vector, x. In this respect, the power
system is observable1 if and only if the measurement matrix H is
of full column rank [15], which was our initial assumption. This
is known as algebraic observability. Due to the P − θ, Q − V
decoupling2 in power systems [16], the observability analysis
can be decoupled by separately studying the observability of
voltage phase angles, using real power measurements, and the
observability of voltage magnitudes, based on reactive power
measurements. Since the two analyses are identical, we focus
here on phase angle observability. To this end, we consider z ∈
Rm to be a vector of real power measurements (bus injections
and line flows), and the state vector x ∈ [−π, π]n to be the
vector of voltage phase angles (in radians). Here, n = N − 1
1 Otherwise, when this observability condition is not met, the power system is
dubbed unobservable.
2 P denotes real power, θ denotes voltage phase angles, Q denotes reactive
power, and V denotes voltage magnitudes.
for a power system with N buses given that the phase angle of
the reference bus is fixed and is taken to be the reference with
respect to which all other phase angles are calculated [15].
An alternative measure of observability, which is equivalent
to algebraic observability, is proposed in [30] and uses graphtheoretic techniques to introduce the concept of topological
observability. In this regard, let the power system 1-line diagram
be represented as a graph G(N , L) in which the set of vertices N ,
|N | = N , represents the set of buses of the power system while
the set of branches L, |L| = L, represents the set of lines. One
key result that was shown in [30] and that will be of relevance
to our work is the following:
Remark 1: A power system is observable if and only if the
set of measurements can be assigned to the edges of the power
system graph, following a set of assignment rules, in a way to
form a spanning tree over this graph.
In this respect, let M be the set of measurements and let
f (.): M → L be an assignment function defined as follows.
Definition 1: f (.): M → L is a measurement assignment
function which assigns measurements in M to lines in L
following a set of assignment rules defined as [30]:
1) If l1 , l2 ∈ L and l1 6= l2 , then f −1 (l1 ) 6= f −1 (l2 ). In other
words, a measurement cannot be simultaneously assigned
to two different lines.
2) If m is a measurement over a transmission line l, then m
can only be assigned to l.
3) If m is an injection measurement over bus η ∈ N , then
m can only be assigned to an unmeasured line l that is
incident to η.
If such a measurement assignment that yields a spanning tree
over the power network G can be found, the power system will
be observable (and vice versa). Fig. 1 shows an example of measurement assignments over the IEEE 14-bus system. This figure
shows the tree branches (marked in solid red lines) to which
measurement where assigned as part of the measurement assignment function. The measurements that were assigned to each
one of these branches are identified using dashed arrow lines
originating from the assigned measurement and pointing to the
line to which this measurement is assigned. This tree is formed
of branches {1, 2, 4, 6, 8, 9, 10, 12, 13, 15, 16, 17, 19} and
spans the whole vertex set N of the power system graph G, and
hence, is a spanning tree. As a result, since this measurement
assignment yields a spanning tree, then the available set of
measurements renders the system observable.
Various algorithms of low complexity have been proposed to
find and build such a spanning tree [30]–[32]. In this regard, the
work in [30] proposes an algorithm to find a spanning tree over
G, which will be used in some of the derivations in the following
sections. This algorithm starts by processing flow measurements
by assigning each flow measurement to its corresponding branch
to form disjoint tree components. Then, injection measurements
are assigned to lines in a way to connect these tree components to
form one spanning tree. Here, we highlight one type of injection
measurements, namely, boundary injections, which will play a
crucial role in our derivations.
Definition 2: A boundary injection is an injection measurement over a bus incident to lines whose flow is measured and
13
19
20
12
14
17
12
11
13
16
18
9
6
10
10
14
11
15
8
9
7
1
2
5
1
3
7
4
4
6
5
2
8
3
IEEE 14-Bus System
Bus injection
measurement
Tree branch
Edge not part
of the tree
Line flow
Measurement
Line number
measurement
assignment
Fig. 1. IEEE 14-bus system with measurement assignment.
Bus number
lines whose flow is not measured [30].
Boundary injections play a major role in connecting these
tree components. Indeed, for a bus which is not incident to
a measured line to be connected to the spanning tree, it has
to be reachable from a boundary injection through a series of
measurement assignments [30]. As such, boundary injections are
considered to be sources and unmeasured buses are considered
to be sinks which must be connected to these sources following
the set of measurement assignment rules.
We next build on the foundation of topological observability to
present a graph-theoretic framework for modeling and studying
the security of the smart grid facing observability attacks and
SDIAs. This framework is based on our proposed concepts of
critical sets and observability sets, which we define and derive
in the next section.
III. O BSERVABILITY ATTACKS
The sustainable and efficient operation of the power system
requires an accurate observability of all its states [15]. Security
attacks that target this observability can cause a limited (or
partial) monitoring ability for the operator over the power system
which can lead to incorrect operational decisions. Hence, studying and modeling attacks which can target the full observability
of the system is indispensable for the sustainable operation
of the grid. In this respect, we define a cyber-physical attack,
dubbed observability attack, that consists of launching a denialof-service (DoS) attack against a set of measurements to make
the system unobservable. We next study this type of attacks by
introducing and characterizing what we define as critical sets and
observability sets and prove that the well-studied stealthy data
injection attack is a subset of our defined observability attacks.
This latter finding will provide us with a unified set of tools to
solve various widely-studied SDIA problems.
A. Critical Sets
Understanding and modeling observability attacks requires an
in-depth understanding of the effect of the loss of any bundle of
measurements on the observability of the system. In this regard,
we next introduce a structured method for identifying, for each
measurement m, the set of additional measurements including m
(denoted as the critical set of m), which when removed renders
the system unobservable. To this end, we first characterize
the set of potential measurements to be investigated, for each
measurement m, and then provide the necessary discussion and
introduce the underlying method for characterizing the critical
set of m. In this process, we prove that characterizing such a
critical set is at most as complex as solving a maximum matching
problem over a bipartite graph. Then, a detailed algorithm is
introduced to provide a step-by-step method for characterizing
such critical sets.
Let MC ⊆ M be the set of measurements which are part
of the assignment function, i.e., the measurements assigned to
form a spanning tree over the power network. We refer to
measurements not part of MC as unassigned measurements.
Such unassigned measurements are, hence, by definition redundant measurements. We consider that the system is originally
observable. Hence, such a spanning tree and its corresponding
set of assigned measurements exist. We also let T (N , B) be
the spanning tree whose set of branches are captured by the
set B ⊆ L. B, in essence, represents the set of lines to which
measurements were assigned as part of the assignment function
f (.). For example, as previously mentioned, in Fig. 1, the
branches of the spanning tree are represented in solid red lines.
Consider an assigned measurement m ∈ MC . Since m is
assigned, its removal will split the tree T into two spanning
trees T1m (N1m , B1m ) and T2m (N2m , B2m ) spanning subgraphs G1m
and G2m , respectively, such that N = N1m ∪ N2m and B =
B1m ∪ B2m ∪ {f (m)}. Fig. 2 provides an illustrative example
of the two spanning trees created by the deletion of the flow
measurement over line 2 (we denote this flow measurement by
F2 ). Fig. 2 represents the same system shown in Fig. 1 and will
be used throughout this work to provide a practical example of
the defined concepts and analytical derivations.
To investigate observability, we define a set of measurements
for each measurement m ∈ MC , which we refer as the
critical set of m and we denote by C m , as follows:
Definition 3: For a measurement m ∈ MC , the critical set
of m, denoted as C m ⊆ M, is a set of measurements which can
be used to reconnect T1m and T2m when m is deleted.
Consequently, when m is removed, any m0 ∈ C m can be
used to reconnect T1m and T2m . For convenience, we consider
m to be part of its own critical set C m . Let Lm be the set of
lines connecting a bus in T1m to a bus in T2m . In addition, let
m
m
⊆ N2m be the set of nodes in, respectively,
N1,2
⊆ N1m and N2,1
m
m
N1 and N2 which are connected to a node in, respectively, N2m
and N1m . An example of these notations is provided in Fig. 2.
Lm is, hence, formally defined as:
m
m
Lm = {l ∈ L | l = (η1 , η2 ), η1 ∈ N1,2
, η2 ∈ N2,1
}.
(3)
We next introduce the set of rules that should be followed to
build the critical set of a certain measurement, following which,
we provide a structured algorithm for building such critical sets.
Based on the measurement assignment rules described in
Section II-B, a necessary condition for a measurement m0
13
19
Redundant measurements
part of
20
12
14
17
12
11
13
Unassigned injection
measurements over
buses in
16
18
9
6
10
14
11
8
1
1
5
9
7
7
3
2
Category 1
15
10
8
4
4
Category 2
Flow measurements
over lines in
Assigned injection
measurements over
buses in
6
Category 3
2
5
3
Measurements in
IEEE 14-Bus System
Unassigned
Boundary Injection
Further investigation needed
to determine whether a
measurement in this set
is part of
Branches
Fig. 3. The three categories of measurements in Mm .
Branches
As a result, since the measurements in Mm
F and the unasm
m
signed injection measurements over buses in N1,2
∪ N2,1
(which
m
are the first two categories of measurements in M , as shown in
Fig. 3) are redundant, they are part of C m . Now, when m0 ∈ Mm
is assigned as part of the original assignment function, i.e.,
m0 ∈ Mm ∩ MC (which corresponds to the third category
of measurements in Mm , indicated in Fig. 3), then additional
investigation is needed to determine whether m0 is redundant,
and hence, whether it can be considered in C m .
In this regard, consider that m0 ∈ Mm is assigned to a
line l0 , that is f (m0 ) = l0 ∈ B. If m0 is to be reassigned to
a line l ∈ Lm , T1m and T2m will be reconnected, but since
m0 was originally assigned as part of the original tree, another
portion of the tree gets disconnected by this reassignment of
m0 to l instead of l0 . Hence, m0 can be part of C m if another
measurement can be used to reconnect the subgraph which
was disconnected by the reassignment of m0 from l0 to l. For
example, consider the injection measurement over bus 13 in
Fig. 2, which we denote by I13 . I13 has been assigned to line 12
as part of the original spanning tree. Hence, if I13 is assigned
to line 20 to reconnect T1F2 and T2F2 after measurement F2 is
removed, it cannot be assigned to line 12 anymore which will
split T1F2 into two subtrees, one formed by buses {12, 13} and
line 19 and the other subtree composed of buses {6, 10, 1, 5}
F2
and lines {13, 10, 1}. We denote these two subtrees as T1,1
F2
and T1,2 , respectively. In this respect, if another measurement
F2
F2
can replace I13 in reconnecting T1,1
and T1,2
, then I13 can
be assigned to line 20 and, hence, should be part of C F2 .
To this end, consider the injection measurement over bus 12,
denoted by I12 , which was not part of the original spanning
tree assignment. I12 can be assigned to line 11 to reconnect
F2
F2
T1,1
and T1,2
in case I13 is reassigned to line 20 instead of
line 12. Hence, I13 is indeed redundant, resulting in I13 ∈ C F2 .
To generalize the analysis in this example, we next provide a
general discussion of measurements in Mm ∩MC (i.e., the third
category of measurements in Mm , shown in Fig. 3) which allows
determining whether a measurement in this set is part of C m .
More generally, consider m0 ∈ Mm ∩ MC to be a measurement assigned to a branch l0 = f (m0 ) in T1m , and let Mm
1 be the
Fig. 2. Critical set of the flow measurement over line 2, F2 , of the IEEE 14-bus
system.
to be in C m is for it to be either a line flow measurement
over a line l ∈ Lm or an injection measurement over a bus
m
m
∪ N2,1
}. We denote this set of measurements, for a
η ∈ {N1,2
measurement m, by Mm . Here, for convenience3 , we consider
Mm not to include measurement m. The measurements in Mm
are the only measurements which can be assigned to a line in
Lm , and hence, are the only measurements with the potential
of reconnecting T1m and T2m . An example of the set Mm for
m := F2 is shown in Fig. 2.
The measurements in Mm can be split into three different
categories: 1) flow measurements over lines in Lm , which we
4
denote as Mm
F , 2) unassigned injection measurements over
m
m
buses in N1,2 ∪ N2,1
, and 3) assigned injection measurements
m
m
over buses in N1,2 ∪ N2,1
, i.e., measurements in Mm ∩ MC . A
representation of this partition is shown in Fig. 3
In this regard, not all measurements in Mm are necessarily
in C m . In fact, for m0 ∈ Mm to be in C m , it must be redundant.
Namely, m0 must be assignable to a line in Lm to reconnect
T1m and T2m , without causing any disconnections within either
T1m or T2m . In this respect, if m0 ∈ Mm is unassigned, i.e
m0 ∈
/ MC , m0 would be part of C m . For example, consider the
injection measurement over bus 4 in Fig. 2, which we denote by
I4 . Measurement I4 is in MF2 and is a redundant measurement
since it was not assigned to any line as part of the original tree
T . Hence, when F2 is removed, I4 can be assigned to line 7 to
reconnect T1F2 and T2F2 without causing any disconnection in
T2F2 , in which it is located. Hence, I4 ∈ C F2 .
3 Given that m is always considered to be part of its critical set C m , m is
always added to C m after investigating the measurements in Mm .
4 Such measurements are unassigned measurements. In fact, if m is a line
measurement, lines in Lm would form a loop with f (m) and hence cannot be
part of the original spanning tree. Moreover, if m is an injection measurement,
Mm
F would be an empty set since, otherwise, based on the spanning tree building
method described in Section II-B and originally presented in [30], one of the
m
measurements in Mm
F would have been assigned to a line in L , and m would
not have been part of MC . As a result, measurements in Mm
F are redundant.
set of measurements in G1m . Reassigning m0 to l ∈ Lm instead
of l0 , to reconnect T1m and T2m , will split T1m into two subtrees
m
m
m
T1,1
and T1,2
. These trees, respectively, span subgraphs G1,1
and
m
m
m
m
G1,2 . Let M1,1 and M1,2 be the sets of measurements in G1,1
m
0
and G1,2 . m can be reassigned to l only if some measurement
m
m
in Mm
1 can reconnect T1,1 and T1,2 . Hence, this corresponds to
finding a measurement assignment that connects the two subtrees
m
m
T1,1
and T1,2
. As discussed in Section II-B, two subtrees can
be interconnected by using a measurement assignment if the
processing of an unassigned boundary injection in one of them
reaches a node in the other. We denote such unassigned boundary
injections as backup boundary injections, which we formally
define as follows:
m
m
Definition 4: A measurement m00 ∈ Mm
1 \ {M1 ∩ M } is a
0
00
backup boundary injection for a measurement m , if m can be
m
m
used to reconnect T1,1
and T1,2
generated by the reassignment
0
m
C
of m ∈ M ∩ M to a line l ∈ Lm instead of its original line
assignment f (m0 ) = l0 . The set of all such backup boundary
injections for this measurement m0 ∈ Mm ∩ MC is referred to
as the backup boundary injection set of m0 and is denoted by
m
Ib−m
0.
Since the algorithm in [30] is based on connecting subtrees –
to build a full spanning tree – by starting from an unassigned
boundary injection in a certain subtree (as a source) to reach
a node in another subtree (as a sink), this algorithm can be
employed to locate a backup boundary injection for a measurement m0 ∈ Mm ∩ MC . To this end, to find a backup boundary
injection of m0 , we run the algorithm in [30] by starting from
m
m
an unassigned boundary injection in either T1,1
or T1,2
and
m
m
checking whether the algorithm reaches a bus in T1,2 or T1,1
,
respectively. As such, using the spanning tree building algorithm
provided in [30, Fig. 1], one can identify the backup boundary
injections for each injection measurement in m0 ∈ Mm ∩ MC .
Here, we note that a backup boundary injection cannot be an
injection measurement in Mm , since if m00 ∈ Mm and m00 is
unassigned, m00 will itself be part of C m , as previously discussed.
For example, consider the injection measurements on buses 2 and
4 of Fig. 2, denoted by I2 and I4 , respectively. I2 is an assigned
injection measurement in MF2 . I2 can be assigned to line 3
instead of line 4 to reconnect T1F2 and T2F2 . However, this will
split bus 2 from the rest of T2F2 . I4 is an unassigned boundary
injection in G2F2 and can reconnect bus 2 to T2F2 by assigning
F2
I4 to line 4. However, I4 is itself an injection at a bus in N2,1
.
F2
Hence, I4 is an unassigned injection in M , and is as a result
part of C F2 . Thus, it cannot be considered a backup boundary
injection for I2 .
Therefore, an assigned injection measurement m0 ∈ Mm ∩
MC is a redundant measurement and is, as a result, part
m
of C m if it has a nonempty boundary injection set Ib−m
0.
However, a boundary injection may be part of multiple backup
boundary injection sets. In this regard, based on the measurement
assignment rules, an injection measurement can be assigned to
only one line at a time. In relation to backup boundary injections,
an unassigned boundary injection can act as a backup boundary
injection for only one measurement in Mm ∩ MC , at a time.
Hence, if two measurements m1 and m2 in Mm ∩ MC have
only one and the same backup boundary injection, only one
Injection Backup Boundary
Measurement
Injection
General Representation
Unmatched
Matched
Fig. 4. Maximum matching over the injection measurements - backup boundary
injections bipartite graph.
of them can be in C m , concurrently. As a result, due to this
one-to-one assignment requirement between backup boundary
injections and injection measurements in Mm ∩MC , finding this
assignment can be performed by solving a maximum matching
problem over a bipartite graph5 , as the one shown in Fig. 4.
We refer to this graph as the injection measurements - backup
boundary injections bipartite graph.
In this bipartite graph, the left-side nodes denote the injection
measurements in Mm ∩ MC and right-side nodes
the
S denote
m
union of their backup boundary injections,
Ib−m
0,
m0 ∈Mm ∩MC
in which each node represents one backup boundary injection.
In this bipartite graph, an edge exists between a node m0 ∈
Mm ∩ MC , on the left-side of the graph, and a boundary
m
injection m00 , on the right-side of the graph, if m00 ∈ Ib−m
0.
Here, we note that a boundary injection can be simultaneously
part of different backup boundary injection sets. Hence, finding
the injection measurements in Mm ∩ MC which are part of
the critical set of m, C m , is equivalent to solving a maximum
matching problem over this bipartite graph, which is a problem
whose exact solution can be obtained efficiently in polynomial
time6 . As a result, the matched left-side nodes in this maximum
matching over the injection measurements - backup boundary
injections bipartite graph are the injection measurements in
Mm ∩ MC which will be part of the critical set of m, C m .
Based on these introduced rules for building critical sets, Algorithm 1 provides a structured step-by-step method for building
the critical set, C m , for each measurement m ∈ MC .
B. Example and Case Analysis
As an example of characterizing the critical sets of the various
measurements in a power system, we consider the IEEE 14bus system in Fig. 1. In this example, we refer to an injection
measurement over bus k as Ik and a flow measurement over line
k as Fk . We first consider the measurement over line 2, F2 , for
which we find the critical set C F2 using Algorithm 1.
From Fig. 2, we can see that removing F2 will result in splitting the original spanning tree into two trees, T1F2 (N1F2 , B1F2 )
5 A matching over a graph is a subset of edges sharing no vertices. A maximum
matching is a matching having the maximum possible number of edges [33].
6 The solution of a maximum matching problem over a bipartite graph can be
efficiently obtained in polynomial time by transforming the matching problem
into a max-flow problem, which can be solved in polynomial time using various
known algorithms such as Ford-Fulkerson [33].
Algorithm 1 Critical sets step-by-step procedure
Input: Power system 1-line diagram G(N , L), measurement set
M, spanning tree T (N , B), set of assigned measurements
MC , assignment function f (.): M → L
Output: Critical set C m for all measurements m ∈ MC
1: for m ∈ MC do
m
m
2:
Characterize T1m , T2m , N1,2
, N2,1
, Lm , Mm
m
3:
Initialize C
4:
Initialize Mm
test
5:
Add m to C m
6:
for m0 ∈ Mm do
7:
if m0 is a flow measurement then
8:
Add m0 to C m
9:
end if
10:
if m0 is an injection measurement then
11:
if m0 ∈
/ MC then
12:
Add m0 to C m
13:
end if
14:
if m0 ∈ MC then
15:
Characterize its backup boundary injection set
m
Ib−m
0
16:
Add m0 to Mm
test
17:
end if
18:
end if
19:
end for
20:
Solve maximum matching over the injection measurements - backup boundary injections bipartite graph
21:
for m0 ∈ Mm
test do
22:
if m0 is a matched node as part of the maximum
matching then
23:
Add m0 to C m
24:
end if
25:
end for
26: end for
27: return Critical set C m for all measurements m ∈ MC
and T2F2 (N2F2 , B2F2 ), such that N1F2 = {1, 5, 6, 10, 12, 13}
and N2F2 = {2, 3, 4, 7, 8, 9, 11, 14} are the sets of nodes of
the two trees and B1F2 = {1, 10, 13, 12, 19} and B2F2 =
{4, 6, 8, 15, 9, 16, 17} are their sets of branches. In addiF2
F2
tion, N1,2
= {1, 5, 10, 13}, N2,1
= {2, 4, 11, 14}, LF2 =
F2
{2, 3, 7, 18, 20}, and M
= {F2 , I4 , I11 , I2 , I1 , I5 , I13 }.
Now, for characterizing the critical set of F2 , we explore the
set MF2 .
The first measurement in MF2 is F2 . F2 is a flow measurement7 . Hence, F2 ∈ C F2 .
The second and third measurements in MF2 are I4 and I11 . I4
and I11 are unassigned injection measurements, i.e. I4 ∈
/ MC
C
F2
and I11 ∈
/ M . Hence, {I4 , I11 } ⊆ C . Indeed, I4 can be
assigned to line 7 to reconnect T1F2 and T2F2 , while I11 can be
assigned to line 18 for that purpose.
I2 is the fourth measurement in MF2 and the last remaining
m
injection measurement on N2,1
to be explored. I2 is an assigned
measurement, originally assigned to line 4 as part of the spanning
tree T . Since I2 ∈ MC , assigning I2 to lines 2 or 3 to reconnect
7F
2
is the only flow measurement in MF2 .
TABLE I
C RITICAL SETS OF THE MEASUREMENTS IN MC .
Measurement (m ∈ MC )
F2
F8
F9
F15
I1
I2
I3
I5
I6
I9
I13
F17
F19
Critical Set (C m )
{F2 , I4 , I11 , I13 }
{F8 , I4 , I7 , I9 }
{F9 , I4 , I7 , I11 , I13 }
{F15 , I7 }
{I1 , I4 , I11 , I13 }
{I2 , I4 , I11 , I13 }
{I3 , I2 , I4 }
{I5 , I11 , I13 }
{I6 , I11 }
{I9 , I11 }
{I6 , I12 , I13 }
{F17 , I9 , I13 }
{F19 , I6 , I12 }
T1F2 and T2F2 will disconnect bus 2 from the rest of T2F2 . Hence,
we next characterize the backup boundary injection set of I2 ,
F2
i.e. Ib−I
. The only unassigned boundary injection in G2F2 that
2
is not part of MF2 is I7 . However, using the algorithm in [30,
Fig. 1], we can observe that starting from I7 , the algorithm does
not reach bus 2. Hence, bus 2 cannot be reconnected to the rest
of T2F2 using any unassigned boundary injections over buses in
F2
G2F2 . Hence, Ib−I
= ∅, and as a result F2 ∈
/ C F2 .
2
Similarly, exploring I1 and I5 – the fifth and sixth measurements in MF2 – which are both assigned measurements,
i.e. {I1 , I5 } ⊆ MC , shows that they both have empty backup
F2
F2
boundary injection sets8 , i.e. Ib−I
= ∅ and Ib−I
= ∅. Hence,
1
5
F2
neither I1 nor I5 are part of C .
The only remaining measurement in MF2 is I13 . I13 is an
assigned measurement, I13 ∈ MC . As previously discussed in
Section III-A, when I13 is reassigned to line 20 to reconnect T1F2
and T2F2 , the subtree containing buses 12 and 13 and line 19 gets
disconnected from the rest of T1F2 . Hence, we next characterize
F2
the backup boundary injection set of I13 , i.e. Ib−I
. To this
13
end, I12 , the only unassigned boundary injection measurement
in G1F2 , can be assigned to line 11 to reconnect the two subtrees,
and is the only boundary injection which can do so. Hence,
F2
Ib−I
= {I12 }.
13
As a result, the injection measurements - backup boundary
injections bipartite graph is composed of only I13 on the leftF2
side connected to Ib−I
= {I12 } on the right-side. Hence, I13
13
is matched to the backup boundary injection I12 . As a result,
I13 ∈ C F2 .
The processing of MF2 is thus complete, resulting in C F2 =
{F2 , I4 , I11 , I13 }.
Similarly, Algorithm 1 can be carried out to characterize the
critical sets of all of the measurements in MC in the IEEE 14bus system in Fig.1. The results are listed in Table I.
We next discuss the value of critical sets with regard to understanding and analyzing observability attacks. We also introduce
the concept of observability sets, a generalization of critical sets,
which provides a holistic modeling of observability attacks.
Notation: We use the following notation in the derivations
0
that ensue. For the Jacobian matrix H, we let H (−K)+(K )
8 If I or I are to be reassigned to lines 2 or 3, respectively, to reconnect
1
5
T1F2 and T2F2 , each of these reassignments will split T1F2 into two subtrees
which cannot be reconnected using the unassigned boundary injection I12 , as
can be shown by a run of the algorithm in [30, Fig. 1]. Here, we note that I12
is the only unassigned boundary injection in G1F2 .
correspond to H but with the removal of the rows corresponding
to measurements in K and the addition of rows corresponding
to measurements in K0 .
would have been part of C m because m0 can reconnect T1m and
T2m . Hence, replacing m by m0 will not affect the connectivity
0
of the spanning tree and, hence, rank(H) = rank(H (−m)+(m ) ).
C. Observability Sets
Theorem 2: For m
∈ MC , removing its critical set, C m ,
(−C m )
results in rank(H
) = rank(H) − 1.
Proof: Since the system is originally fully observable,
m
rank(H) = N − 1. Now, let Mm
1 and M2 be the meam
m
surement
sets in subgraphs
G1 and G2 , respectively, and let
(−C m )
(−C m )
H1
and H 2
be the Jacobian matrices of G1m and G2m ,
m
m
respectively, composed of measurements in Mm
1 \ {M1 ∩ C }
m
m
m
m
and Mm
\
{M
∩
C
}.
Since
T
and
T
respectively
span
2
2
1
2
(−C m )
G1m and G2m
,
this
implies
that
rank(H
) = N1 − 1 and
1
(−C m )
m
rank(H 2
) = N2 − 1. In addition, let m0 ∈ Mm
1 \ {M1 ∩
m
m
C } be an injection measurement over a bus in N1,2 . Since
m
m
0
m0 ∈ Mm
1 \ {M1 ∩ C }, then m is assigned to a certain
0
0
m
branch b = f (m ) ∈ T1 ; otherwise, m0 would have also been
in C m . By Lemma 1, m0 can be replaced by a hypothetical line
(−C m )
flow measurement over b0 without affecting the rank of H 1
.
m 0
m
(−C )
(−C )
As such, let H 1
be the same as H 1
but replacing
m
any row corresponding to an injection measurement in N1,2
by
its corresponding hypothetical line flow measurement.
The
same
(−C m )0
(−C m )
can be done to form Jacobian matrix H 2
from H 2
.
m 0
m
(−C )
(−C )
By Lemma 1, rank(H 1
)=rank(H 1
) = N1 − 1 and
(−C m )0
(−C m )
rank(H 2
)=rank(H 2
) = N2 − 1.
(−C m )
Now, let us return to H
. By rearranging its elements
m
m
to include first the measurements in Mm
1 \ {M1 ∩ C } then
(−C m )
m
m
m
the elements"of M2 \ {M
can be written as
# 2 ∩ C }, H
(−C m )
H1
(−C m )
. In this respect,
H
=
(−C m )
H2
#
"
H (−C m )
m
1 m
rank(H (−C ) ) = rank
(−C )
H2
"
#
H (−C m )0
0
1
= rank
m 0
(−C )
0
H2
Next, in Theorem 1, we show that the derived critical sets are
indispensable for modeling observability attacks.
Theorem 1: For m ∈ MC , removing its critical set, C m ,
renders the system unobservable.
Proof: By topological observability, we know that a system
is observable if and only if a spanning tree could be formed
using an assignment function. When m is removed, the original
spanning tree T is split into two disjoint trees T1m and T2m ,
spanning subgraphs G1m and G2m , respectively. By definition
of C m , the only way to reconnect T1m and T2m is to use
measurements in C m . Hence, if all measurements in C m are
removed, then T1m and T2m cannot be connected using an
assignment function, which implies that a spanning tree cannot
be formed, implying that the system is not observable. As such,
removing a critical set renders the power system unobservable.
Based on Theorem 1, the critical measurements9 of a power
system can be characterized using critical measurement sets, as
shown in the following corollary.
Corollary 1: m is a critical measurement if and only if its
critical set is C m = {m}.
Proof: By definition, if m is a critical measurement, removing it will render the system unobservable. Hence, if the critical
set of m is such that C m ⊃ m, then removing m would not affect
the observabilty of the system since any other measurement
m0 ∈ C m \ {m} can be used to replace m and reconnect the
tree. As such, C m ⊃ {m} ⇒ m is not a critical measurement,
which proves the contrapositive: m is critical ⇒ m is the only
element in its critical set, i.e. {m} = C m . Conversely, if m
is the only element in its critical set, its removal constitutes
removing a complete critical set, which by Theorem 1 renders the
system unobservable. As a result, C m = {m} ⇒ m is a critical
measurement. Thus, m is critical if and only if C m = {m}.
In addition, removing a full critical set decreases the rank of
the Jacobian matrix by 1. This is shown in Theorem 2, which
will be proven next. However, we first present the following preliminary lemma, which is essential for the proof of Theorem 2.
Lemma 1: Let m ∈ M be an injection measurement over a
bus η that is assigned to a line l, f (m) = l. Then, replacing
m by a hypothetical line flow measurement m0 over line l
will not affect
the rank of matrix H. In other words, let
0
H (−m)+(m ) be the the Jacobian matrix with the removal of
the row corresponding to measurement m and the addition of
the row corresponding to the 0hypothetical measurement m0 , then
rank(H) = rank(H (−m)+(m ) ).
Proof: Since m is assigned, i.e. is part of the original
spanning tree measurement assignment, removing it will split
the original spanning tree into two subtrees T1m and T2m . By
the definition of the critical set C m , any measurement in C m
could replace m to reconnect T1m and T2m . If m0 existed, it
9 In power systems, a critical measurement is a single measurement which
when removed renders the system unobservable [15].
= (N1 − 1) + (N2 − 1)
= N − 2 = rank(H) − 1.
Therefore, Theorem 2 shows the effect of the removal of a
single critical set on the rank of the Jacobian matrix. Theorem 1
and Theorem 2 provide a necessary condition for observability
of the power system under observability attacks. In fact, the
contrapositive of Theorem 1 states that if a power system is
fully observable, then the investigated observability attack (i.e.
the removal of measurements) did not result in removing a full
critical set. Next, we extend this concept to account for the
interconnection between multiple critical sets.
In fact, for a measurement m0 to be in the critical set of a
0
measurement m, i.e. m0 ∈ C m , the critical set of m0 , C m ,
0
must contain measurements other than m0 , i.e. C m ⊃ {m0 }.
Otherwise, m0 would not be redundant. For example, consider
injection measurements I6 and I9 . Removing I6 and I9 will
render the system unobservable – even though I6 and I9 do not
form a critical set – since C I6 = {I6 , I11 } and C I9 = {I9 , I11 }.
As such, if I6 is removed, I11 can be used to replace I6
since I11 ∈ C I6 . However, if I9 is also removed, even though
I11 ∈ C I9 , I11 cannot be used to replace I9 since I11 has
already been used as a replacement to I6 . Therefore, removing I9 and I6 does render the system unobservable. Indeed,
rank(H −(I6 )−(I9 ) ) = 12 < N − 1 = 13.
This concept can be extended to the interconnection between
multiple critical sets. For example, consider F2 , I1 , I2 , and I5
and their critical sets shown in Table I. We can see that F2 ,
I1 , and I2 have critical sets sharing measurements I4 , I11 , and
I13 . Hence, if F2 , I1 , and I2 are removed, I4 , I11 , and I13 are
assigned, one to each of these measurements, to preserve system
observability and, hence, cannot be used as part of further critical
sets in case further measurements are removed. Hence, since
C I5 = {I5 , I11 , I13 }, removing F2 , I1 , I2 and I5 will render
the system unobservable, even though {F2 , I1 , I2 , I5 } is not a
critical set.
In this respect, the concept of critical sets must be further
developed to yield a general graph-theoretic concept of observability attacks. This development is provided as follows. We
build a bipartite graph in which each left-side node represents
one of the critical sets of the power system, and each right-side
node represents one measurement of the system. An example
of this bipartite graph is shown in Fig. 5. In this respect, an
edge is drawn between a critical set C i and a measurement j
if j ∈ C i . We refer to this bipartite graph as the critical sets system measurements bipartite graph. Based on this formulation,
a general concept of observability is established in Theorem 3.
Theorem 3: If the system is observable, then the maximum
matching over the critical sets - system measurements bipartite
graph includes all critical sets.
Proof: We prove this theorem by proving its contrapositive
which is the following: if the maximum matching does not
include all critical sets, then the system is not observable.
The contrapositive can be proven as follows. If one critical set
is not matched to any measurements, then this critical set cannot
be used to connect two subgraphs of the system. Since these
two subgraphs can only be connected by this critical set, then
there is no measurement assignment which will connect these
two subgraphs. As a result, a spanning tree cannot be formed,
implying that the system is not observable.
This proves the contrapositive of this theorem and, hence,
proves the theorem.
Theorem 3 can be used to fully characterize observability
attacks as follows. An observability attack is one in which
measurements are removed (i.e. nodes from the right-side of
the critical sets - system measurements bipartite graph) such
that the maximum matching over the bipartite graph does not
include all critical sets (i.e. nodes on the left-side of the critical
sets - system measurements bipartite graph), which renders
the system unobservable. This, as a result, provides a general
analytical characterization of observability attacks and enables
an analytical prediction of the effect of the removal of a subset
of measurements on the observability of the system. For example,
such characterization allows analytical derivation of various
security indices related to observability attacks such as finding
the observability attack of lowest cardinality, or finding the
minimal set of measurements to remove in addition to a certain
measurement to make the system unobservable. This, as a result,
provides analytical tools which are necessary to further assess
the vulnerability of a system against observability attacks as
well as derive defense strategies to thwart such attacks. Indeed,
in what follows, we focus on stealthy data injection attacks –
proving that they are a subset of observability attacks – and show
how our provided analytical characterization of observability
attacks enables solutions of various widely-studied stealthy data
injection attack problems. To this end, we introduce sets of
measurements, dubbed observability sets, as follows, which are
valuable for the analysis of data injection attacks which ensues.
Definition 5: An observability set S ∈ M, is a set of
measurements such that strictly removing S leads the maximum
matching over the critical sets - system measurements bipartite
graph not to include a certain critical set.
The term “strictly” in this definition reflects that adding any
measurement s ∈ S, which was removed, back to the right-side
of the bipartite graph will result in reincluding the previously
unmatched critical set in the maximum matching. A union of
observability sets is, then, defined to be a set of measurements
composed of a number of observability sets such that, when each
of these sets is successively removed, each such removal leads
to excluding one additional critical set from being part of the
maximum matching over the critical sets - system measurements
bipartite graph. Adding back any of the removed measurements
to the right-side of the bipartite graph will result in reincluding
one of the unmatched critical sets in the maximum matching.
These observability sets play a crucial role in characterizing
stealthy data injection attacks, as will be shown next.
We next introduce stealthy data injection attacks and prove that
they are a variant of our introduced observability attacks. This
enables further studying and solving various problems related to
SDIAs using our developed analytical tools.
IV. S TEALTHY DATA I NJECTION ATTACKS
A. Stealthy Data Injection Attacks
Recalling the measurement-state equation in (1), data injection attacks aim at replacing the measurement vector, z by a
manipulated measurement vector z a = z + a, where a ∈ Rm is
the attack vector, resulting in a new state estimate x̂a . However,
typically, the state estimation process is run in conjunction with
what is known as a bad data detector and identifier (BDD).
The BDD aims at detecting and identifying the presence of
outliers in the collected data set, so that such outliers can be
removed preventing them from affecting the estimation outcome.
Such BDDs rely on the statistical analysis of what is known as
measurement residuals, r, defined as [15]:
ẑ = H x̂ = Sz, r = z − ẑ = (I n − S)z = W z,
(4)
where S = H(H T R−1 H)−1 H T R−1 and W = I n − S.
A statistical analysis on the residuals enables analysis of the
magnitudes of the errors associated with each measurement, and
hence, allows the identification of outliers [15]. Regarding data
injection attacks, when data is added to certain measurements,
the adversary aims at keeping the residuals unchanged, so that the
attack cannot be detected by the BDD. Indeed, as shown in [5],
an attack vector that falls in the column-space of the Jacobian
matrix H, i.e. a = Hc, cannot be detected by residual statistical
analysis. Indeed, for a = Hc,
r a = W (z + a) = r + W a
= r + [I n − H(H T R−1 H)−1 H T R−1 ]Hc
= r + Hc − Hc = r.
(5)
As such, given the weighted least squares state estimation
equation in (2), the attack vector a = Hc generates an arbitrary
new state estimate x̂a = x̂ + c by choosing the constant vector
c without inducing any changes to the residual vector, as shown
in (5). Such DIAs are, hence, stealthy and are referred to as
stealthy DIAs. The ability of SDIAs to stealthily manipulate
the state estimates poses various challenges to the operation
of the grid. Hence, understanding and modeling such attacks is
indispensable to the secure and sustainable operation of power
systems.
To this end, we next introduce a holistic graph-theoretic
modeling of SDIAs that is based on the graph-theoretic modeling
of observability attacks introduced in Section III.
B. Graph-Theoretic Modeling of SDIAs
The observability attacks and observability sets introduced in
Section III provide the basis for a graph-theoretic interpretation
of SDIAs as will be shown in Theorem 4. However, before
introducing and proving Theorem 4, we introduce a preliminary
lemma which will be used in the proof of Theorem 4.
Lemma 2: If a DIA is stealthy (i.e. a = Hc), then removing
the attacked measurements renders the system unobservable.
Proof: Since the attack vector a is stealthy, then a = Hc.
Since H is of full rank, then the only solution to Hc = 0 is
c = 0. Hence, a = Hc has zero and nonzero elements for
c 6= 0. Now, if all of the rows of H corresponding to nonzero
elements of a are removed to form matrix H new , then, this results
in anew = H new c = 0 for c 6= 0. Hence, H new is not of full rank
and the power system whose Jacobian matrix is given by H new
is unobservable. Therefore, when the attack is stealthy, removing
the attacked measurements renders the system unobservable.
Here we note, that the result of Lemma 2, provides a one
directional relation stating that if a = Hc, i.e. the attack is
stealthy, then the removal of the nonzero elements of a, i.e. the
attacked measurements, causes the system to be unobservable.
However, the reverse direction does not always hold true. Indeed,
the reverse statement of Lemma 2 states that, if removing a
set of measurements renders the system unobservable, then this
guarantees that a stealthy DIA can be constructed which targets
all of these measurements, and only these measurements. We
next provide a counter example which proves that this reverse
statement does not hold true. In this regard, we consider
the
H0
Jacobian matrix H to be represented as follows: H =
.
H1
We let M0 and M1 represent the subset of measurements
corresponding to the rows of H 0 and H 1 , respectively. Consider
M0 to contain one critical measurement, i.e., one row of H 0 is
independent of all of the other rows of H. As such, removing the
subset of measurements M0 renders the system unobservable. In
addition, consider two measurements m0 ∈ M0 and m1 ∈ M1
such as m0 measures the power flow from bus i to bus j and m1
measures the power flow from bus j to bus i (i.e., m0 and m1 are
installed on the same transmission line but measure the flow in
two opposite directions). In this regard, let h0 and h1 correspond
to the rows of m0 and m1 in, respectively, H 0 and H 1 . Then, we
have10 h0 =−h1 .As aresult,one cannot find a stealthy attack
a0
H0
vector a =
=
c, in which all the elements of
a1
H1
a0 are nonzero and all the elements of a1 are zero, since if
h0 c 6= 0, then h1 c 6= 0, due to the fact that h0 = −h1 . This
implies that for the attack to target all the measurements in M0
and be stealthy, this attack must also target measurements in M1 .
Otherwise, this attack must be limited to a strict subset of M0
and may not target all the measurements in M0 . As a result, even
though removing the measurements in M0 renders the system
unobservable, one cannot necessarily construct a stealthy attack
vector that only targets all the measurements in M0 . Hence, this
provides a counter example of the reverse statement of Lemma 2
proving that this reverse statement does not always hold true.
Theorem 4: A DIA is stealthy if and only if the attacked
measurements constitute a union of observability sets.
Proof: We begin by proving that when the attacked measurements (i.e. nonzero elements of the attack vector a) constitute
a union of observability sets, then a is stealthy (i.e. a can be
represented as a = Hc). As shown in Theorem 3, when an
observability set (equivalently, a union of observability sets)
is removed, the system is unobservable. Hence, consider an
observability set S which has been removed. Let H (−S) be the
system’s Jacobian matrix without the measurements in S and let
C be the critical set which cannot be part of a maximum matching
over the critical sets - system measurements bipartite graph
when S is removed. Since the system is unobservable when
removing S, H (−S) y = 0 for a y 6= 0. However, the addition
of any measurement s ∈ S will reinclude C in the maximum
matching over the critical sets - system measurements bipartite
graph, and hence, reconnect the tree. As such, let H (−S)+(k)
correspond to H (−S) with the addition of a row corresponding to
a measurement k ∈ S. In this regard, since the system is rendered
observable, H (−S)+(k) is of full rank and H (−S)+(k) y will have
one nonzero element corresponding to the row of H (−S)+(k)
pertaining to the added measurement k. This procedure can be
repeated for all k ∈ S. As such, adding the rows corresponding
to S back to the Jacobian matrix results in b = Hy in which
only the elements of b corresponding to measurements in S are
nonzero. As a result, a = b is an attack vector in which only
the observability set S is attacked and is proven to be stealthy.
Now, we prove that, when an attack is stealthy, i.e. a = Hc,
then the nonzero elements of a correspond to a union of
observability sets. In this regard, from Lemma 2, we know that
removing the nonzero elements of a = Hc will render the
system unobservable, which implies that the nonzero elements
of a contain at least one observability set. Let S denote this
observability set, and let H (−S) be the system’s Jacobian matrix
without the measurements in S. Removing S will lead to two
subsystems each of which is fully observable (i.e. it will split the
10 Since P = −P , where P
ij
ji
ij and Pji are the real power flow from bus i
to bus j and from bus j to bus i, respectively, over the same transmission line.
spanning tree, T , into two subtrees each of which spans its own
subgraph). Let H 1 and H 2 be the Jacobian matrices of each of
these two subsystems (we denote these subsystems as subsystem
1 and subsystem 2) and let a1 and a2 correspond to the portions
of a (excluding the measurements of the previously removed
observability set) corresponding to the measurements in H 1 and
H 2 , respectively. In addition, let c1 and c2 correspond to the
portions of c pertaining to nodes in subsystem 1 and subsystem
2, respectively. Now, if ai for i ∈ {1, 2} has nonzero elements,
this implies that removing these elements will make subsystem
i unobservable, which implies that the nonzero elements of ai
contain an observability set. Following this same logic, removing
this observability set will subsequently split subsystem i into two
subsystems, each of which is observable. This process can be
continued recursively until no measurement m corresponding to
a nonzero element of a remains. Hence, this shows that when
a = Hc, then the nonzero elements of a correspond to a union
of observability sets.
This proves both directions of the theorem, and hence, concludes the proof.
Theorem 4 provides an analytical graph-theoretic modeling
of SDIAs using the fundamentals of observability attacks introduced in Section III. This enables a fundamental understanding
of SDIAs since it allows the characterization of the subset of
measurements which would be compromised as part of an SDIA
and hence enables defense against such attacks. In addition, this
analytical characterization of SDIAs enables a more in-depth
analysis of such integrity attacks and allows a unified derivation
of analytical solutions to a wide-range of well-studied problems
in this field, as will be explored in Section IV-C.
Example 1: As an illustrative example of the result11 in
Theorem 4, we consider the IEEE 14-bus system, shown in
Fig. 1, whose line transmission data can be found in [34]. We
consider the stealthy attack a = Hc with c = [1, 0, ..., 0]T ,
which corresponds to having the attack vector equal to the
first column of the Jacobian matrix H given by H(:, 1) =
[−16.9, 0, 0, 0, −16.9, 33.37, −5.05, −5.67, −5.75, zeros(1, 8)]T .
This attack consists of attacking measurement indices
{1, 5, 6, 7, 8, 9} which correspond to {F2 , I1 , I2 , I3 , I4 , I5 }.
In this respect, we next verify whether this attack is stealthy,
following Theorem 4. To this end, Fig. 5 shows a portion of
the critical sets - system measurements bipartite graph that is
relevant to the attacked measurements. The post-attack portion
of Fig. 5 marks the nodes corresponding to measurements
{F2 , I1 , I2 , I3 , I4 , I5 }, on the right-side of the bipartite graph,
as attacked (following the attack vector a). As a result, all
the edges connecting these nodes to the critical sets on the
left-side of the bipartite graph are removed. Then, building a
maximum matching over the post-attack bipartite graph shows
that, indeed, not all the critical sets are matched. Hence, the
removed measurements lead to a maximum matching that
does not include all critical sets. Furthermore, the addition of
a node corresponding to any of the attacked measurements,
i.e. {F2 , I1 , I2 , I3 , I4 , I5 }, would lead to reincluding one of
11 In this example, we index the measurements in Fig. 1 from
1 to 17 in an incremental manner based on the following order
(F2 , F8 , F9 , F15 , I1 , I2 , I3 , I4 , I5 , I6 , I7 , I9 , I11 , I12 , I13 , F17 , F19 ).
Pre-attack
Critical Set Measurements
C F2
F2
C I1
I1
I2
C I2
I3
C I3
I4
C F8
I11
C F9
I13
C I5
F8
I7
I9
F9
I5
Post-attack
Critical Set Measurements
C F2
F2
C I1
I1
C I2
I2
I3
C I3
I4
C F8
I11
C F9
I13
C I5
F8
Unmatched
I7
Matched
I9
Attacked
F
9
Uncompromised
I5
Maximum matching
Fig. 5. Critical sets – system measurements maximum matching and SDIAs.
the unmatched critical sets {C F2 , C I3 , C I5 } in the maximum
matching. This implies that the attack consists of a union of
observability sets which implies that the attack is, indeed,
stealthy.
C. Unified Solution to Diverse SDIA Problems
Theorem 4 provides a basis for studying various SDIA problems from a graph-theoretic perspective. Indeed, this representation provides a unified approach for characterizing analytical
solutions to various widely-studied SDIA problems. In this regard, we next present a set of such SDIA problems and show that
the derivations leading to Theorem 4 enable the understanding
and characterization of analytical solutions to these problems.
SDIA analyses can be categorized based on whether the focus
is on modeling the attack or the defense strategies. As such, we
first present two problems focusing on modeling attack strategies
followed by two problems focusing on the derivation of defense
strategies to thwart SDIAs.
1) Modeling SDIA Attack Strategies: Modeling SDIA attack
strategies enables a vulnerability assessment of the system and
allows anticipating sophisticated attack strategies which can
target the system. This, in turn, allows the derivation of adequate
defense strategies to thwart such attacks. As such, solving SDIA
problems focusing on modeling the attack strategies is indispensable to understanding such attacks and, as a result, defending the
system against them. We next focus on two problems which aim
at modeling potential attack strategies.
Problem 1: If measurement k ∈ M is attacked, what is the
minimal set of measurements which must be attacked along with
k for the attack to be stealthy? In other words, Problem 1 seeks
the solution to the following optimization problem:
min||Hc||0 ,
c
subject to: H(k, :)c = 1.
(6)
Problem 1 has been proposed in [20] and studied in [21]
and [22]. However, the derived solution in [21] is based on an
approximate relaxation method while the solution in [22] focuses
on the special case assuming that the measurement set consists
of all injection measurements at all buses and all line flow measurements at all transmission lines, which limits its generality.
Instead, here, we provide a general analytical characterization of
the solution to this problem using our developped graph-theoretic
framework.
The solution to this problem enables a vulnerability assessment of each measurement against SDIAs since it shows, for
each measurement, what is the minimum number of measurements which must be additionally compromised to potentially
launch an SDIA against the system. This can represent a security
index of that measurement following which, a measurement with
a lower (higher) security index is more (less) vulnerable to
SDIAs. In other words, a measurement which has a low security
index is more easily targeted by SDIAs since the adversary would
not need to comprise a large number of additional measurements
to lunch the stealthy attack. Such knowledge can be used to
improve the security of the system – by adding redundancy or
adding security defense mechanisms (such a encryption techniques or advanced meetering uints) – at the meters which are
deemed the most vulnerable to SDIAs.
The analytical graph-theoretic solution to Problem 1 is characterized in Theorem 5.
Theorem 5: The stealthy attack of smallest cardinality containing measurement k corresponds to attacking the measurements of the critical set of lowest cardinality which contains k.
Proof: First, we show that the attack containing the critical
set of lowest cardinality containing k is, indeed, stealthy. Then,
we prove that this attack corresponds to the stealthy attack
containing k that has a minimum cardinality.
By Theorem 4, for the attack to be stealthy, the removal of
the attacked measurements must lead the maximum matching
over the critical sets - system measurements bipartite graph not
to include all the critical sets (i.e. all the left-side nodes of the
bipartite graph). In other words, the attack must be composed
of a union of obseravbility sets. In this respect, removing the
critical set containing k that is of smallest cardinality is, indeed,
stealthy since removing a whole critical set will disconnect the
node corresponding to this critical set (on the left-hand side of
the critical sets - system measurements bipartite graph) from
the right-side of the bipartite graph which prevents this critical
measurement from being part of a maximum matching.
Next, we prove that there are no stealthy attacks containing
k that have a smaller cardinality. In this regard, for an attack
containing k to be stealthy, it must prevent a critical set, in
which k exists, from being part of the maximum matching
over the critical sets - system measurements bipartite graph. To
this end, a critical set would be excluded from the maximum
matching in two cases: 1) if all the measurements corresponding
to this critical set are attacked, or 2) if all the measurements
corresponding to this critical set are part of a different matching,
which assigns these measurements to other critical sets.
In the first case, considering attacking all the measurements
in a critical set, then attacking the critical set that has the
fewest number of measurements – as stated in this theorem –
corresponds to the minimum cardinality attack.
As for the second case, if a measurement k 0 in a critical set
containing k is matched – as part of a maximum matching – to
another critical set (we denote this set by C p ), then measurement
p must be attacked since, otherwise, C p would have been matched
to p sparing k to be matched to another critical set to maximize
the cardinality of the matching. In other words, matching a
critical set C p with a measurement k 0 6= p while p is not
attacked is contradictory to the assumption that this matching
is maximum. As a result, for a critical set C, such that k ∈ C, to
be discarded from the maximum matching, every measurement
in C must be matched to another critical set. This implies that at
least one measurement of each of these critical sets is attacked.
Thus, the number of attacked measurements will be at least
equal to the number of measurements within C for the attack
to be stealthy. Consequently, the stealthy attack containing k
that has the lowest cardinality corresponds to attacking only the
measurements within the critical set containing k that has the
lowest cardinality.
Solving Problem 1 will also facilitate solving another key
SDIA problem, referred to as Problem 2, and stated as follows.
Problem 2: What is the SDIA with the lowest cardinality? In
other words, which SDIA is a solution to:
min||a||0 ,
a
subject to: a = Hc.
(7)
Similarly to Problem 1, the solution to Problem 2 also provides
a vulnerability assessment of the system against SDIAs. In
fact, Problem 1 focuses on finding the security index associated
with each measurement. On the other hand, Problem 2 focuses
on the system as a whole by focusing on finding, in general,
the sparsest data injection attack which can target the system
and be stealthy. This corresponds to a security index for the
whole system. Indeed, this security index reflects the amount of
effort that an attacker must put to potentially launch an SDIAs
against the system. A low security index shows that, even when
manipulating a small set of measurements, the attack can be
stealthy. In contrast, a high security index reflects the robustness
of the system against SDIAs since the attacker would need to
concurrently manipulate a large number of measurements to
potentially launch a successful SDIA.
The solution to Problem 2 is provided in Proposition 1.
Proposition 1: The stealthy attack with the lowest cardinality
corresponds to attacking the smallest critical set.
Proof: This proof follows directly from the proof of Theorem 5. Indeed, since the stealthy attack containing measurement
k that is of smallest cardinality corresponds to the critical set
of lowest cardinality containing k, then searching for the global
stealthy attack of lowest cardinality should be limited to only
critical sets. Based on this fact, the stealthy attack of lowest
cardinality is the one in which the measurements in the critical
set of lowest cardinality are the only measurements that are
attacked (the only measurements having nonzero corresponding
elements in the attack vector a).
Example 2: For example, by inspecting the critical sets in
Table I, we can solve Problem 1 and Problem 2 for the IEEE
14-bus system shown in Fig. 1. With regard to Problem 1,
the results of Theorem 5 can be readily applied to find the
minimum stealthy attack containing a certain measurement k. For
example, the minimum stealthy attack containing measurement
I4 corresponds to attacking C I3 = {I3 , I2 , I4 }, since that is the
critical set of smallest cardinality containing I4 . As for Problem
2, the stealthy attack of lowest cardinality is one in which either
C F15 = {F15 , I7 }, C I6 = {I6 , I11 }, or C I9 = {I9 , I11 } are
attacked. As such, the minimum possible cardinality of a stealthy
attack for this IEEE 14-bus system is 2. To find such a stealthy
attack, the basis of the null space can be found for matrices
F15
I6
I9
H −(C ) , H −(C ) , or H −(C ) . We refer to these vectors as
nF15 , nI6 , and nI9 , respectively. As a result, these stealthy attack
vectors of minimum cardinality can be obtained as αHnF15 ,
αHnI6 , or αHnI9 , where α is a scalar multiplier.
2) Modeling SDIA Defense Strategies: The knowledge acquired from our introduced graph-theoretic framework enables
the derivation of adequate defense policies which can thwart
potential SDIAs. In this regard, next, two fundamental widelystudied problems for defending the system against SDIAs are
presented and investigated in Problem 3 and Problem 4.
Problem 3: What is the minimum set of measurements that
need to be protected (i.e. made immune to SDIAs) to guarantee
no SDIAs can be successful?
The solution to this problem enables finding a minimum-cost
defense strategy to thwart all potential SDIAs. Hence, this makes
the system robust against all possible SDIAs. However, even
though the solution to this problem provides the minimum-cost
defense strategy (assuming that protecting each measurement is
equally costly), for a practically large power system with several
thousand buses, such a defense strategy is likely to exceed any
practical security budget.
The solution to Problem 3 is presented in Theorem 6.
Theorem 6: The minimum set of measurements that must
be protected to guarantee that no SDIAs can be successful
corresponds to protecting all measurements in MC , i.e. all
measurements that are part of the original assignment function
forming the spanning tree over the power system.
Proof: By Theorem 4, for an attack to be stealthy, it
must lead to a critical set not to be matched as part of a
maximum matching over the critical sets - system measurements
bipartite graph. Hence, to guarantee that no attack can be stealthy,
all critical sets must be guaranteed to be matched. Thus, the
minimum number of measurements to be protected must be at
least equal to the number of critical sets, which is equal to
the number of measurements in MC . In this regard, protecting
every measurement m ∈ MC results in protecting the minimum
possible number of measurements which guarantees that C m can
be matched to m for all m ∈ MC , hence, guaranteeing that no
stealthy attack can be successfully carried out.
This proof can also be carried out equivalently using the
techniques of linear algebra. Indeed, protecting all the measurements in MC will guarantee that these measurements will be
part of the Jacobian matrix H. Since these measurements form
a spanning tree over the power system, their rows in H are
linearly independent. As such, let H C be the Jacobian matrix
corresponding only to measurements in MC , then H C c = 0
has no solution other than c = 0. The rows of H C are a subset
of the rows of H. As such, one cannot find an attack vector
a = Hc such that all the elements of a corresponding to the
rows of H C are zero. Hence, one cannot find a stealthy attack
a = Hc which does not attack the measurements in MC . As
a result, protecting these measurements will guarantee that no
stealthy attack can be carried out.
Here, the minimum defense set, solution to Problem 3, might
not be unique. In other words, the set MC might not be the only
minimum set of measurements which, when defended, makes the
system immune to SDIAs. However, characterizing a solution
to this problem provides important information regarding the
size of investments needed to make a power system immune
to SDIAs. In this regard, regardless of how high the number of
measurements in an N -bus system is, the number of measurements that must be protected to render the system immune to
SDIAs is always equal to N − 1. As such, by assessing the costs
of reinforcing the security of each measurement unit, the solution
to Problem 3 enables the calculation of the cost needed to
make the system robust against SDIAs. However, as the solution
implies, for practical power systems, securing this number of
measurements might exceed practical budget constraints.
Example 3: Applying the results in Theorem 6 to our treated
IEEE 14-bus system case analysis, protecting the measurements
in the first column of Table I is the set of measurements of
minimal cardinality which when protected renders the IEEE 14bus system in Fig. 1 immune to SDIAs.
Theorem 4 and the solutions to Problem 1, Problem 2,
and Problem 3 can be used to solve Problem 4 which was
proposed in [23] and which is presented next. The solution to
Problem 4 in [23] was derived based on an l1 relaxation of the
corresponding optimization problem which leads to approximate,
rather than generally accurate solutions.
Problem 4: What is the minimum set of measurements to
protect as to force the attacker to manipulate at least τa measurements to stay stealthy?
As discussed in Problem 3, making the system completely
robust against SDIAs may be very costly and exceed any practical budgetary constraints. As a result, rather than considering all
theoretically possible SDIAs, the solution to Problem 4 focuses
on defending the system against a large subset of practical
SDIAs in which the attacker’s limited resources prevents its
attack vector’s cardinality from exceeding τa . In other words, an
attacker might not be able to concurrently comprise more than
τa measurement units. This, hence, enables defending the system
against a practically large subset of potential SDIAs. In addition,
the solution to this problem allows using the knowledge about
the resources of potential attackers – which can be potentially
acquired from historical data – to compute adequate defense
policies against such attacks.
The solution to Problem 4 is presented in Proposition 2.
Proposition 2: A minimum set of measurements to protect
so that no attack with cardinality ||a||0 < τa can be stealthy,
corresponds to protecting one distinct measurement from each
critical set whose cardinality is less than τa .
Proof: Solving Problem 4 entails ensuring that all critical
sets of cardinality smaller than τa are part of the maximum
matching. Hence, when one distinct measurement in each of
these sets is secured, it is ensured that these critical sets will
be part of the maximum matching over the critical sets system measurements bipartite graph. As a result, additional
measurements would need to be attacked to target critical sets
of higher cardinality, if a stealthy attack were to be found,
which would require the attacker to manipulate more than τa
measurements. Hence, the solution to Problem 4 is a direct
result of Theorem 6 but by considering critical sets that have
cardinality smaller than τa rather than all critical sets, as is the
case in Theorem 6. As such, the rest of the proof of Proposition 2
follows directly from the proof of Theorem 6.
This result is very important since it allows the defender to
build on some knowledge that it has about the capacity and
resources of the attacker, to build a corresponding defense policy.
In other words, knowing that the attacker does not have the
capacity to concurrently manipulate more than τa measurements
enables the defender to focus on defending a smaller set of measurements rather than aiming to thwart any theoretically possible
SDIA. This can lead to a significant reduction in the needed
resources for such a defense since, as shown in the solution of
Problem 3, the latter defense policy requires committing a large
volume of resources which can exceed practical constraints.
Example 4: For our studied IEEE 14-bus system, consider that
τa = 3. This indicates that a set of measurements to protect must
be found to ensure that no attacker can have a successful stealthy
attack by attacking less than 3 measurements. The critical sets
that have cardinality lower than 3 are C F15 , C I6 , and C I9 which all
have a cardinality of 2. Now, we consider a distinct measurement
in each of these three sets, such that I7 , I6 , and I9 . As a result,
{I7 , I6 , I9 } is a minimum set of measurements which, when
defended, no stealthy attack vector of cardinality less than τa = 3
can be successfully launched.
As such, showing that SDIAs are a subset of our introduced observability attacks, enables using our proposed graphtheoretic framework to model, understand, and thwart such types
of cyber-physical attacks. Indeed, the four problems that we
have discussed show the way our developed framework enables
analytical characterization of the solutions to these various wellstudied SIDA problems. Such analytical characterization allows
assessing the vulnerability of the system against SDIAs as well
as deriving adequate defense strategies.
V. C ONCLUSION AND F UTURE O UTLOOK
In this paper, we have introduced a novel graph-theoretic
framework which enables a fundamental modeling of observability attacks targeting power systems and have proven that
the widely-studied stealthy data injection attacks are a special
case of such observability attacks. Based on this proposed
framework, we have characterized the analytical solutions to
various central observability and data injection attack problems.
These solutions aim at capturing potential attack strategies as
well as suggesting defense policies to thwart such attacks. In
this respect, we have shown that our derived framework enables
characterization of the sparsest stealthy attack as well as the
sparsest stealthy attack including a certain measurement. With
respect to defense policies, we have shown that our graphtheoretic framework enables the analytical characterization of
the minimum measurement set which when defended guarantees
thwarting any potential stealthy attack as well as the minimum
set of measurements whose defense guarantees that no attack
below a certain cardinality can be stealthy.
The proposed graph-theoretic framework provides a general
analytical tool using which a wide set of key observability
attacks and data injection attacks problems can be modeled and
analyzed, and is not limited to the set of problem examples
which are studied in this paper. For example, the problem of
characterizing the sparsest stealthy attack containing a certain
measurement can be extended to studying the sparsest stealthy
attack containing a certain set of measurements. Using our
proposed framework, a solution approach can be investigated
to potentially derive analytical solutions to this critical problem.
The solution of this problem enables a risk assessment of the
power system by quantifying the risk of having a vulnerable set
of measurements and the way that such a vulnerability can be
leveraged by an intelligent malicious attacker. Along the same
lines, for security assessment, a central problem is quantifying
the sparsest stealthy attack possible when a certain set of
measurements is defended. The solution to this problem enables
assessment of the effectiveness and impact of an implemented
defense strategy. This problem has been formalized in [23].
However, the proposed solution approach in [23] relied on an
l1 relaxation of the original optimization problem formulation
which leads to approximate numerical solutions. However, our
introduced graph-theoretic framework can be used to attempt the
characterization of analytical solutions to this problem.
Beyond these one-sided attack and defense problems, the
ability to analytically characterize attack and defense policies
using the proposed framework allows studying problems that
involve interactions between attackers and defenders from a
game-theoretic perspective. Such analyses can account for the
opponent’s potential attack or defense strategies when designing,
respectively, defense policies or attack vectors. As a result,
such analyses allow the modeling and investigation of practical
competitive attack vs. defense settings. This enables studying the
effects of sophisticated observability attacks and data injection
attacks on the system as well as the impact of proposed defense
strategies within various application domains such as electricity markets, congestion management, and contingency analysis,
among others, thus taking the application of our framework
beyond the domain of power systems which motivated this study.
R EFERENCES
[1] Y. Mo, T. H. J. Kim, K. Brancik, D. Dickinson, H. Lee, A. Perrig,
and B. Sinopoli, “Cyber–physical security of a smart grid infrastructure,”
Proceedings of the IEEE, vol. 100, no. 1, pp. 195–209, Jan 2012.
[2] A. Sanjab and W. Saad, Power System Analysis: Competitive Markets,
Demand Management, and Security. Cham: Springer International
Publishing, 2017, pp. 1–38. [Online]. Available: https://doi.org/10.1007/
978-3-319-27335-8 28-1
[3] G. E. Rahi, A. Sanjab, W. Saad, N. B. Mandayam, and H. V. Poor, “Prospect
theory for enhanced smart grid resilience using distributed energy storage,”
in Proc. 54th Annual Allerton Conference on Communication, Control, and
Computing (Allerton), Sept 2016, pp. 248–255.
[4] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against
state estimation in electric power grids,” in Proc. 16th ACM Conference on
Computer and Communications Security, Chicago, Illinois, USA, November 2009, pp. 21–32.
[5] Y. Liu, P. Ning, and M. Reiter, “False data injection attacks against state
estimation in electric power grids,” ACM Transactions on Information and
System Security (TISSEC), vol. 14, no. 1, pp. 1–33, May 2011.
[6] A. Sanjab and W. Saad, “Data injection attacks on smart grids with multiple
adversaries: A game-theoretic perspective,” IEEE Transactions on Smart
Grid, vol. 7, no. 4, pp. 2038–2049, July 2016.
[7] L. Xie, Y. Mo, and B. Sinopoli, “Integrity data attacks in power market
operations,” IEEE Transactions on Smart Grid, vol. 2, no. 4, pp. 659–666,
Dec 2011.
[8] A. Sanjab and W. Saad, “Smart grid data injection attacks: To defend or
not?” in Proc. IEEE International Conference on Smart Grid Communications (SmartGridComm), Nov 2015, pp. 380–385.
[9] Y. W. Law, T. Alpcan, and M. Palaniswami, “Security games for risk
minimization in automatic generation control,” IEEE Transactions on
Power Systems, vol. 30, no. 1, pp. 223–232, Jan 2015.
[10] A. Sanjab and W. Saad, “On bounded rationality in cyber-physical systems
security: Game-theoretic analysis with application to smart grid protection,”
in Proc. Joint Workshop on Cyber- Physical Security and Resilience in
Smart Grids (CPSR-SG), April 2016, pp. 1–6.
[11] F. Pasqualetti, F. Dorfler, and F. Bullo, “Attack detection and identification
in cyber-physical systems,” IEEE Transactions on Automatic Control,
vol. 58, no. 11, pp. 2715–2729, Nov 2013.
[12] Q. Zhu and T. Başar, “Robust and resilient control design for cyberphysical systems with an application to power systems,” in Proc. 50th IEEE
Conference on Decision and Control and European Control Conference
(CDC-ECC), Dec 2011, pp. 4066–4071.
[13] F. Pasqualetti, F. Dorfler, and F. Bullo, “Cyber-physical attacks in power
networks: Models, fundamental limitations and monitor design,” in Proc.
50th IEEE Conference on Decision and Control and European Control
Conference (CDC-ECC), Dec 2011, pp. 2195–2201.
[14] A. Ferdowsi, A. Sanjab, W. Saad, and N. B. Mandayam, “Game theory
for secure critical interdependent gas-power-water infrastructure,” in Proc.
Resilience Week (RWS), Sept 2017, pp. 184–190.
[15] A. Abur and A. G. Exposito, Power System State Estimation: Theory and
Implementation. New York: Marcel Dekker, 2004.
[16] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and
Control. John Wiley & Sons, 2012.
[17] J. B. A. London, L. F. C. Alberto, and N. G. Bretas, “Network observability:
identification of the measurements redundancy level,” in Proc. International
Conference on Power System Technology (PowerCon), vol. 2, 2000, pp.
577–582.
[18] K. C. Sou, H. Sandberg, and K. H. Johansson, “Computing critical k-tuples
in power networks,” IEEE Transactions on Power Systems, vol. 27, no. 3,
pp. 1511–1520, Aug 2012.
[19] O. Kosut, L. Jia, R. J. Thomas, and L. Tong, “Malicious data attacks on the
smart grid,” IEEE Transactions on Smart Grid, vol. 2, no. 4, pp. 645–658,
Dec 2011.
[20] H. Sandberg, A. Teixeira, and K. H. Johansson, “On security indices for
state estimators in power networks,” in First Workshop on Secure Control
Systems (SCS), 2010.
[21] K. C. Sou, H. Sandberg, and K. H. Johansson, “Electric power network
security analysis via minimum cut relaxation,” in Proc. 50th IEEE Conference on Decision and Control and European Control Conference, Dec
2011, pp. 4054–4059.
[22] J. M. Hendrickx, K. H. Johansson, R. M. Jungers, H. Sandberg, and K. C.
Sou, “Efficient computations of a security index for false data attacks in
power networks,” IEEE Transactions on Automatic Control, vol. 59, no. 12,
pp. 3194–3208, Dec 2014.
[23] T. Kim and H. V. Poor, “Strategic protection against data injection attacks
on power grids,” IEEE Transactions on Smart Grid, vol. 2, no. 2, pp. 326–
333, June 2011.
[24] S. Cui, Z. Han, S. Kar, T. T. Kim, H. V. Poor, and A. Tajer, “Coordinated
data-injection attack and detection in the smart grid: A detailed look at
enriching detection solutions,” IEEE Signal Processing Magazine, vol. 29,
no. 5, pp. 106–115, Sept 2012.
[25] J. London, A. Bretas, and N. Bretas, “Algorithms to solve qualitative
problems in power system state estimation,” International Journal of
Electrical Power and Energy Systems, vol. 26, no. 8, pp. 583 – 592, 2004.
[26] E. Castillo, A. J. Conejo, R. E. Pruneda, C. Solares, and J. M. Menendez,
“m − k robust observability in state estimation,” IEEE Transactions on
Power Systems, vol. 23, no. 2, pp. 296–305, May 2008.
[27] K. C. Sou, H. Sandberg, and K. H. Johansson, “On the exact solution to a
smart grid cyber-security analysis problem,” IEEE Transactions on Smart
Grid, vol. 4, no. 2, pp. 856–865, June 2013.
[28] Y. Zhao, A. Goldsmith, and H. V. Poor, “Minimum sparsity of unobservable
power network attacks,” IEEE Transactions on Automatic Control, vol. 62,
no. 7, pp. 3354–3368, July 2017.
[29] ——, “A polynomial-time method to find the sparsest unobservable attacks
in power networks,” in Proc. American Control Conference (ACC), July
2016, pp. 276–282.
[30] G. R. Krumpholz, K. A. Clements, and P. W. Davis, “Power system observability: A practical algorithm using network topology,” IEEE Transactions
on Power Apparatus and Systems, vol. PAS-99, no. 4, pp. 1534–1542, July
1980.
[31] V. H. Quintana, A. Simoes-Costa, and A. Mandel, “Power system topological observability using a direct graph-theoretic approach,” IEEE Transac-
tions on Power Apparatus and Systems, vol. PAS-101, no. 3, pp. 617–626,
March 1982.
[32] A. Bargiela, M. R. Irving, and M. J. H. Sterling, “Observability determination in power system state estimation using a network flow technique,”
IEEE Transactions on Power Systems, vol. 1, no. 2, pp. 108–112, May
1986.
[33] D. B. West, Introduction to Graph Theory, 2nd ed. Upper Saddle River,
N.J: Prentice Hall, 2001.
[34] R. Zimmerman, C. Murillo-Sanchez, and R. Thomas, “Matpower: Steadystate operations, planning, and analysis tools for power systems research
and education,” IEEE Transactions on Power Systems, vol. 26, no. 1, pp.
12–19, Feb 2011.
| 3cs.SY
|
Ten Diverse Formal Models
for a CBTC Automatic Train Supervision System
Franco Mazzanti
Alessio Ferrari
CNR-ISTI
Pisa, Italy
[email protected]
CNR-ISTI
Pisa, Italy
[email protected]
Communications-based Train Control (CBTC) systems are metro signalling platforms, which coordinate and protect the movements of trains within the tracks of a station, and between different stations.
In CBTC platforms, a prominent role is played by the Automatic Train Supervision (ATS) system,
which automatically dispatches and routes trains within the metro network. Among the various functions, an ATS needs to avoid deadlock situations, i.e., cases in which a group of trains block each
other. In the context of a technology transfer study, we designed an algorithm for deadlock avoidance
in train scheduling. In this paper, we present a case study in which the algorithm has been applied.
The case study has been encoded using ten different formal verification environments, namely UMC,
SPIN, NuSMV/nuXmv, mCRL2, CPN Tools, FDR4, CADP, TLA+, UPPAAL and ProB. Based on
our experience, we observe commonalities and differences among the modelling languages considered, and we highlight the impact of the specific characteristics of each language on the presented
models.
1
Introduction
Communications-based Train Control (CBTC) systems are the de-facto standard for metro signalling and
control, including several interacting wayside and onboard components that ensure safety and availability
of trains within the metro network. In the context of a technology transfer project named TRACE-IT,
the authors of the current paper, together with representatives of a large railway company, designed one
of the main components of a CBTC system prototype, namely the Automatic Train Supervision (ATS)
system [10]. This is a wayside system that dispatches and monitor trains along the metro network,
according to a set of predefined missions. The ATS includes a scheduling kernel, which shall ensure
that, regardless of train delays, no deadlock situation occurs, i.e., the missions are designed in such
a way that it never happens that two or more trains block each other from completing their missions.
In the context of the project, we applied formal methods to design and verify a scheduling algorithm
that addresses the deadlock avoidance problem [25]. The application of the algorithm to the TRACEIT case study was initially modelled and verified by means of the UMC tool [30, 4, 20]. Then, the
design of the case study was replicated with other six different formal frameworks – i.e., SPIN [16, 29],
NuSMV/nuXmv [5, 18], mCRL2 [14, 9], CPN Tools [17, 31], FDR4 [13, 27] and CADP [11, 6] – to
explore the potential of formal methods diversity [24]. This is the usage of different formal tools to
validate the same design, to increase the confidence on the verification results [19]. In the current paper,
we present the models discussed in [24], focusing on the differences between the modelling languages,
rather than on formal verification diversity. Furthermore, we provide three additional models, using
TLA+ [21, 7], ProB [2, 15] and UPPAAL [8, 32]. Within the context of this paper, our goal is to
provide some feedback on the differences and traps that should be tackled when changing the reference
John P. Gallagher, Rob van Glabbeek and Wendelin Serwe (Eds):
Models for Formal Analysis of Real Systems (MARS’18)
and Verification and Program Transformation (VPT’18)
EPTCS 268, 2018, pp. 104–149, doi:10.4204/EPTCS.268.4
c F. Mazzanti & A. Ferrari
This work is licensed under the
Creative Commons Attribution License.
F. Mazzanti & A. Ferrari
105
frameworks, and the commonalities that would allow a simple translation from one framework to another.
The models are made available in Appendix A and in attachment to this paper.
The remainder of the paper is structured as follows. In Sect. 2 we provide an overview of the modelled algorithm. In Sect. 3 we present the different models, discussing commonalities and differences
with a focus on syntactic and semantics discrepancies. Sect. 4 concludes the paper. In Appendix A, we
report the different models presented.
2
A Deadlock Avoidance Algorithm for ATS
This section describes basic elements of the modelled algorithm, which was defined in our previous
works [26, 25]. Fig. 1 shows the structure of the railway layout considered in this study. Nodes in the
yard correspond to itinerary endpoints, and the connecting lines correspond to the entry/exit itineraries
to/from those endpoints. Eight trains are placed in the layout. Each train has its own mission to execute,
defined as a sequence of itinerary endpoints. For example, the mission of train0, which traverses
the layout from left to right along top side of the yard, is defined by the mission vector: T0 = [1, 9, 10,
13, 15, 20, 23] (the numbers in the vector refer to the sequence of traversed endpoints in the diagram of
Fig. 1). The mission of train7, which instead traverses the layout from right to left, is defined by
the vector: T7 = [26, 22, 17, 18, 12, 27, 8]. The progress status of each train is represented by the index,
pointing to a position in the mission vector, which allows the identification of the endpoint in which
the train is at a certain moment. We will have 8 variables P0 , . . . , P7 , one for each train, which store the
current index for the train. For example, at the beginning, we have P0 = 0, . . . , P7 = 0, since all the trains
occupy the initial endpoints of their missions – at index 0 in the vector.
Parco della Vittoria
I
III
train0
1
BCA03
2
9
train1
Piazza Dante
I
10
Via Roma
Via Marco Polo
15
1
13
16
train4
II
II
20
24
train5
3
I
II
4
train2
23
11
5
III
BCA05
6
12
18
17
Viale dei Giardini
III
22
25
train6
IV
26
train7
27
train3
7
8
Figure 1: A fragment of the yard layout and the 8 missions of the trains
If the 8 trains are allowed to move freely, i.e., if their next endpoint is free, there is the possibility of
creating deadlocks, i.e., a situation in which the 8 trains block each other in their expected progression.
To solve this problem the scheduling algorithm of the ATS must take into consideration two critical
sections A and B – i.e., zones of the layout in which a deadlock might occur – which have the form of
a ring of length 8 (see Fig. 2), and guarantee that these rings are never saturated with 8 trains – further
information on how critical sections are identified can be found in our previous work [25]. This can be
modelled by using two global counters RA and RB, which record the current number of trains inside these
critical sections, and by updating them whenever a train enters or exits these sections. For this purpose,
each train mission Ti , with i = 0 . . . MISSION LEN (in our case MISSION LEN = 7) , is associated with:
a vector of increments/decrements Ai to be applied to counter RA at each step of progression; a vector Bi
of increments/decrements to be applied to counter RB.
Ten Diverse FMs for an ATS
106
For example, given T0 = [1, 9, 10, 13, 15, 20, 23], and A0 = [0, 0, 0, 1, 0, −1, 0], when train0 moves
from endpoint 10 to endpoint 13 (P0 = 3) we must check that the +1 increment of RA does not saturate
the critical section A, i.e., RA + A0 [P0 ] ≤ LA (in our case, LA = 7); if the check passes then the train can
proceed and safely update the counter RA := RA + A0[P0 ]. The maximum number of trains allowed in
each critical section (i.e., 7), will be expressed as LA and LB in the following.
Parco della Vittoria
I
III
BCA03
9
Piazza Dante
I
10
Via Roma
Via Marco Polo
15
13
16
20
24
I
18
III
BCA05
II
II
II
11
23
12
17
Viale dei Giardini
III
22
25
A
B
IV
26
27
Figure 2: The critical section A and B which must not be saturated by 8 trains
The models presented in Appendix A, which implement the algorithm described above, are deadlockfree, since the verification is being carried on as a final validation of a correct design. The actual
possibility of having deadlocks, if the critical sections management were not supported or incorrectly
implemented, can easily be observed by raising from 7 to 8 the values of the variables LA or LB.
The case study presented here is actually a fragment of the complete TRACE-IT case study. In
the original model the railway layout is much larger and the trains continually repeat cycling round
missions. In that configuration further deadlocks situations may occur and further critical sections have
to be defined and managed. The model considered in this case study represents just one of the three
fragments in which the complete TRACE-IT layout has been decomposed to render the complexity of
the problem amenable for formal verification. This is a typical procedure in the verification of real-world
railway problems [33].
The current design, in which each system state logically corresponds to a set of train progresses and
each train movement logically corresponds to an atomic system evolution step, leads to a state-space of
1,636,535 configurations. This data is useful because it allows the user to cross-check the correctness of
the encoding of this logical design in the various frameworks.
3
Commonalities and Differences
3.1 The basic blackboard model
We want to build a model that describes all the possible evolutions of the system composed by the 8
trains, with purpose of verifying the correctness of the A0, ..., A7 and B0, ..., B7 tables that control
the non saturation of the sections A and B, and the correctness of the assumption that the A and B
sections are the only zones where a deadlock might occur. The design skeleton we have in mind is
that of a blackboard model, where a global space of common variables is read and updated by a set
of atomic transformation operations. An atomic system evolution corresponds to a one-step movement
of one train in the yard, which can occur when the next endpoint is free and when the move does not
saturate neither the A section, nor the B section. We have encoded the above simple skeleton design
F. Mazzanti & A. Ferrari
107
Table 1: Different models developed with associated frameworks.
Framework
CADP
CPN Tools
FDR4
mCRL2
ProB
NuSMV/nuXmv
SPIN
TLA+
UMC
UPPAAL
File Name
Description
cadp oneway8par.lnt
cadp oneway8seq.lnt
cpn-oneway<X>.xml
fdr4 oneway8par.txt
fdr4 oneway8seq.txt
mcrl2 oneway8par.txt
mcrl2 oneway8seq.txt
prob oneway8seq.mch
smv oneway8-SM.smv
spin oneway8.pml
tla oneway8.txt
umc oneway8seq.txt
uppaal-oneway8par.ta
uppaal-oneway8seq.ta
Parallel without shared memory
Sequential
Parallel without shared memory with X trains
Parallel without shared memory
Sequential
Parallel without shared memory
Sequential
Sequential
Sequential
Sequential
Sequential
Sequential
Parallel with shared memory
Sequential
using notations supported by 10 verification frameworks, namely UMC, Promela/SPIN, NuSMV/nuXmv,
mCRL2, CPN Tools, FDR4, CADP, TLA+, UPPAAL and ProB. Within the context of this paper, our
goal is to provide some feedback on the differences and traps that should be tackled when changing the
reference frameworks, and the commonalities that would allow a simple translation from one framework
to another. Each framework surely has its own typical set of features that might lead to the best modelling
and verification of a system, but, in this work, we are not interested in comparing the best way in which
all the 10 frameworks could model the system. Instead, we are interested in seeing if, and to which extent,
our basic design skeleton could be fitted with minimal transformations in all the frameworks taken into
consideration.
In the following subsections we summarise some of the aspects that appear to characterise the differences of the various frameworks as evidenced by our specification problem. These observations can
support the reader in making sense of the different models that are reported in Appendix A1 , and attached
to the current paper. More specifically, for each framework, we provide one or more model variants. The
variants represent different modelling styles, according to the classification provided in Sect. 3.2. In
the case of CPN Tools, the different variants are associated to models with a different number of trains.
Indeed, in our experiments, presented in [24, 23], CPN Tools was not able to verify the case with eight
trains, and models with a lower number or trains were tested. Table 1 provides a brief description of the
different variants considered, with the associated file names.
3.2 System Design Structure
The frameworks taken into account allow different kinds of model structures, which can be seen in our
variants.
Sequential With the sequential design structure the global system status is read and updated by a single
sequential, nondeterministic process. This is the case that more directly reflects our initial design
skeleton, and this structure has been modelled in all the considered frameworks2 , with the excep1 In
Appendix A we report solely the sequential cases – according to the classification in Sect. 3.2 – which are the most
representative for our design.
2 mcrl2 oneway8seq.ta, cadp oneway8seq.lnt, fdr4 oneway8seq.csp, umc oneway8seq.txt, spin oneway8.pml, prob oneway8.mch, tla oneway8.txt, smv oneway8-SM.smv
Ten Diverse FMs for an ATS
108
tion of CPN Tools. Indeed, this modelling style can be reproduced with CPN Tools, but it is not in
line with the typical use of Petri Nets.
Parallel without Shared Memory With this design structure we indicate the case in which different
parallel process interact among themselves in the absence of a common shared memory that could
be directly read and updated by the processes. This is in general the case of concurrent frameworks,
such as UMC, CPN, FDR4, CADP, mCRL2, where sets of processes (or a network layout in the
case of Petri Nets) are used to model the system, and where a single entity might model the
evolutions in time of a specific component of the system status (e.g., a variable). This is not our
main reference scenario, however in the case of mCRL2, CADP, FDR4, CPN Tools we show
alternative modelling examples that follow this design structure3 .
Parallel with Shared Memory With this design structure a set of parallel processes share a common
memory space, and, at the same time, may interact through inter-process communication. SPIN
and UPPAAL are the only frameworks that allow the user to design a system in this way. An
example of this model structure has been shown only in the case of UPPAAL4.
subsectionLanguage Style
Another evident difference among the various frameworks, is the overall style of the language used to
specify the system. For example, if we consider the way in which the transition relation (i.e., the system
evolutions) are described, we can observe that three main approaches are followed by our considered
frameworks. These three language styles can be qualified as imperative, logical and algebraic, and are
exemplified below with small fragments of code in the style of CADP-LNT [12]5 , TLA+ and FDR4,
respectively.
if P0<6 then
P0 := P0+1;
RA := RA+A0[P0];
end;
(P0 < 6) &
(P0’ = P0+1) &
(RA’ = RA+A0[P0+1])
System(P0,RA) =
(P0 <6) ->
System(P0+1,RA+A0[P0+1])
In spite of the apparent difference, if the state transformation to be carried out during a system
evolution is simple (like it happens in our case), the three styles are roughly equivalent, and translation
from one style to the other can be performed with limited effort.
3.3 Arrays and Indexing
In our example we do not have the need to use sophisticated data structures, and our design skeleton is just
based on integer values and fixed-size tables of numbers. Sometimes, e.g., in the case of UMC, SPIN,
NuSMV/nuXmv, CADP-LNT, UPPAAL, TLA+, array-like types and indexing operations are natively
supported by the specification language; other times, e.g., in the case of CPN Tools, FDR4, mCRL2,
arrays should be represented as functions, or sequences, or lists, and the indexing operations possibly
manually encoded as custom recursive functions. For example, in the case of FDR4 we have:
M0 = <1,9,10,13,15,20,23>
select_item(list,index) =
if index==0 then
3 mcrl2
-- list of endpoint for the mission of train0
-- item selection, given an index
-- (assuming index in the appropriate range)
oneway8par.ta, cadp oneway8par.lnt, fdr4 oneway8par.csp, cpn oneway8.xml, cpn oneway6-nocol.xml
4 uppal-oneway8seq.ta
5
LTN is one of the languages supported by CADP, and is the language chosen for our experiment.
F. Mazzanti & A. Ferrari
109
head(list)
else
select_item(tail(list),index-1);
3.4 System Initialisation
The different ways in which the frameworks treat system initialisation point out a difference that might
trick an inexperienced designer. Three different approaches can be recognised when a state variable in
defined by the model, but not explicitly initialised at system startup.
Default Value The uninitialised variables might get some default initial value (typically 0 for integers).
This is the approach found in UMC, SPIN, UPPAAL.
Error The situation can be statically recognised as a design error, and notified to the designer. This is
the approach followed by TLA+, ProB, CPN Tools, FDR4, mCRL2, CADP-LNT.
Nondeterministic Assignment The not explicitly initialised variable may nondeterministically get any
of the possible values allowed by its type. This approach has been encountered only in in NuSMV/nuXmv. From one side this choice provides a powerful and flexible way to specify a rich
set of possible system initial values, from the other side it might trick an inexperienced designer
wrongly thinking that a classical default value (like 0) is used instead.
3.5 The Transition Relation
In all the considered frameworks the transition relation is defined by rules that have the form: guardcondition / state-transformation-effects. A possible question is what happens to the variables that are not
explicitly modified by the state-transformation-effects. The situation is similar to the initialisation issue
previously seen. Also in this case we have three different approaches:
Previous Value The not explicitly assigned state variables preserve their previous value. This is the
approach followed by CPN, UPPAAL, FDR4, mCRL2, SPIN, UMC, ProB, CADP-LNT.
Default Value The not explicitly assigned state variables get a default null value. This is what happens
in the case of TLA+.
Nondeterministic Assignment The not explicitly assigned variable may nondeterministically get any
of the possible values allowed by its type. This is what happens in the NuSMV/nuXmv case.
The difference among the three classes is evident if we compare the fragments of state-transformationeffects as they occur in CADP-LNT, TLA+ and NuSMV/nuXmv:
P0 := P0+1;
(P0’ = P0+1) &
UNCHANGED<<P1,..P7>>
next(P0) in P0+1 &
next(P1) in P1 &
...
next(P7) in P7
While with CADP-LNT it is not needed to make explicit that P1 ... P7 do not change their value,
in TLA+ we need to use the keyword UNCHANGED, and in NuSMV/nuXmv we have to explicitly state, for
each variable, that the next value is equal to the one at the previous execution step.
Another relevant difference among the various frameworks is whether they allow the transition relation to be only partially defined, i.e., are certain inputs and certain states allowed not to trigger a system
evolution?
Ten Diverse FMs for an ATS
110
In our problem this situation actually occurs. For example, when a train cannot proceed because its
next endpoint is occupied by another train, the rule describing the train progress cannot be applied. In all
frameworks, with the exception of NuSMV/nuXmv, this does not represent a problem. It simply means
that from such a system configuration state there is no outgoing edge corresponding to the movement of
that train.
In the case of NuSMV/nuXmv instead the transition relation must be a total function. This means
that if a certain state configuration and a certain system input does not trigger an actual system evolution,
we should equally explicitly state that the next system state is unchanged. If we fail to explicitly state
that, the consequence is that the next state can become any state where all the system state variables nondeterministically get any of the values allowed by their type. Notice that in this way we are introducing
self loops in many states of the graph describing the system behaviour, and this has a certain impact on
the way in which the system properties could be stated and verified. For example, the user might be
constrained to specify fairness constraints, or avoid the use of LTL formulas, or avoid CTL formulas like
AF<statepredicate>. Indeed the verification approach of NuSMV always takes into consideration only
infinite – possibly fair, if requested – traces6 .
3.6
Verification Techniques
In our case the property we want to verify is that for all possible executions all the trains eventually
complete their missions. This property can be easily specified and verified in all the considered frameworks. However each framework provides original advanced verification features not supported by other
frameworks. The possibility to translate a specification from a formalism to another might lead to several
advantages:
• We can increase the confidence of the verification results, given that none of the analysed frameworks are qualified at the highest integrity levels usually required by safety critical standards.
• We can exploit the specific strong points of more than one framework (e.g. the friendliness of a
user interface, the ability to scale well, the possibility of generating program code or performing
model based testing).
• We can verify a wider class of properties. For example, by importing a FDR4 model into ProB we
can verify also LTL/CTL properties, by translating a model into UPPAAL we can introduce and
verify further time related aspects, and so on. Table 2 summarises the basic verification features
that the considered frameworks make available.
3.7 Some Performance Data
It is not a goal of our paper to make a comparative evaluation of the various frameworks in terms of scalability or performance. Nevertheless a summary of the experienced times when evaluating the property
that for all possible executions all the trains eventually complete their missions might still be a useful
approximate indication of the impact of a certain system design approach / formal verification technique
in terms of performance. The verification times presented in Table 3 are expressed as ranges because
they actually depend of the specific design approach adopted, on the specific formulas being evaluated,
and on the specific options used during the tool execution. We refer to [24] for additional details.
6 when
using the -bmc option the behaviour might be different
F. Mazzanti & A. Ferrari
111
Table 2: Verification features supported by the various frameworks
Framework
Supported Verification Techniques
UMC
SPIN
NuSMV/nuXMV
CADP
UPPAAL
TLA+
ProB
mCRL2
FDR4
CPN
model checking CTL-like, state-event based logics
model checking LTL, fairness requirements
LTL, CTL, PSL [1], SMT model checking, fairness requirements
MCL [22], Parametric Mu-Calculus model checking, equivalence checking
MITL [3], time-related, and probability related properties
LTL, Theorem Proving, Proof Validations
LTL, CTL model checking, constraints based checking
Parametric Mu Calculus model checking, equivalence checking
Refinement Checking, fairness requirements
CTL, custom ML properties
Table 3: Indicative Summary of Evaluation Times
Framework
UMC
SPIN
NuSMV/nuXMV
CADP
UPPAAL
TLA+
ProB
mCRL2
FDR4
CPN
4
Range of evalution times
38 - 86 seconds
13 - 47 seconds
2.9 - 43 seconds
29 seconds
16 seconds
3 minutes
32 minutes
2 minutes -19 minutes
15 seconds - 20 minutes
unable to deal with the state-space size
Conclusion
The availability of CBTC systems relies on the existence of smart ATS systems that prevent the occurrence of deadlock situations in the metro network. In this paper, we present different models of a
scheduling algorithm for an ATS, which was designed and verified to avoid deadlocks. Ten different
formal frameworks are used, and different variants of system design structure are presented, according
to the features made available by the frameworks. Differences in terms of language style, allowed data
types, and treatments of the system evolution are observed, based on the developed models. In our future
work, we plan to adapt our design to tools for model-based development such as Simulink/Stateflow, and
SCADE, to explore their potential in terms of modelling styles and verification capabilities, and compare them with the other frameworks. Furthermore, in the context of the EU ASTRail project7 we are
involved in a comparative analysis of formal and semi-formal tools in the railway domain. The experience gained with the different frameworks will be applied to provide diverse models for ERTMS/ETCS
(European Rail Traffic Management System/European Train Control System) Level 3, the next evolution
of ERTMS/ETCS. This will allow us to further stress the capability of the frameworks with a different
design, including time and probabilistic aspects. It shall be noticed that, in the current work, we did not
discuss aspects related to the usability of the various frameworks. This issue is of paramount importance,
as highlighted, among others, by Sirjani [28], and is also going to be considered in the context of the
ASTRail project.
7 http://www.astrail.eu
Ten Diverse FMs for an ATS
112
Acknowledgements This work has been partially funded by the ASTRail project. This project received
funding from the Shift2Rail Joint Undertaking under the European Unions Horizon 2020 research and
innovation programme under grant agreement No 777561. The content of this paper reflects only the
authors view and the Shift2Rail Joint Undertaking is not responsible for any use that may be made of the
included information.
References
[1] (2010): IEEE Standard for Property Specification Language (PSL). IEEE Std 1850-2010 (Revision of IEEE
Std 1850-2005), pp. 1–182, doi:10.1109/IEEESTD.2010.5446004.
[2] Jean-Raymond Abrial (2010): Modeling in Event-B: System and Software Engineering. Cambridge University Press, doi:10.1017/S0956796812000081.
[3] Rajeev Alur, Tomás Feder & Thomas A. Henzinger (1991): The Benefits of Relaxing Punctuality. In: Proceedings of the Tenth Annual ACM Symposium on Principles of Distributed Computing, Montreal, Quebec,
Canada, August 19-21, 1991, pp. 139–152, doi:10.1145/112600.112613.
[4] Maurice H ter Beek, Stefania Gnesi & Franco Mazzanti (2015): From EU projects to a family of model
checkers. In: Software, Services, and Systems, LNCS 8950, Springer, pp. 312–328, doi:10.1007/
978-3-319-15545-6_20.
[5] Roberto Cavada, Alessandro Cimatti, Michele Dorigatti, Alberto Griggio, Alessandro Mariotti, Andrea
Micheli, Sergio Mover, Marco Roveri & Stefano Tonetta (2014): The nuXmv Symbolic Model Checker.
In: CAV, pp. 334–342, doi:10.1007/978-3-319-08867-9_22.
[6] INRIA CONVECS (2018): CADP Home Page. https://cadp.inria.fr.
[7] Microsoft Corp. (2018): The TLA Toolbox Home Page. https://lamport.azurewebsites.net/tla/toolbox.html.
[8] Alexandre David, Kim G. Larsen, Axel Legay, Marius Mikučionis & Danny Bøgsted Poulsen (2015): Uppaal
SMC tutorial. International Journal on Software Tools for Technology Transfer 17(4), pp. 397–415, doi:10.
1007/s10009-014-0361-y.
[9] Technische Universiteit Eindhoven (2018): mCRL2 Home Page. http://www.mcrl2.org/.
[10] Alessio Ferrari, Giorgio O Spagnolo, Giacomo Martelli & Simone Menabeni (2014): From commercial
documents to system requirements: an approach for the engineering of novel CBTC solutions. International Journal on Software Tools for Technology Transfer, STTT 16(6), pp. 647–667, doi:10.1007/
s10009-013-0298-6.
[11] Hubert Garavel, Frédéric Lang, Radu Mateescu & Wendelin Serwe (2013): CADP 2011: a toolbox
for the construction and analysis of distributed processes. STTT 15(2), pp. 89–107, doi:10.1007/
s10009-012-0244-z.
[12] Hubert Garavel, Frédéric Lang & Wendelin Serwe (2017): From LOTOS to LNT. In: ModelEd, TestEd,
TrustEd - Essays Dedicated to Ed Brinksma on the Occasion of His 60th Birthday, pp. 3–26, doi:10.1007/
978-3-319-68270-9_1.
[13] Thomas Gibson-Robinson, Philip Armstrong, Alexandre Boulgakov & Andrew W Roscoe (2014): FDR3 A
modern refinement checker for CSP. In: International Conference on Tools and Algorithms for the Construction and Analysis of Systems, Springer, pp. 187–201, doi:10.1007/978-3-642-54862-8_13.
[14] Jan Friso Groote & Mohammad Reza Mousavi (2014): Modeling and analysis of communicating systems.
MIT Press.
[15] Heinrich-Heine-University
(2018):
The
ProB
https://www3.hhu.de/stups/prob/index.php/Main Page.
Animator
and
Model
Checker.
[16] Gerard Holzmann (2003): The Spin Model Checker: Primer and Reference Manual. Addison-Wesley Professional.
F. Mazzanti & A. Ferrari
113
[17] Kurt Jensen & Lars M Kristensen (2009): Coloured Petri nets: modelling and validation of concurrent
systems. Springer Science & Business Media, doi:10.1007/b95112.
[18] Fondazione Bruno Kessler (2018): The nuXmv model checker Home Page. https://nuxmv.fbk.eu/.
[19] Tuomas Kuismin & Keijo Heljanko (2013): Increasing confidence in liveness model checking results with
proofs. In: Haifa Verification Conference, Springer, pp. 32–43, doi:10.1007/978-3-319-03077-7_
3.
[20] ISTI-FMT Laboratory (2018): KandISTI-UMC Home Page. https://fmt.isti.cnr.it/umc.
[21] Leslie Lamport (2002): Specifying Systems. https://lamport.azurewebsites.net/tla/book-02-08-08.pdf.
[22] Radu Mateescu & Damien Thivolle (2008): A model checking language for concurrent value-passing
systems. In: International Symposium on Formal Methods, Springer, pp. 148–164, doi:10.1007/
978-3-540-68237-0_12.
[23] Franco Mazzanti, Alessio Ferrari & Giorgio Oronzo Spagnolo (2014): Experiments in Formal Modelling of a
Deadlock Avoidance Algorithm for a CBTC System. In: International Symposium on Leveraging Applications
of Formal Methods - ISoLA 2016, Volune Part II, LNCS 9953, Springer, pp. 297–314, doi:10.1007/
978-3-319-47169-3_22.
[24] Franco Mazzanti, Alessio Ferrari & Giorgio Oronzo Spagnolo (2018): Towards Formal Methods Diversity
in Railways: an Experience Report with Seven Frameworks. International Journal on Software Tools for
Technology Transfer, STTT 20(3), doi:10.1007/s10009-018-0488-3.
[25] Franco Mazzanti, Giorgio Oronzo Spagnolo, Simone Della Longa & Alessio Ferrari (2014): Deadlock
avoidance in train scheduling: a model checking approach. In: International Workshop on Formal Methods for Industrial Critical Systems, FMICS 2014, LNCS 8718, Springer, pp. 109–123, doi:10.1007/
978-3-319-10702-8_8.
[26] Franco Mazzanti, Giorgio Oronzo Spagnolo & Alessio Ferrari (2014): Designing a deadlock-free train scheduler: A model checking approach. In: NASA Formal Methods Symposium, LNCS 8430, Springer, pp.
264–269, doi:10.1007/978-3-319-06200-6_22.
[27] University of Oxford (2018):
https://www.cs.ox.ac.uk/projects/fdr/.
FDR4
The
CSP
Refinement
Checker
Home
Page.
[28] Marjan Sirjani (2017): Power is Overrated, Go for Friendliness! Expressiveness, Faithfulness and Usability
in Modeling - The Actor Experience. In: Principles of Modeling -Essays dedicated to Edward A. Lee on the
Occasion of his 60th Birtday. Available at http://rebeca-lang.org/assets/papers/2017/
Friendliness.pdf.
[29] spinroot (2018): Verifying Multi-threaded Software with Spin. http://spinroot.com/spin/whatispin.html.
[30] Maurice H Ter Beek, Alessandro Fantechi, Stefania Gnesi & Franco Mazzanti (2011): A state/event-based
model-checking approach for the analysis of abstract system properties. Science of Computer Programming
76(2), pp. 119–135, doi:10.1016/j.scico.2010.07.002.
[31] CPN Tools (2018): CPN Tools Home Page. http://cpntools.org/.
[32] Uppsala University and Aalborg University (2015): UPPAAL Home Page. http://www.uppaal.org/.
[33] Kirsten Winter & Neil J Robinson (2003): Modelling large railway interlockings and model checking small
ones. In: Proceedings of the 26th Australasian computer science conference-Volume 16, Australian Computer
Society, Inc., pp. 309–316.
Appendix A
This appendix includes the sequential models for the different tools (when a textual representation is
available). The all these models, together with the other graphical models for CPN Tools and ProB, can
be retrieved from the MARS repository.
Ten Diverse FMs for an ATS
114
4.1 CADP-LNT
module CADP_ONEWAY8SEQ is
-----------------------------------------------------------------------------type Train_Number is
range 0 .. 7 of nat
end type
-----------------------------------------------------------------------------type Train_Mission is
array [0 .. 6] of nat
end type
-----------------------------------------------------------------------------type Train_Constraint is
array [0 .. 6] of int -- actually, of range -1 .. 1
end type
-----------------------------------------------------------------------------channel Movement is
(Train : Train_Number)
end channel
-----------------------------------------------------------------------------process MAIN [MOVE : Movement, ARRIVED : none] is
var P0, P1, P2, P3, P4, P5, P6, P7 : nat,
RA, RB : int,
LA, LB : int,
T0, T1, T2, T3, T4, T5, T6, T7 : Train_Mission,
A0, A1, A2, A3, A4, A5, A6, A7 : Train_Constraint,
B0, B1, B2, B3, B4, B5, B6, B7 : Train_Constraint
in
P0 := 0;
P1 := 0;
P2 := 0;
P3 := 0;
P4 := 0;
P5 := 0;
P6 := 0;
P7 := 0;
RA := 1;
RB := 1;
LA := 7; -- limit for region A
LB := 7; -- limit for region B
-- ------------ train missions -----------T0 := Train_Mission ( 1, 9,10,13,15,20,23);
T1 := Train_Mission ( 3, 9,10,13,15,20,24);
T2 := Train_Mission ( 5,27,11,13,16,20,25);
T3 := Train_Mission ( 7,27,11,13,16,20,26);
T4 := Train_Mission (23,22,17,18,11, 9, 2);
T5 := Train_Mission (24,22,17,18,11, 9, 4);
T6 := Train_Mission (25,22,17,18,12,27, 6);
T7 := Train_Mission (26,22,17,18,12,27, 8);
-- ------------------------------------------ ----- region A: train constraints -----A0 := Train_Constraint ( 0, 0, 0, 1, 0,-1, 0);
A1 := Train_Constraint ( 0, 0, 0, 1, 0,-1, 0);
F. Mazzanti & A. Ferrari
115
A2
A3
A4
A5
A6
A7
--
:= Train_Constraint ( 0, 0, 1,-1, 0, 1, 0);
:= Train_Constraint ( 0, 0, 1,-1, 0, 0, 0);
:= Train_Constraint ( 0, 1, 0, 0,-1, 0, 0);
:= Train_Constraint ( 0, 1, 0, 0,-1, 0, 0);
:= Train_Constraint ( 0, 0, 0,-1, 0, 0, 0);
:= Train_Constraint ( 0, 1, 0,-1, 0, 0, 0);
-----------------------------------------
-B0
B1
B2
B3
B4
B5
B6
B7
--
----- region B: train constraints -----:= Train_Constraint ( 0, 0, 0, 1, 0,-1, 0);
:= Train_Constraint ( 0, 0, 0, 1, 0,-1, 0);
:= Train_Constraint ( 0, 0, 1,-1, 0, 0, 0);
:= Train_Constraint ( 0, 0, 1,-1, 0, 1, 0);
:= Train_Constraint ( 0, 1, 0, 0,-1, 0, 0);
:= Train_Constraint ( 0, 1, 0, 0,-1, 0, 0);
:= Train_Constraint ( 0, 1, 0,-1, 0, 0, 0);
:= Train_Constraint ( 0, 0, 0,-1, 0, 0, 0);
-----------------------------------------
loop
select
only if
(P0 < 6) and
(T0 [P0+1] != T1 [P1]) and
(T0 [P0+1] != T2 [P2]) and
(T0 [P0+1] != T3 [P3]) and
(T0 [P0+1] != T4 [P4]) and
(T0 [P0+1] != T5 [P5]) and
(T0 [P0+1] != T6 [P6]) and
(T0 [P0+1] != T7 [P7]) and
(RA + A0 [P0+1] <= LA) and
(RB + B0 [P0+1] <= LB)
then
MOVE (0 of Train_Number);
P0 := P0 + 1;
RA := RA + A0 [P0];
RB := RB + B0 [P0]
end if
[]
only if
(P1 < 6) and
(T1 [P1+1] != T0 [P0]) and
(T1 [P1+1] != T2 [P2]) and
(T1 [P1+1] != T3 [P3]) and
(T1 [P1+1] != T4 [P4]) and
(T1 [P1+1] != T5 [P5]) and
(T1 [P1+1] != T6 [P6]) and
(T1 [P1+1] != T7 [P7]) and
(RA + A1 [P1+1] <= LA) and
(RB + B1 [P1+1] <= LB)
then
MOVE (1 of Train_Number);
P1 := P1 + 1;
RA := RA + A1 [P1];
RB := RB + B1 [P1]
end if
[]
only if
(P2 < 6) and
(T2 [P2+1] != T0 [P0]) and
(T2 [P2+1] != T1 [P1]) and
(T2 [P2+1] != T3 [P3]) and
(T2 [P2+1] != T4 [P4]) and
(T2 [P2+1] != T5 [P5]) and
-- next place of train0 not occupied by train1
-- next place of train0 not occupied by train2
-- next place of train0 not occupied by train7
-- progress of train0 does not saturate RA
-- progress of train0 does not saturate RD
Ten Diverse FMs for an ATS
116
(T2 [P2+1] != T6 [P6]) and
(T2 [P2+1] != T7 [P7]) and
(RA + A2 [P2+1] <= LA) and
(RB + B2 [P2+1] <= LB)
then
MOVE (2 of Train_Number);
P2 := P2 + 1;
--if ( P2 == 13 ) then P2 := 0 end if;
RA := RA + A2 [P2];
RB := RB + B2 [P2]
end if
[]
only if
(P3 < 6) and
(T3 [P3+1] != T0 [P0]) and
(T3 [P3+1] != T1 [P1]) and
(T3 [P3+1] != T2 [P2]) and
(T3 [P3+1] != T4 [P4]) and
(T3 [P3+1] != T5 [P5]) and
(T3 [P3+1] != T6 [P6]) and
(T3 [P3+1] != T7 [P7]) and
(RA + A3 [P3+1] <= LA) and
(RB + B3 [P3+1] <= LB)
then
MOVE (3 of Train_Number);
P3 := P3 + 1;
RA := RA + A3 [P3];
RB := RB + B3 [P3]
end if
[]
only if
(P4 < 6) and
(T4 [P4+1] != T0 [P0]) and
(T4 [P4+1] != T1 [P1]) and
(T4 [P4+1] != T2 [P2]) and
(T4 [P4+1] != T3 [P3]) and
(T4 [P4+1] != T5 [P5]) and
(T4 [P4+1] != T6 [P6]) and
(T4 [P4+1] != T7 [P7]) and
(RA + A4 [P4+1] <= LA) and
(RB + B4 [P4+1] <= LB)
then
MOVE (4 of Train_Number);
P4 := P4 + 1;
RA := RA + A4 [P4];
RB := RB + B4 [P4]
end if
[]
only if
(P5 < 6) and
(T5 [P5+1] != T0 [P0]) and
(T5 [P5+1] != T1 [P1]) and
(T5 [P5+1] != T2 [P2]) and
(T5 [P5+1] != T3 [P3]) and
(T5 [P5+1] != T4 [P4]) and
(T5 [P5+1] != T6 [P6]) and
(T5 [P5+1] != T7 [P7]) and
(RA + A5 [P5+1] <= LA) and
(RB + B5 [P5+1] <= LB)
then
MOVE (5 of Train_Number);
P5 := P5 + 1;
RA := RA + A5 [P5];
RB := RB + B5 [P5]
end if
F. Mazzanti & A. Ferrari
[]
only if
(P6 < 6) and
(T6 [P6+1] != T0 [P0]) and
(T6 [P6+1] != T1 [P1]) and
(T6 [P6+1] != T2 [P2]) and
(T6 [P6+1] != T3 [P3]) and
(T6 [P6+1] != T4 [P4]) and
(T6 [P6+1] != T5 [P5]) and
(T6 [P6+1] != T7 [P7]) and
(RA + A6 [P6+1] <= LA) and
(RB + B6 [P6+1] <= LB)
then
MOVE (6 of Train_Number);
P6 := P6 + 1;
RA := RA + A6 [P6];
RB := RB + B6 [P6]
end if
[]
only if
(P7 < 6) and
(T7 [P7+1] != T0 [P0]) and
(T7 [P7+1] != T1 [P1]) and
(T7 [P7+1] != T2 [P2]) and
(T7 [P7+1] != T3 [P3]) and
(T7 [P7+1] != T4 [P4]) and
(T7 [P7+1] != T5 [P5]) and
(T7 [P7+1] != T6 [P6]) and
(RA + A7 [P7+1] <= LA) and
(RB + B7 [P7+1] <= LB)
then
MOVE (7 of Train_Number);
P7 := P7 + 1;
RA := RA + A7 [P7];
RB := RB + B7 [P7]
end if
[]
-- ALL TRAINS RUNNING
only if (P0 == 6) and (P1 == 6) and (P2 == 6) and (P3 == 6) and
(P4 == 6) and (P5 == 6) and (P6 == 6) and (P7 == 6)
then
ARRIVED
end if
end select
end loop
end var
end process
end module
----------------
lnt.open cadp_oneway8.lnt generator x
bcg_info x.bcg
1_636_545 states
7_134_233 transitions
time lnt.open cadp_oneway8small.lnt evaluator4 cadpafarr.mcl
cadpafarr.mcl == mu XXX.(([not ARRIVED] XXX) and (<true> true))
cadpafarr.mcl == [ true* ] < true* . ARRIVED > true
cadpafarr.mcl == [ true* ] < true> true
> TRUE
>
> real 0m29.648s
117
Ten Diverse FMs for an ATS
118
-- > user 0m28.341s
-- > sys 0m1.078s
-- Evaluator4 Memory 78MB
--
4.2 FDR4
M0
M1
M2
M3
M4
M5
M6
M7
=
=
=
=
=
=
=
=
< 1, 9,10,13,15,20,23>
< 3, 9,10,13,15,20,24>
< 5,27,11,13,16,20,25>
< 7,27,11,13,16,20,26>
<23,22,17,18,11, 9, 2>
<24,22,17,18,11, 9, 4>
<25,22,17,18,12,27, 6>
<26,22,17,18,12,27, 8>
------ region A: train constraints -----A0 = < 0, 0, 0, 1, 0,-1, 0> -- G1
A1 = < 0, 0, 0, 1, 0,-1, 0> -- R1
A2 = < 0, 0, 1,-1, 0, 1, 0> -- Y1
A3 = < 0, 0, 1,-1, 0, 0, 0> -- B1
A4 = < 0, 1, 0, 0,-1, 0, 0> -- G2
A5 = < 0, 1, 0, 0,-1, 0, 0> -- R2
A6 = < 0, 0, 0,-1, 0, 0, 0> -- Y2
A7 = < 0, 1, 0,-1, 0, 0, 0> -- B2
------------------------------------------------ region B: train constraints -----= < 0, 0, 0, 1, 0,-1, 0> -- G1
= < 0, 0, 0, 1, 0,-1, 0> -- R1
= < 0, 0, 1,-1, 0, 0, 0> -- Y1
= < 0, 0, 1,-1, 0, 1, 0> -- B1
= < 0, 1, 0, 0,-1, 0, 0> -- G2
= < 0, 1, 0, 0,-1, 0, 0> -- R2
= < 0, 1, 0,-1, 0, 0, 0> -- Y2
= < 0, 0, 0,-1, 0, 0, 0> -- B2
-----------------------------------------LA = 7
LB = 7
B0
B1
B2
B3
B4
B5
B6
B7
el(y,x) = if x==0 then head(y) else el(tail(y),x-1)
--channel move:{1..27}.{1..27}.{ -1..1}.{ -1..1}
channel move
channel arrived
AllTrains (P0, P1, P2, P3, P4, P5, P6, P7, RA, RB) =
(P0 < 6 and
-- train0 has not yet reached all the steps of its mission
el(T0,P0+1) != el(T1,P1) and -- next place of train0 not occupied by train1
el(T0,P0+1) != el(T2,P2) and -- next place of train0 not occupied by train2
el(T0,P0+1) != el(T3,P3) and
el(T0,P0+1) != el(T4,P4) and
el(T0,P0+1) != el(T5,P5) and
el(T0,P0+1) != el(T6,P6) and
el(T0,P0+1) != el(T7,P7) and -- next place of train0 not occupied by train7
RA + el(A0,P0+1) <= LA and -- progress of train0 does not saturate RA
RB + el(B0,P0+1) <= LB
-- progress of train0 does not saturate RB
) &
move -> AllTrains(P0+1,P1,P2,P3,P4,P5,P6,P7,RA+el(A0,P0+1),RB+el(B0,P0+1))
[]
F. Mazzanti & A. Ferrari
(P1 < 6 and
el(T1,P1+1) != el(T0,P0) and
el(T1,P1+1) != el(T2,P2) and
el(T1,P1+1) != el(T3,P3) and
el(T1,P1+1) != el(T4,P4) and
el(T1,P1+1) != el(T5,P5) and
el(T1,P1+1) != el(T6,P6) and
el(T1,P1+1) != el(T7,P7) and
RA + el(A1,P1+1) <= LA and
RB + el(B1,P1+1) <= LB
) &
move -> AllTrains(P0,P1+1,P2,P3,P4,P5,P6,P7,RA+el(A1,P1+1),RB+el(B1,P1+1))
[]
(P2 < 6 and
el(T2,P2+1) != el(T0,P0) and
el(T2,P2+1) != el(T1,P1) and
el(T2,P2+1) != el(T3,P3) and
el(T2,P2+1) != el(T4,P4) and
el(T2,P2+1) != el(T5,P5) and
el(T2,P2+1) != el(T6,P6) and
el(T2,P2+1) != el(T7,P7) and
RA + el(A2,P2+1) <= LA and
RB + el(B2,P2+1) <= LB
) &
move -> AllTrains(P0,P1,P2+1,P3,P4,P5,P6,P7,RA+el(A2,P2+1),RB+el(B2,P2+1))
[]
(P3 < 6 and
el(T3,P3+1) != el(T0,P0) and
el(T3,P3+1) != el(T1,P1) and
el(T3,P3+1) != el(T2,P2) and
el(T3,P3+1) != el(T4,P4) and
el(T3,P3+1) != el(T5,P5) and
el(T3,P3+1) != el(T6,P6) and
el(T3,P3+1) != el(T7,P7) and
RA + el(A3,P3+1) <= LA and
RB + el(B3,P3+1) <= LB
) &
move -> AllTrains(P0,P1,P2,P3+1,P4,P5,P6,P7,RA+el(A3,P3+1),RB+el(B3,P3+1))
[]
(P4 < 6 and
el(T4,P4+1) != el(T0,P0) and
el(T4,P4+1) != el(T1,P1) and
el(T4,P4+1) != el(T2,P2) and
el(T4,P4+1) != el(T3,P3) and
el(T4,P4+1) != el(T5,P5) and
el(T4,P4+1) != el(T6,P6) and
el(T4,P4+1) != el(T7,P7) and
RA + el(A4,P4+1) <= LA and
RB + el(B4,P4+1) <= LB
) &
move -> AllTrains(P0,P1,P2,P3,P4+1,P5,P6,P7,RA+el(A4,P4+1),RB+el(B4,P4+1))
[]
(P5 < 6 and
el(T5,P5+1) != el(T0,P0) and
el(T5,P5+1) != el(T1,P1) and
el(T5,P5+1) != el(T2,P2) and
el(T5,P5+1) != el(T3,P3) and
el(T5,P5+1) != el(T4,P4) and
el(T5,P5+1) != el(T6,P6) and
el(T5,P5+1) != el(T7,P7) and
RA + el(A5,P5+1) <= LA and
RB + el(B5,P5+1) <= LB
) &
move -> AllTrains(P0,P1,P2,P3,P4,P5+1,P6,P7,RA+el(A5,P5+1),RB+el(B5,P5+1))
[]
119
Ten Diverse FMs for an ATS
120
(P6 < 6 and
el(T6,P6+1) != el(T0,P0) and
el(T6,P6+1) != el(T1,P1) and
el(T6,P6+1) != el(T2,P2) and
el(T6,P6+1) != el(T3,P3) and
el(T6,P6+1) != el(T4,P4) and
el(T6,P6+1) != el(T5,P5) and
el(T6,P6+1) != el(T7,P7) and
RA + el(A6,P6+1) <= LA and
RB + el(B6,P6+1) <= LB
) &
move -> AllTrains(P0,P1,P2,P3,P4,P5,P6+1,P7,RA+el(A6,P6+1),RB+el(B6,P6+1))
[]
(P7 < 6 and
el(T7,P7+1) != el(T0,P0) and
el(T7,P7+1) != el(T1,P1) and
el(T7,P7+1) != el(T2,P2) and
el(T7,P7+1) != el(T3,P3) and
el(T7,P7+1) != el(T4,P4) and
el(T7,P7+1) != el(T5,P5) and
el(T7,P7+1) != el(T6,P6) and
RA + el(A7,P7+1) <= LA and
RB + el(B7,P7+1) <= LB
) &
move -> AllTrains(P0,P1,P2,P3,P4,P5,P6,P7+1,RA+el(A7,P7+1),RB+el(B7,P7+1))
[]
((P0 == 6) and (P1 ==6) and (P2 ==6) and (P3 ==6) and
(P4 ==6) and (P5 ==6) and (P6 ==6) and (P7 ==6)
) &
arrived -> STOP
-------------------------ASYS = AllTrains(0,0,0,0,0,0,0,0, 1,1)\{move}
--------------------------- compression is helpful for two verifications/visualization
-- NSYS = normal(ASYS)
-- assert SPEC [FD= NSYS
--------------------------
-------------------------SPEC = arrived -> STOP
-------------------------assert SPEC [FD= ASYS
-- -------- verfication process : ---------------- time refines --refinement-storage-file-path swapdir fdr4_oneway8seq.txt
--
5
mCRL2
%-------------------------------------------------------------% T0 := [ 1, 9,10,13,15,20,23,22]; -- G1
% T1 := [ 3, 9,10,13,15,20,24,22]; -- R1
% T2 := [ 5,27,11,13,16,20,25,22]; -- Y1
% T3 := [ 7,27,11,13,16,20,26,22]; -- B1
% T4 := [23,22,17,18,11, 9, 2, 1]; -- G2
% T5 := [24,22,17,18,11, 9, 4, 3]; -- R2
F. Mazzanti & A. Ferrari
121
% T6 := [25,22,17,18,12,27, 6, 5]; -- Y2
% T7 := [26,22,17,18,12,27, 8, 7]; -- B2
%-------------------------------------------------------------map T0: Nat -> Nat;
eqn T0(0)= 1; T0(1)= 9; T0(2)=10; T0( 3)=13; T0( 4)=15; T0( 5)=20; T0( 6)=23;
map T1: Nat -> Nat;
eqn T1(0)= 3; T1(1)=9;
T1(2)=10; T1( 3)=13; T1( 4)=15; T1( 5)=20; T1( 6)=24;
map T2: Nat -> Nat;
eqn T2(0)= 5; T2(1)=27; T2(2)=11; T2( 3)=13; T2( 4)=16; T2( 5)=20; T2( 6)=25;
map T3: Nat -> Nat;
eqn T3(0)= 7; T3(1)=27; T3(2)=11; T3( 3)=13; T3( 4)=16; T3( 5)=20; T3( 6)=26;
map T4: Nat -> Nat;
eqn T4(0)=23; T4(1)=22; T4(2)=17; T4( 3)=18; T4( 4)=11; T4( 5)= 9; T4( 6)= 2;
map T5: Nat -> Nat;
eqn T5(0)=24; T5(1)=22; T5(2)=17; T5( 3)=18; T5( 4)=11; T5( 5)=9;
T5( 6)= 4;
map T6: Nat -> Nat;
eqn T6(0)=25; T6(1)=22; T6(2)=17; T6(3)=18; T6(4)=12; T6(5)=27; T6(6)=6;
T6(7)= 5; T6(8)=27; T6(9)=11; T6(10)=13; T6(11)=16; T6(12)=20; T6(13)=25;
map T7: Nat -> Nat;
eqn T7(0)=26; T7(1)=22; T7(2)=17; T7( 3)=18; T7( 4)=12; T7( 5)=27; T7( 6)= 8;
% ------ region A: train constraints -----%
0 1 2 3 4 5 6
% A0 := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
% A1 := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
% A2 := [ 0, 0, 1,-1, 0, 1, 0]; -- Y1
% A3 := [ 0, 0, 1,-1, 0, 0, 0]; -- B1
% A4 := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
% A5 := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
% A6 := [ 0, 0, 0,-1, 0, 0, 0]; -- Y2
% A7 := [ 0, 1, 0,-1, 0, 0, 0]; -- B2
% -----------------------------------------map LA: Nat; % limit for region A
eqn LA = 7;
map A0: Nat -> Int;
eqn A0(0)=0; A0(1)=0; A0(2)=0; A0( 3)= 1; A0( 4)=0; A0( 5)=-1; A0( 6)=0;
map A1: Nat -> Int;
eqn A1(0)=0; A1(1)=0; A1(2)=0; A1( 3)= 1; A1( 4)=0; A1( 5)=-1; A1( 6)=0;
map A2: Nat -> Int;
eqn A2(0)=0; A2(1)=0; A2(2)= 1; A2( 3)=-1; A2( 4)=0; A2( 5)= 1; A2( 6)=0;
map A3: Nat -> Int;
eqn A3(0)=0; A3(1)=0; A3(2)= 1; A3( 3)=-1; A3( 4)=0; A3( 5)= 0; A3( 6)=0;
map A4: Nat -> Int;
eqn A4(0)=0; A4(1)=1; A4(2)=0; A4( 3)=0; A4( 4)=-1; A4( 5)= 0; A4( 6)=0;
map A5: Nat -> Int;
eqn A5(0)=0; A5(1)=1; A5(2)=0; A5( 3)=0; A5( 4)=-1; A5( 5)= 0; A5( 6)=0;
map A6: Nat -> Int;
eqn A6(0)=0; A6(1)=0; A6(2)=0; A6( 3)=-1; A6( 4)=0; A6( 5)= 0; A6( 6)=0;
map A7: Nat -> Int;
Ten Diverse FMs for an ATS
122
eqn A7(0)=0; A7(1)=1; A7(2)=0; A7( 3)=-1; A7( 4)=0; A7( 5)= 0; A7( 6)=0;
% ------- region B: train constraints -----%
0 1 2 3 4 5 6
% B0 := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
% B1 := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
% B2 := [ 0, 0, 1,-1, 0, 0, 0]; -- Y1
% B3 := [ 0, 0, 1,-1, 0, 1, 0]; -- B1
% B4 := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
% B5 := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
% B6 := [ 0, 1, 0,-1, 0, 0, 0]; -- Y2
% B7 := [ 0, 0, 0,-1, 0, 0, 0]; -- B2
% -----------------------------------------map LB: Nat; % limit for region B
eqn LB = 7;
map B0: Nat -> Int;
eqn B0(0)=0; B0(1)=0; B0(2)=0; B0( 3)= 1; B0( 4)=0; B0( 5)=-1; B0( 6)=0;
map B1: Nat -> Int;
eqn B1(0)=0; B1(1)=0; B1(2)=0; B1( 3)= 1; B1( 4)=0; B1( 5)=-1; B1( 6)=0;
map B2: Nat -> Int;
eqn B2(0)=0; B2(1)=0; B2(2)= 1; B2( 3)=-1; B2( 4)=0; B2( 5)= 0; B2( 6)=0;
map B3: Nat -> Int;
eqn B3(0)=0; B3(1)=0; B3(2)= 1; B3( 3)=-1; B3( 4)=0; B3( 5)= 1; B3( 6)=0;
map B4: Nat -> Int;
eqn B4(0)=0; B4(1)=1; B4(2)=0; B4( 3)=-0; B4( 4)=-1; B4( 5)= 0; B4( 6)=0;
map B5: Nat -> Int;
eqn B5(0)=0; B5(1)=1; B5(2)=0; B5( 3)=0; B5( 4)=-1; B5( 5)= 0; B5( 6)=0;
map B6: Nat -> Int;
eqn B6(0)=0; B6(1)=1; B6(2)=0; B6( 3)=-1; B6( 4)=0; B6( 5)= 0; B6( 6)=0;
map B7: Nat -> Int;
eqn B7(0)=0; B7(1)=0; B7(2)=0; B7( 3)=-1; B7( 4)=0; B7( 5)= 0; B7( 6)=0;
act
arrived;
move: Nat;
proc AllTrains(
P0:Nat,P1:Nat,P2:Nat,P3:Nat,P4:Nat,P5:Nat,P6:Nat,P7:Nat,RA: Int,RB: Int) =
(P0 < 6 &&
% train0 has not yet reached all the steps of its mission
T0(P0+1) != T1(P1) && % next place of train0 not occupied by train1
T0(P0+1) != T2(P2) && % next place of train0 not occupied by train2
T0(P0+1) != T3(P3) &&
T0(P0+1) != T4(P4) &&
T0(P0+1) != T5(P5) &&
T0(P0+1) != T6(P6) &&
T0(P0+1) != T7(P7) && % next place of train0 not occupied by train7
RA + A0(P0+1) <= LA && % progress of train0 does not saturate RA
RB + B0(P0+1) <= LB
% progress of train0 does not saturate RB
) ->
move(0). AllTrains(P0+1,P1,P2,P3,P4,P5,P6,P7,RA+A0(P0+1),RB+B0(P0+1))
+
(P1 < 6 &&
T1(P1+1) != T0(P0) &&
T1(P1+1) != T2(P2) &&
T1(P1+1) != T3(P3) &&
T1(P1+1) != T4(P4) &&
T1(P1+1) != T5(P5) &&
T1(P1+1) != T6(P6) &&
F. Mazzanti & A. Ferrari
T1(P1+1) != T7(P7) &&
RA + A1(P1+1) <= LA &&
RB + B1(P1+1) <= LB
) ->
move(1). AllTrains(P0,P1+1,P2,P3,P4,P5,P6,P7,RA+A1(P1+1),RB+B1(P1+1))
+
(P2 < 6 &&
T2(P2+1) != T0(P0) &&
T2(P2+1) != T1(P1) &&
T2(P2+1) != T3(P3) &&
T2(P2+1) != T4(P4) &&
T2(P2+1) != T5(P5) &&
T2(P2+1) != T6(P6) &&
T2(P2+1) != T7(P7) &&
RA + A2(P2+1) <= LA &&
RB + B2(P2+1) <= LB
) ->
move(2). AllTrains(P0,P1,P2+1,P3,P4,P5,P6,P7,RA+A2(P2+1),RB+B2(P2+1))
+
(P3 < 6 &&
T3(P3+1) != T0(P0) &&
T3(P3+1) != T1(P1) &&
T3(P3+1) != T2(P2) &&
T3(P3+1) != T4(P4) &&
T3(P3+1) != T5(P5) &&
T3(P3+1) != T6(P6) &&
T3(P3+1) != T7(P7) &&
RA + A3(P3+1) <= LA &&
RB + B3(P3+1) <= LB
) ->
move(3). AllTrains(P0,P1,P2,P3+1,P4,P5,P6,P7,RA+A3(P3+1),RB+B3(P3+1))
+
(P4 < 6 &&
T4(P4+1) != T0(P0) &&
T4(P4+1) != T1(P1) &&
T4(P4+1) != T2(P2) &&
T4(P4+1) != T3(P3) &&
T4(P4+1) != T5(P5) &&
T4(P4+1) != T6(P6) &&
T4(P4+1) != T7(P7) &&
RA + A4(P4+1) <= LA &&
RB + B4(P4+1) <= LB
) ->
move(4). AllTrains(P0,P1,P2,P3,P4+1,P5,P6,P7,RA+A4(P4+1),RB+B4(P4+1))
+
(P5 < 6 &&
T5(P5+1) != T0(P0) &&
T5(P5+1) != T1(P1) &&
T5(P5+1) != T2(P2) &&
T5(P5+1) != T3(P3) &&
T5(P5+1) != T4(P4) &&
T5(P5+1) != T6(P6) &&
T5(P5+1) != T7(P7) &&
RA + A5(P5+1) <= LA &&
RB + B5(P5+1) <= LB
) ->
move(5). AllTrains(P0,P1,P2,P3,P4,P5+1,P6,P7,RA+A5(P5+1),RB+B5(P5+1))
+
(P6 < 6 &&
T6(P6+1) != T0(P0) &&
T6(P6+1) != T1(P1) &&
T6(P6+1) != T2(P2) &&
T6(P6+1) != T3(P3) &&
T6(P6+1) != T4(P4) &&
T6(P6+1) != T5(P5) &&
123
Ten Diverse FMs for an ATS
124
T6(P6+1) != T7(P7) &&
RA + A6(P6+1) <= LA &&
RB + B6(P6+1) <= LB
) ->
move(6). AllTrains(P0,P1,P2,P3,P4,P5,P6+1,P7,RA+A6(P6+1),RB+B6(P6+1))
+
(P7 < 6 &&
T7(P7+1) != T0(P0) &&
T7(P7+1) != T1(P1) &&
T7(P7+1) != T2(P2) &&
T7(P7+1) != T3(P3) &&
T7(P7+1) != T4(P4) &&
T7(P7+1) != T5(P5) &&
T7(P7+1) != T6(P6) &&
RA + A7(P7+1) <= LA &&
RB + B7(P7+1) <= LB
) ->
move(7). AllTrains(P0,P1,P2,P3,P4,P5,P6,P7+1,RA+A7(P7+1),RB+B7(P7+1))
+
((P0 ==6) && (P1 ==6) && (P2 ==6) && (P3 ==6) &&
(P4 ==6) && (P5 ==6) && (P6 ==6) && (P7 ==6)) ->
arrived . AllTrains(P0,P1,P2,P3,P4,P5,P6,P7,RA,RB);
init
AllTrains(0,0,0,0,0,0,0,0, 1,1);
%%%%%%%%%%% verfication process : %%%%%%%%%%%%%%%%%
%
% mcrl22lps mcrl2_oneway8seq.txt temp.lps
% lps2pbes -fformula.mcf temp.lps temp.pbes
%
formula.mcf= "mu X.(([!arrived]X) && (<true> true))" == AF {arrived}
% time pbes2bool -s2 -vrjittyc temp.pbes
% > real 1m53.996s
% > user 1m52.243s
% > sys
0m1.583s
% > USED VIRTUAL MEMORY 1.06 GB
%
% time pbes2bool -s2 temp.pbes
% > real 19m47.753s
%
% lps2lts -rjittyc --verbose temp.lps temp.lts
% ltsinfo temp.lts
% > state space stats: (49 levels, 1_636_545 states and 7_134_233 transitions)
% >
% >
%
% formula.mcf= "mu X.(([!arrived]X) && (<true> true))" == AF {arrived}
% formula.mcf= "[ true* ] < true* . ARRIVED > true"
== AG EF {arrived}
% pbes2bool -c print evidence on validity of formula, nedded by lpsxsim for
%
% lps2lts --verbose âĂŞrjittyc âĂŞcached -D -F temp.lps temp.lts
%
( -sdepth -sBreadth --suppress --todo=maxnum --no-info
%
% the generation of counter-examples
% pbes2bool -zb
breadth-first
% pbes2bool -zd
depth-first
% pbes2bool --todo -strategy
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
F. Mazzanti & A. Ferrari
6
125
ProB
/* https://www3.hhu.de/stups/prob/index.php/Summary_of_B_Syntax */
MACHINE Oneway
DEFINITIONS
SET_PREF_MAXINT == 1;
SET_PREF_MININT == 0
CONSTANTS
T0,T1,T2,T3,T4,T5,T6,T7,
A0,A1,A2,A3,A4,A5,A6,A7,
B0,B1,B2,B3,B4,B5,B6,B7,
LA,LB
PROPERTIES
/*
------- train missions -------------T0: int[] := [ 1, 9,10,13,15,20,23]; -T1: int[] := [ 3, 9,10,13,15,20,24]; -T2: int[] := [ 5,27,11,13,16,20,25]; -T3: int[] := [ 7,27,11,13,16,20,26]; -T4: int[] := [23,22,17,18,11, 9, 2]; -T5: int[] := [24,22,17,18,11, 9, 4]; -T6: int[] := [25,22,17,18,12,27, 6]; -T7: int[] := [26,22,17,18,12,27, 8]; --
G1
R1
Y1
B1
G2
R2
Y2
B2
------ region A: train constraints -----A0: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
A1: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
A2: int[] := [ 0, 0, 1,-1, 0, 1, 0]; -- Y1
A3: int[] := [ 0, 0, 1,-1, 0, 0, 0]; -- B1
A4: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
A5: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
A6: int[] := [ 0, 0, 0,-1, 0, 0, 0]; -- Y2
A7: int[] := [ 0, 1, 0,-1, 0, 0, 0]; -- B2
------- region B: train constraints -----B0: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
B1: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
B2: int[] := [ 0, 0, 1,-1, 0, 0, 0]; -- Y1
B3: int[] := [ 0, 0, 1,-1, 0, 1, 0]; -- B1
B4: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
B5: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
B6: int[] := [ 0, 1, 0,-1, 0, 0, 0]; -- Y2
B7: int[] := [ 0, 0, 0,-1, 0, 0, 0]; -- B2
-----------------------------------------*/
T0
T1
T2
T3
T4
T5
T6
T7
:
:
:
:
:
:
:
:
0..6
0..6
0..6
0..6
0..6
0..6
0..6
0..6
-->
-->
-->
-->
-->
-->
-->
-->
1..27
1..27
1..27
1..27
1..27
1..27
1..27
1..27
&
&
&
&
&
&
&
&
A0
A1
A2
A3
A4
A5
:
:
:
:
:
:
0..6
0..6
0..6
0..6
0..6
0..6
-->
-->
-->
-->
-->
-->
-1..1
-1..1
-1..1
-1..1
-1..1
-1..1
&
&
&
&
&
&
Ten Diverse FMs for an ATS
126
A6 : 0..6 --> -1..1 &
A7 : 0..6 --> -1..1 &
B0
B1
B2
B3
B4
B5
B6
B7
:
:
:
:
:
:
:
:
0..6
0..6
0..6
0..6
0..6
0..6
0..6
0..6
T0(0)= 1
T1(0)= 3
T2(0)= 5
T3(0)= 7
T4(0)=23
T5(0)=24
T6(0)=25
T7(0)=26
-->
-->
-->
-->
-->
-->
-->
-->
&
&
&
&
&
&
&
&
-1..1
-1..1
-1..1
-1..1
-1..1
-1..1
-1..1
-1..1
&
&
&
&
&
&
&
&
T0(1)= 9
T1(1)= 9
T2(1)=27
T3(1)=27
T4(1)=22
T5(1)=22
T6(1)=22
T7(1)=22
&
&
&
&
&
&
&
&
T0(2)=10
T1(2)=10
T2(2)=11
T3(2)=11
T4(2)=17
T5(2)=17
T6(2)=17
T7(2)=17
&
&
&
&
&
&
&
&
T0(3)=13
T1(3)=13
T2(3)=13
T3(3)=13
T4(3)=18
T5(3)=18
T6(3)=18
T7(3)=18
&
&
&
&
&
&
&
&
T0(4)=15
T1(4)=15
T2(4)=16
T3(4)=16
T4(4)=11
T5(4)=11
T6(4)=12
T7(4)=12
&
&
&
&
&
&
&
&
T0(5)=20
T1(5)=20
T2(5)=20
T3(5)=20
T4(5)= 9
T5(5)= 9
T6(5)=27
T7(5)=27
&
&
&
&
&
&
&
&
T0(6)=23 &
T1(6)=24 &
T2(6)=25 &
T3(6)=26 &
T4(6)=2 &
T5(6)=4 &
T6(6)=6 &
T7(6)=8 &
A0(0)=0
A1(0)=0
A2(0)=0
A3(0)=0
A4(0)=0
A5(0)=0
A6(0)=0
A7(0)=0
&
&
&
&
&
&
&
&
A0(1)=0
A1(1)=0
A2(1)=0
A3(1)=0
A4(1)=1
A5(1)=1
A6(1)=0
A7(1)=1
&
&
&
&
&
&
&
&
A0(2)=0
A1(2)=0
A2(2)=1
A3(2)=1
A4(2)=0
A5(2)=0
A6(2)=0
A7(2)=0
&
&
&
&
&
&
&
&
A0(3)= 1
A1(3)= 1
A2(3)=-1
A3(3)=-1
A4(3)= 0
A5(3)= 0
A6(3)=-1
A7(3)=-1
&
&
&
&
&
&
&
&
A0(4)= 0
A1(4)= 0
A2(4)= 0
A3(4)= 0
A4(4)=-1
A5(4)=-1
A6(4)= 0
A7(4)= 0
&
&
&
&
&
&
&
&
A0(5)=-1
A1(5)=-1
A2(5)= 1
A3(5)= 0
A4(5)= 0
A5(5)= 0
A6(5)= 0
A7(5)= 0
&
&
&
&
&
&
&
&
A0(
A1(
A2(
A3(
A4(
A5(
A6(
A7(
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
&
&
&
&
&
&
&
&
B0(0)=0
B1(0)=0
B2(0)=0
B3(0)=0
B4(0)=0
B5(0)=0
B6(0)=0
B7(0)=0
&
&
&
&
&
&
&
&
B0(1)=0
B1(1)=0
B2(1)=0
B3(1)=0
B4(1)=1
B5(1)=1
B6(1)=1
B7(1)=0
&
&
&
&
&
&
&
&
B0(2)=0
B1(2)=0
B2(2)=1
B3(2)=1
B4(2)=0
B5(2)=0
B6(2)=0
B7(2)=0
&
&
&
&
&
&
&
&
B0(3)= 1
B1(3)= 1
B2(3)=-1
B3(3)=-1
B4(3)=-0
B5(3)= 0
B6(3)=-1
B7(3)=-1
&
&
&
&
&
&
&
&
B0(4)= 0
B1(4)= 0
B2(4)= 0
B3(4)= 0
B4(4)=-1
B5(4)=-1
B6(4)= 0
B7(4)= 0
&
&
&
&
&
&
&
&
B0(5)=-1
B1(5)=-1
B2(5)= 0
B3(5)= 1
B4(5)= 0
B5(5)= 0
B6(5)= 0
B7(5)= 0
&
&
&
&
&
&
&
&
B0(
B1(
B2(
B3(
B4(
B5(
B6(
B7(
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
6)=0
&
&
&
&
&
&
&
&
LA=7 & LB=7
VARIABLES
P0,P1,P2,P3,P4,P5,P6,P7,RA,RB
INVARIANT
P0:0..6 & P1:0..6 & P2:0..6 & P3:0..6 & P4:0..6 & P5:0..6 & P6:0..6 & P7:0..6 &
RA:0..8 & RB:0..8
INITIALISATION
P0:=0; P1:=0; P2:=0; P3:=0; P4:=0; P5:=0; P6:=0; P7:=0; RA:=1; RB:=1
OPERATIONS
move0 =
PRE P0<6 &
T0(P0+1) /= T1(P1) &
T0(P0+1) /= T2(P2 )&
T0(P0+1) /= T3(P3 )&
T0(P0+1) /= T4(P4) &
T0(P0+1) /= T5(P5) &
T0(P0+1) /= T6(P6) &
T0(P0+1) /= T7(P7) &
RA + A0(P0+1) <= LA &
RB + B0(P0+1) <= LB
THEN
P0 := P0+1;
F. Mazzanti & A. Ferrari
RA := RA + A0(P0);
RB := RB + B0(P0)
END ;
move1 =
PRE P1<6 &
T1(P1+1) /= T0(P0) &
T1(P1+1) /= T2(P2) &
T1(P1+1) /= T3(P3) &
T1(P1+1) /= T4(P4) &
T1(P1+1) /= T5(P5) &
T1(P1+1) /= T6(P6) &
T1(P1+1) /= T7(P7) &
RA + A1(P1+1) <= LA &
RB + B1(P1+1) <= LB
THEN
P1 := P1+1;
RA := RA + A1(P1);
RB := RB + B1(P1)
END ;
move2 =
PRE P2<6 &
T2(P2+1) /= T0(P0) &
T2(P2+1) /= T1(P1) &
T2(P2+1) /= T3(P3) &
T2(P2+1) /= T4(P4) &
T2(P2+1) /= T5(P5) &
T2(P2+1) /= T6(P6) &
T2(P2+1) /= T7(P7) &
RA + A2(P2+1) <= LA &
RB + B2(P2+1) <= LB
THEN
P2 := P2+1;
RA := RA + A2(P2);
RB := RB + B2(P2)
END ;
move3 =
PRE P3<6 &
T3(P3+1) /= T0(P0) &
T3(P3+1) /= T1(P1) &
T3(P3+1) /= T2(P2) &
T3(P3+1) /= T4(P4) &
T3(P3+1) /= T5(P5) &
T3(P3+1) /= T6(P6) &
T3(P3+1) /= T7(P7) &
RA + A3(P3+1) <= LA &
RB + B3(P3+1) <= LB
THEN
P3 := P3+1;
RA := RA + A3(P3);
RB := RB + B3(P3)
END ;
move4 =
PRE P4<6 &
T4(P4+1) /= T0(P0) &
T4(P4+1) /= T1(P1) &
T4(P4+1) /= T2(P2) &
T4(P4+1) /= T3(P3) &
T4(P4+1) /= T5(P5) &
T4(P4+1) /= T6(P6) &
T4(P4+1) /= T7(P7) &
RA + A4(P4+1) <= LA &
127
Ten Diverse FMs for an ATS
128
RB + B4(P4+1) <= LB
THEN
P4 := P4+1;
RA := RA + A4(P4);
RB := RB + B4(P4)
END ;
move5 =
PRE P5<6 &
T5(P5+1) /= T0(P0) &
T5(P5+1) /= T1(P1) &
T5(P5+1) /= T2(P2) &
T5(P5+1) /= T3(P3) &
T5(P5+1) /= T4(P4) &
T5(P5+1) /= T6(P6) &
T5(P5+1) /= T7(P7) &
RA + A5(P5+1) <= LA &
RB + B5(P5+1) <= LB
THEN
P5 := P5+1;
RA := RA + A5(P5);
RB := RB + B5(P5)
END ;
move6 =
PRE P6<6 &
T6(P6+1) /= T0(P0) &
T6(P6+1) /= T1(P1) &
T6(P6+1) /= T2(P2) &
T6(P6+1) /= T3(P3) &
T6(P6+1) /= T4(P4) &
T6(P6+1) /= T5(P5) &
T6(P6+1) /= T7(P7) &
RA + A6(P6+1) <= LA &
RB + B6(P6+1) <= LB
THEN
P6 := P6+1;
RA := RA + A6(P6);
RB := RB + B6(P6)
END ;
move7 =
PRE P7<6 &
T7(P7+1) /= T0(P0) &
T7(P7+1) /= T1(P1) &
T7(P7+1) /= T2(P2) &
T7(P7+1) /= T3(P3) &
T7(P7+1) /= T4(P4) &
T7(P7+1) /= T5(P5) &
T7(P7+1) /= T6(P6) &
RA + A7(P7+1) <= LA &
RB + B7(P7+1) <= LB
THEN
P7 := P7+1;
RA := RA + A7(P7);
RB := RB + B7(P7)
END ;
arrived =
PRE
P0=6 & P1=6 & P2=6 & P3=6 & P4=6 & P5=6 & P6=6 & P7=6
THEN
skip
END
END
F. Mazzanti & A. Ferrari
//-------------------// SEARCHING DEADLOCKS:
//-------------------
7
129
1_636_547 states,
7_134_235 trans.
TIME
NuSMV/nuXmv
MODULE main
------- train missions -----------------DEFINE
T0 := [ 1, 9,10,13,15,20,23]; -- G1
T1 := [ 3, 9,10,13,15,20,24]; -- R1
T2 := [ 5,27,11,13,16,20,25]; -- Y1
T3 := [ 7,27,11,13,16,20,26]; -- B1
T4 := [23,22,17,18,11, 9, 2]; -- G2
T5 := [24,22,17,18,11, 9, 4]; -- R2
T6 := [25,22,17,18,12,27, 6]; -- Y2
T7 := [26,22,17,18,12,27, 8]; -- B2
------ region A: train constraints -----A0 := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
A1 := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
A2 := [ 0, 0, 1,-1, 0, 1, 0]; -- Y1
A3 := [ 0, 0, 1,-1, 0, 0, 0]; -- B1
A4 := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
A5 := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
A6 := [ 0, 0, 0,-1, 0, 0, 0]; -- Y2
A7 := [ 0, 1, 0,-1, 0, 0, 0]; -- B2
------------------------------------------
B0
B1
B2
B3
B4
B5
B6
B7
------- region B: train constraints -----:= [ 0, 0, 0, 1, 0,-1, 0]; -- G1
:= [ 0, 0, 0, 1, 0,-1, 0]; -- R1
:= [ 0, 0, 1,-1, 0, 0, 0]; -- Y1
:= [ 0, 0, 1,-1, 0, 1, 0]; -- B1
:= [ 0, 1, 0, 0,-1, 0, 0]; -- G2
:= [ 0, 1, 0, 0,-1, 0, 0]; -- R2
:= [ 0, 1, 0,-1, 0, 0, 0]; -- Y2
:= [ 0, 0, 0,-1, 0, 0, 0]; -- B2
------------------------------------------
LA := 7;
LB := 7;
IVAR
-- (unfair) selector of the train transition
RUNNING: {0,1,2,3,4,5,6,7};
VAR
-- vector of train progesses in the execution of their missions
P0: 0..6;
P1: 0..6;
P2: 0..6;
P3: 0..6;
P4: 0..6;
P5: 0..6;
P6: 0..6;
P7: 0..6;
---- the occupation status for regions A and B
32 min
VMEM
3 GB
Ten Diverse FMs for an ATS
130
RA:
RB:
0..8;
0..8;
ASSIGN
-- the initial vector of train progesses
init(P0) := 0;
init(P1) := 0;
init(P2) := 0;
init(P3) := 0;
init(P4) := 0;
init(P5) := 0;
init(P6) := 0;
init(P7) := 0;
--- the initial occupation status for regions A and B
init(RA) := 1;
init(RB) := 1;
TRANS
-- progression rules for the evolving train 0
RUNNING =0 &
-- the current train has not yet completed its mission
P0 < 6 &
--- the next place is not occupied by other trains
T0[P0+1] != T1[P1] &
T0[P0+1] != T2[P2] &
T0[P0+1] != T3[P3] &
T0[P0+1] != T4[P4] &
T0[P0+1] != T5[P5] &
T0[P0+1] != T6[P6] &
T0[P0+1] != T7[P7] &
--- the progression step of id satisfies all contraints
RA + A0[P0+1] <= LA &
RB + B0[P0+1] <= LB
?
next(P0) in (P0+1) &
next(P1) in P1 &
next(P2) in P2 &
next(P3) in P3 &
next(P4) in P4 &
next(P5) in P5 &
next(P6) in P6 &
next(P7) in P7 &
next(RA) in (RA + A0[P0+1]) &
next(RB) in (RB + B0[P0+1])
:
RUNNING = 1 &
P1 < 6 &
T1[P1+1] != T0[P0] &
T1[P1+1] != T2[P2] &
T1[P1+1] != T3[P3] &
T1[P1+1] != T4[P4] &
T1[P1+1] != T5[P5] &
T1[P1+1] != T6[P6] &
T1[P1+1] != T7[P7] &
RA + A1[P1+1] <= LA &
RB + B1[P1+1] <= LB
?
next(P0) in P0 &
next(P1) in (P1+1) &
next(P2) in P2 &
next(P3) in P3 &
F. Mazzanti & A. Ferrari
next(P4)
next(P5)
next(P6)
next(P7)
next(RA)
next(RB)
in
in
in
in
in
in
131
P4 &
P5 &
P6 &
P7 &
(RA + A1[P1+1]) &
(RB + B1[P1+1])
:
RUNNING =2 &
P2 < 6 &
T2[P2+1] != T0[P0] &
T2[P2+1] != T1[P1] &
T2[P2+1] != T3[P3] &
T2[P2+1] != T4[P4] &
T2[P2+1] != T5[P5] &
T2[P2+1] != T6[P6] &
T2[P2+1] != T7[P7] &
RA + A2[P2+1] <= LA &
RB + B2[P2+1] <= LB
?
next(P0) in P0 &
next(P1) in P1 &
next(P2) in (P2+1) &
next(P3) in P3 &
next(P4) in P4 &
next(P5) in P5 &
next(P6) in P6 &
next(P7) in P7 &
next(RA) in (RA + A2[P2+1]) &
next(RB) in (RB + B2[P2+1])
:
RUNNING =3 &
P3 < 6 &
T3[P3+1] != T0[P0] &
T3[P3+1] != T1[P1] &
T3[P3+1] != T2[P2] &
T3[P3+1] != T3[P3] &
T3[P3+1] != T4[P4] &
T3[P3+1] != T5[P5] &
T3[P3+1] != T6[P6] &
T3[P3+1] != T7[P7] &
RA + A3[P3+1] <= LA &
RB + B3[P3+1] <= LB
?
next(P0) in P0 &
next(P1) in P1 &
next(P2) in P2 &
next(P3) in (P3+1) &
next(P4) in P4 &
next(P5) in P5 &
next(P6) in P6 &
next(P7) in P7 &
next(RA) in (RA + A3[P3+1]) &
next(RB) in (RB + B3[P3+1])
:
RUNNING =4 &
P4 < 6 &
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T4[P4+1] !=
T0[P0]
T1[P1]
T2[P2]
T3[P3]
T4[P4]
T5[P5]
T6[P6]
T7[P7]
&
&
&
&
&
&
&
&
Ten Diverse FMs for an ATS
132
RA + A4[P4+1] <= LA &
RB + B4[P4+1] <= LB
?
next(P0)
next(P1)
next(P2)
next(P3)
next(P4)
next(P5)
next(P6)
next(P7)
next(RA)
next(RB)
in
in
in
in
in
in
in
in
in
in
P0 &
P1 &
P2 &
P3 &
(P4+1) &
P5 &
P6 &
P7 &
(RA + A4[P4+1]) &
(RB + B4[P4+1])
:
RUNNING =5 &
P5 < 6 &
T5[P5+1] != T0[P0] &
T5[P5+1] != T1[P1] &
T5[P5+1] != T2[P2] &
T5[P5+1] != T3[P3] &
T5[P5+1] != T4[P4] &
T5[P5+1] != T6[P6] &
T5[P5+1] != T7[P7] &
RA + A5[P5+1] <= LA &
RB + B5[P5+1] <= LB
?
next(P0) in P0 &
next(P1) in P1 &
next(P2) in P2 &
next(P3) in P3 &
next(P4) in P4 &
next(P5) in (P5+1) &
next(P6) in P6 &
next(P7) in P7 &
next(RA) in (RA + A5[P5+1]) &
next(RB) in (RB + B5[P5+1])
:
RUNNING = 6 &
P6 < 6 &
T6[P6+1] != T0[P0] &
T6[P6+1] != T1[P1] &
T6[P6+1] != T2[P2] &
T6[P6+1] != T3[P3] &
T6[P6+1] != T4[P4] &
T6[P6+1] != T5[P5] &
T6[P6+1] != T6[P6] &
T6[P6+1] != T7[P7] &
RA + A6[P6+1] <= LA &
RB + B6[P6+1] <= LB
?
next(P0) in P0 &
next(P1) in P1 &
next(P2) in P2 &
next(P3) in P3 &
next(P4) in P4 &
next(P5) in P5 &
next(P6) in (P6+1) &
next(P7) in P7 &
next(RA) in (RA + A6[P6+1]) &
next(RB) in (RB + B6[P6+1])
:
RUNNING =7 &
-- the current train has not yet completed its mission
P7 < 6 &
--
F. Mazzanti & A. Ferrari
133
-- the next place is not occupied by other trains
T7[P7+1] != T0[P0] &
T7[P7+1] != T1[P1] &
T7[P7+1] != T2[P2] &
T7[P7+1] != T3[P3] &
T7[P7+1] != T4[P4] &
T7[P7+1] != T5[P5] &
T7[P7+1] != T6[P6] &
T7[P7+1] != T7[P7] &
--- the progression step of id satisfies all contraints
RA + A7[P7+1] <= LA &
RB + B7[P7+1] <= LB
?
next(P0)
next(P1)
next(P2)
next(P3)
next(P4)
next(P5)
next(P6)
next(P7)
next(RA)
next(RB)
in
in
in
in
in
in
in
in
in
in
P0 &
P1 &
P2 &
P3 &
P4 &
P5 &
P6 &
(P7+1) &
(RA + A7[P7+1]) &
(RB + B7[P7+1])
next(P0)
next(P1)
next(P2)
next(P3)
next(P4)
next(P5)
next(P6)
next(P7)
next(RA)
next(RB)
in
in
in
in
in
in
in
in
in
in
P0
P1
P2
P3
P4
P5
P6
P7
RA
RB
:
---------
FAIRNESS
FAIRNESS
FAIRNESS
FAIRNESS
FAIRNESS
FAIRNESS
FAIRNESS
FAIRNESS
RUNNING
RUNNING
RUNNING
RUNNING
RUNNING
RUNNING
RUNNING
RUNNING
=
=
=
=
=
=
=
=
&
&
&
&
&
&
&
&
&
0;
1;
2;
3;
4;
5;
6;
7;
-- CTLSPEC
-AF ((P0=6) & (P1=6) & (P2=6) & (P3=6) &
-- LTLSPEC
-F ((P0=6) & (P1=6) & (P2=6) & (P3=6) &
CTLSPEC
AG EF ((P0=6) & (P1=6) & (P2=6) & (P3=6) &
(P4=6) & (P5=6) & (P6=6) & (P7=6))
(P4=6) & (P5=6) & (P6=6) & (P7=6))
(P4=6) & (P5=6) & (P6=6) & (P7=6))
-------------------------------- end main ----------------------------------------------------------------------------------------------------------------------------------------------------------------------Batch Verification:
----------------------------------------- time nusmv -r -v 1 smv_oneway8-SM.smv
-FAIRNESS RUNNING = 1; ... RUNNING = 7;
-LTLSPEC
F ((0=6) & ... & P7=6))
-- >
-- > reachable states: 1.63654e+06 (2ˆ20.6422) out of 4.66949e+08 (2ˆ28.7987)
Ten Diverse FMs for an ATS
134
-- > Successful termination
-- > real 0m43.609s
-- > user 0m43.431s
----------------------------------------- time nusmv -r -v 1 smv_oneway8-SM.smv
-FAIRNESS RUNNING = 1; ... RUNNING = 7;
-CTLSPEC
AF ((0=6) & ... & P7=6))
-- >
-- > reachable states: 1.63654e+06 (2ˆ20.6422) out of 4.66949e+08 (2ˆ28.7987)
-- > Successful termination
-- > real 0m39.211s
-- > user 0m39.015s
----------------------------------------- time nusmv -r -v 1 smv_oneway8-SM.smv
-CTLSPEC
AG EF ((0=6) & ... & P7=6))
-- >
-- > reachable states: 1.63654e+06 (2ˆ20.6422) out of 4.66949e+08 (2ˆ28.7987)
-- > Successful termination
-- > real 0m2.807s
-- > user 0m2.771s
-- > USED MEMORY 74 MB
----------------------------------------- nusmv -v 2 -ctt -r -is smv_oneway8-SM.smv
--ctt checks totatlity of transition relation function
--r prints actual number of reachable states
--v 1
verbose (1..4)
--is
ignore SPEC properties
--AG
used ad hoc algorithm for AG-only properties
-- nusmv -v 1 -bmc -bmc_length 100 cyclic8-smv.txt
------------------------------------------Interactive Verification:
-- ./NuSMV -int
-- read_model -i smv_oneway8-SM.smv
-- flatten_hierarchy
-- encode_variables
-- build_model
-- check_ctlspec -p "AF (P0 = 0)"
--------------- other commands ---------check_ctlspec [-h] [-m | -o output-file] [-n number | -p
-- "ctl-expr [IN context]" | -P "name"]
--go
--pick_state -i
--simulate -i
---------------------------------------
8
SPIN
/* TRAIN MISSION DATA */
byte T0[14];
byte T1[14];
byte T2[14];
byte T3[14];
byte T4[14];
byte T5[14];
byte T6[14];
byte T7[14];
/* TRAIN PROGRESS DATA */
byte P0,P1,P2,P3,P4,P5,P6,P7;
/*
CONSTRAINTs DATA
FOR REGIONS A,B
*/
F. Mazzanti & A. Ferrari
135
byte RA; // occupancy of region
byte RB; // occupancy of region
byte LA; // limit of region A
byte LB; // limit if region B
short A0[14]; // Constraints of
short A1[14]; // Constraints of
short A2[14]; // Constraints of
short A3[14]; //
...
short A4[14];
short A5[14];
short A6[14];
short A7[14];
short B0[14]; // Constraints of
short B1[14]; // Constraints of
short B2[14]; // Constraints of
short B3[14]; //
...
short B4[14];
short B5[14];
short B6[14];
short B7[14];
A
B
Train 0 for Region A
Train 1 for Region A
Train 2 for Region A
Train 0 for Region B
Train 1 for Region B
Train 2 for Region B
/* INITIALIZATIONS */
init {
atomic {
//-------------------------------------------------------------// T0 := [ 1, 9,10,13,15,20,23,22]; -- G1
// T1 := [ 3, 9,10,13,15,20,24,22]; -- R1
// T2 := [ 5,27,11,13,16,20,25,22]; -- Y1
// T3 := [ 7,27,11,13,16,20,26,22]; -- B1
// T4 := [23,22,17,18,11, 9, 2, 1]; -- G2
// T5 := [24,22,17,18,11, 9, 4, 3]; -- R2
// T6 := [25,22,17,18,12,27, 6, 5]; -- Y2
// T7 := [26,22,17,18,12,27, 8, 7]; -- B2
//-------------------------------------------------------------T0[0]= 1; T0[1]= 9; T0[2]=10; T0[3]=13; T0[4]=15; T0[5]=20; T0[6]=23;
T1[0]= 3; T1[1]=9; T1[2]=10; T1[3]=13; T1[4]=15; T1[5]=20; T1[6]=24;
T2[0]= 5; T2[1]=27; T2[2]=11; T2[3]=13; T2[4]=16; T2[5]=20; T2[6]=25;
T3[0]= 7; T3[1]=27; T3[2]=11; T3[3]=13; T3[4]=16; T3[5]=20; T3[6]=26;
T4[0]=23; T4[1]=22; T4[2]=17; T4[3]=18; T4[4]=11; T4[5]= 9; T4[6]= 2;
T5[0]=24; T5[1]=22; T5[2]=17; T5[3]=18; T5[4]=11; T5[5]= 9; T5[6]= 4;
T6[0]=25; T6[1]=22; T6[2]=17; T6[3]=18; T6[4]=12; T6[5]=27; T6[6]= 6;
T7[0]=26; T7[1]=22; T7[2]=17; T7[3]=18; T7[4]=12; T7[5]=27; T7[6]= 8;
//
// ------ initial train positions
-------// Pi=0 as default value. no need of explicit initialization
// ------ region A: train
A0[3] = 1; A0[5] = -1;
A1[3] = 1; A1[5] = -1;
A2[2] = 1; A2[3] = -1;
A3[2] = 1; A3[3] = -1;
A4[1] = 1; A4[4] = -1;
A5[1] = 1; A5[4] = -1;
A6[3] =-1; A6[ 9] = 1;
A7[1] = 1; A7[3] = -1;
// ------- region B:
B0[3] = 1; B0[5]
B1[3] = 1; B1[5]
B2[2] = 1; B2[3]
B3[2] = 1; B3[3]
B4[1] = 1; B4[4]
B5[1] = 1; B5[4]
constraints
A0[ 7]= 1;
A1[ 7]= 1;
A2[ 5]= 1;
A3[ 7]= 1;
A4[10]= 1;
A5[10]= 1;
A6[10]= -1;
A7[ 9]= 1;
train
= -1;
= -1;
= -1;
= -1;
= -1;
= -1;
-----A0[10]
A1[10]
A2[ 9]
A3[ 9]
A4[12]
A5[12]
A6[12]
A7[10]
=
=
=
=
=
=
=
=
constraints -----B0[ 7] = 1; B0[10]
B1[ 7] = 1; B1[10]
B2[ 7] = 1; B2[ 9]
B3[ 5] = 1; B3[ 9]
B4[10] = 1; B4[12]
B5[10] = 1; B5[12]
-1;
-1;
-1;
-1;
-1;
-1;
1;
-1;
=
=
=
=
=
=
-1;
-1;
-1;
-1;
-1;
-1;
Ten Diverse FMs for an ATS
136
B6[1] = 1; B6[3] = -1; B6[ 9] = 1; B6[10] = -1;
B7[3] = -1; B7[9] = 1; B7[10] = -1; B7[12] = 1;
RA
RB
LA
LB
= 1;
= 1;
=7;
=7;
}
do
:: atomic {
(P0 < 6 &&
T0[P0+1] != T1[P1]
T0[P0+1] != T2[P2]
T0[P0+1] != T3[P3]
T0[P0+1] != T4[P4]
T0[P0+1] != T5[P5]
T0[P0+1] != T6[P6]
T0[P0+1] != T7[P7]
(RA + A0[P0+1]) <=
(RB + B0[P0+1]) <=
) ->
P0 = (P0+1);
RA = RA + A0[P0];
RB = RB + B0[P0];
};
:: atomic {
(P1 < 6 &&
T1[P1+1] != T0[P0]
T1[P1+1] != T2[P2]
T1[P1+1] != T3[P3]
T1[P1+1] != T4[P4]
T1[P1+1] != T5[P5]
T1[P1+1] != T6[P6]
T1[P1+1] != T7[P7]
(RA + A1[P1+1]) <=
(RB + B1[P1+1]) <=
) ->
P1 = (P1+1);
RA = RA + A1[P1];
RB = RB + B1[P1];
};
:: atomic {
(P2 < 6 &&
T2[P2+1] != T0[P0]
T2[P2+1] != T1[P1]
T2[P2+1] != T3[P3]
T2[P2+1] != T4[P4]
T2[P2+1] != T5[P5]
T2[P2+1] != T6[P6]
T2[P2+1] != T7[P7]
(RA + A2[P2+1]) <=
(RB + B2[P2+1]) <=
) ->
P2 = (P2+1);
RA = RA + A2[P2];
RB = RB + B2[P2];
};
:: atomic {
(P3 < 6 &&
T3[P3+1] != T0[P0]
T3[P3+1] != T1[P1]
T3[P3+1] != T2[P2]
T3[P3+1] != T4[P4]
T3[P3+1] != T5[P5]
T3[P3+1] != T6[P6]
&&
&&
&&
&&
&&
&&
&&
LA
LB
// next place of train0 not occupied by train1
// next place of train0 not occupied by train2
// next place of train0 not occupied by train7
&& // progress of train0 does not saturate RA
// progress of train0 does not saturate RB
// update occupancy of RA according to the step
// update occupancy of RB according to the step
&&
&&
&&
&&
&&
&&
&&
LA &&
LB
// progress of train0 does not saturate RD
&&
&&
&&
&&
&&
&&
&&
LA &&
LB
// update occupancy of RA according to the step
&&
&&
&&
&&
&&
&&
F. Mazzanti & A. Ferrari
::
::
::
::
T3[P3+1] != T7[P7]
(RA + A3[P3+1]) <=
(RB + B3[P3+1]) <=
) ->
P3 = (P3+1);
RA = RA + A3[P3];
RB = RB + B3[P3];
};
atomic {
(P4 < 6 &&
T4[P4+1] != T0[P0]
T4[P4+1] != T1[P1]
T4[P4+1] != T2[P2]
T4[P4+1] != T3[P3]
T4[P4+1] != T5[P5]
T4[P4+1] != T6[P6]
T4[P4+1] != T7[P7]
(RA + A4[P4+1]) <=
(RB + B4[P4+1]) <=
) ->
P4 = (P4+1);
RA = RA + A4[P4];
RB = RB + B4[P4];
};
atomic {
(P5 < 6 &&
T5[P5+1] != T0[P0]
T5[P5+1] != T1[P1]
T5[P5+1] != T2[P2]
T5[P5+1] != T3[P3]
T5[P5+1] != T4[P4]
T5[P5+1] != T6[P6]
T5[P5+1] != T7[P7]
(RA + A5[P5+1]) <=
(RB + B5[P5+1]) <=
) ->
P5 = (P5+1);
RA = RA + A5[P5];
RB = RB + B5[P5];
};
atomic {
(P6 < 6 &&
T6[P6+1] != T0[P0]
T6[P6+1] != T1[P1]
T6[P6+1] != T2[P2]
T6[P6+1] != T3[P3]
T6[P6+1] != T4[P4]
T6[P6+1] != T5[P5]
T6[P6+1] != T7[P7]
(RA + A6[P6+1]) <=
(RB + B6[P6+1]) <=
) ->
P6 = (P6+1);
RA = RA + A6[P6];
RB = RB + B6[P6];
};
atomic {
(P7 < 6 &&
T7[P7+1] != T0[P0]
T7[P7+1] != T1[P1]
T7[P7+1] != T2[P2]
T7[P7+1] != T3[P3]
T7[P7+1] != T4[P4]
T7[P7+1] != T5[P5]
T7[P7+1] != T6[P6]
137
&&
LA &&
LB
// update occupancy of RA according to the step
&&
&&
&&
&&
&&
&&
&&
LA &&
LB
// update occupancy of RA according to the step
&&
&&
&&
&&
&&
&&
&&
LA &&
LB
// update occupancy of RA according to the step
&&
&&
&&
&&
&&
&&
&&
LA &&
LB
// update occupancy of RA according to the step
&&
&&
&&
&&
&&
&&
&&
Ten Diverse FMs for an ATS
138
(RA +
(RB +
) ->
P7 =
RA =
RB =
};
:: (P0 == 6)
(P4 == 6)
od;
A7[P7+1]) <= LA &&
B7[P7+1]) <= LB
(P7+1);
RA + A7[P7];
RB + B7[P7];
// update occupancy of RA according to the step
&& (P1 == 6) && (P2 == 6) && (P3 == 6) &&
&& (P5 == 6) && (P6 == 6) && (P7 == 6) -> skip;
}
/* PROPERTIES */
ltl p1
{ <> ((P0==6) && (P1==6) && (P2==6) && (P3==6) &&
(P4==6) && (P5==6) && (P6==6) && (P7==6)) }
/*
verfication process
// DEPTH FIRST
spin -a spin_oneway8small.pml
gcc -O3 -o pan pan.c
time pan -a
>
> Full statespace search for:
>
never claim
+ (p1)
>
>
1636546 states, stored
>
7134234 transitions
>
> real 0m13.110s
> user 0m12.683s
> sys 0m0.411s
> USED VIRTUAL MEMORY (pan): 1.02 GB
// BREADTH FIRST
spin -a spin_oneway8.pml
gcc -O3 -DBFS -DBFS_DISK -DVECTORSZ=256000 -o pan pan.c
gcc -O3 -DBFS -DVECTORSZ=256000 -o pan pan.c
time pan -m500000 -v -w33
>
> Full statespace search for:
>never claim
+ (p1)
>
>
1636545 states, stored
>
7134237 transitions
>
> real 1m3.582s
> user 0m31.621s
> sys
0m29.806s
*/
/*
other commands
spin -t[N] -- follow [Nth] simulation trail, see also -k
pan -c0 -- counts all errors
pan -c
-- saves in the trail file the info for 3rd error
pan -e -c0 -- saves all errors trails each one in file specI.trail
spin -k specI.trail -c spec.pml -- displays the trail for error I
pan -r trailfilename
--read and execute trail in file
pan -rN
-read and execute N-th error trail
pan -C -- read and execute trail - columnated output (can add -v,-n)
pan -r -PN read and execute trail - restrict trail output to proc N
pan (for help on options)
pan -w32 -v -D (dot format!)
------
F. Mazzanti & A. Ferrari
139
*/
9
TLA+
------------------ MODULE oneway --------------EXTENDS Integers
VARIABLE
P0,P1,P2,P3,P4,P5,P6,P7,RA,RB
vars == <<P0,P1,P2,P3,P4,P5,P6,P7,RA,RB>>
TypesOK
== (P0 \in 0..6)
T0
T1
T2
T3
T4
T5
T6
T7
==
==
==
==
==
==
==
==
<< 1, 9,10,13,15,20,23>>
<< 3, 9,10,13,15,20,24>>
<< 5,27,11,13,16,20,25>>
<< 7,27,11,13,16,20,26>>
<<23,22,17,18,11, 9, 2>>
<<24,22,17,18,11, 9, 4>>
<<25,22,17,18,12,27, 6>>
<<26,22,17,18,12,27, 8>>
A0
A1
A2
A3
A4
A5
A6
A7
==
==
==
==
==
==
==
==
<<
<<
<<
<<
<<
<<
<<
<<
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
1,
1,
0,
1,
0, 1, 0,-1,
0, 1, 0,-1,
1,-1, 0, 1,
1,-1, 0, 0,
0, 0,-1, 0,
0, 0,-1, 0,
0,-1, 0, 0,
0,-1, 0, 0,
0>>
0>>
0>>
0>>
0>>
0>>
0>>
0>>
B0
B1
B2
B3
B4
B5
B6
B7
==
==
==
==
==
==
==
==
<<
<<
<<
<<
<<
<<
<<
<<
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
0,
1,
1,
1,
0,
0, 1, 0,-1,
0, 1, 0,-1,
1,-1, 0, 0,
1,-1, 0, 1,
0, 0,-1, 0,
0, 0,-1, 0,
0,-1, 0, 0,
0,-1, 0, 0,
0>>
0>>
0>>
0>>
0>>
0>>
0>>
0>>
LA == 7
LB == 7
LL == 6
Init==
/\ P0=0 /\ P1=0 /\ P2=0 /\ P3=0 /\ P4=0 /\ P5=0 /\ P6=0 /\ P7=0
/\ RA=1 /\ RB=1
Move0
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P0 < LL
T0[P0+2] /= T1[P1+1]
T0[P0+2] /= T2[P2+1]
T0[P0+2] /= T3[P3+1]
T0[P0+2] /= T4[P4+1]
T0[P0+2] /= T5[P5+1]
T0[P0+2] /= T6[P6+1]
T0[P0+2] /= T7[P7+1]
RA + A0[P0+2] <= LA
RB + B0[P0+2] <= LB
P0’ = (P0+1)
Ten Diverse FMs for an ATS
140
/\ RA’ = RA + A0[P0+2]
/\ RB’ = RB + B0[P0+2]
/\ UNCHANGED <<P1,P2,P3,P4,P5,P6,P7>>
Move1
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P1 < LL
T1[P1+2] /= T0[P0+1]
T1[P1+2] /= T2[P2+1]
T1[P1+2] /= T3[P3+1]
T1[P1+2] /= T4[P4+1]
T1[P1+2] /= T5[P5+1]
T1[P1+2] /= T6[P6+1]
T1[P1+2] /= T7[P7+1]
RA + A1[P1+2] <= LA
RB + B1[P1+2] <= LB
P1’ = (P1+1)
RA’ = RA + A1[P1+2]
RB’ = RB + B1[P1+2]
UNCHANGED <<P0,P2,P3,P4,P5,P6,P7>>
Move2
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P2 < LL
T2[P2+2] /= T0[P0+1]
T2[P2+2] /= T1[P1+1]
T2[P2+2] /= T3[P3+1]
T2[P2+2] /= T4[P4+1]
T2[P2+2] /= T5[P5+1]
T2[P2+2] /= T6[P6+1]
T2[P2+2] /= T7[P7+1]
RA + A2[P2+2] <= LA
RB + B2[P2+2] <= LB
P2’ = (P2+1)
RA’ = RA + A2[P2+2]
RB’ = RB + B2[P2+2]
UNCHANGED <<P0,P1,P3,P4,P5,P6,P7>>
Move3
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P3 < LL
T3[P3+2] /= T0[P0+1]
T3[P3+2] /= T1[P1+1]
T3[P3+2] /= T2[P2+1]
T3[P3+2] /= T4[P4+1]
T3[P3+2] /= T5[P5+1]
T3[P3+2] /= T6[P6+1]
T3[P3+2] /= T7[P7+1]
RA + A3[P3+2] <= LA
RB + B3[P3+2] <= LB
P3’ = (P3+1)
RA’ = RA + A3[P3+2]
RB’ = RB + B3[P3+2]
UNCHANGED <<P0,P1,P2,P4,P5,P6,P7>>
Move4
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P4 < LL
T4[P4+2] /= T0[P0+1]
T4[P4+2] /= T1[P1+1]
T4[P4+2] /= T2[P2+1]
T4[P4+2] /= T3[P3+1]
T4[P4+2] /= T5[P5+1]
T4[P4+2] /= T6[P6+1]
T4[P4+2] /= T7[P7+1]
RA + A4[P4+2] <= LA
F. Mazzanti & A. Ferrari
141
/\
/\
/\
/\
/\
RB + B4[P4+2] <= LB
P4’ = (P4+1)
RA’ = RA + A4[P4+2]
RB’ = RB + B4[P4+2]
UNCHANGED <<P0,P1,P2,P3,P5,P6,P7>>
Move5
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P5 < LL
T5[P5+2] /= T0[P0+1]
T5[P5+2] /= T1[P1+1]
T5[P5+2] /= T2[P2+1]
T5[P5+2] /= T3[P3+1]
T5[P5+2] /= T4[P4+1]
T5[P5+2] /= T6[P6+1]
T5[P5+2] /= T7[P7+1]
RA + A5[P5+2] <= LA
RB + B5[P5+2] <= LB
P5’ = (P5+1)
RA’ = RA + A5[P5+2]
RB’ = RB + B5[P5+2]
UNCHANGED <<P0,P1,P2,P3,P4,P6,P7>>
Move6
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P6 < LL
T6[P6+2] /= T0[P0+1]
T6[P6+2] /= T1[P1+1]
T6[P6+2] /= T2[P2+1]
T6[P6+2] /= T3[P3+1]
T6[P6+2] /= T4[P4+1]
T6[P6+2] /= T5[P5+1]
T6[P6+2] /= T7[P7+1]
RA + A6[P6+2] <= LA
RB + B6[P6+2] <= LB
P6’ = (P6+1)
RA’ = RA + A6[P6+2]
RB’ = RB + B6[P6+2]
UNCHANGED <<P0,P1,P2,P3,P4,P5,P7>>
Move7
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
/\
==
P7 < LL
T7[P7+2] /= T0[P0+1]
T7[P7+2] /= T1[P1+1]
T7[P7+2] /= T2[P2+1]
T7[P7+2] /= T3[P3+1]
T7[P7+2] /= T4[P4+1]
T7[P7+2] /= T5[P5+1]
T7[P7+2] /= T6[P6+1]
RA + A7[P7+2] <= LA
RB + B7[P7+2] <= LB
P7’ = (P7+1)
RA’ = RA + A7[P7+2]
RB’ = RB + B7[P7+2]
UNCHANGED <<P0,P1,P2,P3,P4,P5,P6>>
Next==
\/ Move0 \/
\/ Move4 \/
Move1 \/
Move5 \/
Move2 \/
Move6 \/
Move3
Move7
Arrived ==
/\ (P0=LL) /\ (P1=LL) /\ (P2=LL) /\ (P3=LL)
/\ (P4=LL) /\ (P5=LL) /\ (P6=LL) /\ (P7=LL)
Property == <>Arrived
Ten Diverse FMs for an ATS
142
Spec == Init /\ [][Next]_vars
SFairSpec == Init /\ [][Next]_vars /\ SF_vars (Next)
(*for LTL verification*)
(**************************************************)
(* Property: <>Arrived, Behavior Spec: SFairSpec *)
(* States: 1636545, Result: TRUE, Time 3m17s
*)
(**************************************************)
(* Model Overview: setting Temporal formula == "Spec" *)
(* Deadlock Found: trace for P0=6 & P4=6 *)
(* PROPERTIES: <>Arrivedis FALSE, because of implicit stuttering*)
(* Model Overview: setting Temporal formula == "SFairSpec" *)
(* Deadlock Found: trace for P0=6 & P4=6 (stuttering does not avoids deadlocks)*)
(* PROPERTIES: <>Arrived
is TRUE, stuttering ignored *)
===============================================
10 UMC
Class REGION2 is
Vars:
--------------------------------------------------------------T0: int[] := [ 1, 9,10,13,15,20,23]; -- G1
T1: int[] := [ 3, 9,10,13,15,20,24]; -- R1
T2: int[] := [ 5,27,11,13,16,20,25]; -- Y1
T3: int[] := [ 7,27,11,13,16,20,26]; -- B1
T4: int[] := [23,22,17,18,11, 9, 2]; -- G2
T5: int[] := [24,22,17,18,11, 9, 4]; -- R2
T6: int[] := [25,22,17,18,12,27, 6]; -- Y2
T7: int[] := [26,22,17,18,12,27, 8]; -- B2
---------------------------------------------------------------P0: int :=0;
P1: int :=0;
P2: int :=0;
P3: int :=0;
P4: int :=0;
P5: int :=0;
P6: int :=0;
P7: int :=0;
--------------------------------------------------------------------- region A: train constraints -----A0: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- G1
A1: int[] := [ 0, 0, 0, 1, 0,-1, 0]; -- R1
A2: int[] := [ 0, 0, 1,-1, 0, 1, 0]; -- Y1
A3: int[] := [ 0, 0, 1,-1, 0, 0, 0]; -- B1
A4: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- G2
A5: int[] := [ 0, 1, 0, 0,-1, 0, 0]; -- R2
A6: int[] := [ 0, 0, 0,-1, 0, 0, 0]; -- Y2
A7: int[] := [ 0, 1, 0,-1, 0, 0, 0]; -- B2
------------------------------------------------ region
B0: int[] := [
B1: int[] := [
B2: int[] := [
B3: int[] := [
B4: int[] := [
B5: int[] := [
B6: int[] := [
B:
0,
0,
0,
0,
0,
0,
0,
train constraints -----0, 0, 1, 0,-1, 0]; -- G1
0, 0, 1, 0,-1, 0]; -- R1
0, 1,-1, 0, 0, 0]; -- Y1
0, 1,-1, 0, 1, 0]; -- B1
1, 0, 0,-1, 0, 0]; -- G2
1, 0, 0,-1, 0, 0]; -- R2
1, 0,-1, 0, 0, 0]; -- Y2
F. Mazzanti & A. Ferrari
B7: int[] := [ 0, 0, 0,-1, 0, 0, 0]; -- B2
------------------------------------------
------------------------------------------------------------RA: int :=1; -- initial value for region RA
RB: int :=1; -- initial value for region RB
LA: int :=7; -- limit value for region RA
LB: int :=7; -- limit value for region RB
------------------------------------------------------------------State Top =s1
Behavior:
------------------------- train 0 ----------------------------s1 -> s1
{- [P0 < 6 &
T0[P0+1] != T1[P1] &
T0[P0+1] != T2[P2] &
T0[P0+1] != T3[P3] &
T0[P0+1] != T4[P4] &
T0[P0+1] != T5[P5] &
T0[P0+1] != T6[P6] &
T0[P0+1] != T7[P7] &
RA + A0[P0+1] <= LA &
RB + B0[P0+1] <= LB] /
P0 := P0 +1;
RA = RA + A0[P0];
RB = RB + B0[P0];
}
------------------------- train 1 ----------------------------s1 -> s1
{- [P1 < 6 &
T1[P1+1] != T0[P0] &
T1[P1+1] != T2[P2] &
T1[P1+1] != T3[P3] &
T1[P1+1] != T4[P4] &
T1[P1+1] != T5[P5] &
T1[P1+1] != T6[P6] &
T1[P1+1] != T7[P7] &
RA + A1[P1+1] <= LA &
RB + B1[P1+1] <= LB ] /
P1 := P1 +1;
RA = RA + A1[P1];
RB = RB + B1[P1];
}
------------------------- train 2 ----------------------------s1 -> s1
{- [P2 < 6 &
T2[P2+1] != T0[P0] &
T2[P2+1] != T1[P1] &
T2[P2+1] != T3[P3] &
T2[P2+1] != T4[P4] &
T2[P2+1] != T5[P5] &
T2[P2+1] != T6[P6] &
T2[P2+1] != T7[P7] &
RA + A2[P2+1] <= LA &
RB + B2[P2+1] <= LB ] /
P2 := P2 +1;
RA = RA + A2[P2];
RB = RB + B2[P2];
}
143
144
------------------------- train 3 ----------------------------s1 -> s1
{- [P3 < 6 &
T3[P3+1] != T0[P0] &
T3[P3+1] != T1[P1] &
T3[P3+1] != T2[P2] &
T3[P3+1] != T4[P4] &
T3[P3+1] != T5[P5] &
T3[P3+1] != T6[P6] &
T3[P3+1] != T7[P7] &
RA + A3[P3+1] <= LA &
RB + B3[P3+1] <= LB ] /
P3 := P3 +1;
RA = RA + A3[P3];
RB = RB + B3[P3];
}
------------------------- train 4 ----------------------------s1 -> s1
{- [P4 < 6 &
T4[P4+1] != T0[P0] &
T4[P4+1] != T1[P1] &
T4[P4+1] != T2[P2] &
T4[P4+1] != T3[P3] &
T4[P4+1] != T5[P5] &
T4[P4+1] != T6[P6] &
T4[P4+1] != T7[P7] &
RA + A4[P4+1] <= LA &
RB + B4[P4+1] <= LB ] /
P4 := P4 +1;
RA = RA + A4[P4];
RB = RB + B4[P4];
}
------------------------- train 5 ----------------------------s1 -> s1
{- [P5 < 6 &
T5[P5+1] != T0[P0] &
T5[P5+1] != T1[P1] &
T5[P5+1] != T2[P2] &
T5[P5+1] != T3[P3] &
T5[P5+1] != T4[P4] &
T5[P5+1] != T6[P6] &
T5[P5+1] != T7[P7] &
RA + A5[P5+1] <= LA &
RB + B5[P5+1] <= LB] /
P5 := P5 +1;
RA = RA + A5[P5];
RB = RB + B5[P5];
}
------------------------- train 6 ----------------------------s1 -> s1
{- [P6 < 6 &
T6[P6+1] != T0[P0] &
T6[P6+1] != T1[P1] &
T6[P6+1] != T2[P2] &
T6[P6+1] != T3[P3] &
T6[P6+1] != T4[P4] &
T6[P6+1] != T5[P5] &
T6[P6+1] != T7[P7] &
RA + A6[P6+1] <= LA &
RB + B6[P6+1] <= LB ] /
P6 := P6 +1;
Ten Diverse FMs for an ATS
F. Mazzanti & A. Ferrari
RA = RA + A6[P6];
RB = RB + B6[P6];
}
------------------------- train 7 ----------------------------s1 -> s1
{- [P7 < 6 &
T7[P7+1] != T0[P0] &
T7[P7+1] != T1[P1] &
T7[P7+1] != T2[P2] &
T7[P7+1] != T3[P3] &
T7[P7+1] != T4[P4] &
T7[P7+1] != T5[P5] &
T7[P7+1] != T6[P6] &
RA + A7[P7+1] <= LA &
RB + B7[P7+1] <= LB ] /
P7 := P7 +1;
RA = RA + A7[P7];
RB = RB + B7[P7];
}
------------------------- termination ----------------------------s1 -> s1
{- [(P0=6) and (P1=6) and (P2=6) and (P3=6)&
(P4=6) and (P5=6) and (P6=6) and (P7=6)] / ARRIVED}
end REGION2;
Objects:
Count: Token;
SYS: REGION2;
Abstractions {
Action ARRIVED -> ARRIVED
Action Error -> Error
-- State:
-SYS.P0=0 and
-SYS.P1=0 and
-SYS.P2=0 and
-SYS.P3=0 and
-SYS.P4=0 and
-SYS.P5=0 and
-SYS.P6=0 and
-SYS.P7=0 -> Home -- abstract label on final state
}
-- time umc -m3 -100 umc_oneway8.txt AFARR.txt
--- > The Formula: "AF {ARRIVED} true"
-- > is: TRUE
-- > statspace stats: states generated= 1636545 ... evaluation time= 37.538 sec.
--- > real 0m36.980s
-- > user 1m23.800s
-- > sys 0m1.735s
-- USED VIRTUAL MEMORY: 2.98G
--- time mcstats -m3 umc_oneway8.txt
--- AFARR==
"AF {ARRIVED} true"
-------------------------------------------------------------------------------------------------------------------------------------
145
Ten Diverse FMs for an ATS
146
11 UPPAAL
//
//
//
global declarations
//------const int
const int
const int
const int
const int
const int
const int
const int
train
T0[7]
T1[7]
T2[7]
T3[7]
T4[7]
T5[7]
T6[7]
T7[7]
const int LA =7;
const int LB =7;
//------const int
const int
const int
const int
const int
const int
const int
const int
missions -----= { 1, 9,10,13,15,20,23};
= { 3, 9,10,13,15,20,24};
= { 5,27,11,13,16,20,25};
= { 7,27,11,13,16,20,26};
= {23,22,17,18,11, 9, 2};
= {24,22,17,18,11, 9, 4};
= {25,22,17,18,12,27, 6};
= {26,22,17,18,12,27, 8};
// limit value for region RA
// limit value for region RB
region A:
A0[7] = {
A1[7] = {
A2[7] = {
A3[7] = {
A4[7] = {
A5[7] = {
A6[7] = {
A7[7] = {
train
0, 0,
0, 0,
0, 0,
0, 0,
0, 1,
0, 1,
0, 0,
0, 1,
constraints
0, 1, 0,-1,
0, 1, 0,-1,
1,-1, 0, 1,
1,-1, 0, 0,
0, 0,-1, 0,
0, 0,-1, 0,
0,-1, 0, 0,
0,-1, 0, 0,
-----0}; //G1
0}; // R1
0}; // Y1
0}; // B1
0}; // G2
0}; // R2
0}; // Y2
0}; // B2
//------- region B: train constraints -----const int B0[7] = { 0, 0, 0, 1, 0,-1, 0}; //
const int B1[7] = { 0, 0, 0, 1, 0,-1, 0}; //
const int B2[7] = { 0, 0, 1,-1, 0, 0, 0}; //
const int B3[7] = { 0, 0, 1,-1, 0, 1, 0}; //
const int B4[7] = { 0, 1, 0, 0,-1, 0, 0}; //
const int B5[7] = { 0, 1, 0, 0,-1, 0, 0}; //
const int B6[7] = { 0, 1, 0,-1, 0, 0, 0}; //
const int B7[7] = { 0, 0, 0,-1, 0, 0, 0}; //
//-----------------------------------------int
int
int
int
int
int
int
int
int
int
P0
P1
P2
P3
P4
P5
P6
P7
RA
RB
G1
R1
Y1
B1
G2
R2
Y2
B2
:= 0;
:= 0;
:= 0;
:= 0;
:= 0;
:= 0;
:= 0;
:= 0;
:=1; // initial value for region RA
:=1; // initial value for region RB
broadcast chan move0,move1,move2,move3,move4,move5,move6,move7;
//------------ template defintions --------process Uppaal_Model() {
state s0;
urgent s0;
init s0;
trans
s0 -> s0 {
guard
P0 < 6 &&
T0[P0+1] != T1[P1] &&
T0[P0+1] != T2[P2] &&
T0[P0+1] != T3[P3] &&
F. Mazzanti & A. Ferrari
T0[P0+1] != T4[P4] &&
T0[P0+1] != T5[P5] &&
T0[P0+1] != T6[P6] &&
T0[P0+1] != T7[P7] &&
RA + A0[P0+1] <= LA &&
RB + B0[P0+1] <= LB;
sync move0!;
assign
P0 := P0+1,
RA := RA + A0[P0],
RB := RB + B0[P0];
},
s0 -> s0 {
guard
P1 < 6 &&
T1[P1+1] != T0[P0] &&
T1[P1+1] != T2[P2] &&
T1[P1+1] != T3[P3] &&
T1[P1+1] != T4[P4] &&
T1[P1+1] != T5[P5] &&
T1[P1+1] != T6[P6] &&
T1[P1+1] != T7[P7] &&
RA + A1[P1+1] <= LA &&
RB + B1[P1+1] <= LB;
sync move1!;
assign
P1 := P1+1,
RA := RA + A1[P1],
RB := RB + B1[P1];
},
s0 -> s0 {
guard
P2 < 6 &&
T2[P2+1] != T0[P0] &&
T2[P2+1] != T1[P1] &&
T2[P2+1] != T3[P3] &&
T2[P2+1] != T4[P4] &&
T2[P2+1] != T5[P5] &&
T2[P2+1] != T6[P6] &&
T2[P2+1] != T7[P7] &&
RA + A2[P2+1] <= LA &&
RB + B2[P2+1] <= LB;
sync move2!;
assign
P2 := P2+1,
RA := RA + A2[P2],
RB := RB + B2[P2];
},
s0 -> s0 {
guard
P3 < 6 &&
T3[P3+1] != T0[P0] &&
T3[P3+1] != T2[P2] &&
T3[P3+1] != T1[P1] &&
T3[P3+1] != T4[P4] &&
T3[P3+1] != T5[P5] &&
T3[P3+1] != T6[P6] &&
T3[P3+1] != T7[P7] &&
RA + A3[P3+1] <= LA &&
RB + B3[P3+1] <= LB;
sync move3!;
assign
147
Ten Diverse FMs for an ATS
148
P3 := P3+1,
RA := RA + A3[P3],
RB := RB + B3[P3];
},
s0 -> s0 {
guard
P4 < 6 &&
T4[P4+1] != T0[P0] &&
T4[P4+1] != T1[P1] &&
T4[P4+1] != T2[P2] &&
T4[P4+1] != T3[P3] &&
T4[P4+1] != T5[P5] &&
T4[P4+1] != T6[P6] &&
T4[P4+1] != T7[P7] &&
RA + A4[P4+1] <= LA &&
RB + B4[P4+1] <= LB;
sync move4!;
assign
P4 := P4+1,
RA := RA + A4[P4],
RB := RB + B4[P4];
},
s0 -> s0 {
guard
P5 < 6 &&
T5[P5+1] != T0[P0] &&
T5[P5+1] != T1[P1] &&
T5[P5+1] != T2[P2] &&
T5[P5+1] != T3[P3] &&
T5[P5+1] != T4[P4] &&
T5[P5+1] != T6[P6] &&
T5[P5+1] != T7[P7] &&
RA + A5[P5+1] <= LA &&
RB + B5[P5+1] <= LB;
sync move5!;
assign
P5 := P5+1,
RA := RA + A5[P5],
RB := RB + B5[P5];
},
s0 -> s0 {
guard
P6 < 6 &&
T6[P6+1] != T0[P0] &&
T6[P6+1] != T1[P1] &&
T6[P6+1] != T2[P2] &&
T6[P6+1] != T3[P3] &&
T6[P6+1] != T4[P4] &&
T6[P6+1] != T5[P5] &&
T6[P6+1] != T7[P7] &&
RA + A6[P6+1] <= LA &&
RB + B6[P6+1] <= LB;
sync move6!;
assign
P6 := P6+1,
RA := RA + A6[P6],
RB := RB + B6[P6];
},
s0 -> s0 {
guard
P7 < 6 &&
F. Mazzanti & A. Ferrari
T7[P7+1] != T0[P0] &&
T7[P7+1] != T1[P1] &&
T7[P7+1] != T2[P2] &&
T7[P7+1] != T3[P3] &&
T7[P7+1] != T4[P4] &&
T7[P7+1] != T5[P5] &&
T7[P7+1] != T6[P6] &&
RA + A7[P7+1] <= LA &&
RB + B7[P7+1] <= LB;
sync move7!;
assign
P7 := P7+1,
RA := RA + A7[P7],
RB := RB + B7[P7];
};
}
// template instantiations
// List one or more processes to be composed into a system.
system Uppaal_Model;
//
//
//
//
//
//
./verifyta -h
./verifyta uppaal_oneway8seq.ta uppaal_queries.txt
file: uppaal_queries.txt
A<>((P0==6) and (P1==6)and (P2==6)and (P3==6) and
(P4==6) and (P5==6)and (P6==6)and (P7==6))
149
| 3cs.SY
|
1
Learning Spike time codes through Morphological
Learning with Binary Synapses
arXiv:1506.05212v1 [cs.NE] 17 Jun 2015
Subhrajit Roy, Student Member, IEEE, Phyo Phyo San, Shaista Hussain, Lee Wang Wei and Arindam Basu∗ ,
Member, IEEE
Abstract—In this paper, a neuron with nonlinear dendrites
(NNLD) and binary synapses that is able to learn temporal
features of spike input patterns is considered. Since binary
synapses are considered, learning happens through formation and
elimination of connections between the inputs and the dendritic
branches to modify the structure or “morphology” of the NNLD.
A morphological learning algorithm inspired by the ‘Tempotron’,
i.e., a recently proposed temporal learning algorithm–is presented
in this work. Unlike ‘Tempotron’, the proposed learning rule uses
a technique to automatically adapt the NNLD threshold during
training. Experimental results indicate that our NNLD with 1-bit
synapses can obtain similar accuracy as a traditional Tempotron
with 4-bit synapses in classifying single spike random latency
and pair-wise synchrony patterns. Hence, the proposed method is
better suited for robust hardware implementation in the presence
of statistical variations. We also present results of applying this
rule to real life spike classification problems from the field of
tactile sensing.
Index Terms—spiking neuron, tempotron, binary synapse,
dendrites, plasticity, learning
I. I NTRODUCTION
Though the representation of stimulus by the neurons in
our brain is a topic of much ongoing research and debate,
it is widely believed that the timing of the action potentials
or spikes fired by these neurons carry important information
[1]. Spike latency codes i.e. delay in the spike time after
stimulus presentation have been suggested for tactile, olfactory and retinal systems [2]. They are also thought to offer
significant advantages in terms of reducing power needed for
communicating spikes as well as allowing rapid processing
of inputs. Hence, neuromorphic engineers, who aim to mimic
the brain’s processing capabilities in silicon, have also been
interested in spike timing based neural networks.
Several analog CMOS integrated circuits operating in the
sub-threshold regime have been designed in the past to implement somatic and synaptic functions [3]–[6]. However, with
the increase of statistical variations due to the constantly decreasing feature size of transistors, performance of silicon neural networks requiring accurate setting of a “weight” parameter
become strongly compromised. The typical solution for this
Financial Support from MOE through grant MOE2012-T2-2-054 is acknowledged.
Subhrajit Roy, Shaista Hussain and Arindam Basu are with the School
of Electrical and Electronics Engineering, Nanyang Technological University, Singapore (e-mail: [email protected]; [email protected];
[email protected]). Phyo Phyo San is with the Institute for Infocomm
Research, Singapore (e-mail: [email protected]). Lee Wang Wei is
with the SINAPSE laboratory at National University of Singapore (e-mail:
lee wang [email protected]). ∗ Arindam Basu is the corresponding author.
problem is to increase transistor size. However, this alone
might not be sufficient to guarantee good matching across the
chip. For example, [7] demonstrates that 5µmX5µm transistor
based 5-bit DACs fabricated in 0.35µm CMOS exhibit only
1.1 effective bits. Calibration techniques can be used to
improve the accuracy; however, this incurs significant area
penalty due to storage of calibration bits and is unacceptable
in large scale systems. The problem of mismatch is also
exacerbated for several nanometer scale non-CMOS devices
(e.g. memristor [8]–[10] or domain wall magnets (DWM)
[11]) that have potential for use in large scale neuromorphic
applications. For example, DWM synapses are expected to
have typical programming accuracies of 4-bits which can
reduce to 2-3 bits effective resolution due to mismatch [11].
Hence, there is a strong need to develop algorithms and
architectures that retain the performance of earlier spiking
systems but require low-resolution weights.
In this paper, a hardware friendly morphological learning
rule for neurons with nonlinear dendrites (NNLD) and binary
synapses for classifying spatiotemporal spike patterns is presented. Previous studies have shown that NNLD can be successfully applied to learn both mean rate encoded inputs [12],
[13] and spike-timing information [14], [15]. Although in [14],
[15] we proposed a spike time based learning rule for training
a NNLD architecture, however the rule was not optimized for
memory capacity. Here we propose a novel memory capacity
optimized spike timing based learning rule for training NNLD.
Our work is inspired by the Tempotron learning rule for
spiking neurons [1]. However, unlike the Tempotron learning
rule that requires weights with high resolution, the proposed
network uses low-resolution non-negative integer weights and
learns through modifying connections of inputs to dendritic
branches. Thus the ‘morphology’ or structure of the neuron (in
terms of connectivity pattern) reflects the learning. This results
in easier hardware implementation since a low-resolution
non-negative integer weight of W can be implemented by
activating a shared binary synapse W times through time
multiplexing schemes like address event representation (AER)
[16], [17]. Furthermore, the spiking threshold of the neuron
employed in Tempotron is fixed throughout the learning. On
the other hand, the proposed method is equipped with a
threshold adaptation mechanism trying to optimize the number
of false positives and false negatives. Some initial results of
this concept were presented in [19]. In this paper, we present
a novel threshold adaptation technique, more detailed set of
results, comparisons with other work and application to a real
world problem.
2
d inputs
m dendritic branches
1 2 3 45
d
1
b()
2
b()
3
b()
4
b()
m
Vthr
V(t) Vout
III. M ORPHOLOGICAL L EARNING A LGORITHM
b()
Fig. 1: The architecture of neuron with nonlinear dendrites (NNLD).
The organization of this paper is as follows: in Section II
and III, the architecture of the neuron with nonlinear dendrites
and its morphological learning algorithm are introduced. Two
spike time based binary classification tasks are discussed in
Section IVA and performance of our method in comparison
to Tempotron is shown in Section IVB, IV-C and V. Finally,
our work is compared with other classifiers in Section VI and
then a conclusion is drawn in Section VII.
II. N EURON WITH N ONLINEAR D ENDRITES
It has been shown that a neuron with lumped dendritic
nonlinearities possesses higher storage capacity than its counterparts with linear dendritic summation [12]. As presented
in Fig. 1, the structure of NNLD is characterized by m
identical dendritic branches and k excitatory synaptic contacts
per branch. For each branch, the synaptic contact is formed by
one of d dimensions of input afferents where d >> k. At the
relevant times governed by incoming spikes, the synapses are
activated and the membrane voltage is calculated by weighted
sum of postsynaptic potentials (PSPs) as follows:
m
m
d
X
X
X
X
V (t) =
b(vj (t)) =
b
wij
K t − tif (1)
j=1
j=1
threshold voltage (Vthr ). Similar to [1], this indicates that a
neuron fires at least one spike if V (t) crosses (Vthr ); otherwise
it remains quiescent. After the neuron fires a spike, it is
returned to the refractory state and the rest of the spikes in
the incoming pattern do not affect the computation.
i=1
On one hand the proposed learning rule tries to optimize the
connections between the input lines and the synapses while on
the other hand it tries to find a optimal value of the neuron
firing threshold(Vthr ). Below we describe both the processes.
a) Learning the connections: Since we have binary
synapses, a morphological learning rule that can modify the
connections between afferent lines and synapses is needed.
The inspiration of this work comes from the Tempotron
learning rule [1] that learnt the temporal feature of random
spike patterns in two classes by updating its weights so
that the neuron fires a spike for patterns in class P + and
stays silent for the other class. It was shown in [1] that the
Tempotron rule endows a neuron with larger classification
capacity than a perceptron with same number of synapses.
Hence, we also start with a cost function that measures the
deviation between the maximum membrane voltage (Vmax )
generated by misclassified patterns and Vthr defined as:
Vthr − Vmax , if pattern in P + is presented
E=
Vmax − Vthr , if pattern in P − is presented
(3)
where Vmax is the maximal value of postsynaptic potential
V (t) at the time tmax , i.e., Vmax = V (tmax ). According to
the gradient-descent method, the change in synaptic efficacy
for the first case is calculated by:
∆wij = −
∂E
∂wij
f
where wij is the weight of the ith synapse formed on the
j th branch, vj (t) is the input to the j th dendritic nonlinearity,
b (·) is the nonlinear activation function of the dendritic branch
2
which is characterized by b (vj (t)) = vj (t) xthr [13], K
denotes the post-synaptic potential kernel and tif are times
of incoming spikes on the ith afferent. Since we consider
binary synapses, wij {0, 1}. To suppress unrealistically large
values, we include a saturation level xsat at the output of
each dendrite such that for b(vj (t)) > xsat , b(vj (t)) = xsat .
Similar to our earlier work [13], we allow each input afferent
to make multiple synaptic connections on the same dendritic
branch butP
limit the total number of connections per branch by
d
enforcing i=1 wij = k for each j. In this work, we consider
normalized PSP of the form:
K (t − tf ) = V0 (exp[−(t − tf )/τ ]−exp[−(t − tf )/τs ]) (2)
where the parameters τ and τs =τ /4 denote the decay time
constants of membrane integration and synaptic current respectively [1].
For binary classification, the final output voltage Vout is
interpreted to take a value of either “1” or “0” depending on
whether the summed membrane voltage V (t) is crossing a
∂
m
P
Vthr −
j=1
=−
0
= b (vj (tmax ))
b
d
P
wij
P
i=1
f
K tmax −
tif
!!!
∂wij
X
K(tmax − tif )
f
(4)
where b0 (.) denotes the derivative of b(.). The gradient for
second case can be calculated similarly and we do not show
it here for brevity. As mentioned earlier, since we consider
binary weights, i.e., wij = 1 if a connection exists and 0
otherwise, we cannot directly modify the weights by adding
the ∆wij term derived here. Instead, the term ∆wij in Equation 4 is reinterpreted as a correlation term cij (= wij ) and is
used to guide the process of swapping connections. At every
iteration of the learning process, the synapse with the lowest
cij averaged over an entire batch of patterns from a randomly
chosen target set will be replaced (weight changed from 1 to
0) with the highest cij synapse in a candidate replacement
set similar to [13]. To keep the paper self-contained, the
mechanism of the learning process is outlined briefly below:
1) The learning process starts with random initialization
of the connection matrix between input afferents and
dendritic branches.
3
2) In each iteration of the training process, the activation
of synapses on each dendritic branch are determined
and the cell membrane output voltage (V (t)) in (1) is
calculated for all the input patterns.
3) From the calculated V (t), the maximum membrane
voltage (Vmax ) is observed and classification result is
determined, i.e., the patterns are correctly classified if
Vmax > Vthr for P + and Vmax < Vthr for P − .
4) A random set T of nT synapses having weight 1 was
targeted for possible replacement. For all misclassified
patterns, the correlation term, cij is calculated for each
synapse in T and averaged over the entire pattern set,
5) The poorest-performing synapse (minimum cij ) in T ,
Tmin is chosen for replacement.
6) To aid the replacement process, a randomly chosen set R
containing nR of the d afferent lines is forced to make
silent synapses with weight 1 on the dendritic branch
of Tmin . These synapses are “silent” since they do not
contribute PSP to the computation of V (t)–so they do
not alter the classification when the same pattern set is
re-applied. But now cij is calculated for synapses in R
and Tmin is replaced with the best-performing synapse
(maximum cij ) in R.
7) The learning from step (2) to (6) continues until all the
patterns are correctly learnt (or) the maximum number
of iteration is reached.
b) Learning the threshold: Since we do not have an
arbitrary multiplicative weight in our neural model, the range
of maximum voltages obtainable from our model in response
to a fixed temporal spike pattern is limited. This is similar
to the problem faced in [18]. Hence, improper selection of
threshold may largely degrade the classification performance
since a very large Vthr (= m × k × Kmax for example) may
never be crossed by V (t). In [19], we determined the value
of Vthr by noting the maximum value of V (t), i.e., Vmax at
time tmax for a large number of random input spike patterns
and connection matrices. The resultant probability distribution
of Vmax was used to determine the optimal threshold Vthr .
In [19], Vthr was set to be the voltage corresponding to the
peak of probability distribution function. Here, we propose
an automatic mechanism for adapting Vthr during training.
This technique involves updating the value of Vthr after each
iteration which is guided by the following formula:
∆Vthr = η(wf p F P − wf n F N )
(5)
where F P , F N , wf p , wf n and η > 0 are the number of false
positives, number of false negatives, weightage associated with
false positive error, weightage associated with false negative
error and threshold learning rate respectively. In this article, we
keep wf p = wf n = 1. Equation 5 is responsible for balancing
the number of false positives and false negatives. When F P >
F N , the number of negative patterns incorrectly classified as
positive patterns is more than the number of positive patterns
incorrectly classified as negative patterns so to balance F P
and F N Equation 5 increases the value of Vthr . On the other
hand, when F P < F N Equation 5 diminishes the value of
Vthr . The rate of threshold adaptation is controlled by the
threshold learning rate η.
IV. E XPERIMENTS AND R ESULTS
A. Problem Description
In this sub-section, we describe the two tasks used to
demonstrate the performance of our algorithm. The reason for
this choice is that both of these are standard problems shown
in [1] and facilitates comparison.
1) Task I: Classifying Random Latency Patterns: The first
task is binary classification of single spike random latency patterns [1]. To perform the task, P spike patterns were generated
and randomly assigned to one of the two classes P + (Class
1) or P − (Class 2). Each spike pattern X = (x1 , x2 , . . . , xd )
consists of d afferents, where each of them spiked only once
at a time drawn independently from a uniform distribution
between 1 and T ms.
2) Task II: Classifying pairwise Synchrony Patterns: To
examine the ability of our algorithm to learn correlations
in multiple spikes, another data set that consists of pairwise
synchrony events in each pattern is generated. In this data set,
all the d afferents are grouped into (d/2) pairs and afferents
in a given pair fire single spike patterns synchronously. Since
synchronous events occur at random, uniformly distributed
times in both pattern categories, so that class information
is embedded solely in the patterns of synchrony; neither
spike counts nor spike timing of individual neurons carry
any information relevant for the classification task. This task
mimicked spike synchrony-based sensory processing.
For both Task I and II, we have kept d as 500.
B. Results: Performance of NNLD trained by morphological
learning algorithm
Throughout the experiment, the design parameters m, k,
xthr , xsat , τ and T , were chosen as 100, 5, 1, 100, 15 ms
and 400 ms respectively. The detailed criteria for selection of
parameters xthr and xsat is described in [13]. As discussed
in Section III, the firing threshold Vthr is adapted and so it is
randomly initialized before training.
To start with Task I, the NNLD in Fig. 1 was trained
on a small number (= 100) random latency patterns as
generated in Section IV-A1. The results in Fig. 2 (a) and (b)
show that the proposed method can efficiently perform the
classification task. A clear separation between Class 1 and
Class 2 shows that the proposed method is able to respond
to the random single spike latency patterns by shifting Vmax
closer and farther to the Vthr for each pattern in classes 1 and
2 respectively.
The NNLD was also trained for larger number of input
patterns (500 and 1000 patterns) as presented in Fig. 3 (b)
and (c). It shows that the proposed method can perform
the classification task quite well by achieving accuracies of
95.58%(SD = 0.54%) and 86.57%(SD = 0.72%) respectively for these cases. Next, we performed the experiment of
Task II with our network. It was observed that the classification
performances (100%, 100% and 99.71 %(SD = 0.1%) accuracy for 100, 500 and 1000 patterns) are much better than that
of performance in random latency patterns of Task I.Hence,
our method can identify the extra information embedded in
synchrony of neural firings. It also reveals that the learning
4
50
15
10
5
0
90
30
20
10
0
8
10
12
14
16
18
20
100
40
8
10
12
14
16
18
Vmax
Vmax
Accuracy (%)
Number of occurences
Number of occurences
20
80
70
NNLD (adaptive)
NNLD (static)
Tempo. (No quant)
Tempo. (4−bit, DT)
Tempo. (4−bit, AT)
Tempo. (2−bit, DT)
Tempo. (2−bit, AT)
60
50
(a)
(b)
40
105
20
100
100
200
300
400
500
600
700
800
900
1000 1100
800
900
1000 1100
95
10
8
9
10
11
12
13
14
15
Vmax
(a)
260
# of occurences
100
50
0
0
Number of patterns (P)
30
0
# of occurences
30
10
12
Vmax
(b)
14
16
90
85
NNLD (adaptive)
80
NNLD (static)
75
Tempo. (No quant)
Tempo. (4 bit, DT)
70
130
Tempo. (4 bit, AT)
Tempo. (2 bit, DT)
65
Tempo. (2 bit, AT)
60
0
8
Accuracy (%)
# of occurences
Fig. 2: Distribution of Vmax for 100 patterns (a)before and (b) after
training for Task I. Vmax for patterns in P + and P − are shown in
dashed and solid line respectively.
8
10
12
14
16
18
20
Vmax
(c)
Fig. 3: The distribution of Vmax for a set of (a) 100, (b) 500 and (c)
1000 patterns in task I showing 100%, 94.8% and 84.9% accuracy
respectively.
rule does not depend only on a single synchronous pair but
on a huge number of synchronous pairs.
C. Results: Comparison with Tempotron learning
Next, the performance of proposed method is compared with
the Tempotron learning [1] that learns spike time patterns by
using weight updates. The number of synapses used by Tempotron is equal to the number of input afferents d. Since we
are interested in the performance of these algorithms in their
hardware implementations plagued by mismatch, we consider
the performance of the Tempotron when its weight is quantized
at different resolutions. Further, we do the quantization in two
ways: either after training (AT) or as a step during training
(DT). The first method corresponds to the case where weights
trained in software version of the algorithm is downloaded
to the hardware while the second method is analogous to
performing training on-chip. Furthermore, for comparison with
our previous threshold selection technique [19], we have also
calculated the value of Vthr by equating it to the voltage
corresponding to the peak location of Vmax distribution over
10000 random input spike patterns. We term this as Vthr,static .
Fig. 4 depicts that the performance obtained by morphological
learning with adaptive threshold is superior to that of learning
with fixed Vthr,static . Moreover, the comparison results in
Fig. 4 (a) also show that the Tempotron using floatingpoint numbers achieves better performance compared to the
0
100
200
300
400
500
600
700
Number of patterns (P)
Fig. 4: Comparison studies on classification performance, for Tempotron learning [1] and the proposed morphological learning rule
for (a) Task I of Random Latency Patterns and (b) Task II of PairWise Synchrony Patterns. The results have been averaged over 10
independent trials.
proposed method. However, when the high resolution weights
are quantized at 2-bit level, its performance is worse than the
proposed method. Also, it can be seen that the performance is
better when the quantization is performed within the learning
loop. This is to be expected since the learning algorithms can
now try to correct this quantization error as well.
Only at 4-bit quantization level, the classification performance of Tempotron (93.05%(SD = 0.55%) and
89.14%(SD = 0.7%) accuracy for 500 and 1000 patterns)
in Task I is comparable to our proposed method using 1bit or binary weights. This underlines the importance of
our proposed method in robustly implementing spike timing
based classifiers using low-resolution analog synapses available in nano-scale CMOS or non-CMOS devices. Similarly,
the classification performance of NNLD for Task II with 1bit binary weights in Fig. 4 (b) is comparable to performance
of Tempotron at 4-bit quantization which shows about 100 %
and 98.12%(SD = 0.12%) accuracy respectively for 500 and
1000 patterns.
V. A PPLICATION E XAMPLE
Till now the proposed algorithm has been applied to classify
synthetic synchrony patterns. But, to check whether the algorithm can work in real world problems, it is used to classify
Tactile information. The patterns which have been till now
presented for classification had one spike per afferent but real
world scenarios may have multiple spikes per afferent.
5
TABLE I: Performance comparison of morphological learning rule
on NNLD and Tempotron
Cases
Morph. lear. on NNLD (adaptive Vthr )
Morph. lear. on NNLD (for Vthr static )
Tempotron (no quantization) [1]
Tempotron (6-Bit quantization after training)
Tempotron (6-Bit quantization during training)
Tempotron (4-Bit quantization after training)
Tempotron (4-Bit quantization during training)
Tempotron (2-Bit quantization after training)
Tempotron (2-Bit quantization during training)
Accuracy
Mean
SD
96.54
0.6573
95.64
0.6473
97.06
0.5123
95.64
0.7614
96.26
0.6822
78.87
1.8371
85.92
1.3846
51.56
3.765
55.88
2.8731
a) Task Description: A detailed description of the task
can be found in [20]–here, we give a brief description for
completeness. The task requires a flexible, stretchable and
conformable tactile sensor array made up of conductive fabric
which is used for data collection. Two glass spheres (indenters)
of diameter 65 mm and 105 mm were indented onto this
sensor array by a suitably programmed 6-axis robotic arm.
The indentation force used was 4N, as measured using a
load cell placed below the sensor having an accuracy of
0.01N. The signal recording was started just before placing the
indenter onto the sensor array and stopped before removing the
indenter. Between consecutive indentations, a 5 second pause
was provided so that the sensors were able to recover. A total
of 100 recordings were taken per indenter. The collected data
were converted to spikes as described in the next sub-section.
For each indenter, 60 randomly chosen recorded data was used
for training both the proposed algorithm and Tempotron. The
remaining 40 recordings were used for testing.
b) Spike Train Generation: The conversion of the analog
data recorded by the sensor array to spike trains involves
the following steps. First, the analog data is converted to
digital output by applying a fixed threshold of 0.5. This digital
output is inverted to generate another set of digital data. Each
channel of the digital data and its inverted version are passed
through Leaky Integrate and Fire Neuron to generate two sets
of spike trains. These spike trains are combined to form the
spike response to be given as an input to the algorithms. The
usage of both the digital data and its inverted version for spike
generation ensures both low to high and high to low transitions
are captured by a change in firing activity. The combined spike
train patterns given as input to the algorithms consist of 130
afferents. Thus, we allot 130 synapses for both Tempotron and
NNLD.
c) Results: The performance of the proposed method
on this application is compared with Tempotron algorithm at
different quantization levels in Table I. The results, averaged
over 10 independent trials, show that the proposed algorithm
is able to achieve an accuracy of 96.54%(SD = 0.6543%).
Although Tempotron without quantization performs better
(97.06%(SD = 0.5123%)) than the proposed learning rule
on NNLD, but after quantization, at least 6 bits of weight
resolution is needed by Tempotron to match our performance
with 1 bit weights.
VI. DISCUSSION
Here, we compare our method with other reported algorithms and other possible variants of morphological learning.
A. Comparison to other supervised spiking neural classifiers
In recent years, several supervised learning algorithms have
been proposed such as Tempotron [1], ReSuMe [21], SpikeProp [22] and Chronotron [23] for training spiking neurons.
From a pattern recognition viewpoint, these algorithms can be
classified into two types. The general theme of the first type of
algorithms is that a desired output spike train is specified prior
to learning for each class of patterns. SpikeProp, ReSuMe and
Chronotron are examples of this type, among which SpikeProp
can only produce a single output spike whereas the others are
capable of producing multi-spike train. In the second type,
no such desired output spike train is specified beforehand
and the algorithms choose the best time to spike for each
pattern during training. Tempotron and the proposed algorithm
are examples of this type which chooses tmax (defined in
Sec. III) as the time to spike. We have already compared the
performance of our algorithm with Tempotron. We now choose
ReSuMe as a representative of the first type of algorithms and
compare its performance with the proposed learning rule for
the two tasks described in Sec. IV-A. For ReSuMe, the neuron
had to fire one spike at t+ = 350ms for P + patterns and at
t− = 450ms for P − patterns. The classification performance
of ReSuMe (89.58%(SD=1.24%)) and 82.65%(SD=2.22%)
accuracy for 500 and 1000 patterns) in Task I is worse
compared to the proposed method. Similarly, for Task II, the
proposed method performs much better than ReSuMe learning
rule (96.74%(SD = 0.8467%) and 92.12%(SD = 1.1654%)
accuracy for 500 and 1000 patterns).
B. Comparison to other classifiers using dendritic processing
We have already compared the proposed algorithm with
previous works employing NNLD in Section I. Here we
first compare our method with another dendritic algorithm
proposed by Wu et al. in [24]. Unlike [24] which considered
only mean rate encoded inputs, our learning rule can be
applied to arbitrary spike trains.
Second, we compare the proposed algorithm with another
recently reported structural plasticity based algorithm named
Dendritically Enhanced Readout (DER) [14], [15]. In [14],
[15], this structure has been used as the readout stage of
Liquid State Machine. The primary difference of the proposed
algorithm with DER is in the number of data points per pattern
to be memorized for a particular task. If we consider there
are P + and P − patterns of Class 1 and Class 2 respectively,
then the number of data points to be memorized by the
proposed algorithm are Pm N N LD = T × P − + 1 while
the number of data points to be memorized by DER are
Pm DER = T × (P + + P − ) where T is the number of time
points per pattern. So, DER has to memorize almost two times
more data points than the proposed algorithm for the same
number of patterns thereby reducing its memory capacity.
Third, we have also compared our performance with another
recently proposed dendritic algorithm termed as Synaptic
Kernel Inverse Method(SKIM) [25]. The difference mentioned
above in the context of DER also applies to SKIM. Apart from
that, SKIM also uses much more resources than the proposed
NNLD which we will show next. We compare the two methods
6
on Task I for 100 patterns in which case our proposed method
has 100% accuracy. The neural network architecture used
in [25] consists of N presynaptic neurons which connect to
an output spiking neuron, via synaptic connections to its M
dendritic branches. The weights of these synapses are random
and fixed. These synapses along with a subsequent nonlinearity
projects the input to a higher dimension thereby increasing
separability. The dendritic branches sum the synaptic input
currents, and the output from the dendritic branches are
summed at the soma of the output neuron. The weights of the
connection between dendritic branches and the soma are learnt
by Moore pseudo inversion method [25]. Thus, for this network N ×M synaptic resources are required for connecting the
N presynaptic neurons to M postsynaptic dendritic branches
and M synapses are required for connecting the M dendritic
branches to the single output spiking neuron. The number of
input spiking neurons are equal to the number of afferents in
the input data which in our case is 500. Since we have used
500 synapses for NNLD, so initially we keep the number of
postsynaptic dendrites in SKIM as 1 (M = 1) to match the
number of resources used by us. But for M = 1, SKIM fails
miserably and provides only 50% accuracy. When the number
of dendritic branches are increased, SKIM provides better
results and finally is able to provide 100% accuracy when
M = 360. So to provide equivalent result as the proposed
algorithm SKIM requires 361 times more resources than our
proposed NNLD.
VII. CONCLUSION
A morphological learning rule that can be used to find
the optimal morphology of neurons with nonlinear dendrites
(NNLD) and binary synapses is presented. The learning rule
includes a novel threshold adaptation technique. To see the
effectiveness of the proposed method, the NNLD trained with
morphological rule is used to solve two classification tasks
and one real world problem. The results depict that our
proposed method with 1 bit weights can achieve comparable
performance to tempotron learning rule with 4-bit to 6-bit
quantized weights.
ACKNOWLEDGEMENT
The authors acknowledge helpful discussions with Prof.
Nitish Thakor regarding tactile sensing.
R EFERENCES
[1] R. Gutig and H. Sompolinsky, “The tempotron: a neuron that learns
spike timing-based decisions”, Nature Neuroscience, vol. 9, no. 3, pp.
420-428, 2006.
[2] R. Johansson and I. Birznieks, “First spikes in ensembles of human
tactile afferents code complex spatial fingertip events”, Nature Neuroscience, vol. 7, no. 3, pp. 170-177, 2004.
[3] A. Basu and C. Petre, and P. Hasler, “Bifurcations in a Silicon Neuron”,
Proceedings of the International Symposium on Circuits and Systems,
pp. 428-431, 2013.
[4] J. Arthur and K. Boahen, “Synchrony in silicon: The gamma rhythm”,
IEEE Transactions on Neural Networks, no. 6, pp. 1815-1825, Nov.
2007.
[5] S. Millner, A. Grübl, K. Meier, J. Schemmel and M. O. Schwartz, “A
VLSI Implementation of the Adaptive Exponential Integrate-and-Fire
Neuron Model,” Advances in Neural Information Processing Systems,
Vancouver, B.C., Canada, Dec. 2010, pp. 1642-1650.
[6] F. Grassia, L. Buhry, T. Lévi, J. Tomas, A. Destexhe, and S. Saı̈ghi,
“Tunable Neuromimetic Integrated System for Emulating Cortical
Neuron Models,” Frontiers in Neuroscience, vol. 5, no. 134, 2011.
doi:10.3389/fnins.2011.00134
[7] B. Linares-Barranco, T. Serrano-Gotarredona and R. SerranoGotarredona, “Compact Low-power Calibration Mini-DACs for Neural
Arrays with Programmable Weights,” IEEE Transactions on Neural
Networks, vol. 14, no. 5, pp. 1207-15, Nov. 2003.
[8] D. Querlioz, O. Bichler, P. Dollfus and C. Gamrat,“Immunity to Device
Variations in a Spiking Neural Network with Memristive Nanodevices,”
IEEE Transactions on Nanotechnology, vol. 12, no. 3, pp. 288-95, May
2013.
[9] K. H. Kim, S. Gaba, D. Wheeler et.al., “A functional hybrid memristor crossbar-Array/CMOS system for data storage and neuromorphic
applications,” Nano Letters, vol. 12, no. 1, pp. 389-395, 2011.
[10] G. Indiveri, B. linares-Barranco, R. Legenstein, G. Deligeorgis and T.
Prodromakis, “Integration of nanoscale memristor synapses in neuromorphic computing architectures,” Nanotechnology, vol. 24, no. 38,
Sep. 2013.
[11] M. Sharad, C. Augustine, G. Panogopoulos and K. Roy, “Spin-Based
Neuron model with Domain-Wall Magnets as Synapse,” IEEE Transactions on Nanotechnology, vol. 11, pp. 843-53, Jul. 2012.
[12] P. Poirazi and B. W. Mel, “Impact of active dendrites and structural
plasticity on the memory capacity of neural tissue”, Neuron, vol. 9, no.
3, pp. 779-776, 2001.
[13] S. Hussain, R. Gopalakrishnan, A. Basu and S. C. Liu, “A Morphological Learning: Increased Memory Capacity of Neuromorphic
Systems with Binary Synapses exploiting AER based Reconfiguration”,
Proceedings of the International Joint Conference on Neural Networks,
Dallas, Texas, Aug. 2013, pp. 1-7.
[14] S. Roy, A. Basu and S. Hussain, “Hardware efficient, neuromorphic
dendritically enhanced readout for liquid state machines”, Proceedings
of the Biomedical Circuits and Systems Conference, Rotterdam, The
Netherlands, Oct. 2013, pp. 302-305.
[15] S. Roy, A. Banerjee and A. Basu, “Liquid State Machine With
Dendritically Enhanced Readout for Low-Power, Neuromorphic VLSI
Implementations,” IEEE Transactions on Biomedical Circuits and
Systems, vol.8, pp.681-695, Oct. 2014.
[16] K. Boahen, “Point-to-point connectivity between neuromorphic chips
using address events”, IEEE Transactions on Circuits and Systems II,
vol. 47, pp. 416-434, May 2000.
[17] S. Brink, S. Nease, P. Hasler, S. Ramakrishnan, R. Wunderlich, A.Basu
and B. Degnan, “A learning-enabled neuron array IC based upon transistor channel models of biological phenomenon”, IEEE Transactions
on Biomedical Circuits and Systems, vol. 7, Mar. 2013, pp. 71-81.
[18] S. Hussain, A. Basu, M. Wang and T. Hamilton, “DELTRON: Neuromorphic architectures for delay based learning”, Proceedings of the
Asia Pacific Conference on Circuits and Systems, Kaohsiung, Taiwan,
Dec. 2012, pp. 304-307.
[19] P. P. San, A. Basu and S. Hussain, “Spike-timing dependent morphological learning for a neuron with nonlinear active dendrites”, in
Proceedings of the IEEE IJCNN, Beijing, China, Jul. 2014, pp. 3192
- 3196.
[20] W. W. Lee, J. Cabibihan and N. V. Thakor , “Bio-mimetic strategies
for tactile sensing”, in Proceedings of IEEE Sensors Conference,
Baltimore, MD, Nov. 2013, pp. 1-4.
[21] F. Ponulak and A. J. Kasinski, “Supervised Learning in Spiking Neural
Networks with ReSuMe: Sequence Learning, Classification, and Spike
Shifting.’, Neural Computation, vol. 22, no. 2, pp. 467-510, 2010.
[22] S. M. Bohte, J. N. Kok, and H. La Poutre, “Error-backpropagation in
temporally encoded networks of spiking neurons”, Neurocomputing,
vol.48, no. 1, pp. 17-37, 2002.
[23] R. V. Florian, “The chronotron: a neuron that learns to fire temporally
precise spike patterns”, PLoS ONE, vol. 7, no. 8, e40233, 2013.
[24] X.E. Wu and B.W. Mel, “Capacity-enhancing learning rules in a medial
temporal lobe online learning model”, Neuron, vol. 62, no. 1, pp. 31-41,
2009.
[25] J. C. Tapson, G. K. Cohen, S. Afshar et al., “Synthesis of neural
networks for spatio-temporal spike pattern recognition and processing”, Frontiers in Neuroscience, vol. 7, no. 153, Aug. 2013. doi:
10.3389/fnins.2013.00153
| 9cs.NE
|
MAXIMUM-A-POSTERIORI SIGNAL RECOVERY WITH PRIOR INFORMATION:
APPLICATIONS TO COMPRESSIVE SENSING
Ali Bereyhi and Ralf R. Müller
arXiv:1802.05776v1 [cs.IT] 15 Feb 2018
Friedrich-Alexander Universität Erlangen-Nürnberg
[email protected] and [email protected]
ABSTRACT
This paper studies the asymptotic performance of maximuma-posteriori estimation in the presence of prior information.
The problem arises in several applications such as recovery
of signals with non-uniform sparsity pattern from underdetermined measurements. With prior information, the maximuma-posteriori estimator might have asymmetric penalty. We
consider a generic form of this estimator and study its performance via the replica method. Our analyses demonstrate
an asymmetric form of the decoupling property in the largesystem limit. Employing our results, we further investigate
the performance of weighted zero-norm minimization for recovery of a non-uniform sparse signal. Our investigations illustrate that for a given distortion, the minimum number of required measurements can be significantly reduced by choosing weighting coefficients optimally.
Index Terms— Maximum-a-posteriori estimation, compressive sensing, weighted norm minimization, decoupling
property, replica method
1. INTRODUCTION
The problem of estimating x ∈ XN , for some X ⊂ R, from
y = Ax + z,
(1)
with A ∈ RK×N and z ∼ N (0, λ0 IK ), arises in various applications. In presence of prior information, the MaximumA-Posteriori (MAP) estimation approach might deal with an
asymmetric penalty term appearing due to the non-identical
prior distributions. In this paper, we intend to investigate the
asymptotic performance of this class of estimators which encloses several reconstruction schemes in signal processing.
Particular examples of these estimators are the weighted
norm minimization schemes [1] in compressive sensing [2, 3]
which are employed for recovery of signals with non-uniform
sparsity patterns. In this problem, the signal consists of multiple sparse blocks whose sparsity factors are different. A
restricted class of such non-uniform sparse settings, in which
This work was supported by the German Research Foundation, Deutsche
Forschungsgemeinschaft (DFG), under Grant No. MU 3735/2-1.
the signal support is partially known, was addressed in [4],
and the modified-CS scheme was proposed for signal recovery. Weighted ℓ1 -norm minimization was further invoked in
[5] for non-uniform sparse recovery in which different blocks
of signal samples have different sparsity factors. More general settings were investigated in recent studies; see [6–11]
and the references therein.
Due to the nonlinear nature of the MAP estimator, basic tools fail to investigate its large-system performance. Several studies thus invoked the replica method for investigation.
This method was developed for analysis of spin glasses [12]
in the physics literature and accepted as an efficient mathematical tool in information theory; e.g., [13]. The method was
moreover employed to investigate the performance of various recovery schemes in large compressive sensing systems
[14–17]. For non-uniform sparse models, the method was
employed in [18] to study the performance of weighted ℓ1 norm minimization recovery considering noise-free measurements. In this paper, we consider a generic class of estimators
which includes formerly studied schemes such as weighted
ℓ1 -norm minimization and also encloses several other settings
whose performances have not yet been addressed in the literature. Invoking our results we derive an asymmetric version
of the MAP decoupling principle which extends the results
of [19, 20] to a larger class of estimators.
2. PROBLEM FORMULATION
Consider (1) with K/N = α < ∞ as N ↑ ∞. Let [N ] :=
{1, . . . , N } be partitioned into disjoint subsets Nj for j ∈ [J].
J is assumed to be fixed and bounded meaning that J/N ↓ 0
as N grows large. The signal xN ×1 is divided into J blocks.
The block j is denoted by Bj (x) and contains entries whose
indices are in Nj , i.e., Bj (x) := {xn : n ∈ Nj }. We use the
notation j(n) to denote the index of the block to which xn
belongs, i.e., xn ∈ Bj(n) (x). The entries of x are independent, and xn ∼ pj(n) (xn ; ρn ) where {ρn } is a deterministic
sequence over [N ]. The signal is reconstructed from y as
x̂ = argmin
v∈
Xn
1
ky − Avk2 + u(v; c)
2λ
(2)
where λ is the estimation parameter, cN ×1 contains weighting coefficients {cn }, and u(v; c) is a penalty function with
decoupling property, i.e., there exist {uj (vn ; cn )} such that
u(v; c) =
N
X
(3)
uj(n) (vn ; cn ).
n=1
T
T
A is assumed to be random, such that J = A A = UDU
with U being Haar distributed and D denoting the diagonal
matrix of eigenvalues. A trivial example is a matrix with independent and identically distributed (i.i.d.) entries. The empirical distribution of eigenvalues when N ↑ ∞ is denoted by
pJ (λ). For this
distribution, the Stieltjes transform is given by
GJ (s) = E (λ − s)−1 where λ ∼ pJ (λ), and the R-trans−1
form is defined as RJ (ω) := G−1
with G−1
J (−ω) − ω
J (·)
being the inverse with respect to (w.r.t.) composition.
The setting recovers several problems in signal processing.
An example is recovery of non-uniform sparse signals from
noisy measurements in compressive sensing: Let J = 1 and
p1 (xn ; ρn ) = ρn q(xn ) + (1 − ρn )δ(xn )
(4)
for some distribution q(xn ). Then, x models a sparse signal
with non-uniform sparsity pattern whose non-zero entries are
distributed with q(xn ). Consequently, by setting u(vn ; cn ) =
cn |vn |, the estimator reduces to the weighted ℓ1 -norm minimization recovery scheme. For λ0 = 0, the setup recovers the
formerly studied noise-free case, e.g., [5, 8, 18], when λ ↓ 0.
In order to quantify the large-system performance of this
setting, we define the weighted distortion as follows.
Definition 1 (Weighted Distortion): Let wN ×1 enclose the
coefficients {wn }. The weighted distortion w.r.t. the distortion function d(·; ·) for a given w reads
N
D(x; x̂|w) :=
1X
wn E {d(xn ; x̂n )}.
N n=1
(5)
Moreover, the asymptotic weighted distortion is given by taking the limit N ↑ ∞, i.e., Dw := limN ↑∞ D(x; x̂|w).
The weighted distortion recovers various forms of recovery distortions. For instance, setting wn = 1 and d(xn ; x̂n ) =
|xn − x̂n |2 , Dw determines the asymptotic Mean Square Error (MSE). Moreover, it evaluates the average error probability by setting d(xn ; x̂n ) = 1 {xn = x̂n } with 1 {·} being the
indicator function. The main goal of this study is to derive the
weighted distortion in its generic form when N grows large.
3. ASYMPTOTIC PERFORMANCE
Invoking the replica method, Dw is derived in a closed form.
The derivations are briefly sketched in Section 5. For the sake
of compactness, we state the basic form of the result known
as the “Replica Symmetry (RS) solution”. Our derivations
are however in a general form enclosing “Replica Symmetry
Breaking (RSB) solutions”.
3.1. Asymptotic Weighted Distortion
Dw in the large-system limit can be expressed in terms of an
equivalent scalar system. For j ∈ [J], we define the scalar
estimator gjdec (·; c) which for given c and θ reads
gjdec (y; c) = argmin
v∈
X
1
(y − v)2 + uj (v; c)
2θ
(6)
gjdec (y; c) represents a estimator which recovers a scalar from
the single measurement y using the one-dimensional form of
MAP formulation in (2) with the weighting coefficient c and
estimation parameter θ. In order to state the result, we moreover define the effective noise variance θ0 and the equivalent
estimation parameter θ for some scalars χ and p as
h
χ i−1
λ
θ = RJ (− )
λ
i
h
χ i
χ −2 ∂ h
(λ0 χ − λp)RJ (− )
θ0 = RJ (− )
λ
∂χ
λ
(7a)
(7b)
where RJ (·) denotes the R-transform of pJ (t) defined in the
previous section. One should note that θ and θ0 are controlled
by χ and p and are functions of the true estimation parameter
λ, statistics of A and the true noise variance λ0 .
Proposition 1: Let z dec ∼ N (0, θ0 ), and for each n ∈ [N ],
define the decoupled estimation gn as
dec
(xn + z dec ; cn ).
gn := gj(n)
(8)
Then, under some assumptions1 , Dw is given by
(9)
Dw = hwn E {d (xn ; gn )}i[N ]
P
where we define hf (an )iN := |N|−1 n∈N f (an ). The variables p and χ which determine θ and θ0 are moreover calculated from the fixed-point equations
(10a)
p = E (gn − xn )2 [N ] ,
θ0
χ = E (gn − xn ) z dec
θ
[N ]
.
(10b)
Proof: The proof is briefly sketched in Section 5. The details
of the proof, however, are skipped due to the page limitation.
3.2. Asymmetric Decoupling Property
Proposition 1 determines the asymptotic weighted distortion by averaging the scalar systems shown in Fig. 1 over n
w.r.t. w. In fact by setting xn = xn and x̂n = gn in this
diagram, one observes that Dw is the weighted average of
input-output distortions. These scalar systems can be further
shown to describe input-output marginal distributions. This
1 These assumptions
are mainly replica continuity and the replica symmetry which are later introduced in Section 5.
z dec
xn
+
yn
gj(n) (·; cn )
x̂n
Fig. 1: Asymmetric Decoupling Property: The decoupled systems
are dependent on the index n in general.
observation states that the estimator exhibits the decoupling
property in the large-system limit. To illustrate this property,
let us denote the marginal joint distribution of (x̂n , xn ) with
qN (x̂n , xn ) where the subscript indicates the dependency of
the distribution on N . The asymptotic decoupling property
mainly claims that as N grows, qN (x̂n , xn ) converge to a
deterministic distribution described by the input-output distribution of the scalar system in Fig. 1. The previously studied
forms of the property, e.g., [19, 20], have considered identically distributed source entries, i.e., pj(n) (·; ρn ) = p(·; ρ)
and uj(n) (·; cn ) = u(·; c) for some constants ρ and c. For
this case, the limiting distribution is shown to be independent
of n, and thus, the equivalent scalar systems are the same.
The decoupled system derived in this paper, however, can
vary from one index to another. We therefore refer to this
form of decoupling as the “asymmetric decoupling property”
which recovers the previous “symmetric” forms. The property is stated in the following. The proof follows the moment
method and takes a similar path as in [16] with some modifications. It is however omitted for the sake of compactness.
Asymmetric Decoupling: For n ∈ [N ], (x̂n , xn ) converges
in distribution to the pair (x̂n , xn ) in Fig. 1 with gjdec (·; cn )
and z dec being given in Proposition 1.
4. APPLICATIONS OF THE MAIN RESULTS
The asymptotic results presented in Section 3 can be employed to investigate various estimation problems. In the sequel, we give some examples in compressive sensing.
4.1. Recovery of Non-uniform Sparse Signals
Stochastic signals with non-uniform sparsity patterns are described by our setting when J = 1 and the signal entries xn
are distributed as in (4). Several recovery schemes, some of
which have not been addressed in the literature, can then be
investigated by choosing corresponding utility functions. A
trivial approach is to let the utility function be
u(v; c) =
N
X
n=1
cn |vn |p .
(11)
Using Proposition 1, the large-system performance of these
recovery schemes can be studied w.r.t. various forms of distortions. Moreover, the optimal choices for {cn } can be found
in terms of the priors {ρn }, such that the average distortion
is minimized. This investigation widens the scope of analyses in [18] to noisy scenarios and various recovery schemes.
Moreover, it enables us to extend the recent study in [21] to
cases with prior information on the sparsity pattern. To discuss further the application of the results in recovery of nonuniform sparse signals, we consider the following example.
Example 1: Assume that x is a sparse-Gaussian signal with
a non-uniform sparsity pattern, i.e., J = 1 and the distribution of xn for n ∈ [N ] are given by p1 (xn ; ρn ) in (4) with
q(xn ) being the zero-mean and unit-variance Gaussian distribution. To recover the signal, we employ the weighted zeronorm recovery scheme which is given by setting u(vn ; cn ) =
cn 1 {vn 6= 0} in (2). Proposition 1 enables us to investigate the recovery performance in this case and also evaluate
the optimal choice of {cn } in terms of {ρn }. For the sake
of simplicity, consider the scenario in which A is an i.i.d.
matrix whose entries are zero-mean with variance 1/K. In
this case, pJ follows the Marcenko-Pastur law [22], and thus,
RJ (ω) = α(α − ω)−1 which implies θ = λ + α−1 χ and
θ0 = λ + α−1 p. Moreover,
(
yn |yn | > tn
(12)
gdec (yn ; cn ) =
0
|yn | ≤ tn
√
where tn := 2θcn . Consequently, the asymptotic distortion w.r.t. some given distortion function and w is determined
by Proposition 1. As (12) shows, weighted zero-norm recovery decouples asymptotically into a set of hard thresholding
operators whose threshold levels depend on weights cn . By
setting cn = 1 and ρn = ρ for all n ∈ [N ], the decoupled
setups reduce to the symmetric setups reported in [19, 20].
To investigate the performance of weighted zero-norm recovery numerically, we consider the configuration in which
(
ρ0 n ∈ [N/B],
ρn =
(13)
ρ1 n ∈ [N/B + 1 : N ],
for some ρ0 , ρ1 ∈ [0, 1] and some integer B being a divisor
of N . Here, [M : N ] denotes {M, . . . , N }. Moreover, we set
(
1 n ∈ [N/B],
cn =
(14)
c n ∈ [N/B + 1 : N ],
for some c. We denote
the asymptotic average MSE by
mse := limN ↑∞ E kx − x̂k2 /N . Moreover, for a given
mse0 , we define the threshold compression rate Rt (mse0 ) to
be the maximum possible inverse load factor α−1 = N/K
which results in mse ≤ mse0 . Fig. 2 shows the threshold
compression rate as a function of c for mse0 = −25 dB. The
curves have been plotted for ρ0 = 0.1 considering various
choices of ρ1 and B. The noise power is set to be λ0 = 0.01
and λ is tuned such that the MSE is minimized at each load
8
where F (β, h) = E {log Z(β, h)} /N . As evaluating a logarithmic expectation is not a trivial task, we invoke the replica
method. The main idea comes from the Riesz equality [23]
which states E {log x} = limm↓0 log E {xm } /m. Using this
equality, Dw is determined in terms of the mth moment of
Z(β, h). Nevertheless, the moments need to be determined
for real values of m which is still challenging. This challenge
is addressed by assuming “replica continuity” which means
that E {Z m (β, h)} analytically continues from m ∈ Z+ to
m ∈ R+ . After calculating the moments, Dw is given by
X
E wn d(v; xn )pβn (v|xn ) [N ] , (17)
Dw = lim lim
B = 5, ρ1 = 0.010
B = 5, ρ1 = 0.005
B = 7, ρ1 = 0.005
Rt (mse0 = −25 dB)
7
6
5
β↑∞ m↓0
1
1.5
2
2.5
3
c
Fig. 2: Rt (mse0 ) at mse0 = −25 dB versus the norm-weight of
the more sparse block in Example 1. As the figure shows, either the
growth in the size of the block or reduction in its sparsity increases
the degradation caused by uniform recovery, i.e., c = 1.
factor. As the figure shows, the optimal choice of c can significantly increase the threshold compression rate. The curves
moreover indicate that as B grows or ρ1 reduces the gap
between the optimal Rt (mse0 ), maximized over c, and the
threshold compression rate at c = 1 increases. This observation is due to the fact that the growth in B or the reduction in
ρ1 imposes more asymmetry into the setting, and therefore,
increases the loss caused by uniform recovery, i.e., c = 1.
4.2. Non-uniform Sparsity with Multiple Prior States
The non-uniform sparsity model can be extended to signals
with multiple prior states by considering J > 1. In this case,
pj(n) (xn ; ρn ) = ρn qj(n) (xn ) + (1 − ρn )δ(xn )
(15)
represents a signal with non-uniform sparsity pattern whose
non-zero entries are taken from multiple possible prior distributions. This model describes a scenario in which multiple
uncorrelated non-uniform sparse signals are simultaneously
measured, e.g., a network of independent sensors with different prior distributions. An efficient approach for signal
recovery in this case is to set uj(n) (vn ; cn ) = fj(n) (vn ) +
cn 1 {vn 6= 0} for some fj (v). Similar to the case with single state sources, the optimal choice of {cn } as well as the
asymptotic distortion is determined using Proposition 1.
5. LARGE-SYSTEM ANALYSIS
In this section, we briefly sketch the derivations based on the
2
replica method. Consider
PE(v) = ky − Avk /2λ + u(v; c),
and define Z(β, h) = v exp{−βE(v) + hN D(x; v|w)}.
One can then employ large deviation arguments and write
Dw = lim lim
N ↑∞ β↑∞
∂
F (β, h)|h=0 .
∂h
(16)
v
for v ∈ X , where xn is an
× 1 vector with all the entries
Pm
m
being xn and d(v; xn ) := a=1 d(va ; xn ); moreover,
m
e−β [(v−xn ) R(v−xn)+uj(n) (v;cn )]
. (18)
=P
−β [(v−xn )H R(v−xn )+uj(n) (v;cn )]
ve
P
with uj (v; cn ) := m
a=1 uj (va ; cn ), and R := TRJ (−βTQ)
1
λ0
for T= (Im −β 1m ) and some Qm×m which satisfies
2λ
λ
X
β
E pn (v|xn )(v − xn )(v − xn )H [N ] . (19)
Q=
H
pβn (v|xn )
v
In (17), the general replica solution is given. The explicit determination of Dw , however, needs Q to be found such that
(19) is fulfilled. To do so, we need to suppose a structure for
Q. The basic structure is given by RS as Q = χβ −1 Im +p1m
for some χ and p. By substituting Q in (17), Proposition 1 is
concluded after some lines of derivations. The RSB solutions
are further derived by extending the RS structure to
Q=
b
X
χ
Im +
cκ I mβ ⊗ 1 µβκ + p1m ,
µκ
β
κ=1
(20)
for some integer b. The derivations under RSB follow [16,
Appendix D] and are omitted due to the page limitation.
6. CONCLUSION
In this paper, we have studied the asymptotic performance of
a class of MAP-based signal recovery schemes when prior
information is available for reconstruction. Our analysis has
demonstrated an asymmetric version of the decoupling principle for these estimators which generalizes the formerly studied forms of MAP decoupling [19, 20]. Invoking the results,
we have investigated the performance of weighted zero-norm
minimization for recovery of a signal with non-uniform sparsity pattern. The results of this paper can be further employed
to study various problems. A particular example in compressive sensing is to extend the scope of investigations in [21] to
signals with non-uniform sparsity patterns and study the impact of replacing the ℓ1 -norm with an ℓp -norm in the weighted
norm minimization scheme for 0 ≤ p ≤ 1. Currently, the
work in this direction has been started.
7. REFERENCES
[1] Emmanuel J Candes, Michael B Wakin, and Stephen P
Boyd, “Enhancing sparsity by reweighted ℓ1 minimization,” Journal of Fourier analysis and applications, vol.
14, no. 5, pp. 877–905, 2008.
[2] David L Donoho, “Compressed sensing,” IEEE Trans.
on Inf. Theory, vol. 52, no. 4, pp. 1289–1306, 2006.
[3] Emmanuel J Candes, Justin K Romberg, and Terence
Tao, “Stable signal recovery from incomplete and inaccurate measurements,” Comm. on pure and applied
mathematics, vol. 59, no. 8, pp. 1207–1223, 2006.
[4] Namrata Vaswani and Wei Lu, “Modified-CS: Modifying compressive sensing for problems with partially
known support,” IEEE Trans. on Signal Processing, vol.
58, no. 9, pp. 4595–4607, 2010.
[5] M Amin Khajehnejad, Weiyu Xu, A Salman Avestimehr, and Babak Hassibi, “Weighted ℓ1 minimization
for sparse recovery with prior information,” in Inf. Theory (ISIT), IEEE International Symposium on. 2009, pp.
483–487.
[6] Laurent Jacques, “A short note on compressed sensing
with partially known signal support,” Signal Processing,
vol. 90, no. 12, pp. 3308–3312, 2010.
[7] Hassan Mansour and Özgür Yilmaz, “Support driven
reweighted ℓ1 minimization,” in Acoustics, Speech and
Signal Processing (ICASSP), IEEE International Conference on. 2012, pp. 3309–3312.
[8] Jonathan Scarlett, Jamie S Evans, and Subhrakanti
Dey, “Compressed sensing with prior information:
Information-theoretic limits and practical decoders,”
IEEE Trans. on Signal Processing, vol. 61, no. 2, pp.
427–439, 2013.
[9] Samet Oymak, M Amin Khajehnejad, and Babak Hassibi, “Recovery threshold for optimal weight ℓ1 minimization,” in Inf. Theory (ISIT), IEEE International
Symposium on. 2012, pp. 2032–2036.
[10] Bubacarr Bah and Rachel Ward, “The sample complexity of weighted sparse approximation.,” IEEE Trans.
Signal Processing, vol. 64, no. 12, pp. 3145–3155,
2016.
[11] Holger Rauhut and Rachel Ward, “Interpolation via
weighted ℓ1 minimization,” Applied and Computational
Harmonic Analysis, vol. 40, no. 2, pp. 321–351, 2016.
[12] Samuel Frederick Edwards and Phil W Anderson, “Theory of spin glasses,” Journal of Physics F: Metal
Physics, vol. 5, no. 5, pp. 965, 1975.
[13] Toshiyuki Tanaka, “A statistical-mechanics approach
to large-system analysis of CDMA multiuser detectors,”
IEEE Trans. on Inf. Theory, vol. 48, no. 11, pp. 2888–
2910, 2002.
[14] Antonia M Tulino, Giuseppe Caire, Sergio Verdu, and
Shlomo Shamai, “Support recovery with sparsely sampled free random matrices,” IEEE Trans. on Inf. Theory,
vol. 59, no. 7, pp. 4243–4271, 2013.
[15] Mikko Vehkaperä, Yoshiyuki Kabashima, and Saikat
Chatterjee, “Analysis of regularized LS reconstruction
and random matrix ensembles in compressed sensing,”
IEEE Trans. on Inf. Theory, vol. 62, no. 4, pp. 2100–
2124, 2016.
[16] Ali Bereyhi, Ralf R Müller, and Hermann SchulzBaldes,
“Statistical mechanics of MAP estimation: General replica ansatz,”
arXiv preprint
arXiv:1612.01980, 2016.
[17] Ali Bereyhi, Ralf R Müller, and Hermann SchulzBaldes, “Replica symmetry breaking in compressive
sensing,” in IEEE Inf. Theory and Applications Workshop (ITA). 2017, pp. 1–7.
[18] Toshiyuki Tanaka and Jack Raymond, “Optimal incorporation of sparsity information by weighted ℓ1 optimization,” in Information Theory (ISIT), IEEE Int. Sym,
on. 2010, pp. 1598–1602.
[19] Sundeep Rangan, Alyson K Fletcher, and Vivek Goyal,
“Asymptotic analysis of MAP estimation via the replica
method and applications to compressed sensing,” in
IEEE Trans. on Inf. Theory, 2012, pp. 1902–1923.
[20] Ali Bereyhi, Ralf Müller, and Hermann Schulz-Baldes,
“RSB decoupling property of MAP estimators,” in IEEE
Inf. Theory Workshop (ITW). 2016, pp. 379–383.
[21] Le Zheng, Arian Maleki, Haolei Weng, Xiaodong
Wang, and Teng Long, “Does ℓp -minimization outperform ℓ1 -minimization?,” IEEE Trans. on Information
Theory, 2017.
[22] Ralf R Müller, Giusi Alfano, Benjamin M Zaidel, and
Rodrigo de Miguel, “Applications of large random matrices in communications engineering,” arXiv preprint
arXiv:1310.5479, 2013.
[23] Frédéric Riesz, “Sur les valeurs moyennes des fonctions,” Journal of the London Mathematical Society,
vol. 1, no. 2, pp. 120–121, 1930.
| 7cs.IT
|
Harmonious Attention Network for Person Re-Identification
Wei Li1
Xiatian Zhu2
Queen Mary University of London1
arXiv:1802.08122v1 [cs.CV] 22 Feb 2018
{wei.li, s.gong}@qmul.ac.uk
Shaogang Gong1
Vision Semantics Ltd.2
[email protected]
Abstract
Existing person re-identification (re-id) methods either
assume the availability of well-aligned person bounding
box images as model input or rely on constrained attention selection mechanisms to calibrate misaligned images.
They are therefore sub-optimal for re-id matching in arbitrarily aligned person images potentially with large human
pose variations and unconstrained auto-detection errors. In
this work, we show the advantages of jointly learning attention selection and feature representation in a Convolutional
Neural Network (CNN) by maximising the complementary
information of different levels of visual attention subject to
re-id discriminative learning constraints. Specifically, we
formulate a novel Harmonious Attention CNN (HA-CNN)
model for joint learning of soft pixel attention and hard
regional attention along with simultaneous optimisation of
feature representations, dedicated to optimise person re-id
in uncontrolled (misaligned) images. Extensive comparative evaluations validate the superiority of this new HACNN model for person re-id over a wide variety of state-ofthe-art methods on three large-scale benchmarks including
CUHK03, Market-1501, and DukeMTMC-ReID.
Figure 1. Examples of attention selection in auto-detected person
bounding boxes used for person re-id matching.
problem of re-id attention selection within person bounding boxes. One common strategy is local patch calibration and saliency weighting in pairwise image matching
[43, 25, 46, 36]. However, these methods rely on handcrafted features without deep learning jointly more expressive feature representations and matching metric holistically
(end-to-end). A small number of attention deep learning
models for re-id have been recently developed for reducing the negative effect from poor detection and human pose
change [17, 42, 27]. Nevertheless, these deep methods implicitly assume the availability of large labelled training
data by simply adopting existing deep architectures with
high complexity in model design. Additionally, they often consider only coarse region-level attention whilst ignoring the fine-grained pixel-level saliency. Hence, these
techniques are ineffective when only a small set of labelled
data is available for model training whilst also facing noisy
person images of arbitrary misalignment and background
clutter.
In this work, we consider the problem of jointly deep
learning attention selection and feature representation for
optimising person re-id in a more lightweight (with less parameters) network architecture. The contributions of this
work are: (I) We formulate a novel idea of jointly learning
multi-granularity attention selection and feature representation for optimising person re-id in deep learning. To our
knowledge, this is the first attempt of jointly deep learning
multiple complementary attention for solving the person reid problem. (II) We propose a Harmonious Attention Con-
1. Introduction
Person re-identification (re-id) aims to search people
across non-overlapping surveillance camera views deployed
at different locations by matching person images. In practical re-id scenarios, person images are typically automatically detected for scaling up to large visual data [44, 18].
Auto-detected person bounding boxes are typically not optimised for re-id due to misalignment with background clutter, occlusion, missing body parts (Fig. 1). Additionally,
people (uncooperative) are often captured in various poses
across open space and time. These give rise to the notorious image matching misalignment challenge in cross-view
re-id [7]. There is consequently an inevitable need for attention selection within arbitrarily-aligned bounding boxes
as an integral part of model learning for re-id.
There are a few attempts in the literature for solving the
1
volutional Neural Network (HA-CNN) to simultaneously
learn hard region-level and soft pixel-level attention within
arbitrary person bounding boxes along with re-id feature
representations for maximising the correlated complementary information between attention selection and feature
discrimination. This is achieved by devising a lightweight
Harmonious Attention module capable of efficiently and
effectively learning different types of attention from the
shared re-id feature representation in a multi-task and endto-end learning fashion. (III) We introduce a cross-attention
interaction learning scheme for further enhancing the compatibility between attention selection and feature representation given re-id discriminative constraints. Extensive
comparative evaluations demonstrate the superiority of the
proposed HA-CNN model over a wide range of state-of-theart re-id models on three large benchmarks CUHK03 [18],
Market-1501 [44], and DukeMTMC-ReID [47].
2. Related Work
Most existing person re-id methods focus on supervised
learning of identity-discriminative information, including
ranking by pairwise constraints [23, 38], discriminative distance metric learning [13, 45, 40, 20, 41, 4], and deep learning [24, 18, 3, 39, 35, 19]. These methods assume that
person images are well aligned, which is largely invalid
given imperfect detection bounding boxes of changing human poses. To overcome this limitation, attention selection techniques have been developed for improving re-id
by localised patch matching [25, 46] and saliency weighting [36, 43]. These are inherently unsuitable by design to
cope with poorly aligned person images, due to their stringent requirement of tight bounding boxes around the whole
person and high sensitivity of the hand-crafted features.
Recently, a few attention deep learning methods have
been proposed to handle the matching misalignment challenge in re-id [17, 42, 27, 16]. The common strategy of
these methods is to incorporate a regional attention selection sub-network into a deep re-id model. For example,
Su et al. [27] integrate a separately trained pose detection
model (from additional labelled pose ground-truth) into a
part-based re-id model. Li et al. [17] design an end-to-end
trainable part-aligning CNN network for locating latent discriminative regions (i.e. hard attention) and subsequently
extract and exploit these regional features for performing
re-id. Zhao et al. [42] exploit the Spatial Transformer Network [11] as the hard attention model for searching re-id
discriminative parts given a pre-defined spatial constraint.
However, these models fail to consider the noisy information within selected regions at the pixel level, i.e. no soft
attention modelling, which can be important. While soft attention modelling for re-id is considered in [22], this model
assumes tight person boxes thus less suitable for poor detections.
The proposed HA-CNN model is designed particularly
to address the weaknesses of existing deep methods as
above by formulating a joint learning scheme for modelling
both soft and hard attention in a single re-id deep model.
This is the first attempt of modelling multi-level correlated
attention in deep learning for person re-id to our knowledge. In addition, we introduce cross-attention interaction
learning for enhancing the complementary effect between
different levels of attention subject to re-id discriminative
constraints. This is impossible to do for existing methods
due to their inherent single level attention modelling. We
show the benefits of joint modelling multi-level attention in
person re-id in our experiments. Moreover, we also design
an efficient attention CNN architecture for improving the
model deployment scalability, an under-studied but practically important issue for re-id.
3. Harmonious Attention Network
Given n training bounding box images I = {Ii }ni=1
from nid distinct people captured by non-overlapping camera views together with the corresponding identity labels as
Y = {yi }ni=1 (where yi ∈ [1, · · · , nid ]), we aim to learn
a deep feature representation model optimal for person reid matching under significant viewing condition variations.
To this end, we formulate a Harmonious Attention Convolutional Neural Network (HA-CNN) that aims to concurrently
learn a set of harmonious attention, global and local feature
representations for maximising their complementary benefit
and compatibility in terms of both discrimination power and
architecture simplicity. Typically, person parts location information is not provided in person re-id image annotation
(i.e. only weakly labelled without fine-grained). Therefore, the attention model learning is weakly supervised in
the context of optimising re-id performance. Unlike most
existing works that simply adopting a standard deep CNN
network typically with a large number of model parameters (likely overfit given small size labelled data) and high
computational cost in model deployment [15, 26, 30, 8], we
design a lightweight (less parameters) yet deep (maintaining strong discriminative power) CNN architecture by devising a holistic attention mechanism for locating the most
discriminative pixels and regions in order to identify optimal visual patterns for re-id. We avoid simply stacking
many CNN layers to gain model depth.
This is particularly critical for re-id where the label data
is often sparse (large models are more likely to overfit in
training) and the deployment efficiency is very important
(slow feature extraction is not scalable to large surveillance
video data).
HA-CNN Overview We consider a multi-branch network
architecture for our purpose. The overall objective of this
multi-branch scheme and the overall architecture composi-
160 × 64 × 3
80 × 32 × 𝑑&
80 × 32 × 32
32, 3×3,2
𝑑&
24 × 28 × 32
Inception-A
40× 16 × 𝑑&
𝑑&
shared
40× 16 × 𝑑'
𝑑'
HA
12 × 14 × 𝑑&
𝑑'
20× 8 × 𝑑'
20 × 8 × 𝑑(
𝑑(
HA
6 × 7 × 𝑑'
shared
𝑑(
shared
𝑑&
𝑑'
𝑑(
𝑑&
𝑑'
𝑑(
𝑑&
𝑑'
𝑑(
Inception-B
HA
Harmonious Attention
Addition
10 × 4 × 𝑑(
𝐿𝑜𝑠𝑠$
HA
512
3 × 4 × 𝑑(
𝐿𝑜𝑠𝑠)
512
Global Average Pooling
Fully-connected
Figure 2. The Harmonious Attention Convoluntional Neural Network. The symbol dl (l ∈ {1, 2, 3}) denotes the number of convolutional
filter in the corresponding Inception unit at the l-th block.
tion is to minimise the model complexity therefore reduce
the network parameter size whilst maintaining the optimal
network depth. The overall design of our HA-CNN architecture is shown in Fig. 2. This HA-CNN model contains two branches: (1) One local branch (consisting of T
streams of an identical structure): Each stream aims to learn
the most discriminative visual features for one of T local
image regions of a person bounding box image. (2) One
global branch: This aims to learn the optimal global level
features from the entire person image. For both branches,
we select the Inception-A/B units [39, 29] as the basic building blocks1 .
In particular, we used 3 Inception-A and 3 Inception-B
blocks for building the global branch, and 3 Inception-B
blocks for each local stream. The width (channel number)
of each Inception is denoted by d1 , d2 and d3 . The global
network ends with a global average pooling layer and a
fully-connected (FC) feature layer with 512 outputs. For the
local branch, we also use a 512-D FC feature layer which
fuses the global average pooling outputs of all streams. To
reduce the model parameter size, we share the first conv
layer between global and local branches and the samelayer Inceptions among all local streams. For our HA-CNN
model training, we utilise the cross-entropy classification
loss function for both global and local branches, which optimise person identity classification.
For attention selection within each bounding box of
some unknown misalignment, we consider a harmonious
attention learning scheme that aims to jointly learn a set of
complementary attention maps including hard (regional) attention for the local branch and soft (spatial/pixel-level and
channel/scale-level) attention for the global branch.
We further introduce a cross-attention interaction learning scheme between the local and global branches for further enhancing the harmony and compatibility degree whilst
simultaneously optimising per-branch discriminative fea1 This choice is independent of our model design and others can be
readily considered such as AlexNet [15], ResNet [8] and VggNet [26].
ture representations.
We shall now describe more details of each component
of the network design as follows.
ℎ ×𝑤 ×𝑐
𝜃
Reduce
ℎ×𝑤, 1
Tanh
1, 3×3, 2
𝑐
, 1×1, 1
𝑟
2𝑇
(d)
Resize
𝑐, 1×1, 1
1, 1×1, 1
(c)
(b)
1 ×1 ×𝑐
ℎ ×𝑤 ×1
𝑐, 1×1, 1
Sigmoid
(a)
(a) whole soft attention
(b) spatial attention
(c) channel attention
(d) hard attention
ℎ × 𝑤× 𝑐
Figure 3. The structure of each Harmonious Attention module
consists of (a) Soft Attention which includes (b) Spatial Attention
(pixel-wise) and (c) Channel Attention (scale-wise), and (d) Hard
Regional Attention (part-wise). Layer type is indicated by background colour: grey for convolutional (conv), brown for global
average pooling, and blue for fully-connected layers. The three
items in the bracket of a conv layer are: filter number, filter shape,
and stride. The ReLU [15] and Batch Normalisation (BN) [10]
(applied to each conv layer) are not shown for brevity.
3.1. Harmonious Attention Learning
Conceptually, our Harmonious Attention (HA) is a principled union of hard regional attention [11], soft spatial
[34] and channel attention [9]. This simulates functionally the dorsal and ventral attention mechanism of human
brain [33] in the sense of modelling soft and hard attention simultaneously. The soft attention learning aims at
selecting the fine-grained important pixels, while the hard
attention learning is dedicated to searching the coarse latent (weakly supervised) discriminative regions. They are
thus largely complementary with high compatibility to each
other in functionality. Intuitively, their combination can relieve the modelling burden of soft attention and resulting
in more discriminative and robust model learning from the
same (particularly small) training data. We propose a novel
Harmonious Attention joint learning strategy to unite the
three distinct types of attention with only a small number of
additional parameters. We take a block-wise (module-wise)
attention design, that is, each HA module is specifically optimised to attend the input feature representations at its own
level alone. In the CNN hierarchical framework, this naturally allows for hierarchical multi-level attention learning
to progressively refine the attention maps, in the spirit of
the divide and conquer design [5]. As a result, we can significantly reduce the attention search space (i.e. the model
optimisation complexity) whilst allow multi-scale selectiveness of hierarchical features to enrich the final feature representations. Such progressive and holistic attention modelling is both intuitive and essential for person re-id due to
that (1) the surveillance person images often have cluttered
background and uncontrolled appearance variations therefore the optimal attention patterns of different images can be
highly variable, and (2) a re-id model typically needs robust
(generalisable) model learning given very limited training
data (significantly less than common image classification
tasks). Next, we describe the design of our Harmonious Attention module in details.
(I) Soft Spatial-Channel Attention The input to a Harmonious Attention module is a 3-D tensor X l ∈ Rh×w×c
where h, w, and c denote the number of pixel in the height,
width, and channel dimensions respectively; and l indicates
the level of this module in the entire network (multiple such
modules). Soft spatial-channel attention learning aims to
produce a saliency weight map Al ∈ Rh×w×c of the same
size as X. Given the largely independent nature between
spatial (inter-pixel) and channel (inter-scale) attention, we
propose to learn them in a joint but factorised way as:
Al = S l × C l
(1)
where S l ∈ Rh×w×1 and C l ∈ R1×1×c represent the spatial
and channel attention maps, respectively. We perform the
attention tensor factorisation by designing a two-branches
unit (Fig. 3(a)): One branch to model the spatial attention S l (shared across the channel dimension), and another
branch to model the channel attention C l (shared across
both height and width dimensions).
By this design, we can compute efficiently the full soft
attention Al from C l and S l with a tensor multiplication.
Our design is more efficient than common tensor factorisation algorithms [14] since heavy matrix operations are eliminated.
(1) Spatial Attention We model the spatial attention by a
tiny (10 parameters) 4-layers sub-network (Fig. 3(b)). It
consists of a global cross-channel averaging pooling layer
(0 parameter), a conv layer of 3 × 3 filter with stride 2 (9
parameters), a resizing bilinear layer (0 parameter), and a
scaling conv layer (1 parameter). In particular, the global
averaging pooling, defined as,
c
S linput =
1X l
X
c i=1 1:h,1:w,i
(2)
is designed especially to compress the input size of the
subsequent conv layer with merely 1c times of parameters
needed. This cross-channel pooling is reasonable because
in our design all channels share the identical spatial attention map. We finally add the scaling layer for automatically
learning an adaptive fusion scale in order to optimally combining the channel attention described next.
(2) Channel Attention We model the channel attention by
2
a small (2 cr parameters, see more details below) 4-layers
squeeze-and-excitation sub-network (Fig. 3(c)). Specifically, we first perform a squeeze operation via an averaging
pooling layer (0 parameters) for aggregating feature information distributed across the spatial space into a channel
signature as
C linput =
h
w
1 XX l
X
h × w i=1 j=1 i,j,1:c
(3)
This signature conveys the per-channel filter response from
the whole image, therefore providing the complete information for the inter-channel dependency modelling in the
subsequent excitation operation, formulated as
C lexcitation = ReLU( W2ca × ReLU(W1ca C linput ))
c
2
(4)
c
2
where W1ca ∈ R r ×c ( cr parameters) and W2ca ∈ Rc× r ( cr
parameters) denote the parameter matrix of 2 conv layers in
order respectively, and r (16 in our implementation) represents the bottleneck reduction rate. Again, this bottleneck
design is for reducing the model parameter number from c2
2
2
(using one conv layer) to ( cr + cr ), e.g. only need 18 times
of parameters when r = 16.
For facilitating the combination of the spatial attention
and channel attention, we further deploy a 1×1× c convolution (c2 parameters) layer to compute blended full soft
attention after tensor multiplication. This is because the
spatial and channel attention are not mutually exclusive but
with a co-occurring complementary relationship. Finally,
we use the sigmoid operation (0 parameter) to normalise
the full soft attention into the range between 0.5 and 1.
Remarks Our model is similar to the Residual Attention
(RA) [34] and Squeeze-and-Excitation (SE) [9] concepts
but with a number of essential differences: (1) The RA
requires to learn a much more complex soft attention subnetwork which is not only computationally expensive but
𝜃
𝑭𝒈
Input
𝑯
Input
𝜃
Parts
(a) STN [11]
𝑭𝒍
𝑯/𝑭𝒈
Parts
𝑭𝒍
(b) HA-CNN Hard Attention
Figure 4. Schematic comparison between (a) STN [11] and (b)
HA-CNN Hard Attention. Global feature and hard attention are
jointly learned in a multi-task design. “H”: Hard attention module; “Fg ”: Global feature module; “Fg ”: Local feature module.
also less discriminative when the training data size is small
typical in person re-id. (2) The SE considers only the channel attention and implicitly assumes non-cluttered background, therefore significantly restricting its suitability to
re-id tasks under cluttered surveillance viewing conditions.
(3) Both RA and SE consider no hard regional attention
modelling, hence lacking the ability to discover the correlated complementary benefit between soft and hard attention learning.
(II) Hard Regional Attention The hard attention learning
aims to locate latent (weakly supervised) discriminative T
regions/parts (e.g. human body parts) in each input image at
the l-th level. We model this regional attention by learning
a transformation matrix as:
s
0 tx
Al = h
(5)
0 sw ty
which allows for image cropping, translation, and isotropic
scaling operations by varying two scale factors (sh , sw ) and
the 2-D spatial position (tx , ty ). We use pre-defined region
size by fixing sh and sw for limiting the model complex.
Therefore, the effective modelling part of Al is only tx and
ty , with the output dimension as 2×T (T the region number). To perform this learning, we introduce a simple 2layers (2×T ×c parameters) sub-network (Fig. 3(d)). We
exploit the first layer output (a c-D vector) of the channel
attention (Eq. (3)) as the first FC layer (2 × T × c parameters) input for further reducing the parameter size while
sharing the available knowledge in spirit of the multi-task
learning principle [6]. The second layer (0 parameter) performs a tanh scaling (the range of [−1, 1]) to convert the
region position parameters into the percentage so as to allow for positioning individual regions outside of the input
image boundary. This specially takes into account the cases
that only partial person is detected sometimes. Note that,
unlike the soft attention maps that are applied to the input
feature representation X l , the hard regional attention is enforced on that of the corresponding network block to generate T different parts which are subsequently fed into the
corresponding streams of the local branch (see the dashed
arrow on the top of Fig 2).
Remarks The proposed hard attention modelling is conceptually similar to the Spatial Transformer Network (STN)
[11] because both are designed to learn a transformation
matrix for discriminative region identification. However,
they differ significantly in design: (1) The STN attention
is network-wise (one level of attention learning) whilst our
HA is module-wise (multiple levels of attention learning).
The latter not only eases the attention modelling complexity (divide-and-conquer design) and but also provides additional attention refinement in a sequential manner. (2)
The STN utilises a separate large sub-network for attention modelling whilst the HA-CNN exploits a much smaller
sub-network by sharing the majority model learning with
the target-task network using a multi-task learning design
(Fig. 4), therefore superior in both higher efficiency and
lower overfitting risk. (3) The STN considers only hard attention whilst HA-CNN models both soft and hard attention
in an end-to-end fashion so that additional complementary
benefits are exploited.
(III) Cross-Attention Interaction Learning Given the
joint learning of soft and hard attention above, we further
consider a cross-attention interaction mechanism for enriching their joint learning harmony by interacting the attended
local and global features across branches. Specifically, at
(l,k)
the l-th level, we utilise the global-branch feature XG
of the k-th region to enrich the corresponding local-branch
(l,k)
feature XL by tensor addition as
(l,k)
X̃L
(l,k)
= XL
(l,k)
+ XG
(6)
(l,k)
where XG is computed by applying the hard regional attention of the (l + 1)-th level’s HA attention module (see
the dashed arrow in Fig. 2). By doing so, we can simultaneously reduce the complexity of the local branch (fewer
layers) since the learning capability of the global branch
can be partially shared. During model training by backpropagation, the global branch takes gradients from both
the global and local branches as
(l)
(l)
∆WG =
∂LG ∂XG
(l)
∂XG
(l)
∂WG
+
T
X
∂LL
(l,k)
k=1 ∂ X̃L
(l,k)
∂ X̃L
(l)
∂WG
(7)
Therefore, the global LG and local LL loss quantities are
(l)
concurrently utilised in optimising the parameters WG of
the global branch. As such, the learning of the global branch
is interacted with that of the local branch at multiple levels,
whilst both are subject to the same re-id optimisation constraint.
Remarks By design, cross-attention interaction learning is
subsequent to and complementary with the harmonious attention joint reasoning above. Specifically, the latter learns
soft and hard attention from the same input feature representations to maximise their compatibility (joint attention generation), whilst the former optimises the correlated complementary information between attention refined global and
local features under the person re-id matching constraint
(joint attention application). Hence, the composition of
both forms a complete process of joint optimisation of attention selection for person re-id.
3.2. Person Re-ID by HA-CNN
Given a trained HA-CNN model, we obtain a 1,024-D
joint feature vector (deep feature representation) by concatenating the local (512-D) and the global (512-D) feature vectors. For person re-id, we deploy this 1,024-D deep
feature representation using only a generic distance metric
without any camera-pair specific distance metric learning,
e.g. the L2 distance. Specifically, given a test probe image I p from one camera view and a set of test gallery images {Iig } from other non-overlapping camera views: (1)
We first compute their corresponding 1,024-D feature vectors by forward-feeding the images to a trained HA-CNN
model, denoted as xp = [xpg ; xpl ] and {xgi = [xgg ; xgl ]}. (2)
We then compute L2 normalisation on the global and local features, respectively. (3) Lastly, we compute the crosscamera matching distances between xp and xgi by the L2
distance. We then rank all gallery images in ascendant order by their L2 distances to the probe image. The probabilities of true matches of probe person images in Rank-1 and
among the higher ranks indicate the goodness of the learned
HA-CNN deep features for person re-id tasks.
4. Experiments
(a) CUHK03
(b) Market-1501
(c) DukeMTMC
Figure 5. Example cross-view matched pairs from three datasets.
Datasets and Evaluation Protocol For evaluation, we selected three large-scale person re-id benchmarks, Market1501 [37], DukeMTMC-ReID [47] and CUHK03 [18]. Figure 5 shows several example person bounding box images.
We adopted the standard person re-id setting including the
training/test ID split and test protocol (Table 1). For performance measure, we use the cumulative matching characteristic (CMC) and mean Average Precision (mAP) metrics.
Table 1. Re-id evaluation protocol. TS: Test Setting; SS: SingleShot; MS: Multi-Shot. SQ: Single-Query; MQ: Multi-Query.
Dataset
# ID # Train # Test# Image Test Setting
CUHK03
1,467 767
Market-1501 1,501 751
DukeMCMT-ReID 1,402 702
700 14,097
750 32,668
702 36,411
SS
SQ/MQ
SQ
Implementation Details We implemented our HA-CNN
model in the Tensorflow [1] framework. All person images
are resized to 160×64. For HA-CNN architecture, we set the
width of Inception units at the 1st /2nd /3rd levels as: d1 = 128,
d2 = 256 and d3 = 384. Following [19], we use T = 4 regions for hard attention, e.g. a total of 4 local streams. In
each stream, we fix the size of three levels of hard attention
as 24×28, 12×14 and 6×7. For model optimisation, we use
the ADAM [12] algorithm at the initial learning rate 5×10−4
with the two moment terms β1 = 0.9 and β2 = 0.999.
We set the batch size to 32, epoch to 150, momentum to
0.9. Note, we do not adopt any data argumentation methods (e.g. scaling, rotation, flipping, and colour distortion),
neither model pre-training. Existing deep re-id methods
typically benefit significantly from these operations at the
price of not only much higher computational cost but also
notoriously difficult and time-consuming model tuning.
4.1. Comparisons to State-of-the-Art Methods
Evaluation on Market-1501 We evaluated HA-CNN
against 13 existing methods on Market-1501. Table 2 shows
the clear performance superiority of HA-CNN over all stateof-the-arts with significant Rank-1 and mAP advantages.
Specifically, HA-CNN outperforms the 2nd best model
JLML (pre-defined hard attention based) by 6.1% (91.285.1) (SQ) and 4.1% (93.8-89.7) (MQ) in Rank-1; 10.2%
(75.7-65.5) (SQ) and 8.3% (82.8-74.5) (MQ) in mAP. Compared to the only soft attention alternative HPN, our model
improves the Rank-1 by 14.3% (91.2-76.9) (SQ). This indicates the superiority of our factorised spatial and channel
soft attention modelling over HPN’s multi-directional attention mechanism. HA-CNN also surpasses recent hard attention re-id methods (MSCAN, DLPA and PDC), boosting the
Rank-1 by 10.9%, 10.2% and 7.1%, mAP by 18.2%, 12.3%
and 12.3% (SQ), respectively. These validate the significant advantage of our harmonious soft/hand attention joint
and interaction learning over existing methods replying on
either hard or soft attention at a single level.
Evaluation on DukeMTMC-ReID We evaluated HACNN on the recently released DukeMTMC-ReID dataset2 .
Compared to Market-1501, person images from this benchmark have more variations in resolution and background
due to wider camera views and more complex scene layout,
therefore presenting a more challenging re-id task. Table 3
shows that HA-CNN again outperforms all compared stateof-the-arts with clear accuracy advantages, surpassing the
2nd best SVDNet-ResNet50 (without attention modelling)
by 3.8% (80.5-76.7) in Rank-1 and 7.0% (63.8-56.8) in
mAP. This suggests the importance of attention modelling
in re-id and the efficacy of our attention joint learning approach in a more challenging re-id scenario. Importantly,
the performance advantage by our method is achieved at a
2 Only a small number of methods (see Table 3) have been evaluated
and reported on DukeMTMC-ReID.
Table 2. Market-1501 evaluation. 1st /2nd best in red/blue.
Query Type
Measure (%)
XQDA[20]
SCS[2]
DNS[41]
CRAFT[4]
CAN[21]
S-LSTM[32]
G-SCNN[31]
HPN [22]
SVDNet [28]
MSCAN [17]
DLPA [42]
PDC [27]
JLML [19]
HA-CNN
Single-Query
R1
mAP
43.8
22.2
51.9
26.3
61.0
35.6
68.7
42.3
60.3
35.9
65.8
39.5
76.9
82.3
62.1
80.3
57.5
81.0
63.4
84.1
63.4
85.1
65.5
91.2
75.7
Multi-Query
R1
mAP
54.1
28.4
71.5
46.0
77.0
50.3
72.1
47.9
61.6
35.3
76.0
48.4
86.8
66.7
89.7
74.5
93.8
82.8
st
nd
Table 3. DukeMTMC-ReID evaluation. 1 /2 best in red/blue.
Measure (%)
BoW+KISSME [37]
LOMO+XQDA [20]
ResNet50 [8]
ResNet50+LSRO [47]
JLML [19]
SVDNet-CaffeNet [28]
SVDNet-ResNet50 [28]
HA-CNN
R1
25.1
30.8
65.2
67.7
73.3
67.6
76.7
80.5
mAP
12.2
17.0
45.0
47.1
56.4
45.8
56.8
63.8
lower model training and test cost through an much easier
training process. For example, the performance by SVDNet relies on the heavy ResNet50 CNN model (23.5 million
parameters) with the need for model pre-training on the ImageNet data (1.2 million images), whilst HA-CNN has only
2.7 million parameters with no data augmentation.
Evaluation on CUHK03 We evaluated HA-CNN on both
manually labelled and auto-detected (more misalignment)
person bounding boxes of the CUHK03 benchmark. We
utilise the 767/700 identity split rather than 1367/100 since
the former defines a more realistic and challenging reid task. In this setting, the training set is small with
only about 7,300 images (versus 12,936/16,522 in Market1501/DukeMCMT-ReID). This generally imposes a harder
challenge to deep models, particularly when our HA-CNN
does not benefit from any auxiliary data pre-training (e.g.
ImageNet) nor data augmentation. Table 4 shows that HACNN still achieves the best re-id accuracy, outperforming
hand-crafted feature based methods significantly and deep
competitors less so. Our model achieves a small margin
(+0.2% in Rank-1 and +1.3%) over the best alternative
SVDNet-ResNet50 on the detected set. However, it is worth
pointing out that SVDNet-ResNet50 benefits additionally
from not only large ImageNet pre-training but also a much
Table 4. CUHK03 evaluation. The setting is 767/700 training/test
split. 1st /2nd best in red/blue.
Measure (%)
BoW+XQDA [37]
LOMO+XQDA [20]
IDE-C [48]
IDE-C+XQDA [48]
IDE-R [48]
IDE-R+XQDA [48]
SVDNet-CaffeNet [28]
SVDNet-ResNet50 [28]
HA-CNN
Labelled
R1 mAP
7.9
7.3
14.8 13.6
15.6 14.9
21.9 20.0
22.2 21.0
32.0 29.6
44.4 41.0
Detected
R1 mAP
6.4
6.4
12.8 11.5
15.1 14.2
21.1 19.0
21.3 19.7
31.1 28.2
27.7 24.9
41.5 37.3
41.7 38.6
Table 5. Evaluating individual types of attention in our HA model.
Setting: SQ. SSA: Soft Spatial Attention; SCA: Soft Channel Attention; HRA: Hard Regional Attention.
Dataset
Metric (%)
No Attention
SSA
SCA
SSA+SCA
HRA
All
Market-1501
R1
mAP
84.7
65.3
85.5
65.8
86.8
67.9
88.5
70.2
88.2
71.0
91.2
75.7
DukeMTMC-ReID
R1
mAP
72.4
53.4
73.9
54.8
73.7
53.5
76.1
57.2
75.3
58.4
80.5
63.8
larger network and more complex training process. In contrast, HA-CNN is much more lightweight on parameter size
with the advantage of easy training and fast deployment.
This shows that our attention joint learning can be a better replacement of existing complex networks with timeconsuming model training.
4.2. Further Analysis and Discussions
Effect of Different Types of Attention We further evaluated the effect of each individual attention component in our
HA model: Soft Spatial Attention (SSA), Soft Channel Attention (SCA), and Hard Regional Attention (HRA). Table 5
shows that: (1) Any of the three attention in isolation brings
person re-id performance gain; (2) The combination of SSA
and SCA gives further accuracy boost, which suggests the
complementary information between the two soft attention
discovered by our model; (3) When combining the hard and
soft attention, another significant performance gain is obtained. This shows that our method is effective in identifying and exploiting the complementary information between
coarse hard attention and fine-grained soft attention.
Effect of Cross-Attention Interaction Learning We also
evaluated the benefit of cross-attention interaction learning
(CAIL) between global and local branches. Table 6 shows
that CAIL has significant benefit to re-id matching, improving the Rank-1 by 4.6%(91.2-86.6) / 6.5%(80.5-74.0), mAP
by 9.5%(75.7-66.2) / 8.4%(63.8-55.4) on Market-1501 /
Table 6. Evaluating cross-attention interaction learning (CAIL).
Setting: SQ.
Dataset
Metric (%)
Market-1501
R1
mAP
DukeMTMC-ReID
R1
mAP
w/o CAIL
w/ CAIL
86.6
91.2
74.0
80.5
66.2
75.7
55.4
63.8
DukeMTMC-ReID, respectively. This validates our design
is rational that it is necessary to jointly learn the attended
feature representations across soft and hard attention subject to the same re-id label constraint.
Effect of Joint Local and Global Features We evaluated
the effect of joint local and global features by comparing their individual re-id performances against that of the
joint feature. Table 7 shows: (1) Either feature representation alone is already very discriminative for person re-id.
For instance, the global HA-CNN feature outperforms the
best alternative JLML [19] (Table 2) by 4.8%(89.9-85.1) in
Rank-1 and by 7.0%(72.5-65.5) in mAP (SQ) on Market1501. (2) A further performance gain is obtained by joining
the two representations, yielding 6.1%(91.2-85.1) in Rank1 boost and 10.2%(75.7-65.5) in mAP increase. Similar
trends are observed on the DukeMCMT-ReID (Table 3).
These validate the complementary effect of jointly learning
local and global features in harmonious attention context by
our HA-CNN model.
Table 7. Evaluating global-level and local-level features. Setting:
SQ.
Dataset
Metric (%)
Global
Local
Global+Local
Market-1501
R1
mAP
DukeMTMC-ReID
R1
mAP
89.9
88.9
91.2
78.9
77.3
80.5
72.5
71.7
75.7
60.0
59.5
63.8
Visualisation of Harmonious Attention We visualise both
learned soft attention and hard attention at three different
levels of HA-CNN. Figure 6 shows: (1) Hard attention
localises four body parts well at all three levels, approximately corresponding to head+shoulder (red), upper-body
(blue), upper-leg (green) and lower-leg (violet). (2) Soft attention focuses on the discriminative pixel-wise selections
progressively in spatial localisation, e.g. attending hierarchically from the global whole body by the 1st -level spatial
SA (c) to local salient parts (e.g. object associations) by
the 3rd -level spatial SA (g). This shows compellingly the
effectiveness of joint soft and hard attention learning.
Model Complexity We compare the proposed HA-CNN
model with four popular CNN architectures (Alexnet [15],
VGG16 [26], GoogLeNet [30], and ResNet50 [8]) in model
size and complexity. Table 8 shows that HA-CNN has the
smallest model size (2.7 million parameters) and the 2nd
smallest FLOPs (1.09 × 109 ) and yet, still retains the 2nd
deepest structure (39).
(f) (g)
(a) (b) (c) (d) (e)
Figure 6. Visualisation of our harmonious attention in person re-id.
From left to right, (a) the original image, (b) the 1st -level of HA,
(c) the 1st -level of SA, (d) the 2nd -level of HA, (e) the 2nd -level of
SA, (f) the 3rd level of HA, (g) the 3rd -level of SA.
Table 8. Comparisons of model size and complexity. FLOPs: the
number of FLoating-point OPerations; PN: Parameter Number.
Model
AlexNet [15]
VGG16 [26]
ResNet50 [8]
GoogLeNet [30]
JLML
HA-CNN
FLOPs
7.25×108
1.55×1010
3.80×109
1.57×109
1.54×109
1.09×109
PN (million)
58.3
134.2
23.5
6.0
7.2
2.7
Depth
7
16
50
22
39
39
5. Conclusion
In this work, we presented a novel Harmonious Attention
Convolutional Neural Network (HA-CNN) for joint learning of person re-identification attention selection and feature representations in an end-to-end fashion. In contrast to
most existing re-id methods that either ignore the matching
misalignment problem or exploit stringent attention learning algorithms, the proposed model is capable of extracting/exploiting multiple complementary attention and maximising their latent complementary effect for person re-id in
a unified lightweight CNN architecture. This is made possible by the Harmonious Attention module design in combination with a two-branches CNN architecture. Moreover,
we introduce a cross-attention interaction learning mechanism to further optimise joint attention selection and reid feature learning. Extensive evaluations were conducted
on three re-id benchmarks to validate the advantages of the
proposed HA-CNN model over a wide range of state-of-theart methods on both manually labelled and more challenging auto-detected person images. We also provided detailed
model component analysis and discussed HA-CNN’s model
complexity as compared to popular alternatives.
References
[1] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean,
M. Devin, S. Ghemawat, G. Irving, M. Isard, et al. Tensorflow: A system for large-scale machine learning. 6
[2] D. Chen, Z. Yuan, B. Chen, and N. Zheng. Similarity learning with spatial constraints for person re-identification. In
CVPR, 2016. 7
[3] W. Chen, X. Chen, J. Zhang, and K. Huang. A multi-task
deep network for person re-identification. In AAAI, 2017. 2
[4] Y.-C. Chen, X. Zhu, W.-S. Zheng, and J.-H. Lai. Person reidentification by camera correlation aware feature augmentation. IEEE TPAMI, 2017. 2, 7
[5] T. H. Cormen. Introduction to algorithms. MIT press, 2009.
4
[6] T. Evgeniou and M. Pontil. Regularized multi–task learning.
In SIGKDD, 2004. 5
[7] S. Gong, M. Cristani, S. Yan, and C. C. Loy. Person reidentification. Springer, January 2014. 1
[8] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In CVPR, 2016. 2, 3, 7, 8
[9] J. Hu, L. Shen, and G. Sun. Squeeze-and-excitation networks. arXiv, 2017. 3, 4
[10] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift. In
ICML, pages 448–456, 2015. 3
[11] M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial
transformer networks. In NIPS, pages 2017–2025, 2015. 2,
3, 5
[12] D. Kingma and J. Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6
[13] M. Koestinger, M. Hirzer, P. Wohlhart, P. M. Roth, and
H. Bischof. Large scale metric learning from equivalence
constraints. In CVPR, 2012. 2
[14] T. G. Kolda and B. W. Bader. Tensor decompositions and
applications. SIAM review, 51(3):455–500, 2009. 4
[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
NIPS, pages 1097–1105, 2012. 2, 3, 8
[16] X. Lan, H. Wang, S. Gong, and X. Zhu. Deep reinforcement
learning attention selection for person re-identification. In
BMVC, 2017. 2
[17] D. Li, X. Chen, Z. Zhang, and K. Huang. Learning deep
context-aware features over body and latent parts for person
re-identification. In CVPR, 2017. 1, 2, 7
[18] W. Li, R. Zhao, T. Xiao, and X. Wang. Deepreid: Deep filter
pairing neural network for person re-identification. In CVPR,
2014. 1, 2, 6
[19] W. Li, X. Zhu, and S. Gong. Person re-identification by deep
joint learning of multi-loss classification. In IJCAI, 2017. 2,
6, 7, 8
[20] S. Liao, Y. Hu, X. Zhu, and S. Z. Li. Person re-identification
by local maximal occurrence representation and metric
learning. In CVPR, 2015. 2, 7
[21] H. Liu, J. Feng, M. Qi, J. Jiang, and S. Yan. End-to-end
comparative attention networks for person re-identification.
IEEE TIP, 2017. 7
[22] X. Liu, H. Zhao, M. Tian, L. Sheng, J. Shao, S. Yi, J. Yan,
and X. Wang. Hydraplus-net: Attentive deep features for
pedestrian analysis. In ICCV, 2017. 2, 7
[23] S. Paisitkriangkrai, C. Shen, and A. van den Hengel. Learning to rank in person re-identification with metric ensembles.
In CVPR, 2015. 2
[24] X. Qian, Y. Fu, Y.-G. Jiang, T. Xiang, and X. Xue. Multiscale deep learning architectures for person re-identification.
In ICCV, 2017. 2
[25] Y. Shen, W. Lin, J. Yan, M. Xu, J. Wu, and J. Wang. Person
re-identification with correspondence structure learning. In
ICCV, 2015. 1, 2
[26] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. In ICLR, 2015.
2, 3, 8
[27] C. Su, J. Li, S. Zhang, J. Xing, W. Gao, and Q. Tian. Posedriven deep convolutional model for person re-identification.
In ICCV, 2017. 1, 2, 7
[28] Y. Sun, L. Zheng, W. Deng, and S. Wang. Svdnet for pedestrian retrieval. In ICCV, 2017. 7
[29] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi.
Inception-v4, inception-resnet and the impact of residual
connections on learning. In AAAI, 2017. 3
[30] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In CVPR, 2015. 2, 8
[31] R. R. Varior, M. Haloi, and G. Wang. Gated siamese
convolutional neural network architecture for human reidentification. In ECCV, 2016. 7
[32] R. R. Varior, B. Shuai, J. Lu, D. Xu, and G. Wang. A
siamese long short-term memory architecture for human reidentification. In ECCV, 2016. 7
[33] S. Vossel, J. J. Geng, and G. R. Fink. Dorsal and ventral
attention systems: distinct neural circuits but collaborative
roles. The Neuroscientist, 20(2):150–159, 2014. 3
[34] F. Wang, M. Jiang, C. Qian, S. Yang, C. Li, H. Zhang,
X. Wang, and X. Tang. Residual attention network for image
classification. In CVPR, 2017. 3, 4
[35] F. Wang, W. Zuo, L. Lin, D. Zhang, and L. Zhang. Joint
learning of single-image and cross-image representations for
person re-identification. In CVPR, 2016. 2
[36] H. Wang, S. Gong, and T. Xiang. Unsupervised learning
of generative topic saliency for person re-identification. In
BMVC, 2014. 1, 2
[37] H. Wang, S. Gong, and T. Xiang. Highly efficient regression
for scalable person re-identification. In BMVC, 2016. 6, 7
[38] T. Wang, S. Gong, X. Zhu, and S. Wang. Person reidentification by discriminative selection in video ranking.
IEEE TPAMI, 2016. 2
[39] T. Xiao, H. Li, W. Ouyang, and X. Wang. Learning deep feature representations with domain guided dropout for person
re-identification. In CVPR, 2016. 2, 3
[40] F. Xiong, M. Gou, O. Camps, and M. Sznaier. Person reidentification using kernel-based metric learning methods. In
ECCV. 2014. 2
[41] L. Zhang, T. Xiang, and S. Gong. Learning a discriminative
null space for person re-identification. In CVPR, 2016. 2, 7
[42] L. Zhao, X. Li, J. Wang, and Y. Zhuang. Deeply-learned
part-aligned representations for person re-identification. In
ICCV, 2017. 1, 2, 7
[43] R. Zhao, W. Ouyang, and X. Wang. Unsupervised salience
learning for person re-identification. In CVPR, 2013. 1, 2
[44] L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian.
Scalable person re-identification: A benchmark. In ICCV,
2015. 1, 2
[45] W.-S. Zheng, S. Gong, and T. Xiang. Re-identification by
relative distance comparison. IEEE TPAMI, pages 653–668,
March 2013. 2
[46] W.-S. Zheng, X. Li, T. Xiang, S. Liao, J. Lai, and S. Gong.
Partial person re-identification. In ICCV, pages 4678–4686,
2015. 1, 2
[47] Z. Zheng, L. Zheng, and Y. Yang. Unlabeled samples generated by gan improve the person re-identification baseline in
vitro. In ICCV, 2017. 2, 6, 7
[48] Z. Zhong, L. Zheng, D. Cao, and S. Li. Re-ranking person
re-identification with k-reciprocal encoding. In CVPR, 2017.
7
| 1cs.CV
|
1
Opinion Dynamics in Social Networks with Hostile
Camps: Consensus vs. Polarization
arXiv:1508.05034v1 [cs.SY] 20 Aug 2015
Anton V. Proskurnikov, Alexey Matveev and Ming Cao
Abstract—Most of the distributed protocols for multi-agent
consensus assume that the agents are mutually cooperative and
“trustful”, and so the couplings among the agents bring the values
of their states closer. Opinion dynamics in social groups, however,
require beyond these conventional models due to ubiquitous
competition and distrust between some pairs of agents, which
are usually characterized by repulsive couplings and may lead
to clustering of the opinions. A simple yet insightful model of
opinion dynamics with both attractive and repulsive couplings
was proposed recently by C. Altafini, who examined first-order
consensus algorithms over static signed graphs. This protocol
establishes modulus consensus, where the opinions become the
same in modulus but may differ in signs. In this paper, we extend
the modulus consensus model to the case where the network
topology is an arbitrary time-varying signed graph and prove
reaching modulus consensus under mild sufficient conditions of
uniform connectivity of the graph. For cut-balanced graphs,
not only sufficient, but also necessary conditions for modulus
consensus are given.
Index Terms—Opinion dynamics, consensus, clustering, agents
I. I NTRODUCTION
For multi-agent networks, the striking phenomenon of
global consensus caused by only local interactions has attracted long-standing interest from the research community.
The interest is motivated by numerous natural phenomena and
engineering designs related to reaching synchrony or agreement among the agents. Examples include, but not limited
to, intelligence of large biological populations and multi-robot
teams. We refer the reader to [1]–[3] for excellent surveys of
recent research on consensus protocols and their applications,
as well as historical milestones.
Starting from the DeGroot algorithm of “iterative pooling” [4] for distributed decision making, many consensus
algorithms were based on the principle of contraction: every
The work was supported in part by the European Research Council
(ERCStG- 307207), St. Petersburg State University, grant 6.38.230.2015 and
RFBR, grants 13-08-01014 and 14-08-01015. Lemma 13 was obtained by
A.S. Matveev under sole support of Russian Science Fund (RSF) grant 1421-00041 at St. Petersburg State University. Theorems 5 and 6 were obtained
by A.V. Proskurnikov under sole support of RSF grant 14-29-00142 at IPME
RAS. The results were partly reported on IEEE Multiconference on Systems
and Control MSC 2014 (Antibes, France) and IEEE Conference on Decision
and Control CDC 2014 (Los Angeles, CA, USA).
A.V. Proskurnikov is with ENTEG, Faculty of Mathematics and Natural
Sciences, University of Groningen, Groningen, the Netherlands and also with
St. Petersburg State University, ITMO University and Institute for Problems
of Mechanical Engineering of Russian Academy of Sciences (IPME RAS),
St.Petersburg, Russia; [email protected]
A. Matveev is with St. Petersburg State University, St. Petersburg, Russia; [email protected]
M. Cao is with ENTEG, Faculty of Mathematics and Natural Sciences,
University of Groningen, Groningen, The Netherlands; [email protected]
agent’s state constantly evolves to the relative interior of
the convex hull spanned by its own and neighbors’ states.
Hence, the convex hull spanned by the states of the agents,
driven by such a protocol, is shrinking over time. Based on
the Lyapunov-like properties of this convex hull [5], [6] or
relevant results on convergence of infinite products of stochastic matrices [1], [2], [7], stability properties of contracting
iterations were examined intensively with special attention on
the effect of time-varying interaction topologies. Necessary
and sufficient conditions for consensus under bidirectional [5],
[8], [9] and cut-balanced graphs [10] boil down to repeated
joint connectivity of the network. For general directed graphs
the sufficient condition of uniform quasi-strong connectivity
(UQSC) [6], is considered to be “the weakest assumption
on the graph connectivity such that consensus is guaranteed
for arbitrary initial conditions” [11]. This common belief has
recently been confirmed by results in [12], [13] stating that
the UQSC is necessary and sufficient for robust consensus
and consensus with exponential convergence. Many high-order
consensus algorithms either extend their first-order counterparts [1], [3] or are squarely based on them [14].
Unlike teams of agents that achieve a common goal due to
cooperation, networks where agents can both cooperate and
compete (sometimes referred to as coopetitive networks [15],
[16]) still demand more thorough mathematically rigorous
analysis. In social networks, competition, antagonism and
distrust between social actors and their groups are ubiquitous
[17], [18], which are usually modeled by repulsive couplings
or negative ties [19] among the agents. A specific example of
such couplings observed in dyadic interactions, is reactance
[20] which leads to boomerang effects, first described in [21]:
in the process of persuasion, opinions (even close to each
other initially) can become opposite. Analogous phenomenon,
referred to as the group polarization [22], has long been
studied in social psychology: the community divides into two
groups, each reaching consensus; the consensus opinions are
not only opposite, but often further away from each other
than the two initial average opinions of the corresponding
groups. Analysis of real-world social networks (e.g. users of
social web-sites [17]) shows the strong correlation between
polarization and structural balance [17], [18] of positive and
negative ties. The latter property implies that community splits
into two hostile camps (e.g. votaries of two political parties),
where the relations inside each faction are cooperative.
It is known that agents’ repulsion can lead to the clustering
behavior in a complex network [23]. The possibility of clustering in social groups due to negative ties was demonstrated in
[19] (see also references therein); these effects are still waiting
2
for mathematically rigorous analysis. Most of the existing
works on opinion dynamics focus on the persistent disagreement and clustering of opinions caused by bounded confidence
[24], [25] or, more generally, biased assimilation [26]: agents
readily adopt opinions of like-minded neighbors, accepting the
“deviating” opinions with discretion. In [27], [28] Altafini
proposed a simple yet instructive mathematical model of
opinion polarization over structurally balanced graphs, extending conventional first-order consensus algorithms to the case
with antagonistic interactions. These protocols were examined
under the assumption that the interaction graph is static and
strongly connected and shown to establish modulus consensus
[29], where the opinions agree in modulus but may differ in
signs. If the graph is structurally balanced, the modulus of
the final opinion is generally non-zero and opinions either
reach consensus or polarize (“bipartite consensus” [28] is
established); otherwise, opinions converge to zero.
Mathematical examination of polarization behavior due
to antagonism is among the first important steps towards
understanding the dynamics of networks consisting of both
cooperative and competitive agents. Such networks are not
confined to social systems; repulsive interactions play an
important role in e.g. motion control of swarms and other
multi-agent formations, where agents may avoid collision [30],
[31] and distribute evenly on circular or other closed curves
[32]–[34]. So a number of papers have been published recently
studying this class of networks [15], [35]–[39].
The aforementioned papers, however, mainly focus on the
case where the interaction topology is static. In the present
paper, we consider Altafini’s model on opinion dynamics over
general directed time-varying graphs. Removing the restrictions of static topologies not only allows one to analyze
dynamics of real social networks, where the agents may
change their relationships from friendship to hostility and
vice versa, but also enables one to extend the result to nonlinear protocols. In fact we will examine nonlinear algorithms
in the common framework as linear ones, getting rid of the
restrictions such as monotonicity [27], [28], [40].
Our main result states that modulus consensus is established if the topology is uniformly strongly connected. Unlike cooperative networks, the uniform strong connectivity
cannot be relaxed to the uniform quasi-strong connectivity,
which is a commonly adopted condition for consensus over
directed time-varying graph [11], [41]. At the same time, the
condition of uniform strong connectivity is in general not
necessary, and filling the gap between necessary and sufficient
conditions remains a tough problem even in the cooperative
case. However, we fill this gap in the special case of cutbalanced graphs, extending necessary and sufficient consensus
criterion from [10] to modulus consensus over signed graphs.
It should be noticed that results from [10] are not directly
applicable to signed graphs; in the special case of bidirectional
or “reciprocal interactions” they were extended to the signed
case by the lifting technique [37]. We will make further
remarks on this in the corresponding sections. The results were
partly reported in our conference papers [42], [43].
The paper is organized as follows. Section II introduces
some preliminary concepts and notations. Section III gives
the setup of the problem in question. Section IV presents the
main results. Section VI offers the proof of the main results.
II. P RELIMINARIES
Throughout the paper m : n, where m, n are integers and
m ≤ n, stands for the sequence {m, m + 1, . . . , n}. The sign
of a number x ∈ R is denoted by sgn x ∈ {−1, 0, 1}. The
abbreviation “a.a.” stands for “almost all” (except for the set
of zero Lebesgue measure). Given a matrix L = (ljk ), let
∆
∆
abs L P
= (|ljk |). We also introduce the matrix norm |L|∞ =
maxj k |Ljk |. As usual, for a column vector x ∈ RN , one
has |x|∞ = maxj |xj | and it is easily shown that |L|∞ =
∞
sup |Lx|
|x|∞ , where the supremum is over all column vectors
∆
x 6= 0 of appropriate dimensions. Let 1̄N = (1, 1, . . . , 1)T ∈
RN . Given x ∈ R, let x+ = max(x, 0) and x− = (−x)+ ,
hence x = x+ − x− and |x| = x+ + x− .
A. Signed graphs and their properties
A (weighted directed) signed graph is a triple G =
(V, E, A), where V = {v1 , . . . , vN } stands for the set of
nodes, E ⊂ V × V is a set of arcs and A = (ajk ) ∈ RN ×N
is a signed adjacency matrix, i.e. ajk 6= 0 if and only if
(vk , vj ) ∈ E. Throughout the paper, we confine ourselves
to graphs that have no self-loops (ajj = 0 ∀j) and are
digon sign-symmetric [28], i.e. any pair of opposite arcs (if
exists) is identically signed: ajk akj ≥ 0 ∀j, k. Identifying
the set of nodes V with 1 : N , there is a one-to-one
correspondence between signed graphs and their adjacency
matrices A ∈ RN ×N 7→ G[A] = (1 : N, E[A], A), where
E[A] = {(j, k) : akj 6= 0}.
Given ε > 0, let Aε = (aεjk ) stand for the “truncated”
adjacency matrix: aεij = aij when |aij | ≥ ε and aεij = 0
∆
otherwise. The corresponding graph Gε = G[Aε ] is obtained
from G = G[A] by removing arcs of absolute weight less than
ε and we call it ε-skeleton of the graph G.
A path connecting nodes v and v 0 is a sequence of
nodes vi0 := v, vi1 , . . . , vin−1 , vin := v 0 (n ≥ 1) such
that (vik−1 , vik ) ∈ E for k ∈ 1 : n. A path where
vi0 = vin is referred to as a cycle. The cycle is positive if
ai0 i1 ai1 i2 . . . ain−1 in > 0 and negative otherwise. The digonsymmetric strongly connected graph is structurally balanced
if and only if all its oriented cycles are positive [17], [28]. A
node is called root if it can be connected with a route to any
other node of the graph. A graph is strongly connected (SC)
if a path between any two different nodes exists. The graph
is quasi-strongly connected (QSC) if it has at least one root.
Any SC graph is also QSC, each node being a root. A graph
whose ε-skeleton is SC (respectively, QSC) is called strongly
ε-connected (respectively, quasi-strongly ε-connected).
Given a graph G = (V, E, A), its subgraph is a graph G0 =
(V 0 , E 0 , A0 ), where V 0 ⊆ V , E 0 ⊆ (V 0 × V 0 ) ∩ E and A0 =
(aij )i,j∈V 0 stands for the corresponding submatrix of A. We
call a subgraph in-isolated if no arc comes from V \ V 0 to V 0 ,
i.e. aji = 0 ∀i ∈ V 0 , j 6∈ V 0 .
We call two disjoint non-empty sets V1 , V2 ⊆ V hostile
camps in the graph G if ajk ≥ 0 when j, k ∈ V1 or j, k ∈ V2
3
and ajk ≤ 0 whenever j ∈ V1 , k ∈ V2 or j ∈ V2 , k ∈ V1 . The
graph is structurally balanced (SB) [17], [28] if the set of its
nodes can be divided into two hostile camps V = V1 ∪ V2 .
The digon-symmetric SC graph is structurally balanced if and
only if any cycle in it is positive [17], [28].
Following [28], we define the Laplacian matrix L = L[A]
of the signed graph G[A] as follows
(
−ajk , j 6= k
∆
(1)
L[A] = (Ljk )N
Ljk := PN
j,k=1 ,
m=1 |ajm |, j = k.
Equation (1) is a straightforward extension of the conventional definition of the Laplacian matrix of a weighted
graph [44] to the case of signed weights. As implied by the
Gershgorin disk theorem [28], L[A] has no eigenvalues in the
closed left half-plane C̄− = {λ ∈ C : Re λ ≤ 0} except for
possibly λ = 0. Unlike the unsigned case, in general L[A] may
have no zero eigenvalue and hence −L[A] may be a Hurwitz
matrix. For a SC graph G = G[A] this is the case if and only
if G is not structurally balanced [28, Lemma 2].
B. Some important types of time-varying signed graphs
Throughout the paper, the term time-varying (signed) graph
means the graph G[A(t)], where a time-dependent matrix
A(t) ∈ RN ×N is Lebesgue measurable and locally bounded.
Given such a graph G(t) = G[A(t)],
R ∞we say a node j is
essentially connected to a node k if t0 |ajk (s)|ds = ∞ for
some t0 ≥ 0 (the latter inequality then holds for any t0 ≥ 0
since ajk is locally bounded). Let E = E[A(·)] stand for the
set of all such pairs (j, k). Following [9], we call the graph
G[A(·)] = (1 : N, E) the graph of essential interactions and
say that the graph G(·) is essentially strongly connected (ESC)
if G[A(·)] is strongly connected. Likewise, G(·) is essentially
quasi-strongly connected (EQSC) if G[A(·)] is QSC.
The graph G[A(·)] is said to be uniformly strongly connected (USC) if there
R t+Texist constants T > 0 and ε > 0 such
that the graph G[ t
abs A(s) ds] is strongly ε-connected
for any t ≥ 0. By replacing the word “strongly” in the
latter definition with “quasi-strongly”, one defines uniformly
quasi-strongly connected (UQSC) time-varying graph. It may
be easily shown that the USC (respectively UQSC) graph is
always ESC (respectively, EQSC), while the inverse is not
valid.
The graph G[A(·)] is cut-balanced [10] if a constant K ≥ 1
exists such that for any partition of the nodes V 0 ∪V 00 = 1 : N ,
V 0 ∩ V 00 = ∅, the following inequalities hold
X X
X X
X X
K −1
|akj | ≤
|ajk | ≤ K
|akj |.
j∈V 0 k∈V 00
j∈V 0 k∈V 00
j∈V 0 k∈V 00
(2)
A typical example of a cut-balanced graph is the typesymmetric graph [10], which means the existence of K ≥ 1
such that
K −1 |akj (t)| ≤ |ajk (t)| ≤ K|akj (t)| ∀t ≥ 0∀j 6= k.
(3)
Other examples include weight-balanced graphs, see [10] for
details. As implied by [10, Lemma 1], for cut-balanced graphs,
the EQSC property implies ESC; precisely, any quasi-strongly
∆
connected component of the digraph G = G[A(·)] is strongly
connected, and a path between j and k exists if and only if
the path from k to j exists.
III. P ROBLEM S ETUP
Consider a group of N ≥ 2 agents indexed 1 through N ,
the opinion of the ith agent is denoted by xi ∈ R and we
define x := (x1 , . . . , xN )T ∈ RN . The agents update their
opinions in accordance with a distributed protocol as follows:
ẋ(t) = −L[A(t)]x(t), t ≥ 0,
(4)
which can be written componentwise as
ẋj (t) =
N
X
|ajk (t)|(xk (t) sgn ajk (t) − xj (t)) ∀j.
(5)
k=1
Here A(t) = (ajk (t)) is a locally bounded matrix-valued function which describes the interaction topology of the network
and ajj (t) ≡ 0. At time t ≥ 0, the opinion of the jth agent
is influenced by agents for which ajk 6= 0 (“neighbors”).
Unlike conventional consensus protocols [44] this influence
may be either cooperative (when ajk > 0) or competitive
(when ajk < 0). The coupling term |ajk |(xk sgn ajk − xj )
in (5) drives the opinion of the jth agent, respectively, either
towards the opinion of the kth one or against it.
In [28] protocol (4) has been carefully examined, assuming
the interaction graph is constant (A(t) ≡ A) and strongly
connected. It was shown that the steady-state opinions always
agree in modulus, but generally differ in signs; in other words,
the modulus consensus of opinions [29] is established.
Definition 1: The protocol (4) establishes modulus consensus, if for any x(0) a number x∗ ≥ 0 exists such that
lim |xi (t)| = x∗ .
t→+∞
(6)
The following lemma shows that there are two essentially
different types of modulus consensus: “trivial” with x∗ = 0
for all x(0) (the system (4) is asymptotically stable) and “nontrivial”, where x∗ 6= 0 for a.a. x(0).
Lemma 1: Suppose that protocol (4) establishes modulus consensus. Then there exist vectors v, ρ ∈ RN with
ρ1 , . . . , ρN = ±1 such that for any solution of (4) one has
lim x(t) = ρv T x(0) ⇔ lim xj (t) = ρj v T x(0).
t→+∞
t→+∞
(7)
Lemma 1 shows that in the “non-trivial” case v 6= 0, opinions
either reach consensus (ρ1 = . . . = ρN ) or polarize (ρi have
different signs) whenever v T x(0) 6= 0. For both situations we
say that the protocol establishes bipartite consensus.
Definition 2: We call the protocol (4) stabilizing, if
lim xj (t) = 0 ∀j∀x(0). The protocol establishes bipartite
t→∞
consensus if (7) holds with some v 6= 0; it establishes
consensus if additionally ρ = 1̄N or ρ = −1̄N .
It was proved in [28] if A(t) ≡ A, then the protocol is
stabilizing (that is, −L[A] is a Hurwitz matrix) unless the
graph G[A] is structurally balanced (SB). The latter property
implies that a community is divided into two hostile camps
(such as votaries of two political parties), where each agent
cooperates with its camp-mates, competing with agents from
4
the opposite camp. A special case of SB graphs is the graph
with non-negative weights ajk ≥ 0 where one of the camps is
empty. In this case strong connectivity (SC) and even quasistrong connectivity (QSC) imply consensus [1]. The case of
general SB and SC graph is reducible to this case by means
of the gauge transformation [28], which allows to prove (6),
where x∗ depends on the initial conditions. If both hostile
camps are not empty, the opinions polarize. In other words,
structural balance implies bipartite consensus. A generalization
of the gauge transformation from [28] is the lifting approach
from [37], splitting each agent into a pair of virtual agents with
opposite opinions, after which the original dynamics can be
considered as a projection of some larger network with purely
cooperative interactions. This approach can be applied also to
some time-varying networks.
In Section IV-A we further refine Altafini’s results on
modulus consensus over static graphs by discarding the strong
connectivity assumption. We show that for structurally balanced graphs modulus consensus is established if and only if
the graph is QSC; in fact, in this case bipartite consensus is
established. Conversely, bipartite consensus is reached only
when the graph is structurally balanced and QSC. If the
graph has no structurally balanced in-isolated subgraphs, the
protocol is stabilizing. Thus we offer necessary and sufficient
conditions for modulus consensus for general static graph.
The main concern of this paper is modulus consensus
over time-varying signed graphs. In [28] this problem was
considered only for the very special case where the graph
is constantly strongly connected, has time-invariant signs of
the arcs and also weight-balanced (this assumption was not
explicitly mentioned, but in fact was used in the proof which
appeals to [44, Theorem 9]). Below we relax these restrictions.
Dealing with real-world social networks, the time-invariance
of such relationships between individuals as friendship and
hostility is evidently a non-realistic assumption. What is more
important, the opinion dynamics in social networks are usually
considered to be nonlinear [19], [24]. Such models are often
reducible to the linear case by introducing time-varying gains,
depending on the solution; however, the corresponding graphs
can hardly be weight-balanced. Our techniques allow us to
examine both linear and nonlinear consensus protocols from
[27], [28] in the common framework. Although it is a hard
problem to find explicitly the ultimate opinion vector in the
case of time-varying topologies, Lemma 1 shows that there
are similarities with the static case.
A common techniques used to prove consensus in the case
of cooperative agents is the shrinking property of the convex
hull, spanned by agent’s opinions. Under the UQSC property
of the graph, the diameter of this convex hull may serve
as a Lyapunov function [5], [6], [11]. The UQSC condition
is not necessary in general [5], considered as “the weakest
assumption on the graph connectivity such that consensus is
guaranteed for arbitrary initial conditions” [11], and becomes
necessary under additional restriction of uniform convergence
[6]. On the other hand, the EQSC condition is always necessary for consensus yet insufficient in the case of directed
topologies [5]. This gap between necessary and sufficient
conditions has been filled recently for type-symmetric and
other cut-balanced graphs [9], [10] where EQSC is not only
necessary but also sufficient for consensus.
Under antagonistic interactions between the agents, the
convex hull spanned by opinions is not shrinking, and the only
available Lyapunov function is the maximal modulus, which
will be shown to be non-increasing and thus converging to a
limit. However, the UQSC property in general does not guarantee that the minimal modulus also converges to the same limit
(as will be shown by a counter example). To provide this, one
requires stronger USC conditions. Whereas EQSC property
is necessary for bipartite consensus, it is not necessary for
stability, as illustrated by the following trivial example. Let
A = diag(A1 , A2 ), where both graphs G[A1 ] and G[A2 ] are
strongly connected and structurally unbalanced. As follows
from [28], the matrices (−L[A1 ]) and (−L[A2 ]) are Hurwitz,
which also holds for (−L[A]) = − diag(L1 [A], L2 [A]) and
thus the protocol is stabilizing. The static graph G[A] is
not QSC and thus not EQSC. Filling this gap between the
necessary and sufficient conditions for modulus consensus that
is even “wider” than in the cooperative case, is a tough open
problem. However, adopting the techniques from [9], we fill
this gap for cut-balanced graphs by offering necessary and
sufficient conditions of modulus consensus (Subsection IV-C).
IV. M AIN R ESULTS
This section is organized as follows. We start with modulus
consensus criteria for static graphs which extend results from
[28] by discarding the assumption of strong connectivity in
them (Subsection IV-A). We show that the necessary and
sufficient condition for bipartite consensus is structural balance
and QSC, and give also necessary and sufficient conditions
for stability. The next Subsection IV-B deals with the case
of general time-varying graphs. We show sufficiency of the
USC condition for modulus consensus and demonstrate that,
unlike the cooperative case, this condition cannot be relaxed
to UQSC. In the last Subsection IV-C we focus on modulus
consensus over cut-balanced graphs. In this case it is possible
to give necessary and sufficient conditions for both types of
modulus consensus, whereas filling the gap between necessary
and sufficient conditions for modulus consensus under general
directed graphs remains a tough open question.
A. Time-invariant protocols
Throughout this section A(t) ≡ A is a constant matrix. We
start with the case of structurally balanced graph. In this case,
a gauge transformation [28] exists which reduces the protocol
to a cooperative one, whose properties are well established.
Lemma 2: Let G[A] be structurally balanced. Then L[A]
has eigenvalue at 0 and the following claims are equivalent:
1) the graph G[A] is QSC;
2) the linear subspace ker L[A] ⊂ Rn has dimension 1;
3) the protocol (4) establishes modulus consensus.
If these claims hold, then ρ, v from Lemma 1 are respectively the right and the left eigenvectors of L[A] at 0, hence
v T L[A] = L[A]ρ = 0 and v 6= 0, so protocol establishes
bipartite consensus. If A is a non-negative matrix (one of the
5
Fig. 1.
Static QSC graph: modulus consensus is impossible if a31 a32 > 0
hostile camps is empty), the protocol establishes consensus,
and otherwise, opinions polarize.
In the case where G[A] is structurally balanced yet not QSC
(so modulus consensus is not reached), the structure of steadystate opinions may be described in terms of the maximum
out-forest matrix as done in [45] for cooperative agents.
As follows from Lemma 2, a protocol with structurally
balanced graph cannot be stabilizing. Therefore, stability is
also impossible when the graph contains an in-isolated structurally balanced (ISB) subgraph, in other words, the group
involves hostile camps whose members ignore the opinions of
the remaining agents. The nodes of such a subgraph, if existed,
would reach bipartite consensus of opinions independently of
the remaining agents. The following theorem gives a criterion for modulus consensus over static graphs, showing that
existence of ISB subgraphs is the only obstacle for stability.
Theorem 1: Let A(t) = const. The protocol (4) is stabilizing if and only if the graph is neither SB itself nor contains an
ISB subgraph. Bipartite consensus is established if and only if
G[A] is structurally balanced and QSC: if ajk ≥ 0 ∀j, k, then
consensus is established; otherwise, opinions polarize.
Example 1. Consider a team of N = 3 agents with states
x1 (t), x2 (t), x3 (t). Assume that a12 = a21 = −1 (see Fig. 1)
and a31 a32 > 0. Thus the equations are
ẋ1 = (−x2 − x1 ), ẋ2 = (−x1 − x2 ),
ẋ3 = a31 x1 + a32 x2 − (|a31 | + |a32 |)x3 ,
and the graph is not structurally balanced (agents 1 and 2
are constantly antagonistic, so the structural balance requires
agent 3 to cooperate with only one of them, competing with
the other, whereas in reality it cooperates with both agents
1 and 2). According to Theorem 1, modulus consensus is
impossible (since the ISB subgraph with the set of nodes
{1, 2} exists). This can also be shown in a straightforward
∆
way: the system has equilibria (ξ, −ξ, ρξ), with ρ = (a31 −
a32 )/(|a31 | + |a32 |) ∈ (−1; 1), ξ ∈ R.
This simple example illustrates that, unlike the cooperative
case (ajk ≥ 0), the protocol with static QSC graphs in general
does not establish modulus consensus. To guarantee modulus
consensus, one typically requires strong connectivity (assumed
in [28]) or some other property, excluding ISB subgraphs.
B. Protocols over dynamic signed graphs
We start with the following useful lemma, which does not
rely on any connectivity assumptions and shows, in particular,
that solutions to (4) are always bounded.
Lemma 3: For any solution of system (4), the function
|x(t)|∞ = maxi |xi (t)| is monotonically non-increasing:
|x(t)|∞ ≤ |x(t0 )|∞ whenever t ≥ t0 ≥ 0. Equivalently, the
Cauchy evolutionary matrix Φ(t; t0 ) of system (4) satisfies the
inequality |Φ(t; t0 )|∞ ≤ 1 for t ≥ t0 .
Lemma 3 implies, in particular, the existence of the limit
lim |x(t)|∞ . The following theorem shows that under the
t→+∞
uniform strong connectivity property and bounded coupling
gains the modules of all opinions converge to the same limit.
Theorem 2: If ajk (t) are bounded and the graph G[A(·)] is
USC, then the protocol (4) establishes modulus consensus.
As a corollary, we immediately obtain the well-known
criterion for consensus under cooperative protocols.
Corollary 1: If ajk (t) ≥ 0 ∀j, k for a.a. t ≥ 0 and G[A(·)]
is USC, then the protocol (4) establishes consensus.
It is well known that assumptions of Corollary 1 can be
relaxed: in the case where ajk ≥ 0 the UQSC condition is
sufficient for consensus [6], [11], [46]. Moreover, using the
gauge transformation approach from [28], sufficiency of the
UQSC property may be proved for a dynamic structurally
balanced graph, provided that the subdivision into two “hostile
camps” remains unchanged.
Lemma 4: Suppose that V = 1 : N = V1 ∪ V2 , where
ajk (t) ≥ 0 for any t ≥ 0 if j, k ∈ V1 or j, k ∈ V2 ; otherwise,
ajk (t) ≤ 0 for any t ≥ 0. If the graph G[A(·)] is UQSC, the
protocol establishes bipartite consensus (if V1 = ∅ or V2 = ∅)
or bipartite consensus (when V1 , V2 6= ∅).
Remark 1: Lemma 4 obviously remains valid if two hostile
camps V1 , V2 exist only for t ≥ t0 , where t0 ≥ 0. This
observation makes the result of Lemma 4 applicable to topologies that evolve in order to achieve the structural balance in
finite time (after which the signs of arcs remain unchanged);
graph dynamics leading to structural balance in finite time
were proposed in [47], [48].
However, in general the USC condition in Theorem 2 is not
relaxable to UQSC. Example 1 in Subsection IV-A shows that
even for static graphs the QSC property (equivalent to UQSC)
does not guarantee modulus consensus unless the graph is
structurally balanced. Our next example shows that the UQSC
property is not sufficient neither when the graph remains structurally balanced but the relations of friendship and hostility
between the agents evolve over time. We construct a protocol
(4) with periodic piecewise-constant matrix A(t), such that
the graph G[A(t)] is structurally balanced for any t ≥ 0 and
UQSC, and nevertheless modulus consensus is not established.
Example 2. Consider the more general system
ẋ1 (t) = (−x2 (t) − x1 (t)), ẋ2 (t) = (−x1 (t) − x2 (t)),
ẋ3 (t) = a31 (t)(x1 (t) − x3 (t)) + a32 (t)(x2 (t) − x3 (t)).
(8)
The functions a31 , a32 are constructed as follows. Consider
first system (8) with a31 (t) ≡ 1, a32 (t) ≡ 0 and the solution
to (8) launched at the initial state x1 (0) = 1, x2 (0) =
−1, x3 (0) = −1/2. It is evident that x1 (t) = 1 = −x2 (t) for
any t ≥ 0 and x3 (t) ↑ 1 as t → +∞. Therefore, there exists
the first time instant T0 > 0 such that x3 (T0 ) = 1/2. Notice
that in the symmetric situation where a31 (t) ≡ 0, a32 (t) ≡ 1
and x(t) is a solution to (8) starting at x1 (0) = 1, x2 (0) =
−1, x3 (0) = 1/2, one has x3 (t) ↓ −1 and T0 is the first instant
6
where x3 (T0 ) = −1/2. Taking
(
1, t ∈ [0; T0 ) ∪ [2T0 ; 3T0 ) ∪ . . .
a31 (t) = 1−a32 (t) =
0, t ∈ [T0 ; 2T0 ) ∪ [3T0 ; 4T0 ) ∪ . . .
∆
,
one finally gets a 2T0 -periodic matrix A(t), corresponding
to the UQSC graph G[A(·)]. Moreover, this graph is also
quasi-strongly connected and structurally balanced at any time.
Even so the solution to (8) starting at x1 (0) = 1, x2 (0) =
−1, x3 (0) = −1/2 does not achieve modulus consensus. It
can be easily shown that x1 (t) = −x2 (t) = 1 for any
t ≥ 0. Since a31 (t) = 1 and a32 (t) = 0 when t < T0 ,
one has x3 (T0 ) = 1/2 by definition of T0 . On the next
interval t ∈ [T0 ; 2T0 ) one has a31 (t) = 0 and a32 (t) = 1
and hence x3 (2T0 ) = −1/2, so the solution x(t) is periodic
and x3 (t) ∈ [−1/2; 1/2] whereas |x1 (t)| = |x2 (t)| = 1.
Dealing with purely cooperative protocols, UQSC is considered to be not only sufficient for consensus but also “nearly”
necessary. It may also be relaxed to EQSC for some types of
graphs (e.g. cut-balanced ones). On the other hand, the EQSC
condition is always necessary for consensus among cooperating agents, being in general not sufficient. As discussed in
Section IV-A, stability of (4) is possible without the EQSC
property even for static graphs. However, EQSC is required
for bipartite modulus consensus.
Lemma 5: If the protocol (4) establishes bipartite consensus, the graph G[A(·)] is EQSC.
Remark 2: In the general situation, where the graph is USC
but the assumptions of Lemma 4 do not hold, it is difficult
to distinguish between stability and bipartite consensus. A
sufficient condition for stability was proved in [37]: the
protocol
R t+T is stabilizing, if for some ε > 0, T > 0 all the graphs
G[ t
abs A(s) ds] (where t ≥ 0) have strongly connected
and structurally unbalanced ε-skeletons [37]. This is the only
case where the “lifted” network is proved to inherit the USC
property [37]. Generally, this is not the case, so the approach
[37] does not allow to derive Theorem 2.
C. Modulus consensus over cut-balanced graphs
In the previous subsection, we get a sufficient condition
for modulus consensus (the USC condition). In general, this
property is not necessary for modulus consensus. Moreover,
even the weaker EQSC condition is necessary for bipartite
consensus but not for stability. Filling this gap between necessary and sufficient conditions is a hard open problem even
for cooperative protocols. However, this gap has been recently
filled for cooperative protocols with type-symmetric [9] and
more general cut-balanced graphs [10]. The proofs from [10]
are not applicable for signed graphs. However, adopting the
approach from [9], we extend the result from [10] to the case
of modulus consensus over signed graph, giving necessary and
sufficient conditions for each type of modulus consensus.
Throughout this section the graph G[A(·)] is cut-balanced,
i.e. the inequalities (2) hold for some K ≥ 1.RRecall the jth
∞
agent essentially interacts
the kth one if R 0 ajk (t)dt =
R ∞ with
∞ −
+
∞ and hence either 0 ajk (t)dt = ∞ or 0 ajk (t)dt =
∞. We say agents essentially cooperate in the first case and
essentially compete in the second situation (in general, both
relations may hold). Let E = E[A(·)] and E + ⊆ E, E − ⊆ E
be the sets of those pairs of agents (j, k) that respectively
essentially cooperate and essentially compete.
If E + ∩ E − = ∅, we assign the weights +1 and −1 to the
∆
arcs from E + and E − respectively, transforming G = G[A(·)]
±
to a signed graph G = (1 : N, E, (sjk )), sjk = +1 for
(j, k) ∈ E + and sjk = −1 for (j, k) ∈ E − .
As implied by [10, Lemma 1], for cut-balanced graphs
the EQSC property implies ESC, precisely, any quasi-strongly
∆
connected component of the digraph G = G[A(·)] is strongly
connected, and a path between j and k exists if and only if a
path from k to j exists. From Lemma 5, the ESC condition is
necessary for bipartite consensus. In the case of cooperative
agents, this property is also sufficient for consensus [10].
However, in the case of signed graph, ESC is not sufficient
without the “essential” structural balance.
Theorem 3: Assume the graph G[A(·)] is cut-balanced. The
protocol (4) establishes bipartite modulus consensus if and
only if G ± is well-defined (E + ∩ E − = ∅), strongly connected
and structurally balanced; opinions polarize if and only if
E − 6= ∅, and otherwise consensus is established. If G is
strongly connected but E + ∩ E − 6= ∅ or G ± is structurally
unbalanced, the protocol (4) is stabilizing.
In the case of purely cooperative protocol (ajk (t) ≥ 0),
Theorem 3 transforms into the result obtained in [9], [10].
Corollary 2: Cooperative protocol establishes consensus if
and only if the graph G[A(·)] is essentially connected.
Our next result addresses the case where G is not necessarily
connected and thus may be decomposed into several disjoint
strongly connected components G = G1 ∪ G2 ∪ . . . ∪ Gd , Gr =
(Vr , Er ), d ≥ 1. In this case in any component Gr , modulus
consensus is established, and the type of which depends only
on the structure of Gr . Let Er+ := Er ∩ E + and Er− := Er ∩ E − .
If Er+ ∩ Er− = ∅, define a signed graph Gr± by assigning arcs
from Er+ , Er− with weights +1 and −1 respectively.
Theorem 4: For any solution of (5) there exists limits
x†i = lim xi (t) and |x†i | = |x†j | whenever i and j are
t→∞
in the same strongly connected component: i, j ∈ Vr . If
Er+ ∩ Er− = ∅ and Gr± is structurally balanced, the bipartite
consensus is achieved, which comes to consensus if Er− = ∅,
and otherwise the opinions polarize (Vr = Vr1 ∩ Vr2 and
x†i = −x†j ∀i ∈ Vr1 , j ∈ Vr2 ). If Er+ ∩ Er− 6= ∅ or the graph Gr±
is structurally unbalanced, x†i = 0 ∀i ∈ Vr that is, dynamics
of opinions from Vr are stable.
The following criterion of stability is immediate.
Corollary 3: The protocol (4) is stable if and only if for any
strongly connected component Gr , one has either Er+ ∩Er− 6= ∅
or Gr± being structurally unbalanced.
Remark 3: Theorems 3 and 4 were proved in [37] in the
special case of “type-symmetric” graphs, such that (3) holds
for some K ≥ 1. The main idea of the proof is to show that the
latter property remains valid for the “lifted” network, which is
purely cooperative and hence can be examined by techniques
from [10]. We extend the result from [37] to cut-balanced
graphs. Although this extension seems to be provable by
techniques from [37], our proof based on [9] is of independent
interest; we elaborate mathematical techniques to cope with
7
both general and cut-balanced cases in similar ways.
B. Nonlinear Laplacian Flow
V. A PPLICATIONS : N ONLINEAR P ROTOCOLS
In this section we apply our results to some types of
nonlinear consensus protocols, similar to those from [27], [28].
In this subsection we examine the nonlinear consensus
protocol similar to that addressed in [28, Section IV-B]:
ẋi (t) =
A. Additive Laplacian protocols
Our first example concerns nonlinear consensus algorithms
that are referred in [28] as the “additive Laplacian feedback
schemes”. The first of them is
N
X
ẋi (t) =
|aij (t)|(hij (xj (t) sgn aij (t)) − hij (xi (t))), (9)
N
X
|Fij (t, x)|(xj (t) sgn Fij (t, x) − xi (t)),
here i ∈ 1 : N and Fij : [0; ∞) × RN → R are Caratheodory
maps, i.e. Fij (t, ·) are continuous for a.a. t and Fij (·, x) are
measurable for any x. We assume also that for any compact
set K ⊂ [0; ∞) × RN one has
sup{Fij (t, x) : x ∈ K, t ≥ 0} < ∞
j=1
(12)
j=1
∀i, j.
(13)
and the second protocol has the form
ẋi (t) =
N
X
|aij (t)|hij (xj (t) sgn aij (t) − xi (t))) ∀i.
(10)
j=1
We adopt the following assumption about the nonlinearities.
Assumption 1: For any i, j ∈ 1 : N the map hij ∈ C 1 (R)
is strictly increasing (and hence h0ij > 0) with hij (0) = 0.
Defining the functions Hij [y, z] as follows: Hij [y, z] :=
(hij (y)−hij (z))/(y −z) for y 6= z and Hij [z, z] := h0ij (z) so
that Hij [y, z] > 0 for any y, z. Since hij ∈ C 1 , Hij is easily
shown to be a continuous function and hij (y) − hij (z) =
Hij [y, z](y − z) ∀y, z. Under Assumption 1, Theorems 2 and
3 appears to be applicable to the protocols (9), (10) after
the standard trick, replacing nonlinearities with the solutiondependent gains, as shown by the following lemma.
Lemma 6: Let x(t) be a solution to system (9), which is
defined for t ≥ 0. Define the matrix A(t) = (aij (t)) by
aij (t) := aij (t)Hij [xj (t) sgn aij (t), xi (t)]. Then
ẋ(t) = −L[A(t)]x(t).
(11)
If the graph G[A(·)] is EQSC, ESC, UQSC, USC, or cutbalanced, then the same is valid for the graph G[A(·)]. If
the matrix A(·) is globally bounded, the same is valid for
the matrix A(·). These claims also hold for the protocol (10),
taking aij (t) := aij (t)Hij [xj (t) sgn aij (t) − xi (t), 0].
Application of Theorems 2, 3 to (11) yields the following.
Theorem 5: Under Assumption 1, the solutions to systems
(9), (10) exist, are unique and infinitely prolongable for any
initial condition. If the graph G[A(·)] is USC and A(·) is
bounded, or the graph G[A(·)] is ESC and cut-balanced, the
protocols (9), (10) establish modulus consensus.
Comparing the result of Theorem 5 with that of [28,
Theorem 3,4], one notices that our assumption about the nonlinearities hij differs from [28], where they are not assumed to
be smooth, but only monotonic with some integral constraint.
However, unlike [28], functions hij may be heterogeneous and
not necessarily odd; the graph may be time-varying.
Note that Theorem 5 gives only sufficient conditions for
modulus consensus. Necessary conditions such as Lemma 5
are not directly applicable since they assume the matrix A(t)
to be common for all solutions. Extending the concept of
essentially equivalent protocols (see Subsection VI-C) to the
nonlinear case, it is possible to get some necessary conditions
which are, however, beyond the scope of this paper.
Theorem 6: For any initial condition x(0) a solution of (12)
∆
exists for t ≥ 0 and the matrix A(t) = Fij (t, x(t)) is bounded.
If the graph G[A(·)] is USC or ESC and cut-balanced, the
protocol (12) establishes modulus consensus.
Although in general it is hard to verify the uniform or essential strong connectivity of G[A(·)], where A(t) = Fij (t, x(t))
depends on the concrete solution, in special cases such a
property may be proved. For instance, it is implied by the
global strong ε-connectivity [28, Section IV-B]: the graph
G(F̂ij (t, x)) is strongly ε-connected for any t, x. The result
of Theorem 6 extends the result from [28, Section IV-B] in
several ways. First of all, it deals with time-variant gains
Fij (t, x) and does not require them to have a constant sign.
In particular, system (12) does not necessarily generate orderpreserving flow [27]. Moreover, we do not assume that the
graph G[A(t)] is weight-balanced which can hardly be provided for nonlinear functions Fij . In the case of USC graphs,
the balance is not needed at all; in the case of ESC graphs
it is replaced by the much weaker cut-balance condition. At
last, we relax the connectivity assumption.
VI. P ROOFS
We start with the proof of Lemma 1 and then prove
results, concerned with static graphs (Subsection VI-A). To
proceed with the case of dynamic graph, we elaborate some
useful techniques in Subsections VI-B and VI-C, entailing also
Lemmas 3 and 5. The case of cut-balanced graph is considered
in Subsection VI-D. In Subsections VI-E and VI-F we prove
the modulus consensus criterion for directed dynamic graphs
and its implications, dealing with nonlinear protocols.
Throughout the section, Φ(t|t0 ) (where t, t0 ≥ 0) stands
for the Cauchy evolutionary matrix of the system (4), that is,
the solution of the Cauchy problem for (4) with initial data
x(t0 ) = x0 is given by x(t) = x(t|t0 , x0 ) = Φ(t|t0 )x(0).
Proof of Lemma 1: Assume the protocol (4) establishes
modulus consensus. Note that since functions xi (t) are continuous, existence of the limits lim |xi (t)| = x∗ implies
t→+∞
that the limits
lim xi (t) also exist (and equal to ±x∗ ).
t→+∞
Therefore Φ(t) −→ Φ∗ := [φ1 , . . . , φN ] as t → ∞, where
t→∞
each column φj have entries with equal modules. The same
PN
applies for any linear combination j=1 αj φj . If Φ∗ = 0, the
statement of Lemma 1 is evident, taking v = 0. Assume that
8
one of φj , say, φ1 is nonzero, thus φ1 = v1 ρ where v1 6= 0
and ρ is a vector with entries ±1. Notice that for any real
numbers α, β 6= 0 we have |α − β| =
6 |α + β|. Therefore,
if φj 6= 0 for some j 6= 1, all entries of φj − φ1 have the
same module if and only if φj = vj ρ, vj 6= 0. If φj = 0, we
put by definition vj = 0. Therefore, φj = vj ρ for any j and
lim xj (t) = Φ∗ = ρv T x(0), where v := (v1 , . . . , vN )T .
t→∞
A. Proofs of Lemmas 2, 4 and Theorem 1
Proof of Lemma 2: We use a gauge transformation [28].
Suppose that the graph is structurally balanced; let V1 and V2
be hostile camps covering all the nodes. Introducing a diagonal
matrix D = diag(d1 , . . . , dN ) by di = 1 for i ∈ V1 and di =
−1 for i ∈ V2 , one easily shows that the gauge transformation
x 7→ z := Dx transforms the system (4) into
ż(t) = −L[abs A]z(t).
(14)
The properties of cooperative protocols (14) are widely known
[1], [2], [49]. Since the matrix B = abs A is non-negative,
L[B] has zero eigenvalue with eigenvector 1N . The algebraic and geometric multiplicities of this eigenvalue coincide
[49], and it is simple if and only the graph G[B] is QSC
(has oriented spanning tree). If this holds, the protocol (14)
establishes consensus and exp(−L[B]t) −→ 1̄N v. Since
t→+∞
L[B] = DL[A]D and G[A] is QSC if and only if G[B] is
QSC, one immediately obtains the claims of Lemma 2.
Proof of Lemma 4: The proof employs the same idea of
the gauge transformation, retracing the arguments from [28,
Section III-B-2)]. Suppose the graph G[A(t)] is structurally
balanced for any t with static hostile camps V1 and V2 .
The gauge transformation x 7→ z := Dx, introduced in
the foregoing, transforms the system (4) into (14). Since
G[abs A(t)] is UQSC, the protocol (14) establishes consensus
[1], [46] and thus opinions in network (4) polarize.
To proceed with the proof of Theorem 1, note that stability
is equivalent to the asymptotic stability of the system (4),
which means that (−L[A]) is a Hurwitz matrix. By using
the Gershgorin disk theorem, it was shown in [28] that all
eigenvalues of L[A], except for possibly zero, have positive
real parts. Hence protocol is stable if and only if 0 is not an
eigenvalue of L[A], i.e. L[A]ρ = 0 is impossible when ρ 6= 0.
We need the following simple lemma.
Lemma 7: Let L[A]ρ = 0 and |ρ|∞ = 1. Then V1 = {j :
∆
ρj = 1}, V2 = {j : ρj = −1} are hostile camps and V∗ =
V1 ∪ V2 6= ∅. If j ∈ V∗ , k 6∈ V∗ , then ajk = 0.
Proof: By assumption, |xk | ≤ 1 for any k and |xj | =
1 for some j, hence V∗ 6= ∅. For any such j one has
P
P
P
ρj
|ajk | =
ρk ajk . We note that
ρk ajk ≤
k6=j
k6=j
Pk6=j
|ajk |. The equality implies, firstly, that |ρk ajk | = |ajk |
k6=j
(i.e. k ∈ V∗ or ajk = 0) and secondly, all non-zero terms
ρk ajk have the same sign (coinciding with sgn ρj ). In other
words, if j ∈ V∗ , then ajk ρj ρk ≥ 0 ∀k ∈ V∗ , and ajk = 0
when k 6∈ V∗ , from where the statement is straightforward.
Proof of Theorem 1: Sufficiency in the first statement is
immediate from Lemma 7: if no “in-isolated” structurally balanced subgraph exists, then 0 is not an eigenvalue of (−L[A]),
which therefore is a Hurwitz matrix. Necessity follows from
Lemma 4: nodes of the subgraph are independent of the
remaining agents and reach bipartite consensus, and hence
stability of the whole community is impossible. Lemma 4
implies sufficiency in the second statement. To prove necessity,
suppose that bipartite consensus is established. Then 0 is an
eigenvalue of L[A] and there exists an eigenvector ρ, such that
L[A]ρ = 0, and |ρ|∞ = 1. Bipartite consensus implies that for
this vector one has |ρ1 | = . . . = |ρN | = 1. Applying Lemma 7,
we obtain the structural balance of G[A] since V∗ = 1 : N .
B. Ordering permutations
In this Subsection, we elaborate some useful techniques to
be used in the subsequent proofs.
Given a family of scalar functions f1 (t), . . . , fN (t) (where
t ≥ 0), let [k 1 (t), . . . , k N (t)] be the ordering permutation,
sorting the set {f1 (t), . . . , fN (t)} in the ascending order:
fk1 (t) (t) ≤ fk2 (t) (t) ≤ . . . ≤ fkN (t) (t). If fj (t) = fk (t)
for some j,k and t, the permutation is not uniquely defined.
The following technical Lemma shows the permutation may
always be taken in some regular way.
Lemma 8: Assume that fj are locally Lipschitz. Then there
exists such an ordering permutation k 1 (t), . . . , k N (t) that
∆
k j (·) are measurable, functions Fj (t) = fkj (t) (t) are locally
Lipschitz and Ḟj (t) = f˙kj (t) (t) for any j and a.a. t ≥ 0.
To prove Lemma 8, we need the following proposition.
Lemma 9: Let f∗ (t) := maxi∈[1:N ] fi (t) and j(t) be the
index such that f∗ (t) = yj(t) (t) for a.a. t. In the case of nonuniqueness, one always may choose j(·) in a way that it is
measurable and f˙∗ (t) = ẏj(t) (t) for a.a t. The claim remains
valid, replacing max with min.
Proof: As follows from the generalized version of Danskin theorem [50, Theorem 2.1], f∗ (·) is locally Lipschitz
and f˙∗ (t) ∈ {ẏj (t) : yj (t) = f∗ (t)} for a.a. t ≥ 0. The
Filippov-Castaing measurable selector theorem (see e.g. [51,
Theorem 1]) yields that a measurable function j(t) exists such
that f˙∗ (t) = ẏj(t) (t) and yj(t) (t) = f∗ (t). The last claim is
proved by replacement fj 7→ −fj .
Proof of Lemma 8: The proof is via induction on N . For
N = 1, the claim is evident. Let it be true for some N , and
let f1 (·), . . . , fN +1 (·) be locally Lipschitz. The ordering of
the first N functions z1 (·), . . . , zN (·) are locally Lipschitz by
0
the induction hypothesis. The recursion yN
+1 (t) := yN +1 (t),
ν−1
ν
yN
+1 (t) := max{zν (t); yN +1 (t)},
ν−1
zbν (t) := min{zν (t); yN
+1 (t)}, ν ∈ 1 : N
results in the ordering zb1 (t) ≤ . . . ≤ zbN (t) ≤ zbN +1 (t) :=
yN +1,N (t) of the entire set y1 (·), . . . , yN +1 (·). By applying
Lemma 9 at every recursion step, we see that zbν (·) are locally
Lipschitz. For any ν = 1, . . . , N , the sequence
ν
Zν (t) := [b
z1 (t), zb2 (t), . . . , zbν (t), zν+1 (t), . . . , zN (t), yN
+1 (t)]
(15)
is obtained from Zν−1 via a permutation Jν (t) of indices,
which either is the identity one or exchanges the places of the
νth and (N + 1)th entries and may be chosen measurable. By
Lemma 9, the sequence Z0ν (t) that results from replacement of
9
any function in (15) by its derivative is related to Z0ν−1 (t) by
the same permutation for a.a. t. The sequence Z0 (t) is obtained
from Y(t) := [y1 (t), . . . , yN +1 (t)] via a permutation of indices
K(t) = [k 1 (t), . . . , k N (t), N +1]. By the induction hypothesis,
0
K(t) also relates Z00 (t) and Y0 (t) := [y10 (t), . . . , yN
+1 (t)] for
a.a. t. Then KN +1 (t) = JN ◦ · · · ◦ J1 ◦ K(t) transforms Y0 (t)
into Z0N (t) for a.a. t, which proves the induction step.
from where (17) follows since Mi = χki . Using (17), one
has ṀN (t) ≤ 0 for a.a. t ≥ 0 since |θ̃ij | ≤ 1 and Mj ≤
M
R ∞N for any j thus MN (·) is non-increasing. We have also
|ṀN (t)|dt = M (0) − inf t≥0 M (t) ≤ M (0).
0
We also require one additional simple tool which allows
to examine the behavior of system (5) by comparing it
with a simpler system, obtained by ignoring all inessential
interactions between the agents. Consider a protocol
˙ = −L[A(t)]ξ(t)
ξ(t)
C. Some Technical Lemmas and Proofs of Lemmas 3, 5
To start with, we note that since ajk (t) are locally bounded,
this also holds for ẋj and therefore xj (·) are locally Lipschitz.
For any solution x(t) of (4) we introduce functions
χj (t) := |xj (t)|, θij (t) := sgn aij (t) sgn xi (t) sgn xj (t).
The following lemma gives a useful interpretation of dynamics
(4) in terms of the moduli functions χk .
Lemma 10: The functions χj are locally Lipschitz, thus
absolutely continuous. For a.a. t ≥ 0 and any k one has
χ̇k (t) =
N
X
|aki (t)|[χi (t)θki (t) − χk (t)].
(16)
i=1
(since χk ≥ 0, it has sign sgn χk (t) equal to 0 or 1).
Proof: Since ajk are assumed to be locally bounded,
xk (·) are locally Lipschitz. The same applies to χk (·) since
|χk (t1 ) − χk (t2 )| = ||xk (t1 )| − |xk (t2 )|| ≤ |xk (t1 ) − xk (t2 )|.
Therefore χk are absolutely continuous and for a.a. t > 0
the derivative χ̇k (t) exists. For such t we immediately have
χk (t) = 0 =⇒ χ̇k (t) = 0 by the Fermat theorem since 0 is the
global minimum of χk , which proves (16) (indeed, xk (t) = 0
implies that θki (t) = 0 for any i). Let χk (t) > 0. Since
sgn xk (s) = sgn xk (t) for s ≈ t, one has
(5)
χ̇k (t) = ẋk (t) sgn xk (t) =
(5)
=
N
X
Ṁj (t) =
N
X
|ãji (t)|[Mi (t)θ̃ji (t) − Mj (t)].
i=1
Proof: Combining (16) with Lemma 8, we get
=
0
We are going to show that the essentially equivalent protocol
provides the same limit sets for the solutions. Let B1 := {x ∈
RN : |x|1 ≤ 1∀i} be a unit ball in the | · |∞ -norm, positively
invariant by Lemma 11: x(t0 ) ∈ B1 =⇒ x(t) ∈ B1 ∀t ≥ t0 .
Definition 3: Let Ωt0 ,x0 := {y ∈ RN :S ∃tn → ∞ :
Ωt0 ,x0 ⊆ B1
x(tn |t0 , x0 ) −→ y)}. We call the set Ω :=
n→∞
t0 ,x0
the Ω-set of the system (4) (the union is over t0 ≥ 0, x0 ∈ B1 ).
Lemma 12: Suppose the protocols (4) and (18) are essentially equivalent. Then for any ε > 0 there exists T0 = T0 (ε)
such that ξ(T0 ) = x(T0 ) ∈ B =⇒ |x(t)−ξ(t)|∞ ≤ ε ∀t ≥ T0 .
In particular, the systems (4) and (18) have equal Ω-sets.
Proof: Since x(t0 ) ∈ B1 =⇒ x(t) ∈ B1 ∀t ≥ t0 thanks to
Lemma 11, one has
|Φ(t|t0 )|∞ ≤ 1∀t0 ≥ 0∀t ≥ t0 . Let T0 >
R∞
0 be so large that T0 |L[A(t)] − L[A(t)]|∞ dt < ε. Applying
Lemma 11 to (18) implies that ξ(T0 ) ∈ B =⇒ ξ(t) ∈ B∀t ≥
T0 , and hence
R ∞ ∆(t) := (L[A(t)] − L[A(t)])ξ(t) satisfies the
inequality T0 |∆(t)|∞ dt < ε. Since ξ 0 (t) = −L[A(t)]ξ +
∆(t), condition x(T0 ) = ξ(T0 ) ∈ B implies
Z t
ξ(t) − x(t) = ξ(t) − Φ(t|T0 )ξ(T0 ) =
Φ(t|s)∆(s)ds
T0
Rt
which proves (16) since xi (t) = χi (t) sgn xi (t).
Henceforth, we fix some ordering permutation
k 1 (t), . . . , k N (t) for the family χ1 (t), . . . , χN (t) (see
Subsect.VI-B) and put Mj (t) = χkj (t) , in particular,
MN (t) = maxj |xj (t)|. Combining Lemmas 10 and 8, one
gets the following.
Lemma 11: The function MN (t) = max |xi (t)| is noni∈1:N
R∞
increasing and hence 0 |ṀN (t)|dt < ∞. For ãji := akj ki ,
θ̃ji := θkj ki , for a.a. t ≥ 0 and any j ∈ 1 : N one has
Ṁj (t) =
where A(t) = (aij (t)) is locally bounded. We say the protocol
(18) is essentially equivalent to (4) if
Z ∞
|aij (t) − aij (t)|dt < ∞ ∀i, j.
|aki (t)|[xi (t) sgn aki (t) sgn xk (t) − χk (t)],
i=1
N
X
i=1
N
X
i=1
|akj i (t)| [χi (t)θkj i (t) − Mj (t)]
|akj ki (t)| [χki (t)θkj ki (t) − Mj (t)] ,
(17)
(18)
and therefore |ξ(t) − x(t)|∞ ≤ T0 |Φ(t|s)|∞ |∆(s)|∞ ds < ε.
This proves the first claim from where the second one immediate follows: indeed, for any solution x(t|t0 , x0 ) with
t0 ≥ 0, x0 ∈ B and any ε > 0 one can find T0 such that
|ξ(t|T0 , ξ0 ) − x(t|t0 , x0 )|∞ = |ξ(t|T0 , ξ0 ) − x(t|T0 , ξ0 )|∞ ≤ ε
for any t ≥ T0 , where ξ0 := x(T0 |t0 , x0 ) ∈ B. Therefore, any
set Ωt0 ,x0 from Definition 3 belongs to the Ω-set of (18) and
thus the whole Ω-set of (4) belongs to the Ω-set of (18). The
inverse inclusion is proved in the same way.
Remark 4: Such properties of the protocol as modulus
consensus, stability, bipartite consensus, and “partial” modulus
consensus (modulus agreement among a subgroup of agents)
in fact depend only on the Ω-set. For instance, modulus
consensus is established if and only if the Ω-set is comprised
of the set {x ∈ B : |x1 | = . . . = |xN |}. Lemma 12 implies
that those properties are preserved, replacing the protocol with
essentially equivalent one.
We are now going to prove Lemmas 3, 5. Lemma 3 immediately follows from Lemma 11 since MN (t) = maxj |xj (t)|.
Proof of Lemma 5: Suppose the protocol (4) establishes
bipartite consensus. Lemma 12 and Remark 4 show that,
10
without loss of generality, one may assume ajk ≡ 0 unless
j essentially interacts with k, i.e. (j, k) ∈ E[A(·)]. If the
topology is not EQSC, then the graph G[A(·)] is not QSC and
thus, as shown in [5], there exist non-empty disjoint subsets
V1 , V2 ⊂ 1 : N that has no incoming arcs: ajk = 0 if
j ∈ V1 , k 6∈ V1 or j ∈ V2 , k 6∈ V2 . Therefore, the opinions
of agents from V1 are independent on the opinions of agents
from V2 , and hence bipartite consensus is impossible.
D. Proof of Theorems 3, 4 and Corollary 3
Henceforth G[A(·)] is cut-balanced. The cornerstone of the
proofs is the following lemma, based on Lemma 11.
Lemma 13: For a given solution of (4), let η̃ji (t) :=
|ãji (t)|(θ̃ji (t)Mi (t) − Mj (t)), where θ̃ji are the same as in
(16). Then η̃ji ∈ L1 [0; ∞] for any i, j, so that Ṁj ∈ L1 [0; ∞].
Proof: Lemma 11, which is valid for any dynamic graph,
shows that η̃N 1 , . . . , η̃N,N −1 ∈ L1 [0; ∞] since η̃N j ≤ 0 and
ṀN ∈ L1 . We are going to show that η̃N −1,j ∈ L1 .
Note that |η̃N j (t)| ≥ |ãN j (t)|(MN (t) − MN −1 (t)) for any
j < N . Indeed, |θ̃N,j Mj − MN | ≥ |MN | − |θ̃N,j Mj | ≥
MN − MN −1 since Mj ≤ MN −1 . Applying the definition
0
1 2
of cut-balance (2) to
. . , k N −1 } and V 00 =
P V = {k , k , .P
N
{k }, one obtains
|ãjN (t)| ≤ K
|ãN j (t)|. Therefore
j<N
j<N
∆
the function SN (t) = (MN (t) − MN −1 (t))
NP
−1
|ãjN (t)| is
j=1
N
−2
X
+
η̃N −1,j (t)+η̃N
−1,N (t). (19)
j=1
−
Since −η̃N
−1,N (t) ≤ 0 and η̃N −1,j (t) ≤ 0 for any t ≥ 0
1
and the last term in (19)
R ∞is L -summable, we either have
−
1
η̃N −1,N , η̃N −1,j ∈ L or 0 ṀN −1 (t)dt = −∞. The latter is
impossible since MN −1 (t) ≥ 0. Thus η̃N −1,j ∈ L1 [0; ∞] ∀j.
Our next step is to prove that η̃N −2,j ∈ L1 for any
j. We note that for any j ≤ N − 2 and r = N − 1, N
we have |η̃rj (t)| ≥ |ãrj (t)|(MN −1 (t) − MN −2 (t)).
Applying (2) to V 0
=
{k 1 , k 2 , . . . , k N −2 } and
NP
−2 P
N
V 00 = {k N −1 , k N }, we obtain that
|ãjr (t)| ≤
j=1 r=N −1
K
NP
−2
N
P
|ãrj (t)|
and
hence
the
function
j=1 r=N −1
SN −1 (t)
∆
=
(MN −1 (t) − MN −2 (t))
NP
−2
−
−
ṀN −2 (t) = −η̃N
−2,N −1 (t) − η̃N −2,N (t) +
N
P
|ãjr (t)|
j=1 r=N −1
belongs to L1 . If η̃N −2,N −1 (t)
>
0, one has
η̃N −2,N −1 (t) = |ãN −2,N −1 (t)|(MN −1 − MN −2 (t)). There+
fore, η̃N
−2,N −1 ≤ SN −1 . Analogously, if η̃N −2,N (t) > 0
then η̃N −2,N (t) = |ãN −2,N (t)|(MN (t) − MN −1 (t) +
+
MN −1 − MN −2 (t)) and hence η̃N
−2,N ≤ SN −1 + SN . Thus
N
−3
X
η̃N −2,j (t)+
j=1
+
+
+η̃N
−2,N −1 (t) + η̃N −2,N (t).
−
−
Since −η̃N
−2,N (t) ≤ 0, −η̃N −2,N −1 (t) ≤ 0 and η̃N −2,j (t) ≤
0 for
j < N −2, then either all of these functions are summable
R∞
or 0 ṀN −2 (t)dt = −∞. The latter is impossible since
MN −2 (t) ≥ 0. Therefore, η̃N −2,j ∈ L1 [0; ∞] ∀j.
∆
Applying the same procedure, one proves that SN −2 (t) =
NP
−2 P
N
|ãjr (t)| is summable and
(MN −2 (t) − MN −3 (t))
j=1 r=N −2
η̃N −3,j ∈ L1 [0; ∞] ∀j, and so on, η̃ij ∈ L1 [0; ∞].
Corollary 4: For any solution of (4) one has ηij :=
|aij |(θij χj − χi ) ∈ L1 [0; ∞] and χ̇i ∈ L1 [0; ∞], so the finite
limits χ0i = lim χi (t) and x0i := lim xi (t) = ±χ0i exist.
t→+∞
t→+∞
If (i, j) ∈ E[A(·)] then |x0i | = |x0j |; moreover, x0i = x0j when
(i, j) ∈ E + [A(·)] and x0i = −x0j when (i, j) ∈ E − [A(·)]; as a
consequence, x0i = x0j = 0 if (i, j) ∈ E + [A(·)] ∩ E − [A(·)].
Proof: Since k 1 , . .P
. , k N is just P
a permutation of the
set 1 : N , we have
i,j |ηij | =
i,j |η̃ij |, and hence
ηji ∈ L1 [0; ∞] for any i, j. From (16) one immediately
obtains that χ̇i ∈ L1 [0; ∞] for any i, from where the
existence of the finite limits χ0i := lim χi (t) is immediate.
t→+∞
The limits x0i =
summable. Note that if η̃N −1,N (t) > 0 then θ̃N −1,N (t) = 1
and η̃N −1,N (t) = |aN −1,N (t)|(MN (t) − MN −1 (t)). Thus,
+
+
1
η̃N
−1,N (t) ≤ SN (t) and hence η̃N −1,N ∈ L . By invoking
(17) for j = N − 1, one obtains that
−
ṀN −1 (t) = −η̃N
−1,N (t)+
+
+
1
η̃N
−2,N , η̃N −2,N −1 ∈ L [0; ∞]. Applying (17), we get
lim xi (t) = ±χ0i exist since xi are
t→+∞
continuous. From |ηij (t)| ≥ |aij (t)||χj (t) − χi (t)| we know
that if δ := |χ0i − χ0j | > 0, for
R ∞large t > 0 one has
|ηij (t)| ≥ |aij (t)|δ/2 and thus 0 |aij (t)|dt < ∞ (aij
are locally bounded). Therefore, χ0i = χ0j whenever i and
j essentially interact. Suppose that x0i = x0j 6= 0. As
t → ∞, one has sgn xi (t) = sgn x0i = sgn xj (t) and thus if
ajk (t) < 0, we have ηij (t) = |aij (t)|(−xi (t) − xj (t)). Thus
0
|ηij (t)| ≥ |a−
ijR(t)||xi | for t > 0 sufficiently large, from where
−
one has that ij |aij (t)|dt < ∞. So if the agents essentially
compete, the option x0i = x0j 6= 0 is impossible, and thus
x0i = −x0j (with possibility of x0i = x0j = 0). Analogously,
0
one can easily show that |ηij (t)| ≥ |a+
ij (t)||xi | for t > 0
0
0
sufficiently large if xi = −xj 6= 0 which proves the essential
cooperation excludes the possibility of x0i = −x0j 6= 0 and
thus x0i = x0j . At last, simultaneous essential cooperation and
essential competition imply that x0i = x0j = −x0j = 0.
Corollary 5: If the network topology is ESC, the protocol
(4) establishes modulus consensus (χ01 = . . . = χ0N ), and
χ0j = 0 unless G ± is well-defined (E + ∩ E − = ∅) and SB.
Proof: The first statement immediately follows from
Corollary 4 since for any path i1 , i2 , . . . , ir in G one has
χ0i1 = . . . = χ0ir . Assume that for some initial vector x(0) one
has χ0i 6= 0. Corollary 4 implies that no pair of agents may be
both essentially cooperative and essentially competitive, and
thus E + ∩ E − = ∅ so the signed graph G ± is well-defined. We
have to show G ± is structurally balanced, in other words [17],
[28], has no negative cycles. Indeed, the weight of any arc
(i, j) ∈ E is sij = 1 if (i, j) ∈ E + (hence x0i = x0j ) and sij =
−1 if (i, j) ∈ E − (and thus x0i = −x0j ) so that x0i x0j sij ≥ 0.
Given a cycle i1 , . . . , in = i1 , Multiplying the inequalities
11
x0ik x0ik+1 sik ik+1 ≥ 0, where i1 , i2 , . . . , in = i1 is a cycle and
k ∈ 1 : (n − 1), one has si1 i2 si2 i3 . . . sin−1 in (x01 . . . x0n )2 ≥ 0
which means that the cycle is positive.
Proof of Theorem 3: The necessity in the first statement
part follows from Lemma 5 and Corollary 5: indeed, the ESC
condition is necessary for bipartite consensus independent of
the cut-balance property. Under the ESC condition, bipartite
consensus is possible only when G ± exists and is SB. To
prove sufficiency, suppose that G ± is well-defined, strongly
connected and structurally balanced. Thanks to Corollary 5,
the protocol (4) establishes modulus consensus, and it remains
to show it is bipartite consensus. Indeed, consider the protocol
(18), where A(t) = (ajk (t)) and ajk (t) = a+
jk (t) when
−
(j, k) ∈ E + , ajk (t) = a−
(t)
when
(j,
k)
∈
E
,
and
otherwise
jk
ajk ≡ 0. The protocol (18) is equivalent to the protocol (4) and
hence establishes modulus consensus of the same type as (18)
by Lemma 12 and Remark 4. Let V1 , V2 be the hostile camps
of the graph G ± , covering all its nodes. Taking xj = +1
for j ∈ V1 and xj = −1 for j ∈ V2 , one can show that
x = (x1 , . . . , xN )T is an equilibrium point for (18) and hence
bipartite consensus is reached. Accordingly to Corollary 4,
x0i = x0j if i, j ∈ Vr1 or i, j ∈ Vr2 , and x0i = −x0j whenever
i ∈ Vr1 , j ∈ Vr2 ; therefore, opinions polarize unless Er− = ∅.
The claim about stability also follows from Corollary 4.
Remark 5: It may seem that Theorem 3 may be proved by
applying the result for unsigned graphs [10] to the system,
obtained from the protocol (18) just constructed via the gauge
transformation (as Lemma 4 was derived from the relevant
result on cooperative agents). Unfortunately, this is not the
case. The problem is that the graph G[A(·)] is no longer cutbalanced.
This property depends not only on the integrals
R∞
|a
(s)|ds,
but on the whole function A(t), and is lost
jk
0
after removing inessential interactions.
Proof of Theorem 4: The first and the last claims follow
from Corollary 4. The second claim is proved by passing to an
auxiliary protocol (18), constructed in the proof of Theorem 3.
Let Vr1 and Vr2 be hostile camps in Gr± , then taking xi = +1
for i ∈ Vr1 , xi = −1 for i ∈ Vr2 and xi = 0; otherwise, we get
an equilibrium point of (18), and therefore, modulus consensus
in the subcommunity Vr is bipartite. Thanks to Corollary 4,
x0i = x0j if i, j ∈ Vr1 or i, j ∈ Vr2 , and x0i = −x0j whenever
i ∈ Vr1 , j ∈ Vr2 ; therefore, opinions polarize unless Er− = ∅,
when bipartite consensus is established.
Sk (t) ≤ R(t1 − t0 ) ∀t ∈ [t0 ; t1 ]. From (16) one derives that
Zt
N
X
|aki (τ )|χi (τ )dτ e−Sk (t) .
χk (t) ≤ χk (t0 ) + eSk (τ )
0
i=1
(21)
P
Since χi (τ ) ≤ M , one has
|aki (τ )|χi (τ ) ≤ M sk (τ ) and
χk (t) ≤ M + [χk (t0 ) − M ]e−Sk (t) ≤ M + [χk (t0 ) − M ]θ0 .
Lemma 15: Under assumptions of Lemma 14, suppose the
Rt
∆
graph G[ t01 abs A(t)dt] is strongly ε-connected and θ =
2
εθ0 < 1. Let V ( 1 : N be a non-empty set, |x(t0 )|∞ = M
and max |xj (t0 )| = M 0 . Then there exists k ∈ V such that
j6∈V
|xk (t1 )| ≤ θM 0 + (1 − θ)M.
(22)
R t1
Proof: Since the graph G[ t0 abs A(t)dt] is strongly
ε-connected,
there exist i 6∈ V and k ∈ V such that
R t1
|aki (t)|dt ≥ ε. From (20) one has χj (t) ≤ M + (M 0 −
t0
M )θ ∀j 6∈ V and χj (t) ≤ M ∀j ∈ V , which entails that
PN
eSk (τ ) i=1 |aki (τ )|χi (τ ) ≤ M eSk (τ ) sk (τ ) − (M − M 0 )θε
(we used that eSk (τ ) ≥ 1). Inequality (21) yields now that
χk (t1 ) ≤ M + (M 0 − M )θ0 εe−Sk (t1 ) ≤ M + (M 0 − M )θ,
since eSk (t1 ) ≤ eR(t1 −t0 ) = θ0 .
Proof of Theorem 2: By assumption, there exist
R,Rε, T > 0 such that |A(t)|∞ ≤ R and the graph Gt =
t+T
G[ t
abs A(τ )dτ ] is strongly ε-connected for any t ≥ 0.
Let θ0 = e−RT and θ = εθ02 , without loss of generality we
may assume that θ < 1.
We know from Lemma 3 that the maximal modulus MN (t)
always has a limit: MN (t) → M∗ as t → +∞ (here we use
the notation introduced in Subsection VI-C so that Mj (t) is
the j-th modulus in the ascending order, Mj (t) = |xkj (t) (t)|).
Our goal is to show that Mj (t) → M∗ via induction by j =
N, N − 1, . . . , 1 . If M∗ = 0 the latter claim is trivial since
0 ≤ Mj (t) ≤ MN (t); hence we may assume that M∗ > 0.
For j = N our claim holds by definition of M∗ . Suppose
we have proved that MN (t), MN −1 (t), . . . , Mr+1 (t) → M∗
as t → +∞ and have to show that Mr (t) → M∗ . Since
Mr (t) ≤ Mr+1 (t), it suffices to show that lim Mr (t) ≥
t→+∞
M∗ . Assume on the contrary that lim Mr (t) = m < M∗ .
t→+∞
For any δ > 0, we have M∗ + δ > Mj (t) > M∗ − δ for large
t > 0 and j > r. On the other hand, there exist a sequence
tn → ∞, along which Mr (tn ) < m + δ. Assume that δ > 0
is so small that (m + δ)θ + (M∗ + δ)(1 − θ) < M∗ − δ and
E. Proof of Theorem 2 and Corollary 1
(m + δ)θ0 + (M∗ + δ)(1 − θ0 ) < M∗ − δ, that is, 2δ <
(M∗ − m) max(θ0 , θ).
We start with some useful estimates for the solutions.
Accordingly to Lemma 15, applied for t0 = tn and V =
Lemma 14: Suppose that A : [t0 ; t1 ] → RN is a matrixr+1
∆
{k
(tn ), . . . , k N (tn )}, M = M∗ + δ and M 0 = m∗ + δ,
valued function, |A(t)|∞ ≤ R for a.a. ∈ [t0 ; t1 ] and θ0 =
there exists j ∈ V such that |xj (t0 + T )| < M∗ − δ. Similarly,
e−R(t1 −t0 ) . For any solution of (4) one has
for any j 6∈ V we have |xj (tn + T )| < M∗ − δ thanks to
|xk (t)| ≤ θ0 |xk (t0 )|+(1−θ0 )|x(t0 )|∞ ∀k ∈ 1 : N ∀t ∈ [t0 ; t1 ]. Lemma 14. Therefore, at time tn + T there are at least r + 1
(20) agents, whose opinions have moduli less than M∗ − δ and
hence Mr+1 (tn + T ) ≤ M∗ − δ for any n. We arrived at the
∆
Proof: Let M = |x(t0 )|∞ . Thanks to Lemma 3, one has
contradiction
with the induction hypothesis. We thus proved
P
∆
N
|xj (t)| ≤ M for any j, t. Let sk (t) =
j=1 |akj (t)| and
that
M
(t)
→
M∗ ∀j for any solution of (4), that is, modulus
j
∆ Rt
Sk (t) = t0 sk (ξ) dξ. By assumption, sk (t) ≤ R and hence consensus is established.
12
Proof of Corollary 1: By virtue of Theorem 2, modulus
consensus is established. According to Lemma 1 it only three
types of such a consensus are possible, which are stability,
polarization and consensus. The common feature of the first
two types is that for a.a. x(0) there exists i ∈ 1 : N such that
lim xi (t) ≤ 0. It is well known [6], [9], [46] that the convex
polarization). Getting rid of the restriction of static topologies
allows to examine linear and nonlinear dynamics of social
networks, where the agents may change their relations from
friendship to hostility and vice versa. We are currently working
with sociologists to test the theoretical results presented in this
paper using data from human social groups.
hull of the agents’ states ∆(t) = [min xi (t), max xi (t)] is noni
i
expanding over time. Hence if xi (0) ≥ 1 ∀i, then xi (t) ≥ 1
for any t ≥ 0, so the first two options are not possible.
R EFERENCES
t→+∞
F. Proof of Lemma 6 and Theorems 5, 6
We start with proof of Lemma 6, being a basis for Theorem 5.
Proof of Lemma 6: We consider system (9), and the
protocol (10) may be studied in the same way. Equation (11)
is immediate from the definitions of aij and Hij . As follows
from Lemma 3, the solutions of (11) remain bounded since
|x(t)|∞ ≤ |x(0)|∞ . Since Hij (y, z) > 0 are continuous functions and the set {(y, z) : |y|, |z| ≤ |x(0)|∞ } is compact, there
exist M > m > 0 such that m ≤ Hij [y, z] ≤ M whenever
|y|, |z| ≤ |x(0)|∞ . By substituting y := xj (t) sgn aij (t) and
z := xi (t), one shows that m|ajk | ≤ |ajk | ≤ M |ajk |, from
where the claim of Lemma 6 is obvious.
Now we proceed with the proofs of Theorems 5 and 6.
Proof of Theorem 5: Since the right-hand sides of
(9), (10) are smooth in x, the solutions exist locally and
are unique. According to Lemma 3 and (11), the solutions
remain bounded and thus infinitely prolongable. Under the
USC assumption, modulus consensus follows from Theorem 2
and Lemma 6. If the graph G[A(·)] is ESC and cut-balanced,
modulus consensus is implied by Theorem 3.
Proof of Theorem 6: Using Lemma 3, one proves that
the solution is bounded and hence its derivative also remains
bounded due to (13), so any solution is infinitely prolongable.
Since x(t) is bounded, A(t) is also bounded due to (13). The
remaining claims follow now from Theorems 2, 3.
VII. C ONCLUSIONS AND RELATED WORKS
In the present paper, we extend a model of opinion dynamics in social networks with both attractive and repulsive
interactions between the agents, which was proposed in recent papers by C. Altafini, who considered the conventional
first-order consensus protocols over signed graphs. Altafini
showed, in particular, the possibility of opinion polarization
if the interaction graph is structurally balanced. In general,
the protocol establishes modulus consensus, where the agents
agree in modulus but may differ in signs (which not excludes
convergence of all opinions to zero). In the present paper, we
have examined dynamics of Altafini’s protocols with switching
directed topologies and offer sufficient conditions for reaching
modulus consensus that boil down to uniform strong connectivity of the network. Moreover, under the assumption of cutbalance, the uniform connectivity may be further relaxed. In
this case, we have obtained necessary and sufficient conditions
for modulus consensus, classified into stability (converges
of opinions to zero) and bipartite consensus (consensus or
[1] W. Ren and R. Beard, Distributed consensus in multi-vehicle cooperative
control: theory and applications. Springer, 2008.
[2] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multiagent
Networks. Princeton and Oxford: Princeton University Press, 2010.
[3] W. Ren and Y. Cao, Distributed Coordination of Multi-agent Networks.
Springer, 2011.
[4] M. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, vol. 69, pp. 118–121, 1974.
[5] L. Moreau, “Stability of multiagent systems with time-dependent communication links,” IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 169–
182, 2005.
[6] Z. Lin, B. Francis, and M. Maggiore, “State agreement for continuoustime coupled nonlinear systems,” SIAM Journ. of Control and Optimization, vol. 46, no. 1, pp. 288–307, 2007.
[7] M. Cao, A. Morse, and B. Anderson, “Reaching a consensus in a dynamically changing environment: convergence rates, measurement delays,
and asynchronous event,” SIAM Journal of Control and Optimization,
vol. 47, no. 2, pp. 601–623, 2008.
[8] V. Blondel, J. Hendrickx, A. Olshevsky, and J. Tsitsiklis, “Convergence
in multiagent coordination, consensus, and flocking,” in Proc. IEEE
Conf. Decision and Control, 2005, pp. 2996 – 3000.
[9] A. Matveev, I. Novinitsyn, and A. Proskurnikov, “Stability of
continuous-time consensus algorithms for switching networks with bidirectional interaction,” in Proceedings of European Control Conference
ECC-2013, 2013, pp. 1872–1877.
[10] J. Hendricx and J. Tsitsiklis, “Convergence of type-symmetric and cutbalanced consensus seeking systems,” IEEE Trans. Autom. Control,
vol. 58, no. 1, pp. 214–218, 2013.
[11] U. Münz, A. Papachristodoulou, and F. Allgöwer, “Consensus in multiagent systems with coupling delays and switching topology,” IEEE
Trans. Autom. Control, vol. 56, no. 12, pp. 2976–2982, 2011.
[12] G. Shi and K. Johansson, “Robust consensus for continuous-time multiagent dynamics,” SIAM J. Control Optim, vol. 51, no. 5, pp. 3673–3691,
2013.
[13] ——, “The role of persistent graphs in the agreement seeking of social
networks,” IEEE J. On Selected Areas In Communications, vol. 31, no. 9,
pp. 595–606, 2013.
[14] L. Scardovi and R. Sepulchre, “Synchronization in networks of identical
linear systems,” Automatica, vol. 45, no. 11, pp. 2557–2562, 2009.
[15] J. Hu and W. Zheng, “Emergent collective behaviors on coopetition
networks,” Physica A, vol. 378, pp. 1787–1796, 2014.
[16] J. Hu and H. Zhu, “Adaptive bipartite consensus on coopetition networks,” Physica D, vol. 307, pp. 14–21, 2015.
[17] D. Easley and J. Kleinberg, Networks, Crowds and Markets. Reasoning
about a Highly Connected World. Cambridge: Cambridge Univ. Press,
2010.
[18] S. Wasserman and K. Faust, Social Network Analysis: Methods and
Applications. Cambridge: Cambridge Univ. Press, 1994.
[19] A. Fläche and M. Macy, “Small worlds and cultural polarization,”
Journal of Math. Sociology, vol. 35, no. 1–3, pp. 146–176, 2011.
[20] J. Dillard and L. Shen, “On the nature of reactance and its role in persuasive health communication,” Communication Monographs, vol. 72,
no. 2, pp. 144–168, 2005.
[21] C. Hovland, I. Janis, and H. Kelley, Communication and persuasion.
New Haven: Yale Univ. Press, 1953.
[22] E. Aronson, T. Wilson, and R. Akert, Social psychology. Upper Saddle
River, NJ: Prentice Hall, 2010.
[23] W. Xia and M. Cao, “Clustering in diffusively coupled networks,”
Automatica, vol. 47, no. 11, pp. 2395–2405, 2011.
[24] R. Hegselmann and U. Krause, “Opinion dynamics and bounded confidence models, analysis, and simulation,” Journal of Artifical Societies
and Social Simulation (JASSS), vol. 5, no. 3, p. 2, 2002.
[25] G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch, “Mixing beliefs
among interacting agents,” Advances in Complex Systems, vol. 3, pp.
87–98, 2000.
13
[26] P. Dandekar, A. Goel, and D. Lee, “Biased assimilation, homophily, and
the dynamics of polarization,” PNAS, vol. 110, no. 15, pp. 5791–5796,
2013.
[27] C. Altafini, “Dynamics of opinion forming in structurally balanced social
networks,” PLoS ONE, vol. 7, no. 6, p. e38135, 2012.
[28] ——, “Consensus problems on networks with antagonistic interactions,”
IEEE Trans. Autom. Control, vol. 58, no. 4, pp. 935–946, 2013.
[29] Z. Meng, G. Shi, K. Johansson, M. Cao, and Y. Hong, “Modulus
consensus over networks with antagonistic interactions and switching
topologies,” http://arxiv.org/abs/1402.2766, 2014.
[30] P. Romanczuk and L. Schimansky-Geier, “Swarming and pattern formation due to selective attraction and repulsion,” Interface Focus, vol. 2,
pp. 746–756, 2012.
[31] W. Yu, G. Chen, M. Cao, J. Lü, and H. Zhang, “Swarming behaviors
in multi-agent systems with nonlinear dynamics,” Chaos, vol. 23, p.
043118, 2013.
[32] C. Wang, G. Xie, and M. Cao, “Forming circle formations of anonymous
mobile agents with order preservation,” IEEE Trans. Autom. Control,
vol. 58, no. 12, pp. 3248–3254, 2013.
[33] A. Zakhar’eva, A. S. Matveev, M. C. Hoy, and A. V. Savkin, “Distributed control of multiple non-holonomic robots with sector vision and
range-only measurements for target capturing with collision avoidance,”
Robotica, vol. 33, pp. 385–412, 2015.
[34] K. Ovchinnikov, A. Semakova, and A. Matveev, “Decentralized multiagent tracking of unknown environmental level sets by a team of
nonholonomic robots,” in Proc. of 6th Int. Congress on Ultra Modern
Communication and Control Systems (ICUMT), 2014, pp. 352–359.
[35] M. Valcher and P. Misra, “On the consensus and bipartite consensus
in high-order multi-agent dynamical systems with antagonistic interactions,” Systems Control Letters, vol. 66, pp. 94–103, 2014.
[36] H. Zhang and J. Chen, “Bipartite consensus of linear multi-agent systems
over signed digraphs: An output feedback control approach,” in Proc.
of IFAC World Congress, 2014, pp. 2118 – 2123.
[37] J. Hendrickx, “A lifting approach to models of opinion dynamics with
antagonisms,” in Proc. of IEEE Conference on Decision and Control
(CDC), 2014, pp. 2118 – 2123.
[38] C. Altafini and G. Lini, “Predictable dynamics of opinion forming
for networks with antagonistic interactions,” IEEE Trans. on Autom.
Control, vol. 60, no. 2, pp. 342–357, 2015.
[39] W. Xia, M. Cao, and K. Johansson, “Structural balance and opinion
separation in trust–mistrust social networks,” IEEE Trans. on Control of
Networks, 2015 (accepted).
[40] H. Smith, “Systems of ordinary differential equations which generate an
order preserving flow. A survey of results.” SIAM Review, vol. 30, pp.
87 – 113, 1988.
[41] Z. Lin, B. Francis, and M. Maggiore, “Necessary and sufficient graphical
conditions for formation control of unicycles,” IEEE Trans. Autom.
Control, vol. 50, no. 1, pp. 121–127, 2005.
[42] A. Proskurnikov and M. Cao, “Opinion dynamics using Altafini’s model
with a time-varying directed graph,” in Proceedings of IEEE ISIC 2014
(Part of IEEE MSC 2014), Antibes, 2014, pp. 849–854.
[43] A. Proskurnikov, A. Matveev, and M. Cao, “Consensus and polarization
in Altafini’s model with bidirectional time-varying network topologies,”
in Proceedings of IEEE CDC 2014, Los Angeles, 2014, pp. 2112–2117.
[44] R. Olfati-Saber and R. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Trans. Autom.
Control, vol. 49, no. 9, pp. 1520–1533, 2004.
[45] P. Chebotarev and R. Agaev, “The forest consensus theorem,” IEEE
Trans. Autom. Control, vol. 59, no. 9, pp. 2475–2479, 2014.
[46] L. Moreau, “Stability of continuous-time distributed consensus algorithms,” in Proc. IEEE Conf. Decision and Control (CDC 2004), 2004,
pp. 3998 – 4003.
[47] S. Marvel, J. Kleinberg, R. Kleinberg, and S. Strogatz, “Continuous-time
model of structural balance,” PNAS, vol. 108, no. 5, pp. 1771–1776,
2011.
[48] T. Antal, P. Krapivsky, and S. Redner, “Dynamics of social balance on
networks,” Phys. Rev. E, vol. 72, p. 036121, 2005.
[49] R. Agaev and P. Chebotarev, “On the spectra of nonsymmetric laplacian
matrices,” Linear Algebra Appl., vol. 399, pp. 157–168, 2005.
[50] F. Clarke, “Generalized gradients and applications,” Trans. of Amer.
Math. Soc., vol. 205, pp. 247 – 262, 1975.
[51] E. McShane and R. Warfield, “On Filippov’s implicit functions lemma,”
Proc. Amer. Math. Soc., vol. 18, pp. 41–47, 1967.
Anton V. Proskurnikov was born in Leningrad
(now St. Petersburg) in 1982. He received his M.Sc.
and Ph.D. degrees in 2003 and 2005 respectively,
both from St.-Petersburg State University and supervised by Prof. V.A. Yakubovich. From 2003 till 2010
he was an Assistant Professor of the Department
of Mathematics and Mechanics, St.Petersburg State
University. Now A. Proskurnikov is a postdoctoral
researcher at the ENTEG institute the University
of Groningen, The Netherlands. He also occupies
part-time researcher positions at St. Petersburg State
University and Institute for Problems of Mechanical Engineering of Russian
Academy of Sciences, St.Petersburg, Russia. His research interests include
dynamics and control of complex networks, multi-agent and decentralized
control nonlinear control, robust control, optimal control and optimization. In
2009 Anton Proskurnikov was awarded with a medal of Russian Academy of
Sciences for prominent young researchers for a series of works on optimal
tracking and disturbance rejection.
Alexey S. Matveev was born in Leningrad, USSR,
in 1954. He received the M.S. and Ph.D. degrees
in 1976 and 1980, respectively, both from the
Leningrad University. Currently, he is a professor
of the Department of Mathematics and Mechanics,
St. Petersburg State University. His research interests
include estimation and control over communication
networks, hybrid dynamical systems, and navigation
and control of mobile robots.
Ming Cao is currently a tenured associate professor
responsible for the research direction of network
analysis and control with the Faculty of Mathematics and Natural Sciences at the University of
Groningen, the Netherlands, where he started as a
tenure-track assistant professor in 2008. He received
the Bachelor degree in 1999 and the Master degree
in 2002 from Tsinghua University, Beijing, China,
and the PhD degree in 2007 from Yale University,
New Haven, CT, USA, all in electrical engineering.
From September 2007 to August 2008, he was a
postdoctoral research associate with the Department of Mechanical and
Aerospace Engineering at Princeton University, Princeton, NJ, USA. He
worked as a research intern during the summer of 2006 with the Mathematical
Sciences Department at the IBM T. J. Watson Research Center, NY, USA.
His main research interest is in autonomous agents and multi-agent systems,
mobile sensor networks and complex networks. He is an associate editor for
Systems and Control Letters, and for the Conference Editorial Board of the
IEEE Control Systems Society. He is also a member of the IFAC Technical
Committee on Networked Systems.
| 3cs.SY
|
arXiv:1704.03627v2 [cs.HC] 6 Dec 2017
Real-time On-Demand Crowd-powered Entity Extraction
(Supplementary Technical Report)
Ting-Hao (Kenneth) Huang
Yun-Nung (Vivian) Chen
Jeffrey P. Bigham
Carnegie Mellon University
Pittsburgh, PA, USA
[email protected]
National Taiwan University
Taipei, Taiwan
[email protected]
Carnegie Mellon University
Pittsburgh, PA, USA
[email protected]
[Note] This is the supplementary technical report of our
paper “Real-time On-Demand Crowd-powered Entity
Extraction,” which was published at the 5th Edition of
The Collective Intelligence Conference (CI 2017) as an
oral presentation (Huang et al. 2017). The original paper
was only 3-page long, so we decided to share extra technical
details in this report.
Sunday
flights
from New
York to
Las Vegas
Dialog ESP Game
Answer
Aggregate
Destination:
Las Vegas
User
Abstract
Figure 1: The crowd-powered entity extraction with a multiplayer Dialog ESP Game. By aggregating input answers
from all players, our approach is able to provide good quality results in seconds.
Modern dialog systems rely on accurate entity extraction
to understand user utterances. However, entity extraction is
brittle due to data scarcity, language variability, and out-ofvocabulary entities. To bridge this gap, we propose a realtime crowdsourcing solution based on the ESP game for image labeling. When multiple players agree, entities can be reliably extracted from an utterance. This approach is advantageous because it does not require training data. Further, it is
robust to unexpected input and capable of recognizing new
entities. Our approach achieves better F1-scores than that of
the automated baseline for complex queries with a reasonable response time. The proposed method is also evaluated
via Google Hangouts’ text chat and demonstrates the feasibility of real-time crowd-powered entity extraction.
more difficult to acquire the complicated conversational data
required by other alternative dialog technologies, such as
statistical dialog management (Young 2006) or state tracking (Williams et al. 2013). Second, existing entity extraction technologies are not robust enough to identify out-ofvocabulary entities. Even when labeled training data for
the targeted slot could be collected, state-of-the-art supervised learning approaches are brittle in extracting unseen
entities. (Xu and Sarikaya 2014) find that the CRF-based
entity extractor performed significantly worse when dictionary features were not used. Third, challenges are also posed
by language variability. Successful applications process diverse input languages where potential entities are unlimited.
Therefore, to robustly serve arbitrary input, dialog systems
must collect new sources of entities and update accordingly.
Research on dialog systems has focused on utilizing
the Internet resource to extract entities such as movie
names (Wang, Heck, and Hakkani-Tur 2014); Unsupervised
slot-filling approaches have also been developed in recent
years (Chen, Hakkani-Tür, and Tur 2014; Heck, HakkaniTür, and Tür 2013). However, these methods are still underdeveloped.
To address these challenges, we propose to use real-time
crowdsourcing as an entity extractor in dialog systems. To
the best of our knowledge, few previous works have attempted to use crowdsourcing to extract entities from a running conversation. (Wang et al. 2012), for example, studied
various methods to acquire natural language sentences for a
given semantic form by the crowd. (Lasecki, Kamar, and Bohus 2013) utilized crowdsourcing to collect dialog data, and
Introduction
To understand user utterances, modern dialog systems rely
heavily on entity extraction, known as the core task of slot
filling in many dialog system frameworks such as Olympus (Bohus et al. 2007). The goal of slot filling is to identify
from a running dialog different slots, which correspond to
different parameters of the user’s query. For instance, when
a user queries for nearby restaurants, key slots for location
and preferred food are required for a dialog system to retrieve the appropriate information. Thus, the main challenge
in the slot-filling task is to extract the target entity.
Dialog systems face three key challenges in entity extraction. Due to data scarcity, labeled training data, which
many existing technologies require to identify entities such
as Conditional Random Fields (CRF) (Raymond and Riccardi 2007; Xu and Sarikaya 2014) and Recurrent Neural
Networks (Mesnil et al. 2015a), are often unavailable for
the wide variety of dialog system tasks. Furthermore, it is
Copyright © 2016, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
1
a dialog1 . The ESP Game was originally proposed as a
crowdsourcing mechanism to acquire quality image labels (Von Ahn and Dabbish 2004). The original game randomly pairs two players and presents them with the same
image. Each player guesses the labels that the other player
would answer. If the players match labels, each is awarded
1000 points. Our approach replaces the image in the ESP
Game with a dialog chat log and players answer the required entity name within a short time. We also relax the
constraints of player numbers to increase game speed. As
Figure 1 shows, by aggregating input answers from all players, the Dialog ESP Game is able to provide high quality
results in seconds.
Figure 2 shows the worker’s interface. When input dialog
utterances reach the crowd-powered entity extraction component, workers are recruited from crowdsourcing platforms
such as Amazon Mechanical Turk (MTurk). The timer begins counting down when the input utterance arrives, and
the worker sees the remaining time on the top right corner
of the interface (Figure 2). When two workers match answers, a feedback notification is displayed, and the workers
earn 1000 points. When the time is up, the task automatically
closes.
To recruit crowd workers quickly, many approaches have
been used in real-time crowd-powered systems such as
VizWiz (Bigham et al. 2010) and Chorus (Lasecki et al.
2013). The quikTurkit toolkit (quikturkit.googlecode.com)
attracts workers by posting tasks and using old tasks to
queue workers. Similarly, the Retainer Model maintains a
retaining pool of workers-in-waiting, who receive a signal
when tasks become available. Prior research shows that the
Retainer Model is able to recall 75% of workers within 3
seconds (Bernstein et al. 2011). In Experiment 1, we first focus on the speed and performance of the Dialog ESP Game
itself instead of recruiting time. In Experiment 2, we propose a novel approach to recruit workers within 60 seconds
and discuss details of the end-to-end response speed.
Figure 2: The Dialog ESP Game interface is designed to
encourage quick and correct entity identification by crowd
workers. Workers are shown the complete dialog and a description of the entity they should identify.
illustrated CrowdParse, a system that uses the crowd to parse
dialogs into semantic frames. Recently, (Huang, Lasecki,
and Bigham 2015) presented a crowd-powered dialog system called Guardian that uses the crowd to extract information from input utterances. However, none of these works
conducted formal studies on crowd-powered entity extraction in real-time.
Inspired by the ESP game for image labeling (Von Ahn
and Dabbish 2004), we propose a Dialog ESP Game to encourage crowd workers to accurately and quickly perform
entity extraction. The ESP Game matches answers among
different workers to ensure label quality, and we use a timer
on the interface (Figure 2) to ensure input speed. Our method
offers three main advantages: 1) it does not require training
data; 2) it is robust to unexpected input; and 3) it is capable
of recognizing new entities. Furthermore, answers submitted from the crowd can be used as training data to bootstrap
automatic entity extraction algorithms. In this paper, we conduct experiments on a standard dialog dataset and user experiments with 10 users via Google Hangouts’ text chatting
interface. Detailed experiments demonstrate that our crowdpowered approach is robust, effective, and fast.
In sum, the contributions of our work are as follows:
Experiment 1:
Applying Dialog ESP Game
on ATIS Dataset
To evaluate the Dialog ESP Game for entity extraction, we
conducted experiments on MTurk to extract names of destination cities from a flight schedule query dialog dataset, the
Airline Travel Information System (ATIS) dataset.
1. We propose an ESP-game-based real-time crowdsourcing
approach for entity extraction in dialog systems, which
enables accurate entity extraction for a wide variety of
tasks.
2. To strive for real-time dialog systems, we present detailed
experiments to understand the trade-offs between entity
extraction accuracy and time delay.
ATIS Dataset The ATIS dataset contains a set of flight
schedule query sessions, each of which consists of a sequence of spoken queries (utterances). Each query contains
automatic speech recognized transcripts and a set of corresponding SQL queries. All queries in the data set are annotated with the query category: A, D, or X. Class A queries
are context-independent, answerable, and formed mostly in
a single sentence; however, real-world queries are more
3. We demonstrate the feasibility of real-time crowdpowered entity extraction in instant messaging applications.
Real-time Dialog ESP Game
1
The source code of worker interface and the data collected in
Experiment 2 are available at:
https://github.com/windx0303/dialogue-esp-game
We utilize real-time crowdsourcing with a multi-player Dialog ESP Game setting to extract the targeted entity from
2
complex. In the ATIS data set, 32.2% queries are contextdependent (Class D) and 24.0% of the queries are cannot
be evaluated (Class X) (Hirschman 1992). The “contextdependent” Class D queries require information from previous queries to form a complete SQL query. For instance,
in one ATIS session, the first query is “From Montreal to Las
Vegas” (Class A). The second query in the session is “Saturday,” which requires the destination and departure city name
from the first query, and is thus annotated as Class D. Class
X is of all the problematic queries, e.g., hopelessly-vague or
unanswerable.
Time
Const.
20s
Data Pre-processing & Experiment Setting For Class
A, we obtain the preprocessed data used in many slot filling works (Xu and Sarikaya 2014; Mesnil et al. 2015a;
He and Young 2003; Raymond and Riccardi 2007; Tur,
Hakkani-Tur, and Heck 2010), which contain 4,978 queries
for training, 893 queries for testing, and 491 queries for
developing. 200 queries are randomly extracted from the
developing set for our study; For Class D and X, we obtain the original training set of ATIS-3 data (Dahl et al.
1994), which contains 364 sessions and 3,235 queries. 200
Class-D queries are randomly selected from 200 distinct sessions. For each extracted query, all previous queries before
it within the same session are also obtained and displayed in
the worker’s interface (Figure 2). The same process is used
to extract 150 Class-X queries for the experiments. Note that
in this work we focus only on the toloc.city name slot
(name of destination city), which is the most frequent slot
type in ATIS. For each extracted query of Class D and X, we
define the last-mentioned destination city name of the flight
in the query history (including the extracted query) as the
gold-standard slot value.
15s
#
Player
Avg.
Resp.
Time
P
R
F1
ESP+
1st
10
7.837s
.867
.916
.891
5
11.160s
.828
.877
.852
1st
Only
10
5.590s
.713
.753
.732
5
6.924s
.730
.769
.749
ESP
Only
10
7.837s
.867
.916
.891
5
11.160s
.856
.797
.826
ESP+
1st
10
8.129s
.837
.893
.864
5
10.628s
.799
.798
.798
1st
Only
10
5.895s
.739
.764
.751
5
7.136s
.729
.726
.727
ESP
Only
10
8.129s
.860
.865
.863
5
10.628s
.872
.637
.736
Aggregate
Table 1: Dialog ESP Game results in Class A given different
settings of number of players, time constraint (Time Const.),
and the method to aggregate input answers.
could result in different response speed and output quality.
Note that if there are not any input answers, the methods
above will wait until the time constraint and return an empty
label. In the actual experiments, 5 Dialog ESP Games for 5
different Class-A queries are aggregated in one task, with an
extra scripted game at the beginning as a tutorial. When the
first game ends, the timer of the second ESP game starts and
a browser alert informs the worker. All experiments are run
on MTurk; 800 Human Intelligence Tasks (HITs) are posted,
and 588 unique workers participate in this study.
Table 1 shows the results on Class A queries. With 10
players and a 20-second time constraint, the Dialog ESP
Game achieves a best F1-score of 0.891 by the “ESP+1st”
setting, and achieves the fastest average response time of
5.590 seconds by the “1st” setting. The ESP+1st setting
achieves the best F1-score, and the 1st Only setting has
the shortest response time. In most cases, tightening the
time constraint provides a faster response but reduces output quality.
We also analyze the relations among worker numbers,
performance, and response time. First, Figure 4 shows output quality with respect to answer’s input order. On average, earlier input answers are of better quality, unless 10
or more players participate in the game. However, with 10
players, almost all ESP games have at least one matched answer pair so that the ith answer is not solely used. Therefore, for the following experiments, we set i as 1. Second,
in Figure 3(a) we observe the relations between the number of players and average response time. Adding players
reduces the average response time for all settings. Third, the
relations between number of players and output quality are
also analyzed. Figure 3(b) shows that the F1-scores increase
when adding more players, even with the “1st Only” setting.
Understanding Accuracy and Speed Trade-offs
In order to design an effective crowd-powered real-time entity extraction system, it is crucial to understand trade-offs
between accuracy and speed. These trade-offs correspond to
the three main variables in our system: the number of players recruited to answer each query in the Dialog ESP Game,
the time constraint that each player has to answer a query,
and the method to aggregate input answers. We have 3
ways to aggregate the input answers from the ESP game:
• ESP Only: Return the first matched answer. If no answers
match within the given time, return an empty label.
• ith Only: Return the ith input answer (i = 1, 2, ...). For
example, i = 1 means to return the first input answer.
• ESP + ith: Return the first matched answers of the ESP
game. If no answers match within the given time, return
the ith answer.
We recruit 10 players for each ESP game, and randomly
select player results to simulate the conditions of various
player numbers. All results reported in Experiment 1 are the
averages of 20 rounds of this random-pick simulation process. After empirically testing the interface, we run two sets
of studies with time constraints set at 20 and 15 seconds,
respectively. Different methods to aggregate input answers
3
16
12
8
ESP + 1st (20 sec)
ESP + 1st (15 sec)
1st (20 sec)
1st (15 sec)
1.00
0.90
0.80
0.80
0
2
4
6
8
# Player
(a)
0.60
0
10
10 Players
9 Player
8 players
7 Players
6 Players
5 Players
1st Only
(20 sec)
0.70
0.70
0.60
4
ESP + 1st
(20 sec)
0.90
F1-score
ESP + First (20 sec)
ESP + First (15 sec)
First (20 sec)
First (15 sec)
F1-score
Avg. Resp. Time (sec)
20
2
4
6
8
# Player
(b)
10
5
(c)
6
7
8
9
10
11
12
Avg. Response Time (sec)
Figure 3: Trade-off curves between accuracy, average response time and number of players.
10 Players
4 Players
2 Players
F1-score
0.80
0.70
language understanding model, implemented by an recurrent neural network (RNN) with GRU cells (Mesnil et al.
2015b; Chen et al. 2016), on the crowd-labelled or expertlabelled 200 ATIS conversations used in our study, and use
the standard ATIS Class-A testing set to compare the performance in terms of the number of training samples. Figure 5 illustrates the performance curves of models trained
on crowd-labelled and expert-labelled data. The result shows
that crowd-generated labels can be used to train machinelearning models for bootstrapping in new domains, and
achieve comparable performance to expert-labelled data.
0.60
0
1
2
3
4
Input Order (i)
5
6
7
Figure 4: F1-score of the “ith Only” setting. Earlier input
answers are generally of better quality (unless #players >
10, where almost all ESP games have at least one matched
answer and the ith answer might not be solely used.)
80
Evaluation on Complex Queries
70
Based on the study above, for Class D and X queries, we use
the Dialog ESP Game of 10 players with “ESP+1st” and “1st
Only” settings to measure the best F1-score and speed. The
time constraint is set to 20 seconds. The experiments are run
on MTurk and all settings are identical as the previous section. 76 distinct workers participate in Class D experiments,
and 68 distinct workers participate in Class X experiments.
Experimental results are shown in Table 2. An automated
CRF model is implemented as a baseline.2 The CRF model
is trained on the Class-A training set mentioned above by using neighbor words (window size = 2) and POS tag features.
The CRF model is decoded and timed on a laptop with Intel i5-4200U CPU (@1.60GHz) and 8GB RAM. As a result,
the proposed crowd-powered approach largely outperforms
the CRF baseline in terms of F1-score on both Class D and
X queries . Although the CRF approach is well-developed
on Class A data, it is not effective on the remaining data.
Surprisingly, we find similar average response times in
each query category. Note that the text length is different for
each category: the average token number of Class-A queries
is 11.47, of Class-D queries (including the query history)
is 48.64, and of Class-X queries is 67.72. Studies showed
that eyes’ warm-up time (Inhoff and Rayner 1986) and word
frequency influence speed of text comprehension (Rayner
and Duffy 1986; Healy 1976). These factors might reduce
the effect of text length to the reading speed of crowd works.
We also conduct an error analysis on the result of
“ESP+1st” setting, which achieves our best F1-score. The
distribution of error types are shown in Table 3. The
“fromloc.city name” type indicates that the crowd extracts the departure city, rather than destination city; In “In-
60
F1 (%)
50
40
30
Crowd-Labelled
20
Expert-Labelled
10
0
0
20
40
60
80
100
120
140
160
180
200
#Training Samples for RNN
Figure 5: The performance of the extracted destination city
predicted by the RNN-GRU trained on crowd-labelled and
expert-labelled data. Crowd-labelled labels can be used to
train machine-learning models and achieve comparable performance to expert-labelled data.
Finally, Figure 3(c) demonstrates the trade-offs between performance and speed. For a fixed number of players, different
input aggregate methods have different response times and
F1-scores. The ESP game requires more time for input answer matching, but in return output quality increases.
Bootstrapping Automated System Performance
Once our crowd-powered system starts extracting entities,
the collected annotations can serve as training data for
the automated system. In order to see whether the crowdannotated data is good enough for training a machine learning model, and how many of text instances are required to
achieve reasonable performance, we train a state-of-the-art
2
4
Implemented with CRF++: http://taku910.github.io/crfpp/
Query Category
Methods
CRF Baseline
1st Only
ESP + 1st
Class D (context-dependent)
Resp. Time
P
R
F1
0.043s
.776 .307 .440
5.460s
.658 .641 .649
7.118s
.814 .797 .805
Class X (unevaluable)
Resp. Time
P
R
F1
0.061s
.636 .285 .393
6.342s
.563 .577 .570
8.301s
.654 .675 .664
Class A (context-independent)
Resp. Time
P
R
F1
0.019s
.985 .987 .986
5.590s
.713 .753 .732
7.837s
.867 .916 .891
Table 2: Result for Class D, X and A. Crowd-powered entity extraction outperforms the CRF baseline in terms of F1-score on
both Class D and X queries. Although the CRF baseline is well-developed on Class A, it is not effective on complex queries.
Error Type
fromloc.city name
False Negative
Incorrect City
Correct City & Soft Match
False Positive
Class D
Class X
39.53%
18.60%
16.28%
16.28%
9.30%
16.67%
26.67%
18.33%
5.00%
33.33%
recruited workers and log the timestamps of following activities: 1) users’ and workers’ keyboard typing, 2) workers’
task arrival, and 3) the workers’ answer submissions.
To recruit crowd workers, we introduce fleeting task, a recruiting practice inspired by quikturkit (Bigham et al. 2010).
This approach achieves low latency by posting hundreds of
short lifetime tasks, which increases task visibility. Its short
lifetime (e.g., 60 seconds) encourages workers to complete
tasks quickly. A core benefit of the fleeting task approach is
its ease in implementation: the method bypasses the common practices of pre-recruiting workers and maintaining
a waiting pool (Bigham et al. 2010; Lasecki et al. 2013;
Bernstein et al. 2011). In a system deployed at scale, a retainer or push model is likely to work as well.
Class A
40.00%
0.00%
8.00%
12.00%
40.00%
Table 3: Error Analysis for Class D, X and A.
correct City” type, the crowd extracts an incorrect city from
the query history (but not the departure city); “Correct City
& Soft Match” type means the extracted city name is semantically correct but does not match the gold-standard city
name (e.g., “Washington” and “Washington DC”). From the
error analysis, we conclude two directions to improve performance: 1) treat the cases of absent slot more carefully,
and 2) use domain knowledge if available. First, 28% of errors in Class D and 50% in Class X occur when either the
gold-standard label or the predicted label does not exist. It
suggests that a more reliable step to recognize the existence
of the targeted entity might be required. Second, 16.28% of
Class-D queries and 5% of Class-X queries are of the “Soft
Match” cases. By introducing domain knowledge like a list
of city names, a post-processor that finds the most similar
city name of the predicted label can fix this type of error.
User Experiment Setup
We conduct lab-based user experiments to evaluate the proposed technology on extracting “food” entities. Ten Google
Hangouts users enter our lab with their own laptops. We first
ask them to arbitrarily create a list 9 foods, 3 drinks, and 3
countries based on their own preferences. Then we explain
the purpose of the experiments, and introduce five scenarios
of using instant messaging:
1. Eat: You discuss with your friend about what to eat later.
2. Drink: You discuss with an employee a coffee place, bar,
or restaurant to order something to drink.
Experiments 2:
User Experiment via a Real-world Instant
Messaging Interface
3. Cook: You plan to cook later. You discuss the details with
your friend who knows how to cook.
To examine the feasibility of real-time crowd-powered entity
extraction in an actual system, we conduct lab-based user
experiments via Google Hangouts’ instant messaging interface. Our proposed method has a task completion time of
5-8 seconds, per Experiment 1. In this section, we demonstrate our approach is robust and fast enough to support a
real-world instant messaging application, where the average
time gap between conversational turns is 24 seconds (Isaacs
et al. 2002).
4. Chat: You are chatting with your friend.
System Implementation
3. Mentioning: Naturally converse without asking or answering any specific questions.
5. No Food: You are chatting with your friend. You do not
mention food. Instead, you mention a country name.
We also list three types of conversational acts which could
emerge in each scenario:
1. Question: Ask a question.
2. Answer: Answer a question that could be asked under the
current scenario.
We implemented a Google Hangouts chatbot by using the
Hangupsbot3 framework. Users are able to send text chats
to our chatbot via Google Hangouts. The chatbot recruits
crowd workers on MTurk in real-time to perform the Dialog
ESP Game task upon receiving the chat. Figure 6 shows the
overview of our system. We record all answers submitted by
3
Using their laptops, users send one text chat for each combination of [scenario, conversational act] to our chatbot, i.e.,
15 chats in total. In the Eat, Cook, and Chat scenarios, users
must mention one of the foods they listed earlier; in the
Drink scenario, they must mention one of the drinks they
listed. In the No Food scenario, users must mention one of
https://github.com/hangoutsbot/hangoutsbot
5
Fleeting Task Method
1st Response
1st Worker Arrives
Avg User Typing Time = 27.05s
1st Matched Response
37.14 sec
30.83 sec
40.95 sec
0.5
0.3
0.1
-0.1
-0.3
-0.5
User
0
5
Post 120
fleeting tasks
10
15
20
25
30
35
40
45
Tasks become available to workers
Chatbot
Task Posting Time
Worker Searching Time
Task Routing Time
Task Completion Time
Figure 6: Timeline of the Real-time Crowd-powered Entity Extraction System. On average, the first worker takes 30.83 seconds
to reach, the first answer is received at 37.14 seconds, and the first matched answer occurs at 40.95 seconds. A user on average
spends 27.05 seconds to type a chat line, i.e., the perceived response time to users falls within 10-14 seconds.
Acc (%)
Response Time (s)
Mean (Stdev)
1st Only
77.33%
37.14 (14.70)
ESP Only
81.33%
40.95 (13.56)
ESP + 1st
84.00%
40.95 (13.56)
1st Worker Reached Time (s)
30.83 (16.86)
User Type Time (s)
27.05 (25.28)
1st
Entity
Type
Conv.
Act
Table 4: Result of User Experiment. A trade-off between
time and output quality can be observed.
Acc.
(%)
Avg.
Time(s)
Acc.
(%)
36.64
70.00%
40.19
78.89%
Drink
37.43
80.00%
41.37
83.33%
None
38.33
96.67%
42.83
100.00%
Question
34.26
82.00%
37.94
90.00%
Answer
39.90
68.00%
43.88
78.00%
Mention
37.26
82.00%
41.04
84.00%
37.14
77.33%
40.95
84.00%
Food5
Avg.
the countries they listed, and no food names can be mentioned. In total, we collect 150 chat inputs from 10 user experiments. Correspondingly, instructions on the workers’ interface (Figure 2) is modified as “What is the food name
in this dialog?”, and the explanation of food name is modified as “Food name. The full name of the food. Including
any drinks or beverages.” In the experiments, our chatbot
post 120 HITs with a lifetime of 60 seconds to MTurk upon
receiving a text chat. The price of each HIT is $0.1. We use
the interface shown in Figure 2 with a time constraint of 20
seconds.
Table 5: Results of user experiment for each scenario and
conversational act.
knowledge-base, our crowdsourcing approach achieves an
accuracy of 78.89% in extracting food entities and 83.33%
in extracting drink entities. Despite the significant variety of
the input entities6 , our approach extracts most entities correctly. Furthermore, our method is effective in identifying
the absence of entities; Table 5 also shows the robustness
of the proposed method under various linguistic conditions.
The “ESP+1st” setting achieves accuracies of 90.00% in extracting entities from questions, 78.00% in extracting from
answers, and 84.00% in extracting from regular conversations. Qualitatively, our approach can handle complex input,
such as strange restaurant names and beverage names, which
are essentially confusing for automated approaches. For example, “Have you ever tried bibimbap at Green pepper?”
and “I usually have Magic Hat #9”, where Green pepper
and Magic Hat #9 are names of a restaurant and beverage,
respectively.
Experimental Results
Results are shown in Table 4. The “ESP+1st” setting
achieves the best accuracy of 84% with an average response
time of 40.95 seconds. The “1st Only” setting has the shortest average response time of 37.14 seconds with an accuracy
of 77.33%.4 A trade-off between time and output quality can
be observed.This trade-off is similar to the results of Experiment 1 (shown in Figure 3(c)). On average, 14.45 MTurk
workers participated in each trial and submitted 33.81 answers.
Robustness in Out-of-Vocabulary Entities & Language
Variability The results over each entity type are shown
in Table 5. Without using any training data or pre-defined
4
5
ESP + 1st
Avg.
Time(s)
6
The food entities arbitrarily created by our users are quite diverse: From a generic category (e.g., Thai food) to a specific entry (e.g., Magic Hat #9), and from a simple food (e.g., cherry) to
a complex food (e.g., sausage muffin with egg). The list covers
the food of many other countries (e.g., Okonomiyaki, Bibimbap,
Samosa.)
We only consider the answers submitted within 60 seconds.
Including the results from Food, Cook, and Chat scenarios.
6
Error Analysis Table 6 shows the errors in the user experiments (“ESP+1st” setting). 45.83% of errors are caused by
absence of answers, mainly due to the task routing latency
of the MTurk platform. We discuss this in more detail below.
37.50% of errors are due to various system problems such
as the string encoding issues. More interestingly, 12.50% of
incorrect answers are sub-spans of the correct answers. For
instance, the crowd extracts “rice” for “stew pork over rice”,
and “tea” for “bubble tea”. This type of error is similar to
the “Soft Match” error in Experiment 1. Finally, 4.17% of
errors are caused by user typos (e.g., latter for latte), which
the crowd tends to exclude in their answers.
knowledge can help to fix this type of errors. However, unlike automated technology, we do not have a generic method
to update human workers with new knowledge. Thus, our
next step is to incorporate automated components. It is easy
to replace some workers with automated annotators in our
multi-player ESP Game. Despite fragility in extracting unseen entities, automated approaches are robust in identifying known entities and can be easily updated if new data
is collected. We will develop a hybrid approach, which we
believe will be robust in unexpected input and easily incorporate new knowledge.
Conclusion and Future Work
Error Type
No Answers Received
System Problem
Substring of a Multi-token Entity
Typo
%
We have explored using real-time crowdsourcing to extract
entities for dialog systems. By using an ESP Game setting,
our approach is absolute 36.5% and 27.1% better than the
CRF baseline in terms of F1-score for Class D and X queries
in the ATIS dataset, respectively. The timing cost is about 8
seconds, which is slower than machines but still reasonable
given the large gains in accuracy. The proposed method also
has been evaluated via Google Hangouts’ text chat with 10
users. The results demonstrate the robustness and feasibility
of our approach in real-world systems. In the future, we will
generalize our approach by adding automated components,
and also explore the possibility of using audio input.
45.83%
37.50%
12.50%
4.17%
Table 6: Error Analysis for User Experiment.
Response Speed Table 4 shows the average response time
in the user experiment. On average, the first worker takes
30.83 seconds to reach to our Dialog ESP Game, the first
answer is received at 37.14 seconds, and the first matched
answer occurs at 40.95 seconds. For comparison, we illustrate the timeline of our system in Figure 6. In the user experiments, a user on average spends 27.05 seconds to type
a chat line. If we align the user typing time along with
the system timeline, the theoretical perceived response time
to users falls within 10-14 seconds, while the average response time in instant messaging is 24 seconds (Isaacs et
al. 2002). (Baron 2010) reports that 24.5% of instant messages get responses within 11-30 seconds, and 8.2% of messages have even longer response times. The proposed technology proves to be fast enough to support instant messaging applications. The main bottleneck of the end-to-end response speed is the task routing time in Figure 6, which
approximately ranges from 5-40 seconds and changes over
time. The task routing time also causes the major errors in
Table 6. The task lifetime begins when a task reaches the
MTurk server instead of when it becomes visible to workers. When the task routing time is longer than a task’s lifetime, the task could expire before it is selected by workers. Because MTurk requesters can not effectively reduce
the task routing time, pre-recruiting and queuing workers
seems inevitable for applications which require a response
time sharply shorter than 30 seconds.
Appendix: List of Food and Drinks Used in the
Experiment 2
The followings are the lists of 9 food and 3 drinks created
by 10 participants in our user experiment.
Food
1. spaghetti, burger, vindaloo lamb, makhani chicken, kimchee, wheat bread pizza, cornish pasty, mushroom soup
2. burger, french fries, scallion cake, okonomiyaki, oyakodon, gyudon, fried rice, wings, salad
3. Stinky Tofu, Acai Berry Bowl, Tuna Onigiri, Rice Burger,
Seared Salmon, Milkfish Soup, Mapo Tofu, Beef Pho,
Scallion Pancake
4. pizza, fried rice, waffle, alcohol drink, chocolate pie,
cookie, dimsum, burger, milk shake
5. Pho, BBQ, Thai food, beef noodles, steak, Tomato soup,
Spicy hot pot, Soup dumplings, Ramen
6. chocolate, donut, cheesecake, pad thai, seafood pancake,
fish fillets in hot chili, hot pot, bibimbap, japchae
7. chocolate, pancakes, strawberries, fried fish, fried
chicken, sausages, gulaab jamun, paneer tika, samosa
8. Dumplings, noodle, stew pork over rice, Sandwich, pasta,
hot pot, Potato slices with green peppers, Chinese BBQ,
pancakes
9. stinky tofu, stew pork over rice, yakitori, baked cinnamon
apple, apple pie, stew pork with potato and apple, teppanyaki, okonomiyaki, crab hotpot
10. hot pot, cherry, Chinese cabbage, Pumpkin risotto,
Tomato risotto, Boeuf Bourguignon, stinky tofu, sausage
muffin with egg (McDonald), eggplant with basil
Discussion
Incorporating domain-specific knowledge is a major obstacle in generalization of crowdsourcing technologies (Huang,
Lasecki, and Bigham 2015). We think that automation helps
resolve this challenge. One most common errors in our system are the soft match, where the crowd extracts a sub-string
of the target entity instead of the complete string. Domain
7
Drink
[Heck, Hakkani-Tür, and Tür 2013] Heck, L. P.; HakkaniTür, D.; and Tür, G. 2013. Leveraging knowledge graphs
for web-scale unsupervised semantic parsing. In INTERSPEECH, 1594–1598.
[Hirschman 1992] Hirschman, L. 1992. Multi-site data collection for a spoken language corpus. In Proceedings of the
Workshop on Speech and Natural Language, HLT ’91, 7–
14. Stroudsburg, PA, USA: Association for Computational
Linguistics.
[Huang et al. 2017] Huang, T.-H. K.; Chen, Y.-N.; Bigham,
J. P.; et al. 2017. Real-time on-demand crowd-powered entity extraction. In In Proceedings of the 5th Edition Of The
Collective Intelligence Conference (CI 2017, oral presentation).
[Huang, Lasecki, and Bigham 2015] Huang,
T.-H.
K.;
Lasecki, W. S.; and Bigham, J. P. 2015. Guardian: A
crowd-powered spoken dialog system for web apis. In
HCOMP.
[Inhoff and Rayner 1986] Inhoff, A. W., and Rayner, K.
1986. Parafoveal word processing during eye fixations in
reading: Effects of word frequency. Perception & Psychophysics 40(6):431–439.
[Isaacs et al. 2002] Isaacs, E.; Walendowski, A.; Whittaker,
S.; Schiano, D. J.; and Kamm, C. 2002. The character,
functions, and styles of instant messaging in the workplace.
In Proceedings of the 2002 ACM CSCW, 11–20. ACM.
[Lasecki et al. 2013] Lasecki, W. S.; Wesley, R.; Nichols, J.;
Kulkarni, A.; Allen, J. F.; and Bigham, J. P. 2013. Chorus: A
crowd-powered conversational assistant. In UIST ’13, UIST
’13, 151–162. New York, NY, USA: ACM.
[Lasecki, Kamar, and Bohus 2013] Lasecki, W. S.; Kamar,
E.; and Bohus, D. 2013. Conversations in the crowd: Collecting data for task-oriented dialog learning. In HCOMP.
[Mesnil et al. 2015a] Mesnil, G.; Dauphin, Y.; Yao, K.; Bengio, Y.; Deng, L.; Hakkani-Tur, D.; He, X.; Heck, L.; Tur,
G.; Yu, D.; et al. 2015a. Using recurrent neural networks
for slot filling in spoken language understanding. Audio,
Speech, and Language Processing, IEEE/ACM Transactions
on 23(3):530–539.
[Mesnil et al. 2015b] Mesnil, G.; Dauphin, Y.; Yao, K.; Bengio, Y.; Deng, L.; Hakkani-Tur, D.; He, X.; Heck, L.; Tur,
G.; Yu, D.; et al. 2015b. Using recurrent neural networks for
slot filling in spoken language understanding. IEEE/ACM
Transactions on Audio, Speech, and Language Processing
23(3):530–539.
[Raymond and Riccardi 2007] Raymond, C., and Riccardi,
G. 2007. Generative and discriminative algorithms for
spoken language understanding. In INTERSPEECH, 1605–
1608.
[Rayner and Duffy 1986] Rayner, K., and Duffy, S. A. 1986.
Lexical complexity and fixation times in reading: Effects
of word frequency, verb complexity, and lexical ambiguity.
Memory & Cognition 14(3):191–201.
[Tur, Hakkani-Tur, and Heck 2010] Tur, G.; Hakkani-Tur,
D.; and Heck, L. 2010. What is left to be understood in
atis? In SLT, 2010 IEEE, 19–24. IEEE.
1. tea, coke, latte
2. green tea latte, bubble tea, root beer
3. medium latte with non-fat milk, green Tea Latte,
Soymilk
4. water, pepsi, tea
5. Latte with nonfat milk, Magic hat #9, Old fashion
6. vanilla latte, strawberry smoothie, iced tea
7. coffee, milk shake, beer
8. Mocha coffee, beers, orange juice
9. caramel frappuccino, caramel macchiato, coffee with coconut milk
10. ice tea, macha, apple juice
References
[Baron 2010] Baron, N. S. 2010. Discourse structures in instant messaging: The case of utterance breaks. Language@
Internet 7(4):1–32.
[Bernstein et al. 2011] Bernstein, M. S.; Brandt, J.; Miller,
R. C.; and Karger, D. R. 2011. Crowds in two seconds:
Enabling realtime crowd-powered interfaces. In UIST, 33–
42. ACM.
[Bigham et al. 2010] Bigham, J. P.; Jayant, C.; Ji, H.; Little, G.; Miller, A.; Miller, R. C.; Miller, R.; Tatarowicz, A.;
White, B.; White, S.; et al. 2010. Vizwiz: nearly real-time
answers to visual questions. In UIST, 333–342. ACM.
[Bohus et al. 2007] Bohus, D.; Raux, A.; Harris, T. K.; Eskenazi, M.; and Rudnicky, A. I. 2007. Olympus: an opensource framework for conversational spoken language interface research. In Proceedings of the workshop on bridging
the gap: Academic and industrial research in dialog technologies, 32–39. Association for Computational Linguistics.
[Chen et al. 2016] Chen, Y.-N.; Hakkani-Tur, D.; Tur, G.;
Celikyilmaz, A.; Gao, J.; and Deng, L. 2016. Knowledge as
a teacher: Knowledge-guided structural attention networks.
arXiv preprint arXiv:1609.03286.
[Chen, Hakkani-Tür, and Tur 2014] Chen, Y.-N.; HakkaniTür, D.; and Tur, G. 2014. Deriving local relational surface forms from dependency-based entity embeddings for
unsupervised spoken language understanding. Proceedings
of SLT.
[Dahl et al. 1994] Dahl, D. A.; Bates, M.; Brown, M.; Fisher,
W.; Hunicke-Smith, K.; Pallett, D.; Pao, C.; Rudnicky, A.;
and Shriberg, E. 1994. Expanding the scope of the atis
task: The atis-3 corpus. In HLT, 43–48. Association for
Computational Linguistics.
[He and Young 2003] He, Y., and Young, S. 2003. A datadriven spoken language understanding system. In ASRU’03,
583–588. IEEE.
[Healy 1976] Healy, A. F. 1976. Detection errors on the word
the: Evidence for reading units larger than letters. Journal
of Experimental Psychology: Human Perception and Performance 2(2):235.
8
[Von Ahn and Dabbish 2004] Von Ahn, L., and Dabbish, L.
2004. Labeling images with a computer game. In Proceedings of the SIGCHI conference on Human factors in computing systems, 319–326. ACM.
[Wang et al. 2012] Wang, W. Y.; Bohus, D.; Kamar, E.; and
Horvitz, E. 2012. Crowdsourcing the acquisition of natural
language corpora: Methods and observations. In SLT 2012,
73–78. IEEE.
[Wang, Heck, and Hakkani-Tur 2014] Wang, L.; Heck, L.;
and Hakkani-Tur, D. 2014. Leveraging semantic web search
and browse sessions for multi-turn spoken dialog systems.
In ICASSP 2014, 4082–4086. IEEE.
[Williams et al. 2013] Williams, J.; Raux, A.; Ramachandran, D.; and Black, A. 2013. The dialog state tracking
challenge. In Proceedings of the SIGDIAL 2013 Conference,
404–413.
[Xu and Sarikaya 2014] Xu, P., and Sarikaya, R. 2014. Targeted feature dropout for robust slot filling in natural language understanding. In Fifteenth Annual Conference of the
International Speech Communication Association.
[Young 2006] Young, S. 2006. Using pomdps for dialog
management. In SLT, 8–13.
9
| 2cs.AI
|
Maximizing the Strong Triadic Closure in Split Graphs and
Proper Interval Graphs
Athanasios Konstantinidis∗
Charis Papadopoulos†
arXiv:1609.09433v2 [cs.DS] 7 Oct 2016
Abstract
In social networks the Strong Triadic Closure is an assignment of the edges
with strong or weak labels such that any two vertices that have a common neighbor
with a strong edge are adjacent. The problem of maximizing the number of strong edges
that satisfy the strong triadic closure was recently shown to be NP-complete for general
graphs. Here we initiate the study of graph classes for which the problem is solvable. We
show that the problem admits a polynomial-time algorithm for two unrelated classes of
graphs: proper interval graphs and trivially-perfect graphs. To complement our result,
we show that the problem remains NP-complete on split graphs, and consequently also
on chordal graphs. Thus we contribute to define the first border between graph classes
on which the problem is polynomially solvable and on which it remains NP-complete.
1
Introduction
Predicting the behavior of a network is an important concept in the field of social networks
[7]. Understanding the strength and nature of social relationships has found an increasing
usefulness in the last years due to the explosive growth of social networks (see e.g., [2]).
Towards such a direction the Strong Triadic Closure principle enables to understand
the structural properties of the underlying graph: it is not possible for two individuals to have
a strong relationship with a common friend and not know each other [11]. Such a principle
stipulates that if two people in a social network have a “strong friend” in common, then there
is an increased likelihood that they will become friends themselves at some point in the future.
Satisfying the Strong Triadic Closure is to characterize the edges of the underlying
graph into weak and strong such that any two vertices that have a strong neighbour in
common are adjacent. Since users interact and actively engage in social networks by creating
strong relationships, it is natural to consider the MaxSTC problem: maximize the number
of strong edges that satisfy the Strong Triadic Closure. The problem has been shown
to be NP-complete for general graphs whereas its dual problem of minimizing the number of
weak edges admits a constant factor approximation ratio [25].
In this work we initiate the computational complexity study of the MaxSTC problem
in important classes of graphs. If the input graph is a P3 -free graph (i.e., a graph having
no induced path on three vertices which is equivalent with a graph that consists of vertexdisjoint union of cliques) then there is a trivial solution by labeling strong all the edges. Such
an observation might falsely lead into a graph modification problem, known as Cluster
Deletion problem (see e.g., [3, 12]), in which we want to remove the minimum number of
edges that correspond to the weak edges, such that the resulting graph does not contain a
P3 as an induced subgraph. More precisely the obvious reduction would consist in labeling
the deleted edges in the instance of Cluster Deletion as weak, and the remaining ones as
∗
†
Department of Mathematics, University of Ioannina, Greece. E-mail: [email protected]
Department of Mathematics, University of Ioannina, Greece. E-mail: [email protected]
1
strong. However, this reduction fails to be correct due to the fact that the graph obtained
by deleting the weak edges in an optimal solution of MaxSTC may contain an induced P3 ,
so long as those three vertices induce a triangle in the original graph (prior to deleting the
weak edges).
To the best of our knowledge, no previous results were known prior to our work when
restricting the input graph for the MaxSTC problem. It is not difficult to see that for bipartite graphs the MaxSTC problem has a simple polynomial-time solution by considering
a maximum matching that represent the strong edges [20]. It is well-known that a maximum
matching of a graph corresponds to a maximum independent set of its line graph that represents the adjacencies between the edges [8]. As previously noted, for general graphs it is
not necessarily the case that a maximum matching corresponds to the optimal solution for
MaxSTC. Here we show a similar characterization for MaxSTC by considering the adjacencies between the edges of a graph that participate in induced P3 ’s. Such a characterization
allows us to exhibit structural properties towards the computation of an optimal solution of
MaxSTC.
Due to the nature of the P3 existence that enforce the labeling of weak edges, there
is an interesting connection to problems related to the square root of a graph; a graph H
is a square root of a graph G and G is the square of H if two vertices are adjacent in G
whenever they are at distance one or two in H. Any graph does not have a square root (for
example consider a simple path), but every graph contains a subgraph that has a square
root. Although it is NP-complete to determine if a given chordal graph has a square root
[19], there are polynomial-time algorithms when the input is restricted to bipartite graphs
[18], or proper interval graphs [19], or trivially-perfect graphs [22]. The relationship between
MaxSTC and to that of determining square roots can be seen as follows. In the MaxSTC
problem we are given a graph G and we want to select the maximum possible number of
edges, at most one from each induced P3 in G. Thus we need to find the largest subgraph
(in terms of the number of its edges) H of G such that the square of H is a subgraph of G.
However previous results related to square roots were concerned with deciding if the whole
graph has a square root and there are no such equivalent formulations related to the largest
square root.
Our main motivation is to understand the complexity of the problem on subclasses of
chordal graphs, since the class of chordal graphs (i.e., graphs having no chordless cycle of
length at least four) finds important applications in both theoretical and practical areas
related to social networks [1, 16, 23]. More precisely two famous properties can be found in
social networks. For most known social and biological networks their diameter, that is, the
length of the longest shortest path between any two vertices of a graph, is known to be a small
constant [14]. On the other hand it has been shown that the most prominent social network
subgraphs are cliques, whereas highly infrequent induced subgraphs are cycles of length four
[26]. Thus it is evident that subclasses of chordal graphs are close related to such networks,
since they have rather small diameter (e.g., split graphs or trivially-perfect graphs) and are
characterized by the absence of chordless cycles (e.g., proper interval graphs). Towards such
a direction we show that MaxSTC is NP-complete on split graphs and consequently also
on chordal graphs. On the positive side, we present the first polynomial-time algorithm
for computing MaxSTC on proper interval graphs. Proper interval graphs, also known
as unit interval graphs or indifference graphs, form a subclass of interval graphs and they
are unrelated to split graphs [24]. By our result they form the first graph class, other
than bipartite graphs, for which MaxSTC is shown to be polynomial time solvable. In
order to obtain our algorithm, we take advantage of their clique path and apply a dynamic
programming on subproblems defined by passing the clique path in its natural ordering.
Furthermore by considering the equivalent transformation of the problem mentioned earlier,
2
Figure 1: A split graph G is shown to the left side. The right side depicts a solution for
MaxSTC on G where the weak edges are exactly the edges of G that are not shown.
we show that MaxSTC admits a simple polynomial-time solution on trivially-perfect graphs.
Thus we contribute to define the first borderline between graph classes on which the problem
is polynomially solvable and on which it remains NP-complete.
2
Preliminaries
All graphs considered here are simple and undirected. A graph is denoted by G = (V, E)
with vertex set V and edge set E. We use the convention that n = |V | and m = |E|. The
neighborhood of a vertex v of G is N (v) =S{x | vx ∈ E} and the closed neighborhood of v is
N [v] = N (v) ∪ {v}. For S ⊆ V , N (S) = v∈S N (v) \ S and N [S] = N (S) ∪ S. A graph H
is a subgraph of G if V (H) ⊆ V (G) and E(H) ⊆ E(G). For X ⊆ V (G), the subgraph of
G induced by X, G[X], has vertex set X, and for each vertex pair u, v from X, uv is an
edge of G[X] if and only if u 6= v and uv is an edge of G. For R ⊆ E(G), G \ R denotes
the graph (V (G), E(G) \ R), that is a subgraph of G and for S ⊆ V (G), G − S denotes the
graph G[V (G) − S], that is an induced subgraph of G. Two adjacent vertices u and v are
called twins if N [u] = N [v].
A clique of G is a set of pairwise adjacent vertices of G, and a maximal clique of G is a
clique of G that is not properly contained in any clique of G. An independent set of G is a set
of pairwise non-adjacent vertices of G. For k ≥ 2, the chordless path on k vertices is denoted
by Pk and the chordless cycle on k vertices is denoted by Ck . For an induced path Pk , the
vertices of degree one are called endvertices. A vertex v is universal in G if N [v] = V (G)
and v is isolated if N (v) = ∅. For two vertices u and v we say that u sees v if {u, v} ∈ E(G);
otherwise, we say that u misses v. We extend this notion to vertex sets: a set A sees (resp.,
misses) a vertex set B if every vertex of A is adjacent (resp., non-adjacent) to every vertex
of B. We say that two edges are non-adjacent if they have no common endpoint; otherwise
we call them adjacent edges.
Strong Triadic Closure. Given a graph G = (V, E), a strong-weak labeling on the edges
of G is a function λ that assigns to each edge of E(G) one of the labels strong or weak; i.e.,
λ : E(G) → {strong, weak}. An edge that is labeled strong (resp., weak) is simple called
strong (resp. weak). The strong triadic closure of a graph G is a strong-weak labeling λ such
that for any two strong edges {u, v} and {v, w} there is a (weak or strong) edge {u, w}. In
other words, in a strong triadic closure there are no pair of strong edges {u, v} and {v, w}
such that {u, w} ∈
/ E(G).
The problem of computing the maximum strong triadic closure, denoted by MaxSTC,
is to find a strong-weak labeling on the edges of E(G) that satisfies the strong triadic closure
and has the maximum number of strong edges. Note that its dual problem asks for the
minimum number of weak edges. Here we focus on maximizing the number of strong edges
in a strong triadic closure.
Let G be a strong-weak labeled graph. We denote by (ES , EW ) the partition of E(G)
into strong edges ES and weak edges EW . The graph spanned by ES is the graph G \ EW .
3
For a vertex v ∈ V (G) we say that the strong neighbour of v is the other endpoint of a strong
edge incident to v. We denote by NS (v) ⊆ N (v) the strong neighbours of v. Similarly we say
that a vertex u is strongly adjacent to v if u is adjacent to v and the edge {u, v} is strong.
Observation 2.1. Let G be a strong-weak labeled graph. Then G \ EW satisfies the strong
triadic closure if and only if for every P3 in G \ EW , the vertices of the P3 induce a K3 in
G.
Proof. Observe that G \ EW is the graph spanned by the strong edges. If for two strong
edges {u, v} and {v, w}, {u, w} ∈
/ E(G \ EW ) then {u, w} is an edge in G and, thus, u, v, w
induce a K3 in G. On the other hand notice that any two strong edges of G \ EW are either
non-adjacent or share a common vertex. If they share a common vertex then the vertices
must induce a K3 in G.
Therefore in the MaxSTC problem we want to minimize the number of the removal
(weak) edges EW from G such that every three vertices that induce a P3 in G \ EW form a
clique in G. Then it is not difficult to see that G \ EW satisfies the strong triadic closure if
and only if for every vertex v, NS [v] induces a clique in G.
3
MaxSTC on split graphs
Here we provide an NP-hardness result for MaxSTC on split graphs. A graph G = (V, E) is
a split graph if V can be partitioned into a clique C and an independent set I, where (C, I)
is called a split partition of G. Split graphs form a subclass of the larger and widely known
graph class of chordal graphs, which are the graphs that do not contain induced cycles of
length 4 or more as induced subgraphs. It is known that split graphs are self-complementary,
that is, the complement of a split graph remains a split graph. First we show the following
result.
Lemma 3.1. Let G = (V, E) be a split graph with a split partition (C, I). Let ES be the set
of strong edges in an optimal solution for MaxSTC on G and let IW be the vertices of I
that are incident to at least one edge of ES .
1. If every vertex of IW misses at least three vertices of C in G then ES = E(C).
|
2. If every vertex of IW misses exactly one vertex of C in G then |ES | ≤ |E(C)| + b |IW
2 c.
Proof. Let wi be a vertex of I and let Bi be the set of vertices in C that are non-adjacent to
wi . Let Ai be the strong neighbors of wi in an optimal solution. For the edges of the clique,
there are |Ai ||Bi | weak edges due to the strong triadic closure. Moreover any vertex wj of
I \ {wi } cannot have a strong neighbor in Ai . This means that Ai ∩ Aj = ∅. Notice, however,
that both sets Bi ∩ Bj and Ai ∩ Bj are not necessarily empty.
Observe that IW contains the vertices of I that are incident to at least one strong edge.
Let E(A, B) be the set of weak edges that have one endpointP
in Ai and the other endpoint
in Bi , for every 1 ≤ i ≤ |IW |. We show that 2|E(A, B)| ≥ wi ∈IW |Ai ||Bi |. Let {a, b} ∈
E(A, B) such that a ∈ Ai and b ∈ Bi . Assume that there is a pair Aj , Bj such that {a, b} is
an edge between Aj and Bj , for j 6= i. Then a cannot belong to Aj since Ai ∩ Aj = ∅. Thus
a ∈ Ai and b ∈ Aj . Therefore for every edge {a, b} ∈ E(A, B) there are at most two pairs
(Ai , Bi ) and (Aj , Bj ) for which a ∈ AP
i ∪ Bj and b ∈ Bi ∪ Aj . This means that every edge of
E(A, B) is counted at most twice in wi ∈IW |Ai ||Bi |.
For any two edges {u, v}, {v, z} ∈ E(C) \ E(A, B), observe that they satisfy the strong
triadic closure since there is the edge {u, z} in G. Thus the strong edges of the clique are
exactly the set of edges E(C) \ E(A, B). In total by counting the number of strong edges
4
between the independent
P set and the clique, we have |ES | = |E(C) \ E(A, B)| +
Since 2|E(A, B)| ≥ wi ∈IW |Ai ||Bi |, we get
|ES | ≤ |E(C)| +
X
wi ∈IW
|Bi |
.
|Ai | 1 −
2
P
wi ∈IW
|Ai |.
Now the first claim of the lemma holds because |Bi | = 3 so that IW = ∅. For the second
claim we show that for every vertex of IW , |Ai | = 1. Let wi ∈ IW such that |Ai | ≥ 2 and let
Bi = {bi }. Recall that no other vertex of IW has strong neighbours in Ai . Also note that
there is at most one vertex wj in IW that has bi as a strong neighbour. If such a vertex wj
exist and for the vertex bj of the clique that misses wj it holds bj ∈ Ai , then we let v = bj ;
otherwise we choose v as an arbitrary vertex of Ai . Observe that no vertex of I \ {wi } has a
strong neighbour in Ai \ {v} and only wj ∈ IW is strongly adjacent to bi . Then we label weak
the |Ai | − 1 edges between wi and the vertices of Ai \ {v} and we label strong the |Ai | − 1
edges between bi and the vertices of Ai \ {v}. Making strong the edges between bi and the
vertices of Ai \ {v} does not violate the strong triadic closure since every vertex of C ∪ {wj }
is adjacent to every vertex of Ai \ {v}. Therefore for every vertex wi ∈ IW , |Ai | = 1 and by
substituting |Bi | = 1 in the formula for |ES | we get the claimed bound.
In order to give the reduction, we introduce the following problem that we call maximum disjoint non-neighbourhood: given a split graph (C, I) where every vertex of I misses
three vertices from C, we want to find the maximum subset SI of I such that the nonneighbourhoods of the vertices of SI are pairwise disjoint. In the corresponding decision
version, denoted by MaxDisjointNN, we are also given an integer k and the problem asks
whether |SI | ≥ k. The polynomial-time reduction to MaxDisjointNN is given from the
classical NP-complete problem 3-Set Packing [15]: given a universe U of n elements, a
family F of triplets of U, and an integer k, the problem asks for a subfamily F 0 ⊆ F with
|F 0 | ≥ k such that all triplets of F 0 are pairwise disjoint.
Theorem 3.2. MaxDisjointNN is NP-complete on split graphs.
Proof. Given a split graph G = (C, I) and SI ⊆ I, checking whether SI is a solution for
MaxDisjointNN amounts to checking whether every pair of vertices of SI have common
neighbourhood. As this can be done in polynomial time the problem is in NP. We will give
a polynomial-time reduction to MaxDisjointNN from the classical NP-complete problem
3-Set Packing [15]: given a universe U of n elements, a family F of triplets of U, and an
integer k, the problem asks for a subfamily F 0 ⊆ F with |F 0 | ≥ k such that all triplets of F 0
are pairwise disjoint.
Let (U, F, k) be an instance of the 3-Set Packing. We construct a split graph G = (C, I)
as follows. The clique of G is formed by the n elements of U. For every triplet Fi of F we add
a vertex vi in I that is adjacent to every vertex of C except the three vertices that correspond
to the triplet Fi . Thus every vertex of I misses exactly three vertices from C and sees the rest
of C. Now it is not difficult to see that there is a solution F 0 for 3-Set Packing(U, F, k) of
size at least k if and only if there is a solution SI for MaxDisjointNN(G, k) of size at least
k. For every pair (Fi , Fj ) of F 0 we know that Fi ∩ Fj = ∅ which implies that the vertices
vi and vj have disjoint non-neighbourhood since Fi corresponds to the non-neighbourhood
of vi . By the one-to-one mapping between the sets of F and the vertices of I, every set Fi
belongs to F 0 if and only if vi belongs to SI .
Now we turn to our original problem MaxSTC. The decision version of MaxSTC takes
as input a graph G and an integer k and asks whether there is strong-weak labeling of the
edges of G that satisfies the strong triadic closure with at least k strong edges.
5
w1
···
B1
wk
wk+1
Bk
Bk+1
w|I|
· · ·
I
B|I|
···
· · ·
C
y1
···
yk
yk+1
yk+2
···
yn−1
yn
x1
···
xk
xk+1
xk+2
···
xn−1
xn
CY
IX
Figure 2: The split graph (C ∪ CY , I ∪ IX ) given in the polynomial-time reduction. Every
vertex wi misses the vertices of Bi and sees the vertices of (C ∪ CY ) − Bi . Every vertex xi
misses yi and sees the vertices of (C ∪ CY ) − {yi }. The sets B1 , . . . , Bk are pairwise disjoint
whereas for every set Bj , k < j ≤ |I|, there is a set Bi , 1 ≤ i ≤ k, such that Bi ∩ Bj 6= ∅.
The drawn edges correspond to the strong edges between the independent set and the clique,
and the dashed edges are the only weak edges in the clique C ∪ CY .
Theorem 3.3. The decision version of MaxSTC is NP-complete on split graphs.
Proof. Given a strong-weak labeling (ES , EW ) of a split graph G = (C, I), checking whether
(ES , EW ) satisfies the strong triadic closure amounts to check in G − EW whether there is a
non-edge between its endvertices of every P3 according to Observation 2.1. Thus by listing
all P3 ’s of G \ EW the problem belongs to NP. Next we give a polynomial-time reduction
to MaxSTC from the MaxDisjointNN problem on split graphs which is NP-complete by
Theorem 3.2. Let (G, k) be an instance of MaxDisjointNN where G = (C, I) is a split
graph such that every vertex of the independent set I misses exactly three vertices from the
clique C. For a vertex wi ∈ I, we denote by Bi the set of the three vertices in C that are
non-adjacent to wi . Let |C| = n. We extend G and construct another split graph G0 as
follows (see Figure 2):
• We add n vertices y1 , . . . , yn in the clique that consist the set CY .
• We add n vertices x1 , . . . , xn in the independent set that consist the set IX .
• For every 1 ≤ i ≤ n, yi is adjacent to all vertices of (C ∪ CY ∪ I ∪ IX ) \ {xi }.
• For every 1 ≤ i ≤ n, xi is adjacent to all vertices of (C ∪ CY ) \ {yi }.
Thus wi misses only the vertices of Bi from the clique. By construction it is clear that G0 is
a split graph with a split partition (C ∪ CY , I ∪ IX ). Notice that the clique C ∪ CY has 2n
vertices and G = G0 [I ∪ C].
We claim that G has a solution for MaxDisjointNN of size at least k if and only if G0
has a strong triadic closure with at least n(2n − 1) + b n2 c + d k2 e strong edges.
Assume that {w1 , . . . , wk } ⊆ I is a solution for MaxDisjointNN on G of size at least
k. Since the sets B1 , . . . , Bk are pairwise disjoint, there are k distinct vertices y1 , . . . , yk in
CY such that k ≤ n. We will give a strong-weak labeling for the edges of G0 that fulfills the
strong triadic closure and has at least the claimed number of strong edges. For simplicity,
we describe only the strong edges; the edges of G0 that are not given are all labeled weak.
We label the edges incident to each vertex wi , yi , xi and the three vertices of each set Bi , for
1 ≤ i ≤ k as follows:
6
• The edges of the form {yi , v} are labeled strong if v ∈ (C ∪ CY ) \ Bi or v = wi .
• The edges incident to xi and the three vertices of Bi are labeled strong.
Next we label the edges incident to the rest of the vertices. Let IW be the vertices of
I \ {w1 , . . . , wk } and let CW be the vertices of C \ (B1 ∪ · · · ∪ Bk ). No edge incident to a
vertex of IW is labeled strong. For every vertex u ∈ CW we label the edge {u, v} strong
0 = {x
if v ∈ (C ∪ CY ). Let CY0 = {yk+1 , . . . , yn } and let IX
k+1 , . . . , xn }. Recall that every
0
vertex xk+j is adjacent to every vertex of CY \ {yk+j }. Let ` = b n−k
2 c. Let M = {e1 , . . . , e` }
0
0
be a maximal set of pairwise non-adjacent edges in G [CY ] where ej = {yk+2j−1 , yk+2j }, for
j ∈ {1, . . . , `}; note that M is a maximal matching of G0 [CY0 ]. For every vertex y ∈ CY0 ,
we label the edge {y, v} strong if v ∈ (C ∪ CY ) \ {y 0 } such that {y, y 0 } ∈ M . Moreover, for
j ∈ {1, . . . , `}, the edges {xk+2j−1 , yk+2j } and {xk+2j , yk+2j−1 } are labeled strong. Note that
if n − k is odd then no edge incident to the unique vertex yn belongs to M and all edges
between yn and the vertices of C ∪ CY are labeled strong; in such a case also note that no
edge incident to xn is strong.
Let us show that such a labeling fulfills the strong triadic closure. Any labeling for the
edges inside G0 [C ∪ CY ] is satisfied since G0 [C ∪ CY ] is a clique. Also note that there are
no two adjacent strong edges that have a common endpoint in the clique C ∪ CY and the
two other endpoints in the independent set I ∪ IX . If there are two strong edges incident to
the same vertex v of the independent set then v ∈ {x1 , . . . , xk } and NS [v] = Bi is a clique.
Assume that there are two adjacent strong edges {u, v} and {v, z} such that u ∈ I ∪ IX , and
v, z ∈ C ∪ CY .
• If u ∈ {w1 , . . . , wk } then {u, z} ∈ E(G0 ) since every wi misses only the vertices of Bi .
• If u ∈ {x1 , . . . , xk } then v ∈ Bi and {u, z} ∈ E(G0 ) since every vertex xi misses only
yi .
• If u ∈ IX \ {x1 , . . . , xk } then the strong neighbours of v in C ∪ CY are adjacent to u in
G0 since for the only non-neighbour of u in C ∪ CY there is a weak edge incident to v.
Recall that there is no strong edge incident to the vertices of I \ {w1 , . . . , wk }. Therefore the
given strong-weak labeling fulfills the strong triadic closure.
Observe that the number of vertices in C ∪ CY is 2n. There are exactly 3k + ` weak edges
in G0 [C ∪ CY ]. Thus the number of strong edges in G0 [C ∪ CY ] are n(2n − 1) − 3k − `. There
are k strong edges incident to {w1 , . . . , wk }, 3k strong edges incident to {x1 , . . . , xk }, and 2`
strong edges incident to the vertices of IX \ {x1 , . . . , xk }. Thus the total number of strong
edges is n(2n − 1) − 3k − ` + k + 3k + 2` = n(2n − 1) + ` + k and by substituting ` = b n−k
2 c
we get the claimed bound.
For the opposite direction, assume that G0 has a strong triadic closure with at least
n(2n − 1) + b n2 c + d k2 e strong edges. Let ES be the set of strong edges in such a strongweak labeling. Observe that the number of edges in G0 [C ∪ CY ] is n(2n − 1) which implies
that ES contains edges between the independent set I ∪ IX and the clique C ∪ CY . If no
vertex of IX is incident to an edge of ES then the first statement of Lemma 3.1 implies that
|ES | = |E(C ∪ CY )| = n(2n − 1). And if no vertex of I is incident to an edge of ES then
the second statement of Lemma 3.1 shows that |ES | ≤ |E(C ∪ CY )| + b n2 c. Therefore ES
contains edges that are incident to a vertex of I and edges that are incident to a vertex of
IX .
In the graph spanned by ES we denote by SW the set of vertices of I that have strong
neighbours in C ∪ CY . We will show that the non-neighbourhoods of the vertices of SW in
C ∪CY are disjoint in G0 and, since G is an induced subgraph of G0 , their non-neighbourhoods
are also disjoint in G.
7
Claim 3.4. For every wi ∈ SW , NS (wi ) ⊆ CY and there exists a unique vertex x ∈ IX such
that NS (x) = Bi .
Proof: Let wi be a vertex of SW . We first show that NS (wi ) ⊆ CY . Let Wi be the strong
neighbours of wi in C and let Yi be the strong neighbours of wi in CY . Observe that no other
vertex of SW has a strong neighbour in Wi ∪ Yi . Further notice that there are (|Wi | + |Yi |)|Bi |
weak edges since wi is non-adjacent to the vertices of Bi . We show that for every vertex
wi ∈ SW it holds Wi = ∅. For all vertices wi for which Wi 6= ∅ we replace in ES the strong
edges between wi and the vertices of Wi by the edges between the vertices of Bi and Wi .
Notice that making strong the edges between the vertices of Bi and Wi does not violate the
strong triadic closure since no vertex from SW has a strong neighbour in Bi and every vertex
of IX is adjacent to all the vertices of Wi . Let E(W, B) be the set of edges that have one
endpoint in Wi and the other endpoint in Bi , for every
P wi ∈ SW . Notice that the difference
between the two described solutions
is
|E(W,
B)|
−
|Wi |. By Lemma 3.1 and |Bi | = 3, we
P
know that |E(W, B)| ≥ 3/2 |Wi |. Thus such a replacement is safe for the number of edges
of ES and every vertex wi ∈ SW has strong neighbours only in CY .
Let Xi be the set of vertices of IX that have at least one non-neighbour in Yi . By
construction every vertex of Yi is non-adjacent to exactly one vertex of IX , and thus |Xi | =
|Yi |. Since wi has strong neighbours in Yi , every edge between Xi and Yi is weak. By the
previous argument every vertex of SW has strong neighbours only in CY so that NS (Bi )∩I =
∅. Also notice that no two vertices of the independent set have a common strong neighbour in
the clique, which means that there are at most |Bi | strong neighbours between the vertices of
Bi and IX . Choose an arbitrary vertex x ∈ Xi . We replace all strong edges in ES between Bi
and IX by |Bi | strong edges between x and the vertices of Bi . Notice that such a replacement
is safe since the unique non-neighbour of x belongs to Yi and there are weak edges already
in the solution between Bi and Yi because of the strong edges between wi and Yi . Thus
Bi ⊆ NS (x). We focus on the edges between the vertices of (C ∪ CY ) \ (Bi ∪ Yi ) and x. If
a vertex x of Xi has a strong neighbour u in (C ∪ CY ) \ Bi then the edge {u, y} is weak
where y ∈ Yi is the unique non-neighbour of x. Also notice that NS (u) ∩ (I ∪ IX ) = {x},
NS (y) ∩ (I ∪ IX ) = {wi }, and wi is adjacent to u. Then we can safely replace the strong
edge {x, u} by the edge {u, y} and keep the same size of ES . Hence NS (x) = Bi .
♦
Claim 3.5. For every wi ∈ SW , NS (wi ) = {y} where y ∈ CY is the non-neighbour of x with
NS (x) = Bi .
Proof: Let Yi = NS (wi ). By Claim 3.4 we know that Yi ⊆ CY and there exists x ∈ IX
such that NS (x) = Bi . Both wi and x are vertices of the independent set and, thus, no
other vertex of I ∪ IX has strong neighbours in Bi ∪ Yi . This means that if we remove wi
from SW by making weak the edges incident to wi and the vertices of Yi then the edges
between the vertices of Bi and Yi \ {y} are safely turned into strong. Let ES0 be the set
of strong edges in an optimal solution such that all edges incident to wi are weak. Then
|ES | − |ES0 | = |Yi | + |Bi | − |Yi ||Bi | and |ES | > |ES0 | only if |Yi | = 1 because |Bi | > 1. Thus
NS (wi ) contains exactly one vertex y ∈ CY .
♦
We claim that for every pair of vertices wi , wj ∈ SW , Bi ∩ Bj = ∅. Assume for contradiction that Bi ∩ Bj 6= ∅. Applying Claim 3.4 for wi shows that there exists x ∈ IX that has
strong neighbours in every vertex of Bi ∩ Bj . With a similar argument for wj we deduce that
there exists x0 ∈ IX that has strong neighbours in every vertex of Bi ∩ Bj . If x 6= x0 then a
vertex from Bi ∩ Bj has two distinct strong neighbours in IX which is not possible due to
the strong triadic closure. Thus x = x0 . Claim 3.5 implies that the unique non-neighbour
y of x is strongly adjacent to both wi and wj . This however violates the strong triadic
closure for the edges of ES since wi , wj are non-adjacent and we reach a contradiction. Thus
8
G:
G:
G:
a
b
{cde}
{f gh}
i
j
a
b
{cde}
{f gh}
i
j
a
b
{cde}
{f gh}
i
j
Figure 3: A proper interval graph G and its proper interval ordering. The vertices {c, d, e}
and {f, g, h} form twin sets in G. The two lower orderings show two solutions for MaxSTC
on G. A solid edge corresponds to a strong edge, whereas a dashed edge corresponds to a
weak edge. Observe that the upper solution contains larger number of strong edges than the
lower one. Also note that the lower solution consists an optimal solution for the Cluster
Deletion problem on G.
Bi ∩ Bj = ∅. This means that the number of edges in ES is at least n(2n − 1) + b n2 c + d |S2W | e
which is maximized for k = |SW |. Therefore ES contains the maximum number of |SW |
which is a solution for MaxDisjointNN on G, since G is an induced subgraph of G0 .
4
Computing MaxSTC on proper interval graphs
A graph is a proper interval graph if there is a bijection between its vertices and a family
of closed intervals of the real line such that two vertices are adjacent if and only if the two
corresponding intervals overlap and no interval is properly contained in another interval. A
vertex ordering σ is a linear arrangement σ = hv1 , . . . , vn i of the vertices of G. For a vertex
pair x, y we write x y if x = vi and y = vj for some indices i ≤ j; if x 6= y which implies
i < j then we write x ≺ y. The first position in σ will be referred to as the left end of σ, and
the last position as the right end. We will use the expressions to the left of, to the right of,
leftmost, and rightmost accordingly.
A vertex ordering σ for G is called a proper interval ordering if for every vertex triple
x, y, z of G with x ≺ y ≺ z, {x, z} ∈ E(G) implies {x, y}, {y, z} ∈ E(G). Proper interval
graphs are characterized as the graphs that admit such orderings, that is, a graph is a proper
interval graph if and only if it has a proper interval ordering [21]. We only consider this vertex
ordering characterization for proper interval graphs. Moreover it can be decided in linear
time whether a given graph is a proper interval graph, and if so, a proper interval ordering
can be generated in linear time [21]. It is clear that a vertex ordering σ for G is a proper
interval ordering if and only if the reverse of σ is a proper interval ordering. A connected
proper interval graph without twin vertices has a unique proper interval ordering σ up to
reversal [6, 13]. Figure 3 shows a proper interval graph with its proper interval ordering.
9
b
G:
ab
b{cde}
{cde}{f gh}
{f gh}i
a{cde}
b{f gh}
{cde}i
{f gh}j
ij
IGb
b of the proper interval graph G given in Figure 3.
Figure 4: The line-incompatibility graph G
b by taking into account the weight
The set IGb is a maximum weighted independent set of G,
of each node (i.e., an edge of G) that corresponds to the number of the twin vertices of its
endpoints in G (see Lemma 4.2).
Let us turn our attention to the MaxSTC problem. Instead of maximizing the strong
edges of the original graph G, we will look at the maximum independent set of the following
b of G: for every edge of G there is a node
graph that we call the line-incompatibility graph G
b and two nodes of G
b are adjacent if and only if the vertices of the corresponding edges
in G
induce a P3 in G. Note that the line-incompatibility graph of G is a subgraph of the line
graph1 of G.
Proposition 4.1. A subset S of edges E(G) is an optimal solution for MaxSTC of G if
b
and only if S is a maximum independent set of G.
Proof. By Observation 2.1 for every P3 in G at least one of its two edges must be labeled
b and they cannot belong to
weak in S. This means that these two edges are adjacent in G
b
b are adjacent
an independent set of G. On the other hand, by construction two nodes of G
b
if and only if there is a P3 in G. Thus the nodes of an independent set of G can be labeled
strong in G satisfying the strong triadic closure.
Therefore we seek for the optimal solution of G by looking at a solution for a maximum
b As a byproduct, if we are interested in minimizing the number of
independent set of G.
b We denote by I b the maximum
weak edges then we ask for the minimum vertex cover of G.
G
b To distinguish the vertices of G
b with those of G we refer to the
independent set of G.
former as nodes and to the latter as vertices. For an edge {u, v} of G we denote by uv
b Figure 4 shows the line-incompatibility graph of the proper
the corresponding node of G.
interval graph given in Figure 3.
Lemma 4.2. Let x and y be twin vertices of a graph G. Then there is an optimal solution
IGb such that xy ∈ IGb and for every vertex u ∈ N (x), xu ∈ IGb if and only if yu ∈ IGb .
Proof. First we show that xy is an isolated node in IGb . If xy is adjacent to xu then y is nonadjacent to u in G which contradicts the fact that x and y are twins. Thus xy is an isolated
b which implies xy ∈ I b . For the second argument observe that for every vertex
node in G
G
u ∈ N (x), xu and yu are non-adjacent in IGb . Let u ∈ N (x). Then notice that u ∈ N (y).
b We define
This means that if xu ∈ IGb (resp., yu ∈ IGb ) then yu (resp., xu) is a node of G.
b
the following sets of nodes in G:
/ IGb and let Ay be the set of
• Let Ax be the set of nodes xa such that xa ∈ IGb and ya ∈
nodes ya such that xa ∈ Ax .
• Let By be the set of nodes yb such that yb ∈ IGb and xb ∈
/ IGb and let Bx be the set of
nodes xb such that yb ∈ By .
1
The line graph of G is the graph having the edges of G as vertices and two vertices of the line graph are
adjacent if and only if the two original edges are incident in G.
10
It is clear that Ax ⊆ IGb , By ⊆ IGb , and Ax ∩ By = ∅. Also note that |Ax | = |Ay | and
|By | = |Bx |, since N [x] = N [y].
Let Ixy = IGb \ (Ax ∪ By ) so that IGb = Ax ∪ By ∪ Ixy . We show that every node of Ay is
non-adjacent to any node of IGb \ By . Let ya be a node of Ay . If there is a node az ∈ IGb \ By
that is adjacent to ya then z and y are non-adjacent in G which implies that z and x are
non-adjacent in G. This however leads to a contradiction because xa, az ∈ IGb and xa is
b If there is a node yb ∈ I b that is adjacent to ya then a is non-adjacent
adjacent to az in G.
G
b This means that xb ∈
to b in G so that xa is also adjacent to xb in G.
/ IGb implying that
yb ∈ By . Thus every node of Ay is non-adjacent to any node of IGb \ By and with completely
symmetric arguments, every node of Bx is non-adjacent to any node of IGb \ Ax . Hence both
b By the facts
sets I1 = Ax ∪ Ay ∪ Ixy and I2 = Bx ∪ By ∪ Ixy form independent sets in G.
that |Ax | = |Ay | and |By | = |Bx | we have |I1 | ≥ |Ixy | whenever |Ax | ≥ |By | and |I2 | ≥ |Ixy |
whenever |Ax | < |By |. Therefore we can safely replace one of the sets Ax or By by Bx or Ay
and obtain the solutions I2 or I1 , respectively. Now observe that in both solutions I1 and I2
we have xu ∈ Ii if and only if yu ∈ Ii , for i ∈ {1, 2}, and this completes the proof.
Lemma 4.2 suggests to consider a graph G that has no twin vertices as follows. We
partition V (G) into sets of twins. A vertex that has no twin appears in its twin set alone.
For every twin set Wx we choose an arbitrary vertex x and remove all its twin vertices except
x from G. Let G0 be the resulting graph that has no twin vertices. For every edge {x, y}
of G0 we assign a weight equal to the product |Wx | · |Wy |. This value corresponds to all
edges of the original graph G between the vertices of Wx and Wy . The line-incompatibility
c0 of G0 is constructed as defined above with the only difference that a node of G
c0 has
graph G
0
weight equal to the weight of its corresponding edge in G . Let IG
c0 be a maximum weighted
0
c
independent set, that is an independent set of G such that the sum of the weights of its
nodes is maximized. Then by Lemma 4.2 we have IGb = IG
c0 ∪ S(W ) where S(W ) contains
|Wx |(|Wx | − 1)/2 nodes for every twin set Wx . Therefore we are interested in computing
c0 . Also note that G0 is an induced subgraph of
the maximum weighted independent set of G
the original graph G. Since there is no ambiguity in the forthcoming results and in order
c0 as G
b by assuming that G has no twin vertices and
to avoid heavier notation we refer to G
every vertex of G has a positive weight.
Before reaching the details of our algorithm for proper interval graphs, let us highlight
the difference between the optimal solution for MaxSTC and the optimal solution for the
Cluster Deletion. As already explained in the Introduction a solution for Cluster
Deletion fulfils the strong triadic closure, though the converse is not necessarily true. In
fact such an observation carries out for the class of proper interval graphs as shown in the
example given in Figure 3. For the Cluster Deletion problem twin vertices can be grouped
together following a similar characterization with Lemma 4.2, as proved in [3]. Therefore
when restricted to proper interval graphs the optimal solution for Cluster Deletion does
not necessarily imply an optimal solution for MaxSTC.
Let G be a proper interval graph and let σ be a proper interval ordering for G. We say
that a solution IGb has the consecutive strong ordering with respect to σ if for any three
vertices x, y, z of G with x ≺ y ≺ z the following holds: xz ∈ IGb implies xy, yz ∈ IGb . Our
task is to show that such an optimal ordering exists. We start by characterizing the optimal
solution IGb with respect to the proper interval ordering σ.
Lemma 4.3. Let x, y, z be three vertices of a proper interval graph G such that x ≺ y ≺ z.
If xz ∈ IGb then xy ∈ IGb or yz ∈ IGb .
b If
Proof. We show that at least one of xy or yz belongs to IGb . Consider the node xy in G.
xy is adjacent to a node xx` ∈ IGb then {x` , y} ∈
/ E(G). Then observe that x` ≺ y because
11
x ≺ y and {x` , y} ∈
/ E(G). Since both xx` and xz belong to IGb , {x` , z} ∈ E(G). This
however contradicts the proper interval ordering because x` ≺ y ≺ z, {x` , z} ∈ E(G) and
y is non-adjacent to x` . Thus xy is non-adjacent to any node xx` ∈ IGb and, in analogous
fashion, yz is non-adjacent to any node zzr ∈ IGb .
Now assume that xy is adjacent to a node yyr ∈ IGb and yz is adjacent to a node y` y ∈ IGb .
This means that {x, yr } ∈
/ E(G) and {z, y` } ∈
/ E(G). Since {x, z} ∈ E(G), by the proper
interval ordering we have y` ≺ x ≺ y ≺ z ≺ yr . Then notice that {y` , yr } ∈ E(G), because
both yyr , yy` ∈ IGb . By the proper interval ordering we know that both x and z are adjacent
to y` , yr , leading to a contradiction to the assumptions {x, yr } ∈
/ E(G) and {z, y` } ∈
/ E(G).
Therefore at least one of xy or yz belongs to IGb .
Thus by Lemma 4.3 we have two symmetric cases to consider. The next characterization
suggests that there is a fourth vertex with important properties in each corresponding case.
Lemma 4.4. Let x, y, z be three vertices of a proper interval graph G such that x ≺ y ≺ z
and xz ∈ IGb .
• If xy ∈
/ IGb and yz ∈ IGb then xy is non-adjacent to any node x` x ∈ IGb and there is a
vertex w such that yw ∈ IGb , {x, w} ∈
/ E(G), and z ≺ w.
• If xy ∈ IGb and yz ∈
/ IGb then yz is non-adjacent to any node zzr ∈ IGb and there is a
vertex w such that wy ∈ IGb , {w, z} ∈
/ E(G) and w ≺ x.
Proof. Let xy ∈
/ IGb and yz ∈ IGb . The case for xy ∈ IGb and yz ∈
/ IGb is completely symmetric.
Assume for contradiction that there is no vertex w such that yw ∈ IGb , {x, w} ∈
/ E(G), and
z ≺ w. We prove that xy is non-adjacent to any node of IGb , contradicting the optimality of
IGb . Suppose first that xy is adjacent to a node x` x ∈ IGb . Then y is non-adjacent to x` in
G. Notice that x` ≺ x because y is adjacent to x and x ≺ y. Due to the fact that yz ∈ IGb ,
b which implies that {x` , z} ∈ E(G). Since
we have that x` x and xz are non-adjacent in G
x` ≺ x ≺ y ≺ z and {x` , z} ∈ E(G), by the proper interval ordering we get {x` , y} ∈ E(G)
leading to a contradiction. Thus xy is non-adjacent to any node x` x ∈ IGb .
Next assume that xy is adjacent to a node yyr ∈ IGb . Then {x, yr } ∈
/ E(G). By the
assumption that there is no vertex w with yw ∈ IGb , {x, w} ∈
/ E(G), and z ≺ w, we have
yr ≺ z. This particularly means that yr ≺ x or x ≺ yr ≺ z. However both cases lead
to a contradiction to {x, yr } ∈
/ E(G) since in the former case we have {yr , y} ∈ E(G) and
yr ≺ x ≺ y, and in the latter case we know that {x, z} ∈ E(G). Therefore xy has no neighbor
in IGb reaching a contradiction to the optimality of IGb .
Now we are ready to show that that there is an optimal solution that has the described
properties with respect to the given proper interval ordering.
Lemma 4.5. There exists an optimal solution IGb that has the consecutive strong property
with respect to σ.
Proof. Let σ be a proper interval ordering for G. Assume for contradiction that IGb does not
have the consecutive strong property. Then there exists a conflict with respect to σ, that is,
there are three vertices x, y, z with x ≺ y ≺ z and xz ∈ IGb such that xy ∈
/ IGb or yz ∈
/ IGb . We
will show that as long as there are conflicts in σ, we can reduce the number of conflicts in
σ without affecting the value of the optimal solution IGb . Consider such a conflict formed by
the three vertices x ≺ y ≺ z with xz ∈ IGb . By Lemma 4.3 we know that xy ∈ IGb or yz ∈ IGb .
Assume that yz ∈ IGb . Then clearly xy ∈
/ IGb , for otherwise there is no conflict. Then by
Lemma 4.4 there is a vertex w such that yw ∈ IGb , {x, w} ∈
/ E(G), and x ≺ y ≺ z ≺ w.
Notice that both triples x, y, z and y, z, w create conflicts in σ.
12
We start by choosing an appropriate such conflict that is formed by four vertices x, y, z, w
so that x ≺ y ≺ z ≺ w, xz, yz, yw ∈ IGb , and {x, w} ∈
/ E(G). Fix y and z in σ with y, z
being the leftmost and the rightmost vertices, respectively, such that for every vertex v with
y ≺ v ≺ z, yv, vz ∈ IGb holds. Recall that yz ∈ IGb . We choose x as the leftmost vertex such
that xz ∈ IGb and we choose w as the rightmost vertex such that yw ∈ IGb . Observe that
{x, w} ∈
/ E(G) since y and z participate in a conflict. Due to the properties of the considered
conflicts all such vertices exist (see for e.g., Figure 5).
Let W (x) be the set of vertices wi such that ywi ∈ IGb and {x, wi } ∈
/ E(G), and let X(w)
be the set of vertices xj such that xj z ∈ IGb and {xj , w} ∈
/ E(G). Observe that w ∈ W (x)
and x ∈ X(w) such that w is the rightmost vertex in W (x) and x is the leftmost vertex in
X(w). For a vertex wi of W (x) observe the following. If wi ≺ x then {wi , x} ∈ E(G) because
{wi , y} ∈ E(G) and if x ≺ wi ≺ z then {wi , x} ∈ E(G) because {x, z} ∈ E(G). Thus z ≺ wi
which implies that {z, wi } ∈ E(G) since {y, wi } ∈ E(G). If zwi ∈ IGb then by the fact that
xz ∈ IGb we have {x, wi } ∈ E(G) contradicting the definition of W (x). Moreover for every
vertex b1 such that wi b1 ∈ IGb notice that x ≺ b1 since {x, wi } ∈
/ E(G). If x ≺ b1 ≺ wi then
{z, b1 } ∈ E(G) since x ≺ z ≺ wi ; and if wi ≺ b1 then due to the fact that ywi , wi b1 ∈ IGb
and {y, b1 } ∈ E(G) we have again {z, b1 } ∈ E(G) since y ≺ z ≺ b1 . Furthermore consider a
vertex b2 such that z ≺ b2 ≺ w and b2 ∈
/ W (x). This means that yb2 ∈
/ IGb or yb2 ∈ IGb with
{b2 , x} ∈ E(G). The latter case implies that b2 is adjacent to every vertex of X(w), since x
is the leftmost vertex in X(w) and every vertex of X(w) is to the left of z. Hence for every
vertex wi of W (x) the following hold:
• z ≺ wi ,
• zwi ∈
/ IGb ,
• for every node wi b1 ∈ IGb , {z, b1 } ∈ E(G), and
• for every vertex b2 with z ≺ b2 ≺ w and b2 ∈
/ W (x), yb2 ∈
/ IGb or b2 is adjacent to every
vertex of X(w).
With symmetric arguments for every vertex xj of X(w) we have the following:
• xj ≺ y,
• xj y ∈
/ IGb ,
• for every node a1 xj ∈ IGb , {a1 , y} ∈ E(G), and
• for every vertex a2 with x ≺ a2 ≺ y and a2 ∈
/ X(w), a2 z ∈
/ IGb or a2 is adjacent to every
vertex of W (x).
The topmost ordering given in Figure 5 illustrates the corresponding cases.
b such that wi ∈ W (x), and let Zx be the set of nodes
Let Yw be the set of nodes ywi in G
b
xj z in G such that xj ∈ X(w). Observe that Yw , Zx ⊆ IGb by the previous arguments. We
show that removing either Yw or Zx from IGb does not create any new conflict. Let ywi ∈ Yw
and let u be a vertex such that uy ∈ IGb and uwi ∈ IGb . If y ≺ u ≺ wi then no conflict is
created by removing ywi from IGb . Assume that u ≺ y ≺ wi . Observe that x ≺ u ≺ z. Then
xu ∈
/ IGb because {x, wi } ∈
/ E(G). Since xz ∈ IGb and at least one of xu, uz belongs to IGb ,
we have uz ∈ IGb . However this contradicts the leftmost choice for y in x ≺ u ≺ y ≺ z and
there is no such vertex u. Next assume that y ≺ wi ≺ u. Since wi is non-adjacent to x and
wi ≺ u, u is non-adjacent to x, as well. Then according to the definition of W (x), u ∈ W (x)
and yu ∈ Yw . The case for the nodes of Zx is completely symmetric. Thus no conflicts are
created by removing the nodes of Yw or the nodes of Zx from IGb .
13
X(w)
a1
x
W (x)
···
xj
a2
y
u
z
b2
wi
···
x
···
xj
a2
y
u
z
b2
wi
···
X(w)
a1
x
b1
W (x)
X(w)
a1
w
w
b1
W (x)
···
xj
a2
y
u
z
b2
wi
···
w
b1
Figure 5: A proper interval ordering for a graph G with three different solutions considered
b that belongs to I b ,
in the proof of Lemma 4.5. A solid edge corresponds to a node of G
G
which means that such an edge is labeled strong in an optimal strong-weak labeling, whereas
b that does not belong to I b , which means that
a dashed edge corresponds to a node of G
G
such an edge is labeled weak in an optimal strong-weak labeling. Observe that the lowest
two orderings contain less conflicts than the topmost, that is, triple of vertices that violate
the consecutive strong property.
b such that xj ∈ X(w), and let Zw be the set of
Let Yx be the set of nodes xj y in G
b
nodes zwi in G such that wi ∈ W (x). We denote by I(Yx ) and I(Zw ) the following sets of
nodes: I(Yx ) = IGb \ Yw ∪ Yx and I(Zw ) = IGb \ Zx ∪ Zw . We show that both sets form
b Consider the case for I(Yx ). Let xj y be a node of I(Yx ). There are
independent sets in G.
two cases to consider:
there
is
a
node
ux
∈
I
\
Y
∪ Yx and {u, y} ∈
/ E(G) or there is a
j
w
b
G
node yv ∈ IGb \ Yw ∪ Yx and {xj , v} ∈
/ E(G). In the former case we know that uxj ∈ IGb and
for every node uxj ∈ IGb , {u, y} must be an edge of G which leads to a contradiction to the
non-adjacency of u and y. For the latter case observe that yv ∈ IGb \ Yw and v ∈
/ W (x). Since
{xj , v} ∈
/ E(G) and {y, v} ∈ E(G), we have z ≺ v and by the rightmost choice of w for y we
have z ≺ v ≺ w. This however implies that z ≺ v ≺ w, v ∈
/ W (x) and yv ∈ IGb showing that
{xj , v} ∈ E(G). Completely symmetric arguments hold for I(Zw ). The two lowest orderings
given in Figure 5 illustrate the considered cases. Thus I(Yx ) and I(Zw ) form independent
b
sets in G.
Now observe that both I(Yx ) and I(Zw ) have a smaller number of conflicts with respect
to IGb because either x, y, z in I(Yx ) or y, z, w in I(Zw ) satisfy the consecutive strong property.
It is clear that the difference between I(Yx ) and IGb are the nodes of Yx and Yw , whereas the
14
difference between I(Zw ) and IGb are the nodes of Zw and Zx . For a set A of vertices having
positive weights, denote by M (A) the sum of the weights of its vertices. If M (X(w)) ≥
M (Z(x)) then M (I(Yx )) ≥ M (IGb ) and if M (X(w)) < M (Z(x)) then M (I(Zw )) > M (IGb ).
Thus in any case we can replace appropriate set of nodes in IGb and obtain an optimal solution
with a smaller number of conflicts. Therefore by applying such a replacement in every such
conflict, we get an optimal solution that has no conflicts and, thus, it satisfies the consecutive
strong property.
Lemma 4.5 suggests to find an optimal solution that has the consecutive strong property
with respect to σ. In fact by Proposition 4.1 and the proper interval ordering, this reduces
to computing the largest proper interval subgraph H of G such that the vertices of every P3
of H induce a clique in G.
Let G be a proper interval graph and let σ = hv1 , . . . , vn i be its proper interval ordering.
For a vertex vi we denote by v`(i) and vr(i) its leftmost and rightmost neighbors, respectively,
in σ. Observe that for any two vertices vi ≺ vj in σ, v`(i) v`(j) and vr(i) vr(j) [6]. For
1 ≤ i ≤ r(1), let Vi = {v1 , . . . , vi }, that is, Vi contains the first i vertices in σ. Given the set
Vi , let r be an integer such that 1 ≤ i ≤ r ≤ r(1).
Let A(G) be the value of an optimal solution IGb for G. For the set Vi we denote by
B(Vi ) the value that corresponds to the total weight of the edges incident to v1 and each
of v2 , . . . , vi . Observe that any subset of vertices of Vi induces a clique in G. We denote
by C(Vi ) the value that corresponds to the total weight of the edges among all vertices of
Vi . Given the first vertices Vi and i ≤ r ≤ r(1), let A(G, Vi , r) be the value of the optimal
solution I(G, Vi , r) in G such that every edge among the vertices of Vi belongs to I(G, Vi , r)
and uvk ∈
/ I(G, Vi , r) with u ∈ Vi and k > r. As a trivial case observe that if G contains
exactly two vertices v1 , v2 then A(G) = B({v1 , v2 }) = C({v1 , v2 }).
Lemma 4.6. Let G be a proper interval graph and let Vi and r such that 1 ≤ i ≤ r ≤ r(1).
Then, A(G) = A(G, ∅, r) = A(G, {v1 }, r(1)), and
max {A(G − {v1 }, Vj \ {v1 }, r) + B(Vj )} if i < r,
i≤j≤r
A(G, Vi , r) = A(G − Vi , {vi+1 }, r(i + 1)) + C(Vi )
if i = r and i < n,
C(Vi )
otherwise.
Proof. We show that A(G) computes the value of an optimal solution that satisfies the
consecutive strong property. By Lemma 4.5 such an ordering exists. Since there is no edge
between v1 and vk with k > r(1), it follows that A(G) = A(G, {v1 }, r(1)). Observe that
every induced subgraph of a proper interval graph is proper interval, which implies that the
graphs G − {v1 } and G − Vi remain proper interval. Let Vi = {v1 , . . . , vi } be the first i
vertices of G. According to Lemma 4.5 if vi vk ∈
/ I(G, Vi , r) with k > i, then every node vj vk ,
1 ≤ j ≤ i, does not belong to I(G, Vi , r). Again by Lemma 4.5 if v1 vk ∈ I(G, Vi , r) with
k > i, then every node vj vk , 1 ≤ j ≤ i, belongs to I(G, Vi , r).
b corresponding to the edges of G[Vi ]. Observe that by the
Let I(Vi ) be the nodes of G
definition of I(G, Vi , r), I(Vi ) ⊆ I(G, Vi , r). If i = r then it is clear that the nodes of I(Vi )
belong to I(G, Vi , r). Moreover if i = r < n then no node vi vk , with i < k, can be added in
I(G, Vi , r), implying that all the nodes corresponding to the edges between Vi and G − Vi do
not belong to I(G, Vi , r). Thus we can safely add I(Vi ) to an optimal solution for G − Vi ,
that is, I(G − Vi , {vi+1 }, r(i + 1)).
Assume that i < r. To see that I(Vi ) ⊆ I(G, Vi , r), notice that Vi ⊆ Vj for every
i ≤ j ≤ r and C(Vi ) = B(Vi ) + B(Vi \ {v1 }) + · · · + B({vi−1 , vi }). Consider the nodes v1 vj of
b corresponding to the edges incident to v1 . Let v1 vj ∈ I b with j as large as possible. Then
G
G
15
b with 1 ≤ j 0 ≤ j belongs to I b . Thus
i ≤ j ≤ r ≤ r(1). By Lemma 4.5 every node v1 vj 0 of G
G
I(Vj ) ⊆ I(G, Vi , r). Furthermore for every node vj vk , with r ≤ r(1) < k, we know that vj vk
b since {v1 , vk } ∈
is adjacent to v1 vj in G
/ E(G). Thus vj vk ∈
/ IGb which implies that every node
0
0
vj vk with j ≤ j does not belong to IGb . This particularly means that vj vk ∈ I(G, Vi , r) only
if k ≤ r ≤ r(1). Therefore given an optimal solution by removing v1 for all such possible
values for j, we obtain the desired solution.
Now we are equipped with our necessary tools in order to obtain our main result, namely
a polynomial-time algorithm that solves the MaxSTC problem on a proper interval graph
G.
Theorem 4.7. There is a polynomial-time algorithm that computes the MaxSTC of a proper
interval graph.
Proof. Let G be a proper interval graph on n vertices and m edges. We first compute its
proper interval ordering σ in linear time [21]. Then we compute its twin sets by using the
fact that u and v are twins if and only if `(u) = `(v) and r(u) = r(v). Contracting the
twin sets according to Lemma 4.2 results in a proper interval graph in which every vertex
is associated with a positive weight. In order to compute the optimal solution A(G) we use
a dynamic programming approach based on its recursive formulation given in Lemma 4.6.
Correctness follows from Proposition 4.1 and Lemmata 4.5 and 4.6.
Regarding the running time, notice that given the ordering σ we can remove the twin
vertices in linear time. All instances of A(G, Vi , r) can be computed as follows. Given the first
vertex v1 we compute all possible sets Vi which are bounded by n. Since r ≤ r(1) ≤ n, the
number of instances A(G, Vi , r) generated by v1 is at most n2 . Also observe that computing
the values B(Vi ) and C(Vi ) takes O(m) time. Therefore the total running time of the
algorithm is O(n3 m).
5
Concluding remarks
Given the first study with positive and negative results for the MaxSTC problem on restricted input, there are some interesting open problems. Despite the structural properties
that we proved for the solution on proper interval graphs, the complexity of MaxSTC
on interval graphs is still open. Determining the complexity of MaxSTC for other graph
classes towards AT-free graphs seems interesting direction for future work. More precisely,
by Proposition 4.1 it is interesting to consider the line-incompatibility graph of comparability graphs since they admit a well-known similar characterization [17]. Moreover it is
natural to characterize the graphs for which their line-incompatibility graph is perfect. Such
a characterization will lead to to further polynomial cases of MaxSTC, since the maximum
independent set of perfect graphs admits a polynomial solution [10]. A typical example is the
class of bipartite graphs for which their line graph coincides with their line-incompatibility
graph and it is known that the line graph of a bipartite graph is perfect (see for e.g., [4]).
As we show next, another paradigm of this type is the class of trivially-perfect graphs.
A graph G is called trivially-perfect (also known as quasi-threshold) if for each induced
subgraph H of G, the number of maximal cliques of H is equal to the maximum size of an
independent set of H. It is known that the class of trivially-perfect graphs coincides with
the class of (P4 , C4 )-free graphs, that is every trivially-perfect graph has no induced P4 or
C4 [9]. A cograph is a graph without an induced P4 , that is a cograph is a P4 -free graph.
Hence trivially-perfect graphs form a subclass of cographs.
Theorem 5.1. The line-incompatibility graph of a trivially-perfect graph is cograph.
16
Proof. Let G be a trivially-perfect graph, that is G is a (P4 , C4 )-free graph. We will show
b of G is a P4 -free graph. Consider any P3 in G.
b Due
that the line-incompatibility graph G
b
to the construction of G, the P3 has one of the following forms: (i) v1 v2 , v2 v3 , v3 v4 or (ii)
v1 x, v2 x, v3 x. We prove that the P3 has the second form because G has no induced P4 or
C4 . If (i) applies then {v1 , v3 }, {v2 , v4 } ∈
/ E(G) and {v1 , v2 }, {v2 , v3 }, {v3 , v4 } ∈ E(G) which
implies that v4 6= v1 . Thus G contains a P4 or a C4 depending on whether there is the edge
b has the form v1 x, v2 x, v3 x where v1 , v2 , v3 , x are distinct
{v1 , v4 } in G. Hence every P3 in G
b contains a P4 . Then the P4 is of the
vertices of G. Now assume for contradiction that G
form v1 x, v2 x, v3 x, v4 x because it contains two induced P3 ’s. The structure of the P4 implies
that {v1 , v2 }, {v2 , v3 }, {v3 , v4 } ∈
/ E(G) and {v1 , v3 }, {v2 , v4 }, {v4 , v1 } ∈ E(G). This however
shows that the vertices v3 , v1 , v4 , v2 induce a P4 in G leading to a contradiction that G is a
b is a P4 -free graph.
(P4 , C4 )-free graph. Therefore G
By Theorem 5.1 and the fact that the maximum independent set of a cograph can be
computed in linear time [5], MaxSTC can be solved in polynomial time on trivially-perfect
graphs. We would like to note that the line-incompatibility graph of a cograph or a proper
interval graph is not necessarily a perfect graph.
More general there are extensions and variations of the MaxSTC problem that are
interesting to consider as proposed in [25]. A natural graph modification problem is the
problem of adding the minimum number of edges that satisfy the strong triadic closure
having as few as possible weak edges in the original graph. More formally the objective is
to add F edges in G so that the resulting graph satisfies the strong triadic closure with the
minimum number of |F | + |EW |.
References
[1] A. B. Adcock, B. D. Sullivan, M. W. Mahoney. Tree decompositions and social graphs.
Internet Mathematics, 12:315–361, 2016.
[2] L. Backstrom and J. Kleinberg. Romantic partnerships and the dispersion of social ties:
a network analysis of relationship status on facebook. In Proceedings of CSCW 2014,
pp. 831–841, 2014.
[3] F. Bonomo, G. Durán, and M. Valencia-Pabon. Complexity of the cluster deletion
problem on subclasses of chordal graphs. Theoretical Computer Science, 600:59-69,
2015.
[4] A. Brandstädt, V.B. Le, and J. Spinrad. Graph Classes – a Survey. SIAM Monographs
in Discrete Mathematics and Applications, SIAM, Philadelphia, 1999.
[5] D.G. Corneil, Y. Perl, and L.K. Stewart. Complement reducible graphs. Discrete Applied
Mathematics, 3:163–174, 1981.
[6] X. Deng, P. Hell, and J. Huang. Linear-time representation algorithms for proper
circular-arc graphs and proper interval graphs. SIAM J. Comput., 25:390–403, 1996.
[7] D. Easley and J. Kleinberg. Networks, Crowds, and Markets: Reasoning About a Highly
Connected World. Cambridge University Press, 2010.
[8] J. Edmonds. Paths, trees and flowers. Canad. J. Math., 17:449–467, 1965.
[9] M.C. Golumbic. Trivially perfect graphs. Discrete Mathematics, 24:105-107, 1978.
[10] M. Götschel, L. Lovász, and A. Scrijver. Polynomial algorithms for perfect graphs.
Annals of Discrete Mathematics, 21:325–356, 1984.
17
[11] M. Granovetter. The strength of weak ties. American Journal of Sociology, 78:1360–
1380, 1973.
[12] P. Heggernes, D. Lokshtanov, J. Nederlof, C. Paul, and J. A. Telle. Generalized graph
clustering: recognizing (p, q)-cluster graphs. In Proceedings of WG 2010, pp. 171–183,
2010.
[13] L. Ibarra. The clique-separator graph for chordal graphs. Discrete Applied Mathematics,
157:1737-1749, 2009.
[14] M. O. Jackson. Social and economic networks. Princeton University press, vol. 3, 2008.
[15] R. M. Karp. Reducibility among combinatorial problems. Complexity of Computer
Computations, 85–103, 1972.
[16] D. J. Kleitman and R. V. Vohra. Computing the bandwidth of interval graphs. SIAM
J. Disc. Math., 3:373–375, 1990.
[17] D. Kratsch, R. M. McConnell, K. Mehlhorn, and J. P. Spinrad. Certifying algorithms
for recognizing interval graphs and permutation graphs. SIAM J. Comput., 36: 326–353,
2006.
[18] L. C. Lau. Bipartite roots of graphs. ACM Transactions on Algorithms, 2:178–208,
2006.
[19] L. C. Lau and D. G. Corneil. Recognizing powers of proper interval, split, and chordal
graphs. SIAM J. Disc. Math., 18: 83–102, 2004.
[20] F.T. Leighton and S. Rao. An approximate max-flow min-cut theorem for uniform
multicommodity flow problems with applications to approximation algorithms. In Proceedings of FOCS 1988, pp. 422–431, 1998.
[21] P. J. Looges and S. Olariu. Optimal greedy algorithms for indifference graphs. Computers & Mathematics with Applications, 25:1525, 1993.
[22] M. Milanič and O. Schaudt. Computing square roots of trivially perfect and threshold
graphs. Discrete Applied Mathematics, 161:1538-1545, 2013.
[23] J. L. Pfaltz. Chordless cycles in networks. In Proceedings of ICDE Workshops, pp.
223–228, 2013.
[24] F. S. Roberts. Indifference Graphs. In Proof Techniques in Graph Theory, Academic
Press, New York, 139–146, 1969.
[25] S. Sintos and P. Tsaparas. Using strong triadic closure to characterize ties in social
networks, In Proceedings of KDD 2014, pp. 1466–1475, 2014.
[26] J. Ugander, L. Backstrom, and J. Kleinberg. Subgraph frequencies: Mapping the empirical and extremal geography of large graph collections. In Proceedings of WWW 2013,
pp. 1307–1318, 2013.
18
| 8cs.DS
|
arXiv:1603.02177v1 [math.GR] 7 Mar 2016
A Sufficient Condition for Nilpotency of
the Commutator Subgroup
Raimundo Bastos and Pavel Shumyatsky
Abstract. Let G be a finite group with the property that if a, b
are commutators of coprime orders, then |ab| = |a||b|. We show
that G′ is nilpotent.
The following criterion of nilpotency of a finite group was established by B. Baumslag and J. Wiegold [1].
Theorem 1. Let G be a finite group in which |ab| = |a||b| whenever
the elements a, b have coprime orders. Then G is nilpotent.
Here the symbol |x| stands for the order of the element x in a group
G. In the present article we establish a similar criterion of nilpotency
for the commutator subgroup G′ . Recall that an element g ∈ G is a
commutator if g = x−1 y −1xy for suitable x, y ∈ G. By definition, the
commutator subgroup G′ is the subgroup of G generated by all commutators. The purpose of the present article is to prove the following
theorem.
Theorem 2. Let G be a finite group in which |ab| = |a||b| whenever the elements a, b are commutators of coprime orders. Then G′ is
nilpotent.
In view of the above theorem one might suspect that a similar phenomenon holds for other group-words. Recall that a group-word w =
w(x1 , . . . , xs ) is a nontrivial element of the free group F = F (x1 , . . . , xs )
on free generators x1 , . . . , xs . A word is a commutator word if it belongs to the commutator subgroup F ′ . Given a group-word w, it can be
viewed as function defined in any group G. The subgroup of G generated by the w-values is called the verbal subgroup of G corresponding
2010 Mathematics Subject Classification. 20D30, 20F16, 20D25.
Key words and phrases. Finite groups, commutators.
This work was supported by CNPq-Brazil.
1
2
RAIMUNDO BASTOS AND PAVEL SHUMYATSKY
to the word w. It is usually denoted by w(G). One might ask the
following question.
Let w be a commutator word and G a finite group with the property
that if a, b are w-values of coprime order, then |ab| = |a||b|. Is then the
verbal subgroup w(G) nilpotent?
Asking a similar question for noncommutator words would not be
interesting since an easy counter-example is provided just by any nonabelian simple group G, say of exponent e, and the word xn , where
n is a divisor of e such that e/n is prime. Even in the case of commutator words the answer to our question is negative: Kassabov and
Nikolov showed in [3] that for any n ≥ 7 the alternating group An
admits a commutator word all of whose nontrivial values have order
3. We suspect that the answer to the question is positive in the case
of multilinear commutator words, that is, words having a form of a
multilinear Lie monomial (like for example [[x1 , [x2 , x3 ]], [x4 , x5 ]]).
Throughout the remaining part of this short note G denotes a finite
group satisfying the hypothesis of Theorem 2. We denote by X the set
of commutators in G. As usual, π(K) stands for the set of primes
dividing the order of a group K. The Fitting subgroup of K is denoted
by F (K).
Lemma 3. Let x ∈ X and N be a subgroup normalized by x. If
(|x|, |N|) = 1, then [x, N] = 1.
Proof. Choose y ∈ N. The order of the commutator [x, y] is
prime to that of x. Therefore we must have |x[x, y]| = |x||[x, y]|. However x[x, y] = y −1xy. This is a conjugate of x and so |x[x, y]| = |x|.
Therefore [x, y] = 1.
Lemma 4. If G is soluble, then G′ is nilpotent.
Proof. Arguing by induction on |G| we can assume that the second commutator subgroup G′′ is nilpotent. Suppose that there are two
different primes p ∈ π(G′ ) and q ∈ π(G′′ ). Let P be a Sylow p-subgroup
of G′ and Q a Sylow q-subgroup of G′′ . It is straightforward from the
Focal Subgroup Theorem [2, Theorem 7.3.4] that P is generated by
P ∩ X. Lemma 3 tells us that if x ∈ P ∩ X, we have [Q, x] = 1.
Therefore [Q, P ] = 1. It follows that P is normal in P G′′ and therefore
P ≤ F (P G′′). Since P G′′ is normal in G′ , we conclude that P ≤ F (G).
This holds for every prime p such that G′′ is not a p-group. Therefore
if G′′ is divisible by at least two different primes, all Sylow subgroups
of G′ belong to F (G) and so G′ is nilpotent. If G′′ is q-group, it is
obvious that the Sylow q-subgroup of G′ is normal in G and so again
NILPOTENCY
3
we conclude that all Sylow subgroups of G′ belong to F (G). The proof
is complete.
Proof of Theorem 2. Now suppose that G is a counter-example
of minimal order. So all proper subgroups in G are soluble and we can
assume that G = G′ . Let R be the soluble radical in G. It follows
that G/R is nonabelian simple. By Lemma 4 R′ is nilpotent. Suppose
that G is nonsimple and R 6= 1 and, for a prime q, let Q be the Sylow
subgroup of F (G). Let T be the subgroup of G generated by all commutators that are q ′ -elements. In view of the Focal Subgroup Theorem,
all Sylow p-subgroups of G′ for p 6= q are contained in T . Therefore the
commutator subgroup of G/T is a q-group. Since G = G′ , we conclude
that G = T . Combining Lemma 3 with the Focal Subgroup Theorem
we now deduce that F (G) ≤ Z(G).
Further, we remark that for any x ∈ G the subgroup hx, Ri is soluble
and so hx, Ri′ is nilpotent. It follows that [R, x] ≤ F (G) = Z(G)
and therefore R = Z2 (G). In particular, R is nilpotent and therefore
R = Z(G). Thus, our group G is quasisimple.
Since G does not possess a normal 2-complement, it follows from the
Frobenius Theorem [2, Theorem 7.4.5] that G contains a 2-subgroup H
and an element of odd order b ∈ NG (H) such that [H, b] 6= 1. In view
of Thompson’s Theorem [2, Theorem 5.3.11] we can assume that H
is of nilpotency class at most two and H/Z(H) is elementary abelian.
We claim that G contains an element a such that
a is a 2-element;
a ∈ X;
a has order 2 modulo Z(G).
Indeed, since H is of class at most two, all elements in [H, g] are
commutators for any g ∈ H. If for some g ∈ H the subgroup [H, g] is
not contained in Z(G), any element of [H, g] that has order 2 modulo
Z(G) has the required properties. Therefore we assume that [H, g] is
contained in Z(G) for any g ∈ H. In particular H ′ ≤ Z(G). It follows
that [H, b] ∩ CG (b) ≤ Z(G) and all elements in [H, b] are commutators
modulo Z(G). If d ∈ [H, b] such that d 6∈ Z(G) and d2 ∈ Z(G), then
the commutator [d, b] is as required. This proves the existence of an
element a with the above properties.
Now we fix such an element a. Since G/Z(G) is nonabelian simple,
it follows from the Baer-Suzuki Theorem [2, Theorem 3.8.2] that there
exists an element t ∈ G such that the order of [a, t] is odd. On the one
4
RAIMUNDO BASTOS AND PAVEL SHUMYATSKY
hand, it is clear that a inverts [a, t]. On the other hand, by Lemma 3,
a must commute with [a, t]. This is a contradiction.
References
[1] B. Baumslag and J. Wiegold, A Sufficient Condition for Nilpotency in a Finite
Group, preprint available at arXiv:1411.2877v1 [math.GR].
[2] D. Gorenstein, Finite Groups, Chelsea Publishing Company, New York, 1980
[3] M. Kassabov and N. Nikolov, Words with few values in finite simple groups,
The Quarterly Journal of Mathematics, 64, 2013, 1161–1166.
Department of Mathematics, University of Brasilia, Brasilia-DF,
70910-900 Brazil
E-mail address: [email protected]
Department of Mathematics, University of Brasilia, Brasilia-DF,
70910-900 Brazil
E-mail address: [email protected]
| 4math.GR
|
CLAIRE: Combining Sets, Search and Rules
to Better Express Algorithms
Yves Caseau, François-Xavier Josset, François Laburthe
Bouygues SA – e-lab
1 avenue Eugène Freyssinet
78061 St-Quentin en Yvelines cedex, FRANCE
{ycs,fxjosset,flaburth}@challenger.bouygues.fr
Abstract. This paper presents a programming language that includes paradigms that are usually associated
with declarative languages, such as sets, rules and search, into an imperative (functional) language. Although
these paradigms are separately well-known and available under various programming environments, the
originality of the CLAIRE language comes from the tight integration, which yields interesting run-time
performances, and from the richness of this combination, which yields new ways to express complex algorithmic
patterns with few elegant lines. To achieve the opposite goals of a high abstraction level (conciseness and
readability) and run-time performance (CLAIRE is used as a C++ pre-processor), we have developed two kinds
of compiler technology. First, a pattern pre-processor handles iterations over both concrete and abstract sets
(data types and program fragments), in a completely user-extensible manner. Second, an inference compiler
transforms a set of logical rules into a set of functions (demons that are used through procedural attachment).
1. Introduction
The CLAIRE language [CL96a] is born from a research project that started 6 years ago which goal was to study
the relationship between combinatorial optimization and constraints. We decided to evaluate the competitive
advantages and reciprocal benefits of combinatorial optimization (CO) and constraint programming (CP)
techniques, and then to combine them into hybrid algorithms. We followed a systematic approach and looked at
problems such as Resource Allocation, Matching, Scheduling (job-shop, cumulative), Routing (TSP, VRP,
VRPTW) or Time-Tabling. Most real industrial optimization problems are a combination of these more classical
problems, thus another goal of this project was to identify a set of “building blocks” for optimization
applications. For each of these problems, we have produced a library of hybrid algorithms.
The combination of hybrid algorithms can only be made at the source level, using a “white box” approach. The
complexity and the inter-dependence between the various algorithms for propagation, computing lower bounds,
branching, etc. is such that an approach based on components (black boxes) does not seem realistic to us (today)
if one wants to take advantage of all the various techniques provided by the CO field. Thus, we have found the
need for a language to express our different algorithms that would be:
• simple and readable, for many reasons: because reuse of algorithms requires a readable expression of the
ideas, but also because CLAIRE was part of a research and teaching project and we needed an “executable
pseudo-code” to demonstrate algorithms;
• multi-paradigm (supporting logic, imperative and functional programming styles), with few simple and
well-understood concepts, such as objects, functions, rules and versioning for building search trees;
• compiled efficiently so that we could compare our algorithms with state-of-the-art approaches from teams in
the CO community who use C or FORTRAN as their implementation language. We decided to build
CLAIRE as a C++ preprocessor that would generate hand-quality C++ code (CLAIRE has also been ported
to Java).
The paper is organized as follows. The next section is a presentation of CLAIRE at a glance. CLAIRE is an
object-oriented functional language with parametric polymorphism. It is also a modeling language that uses
high-level abstractions such as sets or relations. Last, it includes an inference engine for object-based rules and a
few tools for building tree search algorithms. The third section presents the notion of “set programming” in
CLAIRE, which is derived from a set of techniques that make the use of sets as a programming paradigm both
natural and efficient. The fourth section shows how logical rules are compiled into functions. This is based on
previous work developed for the LAURE language [Cas94] (one of CLAIRE’s ancestors), so we focus on the
application of the technique to the CLAIRE logic and the impressive performance that we now obtain (since
CLAIRE functions are compiled into C++ code with virtually no overhead). The last section shows the
application of CLAIRE to writing optimization algorithms. We show how we can use the built-in features to
write tree search algorithms, and how the strategy of source-to-source optimization introduces new opportunities
for optimization through the notion of composition polymorphism.
1
2. CLAIRE at a Glance
We simply give an overview of the most salient features of the language. A complete description, including a
formal syntax for CLAIRE may be found in [CL96a].
2.1 A Simple Object-Oriented Data Model
We first briefly describe the CLAIRE object-oriented data model. Classes are defined with a list of typed slots,
and only single inheritance is supported. Each class is a first-class object of the system and is accessible at runtime (an optional list of instances is stored for each class). The <: operator is used to define a class (because
subclasses are subtypes in CLAIRE); it consists of appending a list of new slots to the slots of the parent class.
Slots support default values and unknown values, i.e., there is a special unknown object that represents the
absence of value. The following two examples illustrate class definition; the second one shows a parameterized
class (the usual Stack class, parameterized by the type of its elements).
Point <: object(x:integer = 0, y:integer = 0)
Stack<of> <: object(of:type, content:list = nil)
The second example is slightly surprising, compared to other approaches such as template classes, because it
implicitly rely on the fact that types are first-class objects (hence the type). Thus we attach to each stack the type
that must contain the objects that are pushed in the stack through the slot of. The responsibility of checking that
objects in the content of the stack belong to this type is left to the programmer. On the other hand, parameterized
signatures, as we shall see later, can express this dependency to support the strong typing of stack operations.
The CLAIRE data type complements the simplicity of the class hierarchy. A type is either a class, a constant set
of objects, the union or the intersection of two types, a parameterized class, an integer interval or a typed subset
of list, array or set. A parameterized class is the subset of the class such that the parameter belongs to a given
type. For instance, Stack[of:{integer,float}] is the set of stacks whose type parameter is integer or float.
A stack from this parameterized type will contains either integers (if its of parameter is integer) or floats, but
not both. A stack that contains either would be represented by the parameterized type Stack[of:(integer
∪ float)]
There are three kinds of primitive collections: arrays (constant length and ordered), lists (dynamic length,
ordered) and sets (dynamic length, no order and no duplicates). Primitive collection can either be typed, with a
read-only type attribute of, similar to the Stack example, or un-typed and immutable. This means that only
typed lists or sets support updates, which makes sense since we need a type to guarantee the type-safety of
updates. Immutable lists (or sets) receive a dynamic type when they are created.
Consequently, both list and set types come in two flavors: either strongly typed (list<t>, à la C++), or
supporting inclusion polymorphism (list[t]), i.e., list[t1] ≼ list[t2] ⇔ t1 ≼ t2. The type list<t>
contains typed lists l whose type parameter of(l) is precisely t. A list<t> object is a list that contains objects
that belong to t, including instances of subtypes of t.
On the other hand, the type list[t] contains both typed lists l whose type parameter of(l) is a subtype of t, as
well as immutable lists whose member belong to t. This applies equally to set<t> and set[t].
The distinction between these two kinds of list/set types is a key feature in CLAIRE. Since list<t> ≼ list[t],
we can write generic list methods, that apply both to mutable and dynamic, immutable lists. The type list[t]is
interesting because of inclusion polymorphism, and because read operations on lists of type list[t] can be
statically type-checked. On the other hand, list<t> is monomorphic (list<t1> ≼ list<t2> ⇒ t1 = t2), but
both read and write operations can be statically type-checked [Cas95]. Let us briefly illustrate this on an example.
Consider three classes A ≤ B ≤ C and the following methods:
foo(x:list[C], y:B) : void -> x[1] := y
bar(x:list<B>, y:B) : void -> x[1] := y
Although B is a subclass of C, the insertion in the foo method into the list x is not safe, since x may be a
list<A> object; on the other hand, baris safe but less flexible. Thus, the method foo relies on dynamic typing at
run-time to guarantee the safety of the update, whereas the method baris compiled more efficiently.
In addition, the tuple type is used for constant, heterogeneous, typed “array-like” (i.e., indexed) containers; for
instance, tuple(integer,integer,float) represents the set of tuples with exactly three elements: two integers
and a float. For example, tuples are used in CLAIRE to return multiple values from a function.
We can summarize the richness of type expressions with the following grammar (as usual, t[] is the type for
arrays of t (for instance, t = integer) and .. is used to define integer intervals):
2
<type> ::=
<class>
{<object> *}
(<type> ∪ <type>)
set〈<type>〉
set[<type>]
<type>[]
subtype[<type>]
|
|
|
|
|
|
<class>[<parameter>:<type> *]
(<integer> .. <integer>)
(<type> ∩ <type>)
list〈<type>〉
list[<type>]
tuple(<type> *)
|
|
|
|
|
|
Types are conceptually represented as expressions that denote sets of objects, and can be used everywhere in a
program since they are reified. Types as objects are themselves typed with the subtype[...] type; for example,
subtype[integer] contains all types that are included in integer. This set-oriented view is supported by the
fact that type subsumption is defined by the inclusion of the set extensions for all possible evolution of the
object database. Types form an inclusion lattice with syntactical greater lower bound and subsumption
(subtyping). The relationship between the type lattice and the powerset of the object domain is the basis for
using an abstract interpretation scheme [CP93].
CLAIRE inherits the SMALLTALK[GR83] strategy of considering everything as an object, from a class. This
implies that primitive types such as integer, floats, strings, Boolean are seen as primitive classes in CLAIRE.
2.2 Polymorphism and Advanced Typing
Methods in CLAIRE are simply overloaded functions, where each parameter contributes to the overloading (the
so-called multi-methods [CL94]) and each parameter can be typed with any CLAIRE type. For instance, attaching
a method to a class union is useful to palliate the absence of multiple inheritance. Suppose that a class A is a
natural candidate for multiple inheritance with parents B and C. If we chose B as the unique parent of A, we can
define the methods for C that we want to be inherited by A on the explicit union (A U C). This is not equivalent
because of the lack of extensibility (adding a new subclass to (A U C) is only possible through the addition of a
subclass to A or C). However, it is close enough for most practical cases.
Methods may be attached anywhere in the type lattice, not simply to the class hierarchy (more precisely,
anywhere in the lattice of type Cartesian products). This makes CLAIRE closer to a language like Haskell than
to C++. A method is called a restriction of the global function which itself is called a property.
It is important to notice that the class hierarchy is considered as a “law” given by the user (seen as set inclusion)
and is not based on type substitution. As a consequence, there are no constraints imposed on the user when
adding new methods. Covariance is enforced implicitly but not imposed explicitly (adding a new method on a
subdomain may augment the range of the function on a larger domain). If we define:
foo(x:integer) : (0 .. 5) -> ...
// ... stands for “some expression”
foo(x:(1 .. 10)) : (5 .. 10) -> ...
the range of foo on integer is actually the union (0 .. 10). The range in the first definition is taken as a type
restriction for this first definition, not for “foo on any integer” which is defined by the two restrictions.
Covariance comes as no surprise since we are using an overloading approach, where the property is the
“overloaded function” and the methods (i.e., the restrictions of the property for a given signature) are the
“ordinary functions” using Castagna’s terminology [Cas95].
The use of a type lattice does not prevent the occurrence of “inheritance conflicts” where two restrictions have
conflicting definition on overlapping signature. If one domain is included in the other it is assumed to take
precedence (such is the case for the foo example, so foo(3) uses the second definition). Otherwise, a warning is
issued. The lattice structure guarantees that it is possible to solve the inheritance conflict with a proper, more
specialized, definition, since the intersection of overlapping signatures can always be described with types.
Parameterized types can be used to produce parameterized signature. That means that we can use the value of a
parameter, when it is a type, inside the signature, as in the following example (for each operation op, x :op y
is syntactical sugar for x := (x op y)):
push(s:Stack<X>, y:X) : void -> s.content :add y
The use of the type variable X states that the second argument must belong to the type contained in the slot of
from the first argument (the stack). Note that there is an explicit difference between stack[of:{integer}] and
stack[of:subtype[integer]]. The first is the set of stacks with range parameter integer (exactly), whereas
the second can be thought of as the stacks of integers (stacks whose range parameter is a subtype of integer).
There is no parametric subtyping with the first (i.e. stack[of:{integer}] ≼ stack[of:{number}] is false) as
noticed in [DGL+95]. However, there is parametric subtyping with the second (i.e.
stack[of:subtype[integer]] ≼ stack[of:subtype[number]]). To alleviate notations, stack<integer> is
syntactical sugar for stack[of:{integer}]
3
Another important feature is the fact that CLAIRE supports higher-order functions. Properties, that are function
names, may be passed as arguments to a method and they can, in turn, be applied to functions. For instance, one
can write:
exp(f:property, x:any, n:integer) : integer
-> if (n = 0) f(x)
else f(exp(f, x, n - 1))
// exponentiation
This example will not be type-checked by the compiler and CLAIRE will rely on dynamic typing. On the other
hand, most methods are statically type-checked and the C++ code produced by the compiler is very close to
hand-written code (on classical benchmarks for procedural languages, using CLAIRE is similar to using C++, as
is shown in the appendix).
Last, second-order types can be attached to methods. A second-order type is a method annotation, defined by a
lambda abstraction, which represents the relationship between the type of the input arguments and the result.
More precisely, a second-order type is a function such that, if it is applied to any valid type tuple for the input
arguments, its result is a valid type for the result of the original function. Let us consider the two following
examples:
Identity(x:any) : type[x] -> x
top(s:Stack<x>) : type[x] -> last(s.content)
The first classical example states that Identity is its own second-order type. The second one asserts that the top
of a Stack always belong to the type of the Stack elements (of parameter). In the expression e that we introduce
with the type[e] construct, we can use the types of the input variables directly through the variables themselves,
we may also use the extra type variables that were introduced in a parameterized signature (as in the top
example) and we may use the basic CLAIRE operators on types such as ∪ (type union), ∩ (type intersection) and
some useful operations such as member. member is used in particular to describe the range of the enumeration
method set!. This method returns a set, whose members belong to the input class c by definition. Thus, we
know that they must belong to the type member(X) for any type X to which c belongs (If c is a type, member(c) is
the minimal type that contains all possible members of c). This translates into the following definition:
set!(c:class) : type[set[member(c)]] -> instances(c)
In the same way that overloading becomes more powerful with a richer type system, we have introduced tables
to represent binary relations. Tables extend traditional arrays by allowing the domain (and the range) to be any
type. When the domain is not an integer interval (the usual case for arrays) it is implemented through hashing.
The following examples illustrate CLAIRE tables:
square[x:(0 .. 20)] : integer := x * x
dist[x:(0 .. 10), y:(0 .. 10)] : float := 0.0
creator[c:class] : string := “who created that class”
color[s:{car, house, table}] : set<colors> := {}
married[t:tuple(person,person)] : tuple(date,place) U {nil} := nil
The combination of objects, relations (with explicit and implicit inverses), heterogeneous sets and lists makes
CLAIRE a good language for knowledge representation. For instance, entity-relationship modeling is
straightforward.
2.3 Set and Logic Programming
Sets play an important role in CLAIRE. We saw that each type represents a set and may be used as such. A set is
used through testing membership, iteration, or combining with another set and a set operation. Membership is
tested using the operation ∈ (the ASCII representation is ‘%’). Iteration is performed with the for x in S e(x)
control structure. We have also introduced two convenient expressions:
• exists(x in S | P(x)) returns true if there exists an x in S such that P(x) is true, and false otherwise
• some(x in S | P(x)) returns an x from S such that P(x) is true if one exists, and unknown else (as in
ALMA [AS97]).
For instance (refer to [CL96a] for further information), we may write:
x ∈ (string U list)
for y in (1 .. 10) U (20 .. 30)
(if exists(c in (1 .. 10) | ok?(c + y))
when x := some(t in TASK | completed?(t)) in
register(x)
else error(“no completed task was found”)
choose(y))
// t was found
4
Although primitive data types (such as string or list) may be considered as sets, they are seen as infinite sets,
thus trying to enumerate them will yield an error.
New sets can be formed through selection (e.g., {x in S | P(x)}) or set image (e.g., {f(x) | x in S})
[SDD+86]. When duplicates should not be removed, a list image can be formed with list{f(x) | x in S}.
Using a straightforward syntax leads to straightforward examples:
{x in Person | x.age > 18}
exists(p in {x.father | x in (man U woman)} | p.age < 20)
list{x.age | x in Person}
for x in {f(y) | y in (1 .. 10)}
print(x)
sum(list{y.salary | y in {y in Person | y.dept = sales}})
In addition to sets, CLAIRE uses a fragment of an object-oriented logic to support inference rules associated to
events, which will be presented in Section 4. This logic may be seen as an object-oriented extension of binary
DATALOG with object methods, set expressions and negation. The following expression is the logic part of a
CLAIRE rule:
edge(x,y) | exists(z, edge(x,z) & path(z,y))
=>
x.path :add y
The first extension is the introduction of interpreted functions, which are represented by object methods, with a
model similar to [KW89]. These methods are either comparison methods or binary operations. For computing
x.tax from x.salary, we may write:
exists(z, z = x.salary & z > 30000 & z < 50000 & y = 3000 + (z - 30000) * 0.28 )
=> x.tax := y
The second extension is the introduction of set expressions (à la LDL [NT89]). Here is a simple example that
uses a set expression to represent the set of children who are still minors:
z = 2 + size({y in Person | y ∈ x.children & y.age <= 18})
=> x.family_size := z
The last extension is the introduction of negation (using not(P(x,y))). Although this object-oriented logic
supports a top-down complete resolution strategy (with stratified negation), it is only used in CLAIRE to
provide forward-chaining inference rules. A rule is defined by linking a logical condition to a conclusion, which
is any CLAIRE expression. Here is a simple example:
compute_salary(x:Person[age:(18 .. 40)]) :: rule(
x.salary < average_salary({y in Person | x.dept = y.dept})
=> x.next_salary := x.salary * 1.1)
To complete this brief overview of CLAIRE, we need to mention the “versioning” mechanism. A version (also
called a world, using the AI terminology) is a virtual copy of the state of the objects. The goal is to be able to
return to a previously stored state if a hypothetical branch fails during problem solving. As a consequence,
worlds in CLAIRE are organized into a stack and only two operations are allowed: one for copying the current
state of the database and another for returning to the previous state. The part of the objects that supports these
defeasible updates is defined by the programmer and may include slots, tables, global variables or specific data
structures such as lists or arrays (defeasible is an AI term that means “backtrack-able”).
Each time we ask CLAIRE to create a new world, CLAIRE saves the status of defeasible slots (and tables,
variables, ...). Worlds are represented with numbers, and various operations are provided to query the current
world, create and return to previous worlds, with or without “forgetting” the most recent updates. The use of
worlds is also encapsulated with the branch(e) control structure which creates a world, evaluates the expression
e and returns true if the result of e is true, and otherwise backtracks to the previous world and returns false.
Using these programming features makes writing search algorithms, such as branch-and-bound, much simpler in
CLAIRE (or another language with search primitives such as ALMA [AS97]) than with traditional imperative
languages such as Java or C++. For instance, here is a simple program fragment that solves the “n queens”
problem:
D :: (1 .. SIZE)
column[n:D] : D := unknown
possible[x:D, y:D] : boolean := true
event(column)
store(column, possible)
r1(x:D, y:D) :: rule(exists(z:D, column[z] = y)
=>
possible[x,y] := false)
r2(x:D, y:D) :: rule(exists(z:D, column[z] + z = y + x)
=>
possible[x,y] := false)
r3(x:D, y:D) :: rule(exists(z:D, column[z] - z = y - x)
=>
possible[x,y] := false)
5
queens(n:(0 .. SIZE)) : boolean
-> (if (n = 0) true
else exists(p in D |
(possible[n,p] & branch( (column[n] := p, queens(n - 1)) )))
In this program, queens(n) returns true if it is possible to place n queens. The search tree (these embedded
choices) is represented by the stack of the recursive calls to the method queens. At each level of the tree, each
time a decision is made (an assignment to the column table), a new world is created so that we can backtrack
whether this hypothesis (this branch of the tree) leads to a failure.
3. Set-Based Programming
3.1 Sets in CLAIRE
We gave a brief overview of the use of sets in CLAIRE in the previous section. However, there is more to the
concept of “Set-Based Programming”. For us, it comes from the combination of two principles:
• Sets are "first-class citizens" with a choice of multiple representations (abstraction);
• Abstract sets may be used to capture "design patterns" (in the sense of the Object-Oriented Programming
community, that is, a motif that repeats itself in a program because it represents an abstract operation).
The first aspect in CLAIRE is illustrated by the variety of data structures that are available to represent
collections of objects. These include classes, arrays, lists, sorted collections or bit-vectors. Also, we already
mentioned that any type can be used as a collection, as illustrated in the following example:
for x in person[age:(15 .. 18)]
print(x)
Moreover, new collection classes can be introduced. In the following example, we introduce collections that are
implemented with a hash table (hashing an object to get an index in a list is a standard method).
Hset<of> <: collection(of:type, content:list, index:integer)
add(s:Hset<X>, y:X) : void
set!(s:Hset) : set
->
->
let i := hash(s.content, y) in ( ... )
{x in s.content | known?(x)}
In addition to concrete sets, CLAIRE supports set expressions, which can be used freely since they are built
lazily by the compiler so that any useless allocation and construction is avoided, as we shall see later. We
already mentioned set construction by image or selection. New set expressions can also be introduced through
the concept of pattern. A pattern is a function call that can be lazily dealt with. The first step is to define a new
set operator, such as:
but(x:collection,y:any) : set
->
// (x but y) ⇔ x \ {y}
{z in x | z != y}
Now we can treat the pattern (S but x) lazily without building a concrete data structure that would represent it.
The pattern type F[tuple(X,Y)] is the set of function calls with F as the selector (function) and two arguments
(stored in the args list) of respective types X and Y. For instance, we can declare how to compute membership on
the but pattern so that the compilation of the expression (x ∈ (Person but john)) will be performed by code
substitution into (x ∈ Person & x != john).
CLAIRE supports true abstraction for collections since one may substitute a representation by another and not
have to change a line of code in generic methods, such as the following example:
sum(s:subtype[integer]) : integer
=> let d := 0 in (for x in s
// sum is an inline method (=> is the syntactic marker for inline)
d :+ x, d)
On the other hand, re-compiling will be necessary since the efficient implementation of abstraction relies on
source-to-source transformation.
3.2 Customizable Iteration
One of the most important aspects of set programming from a practical perspective is the ability to customize set
iterations. Iteration plays an important role in CLAIRE, either in an explicit form (using the for control structure)
or in an implicit form (an iteration is implied in the image/selection set expressions as well as in the some/exists
structures) as shown in these examples:
{x in (1 .. 10) | f(x) > 0}
{size(c) | c in (class but class)}
// note: class is a class
For each of these iterations, the compiler uses source-to-source optimization and generates the equivalent, but
faster, expressions. The first one is compiled into:
let s := {}, x := 1 in
(while (x <= 10)
(if (f(x) > 0)
s :add x,
x :+ 1),
s)
6
Since the set of collection classes is extensible, the compiler supports the extension of the iteration mechanism
through the definition of the iterate method. For any expression e(v) that contains a variable v, iterate(x,v,e) is
an expression that produces e(v) to be evaluated for all values v in the set x. Here what we would write if x is a
Hset:
iterate(x:Hset, v:Variable, e:any) : any
=>
for v in x.content
(if known?(v)
e)
This is also true for patterns, thus we can explain how the iteration of a pattern should be compiled lazily
without building the set:
iterate(x:but[tuple(collection,any)],v:Variable,e:any) : any
=> for v in args(x)[1] (if (v != args(x)[2]) e)
This mechanism has also been used to implement the lazy iteration of image or selection sets. As a consequence,
an expression such as:
for x in {f(x,0) | x in (myHset but 12)}
print(x)
will be transformed into:
for x in myHset.content
(if (known?(x) & x != 12)
print(f(x,0))
As a consequence, it is both natural and efficient in CLAIRE to use “virtual” sets that are iterated as a
programming pattern. This yields programs that are easier to read since we separate the filtering of the data from
the processing. Note that the iterate optimizing pattern does not carry a type since the type of an iteration will be
determined dynamically after the substitution as occurred.
3.3 Sets and Design Patterns
This section illustrates the previous feature through two examples of using sets as a design and programming
pattern. The first example is the concept of embedded (doubly) linked lists. Linked lists are very common and
useful objects, since they provide insertion and deletion in constant time. The implementation provided by most
libraries is based on cells with pointers to next/previous cells in the list. This is useful but has the inconvenience
of requiring dynamic memory allocation. When performance is critical, most programmer use embedded linked
list, that are implemented by adding the two next/previous slots to the objects that must be chained. This only
works if few lists (e.g., one) need to be kept, since we need a pair of slots for each kind of list, but it has the
double advantage of fewer memory access (twice fewer) and no dynamic allocation as the list grows. For
instance, if we have objects representing tasks and if we need a list of tasks that have to be scheduled, we may
define the class Task as follows:
Task <: object( ..., next:Task, prev:Task, ...)
The obvious drawback is the loss of readability, since the concept of linked list is diluted into pointer chasing. In
CLAIRE, we can use a pattern to represent the concept of a linked chain using these two slots as follows. We
first define chain(x) which is the list of tasks obtained by following the next pointer:
chain(x:Task) : list<Task>
-> let l := list<Task>(x) in
(while known?(next, x)
// x is the first member
(x := x.next, l :add x), l)
Note that the list constructor carries the type of the new list when we used strongly typed lists.
list<integer>(1,2,3) is the list with 3 element (1,2 and 3) whose member type is integer.
We define the iteration of the chain pattern in a very similar way:
iterate(x:chain[tuple(Task)], v:Variable, e:any) : any
=> let v := args(x)[1] in
(while true (e,
// (a,..,c) is a sequence of exprs
if known?(next, v) v := v.next else break()))
We can now use the “abstract” chains as any other collection in CLAIRE:
sum({weight(x) | x in chain(t0)})
count(chain(t1))
The interest of this approach is to keep the benefit of abstraction at no cost (we do not see the pointer slots
anywhere and we can later replace embedded linked list with another representation). The code produced by the
optimizer is precisely the pointer-chasing loop that we would have written in the first place.
The second example shows the power of combining patterns with the concept of iterators. Iterators are a very
popular solution for iterating collection in most Java and C++ libraries. However, for most collection types they
induce a needless overhead (thus, we do not use them in CLAIRE). However, there are more complex data
structures for which iteration can be seen as traversing the structure, and for which the iterator embodies the
traversal strategy. An iterator is simply defined by two methods, start and next, which respectively provide the
7
entry point for the iteration and the succession relationship. The following is an example of a tree iterator. We
introduce a pattern (T by I) which produces the set of nodes in T using the traversing strategy of the iterator I:
Tree <: object(value:any, right:Tree, left:Tree)
TreeIterator <: object(tosee:list, status:boolean)
iterate(x:by[tuple(Tree,TreeIterator)], v:Variable, e:any) : any
=> let v := start(args(x)[2], args(x)[1]) in
while (v != unknown) (e, v := next(args(x)[2], args(x)[1]))
We can define many types of iterators corresponding to the various popular strategies for iterating a tree (DepthFirst Search, Breadth-First Search, etc.):
TreeIteratorDFS <: TreeIterator()
// Depth-First Search strategy
DFS :: TreeIteratorDFS()
start(x:TreeIteratorDFS, y:Tree) -> ...
next(x:TreeIteratorDFS, y:Tree) -> ...
We may then use the pattern (T
by I)
as any other collection and simply write:
sum({y.weight | y in (myTree by DFS)})
which associated optimized code is the following:
let d := 0, y := start(DFS, myTree) in
(while (y != unknown) (d :+ y.weight, v := next(DFS, myTree)), d)
The iterator object plays two roles: its type will determine which sort of tree traversal is required and it also
contains the state of the traversal (the tosee and status slots) that is exploited by the start and next method. Thus,
if the same iterator may be used on different trees, the simultaneous iteration of many trees will require many
iterators.
4. Rule-Based Programming
4.1 Claire Rules
4.1.1 A Logical Language for Assertions
A rule in CLAIRE is obtained by associating a logical condition to an expression. Each time the condition
becomes true for some objects, the expression will be evaluated for these objects (represented by the rule’s free
variables). A condition is defined with a logical language that includes predicate logic to describe the state of
one or many objects and event logic to capture the occurrence of a precise event.
To define a logical condition, we use the logic language defined by the following grammar.
<assertion> ≡ <expression> <comp> <expression>
<variable> :: <class>
|
<value> := < <expression> | (<expression> <- <variable>) >
|
exists(<variable>, <assertion>)
|
|
not(<assertion>)
|
if (<expression> <comp> <expression>) <assertion> else <assertion>
<assertion> & <assertion>
<expression>
≡
<variable>
|
<entity>
|
|
|
<assertion> | <assertion>
<function>(<expression>)
<expression> <operation> <expression>
|
{<var> in <expression> | <assertion>}
|
|
list{<var> in <expression> | <assertion>}
≡
<value>
<rule>
≡
<table>[<expression>]
|
<expression>.<property>
<name>(〈<variable>:<type>〉+) :: rule(<assertion> => <statement>)
The value of a logical expression is a CLAIRE entity. Logical assertions are made by combining expressions. The
most common type of assertion is obtained by comparing two expressions with a comparison operation. A
comparison operation is an operation that returns a Boolean value.
The novelty in CLAIRE 3.0 is the introduction of event logic. There are two events that can be matched
precisely: the update of a slot or a table, and the instantiation of a class. The expression x.r := y is an event
expression that says both that x.r = y and that the last event is actually the update of x.r from a previous value.
Thus, the introduction of event expressions serves two purposes:
8
•
control the triggering of a rule more precisely by telling which event should be considered. Consider the
following example:
x.age = y
&
z.age = 2 * y
This logical expression may become true (and trigger an associated rule) from an update of the age slot in
two manners, either because x.age has changed or because z.age has changed. In most cases, this behavior
is precisely what is expected. In contrast, the following expression :
x.age := y & z.age = 2 * y
•
adds to the previous expression the explicit constraint that the triggering update is applied to the object x.
use the “previous value” of the updated expression x.r in the logic expression. This is achieved with the
expression x.r := (y <- z) which says that the last event is the update of x.r from z to y. For instance,
here is the event expression that states that x.salary crossed the 100000 limit:
x.salary := (z <- y) & y < 100000 & z >= 100000
The other event that can be described with the event logic is the instantiation of a class. The expression x :: C
means that x is a new instance of the class C. As previously, the event expression is the combination of the
statement x ∈ C and the additional constraint that the last event is precisely the creation of the new instance x of
the class C. We use object instantiation to represent external events when writing an application that must be
integrated with other components using a publish & subscribe event model, such as the Java Bean model
[Eng97].
The value of a logical assertion is always a Boolean, thus logical assertions can be combined with & (and) and |
(or). In addition to pre-defined operations such as <= or +, it is possible to use new properties inside logical
assertions but they need to be described using the description table and a set of keywords including comparison,
binary_operation, monoid, group_operation. By declaring the algebraic characterization of a property we are
telling the rewriting system that produces the algebraic representation how to handle the logic expression.
To define a rule, we define a list of free – universally quantified – variables, that are introduced as parameters of
the rule, a condition, which is given as an assertion using the previously defined variables and a conclusion that
is preceded by =>. Here is a classical transitive closure example:
r1(x:Person, y:Person) :: rule(
exists(z, x ∈ z.friends & z ∈ y.friends)
=> y.friends :add x )
4.1.2 Event-Based Rules
Rules are checked (efficiently as we shall later see) each time that an event occurs. Events come in two kinds.
The first kinds of events are updates to slots or tables that are declared as event generating with the event
statement.
<events definition>
≡
*
< event | noevent >( < <table> | <property> > )
The second kinds are instantiations of classes. These include the instantiation of “event objects” classes that
represent “external events” used in a larger-scale application, but also the instantiation of local classes. Note that
the “instantiation event” occurs after the initialization of the object (with default or given values for the slots).
A rule considers as events all slots or tables that have been declared as such before the rule has been declared, so
the event declaration needs to precede the rule declaration, in order to create the “demons” that will watch over
the given relation and fire the rule when needed. The declaration noevent(...) may be used to prevent
explicitly a rule to react to some relations that were declared previously as events for other rules. By spreading
the event declarations (and occasionally noevent declarations) before and between the rules, one has a precise
control over the triggering of the rules. This level of control is further refined with the use of event expressions
within the condition of the rule.
Updates that are considered as events are:
• x.r := y, where r is a slot of x and event(r) has been declared.
• a[x] := y, where a is a table and event(a) has been declared.
• x.r :add y, where r is a multi-valued slot of x (with range bag) and event(r) has been declared.
• a[x] :add y, where a is a multi-valued table and event(a) has been declared.
• The instantiation of a class C, whether static (x :: C(…)) or dynamic (new(C,…)).
The handling of mono- vs. multi- valuation is an important aspect of this logic. Multi-valued slots in CLAIRE
are slots that contains multiple values (i.e., friends is multi-valued since I may have many friends). A multivalued slot is implemented using a container (a list or a set), however, from a semantic point of view, we are
only interested with the multi-valued binary relationship between a person and her friends. Operations on a
multi-valued slots are the addition or the deletion of a new member, whereas a mono-valued slots only supports
9
the update the value (replacing a value by another one). Thus, CLAIRE makes a difference between a multivalued slot (for instance, a slot friends with range set<person>) and a mono-valued slot with range set (we
can use the same example). The management of inverses is different: in the first case, the inverse of friends is a
relation from Person to Person; in the second case, it is a relation from set<Person> to Person. Similarly, the
atomic events are different: in the first case it is the addition of a new friend; in the second case, it is the
replacement of a set of friend by another set. It is our experience that both modeling options are needed to
capture a large set of situations.
4.1.3 Examples
The following first example is taken from our cumulative scheduling program presented in [CL96b]. Cumulative
scheduling is a special planning problem where some tasks have to be scheduled on resources. A resource may
handle several tasks at the same time, but is limited by a finite capacity (capacity). The problem is to perform
all tasks such that the total completion time is minimal. The amount by which a task “consumes” a resource must
fit between two bounds (minUse/maxUse).
The rule eject-up (see below) involves two data structures: Task and Interval. For each Task t, we store (in
slots) the required resource t.res, time bounds t.earliestStart and t.latestEnd, the associated work t.work,
and its extremal resource use levels minUse/maxUse. For each Interval S, we store its bounds (Left/Right), the
common resource of all tasks in the interval (Res), and the sum of the work amount of tasks (Work) in the
Interval.
The rule considers a task t and an interval S. Let A be the earliest start time of the lower bound of S, B the latest
end time of the upper bound of S, and a the earliest start time of t. Whenever t and S are such that:
• they share some common resource r that is actually cumulative (i.e. r.capacity > 1),
• S is not an empty Interval, but t is not included in S,
• t cannot start before A and end before B ((t.work – max(0,A - a)) * t.maxUse > slack), because there
is not enough slack, even with the maximal consumption level for t, in order to perform as much as possible
of t before A.
• t cannot start before A and end after B ((B - A) * t.minUse > slack) because there is not enough slack,
even with the minimal consumption level for t.
In such a case, the rule deduces that t necessarily starts after A, and analyses the amount of t that can be
performed before B in order to update the time bounds of t. We use two update methods
increaseEarliestStart and increaseLatestStart in order to maintain the Interval structure.
[eject-up(t:task, S:Interval) ::
exists(r:resource, S.Res = r & t.res = r & r.capacity > 1 &
not(empty(S)) & not(contains(S,t)) &
exists(a, a = t.earliestStart &
exists(A, A = S.Left.earliestStart &
exists(B, B = S.Right.latestEnd &
exists(slack, slack = (B - A) * r.capacity – S.Work &
(t.work – max(0,A - a)) * t.maxUse > slack &
(B - A) * t.minUse > slack )))))
=> increaseEarliestStart(t, B - (slack div+ t.minUse)),
increaseLatestEnd(t, B - (slack div+ t.maxUse) + t.minDuration) ]
This rule is triggered whenever a time bound earliestStart/latestEnd or the contains relation is modified.
There are multiple propagation patterns and writing the equivalent procedural code is really a tedious and errorprone task.
We will now present a few examples that rely on the event-based logic that we have introduced. First, here are
three (simplified) examples taken from the field of constraint propagation. The first one tells that if two tasks are
consecutive (attached means without idle time in between), we can propagate the increase in minStart (earliest
starting time) directly. The second tells how to propagate a duration increase.
R1(x:task, y:task, z:integer, u:integer) :: rule(
x.attached = y & x.minstart := (z <- u) =>
y.minstart :+ (z – u) )
R2(x:task, I:taskInterval, z:integer, u:integer) :: rule(
x.minStart >= I.minstart & x.maxEnd <= I.maxEnd & x.duration := (z <- u)
=> I.duration :+ (z – u) )
R3(x:task, I:taskInterval, z:integer, u:integer) :: rule(
x.minStart := (z <- u) & z >= I.minstart & u < I.minStart & x.maxEnd <= I.maxEnd
=> I.duration :+ x.duration )
10
The first two rules are similar and illustrate the use of the event pattern to capture a transition and propagate an
incremental change. The second rule also shows that the event pattern forces the rule to trigger when a change is
made on the duration attribute, which is what we intend here, since a different rule (R3) handles the case when a
change is made on the minStart attribute. In R3, the event pattern is used to capture the state transition (x now
belongs to the task interval and did not before).
The next two examples illustrate the use of event-rules to model state-transition systems. They are taken from an
“artificial life” scenario where units move from cell to cell. The first rule says that if a unit moves to a position
where another unit is already installed, a conflict occurs which will result into a failure of the incoming unit if its
strength is less than the strength of the residing unit. The second re-computes the visibility graph for a unit that
gets “out-of-the-woods”.
RS1(x:unit, y:unit) :: rule(
x.location := y.location & x.strength <= y.strength
=>
failure(x))
RS2(x:unit, z:position, u:position) :: rule(
x.location := (z <- u) & u.forest? = true & z.forest? = false
=> computeVisibility(x) )
The use of event logic enables us to represent the dissymmetry between x and y in the first rule, which would not
be possible with sole predicate logic. The second rule is another example of using the event pattern to capture a
state transition (from being in-the-woods to out-of-the-woods). The use of “?” after forest has no special
meaning in CLAIRE, it is a common practice to add “?” to names of slots or methods that represent a Boolean.
4.2 Algebraic Representation of Rules
4.2.1 Principles of Algebraic Rules
CLAIRE rules are compiled into procedural demons in three phases (cf. Figure 1). First, the condition part of the
rule is rewritten into an algebraic expression, which is next differentiated into many functions that are finally
transformed into if-write demons. The process of rule compilation relies on the use of a relational algebra. This
algebra A(R) contains relational terms (elements of R) that represent binary relations; it is generated from a set
of constants (Cartesian products of types), of variables (CLAIRE slots and tables) and a set of operators (similar
to [McL81]). These operators can define unions, intersections and compositions over relational terms, but also
more advanced algebraic functions such as the composition of a binary operation or a comparison operation with
two binary relations, as we shall see in the next section.
Logical
Rules
Rewriting
Algebraic
Rules
Differentiation
if-write
Demons
Propagation
Formulae
Compiling
Figure 1. Compilation scheme of CLAIRE rules.
We will briefly describe each step of the compilation process, and follow the transformation of a rule, from its
logical expression towards the generated procedural demons. Here, we use the closure rule, which performs a
transitive closure over a graph, described through the binary relations edge and path:
closure(x:point, y:point) :: rule(
edge(x,y) | exists(z, edge(x,z) & path(z,y))
•
=>
x.path :add y )
The first phase of the compilation process is the transformation of the conditional part of the rule into a
relational formula that belongs to A(R). Translation into the algebraic form is based on rewriting and
involves a lot of knowledge about object methods. The principle is to solve the equation assertion(x,y),
while considering that x is known and that y is sought. The result is the relational algorithm that explains
how to get y from x, which is represented as a term in A(R). The conditional part or the rule closure is
transformed into the following algebraic term t1:
t1 = edge ∪ (path ο edge)
where ∪ and ο are respectively the union operator and the composition.
11
•
Next, a phase of symbolic differentiation [SZ88] computes the effect of adding a pair of values into a table
or a slot. The differentiation is a function that is applied to the algebraic term t, with respect to each relation
R occurring in the condition, and that produces a term ∂t/∂R in A(R∪{ , }). and are two constant
functions that represent respectively the empty function ( : (x,y) → ∅) and the identity function ( : (x,y)
→ {(x,y)}). A set of differentiation rules (one for each operation on binary relations) is used to compute the
term ∂t/∂R. The term t1 computed above is differentiated with respect to the relation edge and yields the
following function f1:
f1 =
∂t1/∂edge
=
∪ (path ο
)
or in other terms:
f1 : (a,b)
→ {(a,b)} ∪ {(a,y) | (b,y) ∈ [path]}
•
The last phase of the compilation of a rule consists of associating the set of functions (derived from each
relational term occurring in the condition) and the CLAIRE expression in the conclusion, so as to build one
if-write demon per function. The demon is used to propagate each update so that rules are triggered in an
incremental manner, yielding an approach similar to [SKG87].
We now examine these different phases with more details in the following sections.
4.2.2 A Relational Algebra
The relational algebra contains terms that represent binary relations. It is generated from a set of constants
(cartesian products of types), a set of variables (the relations that are described with rules) and a set of relational
operators. This set may be described as follows :
• ∪, ∩ , o are the usual operators for union, intersection and composition (binary join).
• ψ[+](r1, r2) represents the composition of a binary operation (+) with two binary relations r1 and r2.
• φ[<](r1,r2) is the composition of a comparison operation (<) with two binary relations r1 and r2.
• if[φ[<](r1,r2)](r3,r4) is the conditional composition of a test ( φ[<](r1,r2) ) with two binary relations r3 and r4
which represent the two branches of the conditional (r3 if the test succeeds and r4 otherwise).
• {z:r1, r2} represents a relation that associates to an object the set of objects z bound through r1 which also
“satisfy” r2 (hence, r2 is a filtering relation - a φ or a composition of φ’s). This operator is introduced to
handle set creation in logical formulae.
• Set_expansion(r), written r*, is the reciprocal operator which is used to treat a relation r with range set as a
multi-valued relation.
Here is the formal definition of the relational operators:
(x,y) ∈ r1 ∪ r2 ⇔ (x,y) ∈ r1 ∨ (x,y) ∈ r2
(x,y) ∈ r1 ∩ r2 ⇔ (x,y) ∈ r1 ∧ (x,y) ∈ r2
(x,y) ∈ r1 o r2 ⇔ ∃ z, (x,z) ∈ r2 ∧ (z,y) ∈ r1
(x,y) ∈ ψ[+](r1, r2) ⇔ ∃ z1, z2, (x,z1)∈ r1 , (x,z2)∈ r2 ∧ y = z1 + z2
(x,y) ∈ φ[<](r1, r2) ⇔ x = y ∧ ∃ z1, z2, (x,z1)∈ r1, (x,z2)∈ r2 ∧ z1 < z2
(x,y) ∈ if[φ[<](r1,r2)](r3,r4) ⇔
{(x,y) ∈ r3 ∧ ∃ z1, z2, (x,z1)∈ r1, (x,z2)∈ r2 ∧ z1 < z2 } ∨
{(x,y) ∈ r4 ∧ ∃ z1, z2, (x,z1)∈ r1, (x,z2)∈ r2 ∧ ¬(z1 < z2 )}
(x,y) ∈ {z:r1 , r2} ⇔ y = {z | (x,z) ∈ r1 ∧ (z,z)∈ r2}
(x,y) ∈ r* ⇔ ∃ z, (x,z) ∈ r ∧ y ∈ z
To achieve the same expressive power as the logical fragment that we use to define conditions on objects, we
need to introduce the ability to share a sub-term, which will correspond to the introduction of existential
variables. Let us now define A(R) the term algebra generated by the previous set of operators, the constant
cartesian products of types (such as person × {0,1,2}) and the set R of relational variables (that represent
relations from CLAIRE, which are tables or properties). We extend our definition as follows:
• If x is a relation variable that does not belong to R, if t2(R1,...,Rn,x) is a term from A(R ∪ {x}) and
t1(R1,..,Rm) is a term from A(R), then λ(x:t1).t2 is a term from A(R) which represent the relation denoted by
t2 in an new “object system” where x is bound to the relation denoted by t1.
Notice that an “object system” is simply defined by the valuation function q that associates to each relation name
in R the actual binary relation in the current system. We write [t ]q the binary relation denoted by the term t for a
given valuation q. A key result in [Cas94] is that the logic fragment that we use and the relational algebra A(R)
have the same expressive power. This means that for any condition C with at most two free variables x and y
there exists a term t of the algebra such that C(x,y) ⇔ (x,y) ∈ [t ]q . The term is not necessarily unique and the
difficulty is to produce an “optimized” term, which represents the straightest relational computation that is
necessary to produce the set of y starting from the value of x.
12
The transformation of a logic rule into its algebraic representation is based on rewriting, as explained in [Cas94].
This is achieved with rewriting rules that incorporate the algebraic knowledge about objects’methods such as:
group(+,0,-) & x + y = z ⇒ x = z - y
monoid(*,|,/) & x * y = z ⇒ (y | z) & (x = z / y )
The first rule states that if we have a Group (+,0,-) the equation x + y = z can be rewritten into x = z – y where –
is the inverse operation of + (thus, this applies to matrices as well as integers). The second rule states how to
rewrite the same equation for a weaker structure: here * has no complete inverse operation but a partial one, /,
associated by an inversibility relation – divisibility.
In the previous years, we have added the concept of simplification rules, which implement optimizing strategies
(replacing an algebraic term by another one with the same denotation but which is more efficient to compute) in
a manner similar to the optimization of SQL queries. The rules are applied recursively while algebraic terms are
being built. They are mostly based on three ideas:
(1) filtering terms should be ordered from simpler to more complex, and applied as soon as possible from an
evaluation order point of view,
(2) constants sub-terms must be recognized, simplified, and re-organized. For instance, linear sub-term
factorization (2 * x + 3 * x ⇒ 5 * x) is applied through this simplification step.
(3) inversible sub-terms must be detected because they can be used (in their inverse form) to avoid costly
joins.
The motivation for this translation is that we are able to represent propagation as a formal computation over
algebraic terms, called differentiation. Suppose that t is a term for A(R), Ri a relation variable from R and q a
valuation function representing the current object system. We now consider an atomic update, that is the
addition of a pair (x,y) to the relation q(Ri). The term t now represents a new relation in the new object system q’.
The goal of differentiation is to compute [t]q' - [t]q , that is the set of object pairs that “appeared” in the relation
represented by t “because” of the update (since we use a monotonic algebra, we have [t]q' ⊃ [t]q ).
Let us define the functional algebra F(R) as A(R ∪ {
}), where each relation of R is mapped to a constant
function. We use two constants
that represent respectively the “empty function”
((x,y) → ∅ ) and the unit function ((x,y) → {(x,y)}). To represent the propagation of an update in a set-valued
expression, we also need to extend F(R) with a new projection operation π(t) defined as follows.
(x,y) ∈ [π(t)(a,b)]q ⇔ x = y ∧ ∃ z, (x,z) ∈ [t(a,b)]q
Thus, π(t) is a the first projection of the relation represented by t.
Differentiation is a formal operation on A(R) which associates to a pair (t,Ri) ∈ A(R) × R a term ∂t/∂Ri from
another functional algebra F(R) (which represents a function that associates a binary relation to an object pair)
such that :
[t]q' - [t]q ⊆ [∂t/∂Ri(a,b)]q' ⊆ [t]q'
(1)
To guarantee the interest of differentiation we also impose that ∂t/∂Ri(a,b) is the smallest term from F(R) which
satisfies (1) according to the partial order induced by inclusion. However, (1) is a necessary and sufficient
condition to prove the validity of using differentiation to model propagation.
4.2.3 Extending the Algebra to Handle Event-based Logic
The algebra is extended with two operators:
• χ(R) is an “event-filter” : it represents the same relation as R, but it will behave differently when the term is
differentiated.
• R’ represents the “previous” version of the relation R before an atomic update. It cannot be used freely in
the relational algebra, which would require a much more complex implementation, but only in the following
pattern:
λ(z:R’).t{χ(R)}
where t{χ(R)} is a term which contains χ(R) as a sub-term.
The differentiation rules are straightforward:
∂χ(R) /∂R =
, ∂χ(R’) /∂R =
∂χ(λ(z:R’).t{χ(R)}) /∂R = λ(z:R’). ∂t /∂R
The compilation rules (cf. [Cas94]) are two methods, expR and expF, that represent the semantic of the operators
through code generation. Both are defined by structural induction for each operator.
• expR(t,x,y,e(y) ) is a code fragment that applies the expression e(y) to each object y that is linked to x
according to the binary relation represented by the term t. Here is a simple example for a binary composition:
expR(t1 o t2, x, y, e(y)) = expR(t2, x, z, expR(t1,z,y,e(y)))
13
•
expF(∂t/∂R,x,y,u,v,e(u,v) ) is a code fragment that applies the expression e(u,v) for each pair (u,v) that
belongs to the relation which is the value of [∂t/∂R](a,b), where we recall that ∂t/∂R is a term of a
functional algebra that represent a function that associates a binary relation to each pair of objects. Here is
another example for compiling the composition of a derivative and a relation:
expF(∂t/∂R o t ,x,y,u,v,e(u,v) ) = expR(∂t/∂R ,x,y,a,v, expR((t)-1, a,u,e(u,v))
This last example shows that although differentiation rules are simple and intuitive, their actual implementation
through code generation may be more subtle. The ability to inverse any term from the algebra is not easy to
reach, although many terms may be inversed through recursive formulas. For the few but annoying exceptions,
we actually revert to the logical expression and re-apply a translation step (rewriting) while changing the
variable ordering (each translation step takes one free variable as the “base” and the other one as the target). To
achieve this, we need to ensure that for each composition sub-term (r1 o r2), the second relation r2 is always
applied to this “base” variable. We use the associativity of binary join to generate terms like (r1 o (r2 o r3)) as
opposed to ((r1 o r2) o r3).
Similarly, the rule for differentiating a “lambda-term” ((λz.R).t) (i.e., the binding of a sub-term t to a variable z)
looks simple:
∂(λ(z:t1).t2)/∂R = λ(z:∂t1/∂R).t2 ∪ λ(z:t1).∂t2/∂R
However, the actual evaluation of the second part λ(z:t1).∂t2/∂R is subtle, since we need to know the value of this
base variable before being able to apply the sub-term t1 to find z (and compute the whole expression). This
requires a recursive traversal of the derived term to find out exactly when the base variable can be computed.
These compilation methods, expR and expF, are extended to the new operators as follows:
expR(χ(R), x, y, e(y)) = nil
expF(λ(z:[R]).(∂t/∂R), x,y,u,v,e(u,v) )) =
let z = LAST in expF(∂t/∂R, x,y,u,v,e(u,v) )
The first formula tells that χ(R) will be ignored in a rule if it is not differentiated precisely according to R. Thus
it acts as a filter. The second formula shows that we make a direct reference to the previous value LAST of the
attribute that is being updated. This value is stored when we perform the update as we shall later see.
The following is the set of derivation rules for this new logic.
∀ (i,j), ∂Rj/∂Ri = if i = j and otherwise.
-1
-1
if i = j and otherwise.
∀ (i,j), ∂Rj /∂Ri =
∀ p ∈ P, ∂p/∂Ri = .
If S1 and S2 are two types, ∂(S1 × S2)/∂R =
∀ t1, t2 ∈ A(R), ∂(t1 o t2)/∂Ri = (∂t1/∂Ri o t2) ∪ (t1 o ∂t2/∂Ri)
∀ t1, t2 ∈ A(R),∂(t1 ∩ t2)/∂Ri = (∂t1/∂Ri ∩ t2) ∪ (t1 ∩ ∂t2/∂Ri)
∀ t1, t2 ∈ A(R),∂(t1 ∪ t2)/∂Ri = (∂t1/∂Ri ∪ ∂t2/∂Ri )
∀ t1, t2 ∈ A(R), ∀ ƒ ∈ Fo , ∂ψ[ƒ](t1, t2)/∂Ri = ψ[ƒ](∂t1/∂Ri,t2) ∪ ψ[ƒ](t1,∂t2/∂Ri)
∀ t1, t2 ∈ A(R),∀ θ ∈ Fc, ∂φ[θ](t1, t2)/∂Ri = φ[θ](∂t1/∂Ri,t2) ∪ φ[θ](t1,∂t2/∂Ri)
∀ t1, t2 ∈ A(R), ∀ z ∈ V, ∂(λ(z:t1).t2)/∂Ri = λ(z:∂t1/∂Ri).t2 ∪ λ(z:t1).∂t2/∂Ri
∀ t1,t2,tφ ∈ A(R) | tφ=φ[θ](…)),
∂if[tφ](t1,t2) = if[∂tφ/∂Ri](t1,t2) ∪ (∂t1/∂Ri o tφ) ∪ (∂t2/∂Ri o ¬ tφ)
∀ t ∈ A(R), ∂{z:t1, t2}/∂Ri = {z:t1, t2} o π (∂t1/∂Ri ∪ ∂t2/∂Ri)
∀ j, ∂ (Rj*)/∂Ri = (Rj*) o π (∂Rj/∂Ri )
∀ r ∈ R, ∂¬r/∂r = .
∀ (i,j), ∂χ(Rj)/∂Ri = ∂Rj/∂Ri
∀ t ∈ A(R), ∂(λ(z.[R]) t) /∂R = λ(z.[R]) (∂t /∂R)
Figure 2: Differentiation rules.
These rules completely specify differentiation as a formal computation. The analogy with rules for numerical
differential calculus is no coincidence and can be proven using a matrix model for binary relations [Cas94].
14
4.2.4 Examples
For each rule given in Section 4.1.3, we give the equivalent algebraic term t and its derivation ∂t/∂R according to
the appropriate relation. Since these rules use an event pattern to capture one precise pattern, there is only one
relation R such that the derivative ∂t/∂R is not empty.
R1 [t1]:
λ(u:minstart’).λ(z:χ(minstart)).attached
R2 [t2]:
λ(I:task x taskInterval).λ(z:duration).(χ(duration) o
φ[>=](minStart, minStart o I) o φ[<=](maxEnd, maxEnd o I))
R3 [t3]: λ(I:task x taskInterval).(
φ[<=](maxEnd, maxEnd o I) o λ(u:minstart’).
(φ[<](u, minStart o I)) o φ[>=](χ(minstart), minStart o I)))
RS1 [t4]: λ(y:(location–1 o χ(location))). y o φ[>=](strength o y, strength)
RS2 [t5]: λ(u:location’).( φ[=](forest? o u, false) o φ[=](forest? o χ(location), true)
It is easy, although tedious, to verify that these algebraic terms represent indeed the conditions that define rules.
For instance, λ(y:(location–1 o χ(location))). φ[>=](strength o y, strength) (x,z) Ù ∃ w, location(y) = w ∧
location(x) := w ∧ strength(y) >= strength(x). We then apply differentiation with respect to all relations that are
involved in the rules/
Here are some of the derived terms:
∂t1/∂minStart =
∂t2/∂minStart =
∂t4/∂location =
∂t5/∂location =
λ(u:minstart’).((λz. ) attached–1 )
λ(I:task x taskInterval).λ(z:duration’).
(
o φ[>=](minStart, minStart o I) o φ[<=](maxEnd, maxEnd o I))
λ(y:(location–1 o )). φ[>=](strength o y, strength)
λ(u:location’).( φ[=](forest? o u, false) o φ[=](forest?
o
, true)
Each derived term represents a function, which meaning is obtained from the semantic of the functional algebra.
For instance, the first term is a function that takes a pair of objects (a,b), binds z to the value b (from the update
a.minstart := b), binds u to the previous value of a.minstart, and computes the set of objects that are attached to
a. The use of this function will become more evident in the next section, where we generate a demon that
applied the conclusion of the rule to all pairs (a,y) produced by this derived term.
4.3 Code Generation
4.3.1 Generating Demons from Rules
When the compiler is applied to a set of rules, all rules are translated into an algebraic representation and all
derivatives are computed. An algebraic rule is represented as (t => e), where t is the term from the relational
algebra that is equivalent to the logical condition of the rule and e is the “conclusion” of the rule (the action part).
A code fragment is then produced for each rule (t => e) and each derivative ∂t/∂R by applying the exp2 method
as follows:
expF(∂T/∂R,x,y,u,v,e)
The result is a fragment of CLAIRE code that applies the conclusion of the rule to all pairs (u,v) of objects that
satisfies the condition of the rule because of the event “x.R := y”.
These fragments are combined, for each relation R, into a if-write demon that will be used to change the value of
x.R, or R[x] if R is a table (in CLAIRE, tables and slots are abstracted into a binary relations and can be used
indifferently in the logic language). The demon generated for R is based on the following pattern:
if_write[R](x, y)
-> let LAST := get(R,x) in
(if (LAST != y)
(x.R := y,
…
if (x.R = y) …
if (x.R = y) …
…))
// the previous value is stored
// checks that this is indeed an update
// performs the update
st
// propagation fragment from 1 rule
nd
// propagation fragment from 2 rule
rd
// propagation fragment from 3 rule
Note that the propagation fragment obtained from the rules other than the first one are guarded with a test
(x.R = y) to make sure that the rules are only applied if the update is still valid.
CLAIRE makes a difference between mono- and multi-valued relations. For multi-valued relations, the event
model is simpler since the only event that is detected is the addition of a new member in one object’s set of
values. Thus the generation of the demon is slightly simpler in the multi-valued case.
15
For instance, the last phase of the compilation of the closure rule, which we gave as an example in the previous
section, consists of associating the set of functions (derived from each relational term occurring in the condition)
and the CLAIRE expression in the conclusion, so as to build one if-write demon per function. For instance, the
demon associated to the rule closure and the relation edge is compiled as follows (it is evaluated each time a
pair (x,y) is added to the edge relation):
if_write[edge](x, y)
-> if (x.edge add? y)
(x.path :add y,
for z in y.path
// tests if y is not already here
x.path :add z)
The handling of instantiation is straightforward since each instance is attached to its class through an inversible
relationship (isa/instances). Therefore the representation of the event
x :: C
is made with the term χ(minstart). On the other hand, generating a large collection of if-write demons for this
“isa” relation would slow instantiation down for every class; thus, we associated directly the if-write demon
with the class C (this is why C must be a constant in the expression x :: C).
4.3.2 Code Generation Techniques
The code generation phase is mostly the application of the expR and expF methods, which embodies the
application of various optimisation schemes, such as the generation of simpler and faster code if a sub-term
represents a mono-valued relation. Moreover, there are two other techniques that are used to produce procedural
code, which is as close as possible to what the user would write herself to propagate the rules. The first
technique is the use of precise type inference and the second one is the use of modes to distinguish between
different semantics for applying the rules’ conclusions.
Type Inference within the rule compiler is defined by associating a second-order type to each subterm, either
from the relational algebra or the derived functional algebra. These second-order types are functions that infer an
output type from any input type or reciprocally. For instance, if y is bound to x through the relation represented
by t, this function will return the type of y depending on the type of x. The case of a derivative term is slightly
more complex since there are two input type parameters, namely the type of x and y where the update is
“x.R := y”. These functions are obtained in a straightforward manner through the abstract interpretation of the
relational operators on the CLAIRE type domain, which is an inclusion lattice [CC77]. This precise type
inference is key to producing efficient code, and it is more powerful than the type inference performed by
CLAIRE on the fragments generated by the rule compiler.
CLAIRE supports some additional tuning of its rules through the mode declaration. These modes are defined as
follows, and enable the user to remedy some of the problems that occur with non-monotonic and noncommutative rules.
<modes> ≡ mode(
default | once | set |
<integer> )
Rules should not, in general, be written in such a way that the result is order-dependent. The order in which they
are triggered depends on the events (the propagation pattern) and the order in which the rules were entered. If it
becomes necessary to have a more precise control over this order, priorities may be used. A priority is an integer
attached to the rule using the mode declaration. CLAIRE will ensure that rules with higher priority will be
triggered first for each new event. Notice that there is no implicit stack structure for triggering rules: the events
generated by the application of a first rule may cause a new rule to be evaluated before a second rule is applied
to the original event.
The conclusion is applied to any pair of objects that is obtained through a logical derivation of the conclusion
and the update. This assumes that the conclusion can be fired more than once when the logical expression is
redundant (multiple derivation paths for the same pair). However, it may be wrong to apply the conclusion twice
even if a pair is obtained from two different paths. Consider the following example:
strange1(x:Person) :: rule( x.age = 18 | x.age > 10
=>
give(x, $1000))
An update "John.age := 18" may cause the rule to be fired twice. The solution in CLAIRE is to use the
mode(set) declaration before the rule, which will force the computation of the set of pairs before firing the
conclusion (thus eliminating duplicates). Note that the use of an event pattern is precisely another way to get rid
of this problem.
Similarly, rules should have a monotonous behavior, which means that their conclusion should not invalidate the
condition. CLAIRE supports non-monotonous rules but the difficulties linked to the absence of a clear semantic
remain. Let us consider a second example:
strange2(x:Person, y:Person) :: rule(
x.age = 18 & y ∈ x.friends => (x.age := 19, invite(x,y)))
16
Should the rule invite one friend (which one) or all friends when the age of John is set to 18? The default
behavior as well as the "set" behavior will invite all friends. The mode(once) declaration will ensure that the
conclusion is fired only once for each update event. Using non-monotonic rules is tricky, but it is sometimes
very useful.
These modes are implemented as an additional layer in the generation of the code fragments for the demons. The
priorities are used to order the code fragments. The “set” mode builds a temporary set of pairs to be updated
before applying the conclusion. The “once” mode uses an exception handler to make sure that the conclusion is
only fired once.
4.3.3 Application
Here are two examples of demons generated by CLAIRE from the rules that were proposed as examples. The
first one is associated to the minStart slot and represents the two first rules. It is obtained by applying the code
generation method expF to the two derived terms that we have shown in Section 3.3.
if_write[minStart](t, y)
-> let LAST := t.minStart in
(if (LAST != y)
(t.minStart := y,
// the previous value is stored
// checks that this is indeed an update
// performs the update
st
// propagation fragment from 1 rule (R1)
when Y := get(attached,t) in Y.minStart :+ (y – LAST),
// propagation fragment from 2nd rule (R3)
if (t.minStart = y)
for I in taskInterval
if (y >= I.minStart & LAST < I.minStart & t.maxEnd <= I.maxEnd)
I.duration :+ x.duration )
The second example is obtained from rules RS1 and RS2 for the location attribute:
if_write[location](x, y)
-> let LAST := x.location in
(if (LAST != y)
(x.location := y,
// the previous value is stored
// checks that this is indeed an update
// performs the update
st
// propagation fragment from 1 rule (RS1)
for Y in of~(location,y)
if (Y.strength >= x.strength)
// reads the inverse of location
failure(x)
nd
// propagation fragment from 2 rule (RS2)
if (x.location = y)
if (y.forest? & not(LAST.forest?) computeVisibility(x) )
These two examples illustrate our claim that the rule compiler is able to produce procedural code for propagation
that is very close to what a user would write.
4.3.4 Performance Results
We have not tested the new features of ClaireRules from a performance point of view due to the absence of
established benchmarks. On more classical problems, such as those reported in Table 3, the code generation
strategy of CLAIRE is very efficient (between 100 K to 50 Mips = Million inferences per second), even when
solving complex problems. . The CPU time is measured on a Pentium III laptop at 500Mhz, using MSVC++ 6.0.
Since CLAIRE can also generate Java code, we also include the results using Symantec’s Visual Café 3.0.
CLAIRE is between 2 or 3 orders of magnitude faster than what we have measured with RETE-based inference
engines [For82] and various successive improvements such as TREAT [Mi87] or GATOR [HH93]. We have
tested over the last 10 years four major commercial products – they are all based on refinements of the RETE
algorithm - and two academic inference engines. The problems that we use in Table 3 are taken from a small set
of the rare examples that circulate in the inference engine community, they do not exercise other features than
rule propagation – this is not a backtracking benchmark. More empirically, we have noticed that CLAIRE
programs written with production rules yield only about a 10 to 50 % penalty compared with hand-optimized
procedural code.
CPU time
Java version
CPU time
C++ version
Rules/s
C++ version
Simple Filtering
0.01s
0.02s
50 Mips
Planning (Monkey & Banana)
0.03s
0.015s
13 Mips
17
Constraint Propagation (triplet)
0.36s
0.05s
3 Mips
Rules & methods (Airline)
0.05s
0.01s
100Kips
Problem solving (Zebra)
0.043s
0.011
2.7 Mips
Problem solving (Dinner)
0.18s
0.08s
250Kips
Table 3: Benchmarks on rules.
The demons are produced using CLAIRE, taking advantage of CLAIRE high-level of abstraction, which
simplifies the handling and the iteration of set expressions, as well as the reflective nature of the CLAIRE
system, which makes generating new programs much easier. The CLAIRE program is then compiled into C++
or Java. The generation of Java programs is a new feature from the CLAIRE compiler. Our next step is to
generate directly Java beans that exploit the event/listener model. Because we can generate Java code with no
overhead or central control algorithm, this technique is ideally suited for compiling a set of rules into a Java
bean component [Eng97]. This approach can be applied to all other publish & subscribe message architecture,
which makes this technique very interesting for Enterprise Application Integration (EAI). For instance, we plan
to use this approach to implement distributed workflow engines where a set of process rules is compiled into a
group of distributed agents.
5. Search Algorithms
5.1 Search in an Imperative Language
In addition to rules, we saw that CLAIRE also provides the ability to do some hypothetical reasoning. This
possibility to store successive versions of the database and to come back to a previous one is called the world
mechanism. The part of the database that supports these defeasible updates is defined by the programmer and
may include slots, tables, global variables or specific data structures such as lists or arrays. When the data model
is defined, one may specify which slots are defeasible and which are not. Each time we ask CLAIRE to create a
new world, CLAIRE saves the status of defeasible slots (and tables, variables, ...). Worlds are represented with
numbers, and creating a new one is done with choice(). Returning to the previous world is done with
backtrack(). Returning to a previous world n is done with world=(n). Worlds are organized into a stack (one
cannot explore two worlds at the same time) so that save/restore operations are very fast. The current world that
is being used can be found with world?(), which returns an integer. In addition, one may accept the hypothetical
changes that were made within a world while removing the world and keeping the changes. This is done with the
commit() method. commit() decreases the world counter by one, while keeping the updates that were made in
the current world. It can be seen as a collapse of the current world and the previous world. Note that although
the wording is closer to AI concepts, the implementation of worlds using a stack is very close to the trailing
stack of a constraint logic programming system such as CHIP[VH89] and totally similar to the backtracking
stack of a constraint solving tools such as ILOG solver [Pu94].
Here is a real example of the search strategy for the jobshop problem [Lab98]. The functional parameter
tightest is a heuristic that returns the set of tasks (an Interval) that share the same machine and that looks for
the most difficult to process.
solve(tightest:property) : boolean
-> when I:Interval := tightest() in
exists(t in {x in I | canbefirst?(I, t)} |
branch( (for t’ in (I but t) t.After :add t’, solve(tightest) )
else true
// no interval was found
This example shows many interesting features. Some of them have already been mentioned: the iteration of a
selection set (the set of tasks that can be the first in the set I), the use of the pattern (I but t) or the use of a
functional parameter. We also notice the defeasible update t.After :add t’, which adds t’ to the list of tasks
performed after t and triggers the propagation rules that represent the scheduling constraints [Lab98].
Another interesting aspect of the branch control structure is contradiction handling (failure in a sub-tree).
CLAIRE supports exceptions (represented as objects in a hierarchy) with a built-in contradiction subclass.
When a contradiction is raised (for instance, by rule propagation), it is caught by the branch(...) instruction
which backtracks and returns false. Thus, branch support a concise and readable programming style of search
algorithms, an asset for defining easily new variations (other search strategies, instrumentation, ...).
We are also making good use of the extensibility of the iteration compiler. A task interval I is a set of tasks that
share a same resource I.Res whose time windows are included in a given interval [Lab98]. Each task has a
unique index and the set of tasks that belong to I is represented by a bit-vector I.SET. Thus the iteration of a
task interval consists of iterating the set of tasks on the right machine whose indexes belong to the bit-vector.
For instance, the iteration in the previous algorithm will be expanded into:
18
for t in (I.Res).users
(if I.SET[t.index]
(if canbefirst?(I, t)
branch(...) ))
The ability to pass function parameters at no cost from a performance point of view is a key feature to write
parametric algorithms. This is reinforced by the fact that these parameters can be in-line methods, which means
that the original in-line method can be seen as a higher-level function generator. For instance, we can use the
following definition of max to define the method later:
// a generic max method
max(s:subtype[integer], greater:property, default:any) : integer
=> let x := default, empty := true in
(for y in s (if empty (x := y, empty := false)
else if greater(x, y) x := y), x)
before(x:task, y:task) : boolean
=>
later(x:Interval, t:task) : integer
x.atleast ≤ y.atleast
->
max(x but t, before, EndTask)
The code generated for the method later will be:
let t0 := EndTask, empty := true in
(for t1 in (x.Res).users
(if x.SET[t1.index]
(if (t != t1)
(if empty (t0 := t1, empty := false)
else if (t0.atleast ≤ t1.atleast) t0 := t1))),
t0)
In that example, we have used max as a high-level code generator. This is the first step in writing truly reusable
code for complex problem solving algorithms.
5.2 Composition Polymorphism
Multiple layers of rewriting rules lead to the need for a new paradigm to implement simplification (or reduction)
rules. Consider the two following equations:
(A + B)[i,j] = A[i,j] + B[i,j]
x ∈ (s but t) = (x ∈ s) & (x != t)
These equations may be seen as simplification rules since they respectively allow us to extract a cell of a matrix
sum without computing this sum explicitly or evaluate membership to a “but” set without actually computing
this set. Usually, there is no need to represent these rules because the programmer uses them implicitly. However,
this is no longer the case when these sequences of source-to-source rewriting are used. For instance, there may
exist the expression M[i,j] in the body of an inline method that gets transformed into (A + B)[i,j] because of
the form of the input parameters. This is even more obvious with the second example if we apply a set inline
method to the set expression (s but t).
Representing these simplification rules leads to the notion of composition polymorphism. What we need to
represent is that the processing of f(...,x,...) may be different when x is obtained as g(....). For instance, f
is get (get is used implicitly in the form ...[...], e.g. M[i,j] = get(M,i,j)) and g is + in the first case; f is ∈
and g is but in the second case. This type of polymorphism might be captured with a complex abstract data type
in a statically typed language using complex pattern matching rules, but is out of reach with our concrete type
approach. This is why we have introduced the notion of pattern: a pattern is a set of function calls with a given
selector and a list of types (list of types to which must belong the arguments to the call); a pattern in CLAIRE is
written p[tuple(A,B,...)]. Patterns support composition polymorphism through the definition of new inline
methods whose signature are made either of any (for concrete parameters) or patterns (for expressions). The
compiler uses such a restriction when it finds a function call that matches its signature. Here is the application to
the “matrix” example:
get(x:+[tuple(Matrix,Matrix)],i:any,j:any)
=> (args(x)[1][i,j] + args(x)[2][i,j])
Suppose that we now have this situation:
inverseTrace(M:Matrix) => sum(list{1.0 / M[i,i] | i in (1 .. N)})
f(M1:Matrix, M2:Matrix) -> inverseTrace(M1 + M2) > 0.0
The macro-expansion of inverseTrace(M1 + M2) will yield to an expression that involves
This call matches the signature of the optimizing restriction that we defined previously (with
args(x) = (M1,M2)). Thus a further reduction is performed and the result is:
(M1 + M2)[i,i].
x = M1 + M2 and
let x := 0, i := 1 in
(while (i >= N)
(x :+ 1.0 / (M1[i,i] + M2[i,i]), i :+ 1),
x > 0.0)
19
The macro-expansion of inverseTrace(M1 + M2) is performed by the compiler under the same considerations
an any inlining optimization: the compiler will check that no side-effects are involved, or that the parameter of
an inlining pattern is used exactly once.
5.3 Writing Elegant Algorithms
Our last example is taken from a real application that uses the Hungarian Algorithm to solve a bipartite matching
problem (useful for global constraint propagation [Lab98]). Although it does not use search mechanisms, we
believe that it demonstrates the value of CLAIRE high-level features to express a complex algorithm in a more
readable and elegant manner, even if not declarative.
// builds a maximum weight complete matching
match() -> (...,
while (HN != N)
(if not(grow())
dualChange()))
// initialization
// main loop
// a step repeats until the forest is hungarian (return value is false)
// or the matching is improved (return value is true)
// explore is the stack of even nodes that have not been explored yet
grow() : boolean
-> let i := pop(explore) in
// explore is the stack of nodes that need to be explored
(exists(j in {j in GpiSet(i,LastExplored[i] + 1,LastValid[i]) | not(odd?[j])} |
(if (sol-[j] != 0)
(//[SPEAK] grow: add (~S,~S) to forest// i,j,
odd?[j] := true, pushEven+(sol-[j]), tree[j] := i, false)
else (augment(i,j), true))) |
(if not(empty?(explore))
// if we did not find a j to augment the forest,
grow()) )
// grow is called recursively
// change the dual feasible solution, throw a contradiction if there are no perfect matching
dualChange() : integer
-> let e := Min( list{vclose[i] | i in {j in Dom | not(odd?[j])}}) in
(//[SPEAK] DUAL CHANGE: we pick epsilon = ~S // e,
if (e = NMAX) contradiction!(),
for k in stack(even) (pi+[k] :+ e, LastExplored[k] := LastValid[k]),
for j in {j in Dom | odd?[j]} (pi-[j] :- e, vclose[j] := NMAX)),
clear(explore),
for i in stack(even)
let l := Gpi[i], k := size(l), toExplore := false in
(while (LastValid[i] < k)
(k, toExplore) := reduceStep(i,j,l,k,toexplore),
if toExplore push(explore,i)))
// look at edges outside the valid set one at a time
reduceStep(i:Dom,j:Dom,l:list,k:integer,toExplore:boolean) : tuple(integer,boolean)
-> let j := l[k], c := Cpi(i,j) in
(if (c = 0)
(//[SPEAK] dual_change: Add edge ~S,~S // i,j,
Gpiadd(l,i,j,k), toexplore := true)
else (vclose[j] :min c, k :- 1)),
list(k,toexplore))
Let’s briefly comment the three functions above. match builds the matching by repeatedly applying the two steps
grow and dualChange until the complete set of nodes is spanned (this is a primal-dual algorithm). grow builds a
Hungarian forest using a classical search for augmenting chains (with alternate marks even and odd) in the
reduced graph Gpi. dualChange changes the dual values stored in the arrays pi+ and pi- so that the reduced
graph Gpi is augmented. The reduced graph is implicitly defined as the subset of all edges satisfying a given
condition on the dual weights (that the corresponding constraint in the linear model is saturated). It is
represented by the following pattern: GpiSet(i,a,b) refers to the subset of neighbors of the node i in this
reduced graph where a and b are two iteration cursors (we maintain an ordered set of edges incident to i in Gpi;
all edges up to position a have been explored and all edges up to position b have been touched).
A complete explication of the algorithm is out of the scope of this paper, but we can notice:
• The use of patterns such as stack(even) or GpiSet(i,a,b) which encapsulates the complexity due to data
structure handling (e.g., using the even array to implement the stack).
20
•
The use of complex set expressions that are iterated (e.g., the first line of dualChange) and the use of set
operators (exists) and generic functions (min).
• The use of multi-valued functions (reduceStep) to isolate fragments of code that do not return a single
value.
• The use of active comments, which can be used as trace statements.
From our point of view, a description of this complex algorithm in less than 30 lines is very valuable, and is a
strong advocate for using high-level languages. The interest of the techniques presented in this paper is that they
may come with no penalties for run-time performance, since this CLAIRE implementation was checked to be
exactly as efficient as a standard C++ implementation. These 30 lines may be compared with the 60 lines of the
C++ program that was written by hand or the 90 lines that were generated by the CLAIRE compiler. The real
gain is not the concision, it is the ability to grasp the design of the algorithm at a higher level of abstraction
while reviewing the code.
6. Conclusion
Experiments in the previous years have shown that constraint programming was best used in combination with
other optimization techniques for combinatorial problem solving. The fact that hybrid algorithm engineering
cannot be always performed exclusively through the use of components and may require some programming
motivates the use of multi-paradigm programming languages. There have been many proposals for incorporating
“advanced features” such as sets or rules into object-oriented languages, especially in the AI community. On the
other hand, few have managed to incorporate as well a high degree of polymorphism and the use of higher-order
functions. Among those few, the most interesting, from our point of view, are LIFE and Oz.
The major contribution of LIFE [AK93] is to provide with a clean integration of logic, inheritance and functions
(as its name stands). LIFE played an important role when we tried to put some elegance into our “mix of
features”. However, there is a top-down vs. bottom-up difference in the design strategy of CLAIRE and LIFE.
CLAIRE is the bottom-up integration of features that we know how to compile efficiently whereas LIFE is the
top-down implementation of paradigms that are known to blend elegantly. The result is that it is easier, from our
point of view, to write efficient algorithms with CLAIRE than with LIFE.
Oz is a concurrent constraint language [Smo95] belonging to the post-CLP language generation. Its objectoriented kernel offers multiple inheritance and a clean integration of higher order functions, in particular with
the predefined search combinators that offer parametric search procedures [SS94]. These predefined
combinators have been extended into the ability to define search procedures (programming constraint inference
engines [Sch87]) with few lines of elegant code, which is very similar to CLAIRE and LAURE [Cas94], both
from a motivation and a result point of view,
The comparison between Oz and CLAIRE is interesting because Oz is both more sophisticated and simpler than
CLAIRE. On the one hand, Oz is more ambitious than CLAIRE since it proposes a computational model which
is original, centered on constraints and search (stores), and supports distributed computing. CLAIRE uses an
object-oriented data model that is geared towards knowledge representation, with the additional benefits of
providing a strong basis for data types. Oz has been a model for some of CLAIRE features, such as the use of
higher-order functions to capture parametric search strategies. On the other hand, Oz is still an un-typed
language that is too high-level, from our point of view, to be a language of choice to implement propagation
strategies. Oz relies on the ability to use a lower-level language to implement some of the constraint propagation
strategies, while CLAIRE is a self-contained tool.
To summarize, we have presented a high-level language for writing re-usable algorithms. The application of this
language to various optimization problems has resulted in programs that are easier to read (and thus to maintain)
and to reuse (or to combine). This is mostly due to two features:
• Declarative programming is encouraged with the use of logic rules (the programmer is freed from the
concern about propagation patterns), which can be compiled into procedural demons with no run-time
processing overhead.
• Set programming and set iteration raise the level of abstraction and recursively-nested expressions can be
compiled with no run-time processing overhead.
We believe that our experience with CLAIRE may be of interest to the CLP community on three topics: the
technology of compilation-by-differentiation for rules, the definition of a simple multi-paradigm language
supported by real applications and the concern for a general purpose programming language for hybrid
algorithms. On this last point, CLAIRE offers different tools for different concerns: an object oriented type
system for elegant modeling, a data-driven rule compiler for declarative programming, parametric polymorphism,
iteration and patterns for data structure manipulations and a version mechanism (reified in SaLSA [LC98]) for
controlling search trees. This suggests that the scope of CLAIRE is broader than its current use for combinatorial
optimization algorithms and that it could be used for applications that have been associated with logic and rulebased programming.
21
CLAIRE has been used for various deployed industrial applications, in different companies. It was also used as a
support tool of choice for teaching “Constraints and Algorithms”. CLAIRE is written in CLAIRE (10,000 lines of
code), except for a small bootstrapping, run-time kernel, and the optimizer makes heavy use of CLAIRE features
(including garbage collection). CLAIRE is a public domain tool that may be downloaded from various WEB sites.
References
[AK93]
H. Aït-Kaci. An Introduction to LIFE: Programming with Logic, Inheritance, Functions and
Equations. Proceedings of ILPS’93, p.117, 1993.
[AS97]
K.R. Apt, A. Schaerf. Search and Imperative Programming. Proceedings of POPL'97, p. 67-79,
ACM Press, 1997.
[Cas94]
Y. Caseau. Constraint Satisfaction with an Object-Oriented Knowledge Representation Language.
Applied Intelligence, Vol. 4, no. 2, p. 157-184, May 1994.
[Cas95]
G. Castagna. Covariance and Contravariance: Conflict without a Cause. ACM Transactions on
Programming Languages and Systems, Vol 17, no. 3, p. 431-447, May 1995.
[CL94]
C. Chambers, G. Leavens. Typechecking and Modules for Multi-Methods. Proceedings of
OOPSLA’94, ACM Sigplan Notices, p. 805-843, Portland, 1994.
[CL96a]
Y. Caseau, F. Laburthe. Introduction to the CLAIRE Programming Language. LIENS research report
96-16, Ecole Normale Supérieure, 1996.
[CL96b]
Y. Caseau, F. Laburthe. Cumulative Scheduling with Tasks Intervals. Proceedings of JICSLP’96, M.
Maher ed., p. 363-377, The MIT Press, 1996.
[CP93]
Y. Caseau, L. Perron. Attaching Second-Order Types to Methods in an Object-Oriented Language.
Proceedings of ECOOP’93, O.M. Nierstrasz Ed., p.142-160, Kaiserslautern, Germany, 1993.
[DGL95] M. Day, R. Gruber, B. Liskov, A. Myers. Subtypes vs. Where Clauses: Constraining Parametric
Polymorphism. Proceedings of OOPSLA’95, ACM Sigplan Notices, p. 156-168, Austin, TX, 1995.
[Eng97]
R. Englander. Developing Java Beans. O’Reilly, 1997.
[For82]
C.L. Forgy. Rete: A Fast Algorithm for the Many Pattern/Many Object Pattern Match Problem
Artificial Intelligence, 19, p. 17-37, 1982.
[GR83]
A. Goldberg, D. Robson. Smalltalk-80: The Language and its Implementation. Addison-Wesley,
1983.
[HH93]
E. Hanson, M. Hasan. Gator: An Optimized Discrimination Network for Active Database Rule
Condition Testing. Tech report 93-036, CSE CIS Department University of Florida Gainesville, FL,
1993.
[HU94]
U. Holzle, D. Ungar. A Third-Generation Self Implementation: Reconciling Responsiveness with
Performance. Proceedings of the ACM OOPSLA’94 Conference, p. 229-243, Portland, OR,
October 1994.
[Lab98]
F. Laburthe. Constraints and Algorithms in Combinatorial Optimization, Ph.D. Thesis (in French)
University of Paris VII, 1998.
[LC98]
F. Laburthe, Y. Caseau. SaLSA: A Language for Search Algorithms, Proceedings of CP’98,
M.Maher, J.-F. Puget eds., Springer, LNCS 1520, p.310-324, 1998.
[KW89]
M. Kifer, J. Wu. A logic for Object-Oriented Logic Programming (Maier’s O-Logic
Revisited). Proceeding of PODS-89 , Philadelphia, 1989.
[McL81] B.J. MacLennan. Programming with A Relational Calculus. Rep N° NPS52-81-013, Naval
Postgraduate School, September 1981.
[Mir87]
D.P. Miranker. TREAT: A better match algorithm for AI production systems. In Proceedings AAAI87 Sixth National Conference on Artificial Intelligence, pages 42-47. Morgan Kaufmann, San
Francisco, August 1987.
[NT89]
S. Naqvi, S. Tsur. A Logical Language for Data and Knowledge Bases. Computer Science Press,
1989.
[Pug94]
J.-F. Puget. A C++ Implementation of CLP. Ilog Solver Collected papers, Ilog tech. report, 1994.
[Sch87]
C. Schulte. Programming Constraint Inference Engines. Proceedings of CP97, G. Smolka ed.,
Springer, LNCS 1330, p. 519-533, 1997.
[SDD86] J. Schwartz, R. Dewar, E. Dubinsky, E. Schonberg. Programming with Sets: an Introduction to
SETL. Springer, New-York, 1986.
[SKG87] H. Schmidt, W. Kiessling, V. Guntzer, R. Bayer. Compiling Exploratory And Goal-Directed
Deduction into Sloppy Delta-Iteration. Proc. of the Symposium on Logic Programming, p. 233-243,
San Francisco, 1987.
[SS94]
C. Schulte, G. Smolka. Encapsulated Search for Higher-order Concurrent Constraint Programming,
Proceedings of ILPS’94, M. Bruynooghe ed., MIT Press, p. 505-520, 1994.
22
[Smo95]
[SZ88]
[VH89]
G. Smolka. The Oz Programming Model, Proceedings of Computer Science Today, J. Van Leeuwen
ed., LNCS 1000, p. 324-343, Springer, 1995.
D. Sacca, C. Zaniolo. Differential Fixpoint Methods and Stratification of Logic Programs. MCC
Technical Report ACA-ST-032-88, January 1986.
P. Van Hentenryck. Constraint Satisfaction in Logic Programming. The MIT press, Cambridge,
1989.
Appendix: Benchmark Results
The CLAIRE language is not a powerful high-level language whose compiler has progressively been designed
to eliminate undue overhead. On the contrary, CLAIRE was a simple C++ code generator, which was designed
to guarantee the absence of overhead, and that has evolved into a powerful language over the years while we
were improving the compiler technology. The first application of CLAIRE, in 1994, has been job shop
scheduling software; we presented a new state-of-the-art algorithm with a new data structures (task intervals).
Because we did not have the techniques presented here available at that time, the handling of the data structures
was done by hand everywhere in the code. An encapsulation approach based on iterators would have produced a
slow-down of a factor of 10, but we were able to rewrite the algorithm two years later with exactly no overhead,
since the compiler had been tuned to produce precisely the code we were starting from. We have used CLAIRE
to encode many classical graph algorithms (shortest path, flows, matching, etc.). With earlier versions of the
compiler, the encoding was very similar to the C or Pascal algorithms that we were copying from. The
introduction of the techniques presented here allowed us to evolve towards much more elegant and compact
encoding, without any overhead.
This paper contains three sets of benchmarks that we use to monitor CLAIRE performance; the first two are
reported in this appendix and the last one was shown in Section 4.3.4. They illustrate our central claim that
CLAIRE introduces no overhead compared to more traditional procedural languages. We first gathered two
“algorithmic” micro-benchmark suites: respectively the Self and the Stanford benchmarks. Self’s original results
[HU94] are not reproduced since the hardware has changed too much in the last few years; but the comparison
with C++ shows that CLAIRE seems faster than Self.
Fibonacci(38)
Fill a vector of 100,000 integers (x 1,000)
Increment each element of that vector (x 1,000)
Self
benchmarks Add up all the elements of that vector (x 1,000)
Sum iterators of 2 nested 100-loops (x 100,000)
Bubble sort of 10,000 integers
Quick sort of 2,000,000 integers
Tree sort of 5,000 integers
Stanford
Multiplication of 2 300x300-matrices of integers
benchmarks
Multiplication of 2 300x300-matrices of floats
Permutation of the elements of a vector
Time-consuming placement problem (x 500)
CLAIRE
Visual C++ 6
Java (Visual
Café 3)
4,140
1,220
1,800
830
6,970
890
620
660
500
2,080
3,090
7,610
4,090
1,190
1,750
890
6,940
840
560
780
520
1,540
2,520
6,660
4,530
1,200
1,830
970
6,920
1,140
760
1,730
1,190
2,420
2,730
7,470
We also report benchmark results on libraries for collection data structures for the same 3 languages (we chose
the closest C++ and Java libraries to CLAIRE lists and sets). Moreover, multiple experiments have confirmed
what these benchmarks suggest: handling collection data structure is definitely a strong feature of CLAIRE.
23
Creation
Lists of
Iteration
1,000 objects Membership
Deletion and insertion
Creation
Sets of
Iteration
1,000 objects Membership
Deletion and insertion
Creation
Hash sets of Iteration
1,000 objects Membership
Deletion and insertion
CLAIRE
Visual C++ 6
Java (Visual
Café 3)
140
2,690
4,580
1,840
340
2,440
300
4,120
170
3,750
90
60
90
7,260
15,600
5,770
580
7,220
750
4,950
8,770
23,800
8,030
3,170
390
26,700
17,800
6,970
6,090
26,200
133,000
2,450
1,730
29,900
370
870
24
| 2cs.AI
|
1
Gland Instance Segmentation Using
Deep Multichannel Neural Networks
arXiv:1611.06661v3 [cs.CV] 23 Nov 2017
Yan Xu, Yang Li, Yipei Wang, Mingyuan Liu, Yubo Fan, Maode Lai, and Eric I-Chao Chang*
Abstract—Objective: A new image instance segmentation
method is proposed to segment individual glands (instances)
in colon histology images. This process is challenging since
the glands not only need to be segmented from a complex
background, they must also be individually identified. Methods:
We leverage the idea of image-to-image prediction in recent
deep learning by designing an algorithm that automatically
exploits and fuses complex multichannel information - regional,
location and boundary cues - in gland histology images. Our
proposed algorithm, a deep multichannel framework, alleviates
heavy feature design due to the use of convolutional neural networks and is able to meet multifarious requirements by altering
channels. Results: Compared to methods reported in the 2015
MICCAI Gland Segmentation Challenge and other currently
prevalent instance segmentation methods, we observe state-ofthe-art results based on the evaluation metrics. Conclusion: The
proposed deep multichannel algorithm is an effective method
for gland instance segmentation. Significance: The generalization
ability of our model not only enable the algorithm to solve
gland instance segmentation problems, but the channel is also
alternative that can be replaced for a specific task.
Index Terms—Convolutional neural network, instance segmentation, histology image, multichannel, segmentation.
I. I NTRODUCTION
E
XISTING in most organ systems as important structures,
glands secrete proteins and carbohydrates. However,
adenocarcinomas, the most prevalent type of cancer, arises
from the glandular epithelium [1]. The morphology of glands
determines whether they are benign or malignant and the level
of severity [2]. Segmenting glands from the background tissue
is important for analyzing and diagnosing histological images.
In gland labeling/segmentation, each pixel is assigned one
label to represent whether the pixel belongs to the foreground
Manuscript submitted for review on December 24, 2016; accepted on
March 11, 2017. This work is supported by Microsoft Research under the
eHealth program, the Beijing National Science Foundation in China under
Grant 4152033, the Technology and Innovation Commission of Shenzhen
in China under Grant shenfagai2016-627, Beijing Young Talent Project in
China, the Fundamental Research Funds for the Central Universities of China
under Grant SKLSDE-2015ZX-27 from the State Key Laboratory of Software
Development Environment in Beihang University in China.
Yan Xu, Yang Li, Yipei Wang, Mingyuan Liu and Yubo Fan are with
the State Key Laboratory of Software Development Environment and Key
Laboratory of Biomechanics and Mechanobiology of Ministry of Education
and Research Institute of Beihang University in Shenzhen, Beihang University,
Beijing 100191, China.
Yan Xu and *Eric I-Chao Chang are with the Microsoft Research, Beijing
100080, China (*Corresponding author: [email protected]).
Maode Lai is with the Department of Pathology, School of Medicine,
Zhejiang University, China.
Copyright (c) 2016 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending an email to [email protected].
Fig. 1. Gland Haematoxylin and Eosin (H&E) stained slides and ground truth
labels. Images in the first row exemplify different glandular structures. Characteristics such as heterogeneousness and anisochromasia can be observed in
this figure. The second row shows the ground truth. To achieve better visual
effects, each color represents an individual glandular structure.
(gland) or the background. However, which gland the foreground pixel belongs to is still not determined. In order to
analyze the morphology of glands, they need to be recognized
individually. Each pixel needs to be classified and it must
be determined which gland the pixel belongs to, which is
to assign a gland ID to each foreground pixel. We call this
task as gland instance segmentation (as shown in Fig. 1). In
this paper, we aim to solve the gland instance segmentation
problem. We formulate this problem as two subproblems gland labeling/segmentation [3], [4] and instance recognition.
The intrinsic properties of gland histopathological image
pose plenty of challenges in instance segmentation [5]. First
of all, heterogeneous shapes make it difficult to use mathematical shape models to achieve segmentation. As Fig.1
shows, the cytoplasm being filled with mucinogen granule
causes the nucleus to be extruded into a flat shape whereas
the nucleus appears as a round or oval body after secreting.
Second, variability of intra- and extra- cellular matrices often
leads to anisochromasia. Therefore, the background portion
of histopathological images contains more noise like intensity
gradients, compared to natural images. Several problems arise
in our exploration of analyzing gland images: 1) some objects
are very close together making only the tiny gaps between
them visible when zooming in on a particular image area; or
2) one entity borders another making their edges adhesive to
each other. We call this an problem of ‘coalescence’. If these
problems are omitted during instance recognition process, even
if there is only one pixel coalescing with another, the algorithm
will consider two instances as one.
Gland labeling/segmentation, as one subproblem of gland
instance segmentation, is a well-studied field where various
methods have been explored, such as morphology-based methods [6], [7], [8], [9] and graph-based methods [10], [11].
However, glands must be recognized individually to enable
2
the following morphology analysis. Gland segmentation is
insufficient due to its inability to recognize each gland in
histopathological images. MICCAI 2015 Gland Segmentation
Challenge Contest [12] has drawn attention to gland instance
segmentation. The precise gland instance segmentation in
histopathological images is essential for morphology assessment, which is proven to be not only a valuable tool for clinical
diagnosis but also a prerequisite for cancer grading [13].
Although gland instance segmentation is a relatively new
subject, instance segmentation in nature images has attracted
much interest from researchers. Ever since SDS [14] raised
this problem and proposed a basic framework to solve it, other
methods have been proposed thereafter, such as hypercolumn
[15] and MNC [16], which merely optimize and accelerate
the feature extraction process. All of these algorithms fall into
a routine that detects objects first and then segments object
instances inside the detected bounding boxes.
In medical image analysis, traditional methods are more
prevalent for segmenting gland instances instead of learningbased methods. Traditional methods depend heavily on handcraft features and prior knowledge. In natural images, instance
segmentation algorithms are mostly the pipeline of object
detection and masking [14], [15], [16]. The objects in natural
images are regular-shaped, and relatively easy to segment by
first creating bounding boxes for each one. However, most
glands are irregular in shape, which increases the difficulty
of detecting the whole gland structure. Thus the traditional
instance segmentation methods for natural images are not
suitable for gland instance segmentation.
In a broad sense, gland instance segmentation can be viewed
as gland labeling process with commutative labels. Thus gland
labeling can offer useful cues for gland instance segmentation.
The latest advantages in deep learning technologies have led
to explosive growth in machine learning and computer vision
for building systems that have shown significant improvements
in a huge range of applications such as image classification
[17], [18] and object detection [19]. The fully convolutional
neural networks (FCN) [20] permit end-to-end training and
testing for image labeling; holistically-nested edge detector
(HED) [21] detector learns hierarchically embedded multiscale
edge fields to account for the low-, mid-, and high- level
information for contours and object boundaries; Faster RCNN [22] predicts object locations and compensates for the
possible failure of edge prediction. We solve the gland instance
segmentation problem by multitask learning. One task is to
segment the gland images, and another task is to identify the
gland instances. In the gland segmentation subtask, a fully
convolutional neural network (FCN) [20] model is employed
to exploit the advantage of end-to-end training and image-toimage prediction. In the gland instance recognition subtask, a
holistically-nested edge detector (HED) and a Faster R-CNN
object detector are applied to define the instance boundaries.
We make use of multichannel learning to extract region,
boundary and location cues and solve the instance segmentation problem in gland histology images (as shown in Fig. 2).
Our algorithm is evaluated on the dataset provided by the
MICCAI 2015 Gland Segmentation Challenge Contest [12]
and achieves state-of-the-art performance among all partici-
Fig. 2. This illustrates a brief structure of the proposed algorithm. The
foreground segmentation channel distinguishes gland pixels from the background. The edge detection channel outputs the result of boundary detection.
The object detection channel detects glands and their regions in the images.
A convolution neural network concatenates features generated by different
channels and produces segmented instances. The white areas in subimage
“region, edge and boxes” represent the results of the recognized glands, edges
and detected bounding boxes.
pants and other popular methods of instance segmentation.
We conduct a series of ablation experiments and prove the
superiority of the proposed algorithm.
This paper is arranged as follows. We formulate the instance
segmentation problem in Section II. Section III is a review
of related previous works. In section IV, we describe the
complete methodology of the proposed algorithm of gland
instance segmentation. Section V is a detailed evaluation of
our method. Section VI summarizes our conclusion.
II. P ROBLEM
We formulate the instance segmentation problem by two
subproblems, labeling/segmentation and instance recognition.
We denote D = {(Xn , Yn , Zn ), n = 1, 2, ..., N } as the
input training dataset, where N is the image amount. We
subsequently drop the subscript n for notational simplicity,
since we consider each image independently. X = {xj , j =
1, 2, ..., |X|} denotes the raw input image, Y = {yj , j =
1, 2, ..., |X|}, yj ∈ {0, 1} denotes the corresponding segmentation label and Z = {Rk , k = 0, 1, 2, ..., K} denotes the
instance label, in which Rk = {(p, q)} denotes the coordinates
set of pixels inside of region Rk . When k equals 0, it denotes
the background area and it denotes the corresponding instance
3
Fig. 3. This illustrates two subproblems of gland instance segmentation. Gland instance segmentation can be formulated into foreground labeling/segmentation
and gland instance recognition two subproblems. As demonstrated in images of the second column, a small amount of prediction errors have little influence
on the final cost function for the foreground labeling/segmentation subproblem; however, for the gland instance recognition subproblem, even a few pixels
predicted incorrectly can highly increase the cost, which is shown in images of the third column. The bar chart shows the cost of two subproblems.
when k takes other values. K is the total instance number.
Regions in the image satisfy the following relations:
Rk ∩ Rt = ∅, ∀k 6= t,
(1)
∪Rk = Ω.
(2)
Ω denotes the whole image region. Note that instance labels
only count gland instances thus they are commutative. Our
objective is to segment glands while ensuring that all instances
are differentiated. Note that the labeling/segmentation subproblem is a binary classification problem. Ŷ represents the
labeling/segmentation result, the cost function is:
|Y |
Dist(Y, Ŷ ) =
1 X
δ(yj 6= yˆj ).
|Y | j=1
yˆj = arg maxy P (y|X)
(3)
(4)
In the instance recognition subproblem, Ẑ denotes the
instance prediction. The cost function is:
Dist(Z, Ẑ) = 1 −
1
K
K0
X
d
L(R
k0 , Z),
III. R ELATED W ORK
This section is a retrospective introduction about instance
segmentation and gland instance segmentation.
(5)
k0 =0
A. Instance segmentation
where
d0 ∩Rk
1, ∃k 6= 0, R
k
d0 ∪Rk > thre
d
R
0
k
L(Rk , Z) =
0, otherwise
d
b
R
k0 ∈ Z denotes the instance segmentation prediction
region and Rk ∈ Z denotes the instance label region. K 0
represents the total predicted region count. thre is the threshold which is set to 0.5 in this algorithm. When the overlap
ratio of the gland instance in a certain prediction region and
labels is higher than the threshold, this region is considered
an instance prediction by the algorithm. Fig. 3 shows the two
gland instance segmentation subproblems.
Since the cost function of instance recognition is nondifferentiable, it cannot be trained with SGD. We hereby
approximate instance recognition by edge detection and object
detection. We generate edge labels E and object labels O
through Y and Z to train edge detector and object detector, in
which E = {ej , j = 1, 2, ..., |X|}, ej ∈ {0, 1} and ej equals 0
when all four nearest pixels (over, below, right and left) belong
to the same instance. O denotes the smallest bounding box for
each gland instance.
(6)
Instance segmentation, a task distinguishing contour, location, class and the number of objects in an image, is attracting
more and more attention from researchers in image processing
and computer vision. As a complex problem can hardly be
4
solved using traditional algorithms, a growing number of deep
learning approaches have emerged to solve it. For example,
SDS [14] uses a framework that resembles R-CNN [23] to
extract features from both the bounding box of the region and
the region foreground, and then classifies region proposals
and refines the segmentation inside bounding boxes based
on those extracted features. Hypercolumn [15] defines pixel
features as a vector of activations of all CNN units above
that pixel, and then classifies region proposals and refines
region segmentation based on those feature vectors. MNC [16]
integrates three networks designed for detection, segmentation
and classification respectively in a cascaded structure. Unlike
SDS and Hypercolumn, MNC is capable of training in an
end-to-end fashion, since MNC takes advantage of the Region
Proposal Network (RPN) to generate region proposals. Similar
to SDS and hypercolumn, MNC performs segmentation inside
the proposal box as well. In contrast to the above methods,
our method performs segmentation and instance recognition
in a parallel manner.
B. Gland instance segmentation
Gland morphology and structure can vary significantly,
which poses a big challenge in gland instance segmentation.
Researchers have come up with several methods to solve this
problem [24], [25], [12], [26]. Previous works focus on detecting gland structure like nuclei and lumen. Sirinukunwattana
et al. [27] model every gland as a polygon in which the
vertices are located at the nucleus. Cheikh et al. [28] propose
a mathematical morphology method to characterize the spatial
distribution of nuclei in histological images. Nguyen et al.
[29] use texture and structural features to classify the basic
components of glands, and then segment gland instance based
on prior knowledge of gland structure. These methods perform
well in benign images but are comparatively unsatisfactory
when used on malignant images, which has been the impetus
for creating methods based on deep learning [27]. Li et al. [30]
train a window-based binary classifier to segment glands using
both CNN features and hand-crafted features. Kainz et al. [26]
train two separated networks to recognize glands and glandseparating structures respectively. In MICCAI 2015 gland segmentation challenge contest, some teams achieved impressive
performance. DCAN [24] is a multitask learning framework
that combines a down-sampling path and an up-sampling
path together. From the hierarchical layer, the framework is
separated into two branches to generate contour information
and segment objects. Team ExB [12] proposes a multipath
convolutional neural network segmentation algorithm. Each
path consists of different convolutional layers and is designed
to capture different features. All paths are fused by two fully
connected layers to integrate information. Team Freburg [12]
utilizes an off-the-shelf deep convolutional neural network Unet [31], and then performs post-processing of hole-filling
and removes objects less than 100 pixels wide from the final
results.
C. Previous work
An earlier conference version of our approach was presented
in Xu et al. [32]. Here we further illustrate that: (1) we explore
another channel - object detection - in this paper, due to the
edge detection and the object detection channels complementing each other; (2) ablation experiments are carried out to
corroborate the effectiveness of the proposed algorithm; (3)
based on the rotation invariance of histological images, a new
data augmentation strategy is proposed that has proven to be
effective; (4) this algorithm achieves state-of-the-art results on
the dataset provided by the 2015 MICCAI Gland Segmentation
Challenge Contest.
IV. M ETHOD
There are two possible failures for gland instance segmentation. Since the gland-separating tissues are relatively few
and similar to glands in coloration, it is very difficult for
segmentation to rule out those pixels completely. Although
it has little effect on segmentation, it is detrimental to the
instance recognition process. Only one pixel that connects
two glands can mislead the algorithm into recognizing that
they belong to the same gland. Another possible scenario is
that algorithms designed to recognize instances separately may
cause prediction areas to be smaller than the ground truth. In
this case, the objects number and position may be accurate,
but the segmentation performance is substandard. Those two
scenarios are illustrated in Fig. 4.
Fig. 4. Two possible failures of gland instance segmentation. In the instance
results and the ground truth images, different color regions represent different
gland instances. Case 1 and Case 2 are two possible scenarios in which the
algorithm fails to segment gland instances. In Case 1, glands are separated
from the background but instances are not recognized. In Case 2, instances
are labeled yet under the condition of many gland pixels being neglected.
5
We propose a new multichannel algorithm to achieve gland
segmentation and gland instance recognition simultaneously.
Our algorithm consists of three channels and each of them
is designed to undertake different responsibilities. In the
proposed algorithm, we generate one kind of label of the input
image for each channel. Fig. 2 presents the flow chart of the
proposed algorithm. One channel is designed to segment foreground pixels from background pixels. The other two channels
are used to recognize instances. Aiming to determine which
gland each foreground pixel belongs to, we utilize both object
detection and edge detection to define spatial limits of every
gland. The reason for choosing these two channels is based on
the fact that information on contour and location contributes
respectively and complimentarily to instance recognition and
the joint effort will perform much better together than each one
alone. Specifically, edge detection performs a little better than
object detection in instance recognition, but edge detection
fails to complete the task because of the aforementioned
coalescence phenomenon of glands, which affects not only
segmentation but edge detection as well. Gland detection
may perform well for benign and well-shaped glands, but
hardly detect the entire glands accurately for malignant ones.
However, edge detection and object detection can compensate
for each other’s weaknesses and identify instances better. By
integrating the information generated from different channels,
our multichannel framework is capable of instance segmentation. A detailed depiction of our algorithm is presented in Fig
5.
A. Foreground Segmentation Channel
The foreground segmentation channel distinguishes glands
from the background.
The well-suited solutions to image labeling/segmentation in
which each pixel is assigned a label from a pre-specified set
are FCN family models [20], [21]. FCN replaces the fullyconnected layer with a convolutional layer and upsamples the
feature map to the same size as the original images through deconvolution thus an end-to-end training and prediction is guaranteed. Compared to the previous prevalent method, sliding
window [33], [34] in image segmentation, FCN is faster and
simpler. Usually, an FCN model can be regarded as the combination of a feature extractor and a pixel-wise predictor. A
pixel-wise predictor predicts probability masks of segmented
images. The feature extractor is able to abstract high-level
features by down-sampling and convolution. Though useful
high-level features are extracted, details of images sink in the
process of max-pooling and strided convolution. Consequently,
when objects are adjacent to each other, FCN may consider
them as one. Applying FCN to segment images is a logical
choice but instance segmentation is beyond the ability of FCN.
It requires an algorithm to differentiate instances of the same
class even when they are extremely close to each other. Even
so, probability masks produced by FCN still offer valuable
support in solving instance segmentation problems.
To compensate for the resolution reduction of feature maps
due to downsampling, FCN introduces skip architecture to
combine deep semantic information and shallow appearance
information. Nevertheless, Yu et al. [35] propose the dilated
convolution that empowers the network with a wider receptive
field without downsampling. Less downsampling means less
space-invariance brought by downsampling which is beneficial
to increasing segmentation precision.
Our foreground segmentation channel is a modified version
of the FCN-32s [20] of which the strides of pool4 and pool5
are 1 and subsequent convolution layers enlarge the receptive
field with a dilated convolution.
Given an input image X and the parameter of the FCN
network is denoted as ws , thus the output of FCN is
Ps (Y ∗ = k | X; ws ) = µk (hs (X, ws )) ,
(7)
where µ(·) is the softmax function. µk (·) is the output of the
kth category and hs (·) outputs the feature map of the hidden
layer. In this case, there are two categories (foreground/glands
and background), k=2. Y ∗ is the segmentation prediction.
We train the foreground segmentation channel using softmax cross entropy loss.
B. Edge Detection Channel
The edge detection channel detects boundaries between
glands.
To receive precise and clear boundaries, edges are crucial
as proven by DCAN [24]. The effectiveness of edges in
our algorithm can be shown in two ways. First, the edge
compensates for the information loss caused by max-pooling
and strided convolution in FCN. As a result, contours become
more precise and the morphology becomes more similar to the
ground truth. Second, even if the location and the probability
mask are confirmed, it is unavoidable that predicted pixel
regions of adjacent objects are still connected. Edge, however,
is able to differentiate between them. As expected, the synergy
of regions, locations and edges achieves state-of-the-art results.
The edge channel in our model is based on a Holisticallynested Edge Detector (HED) [21]. It is a CNN-based solution
towards edge detection. It learns hierarchically embedded
multiscale edge fields to account for the low-, mid-, and
high- level information of contours and object boundaries. In
edge detection, pixels of labels are much less than pixels of
backgrounds. The imbalance may decrease the convergence
rate or even cause the network being unable to convergence.
To solve the problem, deep supervision [36] is deployed. In
total, there are five side supervisions which are established
before each down-sampling layer.
We denote we as the parameter of HED, thus the mth
prediction of deep supervision is
Pe(m) (E (m)∗ = 1 | X; we ) = σ(h(m)
e (X, we )).
(8)
σ(·) denotes the sigmoid function - the output layer of HED.
(m)
he represents the output of the hidden layer relative to mth
deep supervision and E (m)∗ denotes the mth side output prediction. The weighted sum of M outputs of deep supervision
is the final result of this channel which is denoted as E ∗ , and
the weighted coefficient is α.
Pe (E ∗ = 1|X; we , α) = σ(
M
X
m=1
α(m) , h(m)
e (X, we ))
(9)
6
Fig. 5. This illustrates the structure of this algorithm. For all the channels in this algorithm, FCN for the foreground segmentation channel, Faster R-CNN [22]
for the object detection channel and HED the for edge detection channel, are all based on the VGG16 model, we present this classical five pooling structure
in detail by “Conv Net” at the left side of the figure and represent it as a rectangular block named “Conv Net”. Especially in foreground segmentation and
object detection channels, arrows pointing from “Conv Net” denote the output of the “Conv Net”, whereas in the edge detection channel they represent the
output of deep supervisions. In the foreground segmentation channel, strides of the last two pooling layers of “Conv Net” are set as 1; dilated convolution is
applied to convolution layers leading to the higher resolution of feature maps (as annotated in brackets in blue). In edge detection channel and object detection
channel, the stride of pool4 and pool5 is 2. The ‘×2’ in brackets means that there are two convolutional layers.
This process is delivered through the convolutional layer. The
back propagation enables the network to learn relative levels
of importance of edge predictions under different scales.
We train the edge detection channel using sigmoid cross
entropy loss.
C. Object Detection Channel
The object detection channel detects glands and their locations in the image.
Object detection is helpful in counting and identifying
the range of objects. According to some previous works on
instance segmentation, such as MNC [16], confirmation of the
bounding-box is usually the first step in instance segmentation.
After that, segmentation and other options are carried out
within bounding boxes. Though this method is widely recognized, the loss of context information caused by the limited
receptive field of bounding-box may exacerbate segmentation
results. Consequently, we integrate location information into
the fusion network instead of segmenting instances within
bounding boxes. To obtain location information, Faster RCNN, a state-of-the-art object detection model, is conceived.
Convolutional layers are applied to extract feature maps from
images. After that, the Region Proposal Network (RPN) takes
an arbitrary-sized feature map as input and produces a set of
bounding-boxes with the probability of objects. Region proposals will be converted into regions of interest and classified
to form the final object detection result.
Filling is done in order to transform the bounding box
prediction into a new formation that represents the number
of bounding boxes that every pixel belongs to. The value of
each pixel in regions covered by the bounding boxes equals
the number of bounding boxes it belongs to. For example, if
a pixel is in the overlapping area of three bounding boxes, the
value of that pixel will be three. wd is denoted as the parameter
of Faster R-CNN and φ represents the filling operation. The
output of this channel is
Pd (X, wd ) = φ (hd (X, wd )) .
(10)
hd (·) is the predicted coordinate of the bounding box.
We train the object detection channel using the same loss
as in Faster R-CNN [22]: the sum of a classification loss and
a regression loss.
D. Fusing Multichannel
Merely receiving the information of these three channels is
not the ultimate purpose of our algorithm. As a result, a fusion
algorithm is of great importance to maximize synergies of the
three kinds of information - region, location and boundary
cues. It is hard for an algorithm which is not learning-based
to recognize the patterns of all this information. Naturally, a
CNN based solution is the best choice.
After obtaining outputs of these three channels, a shallow
seven-layer convolutional neural network is used to combine
information and yield the final result. To reduce information
loss and ensure a sufficiently large reception field, we again
replace downsampling with dilated convolution. The architecture of fusion network is designed by cross validation.
We gradually increase the number layers and filters until the
performance no longer improves.
7
We denote wf as the parameter of this network and hf as
the hidden layer. Thus the output of the network is
P (YI∗ = k | Ps , Pd , Pe ; wf ) = µk (hf (Ps , Pd , Pe , wf )) .
(11)
As mentioned above, in this case, there are two categories,
k=2. YI∗ is the instance segmentation prediction.
We train the fusion network using softmax cross entropy
loss.
V. E XPERIMENT
A. Dataset
Our method is evaluated on the dataset provided by the
MICCAI 2015 Gland Segmentation Challenge Contest [12].
The dataset consists of 165 labeled colorectal cancer histological images scanned by Zeiss MIRAX MIDI. The image
resolution is approximately 0.62m per pixel. Original images
are in different sizes, while most of them are 775 × 522. 85
images belong to the training set and 80 are part of test sets
(test set A contains 60 images and test set B contains 20
images). There are 37 benign sections and 48 malignant ones
in the training set, 33 benign sections and 27 malignant ones
in testing set A and 4 benign sections and 16 malignant ones
in testing set B.
B. Data augmentation and Preprocessing
We first preprocess data by performing per channel zero
mean. The next step is to generate edge labels from region labels and perform dilation on edge labels afterwards. A bounding box for a gland is the smallest rectangle that can encircle
k
k
)
,ymax
the gland. Bounding box ground truth (xkmin ,xkmax ,ymin
k
can be generated from segmentation label, in which, xmin =
k
=
min(Px |P ∈ Rk ), xkmax = max(Px |P ∈ Rk ), ymin
k
min(yx |P ∈ Rk ), and ymax = max(Py |P ∈ Rk ). Rk is the
kth region of the instance ground truth and P denotes a
pixel point in Rk . Px and Py represent the X-coordinate
and Y-coordinate of P . Whether a pixel is an edge or not
is decided by its four nearest pixels (over, below, right and
left) in the region label. If all four pixels in the region
label belong to the foreground or in the background, this
pixel does not belong to any edge. To enhance performance
and combat overfitting, copious amounts of training data are
needed. Given the circumstance of the absence of a large
dataset, data augmentation is essential before training. Two
strategies for data augmentation have been carried out and the
improvement of results is strong enough evidence to prove
the efficiency of data augmentation. In Strategy I, horizontal
flipping and rotation operation (0◦ , 90◦ , 180◦ , 270◦ ) are
used in training images. Besides operations in Strategy I,
Strategy II also includes elastic transformation, such as pin
cushion transformation and barrel transformation. Deformation
of original images is beneficial to increasing robustness and the
promotion of the final result. Since the fully-connected layer
is replaced by convolutional layer, FCN takes arbitrary size
images as testing inputs. After data augmentation, a 400 × 400
region is randomly cropped from the original image as input.
C. Hyperparameter
CAFFE [37] is used in our experiments. Experiments are
carried out on K40 GPU and the CUDA edition is 7.0. The
weight decay is 0.002, the momentum is 0.9. While training
the foreground labeling/segmentation channel of the network,
the learning rate is 103 and the parameters are initialized
by pre-trained FCN32s model [20], while the edge detection
channel is trained under the learning rate of 109 and the Xavier
initialization is performed. object detection channel is trained
under the learning rate of 10−3 and initialized by pretrained
Faster R-CNN model. Fusion is learned under the learning rate
of 10−3 and initialized by Xavier initialization.
D. Evaluation
The evaluation method is the same as the competition
requires. Three indicators are used to evaluate the performance
on test A and test B. Indicators assess detection results,
segmentation performance and shape similarity respectively.
The final score is the summation of six rankings and the
smaller the better. Since image amounts of test A and test B
have a significant difference in quantity, we not only calculate
the rank sum as the host of MICCAI 2015 Gland Segmentation
Challenge Contest demands, but we also list the weighted rank
sum. We calculate the weighted average of three evaluation
criteria on test set A and test set B. Since the images in test
A account for 3/4 of the test set and images in test B account
for 1/4, the weighted rank sum is calculated as:
X
X
W eightedRS = 3/4
testARank + 1/4
testBRank.
(12)
The evaluation program is given by the MICCAI 2015 Gland
Segmentation Challenge Contest [12]. The first criterion is
the F1 score, which reflects gland detection accuracy. The
segmented glandular object of True Positive (TP) is the object
that shares more than 50% of areas with the ground truth.
Otherwise, the segmented area will be determined as a False
Positive (FP). Objects of ground truth without corresponding
prediction are considered as False Negatives (FN).
F1
Score =
2 · P recision · Recall
P recision + Recall
(13)
TP
TP + FP
(14)
P recision =
Recall =
TP
TP + FN
(15)
Dice is the second criterion for evaluating segmentation
performance. The dice index of the whole image is
D(G, S) =
2(| G ∩ S |)
,
|G|+|S|
(16)
of which G represents the ground truth and S is the segmented
result. Unfortunately, it is not able to differentiate instances of
the same class. Further, we denote G as a set of all ground
truth objects and S as a set of all segmented objects. Si
denotes the ith segmented object in an image and Gi denotes
a ground truth object that maximally overlaps Si in the image.
8
TABLE I
P ERFORMANCE IN C OMPARISON TO OTHER M ETHODS
Method
FCN [20]
dilated FCN [38]
Ours
CUMedVision2 [24]
ExB3 [12]
ExB2 [12]
ExB1 [12]
Frerburg2 [31]
Frerburg1 [31]
CUMedVision1 [24]
CVIP Dundee
LIB
CVML
vision4GlaS
1
2
Part
Score
0.788
0.854
0.893
0.912
0.896
0.892
0.891
0.870
0.834
0.868
0.863
0.777
0.652
0.635
F1 Score
A
Part
Rank Score
11
0.764
9
0.798
3
0.843
1
0.716
2
0.719
4
0.686
5
0.703
6
0.695
10
0.605
7
0.769
8
0.633
12
0.306
13
0.541
14
0.527
B
Rank
4
2
1
6
5
9
7
8
11
3
10
14
12
13
Part
Score
0.813
0.879
0.908
0.897
0.886
0.884
0.882
0.876
0.875
0.867
0.870
0.781
0.644
0.737
ObjectDice
A
Part
Rank Score
11
0.796
6
0.825
1
0.833
2
0.781
3
0.765
4
0.754
5
0.786
7
0.786
8
0.783
10
0.800
9
0.715
12
0.617
14
0.654
13
0.610
B
Rank
4
2
1
8
9
10
5
6
7
3
11
13
12
14
ObjectHausdorff
Part A
Part B
Score
Rank
Score
Rank
95.054
11
146.2478
4
62.216
9
118.734
2
44.129
1
116.821
1
45.418
2
160.347
9
57.350
6
159.873
8
54.785
3
187.442
11
57.413
7
145.575
3
57.093
4
148.463
6
57.194
5
146.607
5
74.596
10
153.646
7
58.339
8
209.048
13
112.706
13
190.447
12
155.433
14
176.244
10
107.491
12
210.105
14
RS1
WRS2
45
30
8
28
33
41
32
37
46
40
59
76
75
80
27.75
19.5
4.5
9.5
13.75
15.75
16.5
17.75
23
23.5
27.25
37.5
39.25
39.5
RS is the abbreviation for rank sum.
WRS is the abbreviation for weighted rank sum.
e i denotes the ith ground truth object in and image and Sei
G
denotes a segmented object that maximally overlaps in the
image. As a result, an object-level dice score is employed to
evaluate segmentation results. The definition is as follows:
"n
#
nG
S
X
X
e i , Sei ) ,
Dobject (G, S) = 1/2
wi D(Gi , Si ) +
w
ei D(G
i=1
i=1
(17)
| Si |
wi = PnS
,
j=1 | Sj |
(18)
ei |
|G
.
w
ei = PnG
e
j=1 | Gj |
(19)
nS and nG are the numbers of instances in the segmented
results and the ground truth.
Shape similarity reflects the performance on morphology
likelihood which plays a significant role in gland instance
segmentation. Hausdorff distance is exploited to evaluate shape
similarity. To assess glands respectively, the index of Hausdorff distance deforms from the original formation:
H(G, S) = max supinf kx − yk , supinf kx − yk ,
xG yS
yS xG
(20)
to the object-level formation:
"n
#
nG
s
X
X
e i , Sei ) ,
Hobject (S, G) = 1/2
wi H(Gi , Si ) +
w
ei H(G
i=1
where
i=1
(21)
|Si |
wi = PnS
,
j=1 |Sj |
(22)
ei |
|G
w
ei = PnG
.
e
j=1 |Gj |
(23)
Similar to the object-level dice, index nS and nG represent
instances of segmented objects and the ground truth.
E. Result and Discussion
Table I lists results of our proposed algorithm, FCN, dilated FCN and other participants on datasets provided by the
MICCAI 2015 Gland Segmentation Challenge Contest.
In the table, RS and WRS denote rank sum and weighted
rank sum respectively. We rearrange the scores and ranks in
this table. Our method outranks FCN, dilated FCN and other
participants based on both rank sum and weighted rank sum.
Compared to FCN and dilated FCN, our algorithm obtains
better scores which is convincing evidence that our work is
more effective in solving instance segmentation problems in
histological images. Though dilated FCN performs better than
FCN as the dilated convolution process has less pooling and
covers larger receptive fields, our algorithm combines region,
location and edge information to achieve higher scores in the
dataset. The reason our algorithm ranks higher is because
most adjacent glandular structures have been separated, which
is more beneficial to meet the evaluation index of instance
segmentation, whereas in FCN and dilated FCN they are not.
Comparison results are illustrated in Fig. 6.
Ranks of test A are generally higher than test B due
to the inconsistency of data distribution. In test A, most
images are normal ones whereas test B contains a majority
of cancerous images which are more complicated in shape
and larger in size. Hence, a larger receptive field is required
in order to detect cancerous glands. However, before we
exploit dilated convolution, the downsampling layer not only
gives the network a larger receptive field but also makes the
resolution of the feature map decrease, thus it deteriorates
the segmentation results. Dilated convolution empowers the
convolutional neural network with a larger receptive field
with fewer downsampling layers. Our multichannel algorithm
enhances performance based on the dilated FCN by adding
two channels - edge detection and object detection.
Since the differences between background and foreground
in histopathological images are small (3th row of Fig. 6),
FCN and dilated FCN sometimes predict the background pixel
9
Fig. 6. From left to right: original image, ground truth, results of FCN [20], FCN with dilated convolution and the proposed algorithm. Compared to FCN and
dilated FCN, most adjacent glandular structures are separated (as shown inside the red solid boxes) which indicates that our algorithm accomplishes instance
segmentation. Besides, our algorithm is able to correctly judge the small isolated area as non-gland area (as shown inside the red dotted boxes). However, a
few glands that are broken apart escape the detection of our model (as shown inside the black boxes). The bad performance in the last row is due to the fact
that in most samples the white area is recognized as cytoplasm whereas in this sample, the white area is the background.
as gland, raising the false positive rate. The multichannel
algorithm abates the false positive by adding pixel context
while predicting object location.
Compared to CUMedVision1 [24], CUMedVision2 [24]
adds edge information which improves the results of test A
but those of test B deteriorate. Our method improves results
of test A and test B after combining edge and location context.
However, white regions in gland histopathological images
are of two kinds: 1) cytoplasm; and 2) no cell or tissue (background). The difference between these two is that cytoplasm
usually appears surrounded by nuclei or other stained tissue.
In the image of the last row in Fig. 6, glands encircle some
white regions with no existence of cell or tissue causing the
algorithm to mistake them for cytoplasm. As for images of
the 4th and 5th row in Fig. 6, glands are split when cutting
images, which is the reason that cytoplasm is mistaken for
background.
Comparison with instance segmentation methods Cur-
rently, methods suitable for instance segmentation of natural
scene images predict instances based on detection or proposal,
such as SDS [14], Hypercolumn [15] and MNC [16]. One
problem with this logic is its dependence on the precision
of detection or proposal. If the object or a certain pixel of
an object escapes the detection, it will evade the subsequent
segmentation as well. Besides, the segmentation being restricted to a certain bounding box will have little access to
context information hence it impacts the result. Under the
condition of bounding boxes overlapping one another, which
instance the pixel in the overlapping region belongs to cannot
be determined. The overlapping area falls into the category of
the nearest gland in our experiment. The experiment results
are presented in Fig. 7.
To further demonstrate the defect of the cascade architecture, we design a baseline experiment. We first perform gland
detection and then segment gland instances inside bounding
10
Fig. 7. From left to right: original image, ground truth, results of SDS [14], Hypercolumn [15], MNC [16] and the proposed algorithm. Different color regions
represent different gland instances. SDS, Hypercolumn and MNC all perform masking inside bounding boxes produced by object detection, which causes the
coarse boundary of gland instances and even neglects some glands. In the second column, one gland instance is missed by manual labeling but our algorithm
successfully detects its location and segments it with relatively complete shape, yet SDS, Hypercolumn and MNC fail to detect this gland (as shown inside
of the red boxes).
TABLE II
C OMPARISON WITH INSTANCE SEGMENTATION METHODS
Method
HyperColumn [15]
MNC [16]
SDS [14]
BOX->dilated FCN [38]+EDGE3
OURS
F1 Score
Part A
Part B
0.852
0.691
0.856
0.701
0.545
0.322
0.807
0.700
0.893
0.843
boxes. There is a shallow network (same as the fusion network) combining foreground segmentation and edge detection
information to generate the final result. Configurations of all
experiments are set the same as our method. Results are shown
in Table II and less effective than the proposed algorithm.
F. Ablation Experiment
1) Data Augmentation Strategy: Data augmentation contributes to performance enhancement and overfitting elimination. We observe through experiments that adequate transformation of gland images is beneficial to training. This is
because glands naturally form in various shapes and cancerous
glands are more different in morphology. Here we evaluate the
effect on results of the foreground segmentation channel using
Strategy I and Strategy II (as shown in Table III).
2) Plausibility of Channels: In convolutional neural networks, the main purpose of downsampling is to enlarge the
receptive field, but this comes at a cost of decreased resolution
and information loss of original data. Feature maps with low
resolution increase the difficulty of upsample layer training.
The representational ability of feature maps is reduced after
upsampling and further leads to inferior segmentation results.
ObjectDice
Part A
Part B
0.742
0.653
0.793
0.705
0.647
0.495
0.790
0.696
0.908
0.833
ObjectHausdorff
Part A
Part B
119.441
190.384
85.208
190.323
116.833
229.853
114.230
197.360
44.129
116.821
TABLE III
DATA AUGMENTATION S TRATEGY COMPARISON
F1 Score
ObjectDice
ObjectHausdorff
Part A Part B Part A Part B Part A
Part B
FCN [20]
0.709 0.708 0.748 0.779 129.941 159.639
Strategy I
dilated FCN [38] 0.820 0.749 0.843 0.811 79.768 131.639
FCN [20]
0.788 0.764 0.813 0.796 95.054 146.248
Strategy II
dilated FCN [38] 0.854 0.798 0.879 0.825 62.216 118.734
Strategy
Method
Another drawback of downsampling is the space invariance
it introduces whereas segmentation is space sensitive. The
inconsistence between downsampling and image segmentation
is obvious. Dilated convolution empowers the convolutional
neural network with larger receptive field with less downsampling layers.
The comparison between segmentation performances of
FCN with and without dilated convolution shows its effectiveness in enhancing segmentation precision. The foreground
segmentation channel with dilated convolution improves the
performance of the multichannel algorithm. So does the fusion
stage with dilated convolution.
Pixels belonging to the edge occupy an extremely small
11
proportion of the whole image. The imbalance between edge
and non-edge poses a significant barrier to network training
that the network may not convergent. Edge dilation can
alleviate the imbalance and improve edge detection precision.
To prove that these three channels truly improve instance
segmentation performance, we conduct the following two
baseline experiments: a) we launch a foreground segmentation
channel and an edge detection channel; b) we launch a foreground segmentation channel and an object detection channel.
The results favor the three-channel algorithm. Results from the
experiments mentioned above are presented in Table IV.
TABLE IV
P LAUSIBILITY OF C HANNELS .
F1 Score
ObjectDice
Part A Part B Part A Part B
MC: FCN + EDGE1 + BOX
0.863 0.784 0.884 0.833
MC: FCN + EDGE3 + BOX
0.886 0.795 0.901 0.840
MC: dilated FCN + EDGE3 + BOX 0.890 0.816 0.905 0.841
DMC: FCN + EDGE3 + BOX
0.893 0.803 0.903 0.846
DMC: dilated FCN + EDGE3 + BOX 0.893 0.843 0.908 0.833
DMC: dilated FCN + EDGE1 + BOX 0.876 0.824 0.894 0.826
DMC: dilated FCN + BOX
0.876 0.815 0.893 0.808
DMC: dilated FCN + EDGE3
0.874 0.816 0.904 0.832
Method
ObjectHausdorff
Part A Part B
57.519 108.825
49.578 100.681
47.081 107.413
47.510 97.440
44.129 116.821
50.028 123.881
50.823 132.816
46.307 109.174
We denote DMC as the fusion network with dilated convolution [38] and
MC as the fusion network without dilated convolution. EDGE1 represents
that edge label are not dilated whereas EDGE3 represents that edge label
are dilated by a disk filter with radius of 3. BOX indicates that the method
includes object detection [22]. FCN [20] and dilated FCN [38] indicates that
the method includes foreground segmentation.
VI. C ONCLUSION
We propose a new algorithm called deep multichannel
neural networks. The proposed algorithm exploits features of
edge, region and location in a multichannel manner to generate instance segmentation. We observe state-of-the-art results
on the dataset from the MICCAI 2015 Gland Segmentation
Challenge. A series of baseline experiments are conducted to
prove the superiority of this method.
In future work, this algorithm can be expanded to instance
segmentation of other medical images.
ACKNOWLEDGMENT
We thank the MICCAI 2015 Gland Segmentation Challenge
for providing dataset. We thank Zhuowen Tu for all the help.
R EFERENCES
[1] W. D. Travis et al., “International association for the study of lung
cancer/american thoracic society/european respiratory society international multidisciplinary classification of lung adenocarcinoma,” Journal
of Thoracic Oncology, vol. 6, no. 2, pp. 244–285, 2011.
[2] K. Nguyen, A. Sarkar, and A. K. Jain, “Structure and context in prostatic
gland segmentation and classification,” in MICCAI. Springer, 2012, pp.
115–123.
[3] Y. Al-Kofahi et al., “Improved automatic detection and segmentation of
cell nuclei in histopathology images,” IEEE Transactions on Biomedical
Engineering, vol. 57, no. 4, pp. 841–852, 2010.
[4] M. Veta et al., “Automatic nuclei segmentation in h&e stained breast
cancer histopathology images,” PloS one, vol. 8, no. 7, p. e70221, 2013.
[5] S. Dimopoulos et al., “Accurate cell segmentation in microscopy images
using membrane patterns,” Bioinformatics, vol. 30, no. 18, pp. 2644–
2651, 2014.
[6] S. Naik et al., “Gland segmentation and computerized gleason grading
of prostate histology by integrating low-, high-level and domain specific
information,” in MIAAB workshop, 2007, pp. 1–8.
[7] K. Nguyen, A. K. Jain, and R. L. Allen, “Automated gland segmentation
and classification for gleason grading of prostate tissue images,” in
ICPR, 2010, pp. 1497–1500.
[8] S. Naik et al., “Automated gland and nuclei segmentation for grading of
prostate and breast cancer histopathology,” in ISBI, 2008, pp. 284–287.
[9] A. Paul and D. P. Mukherjee, “Gland segmentation from histology
images using informative morphological scale space,” in ICIP, 2016,
pp. 4121–4125.
[10] J. Egger, “Pcg-cut: graph driven segmentation of the prostate central
gland,” PloS one, vol. 8, no. 10, p. e76645, 2013.
[11] A. B. Tosun and C. Gunduz-Demir, “Graph run-length matrices for
histopathological image segmentation,” IEEE Trans. Medical Imaging,
vol. 30, no. 3, pp. 721–732, 2011.
[12] K. Sirinukunwattana et al., “Gland segmentation in colon histology
images: The glas challenge contest,” Medical Image Analysis, vol. 35,
pp. 489–502, 2016.
[13] M. Fleming et al., “Colorectal carcinoma: pathologic aspects,” Journal
of gastrointestinal oncology, vol. 3, no. 3, pp. 153–173, 2012.
[14] B. Hariharan et al., “Simultaneous detection and segmentation,” in
ECCV, 2014, pp. 297–312.
[15] ——, “Hypercolumns for object segmentation and fine-grained localization,” in CVPR, 2015, pp. 447–456.
[16] J. Dai, K. He, and J. Sun, “Instance-aware semantic segmentation via
multi-task network cascades,” in CVPR, 2016, pp. 3150–3158.
[17] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in NIPS, 2012, pp. 1097–
1105.
[18] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” in ICLR, 2015.
[19] R. Girshick, “Fast r-cnn,” in ICCV, 2015, pp. 1440–1448.
[20] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” in CVPR, 2015, pp. 3431–3440.
[21] S. Xie and Z. Tu, “Holistically-nested edge detection,” in ICCV, 2015,
pp. 1395–1403.
[22] S. Ren et al., “Faster r-cnn: Towards real-time object detection with
region proposal networks,” in NIPS, 2015, pp. 91–99.
[23] R. Girshick et al., “Rich feature hierarchies for accurate object detection
and semantic segmentation,” in CVPR, 2014, pp. 580–587.
[24] H. Chen et al., “Dcan: Deep contour-aware networks for accurate gland
segmentation,” in CVPR, 2016, pp. 2487–2496.
[25] L. Jin, Z. Chen, and Z. Tu, “Object detection free instance segmentation
with labeling transformations,” arXiv preprint arXiv:1611.08991, 2016.
[26] P. Kainz, M. Pfeiffer, and M. Urschler, “Semantic segmentation of
colon glands with deep convolutional neural networks and total variation
segmentation,” arXiv preprint arXiv:1511.06919, 2015.
[27] K. Sirinukunwattana, D. R. Snead, and N. M. Rajpoot, “A stochastic
polygons model for glandular structures in colon histology images,”
IEEE Trans. Medical Imaging, vol. 34, no. 11, pp. 2366–2378, 2015.
[28] B. B. Cheikh, P. Bertheau, and D. Racoceanu, “A structure-based
approach for colon gland segmentation in digital pathology,” in SPIE,
2016, pp. 97 910J–97 910J.
[29] K. Nguyen, B. Sabata, and A. K. Jain, “Prostate cancer grading:
Gland segmentation and structural features,” Pattern Recognition Letters,
vol. 33, no. 7, pp. 951–961, 2012.
[30] W. Li et al., “Gland segmentation in colon histology images using handcrafted features and convolutional neural networks,” in ISBI, 2016, pp.
1405–1408.
[31] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
for biomedical image segmentation,” in MICCAI. Springer, 2015, pp.
234–241.
[32] Y. Xu et al., “Gland instance segmentation by deep multichannel side
supervision,” in MICCAI, 2016, pp. 496–504.
[33] P. Sermanet et al., “Overfeat: Integrated recognition, localization and
detection using convolutional networks,” in ICLR, 2014.
[34] D. Ciresan et al., “Deep neural networks segment neuronal membranes
in electron microscopy images,” in NIPS, 2012, pp. 2843–2851.
[35] F. Yu and V. Koltun, “Multi-scale context aggregation by dilated
convolutions,” arXiv preprint arXiv:1511.07122, 2015.
[36] C.-Y. Lee et al., “Deeply-supervised nets,” in AISTATS, 2015, pp. 562–
570.
[37] Y. Jia et al., “Caffe: Convolutional architecture for fast feature embedding,” in Proceedings of the 22nd ACM international conference on
Multimedia. ACM, 2014, pp. 675–678.
[38] L.-C. Chen et al., “Deeplab: Semantic image segmentation with deep
convolutional nets, atrous convolution, and fully connected crfs,” arXiv
preprint arXiv:1606.00915, 2016.
| 1cs.CV
|
RAMIFICATION OF WILD AUTOMORPHISMS OF LAURENT SERIES FIELDS
arXiv:1611.01077v1 [math.NT] 3 Nov 2016
KENZ KALLAL AND HUDSON KIRKPATRICK
Abstract. We study the structure of the Nottingham group N (Fp ) of power series f ∈ X · Fp [[X]]
with f ′ (0) = 1 where p > 2 is prime, which is isomorphic to the group of wild automorphisms of
Fp ((X)). We concern ourselves with power series g ∈ N (Fp ) with g ◦m 6= X for all m ≥ 1 (that
is, power series of infinite order in N (Fp )), and we determine a necessary and sufficient criterion
for such g ∈ N (Fp ) having ramification type in (g) = 3(1 + · · · + pn ) given a finite number of
coefficients of g. We also conjecture analogous results for higher ramification based on the results
of a computer program, generalizing Fransson’s 2016 theorem on 2-ramified power series, as well as
the work on minimally ramified power series used in a 2013 result of Lindahl and a 2015 result of
Lindahl–Rivera-Letelier.
1. Introduction
Let p be an odd prime, and consider the set of formal power series over Fp .
Definition 1.1. N (Fp ) is the group under composition of power series of the form
f (X) = X +
∞
X
ai X i ∈ Fp [[X]]
i=2
with coefficients ai ∈ Fp .
It is easy to see that N (Fp ) is a group by explicitly computing compositional inverses from a given
sequence of coefficients [7], and it is isomorphic to the group of wild automorphisms of Fp ((X)) [17].
N (Fp ), also known as the Nottingham group, is an example of a pro-p group (a compact topological
group which is an inverse limit of p-groups) which has recently received a considerable amount of
attention. The study of pro-p groups was originally motivated by number theory in the context of
Galois cohomology and notably discussed by Serre [20]. Since the 1960s, the Nottingham group has
become increasingly studied as a prototypical example and test case for group theorists studying
pro-p groups due to the ease of computation in the Nottingham group, which essentially amounts to
compositions of power series [3]. In particular, the Nottingham group and the computations associated
with it were introduced to group theory by Jennings [7], Johnson [8] and York [22, 21]. One of the
most important recent group-theoretic results on the Nottingham group is due to Leedham-Green–
Weiss and Camina [2]; it states that every finite p-group can be embedded in N (Fp ), and every
countably-based pro-p group can be embedded as a closed subgroup of N (Fp ).
We will only consider power series f ∈ N (Fp ) such that f ◦m (X) 6= X for all m ≥ 1 (that is,
elements of N (Fp ) of infinite order). This is because of a celebrated theorem of Klopsch [10] which
Date: November 3, 2016.
2010 Mathematics Subject Classification. 11S15, 11D88 (Primary), 20E36, 20E18 (Secondary).
1
classified all elements of N (Fp ) of order p up to conjugacy, and was generalized by Jean [6] and Lubin
[16] to order pn (see also [1]).
The number-theoretic work on the Nottingham group relies on ramification theory, which we now
introduce in part.
Definition 1.2 (Valuation). If f (X) ∈ Fp ((X)), then the X-adic valuation of f , valX (f ), is the
integer i such that f (X) ∈ X i Fp [[X]] and f (X) 6∈ X i+1 Fp [[X]]. If f is identically zero, then we write
that valX (f ) = +∞.
Definition 1.3 (Ramification). The ramification number of f ∈ Fp ((X)) is
i(f ) := valX (f (X) − X) − 1.
n
We will frequently mention the “nth ramification number” of f , that is, in (f ) := i(f ◦p ). It is
immediate that conjugation in N (Fp ) preserves ramification. That is, for all f, g ∈ N (Fp ),
i(g ◦ f ◦ g −1 ) = i(f ).
It is therefore natural to consider the power series with certain sequences of ramification numbers and
to classify them up to conjugacy.
Question 1.4. Let b ∈ N. Which conjugacy classes of power series f of infinite order in N (Fp ) have
ramification of the form in (f ) = b(1 + p + · · · + pn ) for all n ≥ 0 with p ∤ b?
We will justify the choice of ramification of this form (which we will write as “b-ramified”) with
Theorem 2.5, a powerful result of Laubie and Saïne which restricts it to that form under certain
conditions. Laubie, Movahhedi and Salinier [12] made progress on Question 1.4 in the case b = 1 by
using advanced tools in number theory to prove that all such elements of N (Fp ) must be conjugate
to powers of each other over some extension of Fp . Motivated by trying to understand power series
with ramification of this form for b ≥ 1, we consider the following question:
Question 1.5. Let b ∈ N with p ∤ b. Given a sequence {ai }i≥1 of coefficients in Fp , does
f (X) = X +
∞
X
ai X i+b ∈ N (Fp )
i=1
n
have ramification in (f ) = b(1 + · · · + p ) for all n ≥ 0?
We will write a1 , . . . , an as defined in Question 1.5 as the “first n nontrivial coefficients of f .”
Fransson [5] solved Question 1.5 for b = 2 by giving a finite polynomial condition on the ai ’s. This
itself is a generalization of the characterization of minimally ramified power series, which has been
used to prove many important theorems in nonarchimedean dynamical systems [12, 15, 13, 14, 18].
We prove the following result, answering Question 1.5 in the case b = 3:
P∞
Theorem 1.6. Let f (X) = X + i=1 ai X i+3 ∈ N (Fp ) be of infinite order in N (Fp ) for some prime
p > 3. Then in (f ) = 3(1+· · ·+pn ) for all n ≥ 0 if and only if a1 6= 0 and 2a41 −a32 +2a1 a2 a3 −a21 a4 6= 0.
Motivated by results of a computer program (see Appendix A), we also conjecture in Conjecture 4.2
that for all b ∈ N, there exists a polynomial Pb ∈ Z[x1 , . . . , xb+1 ] such that f ∈ N (Fp ) with i(f ) = b
is b-ramified if and only if Pb is nonzero when evaluated at the first b + 1 nontrivial coefficients of
f (as defined in Question 1.5) and p ∤ b. This reduces finding a polynomial criterion modulo large
2
products of primes for b-ramification to a finite computation. Assuming that Conjecture 4.2 is true,
these computations (also carried out by the SAGE program in Appendix A) give the following results:
Theorem 1.7. Assume Conjecture 4.2, and let b ∈ N and suppose
f (X) = X +
∞
X
ai X i+b ∈ N (Fp )
i=1
with i(f ) = b (i.e. a1 6= 0). Then there exists a polynomial Pb (X) ∈ Z[x1 , . . . , xb+1 ] such that
in (f ) = b(1 + · · · + pn ) for all n ≥ 0 if and only if p ∤ b and Pb (a1 , . . . , ab+1 ) 6= 0, where the first few
values of Pb modulo the product of the first 100 such primes (which is greater than 10200 ) are listed
below.
P1 = 2x21 − 2x2
P2 = 3x31 + 2x22 − 2x1 x3
P3 = 4x41 − 2x32 + 4x1 x2 x3 − 2x21 x4
P4 = 5x51 + 2x42 − 6x1 x22 x3 + 2x21 x23 + 4x21 x2 x4 − 2x31 x5
P5 = 6x61 − 2x52 + 8x1 x32 x3 − 6x21 x2 x23 − 6x21 x22 x4 + 4x31 x2 x5 − 2x41 x6
P6 = 7x71 + 2x62 − 10x1 x42 x3 − 11x21 x22 x23 − 2x31 x33 + 8x21 x32 x4 + 11x31 x2 x3 x4 + 2x41 x24
− 6x31 x22 x5 + 4x41 x3 x5 + 4x41 x2 x6 − 2x51 x7 .
Our program is capable of computing the Pb ’s from Theorem 1.7 in a reasonable amount of time
for b ≤ 50, though the number of terms becomes too large to display in this paper.
This paper is organized as follows. In Section 2, we describe the historical number-theoretic results
on ramification numbers of power series leading up to our work. In Sections 3 and 4, we give an explicit
computation of a necessary and sufficient criterion for power series of ramification in (f ) = 3(1+· · ·+pn )
for all n ≥ 0 and conjecture similar results in higher ramification based on the computations of our
computer program in Appendix A. In Section 5, we outline future directions for work on this problem.
2. Ramification of Power Series of Infinite Order
We set out with the goal of understanding the restrictions on the sequences of ramification numbers
of elements of infinite order in N (Fp ). The most important historical result with regard to this is due
to Sen [19, Theorem 1]:
Theorem 2.1 (Sen, 1969). Let f ∈ N (Fp ). Then for all n ∈ N,
in (f ) ≡ in−1 (f )
(mod pn ).
The proof of Sen’s Theorem can be done entirely with computations on power series. The following
corollary immediately follows from power series computations and Sen’s Theorem.
Corollary 2.2. Let f ∈ N (Fp ). Then for all n ≥ 0,
(2.3)
in (f ) ≥ 1 + p + · · · + pn .
3
Power series f ∈ N (Fp ) for which inequality holds in (2.3) are called minimally ramified. Much
more difficult to prove is Keating’s Theorem [9, Theorem 7], which gives a general form for in (f ) given
the first two ramification numbers. under certain conditions.
Theorem 2.4 (Keating, 1992). Suppose f ∈ N (Fp ) has infinite order. If i0 (f ) = 1 and i1 (f ) = 1 + bp
with 1 ≤ b ≤ p − 2, then in (f ) = 1 + bp + · · · + bpn .
We will use a powerful generalization of Theorem 2.4 from [11, Theorem 2].
Theorem 2.5 (Laubie and Saïne 1997). Let f ∈ N (Fp ) be of infinite order. Then the following hold:
(1) If p|i(f ), then in (f ) = pn i(f ) for all n ≥ 0.
(2) Otherwise, if i1 (f ) < (p2 − p + 1)i(f ), then
in (f ) = i(f ) + (1 + · · · + pn−1 )(i1 (f ) − i(f ))
It follows from Theorem 2.5 that if f ∈ N (Fp ) with i(f ) = b, i1 (f ) = b + bp and p ∤ b, then
in (f ) = b(1 + · · · + pn ). We write that a power series with ramification of this form is b-ramified, and
in answering Question 1.5 we seek to classify the b-ramified power series for certain values of b. The
special case b = 2 was solved in [5, Theorem 1]:
Theorem 2.6 (Fransson 2016). A given power series of infinite order given by f (X) = X +
P∞
i+2
∈ N (Fp ) is 2-ramified if and only if 3a33 + 2a24 − 2a2 a5 6= 0 and a1 6= 0.
i=1 ai X
We will prove a similar result for the 3-ramified power series, and conjecture a general result on all
b-ramified power series.
3. Classification of 3-Ramified Power Series
We will exploit Theorem 2.5 to classify the 3-ramified power series f (X) ∈ N (Fp ) based on a
polynomial condition on the first four non-trivial coefficients of f . For the remainder of this section,
we let
f (X) = X + a1 X 4 + a2 X 5 + · · · ∈ N (Fp )
be of infinite order in N (Fp ) for some prime p > 3 with a1 6= 0, so that i(f ) = 3.
Define the power series ∆m (X) such that
∆m+1 = ∆m ◦ f − ∆m ,
with ∆1 = f (X) − X. Note that by induction, valX (∆m ) ≥ 3m + 1. The following lemma is also true
by induction:
Lemma 3.1. Suppose m ∈ N. Then,
∆m (X) =
m
X
(−1)i
i=0
m ◦(m−i)
f
(X)
i
Since Fp has characteristic p, this corollary is immediate:
Corollary 3.2. ∆p (X) = f ◦p (X) − X.
4
Therefore, we let
∆m (X) = Am X 3m+1 + Bm X 3m+2 + Cm X 3m+3 + Dm X 3m+4 + · · · ,
and by Definition 1.3, i1 (f ) = 3(1 + p) if and only if valX (∆p ) = 3p + 4. Hence, by Theorem 2.5, f
is 3-ramified if and only if Ap = Bp = Cp = 0 and Dp 6= 0 in Fp . We wish to compute the values of
these first four coefficients of ∆p (X), so we find that modulo hX 3m+8 i,
∆m+1 (X) = Am (X + a1 X 4 + · · · )3m+1
+ Bm (X + a1 X 4 + · · · )3m+2
+ Cm (X + a1 X 4 + · · · )3m+3
+ Dm (X + a1 X 4 + · · · )3m+4
− (Am X 3m+1 + Bm X 3m+2 + Cm X 3m+3 + Dm X 3m+4 )
= (Am (3m + 1)a1 )X 3m+4
+ (Am (3m + 1)a2 + Bm (3m + 2)a1 )X 3m+5
+ (Am (3m + 1)a3 + Bm (3m + 2)a2 + Cm (3m + 3)a3 )X 3m+6
3m + 1
+ Am (3m + 1)a4
+ Am a21
2
!
+ Bm (3m + 2)a3 + Cm (3m + 3)a2 + Dm (3m + 4)a1 X 3m+7 .
From this, we obtain the recurrence
(3.3)
Am+1
Bm+1
Cm+1 =
Dm+1
Am
a1 (3m+1)
0
0
0
Bm
a2 (3m+1)
a1 (3m+2)
0
0
Cm
a3 (3m+1)
a2 (3m+2) a1 (3m+3)
0
Dm
+a4 (3m+1) a3 (3m+2) a2 (3m+3) a1 (3m+4)
a21 3m+1
2
with initial conditions Am = a1 , Bm = a2 , Cm = a3 , Dm = a4 . Now, we solve this recurrence for
the individual coefficients to compute Ap , Bp , Cp , Dp . The solutions follow from basic tools in solving
difference equations (see [4]), and we ultimately obtain that if n!!! is defined by 0!!! = 1!!! = 2!!! = 1
and n!!! = n(n−3)!!!, then the following hold:
(3.4)
Am = am
1 (3m−2)!!!
(3.5)
a2 am−1
1
Bm =
(3m−1)!!!+
m−1
X
r=1
Cm = a3 am−1
(3m)!!!+a3 am−1
1
1
(3r+1)!!!(3m−1)!!!
(3r+2)!!!
m−1
X
r=1
+a22 am−2
1
(3.6)
m−1
X
r=1
(3.7)
(3.8)
!
m−1
X (3m)!!!(3r+2)!!!
(3m)!!!(3r+1)!!!
+a22 am−2
1
(3r+3)!!!
(3r+3)!!!
r=1
r−1
(3m)!!! X (3n+1)!!!(3r+2)!!!
(3r+3)!!! n=1
(3n+2)!!!
Dm = α(m)am+1
+β(m)am−3
a32 +γ(m)am−2
a2 a3 +δ(m)am−1
a4
1
1
1
1
α(m) =
m−1
X
r=1
3r(3m+1)!!!
2(3r+4)
5
(3.9)
β(m) =
m−1
X
r=1
γ(m) =
m−1
X
r=1
(3.10)
(3.11)
n−1
r−1
X
X
(3m+1)!!!
(3j +1)!!!(3n+2)!!!
(3r+3)!!!
(3n+2)!!!+
(3r+4)!!! n=1 (3n+3)!!!
(3j +2)!!!
j=1
(3m+1)!!!
(3r+2)!!!+(3r+3)!!!
(3r+4)!!!
!
r−1
X
(3n+1)!!!(3r+2)!!! (3n+1)!!!(3r+3)!!!
+
+
(3n+2)!!!
(3n+3)!!!
n=1
δ(m) = (3m+1)!!!+
m−1
X
r=1
(3m+1)!!!
3r+4
It follows from (3.4), (3.5) and (3.6) that Ap = Bp = Cp = 0 in Fp . And since a1 6= 0, Fermat’s
Little Theorem gives
Dm = α(m)a41 +β(m)a32 +γ(m)a1 a2 a3 +δ(m)a21 a4 .
It is our goal to compute the coefficients α(p), β(p), γ(p), δ(p).
First, we define the inverse of 3 in Fp to be
(
(2p+1)/3, if p ≡ 1
t=
(p+1)/3, if p ≡ 2
(mod 3)
(mod 3)
.
We will repeatedly exploit the following simple fact.
Lemma 3.12. Let y ∈ {0, 1, 2} and x ∈ Z. Then in Fp ,
(3x+y)!!! = 3x
(x+yt)!
.
(yt)!
All equalities in this section are stated in Qp and reduced to Fp by dividing by nonzero elements
in Fp and cancelling factors of p. We will also write that nk = 0 when k < 0 or k > n.
We will use the following two lemmas repeatedly in our computations for the coefficients of Dp .
Lemma 3.13. Let 0 < r < p−1. Then
(r+1)! =
(−1)r
.
(p−r−2)!
The proof of Lemma 3.13 follows from Wilson’s Theorem and a simple computation in Fp .
Lemma 3.14.
p−1
X
r=1
(p+t−1)!
(r+1)! = 0.
(r+t+1)!
Proof. p+t−1
r+t+1 is non-zero in Fp when r+t+1 ≥ p, so applying Lemma 3.13 and separately considering
the r = p−1 term, we have
p−1
p−2
X
X
(p+t−1)!
p+t−1
(−1)r
(r+1)! =
(r+t+1)!
r+t+1
r=1
r=p−t−1
6
t−1
X
X
t−1
p+t−1
r−p−t−1 t−1
=
(−1)
=
(−1)
p+r
r
r=0
r=0
t−1
X
t−1
= 0.
= (−1)t
(−1)r
r
r=0
r−p−t−1
Proposition 3.15. α(p) = 2 in Fp .
Proof. By (3.8) and Lemma 3.12,
α(p) =
p−1
p−1
X
3r(3p+1)!!! X 3r(p+t)!
=
.
2(3r+4)
2(r+t+1)t!
r=1
r=1
The summand is nonzero only when r+t+1 = p so we have
3(p−t−1)(p+t)!
3t+3
=−
·(p−1)! = (−2)(−1) = 2
2p·t!
2
α(p) =
by Wilson’s Theorem.
Part of the computation of β(p) is given by the following lemma:
Lemma 3.16.
p−1
r−1
XX
(p−t)! t+p+1
(p+t)! (r+1)!
= 0.
·
·(n+2t)!
(r+t+1)! (n+1)!
(p+t)!
2t
r=1 n=1
Proof. By simple algebraic manipulations, we see that the sum above can be written
and we let
p−1 r−1
(p−t)! (p+t+1)! X X (p+t)! (r+1)! (n+2t)!
·
·
·
,
(p+t)! (p−t+1)! r=1 n=1 (r+t+1)! (n+1)! (2t)!
p+t+1
(p−t)! (p+t+1)!
·
=
∈ Fp ,
(p+t)! (p−t+1)!
p−t+1
which is constant with respect to the indices. Hence our original sum reduces to
c=
c
which is zero by Lemma 3.14.
p−1 X
r−1
X
(p+t)! (r+1)! (n+2t)!
·
·
(r+t+1)!
(n+1)! (2t)!
r=1 n=1
Proposition 3.17. β(p) = −1 in Fp .
Proof. Applying (3.9) gives that β(p) = β1 (p)+β2 (p) where
β1 (p) =
p−1
X
r=1
and
β2 (p) =
p−1
X
r=1
r−1
(3p+1)!!! X (3r+3)!!!
(3n+2)!!!
(3r+4)!!! n=1 (3n+3)!!!
r−1
n−1
(3p+1)!!! X (3r+3)!!! X (3j +1)!!!(3n+2)!!!
.
(3r+4)!!! n=1 (3n+3)!!! j=1
(3j +2)!!!
7
First, by Lemma 3.12
β1 (p) =
=
p−1
r−1
XX
(3p+1)!!!
(3r+3)!!!
·
·(3n+2)!!!
(3(r+1)+1)!!!
(3n+3)!!!
r=1 n=1
p−1
r−1
XX
(p+t)! (r+1)! (n+2t)!
·
·
.
(r+t+1)! (n+1)! (2t)!
r=1 n=1
Continuing our simplification, we have
1 n+2t
(n+2t)!
=
(2t)!(n+1)!
2t n+1
so that
r−1
X
r−1
(n+2t)!
1
r+2t
1 X n+2t
=
−1−2t .
=
(2t)!(n+1)!
2t n=1 n+1
2t
r
n=1
This means that by Lemma 3.14 and the Hockey Stick Identity,
p−1
p−1
1+2t X (p+t)!
t X (p+t−1)!(r+1)! r+2t
−
(r+1)!
β1 (p) =
2t r=1
(r+t+1)!
2t r=1 (r+t+1)!
r
p−1
p−1
1 X (p+t−1)!(r+1)! r+2t
1 X (p+t−1)!(r+2t)!
=
(r+1)
=
2 r=1
(r+t+1)!
2 r=1 (r+t+1)!(2t)!
r
p−1
(p+t−1)! X (r+2t)!
(r+t+1−t)
2(2t)! r=1 (r+t+1)!
p−1
r+2t
r+2t
(p+t−1)!t! X
−
=
r+t
2(2t)!
r+t+1
r=1
(p+t−1)!t!
p+2t
2t+1
p+2t
2t+1
=
−
−
+
.
2(2t)!
p+t−1
t
p+t
t+1
=
It trivially follows by expansion of each of the four terms that β1 (p) = 0.
On the other hand,
β2 (p) =
p−1 X
r−1 n−1
X
X (p+t)! (r+1)! (n+2t)! (k+t)!
·
·
(r+t+1)! (n+1)! (k+2t)! t!
r=1 n=1
k=1
p−1 X
r−1 n−1
X
X (p+t)! (r+1)! (n+2t)! (k+t+p)!
·
·
=
(r+t+1)! (n+1)! (k+2t)! (p+t)!
r=1 n=1 k=1
!
p−1 X
r−1
n−1
X
(p+t)! (r+1)!
(p−t)! X k+t+p
=
·
·(n+2t)!
(r+t+1)! (n+1)!
(p+t)!
k+2t
r=1 n=1
k=1
p−1 X
r−1
X
n+t+p
t+p+1
(p−t)!
(p+t)! (r+1)!
−
.
·
·(n+2t)!
=
(r+t+1)! (n+1)!
(p+t)!
n−1+2t
2t
r=1 n=1
By Lemma 3.16,
p−1 X
r−1
X
(p−t)! t+p+1
(p+t)! (r+1)!
=0
·
·(n+2t)!
(r+t+1)! (n+1)!
(p+t)!
2t
r=1 n=1
8
so from now on we will only consider
β2 (p) =
p−1
r−1
XX
(p+t)! (r+1)!
(p−t)! n+t+p
.
·
·(n+2t)!
(r+t+1)! (n+1)!
(p+t)! n−1+2t
r=1 n=1
Note that
(n+2t)! n+t+p
(n+t+p)!
(n+t+p)!
(n+2t)!
·
·
= (n+2t)·
.
=
(n+1)! n−1+2t
(n+1)! (n−1+2t)!(p−t+1)!
(n+1)!(p−t+1)!
Now accounting for the
(p−t)!
(p+t)!
factor, we have
(p−t)!
(n+t+p)!
n+2t (n+t)!
·(n+2t)·
=
·
.
(p+t)!
(n+1)!(p−t+1)!
p−t+1 t!(n+1)!
Hence,
β2 (p) =
p−1
r−1
XX
n+2t (n+t)!
(p+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
p−1
r−1
XX
(p+t)!
n+2t+(t−1)−(t−1) (n+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
p−1
r−1
XX
(n+t)!
t−1
(p+t)!
n+2t−(t−1)
·
·(r+1)!·
+
=
(r+t+1)!
p−t+1
p−t+1 t!(n+1)!
r=1 n=1
=
=
+
p−1
r−1
XX
(p+t)!
n+2t−(t−1) (n+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
p−1
r−1
XX
t−1
(n+t)!
(p+t)!
·(r+1)!·
·
.
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
Breaking the sum apart, we define
R1 =
and
p−1
r−1
XX
n+2t−(t−1) (n+t)!
(p+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
R2 =
p−1 X
r−1
X
(p+t)!
t−1
(n+t)!
·(r+1)!·
·
.
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
We will show R1 = 0 and R2 = −1. First,
R1 =
p−1
r−1
XX
(p+t)!
n+2t−(t−1) (n+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
p−1
r−1
XX
(p+t)!
1
(n+t+1)!
·(r+1)!·
·
(r+t+1)!
p−t+1 t!(n+1)!
r=1 n=1
p−1
r−1
X (p+t)!
X
n+t+1
1
·(r+1)!
=
p−t+1 r=1 (r+t+1)!
n+1
n=1
p−1
X (p+t)!
1
r+t+1
=
·(r+1)!·
p−t+1 r=1 (r+t+1)!
r
=
9
p−1
X (p+t)!
1
=
·(r+1) = 0.
p−t+1 r=1 (t+1)!
Computing R2 , we have
R2 =
=
p−1
r−1
XX
(p+t)!
t−1
(n+t)!
·(r+1)!·
·
(r+t+1)!
p−t+1
t!(n+1)!
r=1 n=1
p−1
r−1
XX
(p+t)!
t−1 (n+t)!
·(r+1)!·
·
(r+t+1)!
−t+1
t!(n+1)!
r=1 n=1
=−
p−1 X
r−1
X
(p+t)!
(n+t)!
·(r+1)!·
(r+t+1)!
t!(n+1)!
r=1 n=1
p−1
r−1
X
(p+t)!
(n+t)!
1X
·(r+1)!
(r+t+1)!
t
(t−1)!(n+1)!
r=1
n=1
p−1
r−1
X (p+t)!
X n+t
= −3
·(r+1)!
n+1
(r+t+1)!
r=1
n=1
p−1
X (p+t)!
r+t
= −3
·(r+1)!
(r+t+1)!
r
r=1
=−
p−1
3 X (p+t)!
·(r+1).
=−
t! r=1 r+t+1
Every term in this sum is 0 except when r+t+1 = p, so
p−1
3 (p+t)!
3
3 X (p+t)!
·(r+1) = − ·
·(p−t)+0 = − ·(−t!)·(−t) = −3t = −1
−
t! r=1 r+t+1
t!
p
t!
Therefore we finally have β(p) = β1 (p)+β2 (p) = R1 +R2 = −1.
Proposition 3.18. γ(p) = 2 in Fp .
Proof. We compute each of the four terms in the sum given in (3.10). First, by Lemma 3.12,
p−1
X
r=1
p−1
X (p+t)! (r+2t)!
(3p+1)!!!
(3r+2)!!! =
·
(3r+4)!!!
(r+t+1)! (2t)!
r=1
p−1
X r+2t (p+t)!(t−1)!
·
=
t−1
(2t)!
r=1
p−1
(p+t)!(t−1)! X r+2t
=
(2t)!
t−1
r=1
(p+t)!(t−1)!
p+2t
2t+1
=
−
(2t)!
t
t
(p+t)!(t−1)! (p+2t)! (2t+1)!
−
=
(2t)!
(p+t)!t! (t+1)!t!
10
=
−p −p(2t+1)
−
= 0.
t
t(t+1)
Moreover, by Lemma 3.14 we have
p−1
X
(3p+1)!!!
r=1
(3r+4)!!!
(3r+3)!!! = 3
p−1
X
r=1
(p+t)!(r+1)!
= 0.
(r+t+1)!
Computing the double sums, we let
γ1 (p) =
p−1
r−1
XX
(3p+1)!!! (3r+2)!!!
·
·(3n+1)!!!
(3r+4)!!! (3n+2)!!!
r=1 n=1
p−1
r−1
XX
(p+t)!(r+2t)!(n+t)!
(r+t+1)!(n+2t)!t!
r=1 n=1
p−1
r−1
XX
(n+t)!
r+2t
(−p)
=
·(t−1)!
(n+2t)!
r+t+1
r=1 n=1
p−1
r−1
X r+2t X
(p+n+t)!
(t−1)!
=
(n+2t)!
r+t+1 n=1
r=1
p−1
r−1
X r+2t X
(p+n+t)!
= (−1)t
r+t+1 n=1 (n+2t)!(p−t)!
r=1
p−1
r−1
X r+2t X
p+n+t
t
= (−1)
r+t+1 n=1 n+2t
r=1
p−1
p−1
X r+2t
X r+2t p+r+t
t p+t+1
t
.
−(−1)
= (−1)
r+t+1
2t
r+t+1 r+2t−1
r=1
r=1
=
Applying the Hockey Stick Identity to the second term in the sum and checking both cases for t
depending on the value of p (mod 3) immediately gives that the second term is zero. Thus,
γ1 (p) = (−1)t
p−1
X
r=1
= (−1)t
p−1
X
r=1
=
r+2t
p+r+t
r+t+1 r+2t−1
(r+2t)!(p+r+t)!
(r+t+1)!(t−1)!(r+2t−1)!(p−t+1)!
p−1
X (r+2t)(−p)
(−1)t
.
(t−1)!(p−t+1)! r=1 (r+t+1)
Using the fact that (t−1)! = (−1)t /(p−t)! (which follows from Lemma 3.13), we get
p−1
γ1 (p) =
X (r+2t)(−p)
1
(p−t+1) r=1 (r+t+1)
The only nonzero term in the sum occurs when r+t+1 = p, which gives γ1 (p) = (1−t)/(1−t) = 1.
To compute the second double sum given by (3.10), we use Lemma 3.12 again to get
γ2 (p) =
p−1
r−1
XX
(3p+1)!!! (3r+3)!!!
·
·(3n+1)!!!
(3r+4)!!! (3n+3)!!!
r=1 n=1
11
p−1
r−1
XX
(p+t)!(r+1)!(n+t)!
(r+t+1)!(n+1)!t!
r=1 n=1
p−1
r−1
X (p+t)!(r+1)! X
n+t
=
t(r+t+1)! n=1 n+1
r=1
p−1
X (p+t)!(r+1)! r+t
−(t+1)
=
t(r+t+1)!
r
r=1
=
=
p−1
X
r=1
p−1
X
(p+t−1)!(r+1)!(r+t)!
(p+t−1)!(r+1)!
−(t+1)
.
(r+t+1)!r!t!
(r+t+1)!
r=1
The second sum is zero by Lemma 3.14, so
γ2 (p) =
p−1
X
r=1
−p(r+1)
.
t(r+t+1)
The only nonzero term in this sum occurs when r+t+1 = p, so γ2 (p) = (t−p)/t = 1. Therefore,
γ(p) = γ1 (p)+γ2 (p) = 2.
Proposition 3.19. δ(p) = −1 in Fp .
Proof. We compute that
δ(p) = (3p+1)!!!+
p−1
X
p−1
(3p+1)!!! X (p+t)!
(p+t)!
=
=
= (p−1)! = −1
3r+4
(r+t+1)t!
t!
r=1
r=1
by the same argument used in Lemma 3.15.
Theorem 1.6 follows from (3.7) and Propositions 3.15, 3.17, 3.18, and 3.19.
4. Generalizing the Classification
Given a positive integer b and an odd prime p such that p ∤ b, our SAGE program (given in
Appendix A) directly generalizes the logic in the computation of (3.3) to b-ramified elements f ∈
N (Fp ), giving b+1 polynomial conditions on the coefficients of f . For f defined as in Question 1.5,
we give the following recurrence for the first b+1 nontrivial coefficients of ∆m , which is as defined in
Section 3:
a1 (bm+1)
0
0···
0
A1,m
A1,m+1
A
A
a2 (bm+1)
a1 (bm+2)
0···
0
2,m
2,m+1
..
..
..
..
..
..
=
(4.1)
.
.
.
.
.
.
Ab,m
Ab,m+1
ab (bm+1)
···
a1 (bm+b)
0
Ab+1,m
Ab+1,m+1
+ab+1 (bm+1) ab (bm+2)
···
a1 (bm+b+1)
a21 bm+1
2
(note that the initial conditions are, by definition, An,1 = an ). In particular, An,m denotes the degree(bm+n) coefficient of ∆m , and we do not list the others since valx (∆m ) is trivially at least bm+1 by
the same argument used in the proof of (3.3).
12
It follows that by Corollary 3.2, a power series given by
f (X) = X +
∞
X
ai X i+b ∈ N (Fp )
i=1
is therefore b-ramified if and only if A1,p , . . . , Ab,p = 0 and Ab+1,p 6= 0. Our SAGE program treats
a1 , . . . , ab+1 as the indeterminants x0 , . . . , xb and computes the coefficients A1,p , . . . , Ab+1,p as polynomials in Fp [x0 , . . . , xb ] by repeatedly applying (4.1). Running this program for the first 100 odd
primes p such that p ∤ b for b ∈ {1, 2, . . . , 10}, the results suggest the following conjecture:
Conjecture 4.2. For all positive integers b, there exists a polynomial Pb ∈ Z[x1 , . . . , xb+1 ] such that
P
i+b
any f (X) = X + ∞
∈ N (Fp ) of infinite order is b-ramified if and only if p ∤ b, a1 6= 0 and
i=1 ai X
Pb (a1 , . . . , ab+1 ) 6= 0.
In other words, the first b polynomials given by the generalized recurrence (4.1) are identically
zero, and the coefficients of the last polynomial do not depend on p as long as p ∤ b. Although the
first part of this conjecture is most likely relatively simple to prove given that the recurrence is easily
generalizable, the second part, while supported by all of our numerical computations, gives much less
certainty in what direction a proof would proceed.
Assuming Conjecture 4.2 is true, it is possible for our SAGE program to compute the generalized
recurrence for a single prime q in order to guarantee the values of the coefficients of Pb modulo q.
Applying the program to a few large primes guarantees the coefficients modulo a very large product of
primes by the Chinese Remainder Theorem, which is likely to be the actual value of Pb . This makes
it easy to compute probable values of Pb for b up to around 50, each using a reasonably small amount
of computation time. Those results are omitted from this paper due to the large numbers of terms in
the resulting polynomials, though the first few are included in the statement of Theorem 1.7, which
indexes the indeterminants starting at x1 for readability.
5. Conclusion
In this paper, we use methods in number theory to prove a characterization of power series in
N (Fp ) with ramification restricted to the form b(1+· · ·+pn ) for certain values of b, as given by the
conclusion of a powerful theorem of Laubie and Saïne. We find a full proof of Theorem 1.6 for b = 3,
and give a conjecture that results in very probable criteria for higher values of b in Theorem 1.7, as
well as the ability to compute similar criteria for arbitrary ramification types of this form.
The numerical results given in Theorem 1.7 immediately show some patterns in the coefficients of
Pb , which, after considering more numerical data, could lead to a conjecture on the general form of
the ramification criteria in the future. Moreover, our conjecture that the first b polynomials given by
the recurrence are identically zero can most likely be proved in the general case from the tools we use
to prove Theorem 1.6 given the fact that the recurrences for those polynomials are given in a simple
and easily generalizable form. As we did for the 3-ramified case, it will most likely take an explicit
computation in the general case to determine that the coefficients are fixed with respect to p.
13
Appendix A. SAGE Program
The SAGE code below generates the b+1 polynomials from the generalized recurrence (4.1), which
give a necessary and sufficient criterion for b-ramification in N (Fp ) for a given b ∈ N and an odd prime
p such that p ∤ b. In particular, A(b, m, p) generates the matrix given by (4.1), and V (p, b) returns a
vector containing the polynomials A1,p , . . . , Ab+1,p .
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
def A (b , m , p ) :
R = P o l y n o m i a l R i n g ( GF ( p ) , b +1 , ’x ’)
M = [[0 for i in range ( b +1) ] for j in range ( b +1) ]
for i in range ( b +1) :
for j in range ( i +1) :
M [ i ][ j ] = R . gen (i - j )* ( b * m + j + 1)
M [ b ][0] = ( b * m + 1) * ( b * m ) /( GF ( p ) (2) )* ( R . gen (0) ) ^2
+ ( b * m + 1) * ( R . gen ( b ) )
return matrix (R , M )
def V (p , b ) :
R = P o l y n o m i a l R i n g ( GF ( p ) , b +1 , ’x ’)
a = []
for i in range ( b +1) :
a . append ( R . gen ( i ) )
v = vector (R , a )
for m in range (1 , p ) :
v = A (b , m , p ) * v
return ( v )
Acknowledgements
The results in this paper originated from a research lab project in PROMYS 2016. We are deeply
grateful to Laurent Berger and Sandra Rozensztajn for proposing this problem, and for their mentoring. We also thank our PROMYS research lab counselor Alyosha Latyntsev for his support, as well as
Krishanu Sankar, David Fried, Glenn Stevens, the PROMYS Foundation, and the Clay Mathematics
Institute for making this research possible.
References
[1] F. M. Bleher, T. Chinburg, B. Poonen, and P. Symonds. Automorphisms of Harbater-Katz-Gabber curves. ArXiv
e-prints, September 2015. arXiv:1509.02139.
[2] Rachel Camina. Subgroups of the nottingham group. Journal of Algebra, 196(1):101–113, 1997.
[3] M. du Sautoy, D. Segal, and A. Shalev. New Horizons in pro-p Groups. New Horizons in Pro-p Groups. Birkhäuser
Boston, 2000.
[4] Saber Elaydi. An introduction to difference equations. Springer Science & Business Media, 2005.
[5] J. Fransson. Complete classification of 2-ramified power series. ArXiv e-prints, January 2016. arXiv:1601.03622.
[6] Sandrine Jean. Conjugacy classes of series in positive characteristic and witt vectors. Journal de ThÃľorie des
Nombres de Bordeaux, 21(2):263–284, 2009.
[7] S.A. Jennings. Substitution groups of formal power series under substitution. Canad. J. Math, 6:325 – 340, 1954.
[8] D. L. Johnson. The group of formal power series under substitution. Journal of the Australian Mathematical
Society. Series A. Pure Mathematics and Statistics, 45(3):296–302, 04 2009.
14
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
Kevin Keating. Automorphisms and extensions of k((t)). Journal of Number Theory, 41(3):314 – 321, 1992.
Benjamin Klopsch. Automorphisms of the nottingham group. Journal of Algebra, 223(1):37 – 56, 2000.
F Laubie and M Saïne. Ramification of automorphisms of k((t)). Journal of Number Theory, 63(1):143 – 145, 1997.
François Laubie, Abbas Movahhedi, and Alain Salinier. Systèmes dynamiques non archimédiens et corps des norms
(non-archimedean dynamic systems and fields of norms). Compositio Mathematica, 132(1):57–98, 2002.
Karl-Olof Lindahl. The size of quadratic p-adic linearization disks. Advances in Mathematics, 248:872 – 894, 2013.
Karl-Olof Lindahl and Juan Rivera-Letelier. Optimal cycles in ultrametric dynamics and minimally ramified power
series. Compositio Mathematica, 152(1):187–222, 009 2015.
Jonathan Lubin. Nonarchimedean dynamical systems. Compositio Mathematica, 94(3):321–346, 1994.
Jonathan Lubin. Torsion in the nottingham group. Bulletin of the London Mathematical Society, 43(3), 2011.
Benjamin Muckenhoupt. Automorphisms of formal power series under substitution. Transactions of the American
Mathematical Society, 99(3):373–383, 1961.
J. Riviera-Letelier. Dynamique des fonctions rationelles sur des corps locaux. PhD thesis, Universite de Paris XI,
2000.
Shankar Sen. On automorphisms of local fields. Annals of Mathematics, 90(1):33–46, 1969.
J.P. Serre. Galois Cohomology. Berlin, 1997.
I.O. York. The exponent of certain p-groups. Proc. Edinburgh Math. Soc., 33:483–490, 1990.
I.O. York. The group of formal power series under substitution. PhD thesis, Nottingham University, 1990.
15
| 4math.GR
|
A new class of metrics for spike trains
Cătălin V. Rusu1,2,3 and Răzvan V. Florian1
arXiv:1209.2918v3 [cs.IT] 17 Jan 2014
1
Center for Cognitive and Neural Studies (Coneural), Romanian
Institute of Science and Technology, Cluj-Napoca, Romania
2
Computer Science Department, Babeş-Bolyai University,
Cluj-Napoca, Romania
3
Frankfurt Institute for Advanced Studies, Frankfurt am Main,
Germany
Abstract
The distance between a pair of spike trains, quantifying the differences between them, can be measured using various metrics. Here we introduce a new
class of spike train metrics, inspired by the Pompeiu-Hausdorff distance, and
compare them with existing metrics. Some of our new metrics (the modulusmetric and the max-metric) have characteristics that are qualitatively different than those of classical metrics like the van Rossum distance or the Victor
& Purpura distance. The modulus-metric and the max-metric are particularly
suitable for measuring distances between spike trains where information is encoded in bursts, but the number and the timing of spikes inside a burst does
not carry information. The modulus-metric does not depend on any parameters and can be computed using a fast algorithm, in a time that depends linearly on the number of spikes in the two spike trains. We also introduce localized versions of the new metrics, which could have the biologically-relevant
interpretation of measuring the differences between spike trains as they are
perceived at a particular moment in time by a neuron receiving these spike
trains.
An edited version has been published in Neural Computation 26(2), pp. 306–
348, 2014, http://dx.doi.org/10.1162/NECO_a_00545
Rate or review this paper at http://epistemio.com/p/r43uM8ke
1
1 Introduction
In recent years several spike train distances, some inspired by existing mathematical distances while others not, have been proposed and used to measure the variability of the neural activity (Victor and Purpura, 1996, 1997; van Rossum, 2001;
Schreiber et al., 2003; Victor, 2005; Schrauwen et al., 2007; Kreuz et al., 2007, 2009;
Naud et al., 2011; Kreuz et al., 2011, 2013). The distance between two spike trains
reflects their similarity or dissimilarity. Because how information is represented in
the spatio-temporal patterns of spike times exchanged by neurons is still a heavily
debated topic in neuroscience, metrics based on different neural codes are available. Traditionally it was thought that the mean firing rate of neurons encapsulated all the relevant information exchanged by neurons. This idea dates back to
the work of Adrian (1926) who showed that the firing rate of motor neurons is proportional to the force applied. More recently, scientists have revealed increasing
evidence of the importance of precise spike timings to representing information in
the brain (Bohte, 2004; VanRullen et al., 2005; Tiesinga et al., 2008). For example,
temporally-structured multicell spiking patterns were observed in the hippocampus and cortex, and were associated to memory traces (Nádasdy et al., 1999; Ji and
Wilson, 2007) while the coding of information in the phases of spikes relative to a
background oscillation has been observed in many brain regions (Lee et al., 2005;
Jacobs et al., 2007; Fries et al., 2007; Montemurro et al., 2008; Siegel et al., 2009).
This change in viewpoints from rate codes to spike-time codes is also reflected in
spike train metrics.
The most basic metrics are the ones that rely on counting the total number
of spikes within a spike train. A major drawback of such an approach is that all
the temporal structure is lost. Even though binning techniques were introduced
as a way to overcome this loss by dividing the spike train into discrete bins, the
temporally–encoded information within a bin was neglected (Geisler et al., 1991).
Other, more complex spike train metrics can be obtained by focusing on the precise spike timing instead of their total count. One example is the van Rossum
(2001) distance which is calculated by filtering the time series corresponding to
the raw spike train with a smoothing kernel, typically an exponential one, and
then using the standard Euclidean distance. Another metric is the Schreiber et al.
(2003) correlation-based measure, which uses a symmetric (Gaussian) filter. In
both cases, the choice of the kernel’s parameters is arbitrary and has a high influence on the properties of the metric.
Another metric was introduced by Victor and Purpura (1996, 1997). According
to this metric, the distance between two spike trains is given by the minimum cost
of basic operations needed to transform one spike train into the other. The basic
operations are insertion or deletion of spikes, with a cost of 1, and the shifting of
a spike, with a cost of q|δt | where q is a parameter and δt the shifting interval.
The parameter q significantly influences the behavior of the metric: for q = 0 the
metric counts the difference in the total number of spikes, while for large values of
q the metric returns the number of non-coincident spikes.
Kreuz et al. (2007, 2009, 2011, 2013) have introduced more recently a series of
parameter-free and time-scale independent measures of spike train synchrony.
Here we introduce a new class of spike train metrics inspired by the PompeiuHausdorff distance between two non-empty compact sets (Pompeiu, 1905; Hausdorff, 1914). Preliminary results have been presented in abstract form in (Rusu and
Florian, 2010).
2
2 A new class of spike train metrics
We consider bounded, nonempty spike trains of the form
T = {t (1) , . . . , t (n) },
(1)
where t (i ) ∈ R are the ordered spike times and n ∈ N∗ is the number of spikes in
the spike train. We consider spike trains with no overlapping spikes. If n > 1, then
t (i −1) < t (i ) , ∀i > 1. We denote by a and b some bounds of the considered spike
trains, i.e. a ≤ t (i ) ≤ b, ∀t (i ) , with a, b ∈ R, finite, and a < b. We denote by S [a,b]
the set of all possible such spike trains bounded by a and b. We study metrics that
compute the distances between two spike trains T and T̄ from S [a,b] .
The new metrics that we introduce are inspired by the Pompeiu-Hausdorff distance (Pompeiu, 1905; Hausdorff, 1914). When applied to a pair of spike trains, the
Pompeiu-Hausdorff distance h returns the largest difference, in absolute value, between the timing of a spike in one train and the timing of the closest spike in the
other spike train:
)
(
h(T, T̄ ) = max sup inf |t − t̄ |, sup inf |t − t̄ | ,
t̄ ∈T̄ t ∈T
t ∈T t̄ ∈T̄
(2)
or, equivalently, the minimal number ² ≥ 0 such that the closed ²-neighborhood of
T includes T̄ and the closed ²-neighborhood of T̄ includes T :
©
ª
h(T, T̄ ) = inf ² such that |t − t̄ | ≤ ², ∀t ∈ T, ∀t̄ ∈ T̄ .
(3)
Another equivalent form of the Pompeiu-Hausdorff distance is the following (Papadopoulos, 2005, pp. 105–110; Rockafellar and Wets, 2009, pp. 117–118; Deza and
Deza, 2009, pp. 47–48):
¯
¯
¯
¯
h(T, T̄ ) = sup ¯¯ inf |t − x| − inf |t̄ − x|¯¯ .
(4)
x∈R t ∈T
t̄ ∈T̄
We introduce a distance d between an arbitrary timing x ∈ R and a spike train
T:
d (x, T ) = inf |t − x|.
(5)
t ∈T
Eq. 2 can then be rewritten as
)
(
h(T, T̄ ) = max sup d (t , T̄ ), sup d (t̄ , T )
t ∈T
(6)
t̄ ∈T̄
and Eq. 4 as
¯
¯
h(T, T̄ ) = sup ¯d (x, T ) − d (x, T̄ )¯ .
(7)
x∈R
Because the global supremum is achieved on the interval [a, b] (Appendix A), we
also have:
¯
¯
h(T, T̄ ) = sup ¯d (x, T ) − d (x, T̄ )¯ .
(8)
x∈[a, b]
The Pompeiu-Hausdorff metric has a quite poor discriminating power, as for many
variations of the spike trains the distances will be equal and any spike train space
endowed with this metric would be highly clustered. Our new metrics generalize
the form of the Pompeiu-Hausdorff distance given in Eq. 8, by introducing features
that are more sensitive to spike timings.
3
2.1 The max-metric
We consider B to be the space of all continuous, positive functions H : [0, b − a] →
R+ that satisfy the condition, ∀u ∈ [a, b],
Z
b
a
H (|u − s|) ds > 0.
(9)
Because
Z
b
a
H (|u − s|) ds =
u
Z
H (u − s) ds +
a
Z u−a
=
0
H (s) ds +
Z
b
H (s − u) ds
(10)
H (s) ds,
(11)
u
b−u
Z
0
and because H is continuous, a sufficient condition for satisfying Eq. 9 is that
H (0) > 0.
On compact sets, continuous functions are bounded (Protter, 1998, p. 56). We
denote by m the upper bound of H on the interval [0, b − a], i.e.
0 ≤ H (x) < m < ∞, ∀ x ∈ [0, b − a].
(12)
In typical applications, H (x) has a maximum for x = 0 and is a decreasing function
of x, for example an exponential,
³ x´
1
exp − ,
τ
τ
(13)
µ
¶
1
x2
exp − 2 ,
p
2τ
τ 2π
(14)
H E (x) =
or a Gaussian,
H G (x) =
with τ a positive parameter.
We introduce the max-metric as
Z b
©
ª
d m (T, T̄ ) =
sup |d (x, T ) − d (x, T̄ )| H (|s − x|) ds.
(15)
a x∈[a,b]
The max-metric integrates, through the variation of s along the interval [a, b] that
contains the two spike trains, the maximum of the difference, in absolute value, between the distances from a point x in that interval to the two spike trains, weighted
by the kernel H (|s − x|) which focuses locally around s. Figure 1 shows how the
distance between two spike trains is computed using the max-metric.
The max-metric is a generalization of the Pompeiu-Hausdorff distance, since
in the particular case that H (·) = 1/(b − a) we have d m (T, T̄ ) = h(T, T̄ ).
In Appendix B we show that d m is finite and that it satisfies the properties
of a metric. We also show that regardless of the kernel H all the max-metrics
are topologically equivalent to each other (O’Searcoid, 2007, p. 229) because they
are equivalent to the Pompeiu-Hausdorff distance. Each metric will generate the
same topology and thus any topological property is invariant under an homeomorphism. This means that the metrics generate the same convergent sequences
in the space of spike trains S [a,b] . Thus, learning rules derived from these metrics
will converge in the same way, regardless of the choice of H .
4
A
T
B
T̄
d(x,T)
C
d(x,T̄)
D
E
F
G
0
100
200
Time (ms)
300
400
500
Figure 1: The modulus-metric and the max-metric.
(A) Spike train T =
{20, 150, 350, 400, 440} ms. Each spike time is represented as a vertical bar. (B)
Spike train T̄ = {100, 270, 300, 370, 480} ms. (C) The distances between a timing x and ¯the spike trains,¯ d (x, T ) and d (x, T̄ ), as a function of x. (D) The
difference ¯d (x, T ) − d (x, T̄ )¯ as a function of x. The modulus-metric distance
d o (T, T̄ ) is the area under this curve. (E) The kernel H (|s − x|) as a function
of s with a fixed x = 250 ms. H is an exponential
(Eq. 13)¯ with a decay con¯
stant τ = 50 ms. (F) The weighted difference ¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) as function of s for
©¯ discrete values¯ of x. (G) ªThe supremum of the weighted difference,
supx∈[a,b] ¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) , as a function of s. The max-metric distance d m (T, T̄ ) is the area under this curve. In (C), (D), and (G), the dashed vertical
lines represent the timing of spikes in T and T̄ .
5
A
T
B
T̄
f(x)
C
f̄(x)
D
E
F
G
0
100
200
Time (ms)
300
400
500
Figure 2: The convolution max-metric. (A) Spike train T = {20, 150, 350, 400, 440}
ms. Each spike time is represented as a vertical bar. (B) Spike train T̄ =
{100, 270, 300, 370, 480} ms. (C) The spike trains T and T̄ filtered with an
¯ exponen-¯
tial kernel (Eq. 18) with a decay constant τ = 10 ms. (D) The difference ¯ f (x) − f¯(x)¯
as a function of x. (E) The kernel H (|s − x|) as a function of s with a fixed x = 250
ms. H is an exponential
(Eq.¯ 13) with a decay constant τ = 50 ms. (F) The
¯
weighted difference ¯ f (x) − f¯(x)¯ H (|s − x|) as function of ©¯
s for discrete¯ values of x.
ª
(G) The supremum of the weighted difference, supx∈[a,b] ¯ f (x) − f¯(x)¯ H (|s − x|) .
The convolution max-metric distance d c (T, T̄ ) is the area under this curve. In (C),
(D), and (G), the dashed vertical lines represent the timing of spikes in T and T̄ .
6
2.2 The modulus-metric
We define the modulus-metric as
Z
d o (T, T̄ ) =
b
a
|d (s, T ) − d (s, T̄ )| ds.
(16)
Figure 1 A–D shows how the distance between two spike trains is computed using
the modulus-metric. The modulus-metric is a particular case of the max-metric in
the limit that H is
(
1, if x = 0,
H (x) =
(17)
0, otherwise.
The modulus-metric does not depend on any kernels or parameters and it also
allows a fast computer implementation in linear complexity. This is because the
graph of the function φ(s) = |d (s, T )−d (s, T̄ )| is made out of line segments that join
or end in the following points: all timings of spikes in the two spike trains T and T̄ ;
the time moments that lie at the middle of the interval between two neighboring
spikes from the same spike train; the time moments that lie at the middle of the
interval between a pair of neighboring spikes where the two spikes belong to different spike trains; and the bounds a and b. This is exemplified in Figures 1 D and
3 D. We denote by P the set of these points. In order to compute the integral of
Rb
this function d o = a φ(s) ds, it is sufficient to compute the function at the points
from P . Since between these points the function is linear, the integral can be then
computed exactly.
Algorithm 1 presents an implementation of the d o metric in pseudo-code. In
this algorithm, the function φ is computed in a set of points that includes P but
also other points. In a second Algorithm 2, the set P as well as the value of φ in
the points of P is computed with a single pass through the spikes in the two spike
trains. The algorithms’ duration depends linearly on the number of spikes in the
two spike trains, n + n̄. Implementations in Python and C++ of the two algorithms
are freely available at https://github.com/modulus-metric/.
It can be shown that the distance d o is finite and that it satisfies the properties
of a metric by particularizing the proofs in Appendix E with L (x) = 1, ∀x ∈ [0, b−a].
2.3 The convolution max-metric
The max-metric can also be given in a convolution form. To construct this form of
the metric we consider an arbitrary continuous, smooth, positive kernel K : R →
R+ , with the properties that it is strictly increasing for x < 0 and strictly decreasing
for x > 0, with K (0) = p finite and positive. We thus have 0 ≤ K (x) ≤ p, ∀x ∈ R.
Typically, K is an exponential,
¶
µ
|x|
,
(18)
K E (x) = exp −
τ
with τ being a positive parameter. We convolve the two spike trains T and T̄ with
the filtering kernel K to obtain
f (x) =
n
X
K (x − t (i ) )
(19)
K (x − t̄ (i ) ).
(20)
i =1
f¯(x) =
n̄
X
i =1
7
We denote by F[a,b] the set of all possible filtered spike trains from S [a,b] .
We also consider a function H ∈ B that is strictly positive, that is derivable on
(0, b − a) and that has a bounded derivative.
The convolution max-metric is defined as
Z b
©
ª
(21)
d c (T, T̄ ) =
sup | f (x) − f¯(x)| H (|s − x|) ds.
a x∈[a,b]
Figure 2 shows how the distance d c between two spike trains is computed. In Appendix C we show that d c is finite and that it satisfies the properties of a metric.
3 Localized metrics
In the case of the max metric, with or without convolution, the use of the kernel
H served the purpose of providing a local perspective, around each point within
[a, b], of the distance between the spike trains. These local perspectives were then
integrated in the final distance. In this section we introduce different metrics that
also depend on a kernel L , but for which the kernel has a different purpose. Here,
the kernel may be regarded as a magnifying glass to be used to focus on one specific area of the spike trains. The kernel should be a continuous, positive function,
L : [0, b − a] → R+ . Similarly with H , because L is a continuous function with
bounded support, it is bounded, i.e.
0 ≤ L (x) < m < ∞, ∀ x ∈ [0, b − a].
(22)
Such a metric is biologically relevant if, for example, we take into consideration
how a neuron responds to input spikes. Recent spikes influence more the neuron
than old ones. If we would like to measure the distance between two spike trains
according to how the differences between them influence the activity of a neuron
at a particular moment of time, recent differences should be taken into account
with a greater weight than differences in the distant past. For the localized metrics,
L could thus model the shape of postsynaptic potentials (PSP) that reflects the
dynamics of the effect of one presynaptic spike on the studied neuron. Thus, L
could typically be an exponential, L E = H E (Eq. 13), an alpha function,
L α (x) =
³ x´
x
exp
− ,
τ2
τ
a difference between two exponentials,
·
µ
¶¸
³ x´
τ
x
L D (x) =
exp − − exp −
,
τ − τs
τ
τs
(23)
(24)
or, if we model the postsynaptic potential generated in an integrate-and-fire neuron by a synaptic current that is a difference between two exponentials,
½
·
µ
¶¸
³ x´
τ
τs
x
L I (x) =
exp − − exp −
τ s − τr τ − τ s
τ
τ
·
µ s ¶¸¾
³ x´
τr
x
−
exp − − exp −
,
(25)
τ − τr
τ
τr
where τ, τs , and τr are positive parameters.
8
3.1 Localized max-metric
We introduce the localized max-metric as
Z b
d l (T, T̄ ) =
L (b − s) sup |d (x, T ) − d (x, T̄ )| ds.
a
(26)
x∈[s,b]
Figure 3 shows how the distance d l between two spike trains is computed. The differences between the spike trains that account the most for the distance are those
that are close to b. The shape of L has a high impact on the behavior of the metric.
In Appendix D we show that the distance d l is finite and that it satisfies the
properties of a metric.
3.2 Localized modulus-metric
The modulus metric can also be given in a localized form:
b
Z
d n (T, T̄ ) =
a
|d (s, T ) − d (s, T̄ )| L (b − s) ds.
(27)
In Appendix E we show that d n is finite and that it satisfies the properties of a
metric.
3.3 Localizing the van Rossum metric
A localization by a kernel L similar to the one we applied in Eqs. 26 and 27 can
also be applied to existing metrics. Let T, T̄ ∈ S [a,b] . Consider the van Rossum
(2001) distance defined as
Z ∞
d R (T, T̄ ) =
(g (s) − ḡ (s))2 ds,
(28)
−∞
where
n
X
g (s) =
H (s − t (i ) ) K E (s − t (i ) )
(29)
H (s − t̄ (i ) ) K E (s − t̄ (i ) ),
(30)
i =1
n̄
X
ḡ (s) =
i =1
H is the Heaviside step function, H (x) = 0 if x < 0 and H (x) = 1 if x ≥ 0, and K E is
defined in Eq. 18. When localized with L the distance becomes
Z
d Rl (T, T̄ ) =
b
−∞
(g (s) − ḡ (s))2 L (b − s) ds.
(31)
Here L may be chosen to have the same qualitative properties as the kernels used
in Eqs. 23–25.
4 Simulation results
We analyzed the behavior of the introduced metrics through computer simulations
using simple setups. Across all simulations, H was an exponential (Eq. 13) with
9
A
T
B
T̄
d(x,T)
C
d(x,T̄)
D
E
F
G
0
100
200
Time (ms)
300
400
500
Figure 3: The localized max-metric. (A) Spike train T = {30, 190, 200, 320, 410, 480}
ms. Each spike time is represented as a vertical bar. (B) Spike train T̄ =
{20, 120, 200, 300, 350, 460} ms. (C) The distances between a timing
¯ x and the spike¯
trains, d (x, T ) and d (x, T̄ ), as function of x. (D) The¯ difference ¯d (x,¯ T ) − d (x, T̄ )¯
as a function of x. (E) The supremum, supx∈[s,b] ¯d (x, T ) − d (x, T̄ )¯, as a function of s. (F) The kernel L (b − s) as function of s, an exponential (Eq. 13)
with a decay constant τ = 50 ms. (G) The supremum weighted by the kernel,
L (b − s) supx∈[s,b] |d (x, T ) − d (x, T̄ )|, as a function of s. The localized max-metric
distance d l (T, T̄ ) is the area under this curve. In (C), (D), and (G), the dashed vertical lines represent the timing of spikes in T and T̄ .
10
τ = 10 ms. For the localized metrics d l and d n , L also was an exponential (Eq.
13) with τ = 20 ms. The convolution kernels for the d c and van Rossum distances
were chosen as exponentials (Eq. 18) with τ = 10 ms. The width of the Gaussian
filter for the Schreiber et al. distance was 10 ms. For the Victor & Purpura distance
we set q = 0.2 ms−1 , except in Figs. 7 – 9 where we also used q = 0.001 ms−1 as in
(Dauwels et al., 2009). We set a = 0 ms and b the maximum length of spike trains
(either 200 or 500 ms, except in Fig. 12 where the length was variable).
4.1 Inserting or shifting one spike
We computed the distances between a particular spike train T and a spike train T̄
obtained from it by either inserting or shifting one spike. In the insertion case, T̄
was generated by inserting a spike into T at various timings. In the shifting case, T̄
was generated by shifting a particular spike of T . The distance was plotted against
the time of the inserted spike or of the shifted spike to see how the change is reflected by the metrics. To compute the distance we used the introduced metrics,
a simple spike count distance (c), the van Rossum (2001) (d R ), Victor and Purpura (1996, 1997) (dV P ), Schreiber et al. (2003) (s), and Pompeiu-Hausdorff (h)
distances, as well as the ISI-distance (k i ) and the improved SPIKE-distance (k s )
by Kreuz et al. (2007, 2013). The spike count distance is defined as
c(T, T̄ ) =
|n − n̄|
,
max(n, n̄)
(32)
where n and n̄ are the number of spikes in each train. The spike trains were 200
ms long.
The results for the insertion case are presented in Figure 4. The Victor & Purpura distance was constant since the cost of adding and removing a spike is fixed
at 1 regardless of its timing. Similarly, the van Rossum metric was insensitive to
the time of the inserted spike, a result which can be also shown analytically (van
Rossum, 2001). The spike count distance remained constant regardless of where
the spike was inserted. The results were qualitatively different in the case of our
newly introduced distances, with the exception of the convolution max-metric,
and in the case of the Kreuz et al. and Schreiber et al. metrics. In the case of the
Pompeiu-Hausdorff distance, max-metric, modulus-metric, of the localized variants of the max-metric and the modulus-metric, and of the Schreiber et al. and
Kreuz et al. metrics, the insertion time of the spike had a significant impact on
the outcome (Figure 4). When the inserted spike overlapped an existing spike, the
Schreiber et al. distance had a low value but remained non-zero, while our new
metrics, with the exception of the convolution max-metric, as well as the Kreuz et
al. metrics, returned a zero distance. It can also be seen that the localized distances
were strongly influenced by the shape of the L kernel.
The results for the shifting case are presented in Figure 5. When the spike at
t (4) was shifted, the Victor & Purpura and van Rossum distances were dependent
only on the width of the shifting interval. These results are confirmed by analytical derivations (Victor and Purpura, 1996; van Rossum, 2001). As in the previous
case, the spike count distance was insensitive to the shift operation and remained
zero since the number of spikes did not change. In contrast, our newly introduced
distances, with the exception of the convolution max-metric, and the Schreiber et
al. and Kreuz et al. metrics showed a dependence not only on the width of the
11
shifting interval but also on the particular timing of the shifted spike. The results
are similar to the ones in Fig. 4 of (Kreuz et al., 2011).
4.2 Bursts
We generated a template spike train T containing three bursts and one isolated
spike. We computed, in 6 setups, using various metrics, the distance between T
and another spike train T̄ obtained from T by either: removing one spike from a
burst; inserting one spike into a burst; or removing or inserting one spike not belonging to the bursts. After computing the distances for each of the setups, the
distances for each metric were normalized to the maximum distance for that particular metric among the setups. The normalized distances are shown, for each
setup, in Figure 6. When a spike is added to a burst or removed from a burst, the
max-metric and modulus-metric distances, as well as the Kreuz et al. distances,
are close to zero. Those distances become non-negligible when a solitary spike,
far from a burst, is removed, or a new spike is added far from a burst. The normalized convolution max-metric and the van Rossum distances are close to 1, and
the Victor & Purpura distance is exactly 1, in all setups. The Schreiber et al. distance exhibited an intermediate behavior, but it also remained non-negligible in
all setups.
4.3 Discriminating timing precision vs. event reliability
Timing precision and event reliability are distinct characteristics of the variation of
a spike train (Tiesinga et al., 2008). In computational neuroscience there is a need
for spike train metrics to both characterize the overall variability as well as for parsing out the precision and reliability separately (Toups et al., 2012). We investigated
how our new metrics and existing metrics perform in discriminating precision vs.
event reliability (Dauwels et al., 2009). We generated a 200 ms template spike train
T using a Poisson process with a rate of 100 Hz. We also generated various spike
trains T̄ obtained from T by both: (a) applying to spikes from T a Gaussian jitter
with zero mean and variance σ in the range of 0 and σm = 20 ms and (b) removing
spikes with a probability p ranging from 0 to p m = 80%. For each (σ, p) pair we
ran 100 trials, where in each trial we generated randomly a new T̄ . For each trial,
we computed the distances between T and T̄ using various metrics and we then
averaged the results over the trials. For each metric, average distances were then
normalized to the maximum average distance across the (σ, p) parameters. The
results are presented in Figs. 7–10. Fig. 7 represents the normalized average distances as a function of σ and p. Figs. 8 and 9 represent sections trough the graphs
in Fig. 7, for illustrating more clearly the dependencies of distances on reliability
and precision.
If Fig. 10, we tried to illustrate how reliability compares to the precision in determining the variability of the distances. For a metric ρ, it is not possible to compare directly the partial derivatives ∂ρ/∂p and ∂ρ/∂σ because they are variables
of different physical dimensions: p is dimensionless, while σ represents a time
interval. In order to compare them, we take into consideration that the relevant
intervals on which p and σ vary are practically bounded. Choosing σm and p m is
theoretically arbitrary, but in practice p m is something slightly less than 1 and the
relevant σm is constrained by the time scales of the considered spike trains and of
the distances that depend on time-scale-like parameters. We use these practical
12
T
A
dR
B
dVP
C
c
D
s
E
ks
F
ki
G
h
H
dm
I
dc
J
do
K
dl
L
dn
M
0
50
100
Time (ms)
150
200
Figure 4: Metric comparison - inserting a spike. We computed the distance between the spike train T = {20, 50, 75, 125, 180} ms and one obtained from this spike
train by inserting a spike at different locations. At each time x ∈ [0, 200] ms a spike
was inserted to generate T̄ and the distance between T and T̄ was computed and
plotted against x. (A) The spike train T . (B) The van Rossum distance. (C) The
Victor & Purpura distance. (D) The spike count distance. (E) The Schreiber et al.
distance. (F) The Kreuz et al. improved SPIKE-distance. (G) The Kreuz et al. ISIdistance. (H) The Pompeiu-Hausdorff distance. (I) The max-metric. (J) The convolution max-metric. (K) The modulus-metric. (L) The localized max-metric. (M)
The localized modulus-metric.
13
T
A
dR
B
dVP
C
c
D
s
E
ks
F
ki
G
h
H
dm
I
dc
J
do
K
dl
L
dn
M
0
50
100
Time (ms)
150
200
Figure 5: Metric comparison - shifting a spike. We computed the distance between
the spike train T = {10, 50, 75, 125, 150} ms and one obtained from this spike train
by shifting the spike at t (4) = 125 ms. The spike was shifted at timings x ∈ [0, 200]
ms to generate T̄ and the distance between T and T̄ was computed and plotted
against x. (A) The spike train T . (B) The van Rossum distance. (C) The Victor &
Purpura distance. (D) The spike count distance. (E) The Schreiber et al. distance.
(F) The Kreuz et al. improved SPIKE-distance. (G) The Kreuz et al. ISI-distance. (H)
The Pompeiu-Hausdorff distance. (I) The max-metric. (J) The convolution maxmetric. (K) The modulus-metric. (L) The localized max-metric. (M) The localized
modulus-metric.
14
Figure 6: Dealing with bursts. We measured distances between a template spike
train that includes bursts and a spike train obtained from the template by adding
or removing one spike. Arrows indicate where the spike was added or removed.
Left: the pair of spike trains. Right: the normalized distances (d o - modulus metric;
d m - max metric; d c - convolution max-metric; d R - van Rossum; dV P - Victor
& Purpura; s - Schreiber et al.; k s - Kreuz et al. improved SPIKE-distance; k i Kreuz et al. ISI-distance). (A), (B) One spike removed from a burst. (C), (D) One
spike inserted into a burst. (E) One isolated spike removed. (F) One isolated spike
inserted.
15
bounds in order to commensurate the variance with respect to the two parameters. By dividing the intervals [0, p m ] and [0, σm ] into the same number of bins N
we get two-dimensional pixels of size δp = p m /N and δσ = σm /N on which we
can consider that the comparison of the variation of d along the two axes can be
compared. We computed
δρ(p, σ) =
∂ρ(p, σ)
∂ρ(p, σ)
δp −
δσ,
∂p
∂σ
(33)
where ρ was replaced with the considered metrics. In areas of the p, σ space where
δρ is positive, we may say that the metric is more sensitive to reliability, while in
areas where it is negative the metric is more sensitive to the precision of spikes.
The interpretation of results should take into consideration the caveat that changes
of the σm /p m ratio may change the sign of δρ.
As expected, the Victor & Purpura distance with q = 0.001 ms−1 does not depend at all on the precision, but just on reliability, since it basically counts the
difference in the number of spikes between the spike trains. Our modulus-metric
and max-metrics have a stronger dependence on reliability than on precision, on
most of the considered range, except for high reliability (very low p) where the dependence on precision is still dominant. For the van Rossum and Schreiber et al.
distances, the dependence on reliability and precision is somehow balanced. The
Kreuz et al. distances and the Victor & Purpura distance with q = 0.2 ms−1 have
a stronger dependence on precision than on reliability. For all distances, except
for the Victor & Purpura distance with q = 0.001 ms−1 , the sensitivity to reliability
increases with the unreliability (with the probability p of spikes not being fired).
These results may be different for a different choice of the σm /p m ratio, of the
time scales of the considered spike trains and of the parameters of the parameterdependent distances. Our results are similar to the ones of Dauwels et al. (2009)
for the Victor & Purpura distance q = 0.001 ms−1 , but different for the Schreiber et
al. distance, probably because Dauwels et al. (2009) used a different approach for
modeling unreliability of spikes.
4.4 Correlations
We explored the correlation between the newly introduced metrics and the classical Victor & Purpura and van Rossum distances. We generated a 500 ms Poisson
spike train with a firing rate of 20 Hz. From this spike train, we generated a new
one by adding a Gaussian jitter with zero mean and 20 ms variance. We considered only generated and jittered spike trains that contained 10 spikes. We then
measured the distance between the original and the jittered spike train using various metrics. We repeated this in 1,000 trials, where for each trial the original spike
train and the jitter were generated randomly. For each metric, distances were normalized to the mean value across samples. The results are displayed in Figure 11.
Table 1 shows the correlation coefficients between the max-metric, the modulusmetric, the convolution max-metric, the Schreiber et al. distance, the Kreuz et al.
distances and, respectively, the van Rossum and Victor & Purpura distances.
4.5 Computation speed
We computed the distances between pairs of randomly generated spike trains, the
spike trains within a pair having the same number of spikes n. We varied n from
16
Figure 7: Normalized average distances between a template spike train and another one generated from it, as a function of the probability p of removing spikes
from the template spike train and of jitter σ added to spikes. Distances are normalized to their maximum over the considered range. (A) Modulus-metric. (B) Maxmetric. (C) Victor & Purpura, q = 0.2 ms−1 . (D) Victor & Purpura, q = 0.001 ms−1 .
(E) Van Rossum. (F) Schreiber et al. (G) Kreuz et al. improved SPIKE-distance. (H)
Kreuz et al. ISI-distance.
17
do
1
Jitter 0 (ms)
Jitter 5 (ms)
Jitter 10 (ms)
Jitter 20 (ms)
B
1
dm
A
0.5
0
C
0
40
80
Spikes removed (%)
0
D
0
F
1
0.5
0.5
G
s
dR
0
40
80
Spikes removed (%)
0.5
1
0
ks
dV P
0.5
E
0
40
80
Spikes removed (%)
0
H
1
1
0.5
0.5
0
0
40
80
Spikes removed (%)
1
ki
dV P
1
0
0.5
0
0
40
80
Spikes removed (%)
0
40
80
Spikes removed (%)
0
40
80
Spikes removed (%)
0
40
80
Spikes removed (%)
Figure 8: Normalized average distances as a function of the probability of removing
spikes p, for several levels of jitter σ (sections trough the panels in Fig. 7). (A)
Modulus-metric. (B) Max-metric. (C) Victor & Purpura, q = 0.2 ms−1 . (D) Victor
& Purpura, q = 0.001 ms−1 . (E) Van Rossum. (F) Schreiber et al. (G) Kreuz et al.
improved SPIKE-distance. (H) Kreuz et al. ISI-distance.
18
A
B
p = 0%
p = 25%
p = 50%
1
dm
do
1
0.5
0
0
C
10
Jitter (ms)
0
20
dV P
dV P
0
20
0
10
Jitter (ms)
20
0
10
Jitter (ms)
20
0
10
Jitter (ms)
20
0.5
0
20
F
1
1
0.5
0.5
0
G
10
Jitter (ms)
0
20
H
1
1
0.5
0.5
ki
ks
10
Jitter (ms)
s
dR
E
0
10
Jitter (ms)
1
0.5
0
0
D
1
0
0.5
0
10
Jitter (ms)
0
20
Figure 9: Normalized average distances as a function of jitter σ, for several levels
of the probability of removing spikes p (sections trough the panels in Fig. 7). (A)
Modulus-metric. (B) Max-metric. (C) Victor & Purpura, q = 0.2 ms−1 . (D) Victor
& Purpura, q = 0.001 ms−1 . (E) Van Rossum. (F) Schreiber et al. (G) Kreuz et al.
improved SPIKE-distance. (H) Kreuz et al. ISI-distance.
19
Figure 10: Illustration of the dominance of reliability over precision for determining the variability of distances (see main text). In white areas, the distance varies
more with respect to reliability than with respect to precision. In dark areas, the
dominance is reversed. (A) Modulus-metric. (B) Max-metric. (C) Victor & Purpura,
q = 0.2 ms−1 . (D) Victor & Purpura, q = 0.001 ms−1 . (E) Van Rossum. (F) Schreiber
et al. (G) Kreuz et al. improved SPIKE-distance. (H) Kreuz et al. ISI-distance.
20
Figure 11: Correlations between the Victor & Purpura (left column) and the van
Rossum (right column) metrics and, respectively, the distances computed with our
newly introduced metrics, the Schreiber et al. and the Kreuz et al. metrics. We
computed distances between Poisson spike trains and a jittered version of them.
Figures represent the distribution of pairs of the normalized values of distances
measured with the different metrics. (A) Max-metric. (B) Convolution max-metric.
(C) Modulus-metric. (D) Schreiber et al. (E) Kreuz et al. improved SPIKE-distance.
(F) Kreuz et al. ISI-distance. (G) Victor & Purpura / van Rossum.
21
Table 1: Correlation coefficients between the van Rossum and Victor & Purpura
distances and, respectively, other distances, computed from data presented in Figure 11.
Distance d R correlation coefficient dV P correlation coefficient
dm
0.54
0.48
do
0.54
0.47
dc
0.84
0.81
s
0.87
0.88
ks
0.51
0.70
ki
0.43
0.44
dR
1.00
0.78
dV P
0.78
1.00
5 to 500 while keeping constant the firing rate of the spike trains. The trains were
generated by randomly choosing the n firing times from a uniform distribution
between 0 and n T , where T=35 ms. We measured the average time needed to
compute distances, as a function of the number of spikes. Performance was measured using C++ implementations of the metrics, running on a Intel Core 2 processor. The Victor & Purpura metric was computed using the algorithm in (Victor
and Purpura, 1996). The van Rossum metric was implemented using the exact,
optimized algorithm presented in (Houghton and Kreuz, 2012) (A1) and a discretetime integration with a timestep of 1 ms (A2), which turned out to be slightly faster
than the optimized one. The Schreiber et al. and the Kreuz et al. metrics have also
been computed using a 1 ms integration timestep. The modulus-metric was implemented using Algorithms 1 and 2. The code used for all metrics is freely available
at https://github.com/modulus-metric/. The results are presented in Figure 12. In
panel (A), the results were averaged over 1,000 trials, while in the other panels the
results were averaged over 10,000 trials.
The simulations showed that the max-metric and the Schreiber et al. metric
are relatively slow to compute. Those two metrics and the Victor & Purpura metric require a computation time that grows more than linearly with the number of
spikes. The other metrics have an approximately linear dependence on the number of spikes; we fitted them with a line and computed the proportionality coefficients in Table 2. The fastest distances or algorithms were, in order: modulusmetric (Algorithm 2); van Rossum (A2); Kreuz et al. ISI-distance; van Rossum (A1);
modulus-metric (Algorithm 1); Kreuz et al. improved SPIKE-distance. It should
be noted that, while the modulus-metric algorithms and the van Rossum A1 algorithm compute the distances exactly (within machine numerical precision), the
other linear-time algorithms compute numerical approximations of the distances
through discrete-time integration, with a precision that depends on the integration
timestep.
5 Discussion
The max-metric and the modulus-metric behave in a qualitatively different way
than the classical van Rossum and Victor & Purpura distances (Figures 4-6), but
similar to the Kreuz et al. distances. Within a set of spike trains that are consid-
22
A
B
2500
3
dV P
ks
Time (ms)
Time (ms)
dm
s
1250
0
0
250
Spike count
0
500
0
250
Spike count
500
D
C
0.4
0.4
dR A1
ki
dR A2
do A1
do A2
Time (ms)
Time (ms)
1.5
0.2
0
0
250
Spike count
0.2
0
500
0
250
Spike count
500
Figure 12: Time needed for computing various distances, as a function of the number of spikes in the spike trains. Note the different time scales. (A) d m : max-metric;
s: Schreiber et al. (B) dV P : Victor & Purpura; k s : Kreuz et al. improved SPIKEdistance. (C) d R : van Rossum; A1: algorithm by (Houghton and Kreuz, 2012); A2:
discrete-time integration; k i - Kreuz et al. ISI-distance. (D) d o : modulus metric;
A1: Algorithm 1; A2: Algorithm 2.
23
Table 2: Average computing time per spike in a spike train, for various distances.
Notation and data as in Figure 12.
Distance (and algorithm) Computing time per spike (ms)
d o A2
0.001570
d R A2
0.001831
ki
0.001965
d R A1
0.002437
d o A1
0.003323
ks
0.010910
ered by the van Rossum and Victor & Purpura metrics to be at equal distance from
a reference spike train, the max-metric, the modulus-metric and the Kreuz et al.
metrics can distinguish a range of distances that reflect similarities in the structure
of the compared spike trains (Figures 4, 5). When comparing spike trains that include bursts, the max-metric, the modulus-metric and the Kreuz et al. metrics ignore differences in the number and position of spikes inside the bursts, while this
kind of differences are considered by the van Rossum and Victor & Purpura metrics
as significant as in the case that differences exist outside the bursts (Figure 6). This
makes the max-metric, the modulus-metric and the Kreuz et al. metrics particularly suitable for measuring distances in spike trains where information is encoded
in the timing of bursts or solitary spikes, but not in the internal structure of bursts.
This is the case in some experiments, such as (Reinagel et al., 1998), where bursts
regarded as unitary events encoded more information per event than otherwise, or
such as (Kepecs and Lisman, 2003), where the timing of the first burst spike carried
70% of the information and the spike count only 22% of the information. However,
in other cases the internal structure of bursts does carry information (Krahe and
Gabbiani, 2004). The Schreiber et al. metric has a behaviour that is intermediate
between the one of the max-metric, the modulus-metric and the Kreuz et al. metrics, on one side, and the van Rossum and Victor & Purpura metrics, on another
side.
While the max-metric depends on a kernel H which can be particularized to
cause distinct behaviors, and the van Rossum, Victor & Purpura and Schreiber et
al. distances also depend on parameters that must be chosen by their users, the
modulus-metric does not depend on any parameters, similarly to the Kreuz et al.
(2009, 2011, 2013) metrics. The lack of parameters allows one to start analyzing
data immediately, without the need of preprocessing it in order to find the appropriate parameters. A parameter-free distance also gives a more objective measure,
that does not depend on any assumptions to be made by the experimenter. In
some cases, when the time scales vary during an experiment, no single time scale
may characterize the spike trains, and thus a time-scale-independent measure may
be preferable.
We have also shown that the modulus-metric can be computed faster than any
of the other considered metrics, through an algorithm that operates in a time that
depends linearly on the number of spikes in the considered spike trains. This
fast algorithm computes the distance exactly (within machine precision), not as a
numerically-approximated discrete-time integration, as some algorithms for other
metrics do.
The convolution-metric that we introduced, although analytically similar to the
24
max-metric, is qualitatively similar to the van Rossum distance. A qualitative difference between the convolution-metric and the van Rossum distance appears when
the differences between the spike trains are localized near the ends of the integration interval, and this is a simple consequence of the difference between the
bounded integration interval for the convolution-metric and the infinite integration interval for the van Rossum distance.
We have considered only spike trains having non-overlapping spikes. If we relax
this constraint, for our newly introduced distances, with the exception of the convolution max-metric, we get a zero distance between a spike train and a second
one generated from the first by adding an extra spike to the first, overlapping an
existing spike (Figure 4). This is due to the distance d between an arbitrary timing
and a spike train in Eq. 5 which does not distinguish between overlapping spikes
in a train. Thus, if we relax the constraint of not allowing non-overlapping spikes,
these distances become pseudo-metrics because there might be a zero distance
between two spike trains that differ through overlapping spikes. However, the case
of overlapping spikes is biologically implausible if we consider spike trains fired by
single neurons. If it is enough that the distances are pseudo-metrics, we may also
relax some of the conditions of the kernels, such as the requirement for L to be
strictly positive on (0, b − a] or the conditions on K .
For our metrics, when the integrating interval [a, b] extends beyond the interval covered by extremes of the spike trains, e.g. [min(t (1) , t̄ (1) ), max(t (n) , t̄ (n̄) )] for
a pair of spike trains, the result of the integration in the area not covered by extremes of the spike trains adds to the distance without contributing information
about the spike trains. Thus, the integrating interval should preferably be chosen
as the interval covered by the extremes of the considered set of spike trains. Alternatively, one may artificially add to all considered spike trains two extra spikes at
the extremities a and b of the integrating interval, a procedure that is also used by
the Kreuz et al. metrics (Kreuz et al., 2013).
We have introduced localized versions of our metrics, which, depending on the
localization kernel L , could have a biologically-relevant interpretation of measuring the differences between two spike trains as they are perceived at a particular
moment in time by a neuron receiving these spike trains.
6 Conclusion
We have introduced here a new class of spike train metrics, inspired by the PompeiuHausdorff distance. The max-metric and the modulus-metric behave in a qualitatively different way than classical metrics and are particularly suitable for measuring distances in spike trains where information is encoded in the identity of bursts,
as unitary events. The modulus-metric does not depend on any parameters and
can be computed faster than other distances, in a time that depends linearly on
the number of spikes in the compared spike trains.
Acknowledgements
We acknowledge the useful suggestions of Adriana Nicolae and Ovidiu Jurjuţ. This
work was funded by the Sectorial Operational Programme Human Resources Development (POSDRU, contract no. 6/1.5/S/3, "Doctoral studies: through science
25
towards society"), by a grant of the Romanian National Authority for Scientific Research (PNCDI II, Parteneriate, contract no. 11-039/2007) and by the Max Planck Coneural PartnerGroup.
Appendix
A Equivalent Hausdorff form
Proposition A.1. The Pompeiu-Hausdorff metric
¯
¯
h(T, T̄ ) = sup ¯d (x, T ) − d (x, T̄ )¯
(34)
x∈R
can be equivalently expressed as
¯
¯
h(T, T̄ ) = sup ¯d (x, T ) − d (x, T̄ )¯ .
(35)
x∈[a, b]
Proof. The result follows by showing that the global supremum is achieved on the
interval [a, b]. We need to show that restricting the interval in Eq. 34 is acceptable
since the supremum is achieved on the interval [a, b].
We first consider x ∈ (−∞, a]. Because ∀ t ∈ T , x ≤ a ≤ t (1) ≤ t , we have d (x, T ) =
(1)
t − x and d (a, T ) = t (1) − a. Because ∀ t̄ ∈ T̄ , x ≤ a ≤ t̄ (1) ≤ t̄ , we have d (x, T̄ ) =
t̄ (1) − x and d (a, T̄ ) = t̄ (1) − a. Thus
sup
¯
¯
¯d (x, T ) − d (x, T̄ )¯ =
x∈(−∞,a]
sup
¯ (1)
¯ ¯
¯
¯(t − x) − (t̄ (1) − x)¯ = ¯t (1) − t̄ (1) ¯
sup
(36)
x∈(−∞,a]
¯
¯ ¯
¯
¯d (x, T ) − d (x, T̄ )¯ = ¯d (a, T ) − d (a, T̄ )¯ .
(37)
x∈(−∞,a]
Analogously, for x ∈ [b, ∞) we have
¯
¯ ¯
¯
sup ¯d (x, T ) − d (x, T̄ )¯ = ¯d (b, T ) − d (b, T̄ )¯ .
(38)
x∈[b,∞)
From Eqs. 37 and 38,
¯
¯
¯
¯
sup ¯d (x, T ) − d (x, T̄ )¯ = sup ¯d (x, T ) − d (x, T̄ )¯ .
x∈R
(39)
x∈[a, b]
B Analysis of the max-metric
Proposition B.1. d m (T, T̄ ) < ∞.
Proof. From Eq. 8, for every x ∈ [a, b] we have
¯
¯
¯d (x, T ) − d (x, T̄ )¯ ≤ h(T, T̄ ).
(40)
Multiplying by H (·), which is positive, we obtain, ∀s ∈ [a, b],
¯
¯
¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) ≤ h(T, T̄ ) H (|s − x|).
26
(41)
By taking the supremum and integrating, the equation above becomes
Z
b
sup
a x∈[a,b]
¯
©¯
ª
¯d (t , T ) − d (t , T̄ )¯ H (|s − x|) ds ≤
Z
b
sup
a x∈[a,b]
©
ª
h(T, T̄ ) H (|s − x|) ds. (42)
The left side of the equation above is the max-metric d m (Eq. 15). Because h(T, T̄ )
is independent of s and x,
Z
d m (T, T̄ ) ≤ h(T, T̄ )
b
sup H (|s − x|)ds.
(43)
a x∈[a,b]
Since h(T, T̄ ) ≤ b − a and H (y) < m, ∀y ∈ [0, b − a] (Eq. 12), we obtain
d m (T, T̄ ) ≤ (b − a)2 m < ∞.
(44)
Proposition B.2. d m : S [a,b] × S [a,b] → R+ is a metric.
Proof. In order to show that d m (·, ·) is a metric we need to prove that it is nonnegative, that d m (T, T̄ ) = 0 ⇔ T = T̄ for any T, T̄ ∈ S [a,b] , that is symmetric, and
that it satisfies the triangle inequality.
It is trivial to show that d m (T, T̄ ) is non-negative, symmetric, and that if T =
T̄ ⇒ d m (T, T̄ ) = 0.
In order to prove that d m (T, T̄ ) = 0 ⇒ T = T̄ we use a reductio ad absurdum
argument. Assume that d m (T, T̄ ) = 0 with T 6= T̄ . Then, there must be at least one
spike in one of the two spike trains that is not in the other, since we do not allow
overlapping spikes within the spike trains. Consider that this spike belongs to T ;
in the case that it belongs to T̄ , the proof is analogous. Let u be the timing of this
spike, u ∈ T \ T̄ . Because u ∈ [a, b], we have, ∀s ∈ [a, b],
¯
¯
ª ¯
©¯
sup ¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) ≥ ¯d (u, T ) − d (u, T̄ )¯ H (|s − u|).
(45)
x∈[a,b]
Because u ∈ T , d (u, T ) = 0. Thus
sup
¯
©¯
ª
¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) ≥ d (u, T̄ ) H (|s − u|).
(46)
x∈[a,b]
By integrating the equation above, we obtain
Z
d m (T, T̄ ) ≥ d (u, T̄ )
b
a
H (|s − u|) ds.
(47)
Because u ∉ T̄ , d (u, T̄ ) > 0. Also considering Eq. 9, we have
Z
b
d (u, T̄ )
a
H (|s − u|) ds > 0.
(48)
Thus, from Eqs. 47 and 48, d m (T, T̄ ) > 0. Since we considered that d m (T, T̄ ) = 0, this
cannot be true. Hence, T ⊆ T̄ . Likewise, one can show that T̄ ⊆ T and so T = T̄ .
In order to prove the triangle inequality, consider T̂ ∈ S [a,b] . We have, ∀x ∈ R,
|d (x, T ) − d (x, T̂ ) + d (x, T̂ ) − d (x, T̄ )| ≤ |d (x, T ) − d (x, T̂ )| + |d (x, T̂ ) − d (x, T̄ )|. (49)
27
Because supx ( f (x) + g (x)) ≤ supx ( f (x)) + supx (g (x)) for any two functions f and g ,
it follows that
¯
¯
©¯
©¯
ª
ª
sup ¯d (x, T ) − d (x, T̄ )¯ H (|s − x|) ≤ sup ¯d (x, T ) − d (x, T̂ )¯ H (|s − x|) +
x∈[a,b]
x∈[a,b]
sup
¯
©¯
ª
¯d (x, T̂ ) − d (x, T̄ )¯ H (|s − x|) . (50)
x∈[a,b]
After integration, it results that
d m (T, T̄ ) ≤ d m (T, T̂ ) + d m (T̂ , T̄ ).
(51)
With this final equality we have shown that the distance is indeed a metric and
the proof ends.
Proposition B.3. The metric d m : S [a,b] × S [a,b] → R is topologically equivalent to
the Pompeiu-Hausdorff distance.
Proof. In order to show that the metrics d m and h are topologically equivalent
it is sufficient to prove that the identity function i S [a,b] : (S [a,b] , d m ) → (S [a,b] , h),
i S [a,b] (T ) = T and its inverse are both continuous (O’Searcoid, 2007, p. 229; Deza
and Deza, 2009, p. 12).
We first show that i S [a,b] is continuous, which is equivalent to: ∀ T ∈ S [a,b] , ∀ ² >
0, ∃ δ(²) > 0 such that ∀ T̄ ∈ S [a,b] with d m (T, T̄ ) ≤ δ we have h(T, T̄ ) ≤ ². We
choose δ(²) = ² A(T, T̄ ) with
Z b
H (|s − t |) ds.
(52)
A(T, T̄ ) = inf
t ∈T ∪T̄ a
From Eq. 9, we have that A(T, T̄ ) > 0. From Eq. 15, for all t ∈ T ,
b
Z
d m (T, T̄ ) ≥
=
a
Z b
a
|d (t , T ) − d (t , T̄ )| H (|s − t |)ds
(53)
d (t , T̄ ) H (|s − t |)ds
(54)
b
Z
≥ d (t , T̄ ) inf
t 0 ∈T a
H (|s − t 0 |) ds
(55)
and thus
Z
d m (T, T̄ ) ≥ sup d (t , T̄ ) inf
b
t 0 ∈T a
t ∈T
H (|s − t 0 |) ds
Z
d m (T, T̄ ) ≥ sup d (t , T̄ )
t ∈T
b
inf
t 0 ∈T ∪T̄ a
(56)
H (|s − t 0 |) ds.
(57)
H (|s − t 0 |) ds.
(58)
Analogously, for all t̄ ∈ T̄ ,
Z
d m (T, T̄ ) ≥ sup d (t̄ , T )
t̄ ∈T̄
b
inf
t 0 ∈T ∪T̄ a
Taking the max value in Eqs. 57 and 58 we obtain
(
)
d m (T, T̄ ) ≥ max sup d (t , T̄ ), sup d (t̄ , T )
t ∈T
t̄ ∈T̄
28
Z
inf
b
t 0 ∈T ∪T̄ a
H (|s − t 0 |) ds.
(59)
From Eq. 6 it follows that
Z
b
d m (T, T̄ ) ≥ h(T, T̄ ) inf
t ∈T ∪T̄ a
H (|s − t |) ds.
(60)
Because ² A(T, T̄ ) = δ(²) ≥ d m (T, T̄ ), from the last equation we have that ² A(T, T̄ ) ≥
h(T, T̄ ) A(T, T̄ ). Thus h(T, T̄ ) ≤ ² since A(T, T̄ ) > 0.
−1
It remains to show that i S
is continuous, which is equivalent to: ∀ T ∈ S [a,b] , ∀ ² >
[a,b]
0, ∃ δ(²) > 0 such that ∀ T̄ ∈ S [a,b] with h(T, T̄ ) ≤ δ we have d m (T, T̄ ) ≤ ². We
choose δ(²) = ²/B (a, b) with
Z b
B (a, b) =
sup H (|s − x|) ds.
(61)
a x∈[a,b]
We have
Z
b
B (a, b) ≥ sup
x∈[a,b] a
H (|s − x|) ds.
Considering Eq. 9, we have B (a, b) > 0. From Eq. 15 we have
Z b
¯
¯
d m (T, T̄ ) ≤
sup ¯d (x, T ) − d (x, T̄ )¯ sup H (|s − x|) ds
a x∈[a,b]
(62)
(63)
x∈[a,b]
¯
¯
d m (T, T̄ ) ≤ sup ¯d (x, T ) − d (x, T̄ )¯
x∈[a,b]
Z
b
sup H (|s − x|) ds.
(64)
a x∈[a,b]
From Eqs. 8 and 62 we have
d m (T, T̄ ) ≤ h(T, T̄ ) B (a, b).
(65)
Since h(T, T̄ ) ≤ δ(²) = ²/B (a, b) and B (a, b) > 0, we get h(T, T̄ ) B (a, b) ≤ ² and, finally, d m (T, T̄ ) ≤ ².
C Analysis of the convolution max-metric
Lemma C.1. Let g : [a, b] → R be a continuous function and h : [0, b − a] → R be
a continuous function which is derivable on (0, b − a) and has bounded derivative.
Then the function q : [a, b] → R,
£
¤
q(s) = sup g (x) h(|s − x|)
(66)
x∈[a,b]
is continuous on [a, b].
Proof. Consider s 0 ∈ [a, b]. We need to show that ∀ ² > 0, ∃ δ(²) > 0 such that ∀ s ∈
T
(s 0 − δ, s 0 + δ) [a, b], we have
¯
¯
¯
£
¤
£
¤¯¯
¯
(67)
¯ sup g (x) h(|s − x|) − sup g (x) h(|s 0 − x|) ¯ < ².
¯x∈[a,b]
¯
x∈[a,b]
We have
¯
¯
g (x) h(|s − x|) − g (x) h(|s 0 − x|) ≤ ¯g (x) h(|s − x|) − g (x) h(|s 0 − x|)¯
¯
¯
g (x) h(|s − x|) ≤ ¯g (x) [h(|s − x|) − h(|s 0 − x|)]¯ + g (x) h(|s 0 − x|)
(69)
g (x) h(|s − x|) ≤ |g (x)| |h(|s − x|) − h(|s 0 − x|)| + g (x) h(|s 0 − x|).
(70)
29
(68)
The function g is bounded since it is continuous on a compact interval (Protter,
1998, p. 56). We denote by M the bound of the absolute value of g , i.e. |g (x)| ≤ M ,
∀x ∈ [a, b]. We denote by L the bound of the absolute value of the derivative of h,
i.e. |h(x) − h(y)| ≤ L |x − y|, ∀ x, y ∈ [0, b − a]. Let ² > 0 and δ = ²/(M L). Then for all
T
s ∈ (s 0 − δ, s 0 + δ) [a, b],
|h(|s − x|) − h(|s 0 − x|)| ≤ L ||s − x| − |s 0 − x||
|g (x)| |h(|s − x|) − h(|s 0 − x|)| ≤ M L ||s − x| − |s 0 − x|| .
(71)
(72)
From Eqs. 70 and 72,
g (x) h(|s − x|) ≤ M L ||s − x| − |s 0 − x|| + g (x) h(|s 0 − x|).
(73)
Because ∀u, v ∈ R we have ||u| − |v|| ≤ |u − v|, it follows that
g (x) h(|s − x|) ≤ M L |s − s 0 | + g (x) h(|s 0 − x|).
(74)
Because |s − s 0 | < δ = ²/(M L),
²
+ g (x) h(|s 0 − x|)
ML
= ² + g (x) h(|s 0 − x|).
g (x) h(|s − x|) < M L
Applying supremum to the equation above we obtain
£
¤
£
¤
sup g (x) h(|s − x|) < ² + sup g (x) h(|s 0 − x|) .
x∈[a,b]
(75)
(76)
x∈[a,b]
It follows that
sup
£
¤
£
¤
g (x) h(|s − x|) − sup g (x) h(|s 0 − x|) < ².
x∈[a,b]
(77)
x∈[a,b]
Analogously, by switching s and s 0 in Eq. 68 and the ensuing equations, we get
£
¤
£
¤
sup g (x) h(|s 0 − x|) − sup g (x) h(|s − x|) < ².
(78)
x∈[a,b]
x∈[a,b]
Thus we have proved Eq. 67 and the proof ends.
Proposition C.1. d c (T, T̄ ) < ∞.
Proof. Because K is positive, we have f ≥ 0 and f¯ ≥ 0. Moreover, ∀x ∈ R,
¯
¯ ¯
¯ ¯
¯
¯ f (x) − f¯(x)¯ ≤ ¯ f (x)¯ + ¯ f¯(x)¯ = f (x) + f¯(x)
n
n̄
X
¯
¯ X
¯ f (x) − f¯(x)¯ ≤
K (x − t (i ) ) +
K (x − t̄ (i ) ).
i =1
Since K (x) ≤ p, ∀ x ∈ R,
(79)
i =1
¯
¯
¯ f (x) − f¯(x)¯ ≤ p (n + n̄).
(80)
From Eq. 21,
Z
d c (T, T̄ ) ≤
b
¯
¯
sup ¯ f (x) − f¯(x)¯ sup H (|s − x|) ds
a x∈[a,b]
Z
d c (T, T̄ ) ≤p (n + n̄)
(81)
x∈[a,b]
b
sup H (|s − x|) ds.
a x∈[a,b]
30
(82)
Because H (y) < m, ∀ y ∈ [0, b − a] (Eq. 12), it follows that
d c (T, T̄ ) ≤ p (n + n̄) (b − a) m < ∞.
(83)
Proposition C.2. d c : S [a,b] × S [a,b] → R+ is a metric.
Proof. In order to show that d c (·, ·) is a metric we need to prove that it is nonnegative, that d c (T, T̄ ) = 0 ⇔ T = T̄ for any T, T̄ ∈ S [a,b] , that it is symmetric, and
that it satisfies the triangle inequality.
It is trivial to show that d c (T, T̄ ) is non-negative, symmetric, and that if T =
T̄ ⇒ d c (T, T̄ ) = 0. In order to prove that d c (T, T̄ ) = 0 ⇒ T = T̄ we use a reductio ad
absurdum argument. Assume that d c (T, T̄ ) = 0 with T 6= T̄ . For s ∈ [a, b] let
q(s) = sup
¯
©¯
ª
¯ f (x) − f¯(x)¯ H (|s − x|) .
(84)
x∈[a,b]
Because | f (x) − f¯(x)| is continuous, from the properties of H and Lemma C.1 we
obtain that q is continuous. Because of the properties of K , T 6= T̄ ⇒ ∃ x ∈ [a, b]
such that f (x) 6= f¯(x); and because H is strictly positive, it follows that q is not
zero everywhere, ∃x ∈ [a, b] such that q(x) > 0. Because q is continuous, it follows
that
Z
b
d c (T, T̄ ) =
q(s) ds > 0
a
(85)
which contradicts the hypothesis that d c (T, T̄ ) = 0. Hence, T = T̄ .
In order to prove the triangle inequality, consider T̂ ∈ S [a,b] . We have, ∀x, s ∈
[a, b],
¯ ¯
¯¢
¯
¯
¡¯
¯ f (x) − fˆ(x) + fˆ(x) − f¯(x)¯ H (|s − x|) ≤ ¯ f (x) − fˆ(x)¯ + ¯ fˆ(x) − f¯(x)¯ H (|s − x|).
(86)
Because sups (g (s)+h(s)) ≤ sups (g (s))+sups (h(s)) for any two functions g and h, it
follows that, ∀s ∈ [a, b],
sup
¯
¯
©¯
ª
©¯
ª
¯ f (x) − f¯(x)¯ H (|s − x|) ≤ sup ¯ f (x) − fˆ(x)¯ H (|s − x|) +
x∈[a,b]
x∈[a,b]
sup
¯
©¯
ª
¯ fˆ(x) − f¯(x)¯ H (|s − x|) .
(87)
x∈[a,b]
After integration, it results that
d c (T, T̄ ) ≤ d c (T, T̂ ) + d c (T̂ , T̄ ).
(88)
With this final equality we have shown that the distance is indeed a metric and
the proof ends.
D Analysis of the localized max-metric
Proposition D.1. d l (T, T̄ ) < ∞.
31
Proof. For every s ∈ [a, b] we have
¯
¯
h(T, T̄ ) = sup ¯d (x, T ) − d (x, T̄ )¯
x∈[a,b]
¯
¯
≥ sup ¯d (x, T ) − d (x, T̄ )¯
(89)
x∈[s,b]
¯
¯
h(T, T̄ ) L (b − s) ≥ L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ .
(90)
x∈[s,b]
By integrating the last equation above, we obtain
b
Z
a
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ds ≤ h(T, T̄ )
Z
x∈[s,b]
b
a
L (b − s)ds.
(91)
Since h(T, T̄ ) ≤ b − a and L (y) < m, ∀y ∈ [0, b − a] (Eq. 22), it follows that
d l (T, T̄ ) ≤ m (b − a)2 < ∞.
(92)
Proposition D.2. d l : S [a,b] × S [a,b] → R+ is a metric.
Proof. In order to show that d l (·, ·) is a metric we need to prove that it is nonnegative, that d l (T, T̄ ) = 0 ⇔ T = T̄ for any T, T̄ ∈ S [a,b] , that it is symmetric, and
that it satisfies the triangle inequality.
It is trivial to show that d l (T, T̄ ) is non-negative, symmetric, and that if T = T̄ ⇒
d l (T, T̄ ) = 0.
In order to prove that d l (T, T̄ ) = 0 ⇒ T = T̄ we use a reductio ad absurdum
argument. Assume that d l (T, T̄ ) = 0 with T 6= T̄ . Then, there must be at least one
spike in one of the two spike trains that is not in the other. Consider that this spike
belongs to T ; in the case it belongs to T̄ , the proof is analogous. Let u be the timing
of this spike, u ∈ T \ T̄ .
First, we consider the case u > a. We have, ∀s ∈ [a, u],
¯
¯ ¯
¯
sup ¯d (x, T ) − d (x, T̄ )¯ ≥ ¯d (u, T ) − d (u, T̄ )¯ .
(93)
x∈[s,b]
Because u ∈ T \ T̄ , d (u, T ) = 0 and d (u, T̄ ) > 0. Thus,
¯
¯
sup ¯d (x, T ) − d (x, T̄ )¯ ≥ d (u, T̄ ).
(94)
x∈[s,b]
Multiplying by L (·) and integrating the above equation, we obtain
Z u
Z u
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ds ≥ d (u, T̄ )
L (b − s) ds.
a
We also have
Z
b
u
(95)
a
x∈[s,b]
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ds ≥ 0.
(96)
x∈[s,b]
By adding the last two inequations above, and considering Eq. 26, we have:
Z u
d l (T, T̄ ) ≥ d (u, T̄ )
L (b − s) ds.
a
32
(97)
Because L is strictly positive on (0, b −a] and continuous, we have
0. Because we have d (u, T̄ ) > 0, we get
Ru
a
L (b −s) ds >
d l (T, T̄ ) > 0.
(98)
Since we have considered d l (T, T̄ ) = 0, it follows that Eq. 98 cannot be true. Hence,
T ⊆ T̄ .
Second, we consider the case u = a. Let v be the timing of the first spike in
either T or T̄ , other than u. Since T 6= T̄ , v > u. Because
b
Z
(u+v)/2
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ds ≥ 0,
(99)
x∈[s,b]
from Eq. 26 we have
(u+v)/2
Z
d l (T, T̄ ) ≥
a
d l (T, T̄ ) ≥
a
(100)
x∈[s,b]
(u+v)/2
Z
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ds
L (b − s)
¯
¯
¯d (x, T ) − d (x, T̄ )¯ ds.
sup
(101)
x∈[s,(u+v)/2]
For all x ∈ [u, (u + v)/2) we have¯ d (x, T ) = x −u, ¯d (x, T̄ ) ≥ v − x, and because on this
interval v − x > x − u, we have ¯d (x, T ) − d (x, T̄ )¯ > 0. Because L is strictly positive
on (0, b − a], we get
Z
(u+v)/2
a
L (b − s)
¯
¯
¯d (x, T ) − d (x, T̄ )¯ ds > 0
sup
(102)
x∈[s,(u+v)/2]
and thus
d l (T, T̄ ) > 0.
(103)
Since we have considered d l (T, T̄ ) = 0, it follows that Eq. 103 cannot be true. Hence,
T ⊆ T̄ .
Thus, we have shown that in both the case u > a and the case u = a we have
T ⊆ T̄ . Likewise, one can show that T̄ ⊆ T and so T = T̄ if d l (T, T̄ ) = 0.
In order to prove the triangle inequality consider T̂ ∈ S [a,b] . We have, ∀x ∈
[a, b],
|d (x, T ) − d (x, T̂ ) + d (x, T̂ ) − d (x, T̄ )| ≤ |d (x, T ) − d (x, T̂ )| + |d (x, T̂ ) − d (x, T̄ )|. (104)
Because supx (g (x) + h(x)) ≤ supx (g (x)) + supx (h(x)) for any two functions g and h,
it follows that, ∀s ∈ [a, b],
¯
¯
¯
¯
L (b − s) sup ¯d (x, T ) − d (x, T̄ )¯ ≤ L (b − s) sup ¯d (x, T ) − d (x, T̂ )¯ +
x∈[s,b]
x∈[s,b]
¯
¯
L (b − s) sup ¯d (x, T̂ ) − d (x, T̄ )¯ .
(105)
x∈[s,b]
After integration, it results that d l (T, T̄ ) ≤ d l (T, T̂ ) + d l (T̂ , T̄ ).
With this final equality we have shown that the distance is indeed a metric and
the proof ends.
33
E Analysis of the localized modulus-metric
Proposition E.1. d n (T, T̄ ) < ∞.
Proof. From Eq. 8, we have, ∀s ∈ [a, b],
¯
¯
h(T, T̄ ) ≥ ¯d (s, T ) − d (s, T̄ )¯ .
(106)
Multiplying by L (·), which is positive, we obtain
¯
¯
h(T, T̄ ) L (b − s) ≥ L (b − s) ¯d (s, T ) − d (s, T̄ )¯ .
(107)
By integrating the above equation, we obtain
Z b
Z b
¯
¯
h(T, T̄ )
L (b − s) ds ≥
L (b − s) ¯d (s, T ) − d (s, T̄ )¯ ds.
(108)
a
a
Since h(T, T̄ ) ≤ b − a and L (x) < m, ∀x ∈ [0, b − a] it follows that
d n (T, T̄ ) < m (b − a)2 < ∞.
(109)
Proposition E.2. d n : S [a,b] × S [a,b] → R+ is a metric.
Proof. In order to show that d n (·, ·) is a metric we need to prove that it is nonnegative, that d n (T, T̄ ) = 0 ⇔ T = T̄ for any T, T̄ , that it is symmetric, and that it
satisfies the triangle inequality.
Let T, T̄ ∈ S [a,b] . It is trivial to show that d n (T, T̄ ) is non-negative and symmetric, and that T = T̄ ⇒ d l (T, T̄ ) = 0. In order to prove the converse we use a reductio
ad absurdum argument. Assume T 6= T̄ with d n (T, T̄ ) = 0. For s ∈ [a, b] let
¯
¯
q(s) = L (b − s) ¯d (s, T ) − d (s, T̄ )¯ .
(110)
¯
¯
Because ¯d (s, T ) − d (s, T̄ )¯ is continuous and L is continuous it results that q is
continuous. Because T 6= T̄ ⇒ ∃ s ∈ [a, b] such that d (s, T ) 6= d (s, T̄ ); because L is
strictly positive on (0, b −a], it follows that q is not zero everywhere, ∃s ∈ [a, b) such
that q(s) > 0. Because q is continuous, it follows that
Z b
d n (T, T̄ ) =
q(s) ds > 0
(111)
a
which contradicts the hypothesis that d n (T, T̄ ) = 0. Hence, T = T̄ .
In order to prove the triangle inequality consider T̂ ∈ S [a,b] . We have, ∀s ∈ R,
|d (s, T ) − d (s, T̂ ) + d (s, T̂ ) − d (s, T̄ )| ≤ |d (s, T ) − d (s, T̂ )| + |d (s, T̂ ) − d (s, T̄ )|.
(112)
Multiplying by L (·) and integrating the above equation, we obtain
Z b
¯
¯
¯d (s, T ) − d (s, T̄ )¯ L (b − s) ds ≤
Z
a
b¯
a
¯
¯d (s, T ) − d (s, T̂ )¯ L (b − s) ds +
Z
b¯
a
¯
¯d (s, T̂ ) − d (s, T̄ )¯ L (b − s) ds
(113)
It results that
d n (T, T̄ ) ≤ d n (T, T̂ ) + d n (T̂ , T̄ ).
(114)
With this final inequality we have shown that the distance is indeed a metric and
the proof ends.
34
F Algorithms
35
Input: The pair of nonempty spike trains T1 , T2 and the bounds a and b.
Output: The distance d o (T1 , T2 ) between the spike trains.
〈T1 , T2 and P are ordered sets of real numbers, indexed starting from 0.〉
n 1 := length(T1 ); n 2 := length(T2 );
S
P := sort(T1 T2 );
M := ;;
for i := 1 . . . n 1 − 1 do
S
M := M {(T1 [i ] + T1 [i − 1])/2};
for i := 1 . . . n 2 − 1 do
S
M := M {(T2 [i ] + T2 [i − 1])/2};
for i := 1 . . . length(P ) − 1 do
S
M := M {(P [i ] + P [i − 1])/2};
S
S
P := sort(P M {a, b});
d o := 0; i 1 := 0; i 2 := 0;
〈s is the currently considered point from P . φ is the value at s of the integrated
function |d (s, T1 ) − d (s, T2 )|. s p is the previously considered point from P , and φp is
the value at s p of the integrated function. i 1 is the index of the first spike in T1
having a timing that is greater than s, if there is such a spike, or the index of the last
spike of T1 , otherwise. i 2 is computed analogously for T2 .〉
s p := a; φp := |T1 [0] − T2 [0]|;
for i := 1 . . . length(P ) − 1 do
s := P [i ];
while s ≥ T1 [i 1 ] and i 1 < n 1 − 1 do
i 1 := i 1 + 1;
while s ≥ T2 [i 2 ] and i 2 < n 2 − 1 do
i 2 := i 2 + 1;
d 1 := b − a; d 2 := b − a;
if i 1 > 0 then
d 1 := s − T1 [i 1 − 1];
d 10 = |T1 [i 1 ] − s|;
if d 10 < d 1 then
d 1 := d 10 ;
if i 2 > 0 then
d 2 := s − T2 [i 2 − 1];
d 20 = |T2 [i 2 ] − s|;
if d 20 < d 2 then
d 2 := d 20 ;
〈We now have d 1 = d (s, T1 ) and d 2 = d (s, T2 ) and we can compute the value of φ
at s:〉
φ := |d 1 − d 2 |;
〈The integration is performed here:〉
d o := d o + (s − s p )(φ + φp )/2;
s p := s; φp := φ;
return d o ;
Algorithm 1: An algorithm for computing the modulus-metric distance d o between two spike trains T1 and T2 . The text surrounded by 〈. . .〉 represents comments.
36
Input: The pair of non-empty spike trains T1 , T2 and the bounds a and b.
Output: The distance d o (T1 , T2 ) between the spike trains.
〈T1 and T2 are ordered sets of real numbers, indexed starting from 0. i 1 and i 2 are
the indices of the currently processed spikes in the two spike trains. p 1 and p 2 are
the indices of the previously processed spikes in the two spike trains. p is the index
of the spike train to which the previously processed spike belonged (1 or 2), after at
least one spike has been processed, or 0 otherwise.〉
i 1 := 0; i 2 := 0; p 1 := 0; p 2 := 0; p := 0;
n 1 := length(T1 ); n 2 := length(T2 );
〈P is©¡
an array of structures (s, φ)¢ªconsisting of an ordered pair of numbers.〉
P := s 7→ a, φ 7→ |T1 [0] − T2 [0]| ;
〈Process the spikes until the end of one of the spike trains is reached.〉
while i 1 < n 1 and i 2 < n 2 do
if T1 [i 1 ] ≤ T2 [i 2 ] then
proc1(1, 2);
else
proc1(2, 1);
〈Process the rest of the spikes in the spike train that has not been fully processed.〉
while i 1 < n 1 do
proc2(1, 2);
while i 2 < n 2 do
proc2(2, 1);
¢ª
S ©¡
P := P
s 7→ b, φ 7→ |T1 [n 1 − 1] − T2 [n 2 − 1]| ;
〈Sort P . Elements of P are sorted according to their value of s.〉
P := sort(P );
〈Perform the integration.〉
d o := 0;
for i := 1 . . . length(P ) − 1 do
d o := d o + (P [i ].s − P [i − 1].s)(P [i ].φ + P [i − 1].φ)/2;
return d o ;
Algorithm 2: Another algorithm for computing the modulus-metric distance d o
between two spike trains T1 and T2 . The text surrounded by 〈. . .〉 represents comments.
37
Input: The indices j and k of the two sorted spike trains T j , Tk ; j , k ∈ {1, 2}, j 6= k.
Data: Uses as global variables: the indices i j , i k of the current spikes in the two spike
trains; the indices p j , p k of the previously processed spikes in the two spike
trains; the index p of the spike train to which the previously processed spike
belonged (1 or 2; if no spike has been previously processed, p = 0); the data
structure P . We should have here T j [i j ] ≤ Tk [i k ]. If p 6= 0, we should have
T p [i p ] ≤ T j [i j ].
Result: Performs part of the processing needed for creating P . The procedure is used
when processing has not reached the end of one of the spike trains.
if i j > 0 then
〈Adds to P the timing situated at the middle of the interval between the currently
processed spike and the previous spike in the same spike train.〉
t := (T j [i j ] + T j [i j − 1])/2;
〈We have d (t , T j ) = T j [i j ] − t = t − T j [i j − 1] = (T j [i j ] − T j [i j − 1])/2.〉
¯
¯´o
S n³
¯
¯
P := P
s 7→ t , φ 7→ ¯(T j [i j ] − T j [i j − 1])/2) − d (t , k, i k )¯ ;
if p = k then
〈If the previously processed spike was one from the other spike train than the
spike currently processed, adds to P the timing situated at the middle of the
interval between the currently processed spike and the previously processed
spike.〉
t := (T j [i j ] + Tk [p k ])/2;
〈Since t is at equal distance to the closest spikes in the two spike trains, T j [i j ]
and Tk [p k ], we have d (t , T j ) = d (t , Tk ) and φ(t ) = 0.〉
¢ª
S ©¡
P := P
s 7→ t , φ 7→ 0 ;
〈Adds to P the currently processed spike.〉
t := T j [i j ];
〈We have d (t , T j ) = 0. If at least one spike from Tk has been processed, we have
Tk [p k ] ≤ t ≤ Tk [i k ], with i k = p k + 1, and thus d (t , Tk ) = min(|t − Tk [p k ]|, Tk [i k ] − t ).
If no spike from Tk has been processed, we have p k = i k = 0, and the previous
formula for
d (t , Tk ) still holds.〉
¢ª
S ©¡
P := P
s 7→ t , φ 7→ min(|t − Tk [p k ]|, Tk [i k ] − t ) ;
p j := i j ;
i j := i j + 1;
p := j ;
Procedure proc1( j , k).
38
Input: The indices j and k of the two sorted spike trains T j , Tk ; j , k ∈ {1, 2}, j 6= k.
Data: Uses as global variables: the index i j of the current spike in T j ; the index p k of
the previously processed spike in Tk ; the index p of the spike train to which
the previously processed spike belonged (1 or 2); the data structure P . Here, p k
should be the index of the last spike in spike train Tk . We should have
Tk [p k ] ≤ T j [i j ].
Result: Performs part of the processing needed for creating P . The procedure is used
when processing has reached the end of spike train Tk .
if i j > 0 then
〈Adds to P the timing situated at the middle of the interval between the currently
processed spike and the previous spike in the same spike train.〉
t := (T j [i j ] + T j [i j − 1])/2;
〈We have d (t , T j ) = T j [i j ] − t = t − T j [i j − 1] = (T j [i j ] − T j [i j − 1])/2.〉
¯
¯´o
S n³
¯
¯
P := P
s 7→ t , φ 7→ ¯(T j [i j ] − T j [i j − 1])/2) − d (t , k, p k )¯ ;
if p = k then
〈If the previously processed spike was one from the other spike train than the
spike currently processed (i.e., the last spike in the spike train that has been fully
processed), adds to P the timing situated at the middle of the interval between
the currently processed spike and the previously processed spike.〉
t := (T j [i j ] + Tk [p k ])/2;
〈Since t is at equal distance to the closest spikes in the two spike trains, T j [i j ]
and Tk [p k ], we have d (t , T j ) = d (t , Tk ) and φ(t ) = 0.〉
¢ª
S ©¡
P := P
s 7→ t , φ 7→ 0 ;
〈 Adds to P the currently processed spike. 〉
t := T j [i j ];
〈 We have d (t , T j ) = 0. We have Tk [p k ] ≤ t and the spike at p k is the last one in Tk ,
and thus ©¡
d (t , Tk ) = t − Tk [p k ].〉 ¢ª
S
P := P
s 7→ t , φ 7→ t − Tk [p k ] ;
i j := i j + 1;
p := j ;
Procedure proc2( j , k).
39
Input: A timing t , the index k ∈ {1, 2} of a sorted spike train Tk , and the index i of a
spike in Tk , such that either t ≤ Tk [i ] or i is the index of the last spike of Tk .
Output: The distance d (t , Tk ) between the timing t and the spike train Tk .
d := |Tk [i ] − t |;
j := i − 1;
while j ≥ 0 and |Tk [ j ] − t | ≤ d do
d := |Tk [ j ] − t |;
j := j − 1;
return d ;
Function d(t , k, i ).
References
Adrian, E. D. (1926), ‘The impulses produced by sensory nerve endings’, Journal of
Physiology 61, 49–72.
Bohte, S. M. (2004), ‘The evidence for neural information processing with precise
spike-times: A survey’, Natural Computing 3(2), 195–206.
Dauwels, J., Vialatte, F., Weber, T. and Cichocki, A. (2009), On similarity measures for spike trains, in ‘Advances in Neuro-Information Processing. 15th International Conference, ICONIP 2008, Auckland, New Zealand, November 25-28,
2008, Revised Selected Papers, Part I’, Vol. 5506 of Lecture Notes in Computer Science, Springer, pp. 177–185.
Deza, M. M. and Deza, E. (2009), Encyclopedia of Distances, Springer.
Fries, P., Nikolić, D. and Singer, W. (2007), ‘The gamma cycle’, Trends in Neurosciences 30(7), 309–316.
Geisler, W. S., Albrecht, D. G., Salvi, R. J. and Saunders, S. S. (1991), ‘Discrimination performance of single neurons: Rate and temporal information’, Journal of
Neurophysiology 66, 334–362.
Hausdorff, F. (1914), Grundzüge der Mengenlehre, Verlag von Veit & Comp, Leipzig.
Houghton, C. and Kreuz, T. (2012), ‘On the efficient calculation of van Rossum distances’, Network: Computation in Neural Systems 23, 48–58.
Jacobs, J., Kahana, M. J., Ekstrom, A. D. and Fried, I. (2007), ‘Brain oscillations
control timing of single-neuron activity in humans’, Journal of Neuroscience
27(14), 3839–3844.
Ji, D. and Wilson, M. A. (2007), ‘Coordinated memory replay in the visual cortex
and hippocampus during sleep’, Nature Neuroscience 10(1), 100–107.
Kepecs, A. and Lisman, J. (2003), ‘Information encoding and computation with spikes and bursts’, Network: Computation in Neural Systems 14(1), 103–118.
Krahe, R. and Gabbiani, F. (2004), ‘Burst firing in sensory systems’, Nature Reviews
Neuroscience 5, 13–24.
40
Kreuz, T., Chicharro, D., Andrzejak, R., Haas, J. and Abarbanel, H. (2009), ‘Measuring multiple spike train synchrony’, Journal of Neuroscience Methods 183(2), 287–
299.
Kreuz, T., Chicharro, D., Greschner, M. and Andrzejak, R. (2011), ‘Time-resolved
and time-scale adaptive measures of spike train synchrony’, Journal of Neuroscience Methods 195(1), 92–106.
Kreuz, T., Chicharro, D., Houghton, C., Andrzejak, R. G. and Mormann, F. (2013),
‘Monitoring spike train synchrony’, Journal of Neurophysiology 109(5), 1457–
1472.
Kreuz, T., Haas, J. S., Morelli, A., Abarbanel, H. D. and Politi, A. (2007), ‘Measuring
spike train synchrony’, Journal of Neuroscience Methods 165(1), 151–161.
Lee, H., Simpson, G., Logothetis, N. and Rainer, G. (2005), ‘Phase locking of single
neuron activity to theta oscillations during working memory in monkey extrastriate visual cortex’, Neuron 45(1), 147–156.
Montemurro, M. A., Rasch, M. J., Murayama, Y., Logothetis, N. K. and Panzeri, S.
(2008), ‘Phase-of-firing coding of natural visual stimuli in primary visual cortex’,
Current Biology 18(5), 375–380.
Nádasdy, Z., Hirase, H., Czurkó, A., Csicsvari, J. and Buzsáki, G. (1999), ‘Replay and
time compression of recurring spike sequences in the hippocampus’, Journal of
Neuroscience 19(21), 9497–9507.
Naud, R., Gerhard, F., Mensi, S. and Gerstner, W. (2011), ‘Improved similarity measures for small sets of spike trains’, Neural Computation 23, 3016–3069.
O’Searcoid, M. (2007), Metric Spaces, Springer Undergraduate Mathematics Series,
Springer.
Papadopoulos, A. (2005), Metric spaces, convexity and nonpositive curvature, IRMA
lectures in mathematics and theoretical physics, European Mathematical Society.
Pompeiu, D. (1905), ‘Sur la continuité des fonctions de variables complexes’, Annales de la Faculté des Sciences de Toulouse 2/7(3), 265–315.
Protter, M. H. (1998), Basic elements of real analysis, Springer-Verlag, New York,
USA.
Reinagel, P., Godwin, D., Sherman, S. M. and Koch, C. (1998), ‘Encoding of visual
information by LGN bursts’, Journal of Neurophysiology 81(5), 2558–2569.
Rockafellar, R. T. and Wets, R. J.-B. (2009), Variational Analysis, Springer.
Rusu, C. V. and Florian, R. V. (2010), ‘A new spike train metric’, BMC Neuroscience
11(Suppl. 1), P169. Nineteenth Annual Computational Neuroscience Meeting:
CNS*2010.
Schrauwen, B., and Van Campenhout, J. (2007), ‘Linking non-binned spike train
kernels to several existing spike train metrics’, Neurocomputing 70(7–9), 1247–
1253.
41
Schreiber, S., Fellous, J., Whitmer, D., Tiesinga, P. and Sejnowski, T. (2003), ‘A
new correlation-based measure of spike timing reliability’, Neurocomputing 52–
54, 925–931.
Siegel, M., Warden, M. and Miller, E. (2009), ‘Phase-dependent neuronal coding of
objects in short-term memory’, Proceedings of the National Academy of Sciences
of the USA 106(50), 21341–21346.
Tiesinga, P., Fellous, J. and Sejnowski, T. (2008), ‘Regulation of spike timing in visual
cortical circuits’, Nature Reviews Neuroscience 9(2), 97–107.
Toups, J. V., Fellous, J.-M., Thomas, P. J., Sejnowski, T. J. and Tiesinga, P. H. (2012),
‘Multiple spike time patterns occur at bifurcation points of membrane potential
dynamics’, PLoS Computational Biology 8(10), e1002615.
van Rossum, M. C. W. (2001), ‘A novel spike distance’, Neural Computation 13, 751–
763.
VanRullen, R., Guyonneau, R. and Thorpe, S. J. (2005), ‘Spike times make sense’,
Trends in Neurosciences 28(1), 1–4.
Victor, J. D. (2005), ‘Spike train metrics’, Current Opinion in Neurobiology 15, 585–
592.
Victor, J. D. and Purpura, K. P. (1996), ‘Nature and precision of temporal coding
in visual cortex: A metric-space analysis’, Journal of Neurophysiology 76, 1310–
1326.
Victor, J. D. and Purpura, K. P. (1997), ‘Metric-space analysis of spike trains: Theory,
algorithms and application’, Network: Computation in Neural Systems 8, 127–
164.
42
| 9cs.NE
|
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
arXiv:1705.10718v2 [math.AC] 31 Jan 2018
STEVEN V SAM AND ANDREW SNOWDEN
Abstract. Suppose that for each n ≥ 0 we have a representation Mn of the symmetric
group Sn . Such sequences arise in a wide variety of contexts, and often exhibit uniformity
in some way. We prove a number of general results along these lines in this paper: our
prototypical theorem states that if {Mn } can be given a suitable module structure over a
twisted commutative algebra then the sequence {Mn } follows a predictable pattern. We
phrase these results precisely in the language of Hilbert series (or Poincaré series, or formal
characters) of modules over tca’s.
Contents
1. Introduction
2. Preliminaries
3. Formal characters
4. The main theorems on Hilbert and Poincaré series
5. A categorification of rationality
6. D-finiteness of Hilbert series of bounded modules
7. Examples and applications
References
1
3
5
10
15
18
20
27
1. Introduction
Suppose that for each n ≥ 0 we have a complex representation Mn of the symmetric group
Sn . Such sequences of symmetric group representations arise in a variety of contexts, and
naturally occurring examples tend to exhibit some kind of uniformity. For example:
(a) Fix a manifold X of dimension ≥ 2 and a non-negative integer i. Let Mn be the
ith cohomology group of the configuration space of n labeled points on X. This
example was studied in [CEF]. One of the theorems in loc. cit. states that, under
mild assumptions on X, there is a single character polynomial that gives the character
of Mn for all sufficiently large n.
(b) Fix p and r. Let Mn be the space of p-syzygies of the Segre embedding (Pr )n →
Prn+n−1 . This case was studied in [Sno]. One of the theorems in loc. cit. states that
the generating function of the sequence {dim(Mn )} is rational.
(c) Suppose that N is a functor associating to each finite dimensional vector space V a
module N(V ) over the ring Sym(Cd ⊗ V ), for some fixed d (and that satisfies some
Date: November 1, 2017.
2010 Mathematics Subject Classification. 05E05, 13A50.
SS was supported by a Miller research fellowship and NSF grant DMS-1500069.
AS was supported by NSF grants DMS-1303082 and DMS-1453893 and a Sloan Fellowship.
1
2
STEVEN V SAM AND ANDREW SNOWDEN
technical conditions). For example, one could take N(V ) to be the coordinate ring
of the rth determinantal variety for some 0 ≤ r ≤ d. Now let Mn be the Schur–
Weyl dual of the degree n piece of N. This example is studied in [Sno] and, in much
greater detail, [SS5]. Again, it is known that the generating function of the dimension
sequence is rational.
(d) Let E be a representation of a reductive group G, and take Mn = (E ⊗n )G . The
generating function of {dim(Mn )} is D-finite (and typically not rational, or even
algebraic). This is presumably well-known, though we do not know a reference. See
§7.4 for a proof.
In this paper, we study the uniformity properties of sequences {Mn } with the aim of strengthening and generalizing results like those mentioned above. We prove a number of results of
the form “if {Mn } admits a suitable module structure over a twisted commutative algebra
then the sequence {Mn } follows a predictable pattern.” The specific results give various
precise meanings to “suitable” and “predictable pattern.”
1.1. Statement of results. We now state some of our results in a little more detail. Let k
be a field of characteristic 0. AL
twisted commutative algebra (tca) over k is an associative
unital graded k-algebra A = n≥0 An equipped with an action of the symmetric group Sn
on An , such that the multiplication is commutative up L
to a “twist” by the symmetric group.
A module over a tca A is a graded vector space M = n≥0 Mn equipped with an action of
Sn on Mn and a multiplication An × Mm → Mn+m satisfying suitable axioms. Note that a
module M gives rise to a sequence of representations {Mn } as discussed above. TCA’s and
their modules have been a central object of study in the developing field of representation
stability. For example, k[t] can be regarded as a tca (with t of degree 1, and all Sn actions
trivial), and modules over it are equivalent to the FI-modules of Church–Ellenberg–Farb
[CEF].
Suppose that M is a module over a tca. We define its Hilbert series by
X
tn
HM (t) =
dim(Mn ) .
n!
n≥0
The following theorem was proved in [Sno], and later reproved in [SS4]:
Theorem 1.1. Suppose M is a finitely generated module over a tca finitely generated in
degree 1. Then HM (t) is a polynomial in t and et .
This theorem implies that the generating function for {dim Mn } is rational, but is even
stronger than this. The results of this paper generalize and strengthen this theorem in
various ways.
We now informally describe some of the main results of this paper. In what follows, A
denotes a tca generated in degree 1 and M denotes a finitely generated A-module.
(a) We introduce the formal character ΘM of M. This records the character of each
representation Mn , and thus contains much more information than the Hilbert series.
Using the structure theory of A-modules developed in [SS5], we give a very precise
description of ΘM . Our result shows that there is a finite expression for ΘM , and
moreover that the pieces in this expression reflect the structure of M. As a corollary,
we see that if M and N are finitely generated A-modules such that Mn and Nn
have the same character for all n, then M and N represent the same class in the
Grothendieck group of A-modules.
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
3
(b) Using the method of [Sno], we give a different proof of a rationality result for ΘM .
Actually, we work with the enhanced Hilbert series introduced in [SS1], which is
equivalent to the formal character. This is less precise than the proof described in
(a), but does not rely on the theory from [SS5].
(c) We show how the Fourier transform from [SS5] affects (enhanced) Hilbert series.
(d) We explain how all of the results on (enhanced) Hilbert series carry over to the more
subtle Poincaré series.
(e) Theorem 1.1 can be stated equivalently as: there exists a polynomial p(T ) with nonnegative integer roots such that p( dtd )HM (t) = 0. We prove a categorification of this
result, in which dtd is replaced with the Schur derivative.
Additionally, we prove a result for tca’s not necessarily generated in degree 1:
(f) If A is a finitely generated tca with A0 = C and M is a finitely generated and bounded
A-module, then HM (t) is a D-finite power series.
1.2. Open problems. We now list some questions and open problems related to the work
in this paper.
• To what extent do the results here carry over to positive characteristic? Theorem 1.1
is known to hold in positive characteristic by [SS4, Corollary 7.1.7], but the other
results of this paper are not known.
• Result (a) above gives a finite expression for the formal character ΘM in the case
where M is a finitely generated module over a tca A finitely generated in degree 1.
Is there a more general result if one assumes that A is bounded instead of generated
in degree 1?
• Result (e) categorifies Theorem 1.1. Can the rationality theorem for the formal character be similarly categorified? Presumably, such a categorification should make use
of the more general Schur derivatives Dλ .
• The discussion in §5.4 raises a number of questions about the categorified rationality.
For example: in the tensor category of differential operators, what can one say about
the ideal annihilating a given module? Result (e) ensures that it is non-zero.
• What can be said about the Hilbert series of a finitely generated module over an
arbitrary finitely generated tca? For example, is it D-finite?
• There are interesting sequences {Mn } of symmetric group representations that do
not come from tca’s, but related structures, such as FSop -modules (see [SS4, §8]). To
what extent do the results here extend to those sequences?
1.3. Outline. In §2 we review some background material that we will require. In §3 we
introduce the formal character and prove our “rationality” theorem for it. In §4.1 we translate
our results about formal characters to Hilbert series (and Poincaré series), and also give
an elementary proof of the rationality theorem in this setting. In §5 we prove a result
categorifying the rationality theorem. In §6 we prove D-finiteness for Hilbert series over
bounded tca’s not necessarily generated in degree 1. Finally, in §7 we give some examples
and applications of our results.
2. Preliminaries
This paper is a continuation of [SS5] and hence will use the same notation.
4
STEVEN V SAM AND ANDREW SNOWDEN
Throughout, X denotes a separated, noetherian C-scheme of finite Krull dimension, and
E is a vector bundle over X. We will primarily be interested in the case when X is a point,
but in order to simplify some of the arguments, it will be necessary to allow X to be a
Grassmannian. Any extra generality beyond that will not be used, but does not present any
additional difficulties. The reader unfamiliar with the language of sheaf theory can assume
X is a point for most of the paper, in which case quasi-coherent sheaves specialize to complex
vector spaces, and coherent sheaves specialize to finite-dimensional vector spaces.
Throughout, we make use of the category VX , which has several equivalent models. The
two that we use here are the category of sequences of symmetric group representations on
quasi-coherent OX -modules, and the category of polynomial functors from the category of
vector spaces to quasi-coherent OX -modules. See [SS2, §5] for more detailsLin the case
X = Spec(C). In the first model, every object VX admits a decomposition
λ Mλ ⊗ Fλ
where Mλ is the usual Specht module over C and Fλ is a quasi-coherent sheaf on X; in the
second model, we get a similar decomposition, but with Mλ replaced by the Schur functor
Sλ . Given F ∈ VX in the second model, we use ℓ(F ) to denote the maximum number of
parts of any λ such that Sλ has a nonzero multiplicity space. We say that F is bounded
if ℓ(F ) < ∞. We let Vdfg
X denote the subcategory of VX where the multiplicity spaces
Fλ are coherent. (The “dfg” superscript means “degreewise finitely generated.”) We let
Vfg
of Vdfg
X be the subcategory
X where all but finitely many of the Fλ vanish. We also let
S
∞
n
V = C = n≥1 C be the standard representation of GL∞ ; evaluating a Schur functor
on V gives an equivalence between the category of polynomial functors and the category of
polynomial representations of GL∞ .
Since we are working over a field of characteristic 0, we can use an alternative description
of tca’s: they are polynomial functors from the category of vector spaces to the category of
(sheaves of) commutative algebras. We let A(E) denote the tca which sends a vector space
V to the commutative algebra Sym(E ⊗ V ).
2.1. K-theory of relative Grassmannians. For a non-negative integer r, Grr (E) denotes
the relative Grassmannian of rank r quotients of E. Let πr : Grr (E) → X be the structure
map, and let Q = Qr be the tautological quotient bundle on Grr (E) and let R = Rr denote
the tautological subbundle, so that we have a short exact sequence
0 → R → π∗E → Q → 0
on Grr (E). Define
ir : K(Grr (E)) → K(A(E)),
ir ([V ]) = [Rπr∗ (V ⊗ A(Qr ))].
Let Λ denote the ring of symmetric functions, and let Λd be the space of symmetric functions
which are homogeneous of degree d. The group K(A(E)) is naturally a Λ-module via sλ ·[M] =
[Sλ (V) ⊗ M]. The following can be found in [SS5, Theorem 6.19]:
Theorem 2.1. The maps ir induce an isomorphism of Λ-modules
rank
ME
r=0
Λ ⊗ K(Grr (E)) → K(A(E)).
2.2. Integration on the torus. Let T be the diagonal torus in GL(d). Denote by α1 , . . . , αd
the standard characters T → Gm , and identify the representation
ring K(T ) with the ring
R
of Laurent polynomials Q[α1 , α1−1, . . . , αd , αd−1 ]. We let f 7→ T f dα be the projection map
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
5
K(T ) → Q which takes the constant term of f . We let f 7→ f be the ring homomorphism
K(T ) → K(T ) defined by sending αi to αi−1 . Define
Y
(αi − αj ),
|∆|2 = ∆∆.
∆(α) =
1≤i<j≤d
Weyl’s integration formula (see [FH, §26.2]) can then be stated as follows: if f, g ∈ K(T ) are
the characters of irreducible representations of GL(d), then
(
Z
1 if f = g
1
.
(2.2)
f (α)g(α)|∆(α)|2 dα =
d! T
0 if f 6= g
2.3. Symmetric groups and symmetric functions. Fix a nonnegative integer n. Let
λ be an integer partition of n and let cλ be the conjugacy class of permutations with cycle
type λ. Also, partitions parametrize the irreducible complex representations Mλ . We let
tr(cµ |Mλ ) denote the trace of any element of cµ acting on Mλ . We refer to [SS2, §2] for basic
properties and further references.
Given an integer partition
let mi (λ) denote the number of λj that are
Q λ = (λ1 , . . . , λr ),Q
equal to i, and set λ! = i mi (λ)! and zλ = λ! i imi (λ) .
P
For an integer n, let pn = pn (α) denote the power sum symmetric polynomial di=1 αin .
For a partition λ = (λ1 , . . . , λr ), let pλ denote the polynomial pλ1 · · · pλr . We allow d to be
infinite. The notation sλ is reserved for the Schur function. By [Sta, 7.17.5, 7.18.5], we have
X
(2.3)
sλ =
tr(cµ |Mλ )zµ−1 pµ .
|µ|=|λ|
3. Formal characters
In this section, we will assume all schemes X have an ample
line bundle L.1
L
dfg
Let V be an object of VX . Recall that V decomposes as λ Vλ ⊗Sλ (V) where Vλ ∈ Modfg
X.
We define the formal character of V as
X
ΘV =
[Vλ ]sλ ,
λ
where, as usual, [Vλ ] is the class of Vλ in K(X) and sλ ∈ Λ is the Schur function. Thus ΘV
is a potentially infinite series whose terms belong to Λ ⊗ K(X). Since Λ is the polynomial
ring in the complete homogeneous symmetric functions {sn }n≥1 , one can also think of ΘV
as a power series in these variables with coefficients in K(X). When X is a point, dim(Vλ )
is the class of the vector space Vλ in K(X) ∼
= Z.
′
We write K (X) for the Grothendieck group of vector bundles on X. The group K(X) is
a module over K′ (X), and there is a natural map K′ (X) → K(X).
Remark 3.1 (Splitting principle). Given a symmetric polynomial f (x1 , . . . , xn ) and a vector
bundle E of rank n on X, then we define f ([E]) ∈ K′ (X) as follows: if E has a filtration
whose associated graded is L1 ⊕ · · · ⊕ Ln , where the Li are line bundles, then f ([E]) =
f ([L1 ], . . . , [Ln ]). In general, consider the relative flag variety π : Flag(E) → X. In this
case, π ∗ E has a filtration by line bundles, and we define f ([E]) := Rπ∗ (f ([π ∗ E])). (See [Fu,
1Recall
that this means that for every coherent sheaf F, F ⊗ Ln is generated by global sections for n ≫ 0.
This is in particular satisfied if X is a quasi-projective variety, or an affine scheme.
6
STEVEN V SAM AND ANDREW SNOWDEN
§3.2]; note that when E is a line bundle, Flag(E) = X and π ∗ and π∗ are both identity
maps.)
We now introducePsome notation
needed to state our main result on formal characters.
n
For k ≥ 0, let σk = n≥k k sn . (In the definition of σ0 we use the convention s0 = 1.) For
a partition λ, we put
Y
σλ =
σnmn (λ)
n≥1
where mn (λ) is the multiplicity of n in λ.
Lemma 3.2. The elements σ0 , σ1 , . . . are algebraically independent over Q.
Proof. Let md0 ,...,dr be the monomial (σ0 −1)d0 σ1d1 · · · σrdr . Let fd0 ,...,dr (k) denote the sum of the
coefficients of all sλ with |λ| = k. Then fd0 ,...,dr (k) is a polynomial of degree d1 +2d2 +· · ·+rdr .
Moreover, the degree of the leading term (under the usual grading where deg sλ = |λ|) of
md0 ,...,dr is d0 + d1 + 2d2 + · · · + rdr . Hence any linear dependence among the m’s can be
made homogeneous if we assign md0 ,...,dr the bidegree (d0 , d1 + 2d2 + · · · + rdr ).
Since σ0 − 1 is a nonzerodivisor, it suffices to handle the case d0 = 0. But now we can
simply observe that the σk ’s with k ≥ 1 are related to the sn ’s with n ≥ 1 by an upper
triangular change of variables, and the sn ’s are algebraically independent.
(r)
e = Λ ⊗ Z[σ0 , σ1 , . . . ] be the polynomial ring in the sn ’s and σn ’s. Let m be the
Let Λ
λ
monomial symmetric polynomial in r variables associated to the partition λ, and define
(r)
(r)
Mλ (x1 , . . . , xr ) = mλ (x1 − 1, . . . , xr − 1).
For a proper map π : Y → X, we define a bilinear pairing
h, i : K′ (Y ) × K(Y ) → K(X),
h[E], [F]i = [Rπ∗ (E ⊗ F)].
We define a map
e ⊗ K(X),
θr : Λ ⊗ K(Grr (E)) → Λ
X
r−ℓ(λ)
(r)
sµ ⊗ [F] 7→ sµ
σ λ · σ0
· hMλ ([Q]), [F]i.
ℓ(λ)≤r
Here Q is the tautological quotient bundle on Grr (E). We will see below that this sum is
finite: the maximum value of λ1 for which the term can be non-zero is bounded by a function
of dim(X) and rank(E). We can now state our main result:
P
E
Theorem 3.3. Let M ∈ Dbfg (A(E)), and write [M] = rank
r=0 cr with cr ∈ Λ ⊗ K(Grr (E))
per Theorem 2.1. Then
rank
XE
ΘM =
θr (cr ).
r=0
In particular, ΘM is a polynomial in the sn ’s and σn ’s with coefficients in K(X).
Before proving the theorem, we note a few corollaries. First, in Lemma 3.8 below, we
show that θr is injective. Since everything in the image of θr has total degree r in the σ’s, it
follows that the images of the θr ’s are linearly independent. We conclude:
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
7
Corollary 3.4. The map
e ⊗ K(X)
Θ : K(A(E)) → Λ
is injective. Thus if M and N are two finitely generated A(E)-modules then [M] = [N] in
K(A(E)) if and only if [Mλ ] = [Nλ ] in K(X) for all λ.
Let F • K(Grr (E)) be the filtration by codimension of support. We also consider the topo•
logical filtration Ftop
K′ (Grr (E)) on K′ (Grr (E)) which is a ring filtration (see [FL, §V.3]). In
(r)
|λ|
particular, Mλ ([Q]) ∈ Ftop K′ (Grr (E)); by definition, given a subvariety of dimension < |λ|,
(r)
Mλ ([Q]) can be represented by a complex whose homology is 0 along that subvariety. It
(r)
follows that hMλ ([Q]), [F]i = 0 if F has support dimension < |λ|.
For the following, let ModA,≤r be the subcategory of A-modules which are locally annihilated by powers of the rth determinantal ideal ar , and let T>r : ModA → ModA / ModA,≤r be
the localization functor. It admits a section functor (right adjoint) S>r : ModA / ModA,≤r →
ModA and we set Σ>r = S>r ◦ T>r . This is a left exact functor, and induces a functor RΓ>r
on D(A). Next, we define Γ≤r to be the functor that assigns a module M to the maximal
submodule which is locally annihilated by ar . This is also left exact and induces a functor
RΓ≤r on D(A). Finally, we let Dbfg (A)r be the full subcategory of Dbfg (A) of objects which
are sent to 0 by RΓ≤r−1 and RΣ>r . See [SS5, §6.1] for more details.
Corollary 3.5. Suppose M ∈ Dbfg (A(E))r and [M] belongs to Λ ⊗ F k K(Grr (E)). Then ΘM
P
r−ℓ(λ) λ
has the form |λ|≤k cλ σ0
σ , where cλ ∈ Λ ⊗ K(X).
The point of the corollary is that, if one gives σn degree n, then the degree of ΘM is related
to the support dimension of M on Grr (E). Of course, if M is not in Dbfg (A(E))r then one
can apply the corollary separately to each projection RΣ≥r RΓ≤r (M).
We now begin with the proof of the theorem. We start with a few lemmas. If V ∈ Vdfg
X
is OX -flat, we let Θ′V be defined like ΘV except we use the class of Vλ in K′ (X). Under
the natural map K′ (X) → K(X), we have that Θ′V maps to ΘV . Thus Θ′V contains more
information than ΘV , when it is defined. We note that if V, W ∈ Vdfg
X and V is OX -flat
then ΘV ⊗W = Θ′V · ΘW , where the multiplication on the right side uses the K′ (X)-module
structure on K(X).
Lemma 3.6. Let L be a line bundle on Y . Let ℓ = [L] − 1 ∈ K′ (Y ) and d = dim(Y ). Then
Θ′A(L) =
d
X
ℓm σm .
m=0
Proof. By [FL, Corollary V.3.10], ℓd+1 = 0. We have
Θ′A(L) =
X
n≥0
[L]n sn =
X
(ℓ + 1)n sn =
d
XX
n
n≥0 m=0
n≥0
m
ℓ m sn =
d
X
m=0
ℓm
#
X n
sn .
m
n≥m
"
Lemma 3.7. Let Q be a locally free coherent sheaf of rank r on Y . Then
X
(r)
r−ℓ(λ)
Θ′A(Q) =
Mλ ([Q])σ λ σ0
.
ℓ(λ)≤r
8
STEVEN V SAM AND ANDREW SNOWDEN
Proof. By the splitting principle, we
Q may assume [Q] = [L1 ] + · · · + [Lr ] for line bundles Li .
Put ℓi = [Li ] − 1. Since [A(Q)] = ri=1 [A(Li )], we use Lemma 3.6 to get
" d
#
r
r
Y
Y
X
X (r)
r−ℓ(λ)
mλ (ℓ1 , . . . , ℓr )σ λ σ0
.
Θ′A(Q) =
Θ′A(Li ) =
ℓni σn =
i=1
i=1
n=0
λ1 ≤d,
ℓ(λ)≤r
Since ℓd+1
= 0, the bound λ1 ≤ d in the index for the sum is superfluous. We note that
i
(r)
(r)
mλ (ℓ1 , . . . , ℓr ) = Mλ ([Q]), by definition.
Proof of Theorem 3.3. Fix r, let Y = Grr (E), let π : Y → X be the structure map, and let
Q be the tautological bundle on Y . By [SS5, Corollary 6.16], it suffices to prove the result
for M = Rπ∗ (V ⊗ A(Q)) with V ∈ Vfg
Y . In fact, since everything in sight is V- or Λ- linear,
fg
it suffices to treat the case V ∈ ModY . By Lemma 3.7, we have
X
(r)
r−ℓ(λ)
ΘV ⊗A(Q) = ΘV · Θ′A(Q) =
Mλ ([Q])[V ]σ λ σ0
.
ℓ(λ)≤r
Since formation of formal characters is compatible with derived pushforward, we find
X
(r)
r−ℓ(λ)
ΘM = Rπ∗ (ΘV ⊗A(Q) ) =
hMλ ([Q]), [V ]iσ λ σ0
.
ℓ(λ)≤r
Prank E
Now, in the decomposition [M] = r=0 cr with cr ∈ Λ ⊗ K(Grr (E)), we have ci = 0 for
P
E
i 6= r and cr = [V ]. Thus the above is exactly equal to rank
r=0 θr (cr ), which proves the
theorem.
Lemma 3.8. The map θr is injective.
P
(r)
e if
Proof. By definition of Λ,
i sµi ⊗ [Fi ] maps to 0, then hMλ ([Q]), [F]i = 0 for all λ with
ℓ(λ) ≤ r. Via a change of basis, this implies that h[Sλ (Q)], [F]i = 0 for all such λ.
Also, K(Grr (E)) is generated by [Sλ† (R∗ )⊗π ∗ G] where [G] ∈ K(X), λ ⊆ r ×(d−r), and we
whenever λ, µ ⊆ r × (d − r) by [SS5, Corollary A.3].
have h[Sλ† (R∗ ) ⊗ π ∗ G], [Sµ (Q)]i
P = δλ,µ [G]
∗
∗
In particular, write [F] =
λ [Sλ† (R ) ⊗ π Fλ ]. Applying h−, [Sµ (Q)]i, we conclude that
[Fµ ] = 0 for all µ, so [F] = 0.
In [SS5, §7], we define an equivalence of categories
FE : Dbfg (A(E))op → Dbfg (A(E∨ ))
called the “Fourier transform.” To close this section, we examine how FE affects formal
characters. In what follows, we write F in place of FE .
The definition of F depends on the choice of a dualizing complex ωX on X, so we fix
a choice now. We let D be the induced duality on Dbfg (X) and K(X) and K′ (X). Let
(−)‡ : Λ → Λ be the ring homomorphism defined by s‡λ = (−1)|λ| sλ† . We extend (−)‡ to
infinite formal linear combinations of sλ ’s in the obvious manner.
Lemma 3.9. We have the identity
X
n≥0
σn‡ tn =
"
X
n≥0
σn tn
#−1
e
In particular, σ0‡ = σ0−1 and σn‡ ∈ Λ[1/σ
0 ] for all n ≥ 0.
.
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
Proof. We have
X
X k
tn sk =
(1 + t)n sn ,
σn t =
n
n≥0
k≥0
n,k≥0
X
and so
9
n
X
σn‡ tn =
n≥0
X
(1 + t)n (−1)n s1n .
k≥0
Now, we have the basic identity
X
n≥0
sn
!
X
(−1)n s1n
n≥0
!
= 1.
Since Λ is graded with sn and s1n of degree n, the identity continues to hold if we replace
sn by (1 + t)n sn and s1n by (1 + t)n s1n .
Our main result on the Fourier transform is:
Proposition 3.10. Let M ∈ Dbfg (A(E)). Then ΘF (M ) = D(Θ′A(E) · Θ‡M ).
Proof. In [SS5, §7.1.1], we define the Koszul duality functor K = KE . Let T (M) = M ⊗LA(E)
OX , and let T (M)i denote the degree i piece of this object, regarded as a complex in VX .
By [SS5, Prop 7.1], we have
M
A(E)
Hn (K (M)) =
Tori−n (M, OX ),
i∈Z
and so
ΘK (M ) =
X
(−1)i+n ΘT (M )i .
i,n∈Z
We define F (M) by applying D and (−)† to K (M). Note that the above sum is ΘT (M )
with the terms of odd degree multiplied by −1; when we apply (−)† to this, we get exactly
Θ‡T (M ) . We thus have the identity
ΘF (M ) = D(Θ‡T (M ) ).
V
Now, the complex • (E ⊗ V) ⊗ M computes T (M). The formal character of this complex
is exactly (Θ′A(E) )‡ · ΘM . We thus have the identity
ΘT (M ) = (Θ′A(E) )‡ · ΘM .
Combining with the previous equation, we obtain the stated result.
Corollary 3.11. Suppose E is trivial of rank n. Then ΘF (M ) = σ0n · D(Θ‡M ).
e ⊗ K′ (X).
Proof. In this case, Θ′A(E) = (Θ′A(C) )n , and it is clear that Θ′A (C) = σ0 ⊗ 1 ∈ Λ
10
STEVEN V SAM AND ANDREW SNOWDEN
4. The main theorems on Hilbert and Poincaré series
4.1. Standard Hilbert series. Pick V ∈ Vdfg
of as a sequence of Sn X , which we think L
representations on coherent OX -modules, and decompose it as V = λ Vλ ⊗ Mλ . We define
its Hilbert series by
X
t|λ|
HV (t) =
[Vλ ] dim(Mλ )
,
|λ|!
λ
where [Vλ ] is the class of Vλ in K(X). Thus HV (t) is a power series with coefficients in the
group K(X). In [Sno], the following theorem was proved:
Theorem 4.1. Suppose X is a point and that M is a finitely generated A(E)-module, where
P
E is a d-dimensional vector space. Then HM (t) = di=0 pi (t)eit for some polynomials pi (t).
In this section, we generalize this theorem, and examine how the form of HM (t) relates to
the structure of M. To a large extent, we answer [Sno, Question 3].
For a vector bundle E on X, define
µr,0 : K(Grr (E)) → K(X)
X
hmλ ([Q]∗ ), [F]i
[F] 7→
λ1 ≤d,
ℓ(λ)≤r
t|λ|
ert .
λ1 ! · · · λr !
We define a map
µr (sλ ⊗ x) = t|λ| µr,0(x).
µr : Λ ⊗ K(Grr (E)) → K(X)[t],
b = Q Λd . From [Sta, §7.8], we have a ring homomorphism
Let Λ
d≥0
In particular, we have
b → Q[[t]]
ex : Λ
X
tn
f 7→
(coefficient of x1 x2 · · · xn in f ) .
n!
n≥0
t|λ|
ex(sλ ) = dim(Mλ )
,
|λ|!
tk t
ex(Sk ) = e ,
k!
and hence ex(ΘM ) = HM (t). From the above discussion and the definitions of θr and µr , we
have ex ◦ θr = µr .
Our main theorem on Hilbert series is then:
P
Theorem 4.2. Let M be a finitely generated A(E)-module, and write [M] = dr=0 cr with
cr ∈ Λ ⊗ K(Grr (E)) per Theorem 2.1. Then
HM (t) =
d
X
µr (cr )ert .
r=0
Proof. By Theorem 3.3, we have ΘM =
Pd
r=0 θr (cr ).
Now apply ex to both sides.
One consequence of this theorem is that the coefficient of ert in HM (t) depends only on the
projection RΠr (M) of M to Dbfg (A)r . We now examine how the Fourier transform interacts
with Hilbert series.
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
11
Proposition 4.3. Let M be a finitely generated A(E)-module. Then
HFE (M ) (t) = D(edt HM (−t)).
P
P
Thus if HM (t) = dr=0 pr (t)ert , with pr (t) ∈ K(X)[t], then HFE (M ) (t) = dr=0 p∗d−r (−t)ert ,
where p∗r (t) is obtained from pr (t) by applying D to its coefficients.
Proof. Apply ex to Proposition 3.10.
4.2. Enhanced Hilbert series. Let λ be a partition. Recall that mi (λ) is the number of
times i occurs in λ. Define
m (λ) m2 (λ)
t2
tλ = t1 1
··· ,
and
Xλ (t) =
λ! = m1 (λ)!m2 (λ)! · · ·
X
|µ|=|λ|
tr(cµ |Mλ ) ·
tµ
.
µ!
Thus Xλ (t) is a polynomial in the variables ti that encodes the character of Mλ .
Lemma 4.4. The map
is an isomorphism of rings.
ϕ : Λ ⊗ Q → Q[ti ],
sλ 7→ Xλ
Proof. The pn are algebraically independent generators for Λ ⊗ Q, so as can be seen from
(2.3), ϕ can be described as pn 7→ ntn , which is evidently a ring isomorphism.
We can also extend ϕ to a map
b ⊗ Q → Q[[ti ]].
ϕ: Λ
dfg
Now
Llet V ∈ VX . Thinking of V in the symmetric group model, we can decompose it as
V = λ Vλ ⊗ Mλ . We define the enhanced Hilbert series of V by
X
e V (t) =
H
[Vλ ]Xλ (t).
λ
This is a power series in the variables ti with coefficients in K(X). This series was introduced
in [SS1] as an improvement of the Hilbert series, the idea being that it records the character
e V (t) is obtained
of the Sn -representation of Vn rather than just its dimension. Clearly, H
from the formal character ΘV by applying the ring isomorphism ϕ. Thus the two invariants
contain the same information and are just packaged in a different manner. Our results on
the formal character can be easily translated to the language of enhanced
Hilbert series. We
P
n
just state the rationality theorem here. For k ≥ 0, put Tk = n≥k k tk . It is a simple
exercise to show that ϕ(σn ) has the form p(T ) exp(T0 ) where p(T ) is a polynomial in the Tk
with k ≥ 1. Applying this to Theorem 3.3 yields:
Theorem 4.5. Let M be a finitely generated A(E)-module. Then
e M (t) =
H
d
X
pi (t, T )erT0
r=0
where pi (t, T ) is a polynomial in t1 , t2 , . . . and T1 , T2 , . . . with coefficients in K(X).
12
STEVEN V SAM AND ANDREW SNOWDEN
Remark 4.6. One obtains the usual Hilbert series HM (t) from the enhanced Hilbert series
e M (t) by setting ti = 0 for i ≥ 2. Thus the above theorem recovers our earlier rationality
H
result for the non-enhanced Hilbert series.
4.3. Elementary proof of Theorem 4.5. We now give an elementary proof (i.e., not using
the results of [SS5]) of our main theorem on enhanced Hilbert series. The proof follows the
e M (t)
proof of rationality of the usual Hilbert series given in [Sno, §3.1]. Namely, we express H
d
in terms of the T -equivariant Hilbert series of M(C ), where T is the standard maximal torus
in GL(d) for d sufficiently large. Once this expression is obtained, a formal manipulation
gives the theorem.
Let M be a finitely generated A-module, where A is a tca finitely generated in degree 1.
We assume (without loss of generality) that M is a finitely generated module over A = A(U)
for some finite dimensional vector space U. Fix an integer d ≥ max(ℓ(A), ℓ(M)).
L
Let V be a vector space (possibly infinite dimensional) on which T acts. Write V =
Vα ,
where the sum is over the characters α of T and Vα denotes the α-weight space. Assuming
each Vα is finite dimensional, we define the T -equivariant Hilbert series of V by
X
HV,T (α) =
dim(Vα )α.
α
αi±1 ,
Each character α is a monomial in the
and so HV,T (α) can be regarded as a formal
series in these variables. Note that if V comes from a polynomial representation of GL(d)
then no αi−1 ’s appear in HV,T (α).
Regard A and M as Schur functors and evaluate on the vector space Cd . We obtain a
finitely generated C-algebra A(Cd ) and a finitely generated A(Cd )-module M(Cd ), equipped
with compatible actions of GL(d). We can form the T -equivariant Hilbert series of M(Cd ),
which we denote by HM (Cd ),T (α). By the comments of the last paragraph, this is a formal
series in the αi . A simple argument with equivariant resolutions over A(Cd ) (see [Sno,
Lemma 3.3]), shows that
h(α)
HM (Cd ),T (α) = Qd
,
n
i=1 (1 − αi )
where h(α) is a polynomial in the αi ’s and n = dim(U).
Define
X
tλ
K(t, α) =
pλ (α) ,
λ!
λ
where the sum is taken over all partitions λ. The following lemma gives a different expression
for K(t, α).
Lemma 4.7. For a partition λ, let sλ (α) ∈ K(T ) denote the character of Sλ (Cd ). Then
X
e M (t),
K(t, α) =
sλ (α)H
λ
λ
where the sum is over all partitions λ with ℓ(λ) ≤ d.
Proof. By [Sta, Corollary 7.17.4], we have
X
pµ =
tr(cµ |Mλ )sλ
λ
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
13
as elements of K(T ), where the sum is over partitions λ of |µ| with ℓ(λ) ≤ d. Multiplying
by tµ /µ! and summing over µ, we find
X
tµ X e
sλ HMλ .
K(t, α) =
tr(cµ |Mλ )sλ =
µ!
λ
λ,µ
e M (t) to the more accessible object HM (Cd ),T .
The following is the key lemma, relating H
Lemma 4.8. We have
e M (t) = 1
H
d!
Z
HM (Cd ),T (α)K(t, α)|∆(α)|2dα.
T
Proof. Let aλ be the multiplicity of Mλ in M. Note that aλ = 0 if d < ℓ(λ), by definition of
d. We have
X
M ⊕a
e M (t).
e M (t) =
aλ H
and
H
Mλ λ ,
M=
λ
λ
λ
On the other hand,
M(Cd ) =
M
Sλ (Cd )⊕aλ ,
and
HM (Cd ),T (α) =
λ
We therefore have
e M (t) =
H
1
=
d!
=
1
d!
aλ sλ (α).
λ
X
λ
X
e M (t)
aλ H
λ
Z
T
Z
X
λ
aλ sλ (α)
!
X
µ
!
e Mµ (t) |∆(α)|2 dα
sµ (α)H
HM (Cd ),T (α)K(t, α)|∆(α)|2dα.
T
To go from the first line to the second, we used Weyl’s integration formula (2.2). Both sums
in the second equation are over partitions with ℓ ≤ d; in particular, both sλ and sµ are
characters of (non-zero) irreducible representations of GL(d). To go from the second line to
the third, we used Lemma 4.7.
We need one more lemma before proving the theorem.
Lemma 4.9. Let u1 , . . . , ud be indeterminates. We then have
!
Z
d
∞
Y
X
1
K(t, α)dα =
exp
uni tn .
Qd
(1
−
α
u
)
T
i i
i=1
i=1
n=1
Proof. For x ∈ Zd≥0 , let αx denote α1x1 · · · αdxd , and similarly define ux . Then we have
Z
X
1
K(t, α)dα =
(the coefficient of αx in K(t, α))ux
Qd
(1
−
α
u
)
T
i i
i=1
x
X
ux tλ
.
=
(the coefficient of αx in pλ )
λ!
x,λ
14
STEVEN V SAM AND ANDREW SNOWDEN
Here x varies over Zd≥0 , while λ varies over all partitions. Now, a simple computation shows
that the coefficient of αx in pλ is given by
X
λ!
,
µ
!
·
·
·
µ
!
1
d
µ ,...,µ
1
d
where the sum is taken over partitions µ1 , . . . , µd such that µ1 ∪ · · · ∪ µd = λ and |µi | = xi .
We thus find
µ1
µd
X |µ |
X
ux tλ
|µ | t · · · t
(the coefficient of αx in pλ )
=
u1 1 · · · ud d
λ!
µ1 ! · · · µd !
µ1 ,...,µd
x,λ
!
!
X |µ| tµ
X |µ| tµ
···
.
ud
=
u1
µ!
µ!
µ
µ
Now, we have
X
µ
z |µ|
tµ
=
µ! n
=
X
1 ≥0,
∞
Y
k=1
z n1 +2n2 +···
n2 ≥0, ...
tn1 1 tn2 2
···
n1 ! n2 !
#
∞
n
Y
t
kn k
exp(z k tk ).
=
z
n!
n≥0
"
X
k=1
Combining the previous two equations gives the stated result.
Proof of Theorem 4.5. Put
ϕ0 (x, t) =
X
xi ti ,
i≥1
and, more generally,
ϕn (x, t) =
X
i≥1
i(i − 1) · · · (i − n + 1)xi−n ti =
dn
ϕ0 (x, t).
dxn
We have ϕn (1, t) = Tn and ϕn (0, t) = n!tn (with t0 interpreted as 0). Lemma 4.9 can be
restated as
Z
1
K(t, α)dα = exp(ϕ0 (u1 , t) + · · · + ϕ0 (ud , t)).
(4.10)
Qd
T
i=1 (1 − αi ui )
Finally, suppose we are given an expression
Z
p(α)
K(t, α)dα
Qd
ni
T
i=1 (1 − αi )
where p(α) is a polynomial. We claim that it is a polynomial in exp(T0 ) and {Ti , ti | i ≥ 1}.
e M (t) can be expressed as such an integral
This claim proves the theorem, as we know that H
by Lemma 4.8.
We now prove the claim. When ni = 0 for all i, the integral in question is a polynomial
in the ti , so there is nothing to show in that case. By linearity and applying polynomial
long division in each variable separately, we may assume that p(α) is a sum of monomials
α1m1 · · · αdmd where mi < ni for all i. Now apply ( ∂u∂ 1 )n1 −1 · · · ( ∂u∂ d )nd −1 to (4.10) and then set
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
15
ui = 1 for all i. The result is a polynomial in exp(T0 ) and {Ti | i ≥ 1}, which can be seen
by using the right side of (4.10). The left side becomes a scalar multiple of
Z n1 −1
α1
· · · αdnd −1
K(t, α)dα.
Qd
ni
T
i=1 (1 − αi )
For any mi ≤ ni , we then also know that
Z Qd
mi −1
(1 − αi )ni −mi
i=1 αi
K(t, α)dα
Qd
ni
T
i=1 (1 − αi )
is a polynomial in exp(T0 ) and {Ti | i ≥ 1}. Finally, for fixed ni > 0, note that the set
)
( d
Y
αimi −1 (1 − αi )ni −mi 1 ≤ mi ≤ ni
i=1
is a basis for the span of monomials in α1k1 · · · αdkd such that 0 ≤ ki < ni . Hence by linearity
and combining our earlier reduction, we conclude that our original expression is a polynomial
in exp(T0 ) and {Ti , ti | i ≥ 1}.
4.4. Poincaré series. Given an A-module M, define its Poincaré series by
X
(−q)n HTorAn (M,C) (t).
PM (t, q) =
n≥0
Setting q = 1 and multiplying by HA (t) = edt recovers the Hilbert series HM (t). Note that
the Poincaré series has non-trivial information about the A-module structure of M, whereas
the Hilbert series only knows about the underlying object of VX . The Poincaré series does
not factor through K-theory, so it is much harder to study than the Hilbert series. The
following is our main result about it:
Theorem 4.11. Let M be a finitely generated A-module. There exist fr (t, q) ∈ K(X)[t, q, q −1]
such that
d
X
PM (t, q) =
fr (t, q)e−rqt .
r=0
Proof. Let Nk = H−k (F (M)), where
F (M) is the Fourier transform of M. A simple formal
P
manipulation gives PM (t, q) = k≥0 (−q)−k HNk (−qt). The result now follows from the fact
that each Nk is finitely generated and only finitely many Nk are non-zero [SS5, Theorem 7.7]
and the structure theorem for HNk (Theorem 4.2).
There is much more one could say about Poincaré series and variants using the methods of
this paper: for example, one could define an “enhanced Poincaré series” or a Poincaré-variant
of the formal character, and prove results about them.
5. A categorification of rationality
5.1. Motivation. Let D : V → V be the Schur derivative. If F is a polynomial functor, then
DF is the functor assigning to a vector space V the subspace of F (V ⊕ C) on which Gm
acts through its standard character (this copy of Gm acts by multiplication on C only). In
terms of sequences of symmetric group representations, the Schur derivative takes a sequence
M = (Mn )n≥0 to the sequence DM given by (DM)n = Mn+1 |Sn . See [SS2, §6.4] for more
16
STEVEN V SAM AND ANDREW SNOWDEN
details. (There the Schur derivative is denoted D, but this conflicts with our notation for
duality functors.)
Let A = A(E) and suppose M is a finitely generated A-module. One observes that
HD(M ) (t) = dtd HM (t), and so the Schur derivative can be thought of as a categorification of
d
. There is a natural map E ⊗M → D(M). For a subspace V of E, let ∂V (M) be the 2-term
dt
complex [V ⊗ M → D(M)], where the differential is just the restriction of the previously
mentioned map to V ⊗ M. Then H∂V (M ) (t) = ( dtd − d)HM (t), where d = dim(V ), and so ∂V
is a categorification of the operator dtd − d.
The main result about HM (t) is that it isQa polynomial in t and et . It follows that HM (t)
is annihilated by an operator of the form ni=1 ( dtd − di ) for non-negative integers di . The
discussion of the previous paragraph thus suggests a stronger result, namely, that there exists
subspaces V1 , . . . , Vn such that ∂V1 · · · ∂Vn annihilates M. We will show that this is indeed
the case. This is a categorification of the rationality theorem for HM (t).
One may view this as an analogue of the existence of a system of parameters for a finitely
generated module. For some discussion of this analogy, see [SS1, Remark 5.4.2].
5.2. The main theorem. Let A = A(E) and let M be an A-module. There is then a
natural map E ⊗ M → D(M). For a subspace V of E, we let ∂V (M) be the 2-term complex
[V ⊗ M → D(M)]. More generally, for a complex M of A-modules, we let ∂V (M) be the
cone on the map V ⊗ M → D(M). This defines an endofunctor of Dbfg (A). (To actually get
a functor, one should work with the homotopy category of injective objects.) We define a
differential operator on Dbfg (A) to be a finite composition of endofunctors of the form ∂V .
We say that M satisfies a differential equation if there exists a differential operator ∂
such that ∂(M) = 0.
Theorem 5.1. Every object of Dbfg (A) satisfies a differential equation.
Lemma 5.2. Any two differential operators commute (up to isomorphism). In particular,
if ∂ is a differential operator and M satisfies a differential equation then ∂(M) also satisfies
a differential equation.
Lemma 5.3. In an exact triangle in D(A), if two terms satisfy differential equations then
so does the third.
Let r ≥ 0 be an integer, let Y = Grr (E), let π : Y → Spec(C) be the structure map, let
Q be the tautological bundle on Y , and let B = A(Q). For V ⊂ E, we define the differential
operator ∂V on B-modules just as for A-modules, and we define a differential operator on
D(B) to be a composition of ∂V ’s.
Lemma 5.4. Suppose M ∈ Db (B) satisfies a differential equation and F ∈ Vfg
Y . Then
F ⊗LOY M also satisfies a differential equation.
Proof. By induction on Tor-dimension, we can assume F is OY -flat. We have
D(F ⊗ M) = (D(F) ⊗ M) ⊕ (F ⊗ D(M)).
The map E ⊗ (F ⊗ M) → D(F ⊗ M) maps into the factor F ⊗ D(M). We thus see that
∂V (F ⊗ M) = (F ⊗ ∂V (M)) ⊕ (D(F) ⊗ M).
We thus see that if ∂ is a differential operator annihilating M then ∂(F ⊗ M) is a direct sum
of objects of the form F ′ ⊗ M ′ , where F ′ has strictly lower degree than F and M ′ is the result
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
17
of applying some differential operator to M. Since each M ′ satisfies a differential equation,
the result follows by induction on the degree of F.
Lemma 5.5. B satisfies a differential equation.
Proof. We have ∂V (B) = [V → Q] ⊗ B. Since [V → Q] is a complex of OY -modules, it pulls
out of differential operators. We thus see that if ∂ = ∂V1 · · · ∂Vn then
∂(B) = [V1 → Q] ⊗ · · · ⊗ [Vn → Q] ⊗ B.
For appropriate choices of the Vi ’s, the above complex is acyclic: indeed, simply choose the
Vi ’s so that for every y ∈ Y there is some i for which Vi (y) → Q(y) is an isomorphism.
Proof of Theorem 5.1. It is clear that differential operators commute with Rπ∗ . We thus
L
b
see that if F ∈ Vfg
Y then the object Rπ∗ (F ⊗OY B) of Dfg (A) satisfies a differential equation.
These objects generate Dbfg (A) by [SS5, Corollary 6.16] (note that by [SS5, Remark 6.15], one
can take the F to be OY -flat). Thus all objects of Dbfg (A) satisfy a differential equation.
5.3. Differential operators and duality. Let M be a finitely generated A(E)-module.
By Proposition 4.3, we have
HFE (M ) (t) = edt HM (−t).
For any differentiable function f (t), we have the identity
d
d
− d edt f (−t) = −edt f (−t).
dt
dt
e obtained by applying the
Hence, if ∂ is a differential operator that annihilates f (t), then ∂,
d
d
substitution dt 7→ dt − d to ∂, is a differential operator that annihilates edt f (−t).
We can realize this identity using differential operators as follows. Pick a minimal free
resolution F• of M. Note that
D(V ⊗ A) = (D(V ) ⊗ A) ⊕ (V ⊗ E ⊗ A).
Furthermore, for any subspace V ⊂ E, we have FE ◦∂V [1] = ∂(E/V )∗ ◦FE (see lemma below).
This implies that if ∂ annihilates M, then ∂e annihilates FE (M), where ∂e is obtained from
∂ by substitution of ∂(E/V )∗ into ∂V .
Lemma 5.6. For V ⊂ E we have an isomorphism of endofunctors FE ◦∂V [1] ∼
= ∂(E/V )∗ ◦FE
of D(A).
Proof. (We omit many of the details in this proof to keep it short.) By definition, we have
FE (M) = (M ⊗A K)R,∨,† ,
where K = K(E) is the Koszul complex and (−)∨ denotes vector space duality on multiplicity spaces. Using the fact that D satisfies the Leibniz rule and D(Sym(E[1])) =
E[1] ⊗ Sym(E[1]), we find
D((M ⊗A K)R,∨,† ) = ((D(M) ⊕ (E[1] ⊗ M)) ⊗A K)R,∨,† = FE (D(M) ⊕ (E[1] ⊗ M)).
The above identity neglects the differentials: in fact, the complex inside the Fourier transform
is actually the cone on the canonical map E ⊗ M → D(M), namely ∂E (M). We thus find
that ∂(E/V )∗ (FE (M)) is the cone on the natural map
FE ((E/V ) ⊗ M) = (E/V )∗ ⊗ FE (M) → D(FE (M)) = FE (∂E (M)).
18
STEVEN V SAM AND ANDREW SNOWDEN
A computation shows that this map is just FE applied to the canonical map ∂E (M) →
E/V ⊗ M. (Recall that ∂E (M) is the cone of the map E ⊗ M → D(M); the map here comes
from the canonical map E ⊗ M → E/V ⊗ M.) The cone of the map ∂E (M) → E/V ⊗ M is
quasi-isomorphic to the cone of V ⊗ M → D(M), i.e., ∂V (M). This completes the proof.
5.4. The category of differential operators. We end this section with some additional
thoughts on differential operators. It could be interesting to pursue these observations further
in the future.
Write E for the endofunctor M 7→ E ⊗ M of ModA , and write D for the endofunctor
given by the Schur derivative. There is then a natural transformation E → D, from which
all of the differential operator structure derives. Let G ⊂ GL(E ⊕ V) be the subgroup fixing
all elements of E, and let C be the category of polynomial representations of G. Then C
is the universal tensor category having an object X and a morphism E → X (one takes
X = E ⊕ V): that is, given any tensor category D and a morphism E → X in D, there is a
unique left-exact tensor functor C → D mapping E → E ⊕ V to E → X. (The group G is
an example of a “general affine group,” and the universal property of C, in the d = 1 case,
can be found in [SS3, §5.4]. Note that E is simply isomorphic to a direct sum of d copies of
the unit object of C.)
The universal property of C furnishes a functor C → End(ModA ) taking E → E ⊕ V
to the transformation E → D discussed above. Alternatively, we can think of this functor
as an action C × ModA → ModA . This action induces one on derived categories Db (C) ×
Db (A) → Db(A). For V ⊂ E, the 2-term complex [V → X] in Db (C) acts on Db (A) as the
differential operator ∂V . The tensor product structure on Db (C) corresponds to composition
of differential operators.
Our main theorem on differential equations states that every object of Dbfg (A) is annihilated
by some object of Dbfg (C). It would be interesting if this result could be made more precise.
For example, given M ∈ Dbfg (A), what does the annihilator in Dbfg (C) look like? Is it principal,
as a tensor ideal?
Another potentially interesting observation: C is equivalent to both the category Mod0A
of A-modules supported at 0 and the category Modgen
A of “generic” A-modules (see [SS5,
Propositions 5.4, 5.6]). Is this more than a coincidence? Does the action of C on ModA come
from, or extend to, an interesting action of ModA on itself? We have not fully investigated
these questions.
6. D-finiteness of Hilbert series of bounded modules
Recall that a power series f ∈ CJtK is D-finite if it satisfies a differential equation of the
form
n
X
di f
pi (t) i = 0,
dt
i=0
where each pi is a polynomial in t. See [Sta, §6.4] for some basic properties and examples.
In this section, we prove the following theorem.
Theorem 6.1. Let A be a finitely generated tca with A0 = C and let M be a finitely generated
A-module which is bounded. Then HM (t) is D-finite.
We can assume that A is a polynomial tca. Recall the notation from §4.3.
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
19
Lemma 6.2. Suppose max(ℓ(A), ℓ(M)) < ∞. For d ≥ max(ℓ(A), ℓ(M)), we have
Z
P
HM (t) =
HM (Cd ),T (α)e αi t |∆|2 dα.
T
e M (t) by doing the substitution
Proof. HM (t) is obtained from the enhanced Hilbert series H
ti 7→ 0 for i ≥ 2 and t1 7→ t. Apply this substitution to Lemma 4.8. Then K(t, α) becomes
X
λ
dim(Mλ )sλ (α)
X
t|λ|
tn
=
(α1 + · · · + αd )n = exp(α1 t + · · · + αd t)
|λ|! n≥0
n!
where the first equality comes from Schur–Weyl duality (or see [Sta, Corollary 7.12.5]).
We have A(Cd ) = Sym(V ) (we warn the reader that this is evaluation of A on the vector
space Cd , not to be confused with A(Cd )) forLsome polynomial representation V of GL(d).
n
Ai
d
Ai
Write a T -equivariant decomposition V =
means
i=1 Cα , where Ai ∈ Z≥0 and α
Ai,d
Ai,1
α1 · · · αd . We then have
HM (Cd ),T (α)|∆|2 = Q
p(α)
1 − α Ai
for some polynomial p. It suffices to treat the case where p is a monomial, say p(α) = αb ,
with b ∈ Zd . We then have
X
αxA+b ,
HM (Cd ),T (α)|∆|2 =
x∈Zn
≥0
and hence
X
t|xA+b|
,
(xA + b)!
x∈Zn
≥0
P
Q
where for y ∈ Zd≥0 we write |y| = i yi and y! = i yi !. Put
X
X
#{x ∈ Zn≥0 | xA + b = y} · uy .
uxA+b =
F (u1 , . . . , ud ) =
HM (t) =
x∈Zn
≥0
y∈Zd≥0
Then F is a rational function of the ui . Indeed, let R be the ring C[ξ1 , . . . , ξn ] and give
R a Zd≥0 grading by deg(ξi ) = Ai . Then F is the T -equivariant Hilbert series of the free
R-module with one generator of degree b, and is therefore rational [MS, Theorem 8.20]. The
theorem then results from the following general result:
Proposition 6.3. Suppose that
X
ay uy
X
ay
y∈Zd≥0
is a rational function of the ui . Then
y∈Zd≥0
is D-finite.
t|y|
y!
20
STEVEN V SAM AND ANDREW SNOWDEN
Proof. Define the Hadamard product ∗ on multivariate generating functions by the formula
X
X
X
αy βy uy .
αy u y ∗
βy uy =
y∈Zd≥0
y∈Zd≥0
y∈Zd≥0
Furthermore, we define a multivariate generating function F to be D-finite if the vector
space (over the field of rational functions) spanned by all partial derivatives of F is finitedimensional. By [Lip, Remark 2], the Hadamard product of two D-finite generating functions
is D-finite. It is clear that a rational function is D-finite and the exponential function exp(u1 +
·P
· · + ud ) is also D-finite. Hence we conclude that their Hadamard product F (u1 , . . . , ud) =
uy
y∈Zd≥0 ay y! is also D-finite. Now do the substitution ui 7→ t for i = 1, . . . , d. By the chain
rule,
d
X
∂F
d
F (t, . . . , t) =
,
dt
∂ui
i=1
so the result is also D-finite.
7. Examples and applications
7.1. Hilbert series of polynomial tca’s. For an integer n ≥ 1, define a linear map
Q[ti ] → Q[ti ], denoted f (t) 7→ f (t[n] ), by taking tλ to tnλ , where nλ = (nλ1 , nλ2 , . . .). We
extend this map to series in the obvious manner. The following is the main result of this
section:
Proposition 7.1. Let V be a finite length object of V concentrated in degree d 6= 0 and let
A = Sym(V ). Then
!
X
e A (t) = exp
e V (t[n] ) .
H
nd−1 H
n≥1
Proof. Both sides convert direct sums in V to products, so it suffices to treat the case where
V = Sλ (C∞ ). Recall from (2.3) that
X
sλ =
zµ−1 tr(cµ |Mλ )pµ .
Q
µ
The GL-character of A is T (1 − xT )−1 where the sum is over all semistandard Young
m (T )
tableaux T of shape λ and xT is the product of xi i
where mi (T ) is the multiplicity of i
in T . Apply log to this expression:
"
#
X
X X xn
Y
T
log
log(1 − xT )−1 =
(1 − xT )−1 =
n
T
T n≥1
T
X1
X1X
=
zµ−1 tr(cµ |Mλ )pnµ (x).
sλ (xn1 , xn2 , . . . ) =
n
n µ
n≥1
n≥1
Making the substitution pi 7→ iti , we obtain the stated result (see Lemma 4.4).
e V (t) = td , and so H
e V (t[n] ) = td . We thus obtain
Example 7.2. Take V = (C∞ )⊗d . Then H
1
n
X
e A (t) = exp
H
nd−1 tdn .
n≥1
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
Note that for d = 1 this is exp(T0 ).
21
e V = 1 t2 + t2 , and so H
e V (t[n] ) =
Example 7.3. Take V = Sym2 (C∞ ), so d = 2. Then H
2 1
1 2
t + t2n . We thus obtain
2 n
X
2
n
e A (t) = exp
t + nt2n .
H
2 n
n≥1
In particular, HA (t) = exp( 21 t2 ). The case V =
to a − sign.
V2
(C∞ ) is similar: just change the + sign
Remark 7.4. In the notation of the proposition, the tca A is generated in degree d, and
so Theorem 4.5 cannot be applied to it for d > 1; in fact, the above calculations show that
e A computed
the conclusion of the theorem is false in this case. Nonetheless, the values of H
above are sufficiently nice that one might hope for a good generalization of Theorem 4.5
which includes these cases.
7.2. Formal characters of determinantal rings. Let A be the tca A(E), where E is a
d-dimensional vector
space. V
Let ar ⊂ A be the rth determinantal ideal, generated by the
V
representation r+1 (C∞ ) ⊗ r+1 (E). In this section, we derive a formula for the formal
character of A/ar .
Recall that we have an isomorphism of Grothendieck groups
(7.5)
K(A) =
d
M
r=0
Λ ⊗ K(Grr (E)).
To apply Theorem 3.3, we need to understand the class [A/ar ] in K(A) under this identification. The following lemma gives us what we need.
Lemma 7.6. Under the isomorphism (7.5), the class of [A/ar ] corresponds to 1 ⊗ [OGrr (E) ].
Proof. Let Qr be the tautological bundle on Yr = Grr (E), and let πr : Yr → Spec(C) be the
structure map. Let ir : K(Grr (E)) → K(A) be the map defined by ir ([V ]) = [R(πr )∗ (V ⊗
A(Qr ))]. According to [SS5, Theorem 6.19], the maps ir induce the isomorphism (7.5). Now,
we have
ir ([OYr ]) = [R(πr )∗ (A(Qr ))]
and
M
A(Qr ) =
Sλ (C∞ ) ⊗ Sλ (Qr ).
ℓ(λ)≤r
The vector bundle Sλ (Qr ) has no higher cohomology, and its space of sections is Sλ (E). We
thus find that R(πr )∗ (A(Qr )) = A/ar , which proves the lemma.
For the rest of the section, we fix r and put Y = Grr (E). We let Q be the tautological bundle on Y and π : Y → Spec(C) the structure map. From the above lemma and Theorem 3.3,
we see that the formal character ΘA/ar is given by θr ([OY ]). We now compute θr ([OY ]), at
(r)
least to some extent. To do this, we must compute the inner product hMλ ([Q]), [OY ]i. To
(r)
(r)
(r)
start, define Sλ (x1 , . . . , xr ) = sλ (x1 −1, . . . , xr −1) where sλ denotes the Schur polynomial
in r variables.
(r)
Lemma 7.7. If ℓ(λ) ≤ r, then hSλ ([Q]), [OY ]i = dim Sλ† (Cd−r ).
22
STEVEN V SAM AND ANDREW SNOWDEN
Proof. We first verify this when λ = (n). In that case, it follows from the definitions (or the
general formula [Mac, p. 47, Example I.3.10]) that
n
X
(r)
(r)
n−i n + r − 1
si (x1 , . . . , xr ),
sn (x1 − 1, . . . , xr − 1) =
(−1)
i
+
r
−
1
i=0
and so
Sn(r) ([Q])
=
n
X
(−1)
n−i
i=0
Now,
(r)
si ([Q])
i
n + r − 1 (r)
s ([Q]).
i+r−1 i
is simply [Sym (Q)], and so we see
d+i−1
.
]i = [Rπ∗ (Sym (Q))] =
i
Here we used the fact that π∗ (Symi (Q)) has dimension d+i−1
, and the higher pushforwards
i
vanish. We thus find
n
X
d+i−1
n−i n + r − 1
(r)
hSn ([Q]), [OY ]i =
(−1)
i
i+r−1
i=0
n
X
n + r − 1 −d
(−1)n
=
i
n−i
i=0
n−d+r−1
= (−1)n
n
d−r
= dim S(n)† (Cd−r ).
=
n
a
For the second and fourth equalities, we used the identities ab = a−b
and a+b−1
=
b
b −a
(−1) b , and the third identity is the Chu–Vandermonde identity. This proves the lemma
for λ = (n).
(r)
The above reasoning applies equally well to r-fold products of Sn . That is, we have
d−r
d−r
(r)
([Q]),
[O
]i
=
dim
S
† (C
)
⊗
·
·
·
⊗
S
† (C
)
hSn(r)
([Q])
·
·
·
S
Y
(n
)
(n
)
n
r
1
r
1
(r)
hsi ([Q]), [OY
i
for any n1 , . . . , nr . The key point here is that
[Rπ∗ (Symn1 (Q) ⊗ · · · ⊗ Symnr (Q))] = dim Symn1 (Cd ) ⊗ · · · ⊗ Symnr (Cd ) .
To see this, we first decompose the tensor product Symn1 (Q) ⊗ · · · ⊗ Symnr (Q) into a direct
sum of Schur functors Sλ (Q) with ℓ(λ) ≤ r = rank(Q) where the bound comes from the
Pieri rule. Now Borel–Weil–Bott (see [SS5, Theorem A.1] for a convenient formulation) says
that Rπ∗ Sλ (Q) = Sλ (Cd ), and we get a direct sum which is the same as the tensor product
Symn1 (Cd ) ⊗ · · · ⊗ Symnr (Cd ) (it is important that there are at most r factors in the tensor
product so that all Schur functors appearing in it have at most r rows, and therefore do not
annihilate Q).
We now prove the lemma for general λ by induction on dominance order (i.e., λ dominates
µ if |λ| = |µ| and λ1 + · · · + λi ≥ µ1 + · · · + µi for all i). The base case is when λ has a single
part, which we have already done. In general, sλ = sλ1 · · · sλr + C where C is a sum of sµ
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
23
(r)
with µ less dominant than λ. By induction, hSµ ([Q]), [OY ]i = dim Sµ† (Cd−r ) for all such µ.
By the previous paragraph,
hSλ1 ([Q]) · · · Sλr ([Q]), [OY ]i = dim(S(λ1 )† (Cd−r ) ⊗ · · · ⊗ S(λr )† (Cd−r )).
Hence the statement also holds for λ.
Let Kλ,µ be the Kostka number, i.e., the number of semistandard Young tableaux of shape
λ and content µ, so that
X
sλ =
Kλ,µ mµ .
µ
The matrix (Kλ,µ ) is invertible; let
−1
Kλ,µ
denote the entries of the inverse matrix, so that
X
−1
mλ =
Kλ,µ
sµ .
µ
The above identities also hold after specializing to any finite number of variables, as this
operation is linear. We thus find
X
X
(r)
−1
−1 (d−r)
hMλ ([Q]), [OY ]i =
Kλ,µ
hSµ(r) ([Q]), [OY ]i =
Kλ,µ
sµ† (1, . . . , 1).
ℓ(µ)≤r
ℓ(µ)≤r
We finally reach our main result:
Theorem 7.8. We have
X
ΘA/ar =
r−ℓ(λ)
cλ σ λ σ0
λ⊆r×d
where cλ is the integer given by
cλ =
X
−1
Kλ,µ
dim Sµ† (Cd−r ).
µ⊆r×(d−r)
In the case r = 1, we can give a simpler expression for the formal character by computing
directly:
P
d−1
Proposition 7.9. We have ΘA/a1 = d−1
σn .
n=0
n
Proof. We have
A/a1 =
M
n≥0
and so
ΘA/a1
Symn (E) ⊗ Symn (C∞ ),
X d + n − 1
sn ,
=
d
−
1
n≥0
where the binomial coefficient is the dimension of Symn (E). Appealing to the identity
X
d−1
n
d−1
d+n−1
,
=
d
−
i
−
1
i
d−1
i=0
we find
ΘA/a1 =
# X
d−1
d−1
n
σd−1−i ,
sn =
i
d−i−1
i=0
"
d−1
X
d−1 X
i=0
i
n≥0
from which the stated formula follows.
24
STEVEN V SAM AND ANDREW SNOWDEN
7.3. Enhanced Hilbert series of determinantal rings. Keep the same notation as the
e A/ar by a change of variables. In
previous section. The formula for ΘA/ar gives a formula for H
e A/ar directly, without using any of the theory behind Theorem 3.3.
this section, we compute H
L
n
∞ ⊕ dim Symn (E)
First consider the case r = 1. Then A/a1 =
, and so the
n≥0 Sym (C )
enhanced Hilbert series is
X |λ| + d − 1 tλ
(7.10)
.
|λ|
λ!
λ
Lemma 7.11.
X |λ| + d − 1 tλ
λ
|λ|
λ!
=
∂ d−1 sd−1
exp(st1 + s2 t2 + s3 t3 + · · · )
∂sd−1 (d − 1)!
.
s=1
Proof. The coefficient of tn1 · · · tnr in exp(st1 + s2 t2 + · · · ) is
sn1 +2n2 +···+rnr
.
n1 ! · · · nr !
d−1
s
Set n = n1 + 2n2 + · · · + rnr . If we multiply by (d−1)!
, take the (d − 1)st derivative, and
evaluate at s = 1, we get
1
n+d−1
.
n1 ! · · · nr !
n
Finally, given a partition λ, and setting ni = mi (λ), we have n = |λ|.
Using the product rule for derivatives, this can be written as
X d − 1 sj d j
exp(st1 + s2 t2 + · · · )
j
j!
ds
i
i+j=d−1
.
s=1
i
d
2
Note that Ti = i!1 ds
i (st1 + s t2 + · · · )|s=1 , so we see that this expression is a polynomial in
T1 , . . . , Tm−1 times exp(T0 ).
We list the first few derivatives of exp(f (s)) (they are of the form exp(f (s)) · F so we just
list F ):
f ′ (s)
f ′′ (s) + f ′ (s)2
f (3) (s) + f ′ (s)3 + 3f ′(s)f ′′ (s)
f (4) (s) + 3f ′′ (s)2 + f ′ (s)4 + 4f (3) (s)f ′ (s) + 6f ′ (s)2 f ′′ (s)
In particular, we get the following expressions for
the form exp(T0 ) · F so we just list F ):
dj
dsj
exp(st1 + s2 t2 + · · · )|s=1 (they are of
T1
2T2 + T12
6T3 + T13 + 6T1 T2
24T4 + 12T22 + T14 + 24T3 T1 + 12T12T2
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
25
So we have the following evaluations of (7.10) for small d:
d (7.10)
1
2
3
4
5
exp(T0 )
(T1 + 1) exp(T0 )
( 12 (2T2 + T12 ) + 2T1 + 1) exp(T0 )
( 16 (6T3 + T13 + 6T1 T2 ) + 32 (2T2 + T12 ) + 3T1 + 1) exp(T0 )
1
(24T4 + 12T22 + T14 + 24T3 T1 + 12T12 T2 ) + 32 (6T3 + T13 + 6T1 T2 ) + 3(2T2 + T12 ) + 4T1 + 1) exp(T0 )
( 24
One can verify that these quantities correspond to those in Proposition 7.9 under the
appropriate change of variables.
Using [Ges, Theorem 16], we can give a determinantal formula for the enhanced Hilbert
series of A/ar , in general, in terms of certain modifications of (7.10). Define, for any i ∈ Z,
X |λ| + i + d − 1 tλ
ai =
.
|λ| + i
λ!
λ, |λ|≥−i
Then a0 is the same as (7.10), and the enhanced Hilbert series of A/ar is the determinant
det(aj−i )ri,j=1 .
From this formula, we deduce that the enhanced Hilbert series is a polynomial in T1 , . . . , Td−1
times exp(rT0 ). This gives an independent verification of the rationality theorem in this case.
7.4. Hilbert series of invariant rings. Let E be a finite dimensional representation of a
reductive group G and let A be the tca A(E)G . Using the symmetric group model for V, A
is given by An = (E ⊗n )G . Since A(E) is bounded, so is A, and so Theorem 6.1 assures us
that HA is D-finite.
P
n
For the purposes of this section, we let H∗A (t) =
n≥0 dim(An )t . This is the nonexponential form ofPHA (t). The D-finiteness
of HA is equivalent to the D-finiteness of H∗A :
P
n
n
each of the series
n t /n! and
n n!t is D-finite, and D-finiteness is preserved under
Hadamard product by [Lip, Remark 2] (see also [Sta, Theorem 6.4.12] for a simpler reason
for univariate series).
⊗n G
Example 7.12. Take G = SL(2) and E = C2 . Then the dimension of
(E ) is 0 for n
2k
1
odd and the Catalan number Cn/2 for n even. (Recall that Ck = k+1 k .) We thus obtain
H∗A (t) =
X
Ck t2k =
k≥0
2
√
.
1 + 1 − 4t2
We note that any algebraic function is D-finite. We have
X (2k)!
t2k .
HA (t) =
k!(k + 1)!
k≥0
This is (essentially) a modified Bessel function of the first kind, and satisfies the differential
equation
dy
d2 y
t2 2 + 3t − 4t2 y = 0.
dt
dt
26
STEVEN V SAM AND ANDREW SNOWDEN
Example 7.13. Take G = SL(2) × SL(2) and E = C2 ⊗ C2 . Then (E ⊗n )G is the tensor
square of the space ((C2 )⊗n )SL(2) from the previous example, and so we find
2
X
X
1
2k
∗
2 2k
HA (t) =
Ck t =
t2k .
2
(k + 1) k
k≥0
k≥0
This series is not
P algebraic:2 2kindeed, if it were then the Hadamard product of it and the
rational series k≥0 (k + 1) t would also be algebraic [Sta, Proposition 6.1.11], but this is
2 2k
P
the series k≥0 2k
t , which is known to be transcendental [WS, Theorem 3.3]. We thus
k
∗
see that HA (t) need not always be algebraic. Thus Theorem 6.1 is optimal, in a sense.
Remark 7.14. We point out that the D-finiteness of H∗A (t) can be seen directly in this case.
Let T be a maximal torus of G, and let α1 , . . . , αr be characters of T giving an isomorphism
with Grm . Let χ be the character of E,Pregarded as a function on T (and thus a Laurent
polynomial in the αi ’s). Let F (t, α) = n≥0 tn χn = (1 − tχ)−1 . This is a rational function
of t and the α’s. By the Weyl integration formula (see §2.2 for the GL(n) case and [FH,
§26.2] for the general case), we have
Z
1
∗
HA (t) =
F (t, α)|∆|2dα,
|W | T
where W is the Weyl group of G and ∆ is a certain Laurent polynomial in the αi . The
D-finiteness of H∗A (t) now follows from the following general fact: taking the constant term
(with respect to the αi ) of the the product of a D-finite function with a Laurent polynomial
is again D-finite.
Remark 7.15. There are some similar situations where one does not have D-finiteness. For
k ≥ n, the canonical map
((Ck )⊗n )Sk → ((Ck+1 )⊗n )Sk+1
is an isomorphism. Let An be the stable value of this vector space (precisely, we could define
it as the direct limit). The dimension of AP
n is the Bell number Bn . These numbers do not
n
t
have a D-finite generating series: one has n≥0 Bn tn! = ee −1 .
This example can be made to look more like the previous ones by using the category
Rep(S) of algebraic representations of the infinite symmetric group S, as studied in [SS3].
Let E = C∞ be the permutation representation of S. Then the sequence (An )n≥0 is identified
with A(E)S . Thus the conclusion of Theorem 6.1 does not apply to A(E)S.
We thank a referee for suggesting this example.
7.5. Generalizing character polynomials. Let E be a vector space of dimension m, and
let M be a finitely generated module over A(E). Theorem 4.5 tells us that the enhanced
Hilbert series can be written in the form
e M (t) =
H
d
X
pi (t, T )erT0
r=0
where pi (t, T ) is a polynomial in t1 , t2 , . . . and T1 , T2 , . . .. In fact, using Corollary 3.5, we can
be more precise: if we set deg(Td ) = d and deg(td ) = 0, then each polynomial pi has degree
≤ i(m − i).
HILBERT SERIES FOR TWISTED COMMUTATIVE ALGEBRAS
27
We define the kth umbral substitution to be the linear map
Y
Y
Q[ti ] → Q[ai ],
tdi i 7→
k −di (ai )di ,
i
i
where (x)d = x(x − 1) · · · (x − d + 1). This extends to a linear map on polynomial expressions
in the ti and Ti , and we denote this operation by p 7→↓k p. When k = 1, this was denoted ↓
in [SS1, §5.2].
Proposition 7.16. Given M, there exist polynomials p1 , . . . , pd in t1 , t2 , . . . , T1 , T2 , . . . such
that deg(pi ) ≤ i(m − i), and
tr(cλ |M) =
when |λ| ≫ 0.
d
X
i=1
iℓ(λ) ↓i pi |tk =mk (λ)
Proof. Consider an expression t1α1 · · · tαr r T1β1 · · · Tsβs exp(iT0 ). The coefficient of tλ /λ! is
s Y
r
X
Y
λ!
vi,j
ℓ(λ)−|α|−|β|
i
(λ − α − β)! i=1 j=1 i
r
v ,...,v ∈Z
1
s
≥0
where the sum is over all (v1 , . . . , vs ) such that
Qr |vi | = βi , and we define |α| = α1 + · · · + αr ,
|β| = β1 + · · · + βs , and (λ − α − β)! = i=1 (mi (λ) − αi − v1,i − · · · − vs,i )! when the
arguments are nonzero, and otherwise, the term does not appear. This is the same as
e M (t) above if we
↓i (tα1 1 · · · tαr r T1β1 · · · Tsβs ), and hence the result follows from the form of H
only consider |λ| larger than deg(p0 (t, T )) where we define deg(ti ) = i.
When d = 1, the T1 , T2 , . . . do not show up, and ↓1 p1 is a polynomial. This is the
character polynomial discussed in [SS1, §5.2] and [CEF, Theorem 1.5]. In the general case,
we need to use a formal series.
References
[CEF]
[Fu]
[FH]
[FL]
[Ges]
[Lip]
[Mac]
[MS]
[SS1]
[SS2]
Thomas Church, Jordan S. Ellenberg, Benson Farb, FI-modules: a new approach to stability for
Sn -representations, Duke Math. J. 164, no. 9 (2015), 1833–1910, arXiv:1204.4533v4.
William Fulton, Intersection Theory, second edition, Springer-Verlag, Berlin, 1998.
William Fulton, Joe Harris, Representation Theory: A First Course, Graduate Texts in Mathematics, 129, Springer–Verlag, New York, 1991.
William Fulton, Serge Lang, Riemann-Roch Algebra, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences] 277, Springer-Verlag, New York,
1985.
Ira M. Gessel, Symmetric functions and P-recursiveness, J. Combin. Theory Ser. A 53 (1990), no. 2,
257–285.
L. Lipshitz, The diagonal of a D-finite power series is D-finite, J. Algebra 113 (1988), no. 2, 373–378.
I. G. Macdonald, Symmetric Functions and Hall Polynomials, second edition, Oxford Mathematical
Monographs, Oxford, 1995.
Ezra Miller, Bernd Sturmfels, Combinatorial Commutative Algebra, Graduate Texts in Mathematics
227, Springer-Verlag, New York, 2005.
Steven V Sam, Andrew Snowden, GL-equivariant modules over polynomial rings in infinitely many
variables, Trans. Amer. Math. Soc. 368 (2016), 1097–1158, arXiv:1206.2233v3.
Steven V Sam, Andrew Snowden, Introduction to twisted commutative algebras,
arXiv:1209.5122v1.
28
STEVEN V SAM AND ANDREW SNOWDEN
[SS3]
[SS4]
[SS5]
[Sno]
[Sta]
[WS]
Steven V Sam, Andrew Snowden, Stability patterns in representation theory, Forum Math. Sigma
3 (2015), e11, 108 pp., arXiv:1302.5859v2.
Steven V Sam, Andrew Snowden, Gröbner methods for representations of combinatorial categories,
J. Amer. Math. Soc. 30 (2017), 159–203, arXiv:1409.1670v3.
Steven V Sam, Andrew Snowden, GL-equivariant modules over polynomial rings in infinitely many
variables II, arXiv:1703.04516v1.
Andrew Snowden, Syzygies of Segre embeddings and ∆-modules, Duke Math. J. 162 (2013), no. 2,
225–277, arXiv:1006.5248v4.
Richard P. Stanley, Enumerative Combinatorics, Vol. 2, with a foreword by Gian-Carlo Rota and appendix 1 by Sergey Fomin, Cambridge Studies in Advanced Mathematics 62, Cambridge University
Press, Cambridge, 1999.
Christopher F. Woodcock, Habib Sharif, On the transcendence of certain series, J. Algebra 121
(1989), no. 2, 364–369.
Department of Mathematics, University of Wisconsin, Madison, WI
E-mail address: [email protected]
URL: http://math.wisc.edu/~svs/
Department of Mathematics, University of Michigan, Ann Arbor, MI
E-mail address: [email protected]
URL: http://www-personal.umich.edu/~asnowden/
| 0math.AC
|
Learning ReLUs via Gradient Descent
arXiv:1705.04591v2 [cs.LG] 1 Aug 2017
Mahdi Soltanolkotabi
Ming Hsieh Department of Electrical Engineering
University of Southern California, Los Angeles, CA, 90089
May 2017
Abstract
In this paper we study the problem of learning Rectified Linear Units (ReLUs) which are
functions of the form x ↦ max(0, ⟨w, x⟩) with w ∈ Rd denoting the weight vector. We study
this problem in the high-dimensional regime where the number of observations are fewer than
the dimension of the weight vector. We assume that the weight vector belongs to some closed
set (convex or nonconvex) which captures known side-information about its structure. We
focus on the realizable model where the inputs are chosen i.i.d. from a Gaussian distribution
and the labels are generated according to a planted weight vector. We show that projected
gradient descent, when initialization at 0, converges at a linear rate to the planted model with
a number of samples that is optimal up to numerical constants. Our results on the dynamics of
convergence of these very shallow neural nets may provide some insights towards understanding
the dynamics of deeper architectures.
1
Introduction
Nonlinear data-fitting problems are fundamental to many supervised learning tasks in signal processing and machine learning. Given training data consisting of n pairs of input features xi ∈ Rd and
desired outputs yi ∈ R we wish to infer a function that best explains the training data. In this paper
we focus on fitting Rectified Linear Units (ReLUs) to the data which are functions φw ∶ Rd → R of
the form
φw (x) = max (0, ⟨w, x⟩) .
A natural approach to fitting ReLUs to data is via minimizing the least-squares misfit aggregated
over the data. This optimization problem takes the form
min
w∈Rd
L(w) ∶=
1 n
2
∑ (max (0, ⟨w, xi ⟩) − yi )
n i=1
subject to R(w) ≤ R,
(1.1)
with R ∶ Rd → R denoting a regularization function that encodes prior information on the weight
vector.
Fitting nonlinear models such as ReLUs have a rich history in statistics and learning theory
[11] with interesting new developments emerging [6] (we shall discuss all these results in greater
detail in Section 4). Most recently, nonlinear data fitting problems in the form of neural networks
(a.k.a. deep learning) have emerged as powerful tools for automatically extracting interpretable and
1
actionable information from raw forms of data, leading to striking breakthroughs in a multitude of
applications [12, 13, 4]. In these and many other empirical domains it is common to use local search
heuristics such as gradient or stochastic gradient descent for nonlinear data fitting. These local
search heuristics are surprisingly effective on real or randomly generated data. However, despite
their empirical success the reasons for their effectiveness remains mysterious.
Focusing on fitting ReLUs, a-priori it is completely unclear why local search heuristics such
as gradient descent should converge for problems of the form (1.1), as not only the regularization
function maybe nonconvex but also the loss function! Efficient fitting of ReLUs in this highdimensional setting poses new challenges: When are the iterates able to escape local optima and
saddle points and converge to global optima? How many samples do we need? How does the number
of samples depend on the a-priori prior knowledge available about the weights? What regularizer is
best suited to utilizing a particular form of prior knowledge? How many passes (or iterations) of the
algorithm is required to get to an accurate solution? At the heart of answering these questions is the
ability to predict convergence behavior/rate of (non)convex constrained optimization algorithms.
In this paper we build up on a new framework developed by the author in [17] for analyzing
nonconvex optimization problems to address such challenges.
2
Precise measures for statistical resources
We wish to characterize the rates of convergence for the projected gradient updates (3.2) as a
function of the number of samples, the available prior knowledge and the choice of the regularizer.
To make these connections precise and quantitative we need a few definitions. Naturally the required
number of samples for reliable data fitting depends on how well the regularization function R can
capture the properties of the weight vector w. For example, if we know that the weight vector
is approximately sparse, naturally using an ℓ1 norm for the regularizer is superior to using an ℓ2
regularizer. To quantify this capability we first need a couple of standard definitions which we
adapt from [14, 15, 17].
Definition 2.1 (Descent set and cone) The set of descent of a function R at a point w ∗ is
defined as
DR (w ∗ ) = {h ∶ R(w ∗ + h) ≤ R(w ∗ )}.
The cone of descent is defined as a closed cone CR (w ∗ ) that contains the descent set, i.e. DR (w ∗ ) ⊂
CR (w ∗ ). The tangent cone is the conic hull of the descent set. That is, the smallest closed cone
CR (w ∗ ) obeying DR (w ∗ ) ⊂ CR (w ∗ ).
We note that the capability of the regularizer R in capturing the properties of the unknown weight
vector w ∗ depends on the size of the descent cone CR (w ∗ ). The smaller this cone is the more suited
the function R is at capturing the properties of w ∗ . To quantify the size of this set we shall use
the notion of mean width.
Definition 2.2 (Gaussian width) The Gaussian width of a set C ∈ Rd is defined as:
ω(C) ∶= Eg [sup ⟨g, z⟩],
z∈C
where the expectation is taken over g ∼ N (0, Ip ). Throughout we use B d /Sd−1 to denote the the
unit ball/sphere of Rd .
2
We now have all the definitions in place to quantify the capability of the function R in capturing the
properties of the unknown parameter w ∗. This naturally leads us to the definition of the minimum
required number of samples.
Definition 2.3 (minimal number of samples) Let CR (w ∗ ) be a cone of descent of R at w ∗.
We define the minimal sample function as
M(R, w ∗ ) = ω 2 (CR (w ∗ ) ∩ B d ).
We shall often use the short hand n0 = M(R, w ∗ ) with the dependence on R, w ∗ implied.
We note that n0 is exactly the minimum number of samples required for structured signal recovery
from linear measurements when using convex regularizers [3, 1]. Specifically, the optimization
problem
n
∗
∑ (yr − ⟨xi , w ⟩)
i=1
2
subject to R(w) ≤ R(w ∗ ),
(2.1)
succeeds at recovering an unknown weight vector w ∗ with high probability from n observations of
the form yi = ⟨ai , w ∗ ⟩ if and only if n ≥ n0 .1 While this result is only known to be true for convex
regularization functions we believe that n0 also characterizes the minimal number of samples even
for nonconvex regularizers in (2.1). See [14] for some results in the nonconvex case as well as the role
this quantity plays in the computational complexity of projected gradient schemes for linear inverse
problems. Given that with nonlinear samples we have less information (we loose some information
compared to linear observations) we can not hope to recover the weight vector from n ≤ n0 when
using (1.1). Therefore, we can use n0 as a lower-bound on the minimum number of observations
required for projected gradient descent iterations (3.2) to succeed at finding the right model.
3
Theoretical results for learning ReLUs
A simple heuristic for optimizing (1.1) is to use gradient descent. One challenging aspect of the
above loss function is that it is not differentiable and it is not clear how to run projected gradient
descent. However, this does not pose a fundamental challenge as the loss function is differentiable
except for isolated points and we can use the notion of generalized gradients to define the gradient
at a non-differentiable point as one of the limit points of the gradient in a local neighborhood of
the non-differentiable point. For the loss in (1.1) the generalized gradient takes the form
∇L(w) ∶=
2 n
∑ (ReLU (⟨w, xi ⟩) − yi ) (1 + sgn(⟨w, xi ⟩)) xi .
n i=1
(3.1)
Therefore, projected gradient descent takes the form
wτ +1 = PK (wτ − µτ ∇L(wτ )) ,
(3.2)
where µτ is the step size and K = {w ∈ Rd ∶ R(w) ≤ R} is the constraint set with PK denoting the
Euclidean projection onto this set.
1
We would like to note that n0 only approximately characterizes the minimum number of samples required. A more
√ Γ( t+1 ) √
precise characterization is φ−1 (ω 2 (CR (w∗ ) ∩ Bd )) ≈ ω 2 (CR (w∗ ) ∩ Bd ) where φ(t) = 2 Γ( 2t ) ≈ t. However, since
2
our results have unspecified constants we avoid this more accurate characterization.
3
Theorem 3.1 Let w ∗ ∈ Rd be an arbitrary weight vector and R ∶ Rd → R be a proper function
(convex or nonconvex). Suppose the feature vectors xi ∈ Rd are i.i.d. Gaussian random vectors
distributed as N (0, I) with the corresponding labels given by
yi = max (0, ⟨xi , w ∗ ⟩) .
To estimate w ∗, we start from the initial point w0 = 0 and apply the Projected Gradient (PGD)
updates of the form
wτ +1 = PK (wτ − µτ ∇L(wτ )) ,
(3.3)
with K ∶= {w ∈ Rd ∶ R(w) ≤ R(w ∗ )} and ∇L defined via (3.1). Also set the learning parameter
sequence µτ = 1 for all τ = 0, 1, 2, . . . and let n0 = M(R, w ∗ ), defined by 2.3, be our lower bound on
the number of measurements. Also assume
n > cn0 ,
(3.4)
holds for a fixed numerical constant c. Then there is an event of probability at least 1 − 9e−γn such
that on this event the updates (3.3) obey
1 τ
∥wτ − w ∗∥ℓ2 ≤ ( ) ∥w ∗∥ℓ2 .
2
(3.5)
Here γ is a fixed numerical constant.
The first interesting and perhaps surprising aspect of this result is its generality: it applies not
only to convex regularization functions but also nonconvex ones! As we mentioned earlier the
optimization problem in (1.1) is not known to be tractable even for convex regularizers. Despite
the nonconvexity of both the objective and regularizer, the theorem above shows that with a near
minimal number of data samples, projected gradient descent provably learns the original weight
vector w ∗ without getting trapped in any local optima.
Another interesting aspect of the above result is that the convergence rate is linear. Therefore,
to achieve a relative error of ǫ the total number of iterations is on the order of O(log(1/ǫ)). Thus
the overall computational complexity is on the order of O (nd log(1/ǫ)) (in general the cost is
the total number of iterations multiplied by the cost of applying the feature matrix X and its
transpose). As a result, the computational complexity is also now optimal in terms of dependence
on the matrix dimensions. Indeed, for a dense matrix even verifying that a good solution has been
achieved requires one matrix-vector multiplication which takes O(nd) time.
4
Discussions and prior art
There is a large body of work on learning nonlinear models. A particular class of such problems
that have been studied are the so called idealized Single Index Models (SIMs) [8, 9]. In these
problems the inputs are labeled examples {(xi , yi )}ni=1 ∈ Rd × R which are guaranteed to satisfy
yi = f (⟨w, xi ⟩) for some w ∈ Rd and nondecreasing (Lipchitz continuous) f ∶ R → R. The goal in
this problem is to find a (nearly) accurate such f and w. An interesting polynomial-time algorithm
called the Isotron exists for this problem [11, 10]. In principle, this approach can also be used to
fit ReLUs. However, these results differ from ours in term of both assumptions and results. On
4
the one had, the assumptions are slightly more restrictive as they require bounded features xi ,
outputs yi and weights. On the other hand, these result hold for much more general distributions
and more general models than the realizable model studied in this paper. These results also do
not apply in the high dimensional regime where the number of observations is significantly smaller
than the number of parameters (see [5] for some results in this direction). In the realizable case,
the Isotron result require O( 1ǫ ) iterations to achieve ǫ error in objective value. In comparison, our
results guarantee convergence to a solution with relative error ǫ (∥wτ − w ∗ ∥ℓ2 / ∥w ∗ ∥ℓ2 ≤ ǫ) after
log (1/ǫ) iterations. Focusing on the specific case of ReLU functions, an interesting recent result
[6] shows that reliable learning of ReLUs is possible under very general but bounded distributional
assumptions. To achieve an accuracy of ǫ the algorithm runs in poly(1/ǫ) time. In comparison, as
mentioned earlier our result rquires log(1/ǫ) iterations for reliable parameter estimation. We note
however we study the problem in different settings and a direct comparison is not possible between
the two results.
5
5.1
Proofs
Preliminaries
In this section we gather some useful results on concentration of stochastic processes which will be
crucial in our proofs. These results are mostly adapted from [17, 2, 16]. We begin with a lemma
which is a direct consequence of Gordon’s escape from the mesh lemma [7].
Lemma 5.1 Assume C ⊂ Rd is a cone and Sd−1 is the unit sphere of Rd . Also assume that
n ≥ max (20
ω 2 (C ∩ Sd−1 ) 1
,
− 1) ,
δ2
2δ
for a fixed numerical constant c. Then for all h ∈ C
1 n
∣ ∑(⟨xi , h⟩)2 − ∥h∥2ℓ2 ∣ ≤ δ ∥h∥2ℓ2 ,
n i=1
δ2
holds with probability at least 1 − 2e− 360 n .
We also need a generalization of the above lemma stated below.
Lemma 5.2 ([17]) Assume C ⊂ Rd is a cone (not necessarily convex) and Sd−1 is the unit sphere
of Rd . Also assume that
n ≥ max (80
ω 2 (C ∩ Sd−1 ) 2
, − 1) ,
δ2
δ
for a fixed numerical constant c. Then for all u, h ∈ C
1 n
∣ ∑⟨xi , u⟩⟨xi , h⟩ − u∗ h∣ ≤ δ ∥u∥ℓ2 ∥h∥ℓ2 ,
n i=1
δ2
holds with probability at least 1 − 6e− 1440 n .
5
We next state a generalization of Gordon’s escape through the mesh lemma also from [17].
Lemma 5.3 ([17]) Let s ∈ Rd be fixed vector with nonzero entries and construct the diagonal
matrix S = diag(s). Also, let X ∈ Rn×d have i.i.d. N (0, 1) entries. Furthermore, assume T ⊂ Rd
and define
bd (s) = E[∥Sg∥ℓ2 ],
where g ∈ Rd is distributed as N (0, In ). Define
σ(T ) ∶= max ∥v∥ℓ2 ,
then for all u ∈ T
v∈T
∣∥SAu∥ℓ2 − bd (s) ∥u∥ℓ2 ∣ ≤ ∥s∥ℓ∞ ω(T ) + η,
holds with probability at least
−
1 − 6e
η2
8∥s∥2 σ 2 (T )
ℓ∞
.
The previous lemma leads to the following Corollary.
Corollary 5.4 Let s ∈ Rd be fixed vector with nonzero entries and assume T ⊂ B d . Furthermore,
assume
Then for all u ∈ T ,
∥s∥2ℓ2 ≥ max (20 ∥s∥2ℓ∞
ω 2 (T ) 3
,
− 1) .
δ2
2δ
RRR n 2
RR
2
RRR ∑i=1 si (⟨xi , u⟩) − ∥u∥2 RRRR ≤ δ,
RRR
ℓ2 RR
RR
∥s∥2ℓ2
RR
R
δ2
holds with probability at least 1 − 6e− 1440 ∥s∥ℓ2 .
5.2
2
Convergence proof (Proof of Theorem 3.1)
In this section we shall prove Theorem 3.1. Throughout, we use the shorthand C to denote the
descent cone of R at w ∗, i.e. C = CR (w ∗ ). We begin by analyzing the first iteration. Using w0 = 0
we have
2 n
2 n
w1 ∶= PK (w0 − µ0 ∇L(w0 )) = PK ( ∑ yi xi ) = PK ( ∑ ReLU(⟨xi , w ∗ ⟩)xi ) .
n i=1
n i=1
We use the argument of [17][Page 25, inequality (7.34)] which shows that
2 n
∥w1 − w ∗ ∥ℓ2 ≤ 2 ⋅ sup uT ( ∑ ReLU(⟨xi , w ∗ ⟩)xi − w ∗) .
n i=1
u∈C∩B d
6
(5.1)
Using ReLU(z) =
z+∣z∣
2
we have
2 n
1 n
∗
∗
T 1
∗
T
∗
∑ ReLU(⟨xi , w ⟩)⟨xi , u⟩ − ⟨u, w ⟩ = u ( X X − I) w + ∑ ∣⟨xi , w ⟩∣ ⟨xi , u⟩.
n i=1
n
n i=1
(5.2)
We proceed by bounding the first term in the above equality. To this aim we decompose u in the
direction parallel/perpendicular to that of w ∗ and arrive at
⎛
w ∗ (w ∗ )T ⎞
(uT w ∗)
1
1
T
∗
∗ T 1
I
−
X
X
−
I)
w
+
⟨X
(w
)
(
uT ( X T X − I) w ∗ =
u, Xw ∗ ⟩,
n
n
n
⎝
∥w ∗∥2ℓ
∥w ∗∥2ℓ ⎠
2
2
2
⎞ ∥w ∗∥ℓ
⎛
⎛ ∥g∥ℓ2
w ∗ (w ∗)T ⎞
− 1 + √ 2 aT I −
∼(uT w ∗)
u,
n
⎠
⎝
⎝ n
∥w ∗∥2ℓ2 ⎠
RRR ∥w ∗∥
RRR ∥g∥2
w ∗ (w ∗ )T ⎞
ℓ2
ℓ2
RR
RR
T ⎛
∗
u,
− 1RRR + √
sup a I −
≤ ∥w ∥ℓ2 RRR
n u∈C∩Bd
RRR
⎝
RRR n
∥w ∗∥2ℓ2 ⎠
(5.3)
with g ∈ Rn and a ∈ Rd are independent random Gaussian random vectors distributed as N (0, Id )
and N (0, In ). By concentration of Chi-squared random variables
holds with probability at least 1 − 2e−n
∆2
8
RRR ∥g∥2
RRR
ℓ2
RRR
RR
RRR n − 1RRRR ≤ ∆,
RR
RR
(5.4)
. Also,
⎛
w ∗ (w ∗)T ⎞
1
1
√ aT I −
u ≤ √ (ω (C ∩ B d ) + η) ,
2
n ⎝
n
∥w ∗∥ℓ ⎠
(5.5)
2
holds with probability at least 1 − e−
(5.3), as long as
η2
2
. Plugging (5.4) with ∆ =
n≥
then
sup
u∈C∩B d
δ
6
and (5.5) with η =
δ√
6 n
into
36 2
ω (C ∩ B d ) ,
δ2
1
δ
uT ( X T X − I) w ∗ ≤ ∥w ∗ ∥ℓ2 ,
n
2
(5.6)
δ2
holds with probability at least 1 − 3e−n 288 .
We now focus on bounding the second term in (5.2). To this aim we decompose u in the
direction parallel/perpendicular to that of w ∗ and arrive at
RRR
RRR
n
1 n
∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩ 1 n
RR T ∗ 1
R
∗
∗
∣ ∑ ∣⟨xi , w ⟩∣ ⟨xi , u⟩∣ = RRR(u w ) ∑
+ ∑ ∣⟨xi , w ⟩∣ ⟨xi , u⊥ ⟩RRRR ,
2
∗
n i=1
n i=1
n i=1
RRR
RRR
∥w ∥ℓ2
RRR n
R
1
∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩ RRRR 1 n
≤ ∥w ∗∥ℓ2 RRRR ∑
RR + ∣ ∑ ∣⟨xi , w ∗ ⟩∣ ⟨xi , u⊥ ⟩∣ .
2
RRR n i=1
RRR n i=1
∗
∥w ∥ℓ2
R
R
7
(5.7)
with u⊥ = (I −
by
w ∗ (w ∗ )T
∥w ∗ ∥2ℓ
2
) u. Now note that
∥
∣⟨xi ,w ∗ ⟩∣⟨xi ,w ∗ ⟩
∥w ∗ ∥2ℓ
is sub-exponential with norm bounded
2
∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩
∥w ∗∥2ℓ2
∥
≤ c,
ψ1
with fixed numerical constant. Thus by Bernstein’s type inequality ([18][Proposition 5.16])
RR n
R
RRR 1
∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩ RRRR
RR ≤ t,
RR ∑
RRR n i=1
RRR
∗ ∥2
∥w
ℓ
2
R
R
holds with probability at least 1 − 2e−γn min(t
2
Furthermore,
(5.8)
,t)
with γ a fixed numerical constant. Also note that
¿
Á1 n
1 n
∗
À ∑ ∣⟨xi , w ∗ ⟩∣2 √1 ⟨g, u⊥ ⟩.
∑ ∣⟨xi , w ⟩∣ ⟨xi , u⊥ ⟩ ∼ Á
n i=1
n i=1
n
1 n
∗ 2
∗ 2
∑ ∣⟨xi , w ⟩∣ ≤ (1 + ∆) ∥w ∥ℓ2 ,
n i=1
holds with probability at least 1 − 2e−n
sup
u∈C∩Sd−1
∆2
8
and
(2ω (C ∩ Sd−1 ) + η)
1
√ ∣⟨g, u⊥ ⟩∣ ≤
√
,
n
n
holds with probability at least 1 − e−
η2
2
. Combining the last two inequalities we conclude that
√
(2ω (C ∩ Sd−1 ) + η)
1 n
√
∥w ∗∥ℓ2 ,
∣ ∑ ∣⟨xi , w ∗ ⟩∣ ⟨xi , u⊥ ⟩∣ ≤ 1 + ∆
n i=1
n
holds with probability at least 1 − 2e−n
δ √
η= √
n into (5.7)
∆2
8
− e−
η2
2
(5.9)
. Plugging (5.8) and (5.9) with t = 6δ , ∆ = 1, and
6 2
δ
1 n
∣ ∑ ∣⟨xi , w ∗ ⟩∣ ⟨xi , u⟩∣ ≤ ∥w ∗∥ℓ2 ,
n i=1
2
holds with probability at least 1 − 3e−γnδ − 2e− 8 as long as
2
n ≥ 288
n
ω 2 (C ∩ Sd−1 )
δ2
.
Thus pluggin (5.6) and (5.10) into (5.1) we conclude that for δ = 7/400
2 n
∥w1 − w ∗∥ℓ2 ≤2 ⋅ sup uT ( ∑ ReLU(⟨xi , w ∗ ⟩)xi − w ∗ ) ,
n i=1
u∈C∩B d
≤2δ ∥w ∗ ∥ℓ2 ,
7
∥w ∗ ∥ℓ2 ,
≤
200
8
(5.10)
holds with probability at least 1 − 8e−γn as long as
n ≥ cω 2 (C ∩ Sd−1 ) ,
for a fixed numerical constant c.
To introduce our general convergence analysis we begin by defining
E(ǫ) = {w ∈ Rd ∶ R(w) ≤ R(w ∗ ), ∥w − w ∗ ∥ℓ2 ≤ ǫ ∥w ∗∥ℓ2 } with
ǫ=
7
.
200
To prove Theorem 3.1 we use the argument of [17][Page 25, inequality (7.34)] which shows that if
we apply the projected gradient descent update
the error hτ = wτ − w ∗ obeys
wτ +1 = PK (wτ − ∇L(wτ )) ,
∥hτ +1 ∥ℓ2 = ∥wτ +1 − w ∗ ∥ℓ2 ≤ 2 ⋅ sup u∗ (hτ − ∇L(wτ )) .
(5.11)
u∈C∩B n
To complete the convergence analysis it is then sufficient to prove
sup u∗ (hτ − ∇L(wτ )) ≤
u∈C∩B n
1
1
∥hτ ∥ℓ2 = ∥wτ − w ∗ ∥ℓ2 .
4
4
(5.12)
We will instead prove that the following stronger result holds for all u ∈ C ∩ B n and w ∈ E(ǫ)
u∗ (w − w ∗ − ∇L(w)) ≤
1
∥w − w ∗∥ℓ2 .
4
(5.13)
The equation (5.13) above implies (5.12) which when combined with (5.11) proves the convergence
result of the Theorem (specifically equation (3.5)).
The rest of this section is dedicated to proving (5.13). To this aim note that ReLU(⟨xi , w⟩) =
⟨xi ,w⟩+∣⟨xi ,w⟩∣
. Therefore, the loss function can alternatively be written as
2
L(w) =
1 n
1 n
1 n
2
∗
∗ 2
∗
∗
∑ (∣⟨xi , w⟩∣ − ∣⟨xi , w ⟩∣) +
∑ (⟨xi , w − w ⟩) +
∑ (∣⟨xi , w⟩∣ − ∣⟨xi , w ⟩∣) (⟨xi , w − w ⟩) .
4n i=1
4n i=1
2n i=1
9
Thus
⟨∇L(w), u⟩ =
1 n
∗
∑ (⟨xi , w⟩ − sgn(⟨xi , w⟩) ∣⟨xi , w ⟩∣) ⟨xi , u⟩
2n i=1
+
+
1 n
1 n
∗
∗
∑⟨xi , w − w ⟩⟨xi , u⟩ +
∑ (∣⟨xi , w⟩∣ − ∣⟨xi , w ⟩∣) ⟨xi , u⟩
2n i=1
2n i=1
1 n
∗
∑ sgn(⟨xi , w⟩)⟨xi , w − w ⟩⟨xi , u⟩
2n i=1
1 n
1 n
= ∑⟨xi , w − w ∗⟩⟨xi , u⟩ + ∑ (∣⟨xi , w⟩∣ − ∣⟨xi , w ∗ ⟩∣) ⟨xi , u⟩
n i=1
n i=1
+
1 n
∗
∗
∗
∑ (1 − sgn(⟨xi , w ⟩)sgn(⟨xi , w⟩) + sgn(⟨xi , w ⟩) − sgn(⟨xi , w⟩)) ⟨xi , w ⟩⟨xi , u⟩
2n i=1
1 n
1 n
= ∑⟨xi , w − w ∗⟩⟨xi , u⟩ + ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , w − w ∗ ⟩⟨xi , u⟩
n i=1
n i=1
+
+
1 n
∗
∗
∑ (sgn(⟨xi , w⟩) − sgn(⟨xi , w ⟩)) ⟨xi , w − w ⟩⟨xi , u⟩
n i=1
1 n
∗
∗
∗
∑ (1 − sgn(⟨xi , w ⟩)sgn(⟨xi , w⟩) − sgn(⟨xi , w ⟩) + sgn(⟨xi , w⟩)) ⟨xi , w ⟩⟨xi , u⟩
2n i=1
1 n
1 n
= ∑⟨xi , w − w ∗⟩⟨xi , u⟩ + ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , w − w ∗ ⟩⟨xi , u⟩
n i=1
n i=1
+
+
1 n
∗
∗
∑ (sgn(⟨xi , w⟩) − sgn(⟨xi , w ⟩)) ⟨xi , w − w ⟩⟨xi , u⟩
n i=1
1 n
∗
∗
∗
∑ (1 − sgn(⟨xi , w ⟩)) (sgn(⟨xi , w ⟩) − sgn(⟨xi , w⟩)) ∣⟨xi , w ⟩∣ ⟨xi , u⟩
2n i=1
Now defining h = w − w ∗ we conclude that
⟨u, w − w ∗ − ∇L(w)⟩ =⟨u, h − ∇L(w)⟩,
=uT (I −
+
+
1 n
1
XX T ) h − ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⟩⟨xi , u⟩,
n
n i=1
1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⟩⟨xi , u⟩,
n i=1
1 n
∗
∗
∗
∑ sgn(⟨xi , w⟩) (1 − sgn(⟨xi , w ⟩)) (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) ∣⟨xi , w ⟩∣ ⟨xi , u⟩.
2n i=1
10
Now define h⊥ = h −
(hT w ∗ ) ∗
w .
∥w ∗ ∥2ℓ
Using this we can rewrite the previous expression in the form
2
⟨u, w − w ∗ − ∇L(w)⟩ =uT (I −
+
+
1 n
1
XX T ) h − ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⟩⟨xi , u⟩,
n
n i=1
1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩,
n i=1
⟨h, w ∗ ⟩ 1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) ∣⟨xi , w ⟩∣ ⟨xi , u⟩,
∥w ∗ ∥2ℓ n i=1
2
1 n
∗
∗
∗
+
∑ sgn(⟨xi , w⟩) (1 − sgn(⟨xi , w ⟩)) (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) ∣⟨xi , w ⟩∣ ⟨xi , u⟩
2n i=1
=uT (I −
+
+
1 n
1
XX T ) h − ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⟩⟨xi , u⟩,
n
n i=1
1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩,
n i=1
1 n sgn(⟨xi , w⟩)
⟨h, w ∗ ⟩
(1 − sgn(⟨xi , w ∗ ⟩)) +
]
∑[
n i=1
2
∥w ∗ ∥2ℓ
2
(1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ∗ ⟩)) ∣⟨xi , w ∗ ⟩∣ ⟨xi , u⟩
(5.14)
We now proceed by bounding each of the four terms in (5.14) and then combine them in Section
5.2.5.
5.2.1
Bounding the first term in (5.14)
To bound the first term we use Lemma 5.2, which implies that as long as
then for all u ∈ C ∩ B n and h ∈ E(ǫ)
n ≥ max (80
u∗ (I −
δ2
n0 2
, − 1) ,
δ2 δ
1 ∗
X X) h ≤ δ ∥h∥ℓ2 ,
n
(5.15)
holds with probability at least 1 − 6e− 1440 n .
5.2.2
Bounding the second term in (5.14)
To bound the second term we first define
u⊥ = u −
(uT w ∗)
∥w ∗∥2ℓ2
w∗
and h⊥ = h −
11
(hT w ∗)
∥w ∗ ∥2ℓ2
w∗.
Now note that
1 n
(uT w ∗)(hT w ∗) 1 n
∗
∗
− ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⟩⟨xi , u⟩ = −
∑ ∣⟨xi , w ⟩∣ ⟨xi , w ⟩
4
∗
n i=1
n
∥w ∥ℓ2
i=1
−
(uT w ∗) 1 n
∗
∑ ∣⟨xi , w ⟩∣ ⟨xi , h⊥ ⟩
∥w ∗∥2ℓ n i=1
2
(h w ) 1 n
∗
∑ ∣⟨xi , w ⟩∣ ⟨xi , u⊥ ⟩
∥w ∗ ∥2ℓ n i=1
∗
T
−
2
1 n
− ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩,
n i=1
1 n ∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩
≤ ∥h∥ℓ2 ∣ ∑
∣
n i=1 ∥w ∗∥ℓ2 ∥w ∗∥ℓ2
1 n ∣⟨xi , w ∗ ⟩∣
1 n ∣⟨xi , w ∗ ⟩∣
⟨x
,
h
⟩∣
+
∥h∥
∣
⟨xi , u⊥ ⟩∣
+∣ ∑
∑
i
⊥
ℓ
2
n i=1 ∥w ∗ ∥ℓ2
n i=1 ∥w ∗ ∥ℓ2
1 n
+ ∣ ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩∣ .
n i=1
(5.16)
We now proceed by bound the four terms on the right-hand side of the above inequality. To bound
the first term we use (5.8) to conclude that
R
RRR n
∣⟨xi , w ∗ ⟩∣ ⟨xi , w ∗ ⟩ RRRR
RR 1
RRR ≤ t,
RRR ∑
RRR
RRR n i=1
∥w ∗∥2ℓ2
holds with probability at least 1 − 2e−γn min(t
use (5.9) to conclude that
2
,t)
. To bound the second and third terms in (5.16) we
√
√
√
(2 n0 + η n)
1 n ∣⟨xi , w ∗ ⟩∣
∣ ∑
⟨x
,
h
⟩∣
≤
1
+
∆
√
∥h∥ℓ2 ,
i
⊥
n i=1 ∥w ∗∥ℓ2
n
√
√
√
(2 n0 + η n)
1 n ∣⟨xi , w ∗ ⟩∣
√
⟨xi , u⊥ ⟩∣ ≤ 1 + ∆
,
∣ ∑
n i=1 ∥w ∗∥ℓ2
n
holds with probability at least 1 − 2e−n
variables independent from xi . Then,
Define T = (I −
∆2
8
− e−
(5.17)
η2
n
2
(5.18)
. To bound the last term let ǫi be i.i.d. ±1 random
1 n
1 n
∗
∑ sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩ ∼ ∑ ǫi ⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩.
n i=1
n i=1
w ∗ (w ∗ )T
∥w ∗ ∥2ℓ
2
) C ∩ Sd−1 and note that
1 n
1 n
∣ ∑ ǫi ⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩∣ ≤ ( sup ∣ ∑ ǫi ⟨xi , u⟩⟨xi , v⟩∣) ∥h∥ℓ2 .
n i=1
u,v∈T n i=1
12
(5.19)
Now to bound the term in the parenthesis note that by concentration of sums of sub-Gaussian
random variables [18, Proposition 5.10]
¿
n
∆Á
1 n
À 1 ∑(⟨xi , u⟩)2 (⟨xi , v⟩)2 ,
(5.20)
∣ ∑ ǫi ⟨xi , u⟩⟨xi , v⟩∣ ≤ Á
n i=1
n
n i=1
holds with probability at least 1 − 2e−γn∆ . Now note that
√
√
sup ∣⟨xi , u⟩∣ ≤ 2ω(T ) + η ≤ 2 n0 + η n,
2
u∈T
holds with probability at least 1 − e−
sup
η2
n
2
. Thus using the union bound
√
√
√
sup ∣⟨xi , u⟩∣ ≤ 2ω(T ) + η n ≤ 2 n0 + η n,
i=1,2,...,n u∈T
holds with probability at least 1 − ne−
S = I, we conclude that
η2
n
2
. Plugging this into (5.20) and using Lemma 5.3 with
¿
n
√ Á
∆ √
1 n
Àsup 1 ∑(xT v)2 ,
sup ∣ ∑ ǫi ⟨xi , u⟩⟨xi , v⟩∣ ≤ (2 n0 + η n) Á
i
n
u,v∈T n i=1
v∈T n i=1
√
√
√
∆ √
≤ (2 n0 + η n) ( n0 + (η + 1) n) ,
n
η2
η2
holds with probability at least 1 − 2e−γn∆ − ne− 2 n − 6e− 8 n . Using the latter inequality in (5.19)
we conclude that for all h ∈ C and u ∈ C ∩ Sn−1
√
√
√
∆ √
1 n
(5.21)
∣ ∑ ǫi ⟨xi , h⊥ ⟩⟨xi , u⊥ ⟩∣ ≤ (2 n0 + η n) ( n0 + (η + 1) n) ∥h∥ℓ2
n i=1
n
2
η2
η2
holds with probability at least 1 − 2e−γn∆ − ne− 2 n − 6e− 8 n , completing the bound of the last term
of (5.16). Combining (5.17), (5.18), and (5.21) with η = ∆ = 1, we conclude that
2
1 n
− ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⟩⟨xi , u⟩ ≤ δ ∥h∥ℓ2 ,
n i=1
(5.22)
− (n + 10)e−γn as long as
n0
n≥c 2.
δ
This completes the bound on the second term in (5.14).
holds with probability at least 1 − 2e−γδ
5.2.3
2
n
Bounding the third term in (5.14)
To bound the third term note that
1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩))sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩
n i=1
2 n
= ∑ sgn(⟨xi , w ∗ ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩1{⟨xi ,w⟩⟨xi ,w∗ ⟩≤0} ,
n i=1
¿
¿
Á1 n
(a) Á 1 n
2
Á
À ∑ ∣⟨xi , u⟩∣2 , (5.23)
À ∑ ∣⟨xi , h⊥ ⟩∣ 1
≤ 2Á
{⟨xi ,w⟩⟨xi ,w ∗ ⟩≤0}
n i=1
n i=1
13
where the last inequality follows from Cauchy Schwarz. Note that by Lemma 5.1 as long as
n ≥ max (80 nδ20 , 2δ − 1), then
1 n
2
∑ ∣⟨xi , u⟩∣ ≤ 1 + δ,
n i=1
holds for all u ∈ C ∩ Sd−1 with probability at least 1 − 2e−γδ n . Combining the latter inequality with
(5.23) we conclude that with high probability
2
1 n
∗
∗
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩
n i=1
¿
√
Á1 n
À ∑ ∣⟨xi , h⊥ ⟩∣2 1
≤ 2 1 + δÁ
{⟨xi ,w⟩⟨xi ,w ∗ ⟩≤0} .
n i=1
(5.24)
We now turn our attention to bounding the right-hand side of (5.24). To this aim first note that
as long as ∥h∥ℓ2 ≤ ǫ ∥w ∗∥ℓ2 we have (w T w ∗ ) ≥ (1 − ǫ) ∥w ∗ ∥2ℓ2 . Thus, we have the following chain of
inequalities
1 n
1 n
2
2
∑ ∣⟨xi , h⊥ ⟩∣ 1{⟨xi ,w⟩⟨xi ,w∗⟩≤0} = ∑ ∣⟨xi , h⊥ ⟩∣ 1
⎛ T ∗
n i=1
n i=1
{ (w w ) (xT w∗ )+xT h
⎝
∥w∗ ∥2
ℓ2
⊥
i
i
⎞ T ∗
(x w )≤0}
⎠ i
1 n
= ∑ ∣⟨xi , h⊥ ⟩∣2 1
T ∗
n i=1
{ (w w ) (xT w∗ )2 ≤−(xT h
∥w∗ ∥2
ℓ2
i
1 n
≤ ∑ ∣⟨xi , h⊥ ⟩∣2 1
n i=1
{∣xT w∗ ∣≤∥w∗ ∥2
∣xT h⊥ ∣
i
ℓ2 (wT w∗ )
i
1 n
≤ ∑ ∣⟨xi , h⊥ ⟩∣2 1
.
n i=1
{(1−ǫ)∣xTi w∗ ∣≤∣xTi h⊥ ∣}
}
T ∗
⊥ )(xi w )
i
}
(5.25)
We now proceed by using the following result from [17].
Lemma 5.5 [17] Let C ∈ Rd be a closed cone. Also let w ∗ be a fixed vector in Rd . Furthermore,
assume
n ≥ c ⋅ ω 2 (C ∩ Sd−1 ),
with c a fixed numerical constant. Then
⎛
1 n
2
≤ δ+
∑ ∣⟨xi , h⊥ ⟩∣ 1
n i=1
{(1−ǫ)∣xTi w∗ ∣≤∣xTi h⊥ ∣} ⎝
holds with probability at least 1 − 2e−γδ
2
n
√
21 ⎞
ǫ ∥h∥2ℓ2 ,
20 ⎠
for all vectors h ∈ C obeying
∥h∥ℓ2 ≤ ǫ ∥w ∗ ∥ℓ2 .
14
2
Proof This lemma follows from the argument on pages 27-30 of [17].
Combining the lemma with equations (5.23), (5.24), and (5.25) we conclude that as long as n ≥ cn0 ,
then
√
√
⎛
21 ⎞
1 n
∗
∗
ǫ ∥h∥ℓ2 ,
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩ ≤ 2 1 + δ δ +
n i=1
20 ⎠
⎝
(5.26)
holds for all u ∈ C ∩ Sd−1 and h ∈ E(ǫ) with probability at least 1 − 4e−γnδ . This completes the
bound on the third term of (5.14).
2
5.2.4
Bounding the fourth term in (5.14)
To bound the fourth term of (5.14) note that by the argument leading to (5.23)
1 n sgn(⟨xi , w⟩)
⟨h, w ∗ ⟩
(1 − sgn(⟨xi , w ∗ ⟩)) +
] (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ∗ ⟩)) ∣⟨xi , w ∗ ⟩∣ ⟨xi , u⟩,
∑[
n i=1
2
∥w ∗∥2ℓ2
¿
R2
Á n RRR
√
sgn(⟨xi , w⟩)
⟨h, w ∗ ⟩ RRRR
Á1
À ∑ RRRR
R ∣⟨xi , w ∗ ⟩∣2 1{⟨xi ,w⟩⟨xi ,w∗ ⟩≤0} ,
(1 − sgn(⟨xi , w ∗ ⟩)) +
≤ 2 1 + δÁ
2 RR
∗
n i=1 RRR
2
∥w ∥ℓ2 RRR
R
¿
√
Á1 n
À ∑ ∣⟨xi , w ∗ ⟩∣2 1{⟨x ,w⟩⟨x ,w∗ ⟩≤0} ,
≤ 4 1 + δÁ
(5.27)
i
i
n i=1
holds for all u ∈ C ∩ Sd−1 with probability at least 1 − 2e−γδ n . We again proceed by using the
following result from [17].
2
Lemma 5.6 [17] Let C ∈ Rd be a closed cone. Also let w ∗ be a fixed vector in Rd . Furthermore,
assume
n ≥ c ⋅ ω 2 (C ∩ Sd−1 ),
with c a fixed numerical constant. Then
⎛
1
1 n
∗ 2
δ+
∑ ∣⟨xi , w ⟩∣ 1{⟨xi ,w⟩⟨xi ,w∗⟩≤0} ≤
2
n i=1
(1 − ǫ) ⎝
holds with probability at least 1 − 2e−γδ
2
n
√
21 ⎞
ǫ ∥h∥2ℓ2 ,
20 ⎠
2
for all vectors h ∈ C obeying
∥h∥ℓ2 ≤ ǫ ∥w ∗ ∥ℓ2 .
Proof This lemma follows from the argument on pages 27-30 of [17].
Combining the lemma with (5.27) we conclude that as long as n ≥ cn0 , then
√
√
1 n
21 ⎞
4 1+δ ⎛
∗
∗
ǫ ∥h∥ℓ2 ,
δ+
∑ (1 − sgn(⟨xi , w⟩)sgn(⟨xi , w ⟩)) sgn(⟨xi , w ⟩)⟨xi , h⊥ ⟩⟨xi , u⟩ ≤
2
n i=1
(1 − ǫ) ⎝
20 ⎠
(5.28)
holds for all u ∈ C ∩ Sd−1 and h ∈ E(ǫ) with probability at least 1 − 4e−γnδ . This completes the
bound on the fourth term of (5.14).
2
15
5.2.5
Putting the bounds together
In this Section we put together the bounds of the previous sections. Combining (5.15), (5.22),
(5.26), and (5.28) we conclude that
√
⎛
⎛ √
21 ⎞⎞
2
∗
) δ+
ǫ ∥w − w ∗ ∥ℓ2 ,
⟨u, w − w − ∇L(w)⟩ ≤ 2 δ + 1 + δ (1 +
2
(1 − ǫ) ⎝
20 ⎠⎠
⎝
holds for all u ∈ C ∩ Sd−1 and w ∈ E(ǫ) with probability at least 1 − 16e−γδ
this inequality with δ = 10−4 and ǫ = 7/200 we conclude that
⟨u, w − w ∗ − ∇L(w)⟩ ≤
1
∥w − w ∗∥ℓ2 ,
4
2
n
− (n + 10)e−γn . Using
holds for all u ∈ C ∩ Sd−1 and w ∈ E(ǫ) with high probability.
Acknowledgements
M.S. would like to thank Adam Klivans and Matus Telgarsky for discussions related to [6] and the
Isotron algorithm. This work was done in part while the author was visiting the Simons Institute
for the Theory of Computing.
References
[1] D. Amelunxen, M. Lotz, M. B. McCoy, and J. A. Tropp. Living on the edge: Phase transitions
in convex programs with random data. Information and Inference, 2014.
[2] E. J. Candes, X. Li, and M. Soltanolkotabi. Phase retrieval via Wirtinger flow: Theory and
algorithms. IEEE Transactions on Information Theory, 61(4):1985–2007, 2015.
[3] V. Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. Willsky. The convex geometry of linear
inverse problems. Foundations of Computational Mathematics, 12(6):805–849, 2012.
[4] R. Collobert and J. Weston. A unified architecture for natural language processing: Deep
neural networks with multitask learning. In Proceedings of the 25th international conference
on Machine learning, pages 160–167. ACM, 2008.
[5] R. Ganti, N. Rao, R. M. Willett, and R. Nowak. Learning single index models in high dimensions. arXiv preprint arXiv:1506.08910, 2015.
[6] S. Goel, V. Kanade, A. Klivans, and J. Thaler. Reliably learning the ReLU in polynomial
time. arXiv preprint arXiv:1611.10258, 2016.
[7] Y. Gordon. On Milman’s inequality and random subspaces which escape through a mesh in Rn .
Springer, 1988.
[8] J. L. Horowitz and W. Hardle. Direct semiparametric estimation of single-index models with
discrete covariates. Journal of the American Statistical Association, 91(436):1632–1640, 1996.
16
[9] H. Ichimura. Semiparametric least squares (SLS) and weighted SLS estimation of single-index
models. Journal of Econometrics, 58(1-2):71–120, 1993.
[10] S. M. Kakade, V. Kanade, O. Shamir, and A. Kalai. Efficient learning of generalized linear and
single index models with isotonic regression. In Advances in Neural Information Processing
Systems, pages 927–935, 2011.
[11] A. T. Kalai and R. Sastry. The isotron algorithm: High-dimensional isotonic regression. In
COLT, 2009.
[12] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
[13] A. Mohamed, G. E. Dahl, and G. Hinton. Acoustic modeling using deep belief networks. IEEE
Transactions on Audio, Speech, and Language Processing, 20(1):14–22, 2012.
[14] S. Oymak, B. Recht, and M. Soltanolkotabi. Sharp time–data tradeoffs for linear inverse
problems. arXiv preprint arXiv:1507.04793, 2015.
[15] S. Oymak and M. Soltanolkotabi. Fast and reliable parameter estimation from nonlinear
observations. arXiv preprint arXiv:1610.07108, 2016.
[16] M. Soltanolkotabi. Algorithms and Theory for Clustering and Nonconvex Quadratic Programming. PhD thesis, Stanford University, 2014.
[17] M. Soltanolkotabi. Structured signal recovery from quadratic measurements: Breaking sample
complexity barriers via nonconvex optimization. arXiv preprint arXiv:1702.06175, 2017.
[18] R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv preprint
arXiv:1011.3027, 2010.
17
| 7cs.IT
|
Syntax and Semantics of Linear Dependent Types
Technical Report
Matthijs Vákár
Oxford, UK, January 19, 2015
arXiv:1405.0033v4 [cs.LO] 16 Jan 2015
Abstract
A type theory is presented that combines (intuitionistic) linear types with type dependency,
thus properly generalising both intuitionistic dependent type theory and full linear logic. A syntax
and complete categorical semantics are developed, the latter in terms of (strict) indexed symmetric
monoidal categories with comprehension. Various optional type formers are treated in a modular
way. In particular, we will see that the historically much-debated multiplicative quantifiers and
identity types arise naturally from categorical considerations. These new multiplicative connectives
are further characterised by several identities relating them to the usual connectives from dependent
type theory and linear logic. Finally, one important class of models, given by families with values in
some symmetric monoidal category, is investigated in detail.
1
Disclaimer and Acknowledgements
The concept of a syntax with linear dependent types is not new. Such a calculus was first considered in
[1], in the context of a linear extension of the Logical Framework (LF). In terms of the semantics, the
author would like to point to [2] and [3], which study very similar semantic semantic objects, although
without the notion of comprehension. Finally, the author would like to thank Urs Schreiber for sparking
his interest in the topic through many enthusiastic posts on the nLab and nForum.
The contribution of the present work is as follows.
1. The presentation of a syntax in a style that is very close both to the dual intuitionistic linear logic
of [4] and to that of intuitionistic dependent type theory in [5]. This clarifies, from a syntactic
point of view, how exactly linear dependent types fit in with the work in both traditions.
2. Adding various structural rules, that allow us to consider the more basic setting without the
various type formers of the Linear Logical Framework (LLF). In the rich setting of LLF, these
become admissible.
3. The addition of various type formers to the syntax (most notably, Σ-, !-, and Id-types).
4. The development of the first categorical semantics for linear dependent types - although there have
been some suggestions about this on the nLab, by Mike Shulman and Urs Schreiber, in particular,
as far as the author is aware, no account of this has been published. The semantics is developed
in a style that combines features of the linear-non-linear adjunctions of [4] and the comprehension
categories of [6]. Among other things, it shows that multiplicative quantifiers arise very naturally,
as adjoints to substitution, thereby offering a new point of view on long-debated issue of quantifiers
in linear logic.
2
Contents
0 Preliminaries
0.1 (Non-Linear) Intuitionistic Dependent Type Theory (DTT) . . . . . . . . . . . . . . . . . .
0.2 Intuitionistic Linear (Non-Dependent) Type Theory (ILTT) . . . . . . . . . . . . . . . . . .
5
5
6
1 Intuitionistic Linear Dependent Type Theory?
7
2 Syntax of ILDTT
8
3 Semantics of ILDTT
3.1 Tautological models of ILDTT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Categorical Semantics of ILDTT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
16
17
4 Some Discrete Models: Monoidal Families
31
5 Conclusions and Future Work
34
3
Introduction
To put this work in context, the best point of departure may be Church’s simply typed λ-calculus (or
intuitionistic propositional type theory), [7], which according to the Curry-Howard correspondence (e.g.
[8]) can be thought of as a proof calculus for intuitionistic propositional logic. To this, Lambek (e.g. [9])
added the new idea that it can also be viewed as a syntax for describing Cartesian closed categories.
From this starting point, two traditions depart that are particularly relevant for us.
On the one hand, so-called dependent type theory, can be seen to extend along the Curry-Howard
correspondence, to provide a proof calculus for intuitionistic predicate logic. See e.g. [10]. Various
flavours of categorical semantics for this calculus have been given, but they almost all boil down to the
same essential ingredients of talking about certain fibred or, equivalently, indexed Cartesian categories,
usually (depending on the exact formulation of the type theory) with a notion of comprehension that
relates the fibres to the base category.
On the other hand, there is a school of logic, initiated by Girard, so-called linear logic, that has
pursued a further, resource sensitive, analysis of the intuitionistic propositional type theory by exposing
precisely how many times each assumption is used in proofs. See e.g. [11]. Later, the essence of Girard’s
system was seen to be captured by a less restrictive system called (multiplicative) intuitionistic linear
logic (where Girard’s original system is now referred to as classical linear logic). This will be the sense
in which we use the word ’linear’. A satisfactory proof calculus and categorical semantics, in terms of
symmetric monoidal closed categories, are easily given. The final essential element of linear type theories,
the resource modality providing the connection with intuitionistic type theories, can be given categorical
semantics as a monoidal adjunction between an intuitionistic and a linear model. [4]
The question rises if this linear analysis can be extended to predicate logic. Although some work has
been done in this direction, the author feels a satisfactory answer has not been given.
Although Girard’s early work in linear logic already talks about quantifiers, the analysis appears to
have stayed rather superficial. In particular, an account of internal quantification, or a linear variant
of Martin-Löf’s type theory, was missing, let alone a Curry-Howard correspondence. Later, linear types
and dependent types were first combined in [1], where a syntax was presented that extends LF (Logical
Framework) with linear types (that depend on terms of intuitionistic types). This has given rise to a line
of work in the computer science community. See e.g. [12, 13, 14]. All the work seems to be syntactic in
nature, however.
On the other hand, very similar ideas, this time at the level of categorical semantics and specific models
(coming from homotopy theory, algebra, and mathematical physics), have emerged in the mathematical
community lately, perhaps independently so. Relevant literature is e.g. [15, 2, 3, 16].
Although, in the past months, some suggestions have been made on the nLab and nForum, of possible
connections between both lines of work, no account of the correspondence was ever published, as far as
the author is aware. Moreover, the syntactic tradition seems to have stayed restricted to the situation
of functional type theory, in which one has Π- and ⊸-types, while we are really interested in the general
case of algebraic type theory, which is of fundamental importance from the point of view of mathematics
where structures seldom admit internal homs. Simultaneously, in the syntactic traditions there seems
to be a lack of other type formers (like Σ-, !-, and Id-types). The semantic tradition, however, does not
seem to have given a sufficient account of the notion of comprehension. The present text will take some
steps to close this gap in the existing literature.
The point of this paper is to illustrate how linear and dependent types can be combined straightforwardly and in great generality. Firstly, in section 1, we start with some general discussion on the matter
of combining linear and dependent types. Secondly, in section 2, we present a syntax, intuitionistic linear
dependent type theory (ILDTT), a natural blend of the dual intuitionistic linear logic (DILL) of [4] and
dependent type theory (DTT) of e.g. [5]. Thirdly, in section 3, we present a complete categorical semantics that is the obvious combination of linear/non-linear adjunctions of e.g. [4] and indexed Cartesian
categories with comprehension of e.g. [6]. Finally, in section 4, an important class of (rather discrete)
models is investigated, in terms of families with values in a fixed symmetric monoidal category.
This paper forms the first report of a research programme that aims to explore the interplay between
type dependency and linearity. We we will therefore end with a brief discussion of some our work in
progress and planned work on related topics.
4
0
Preliminaries
We will start these notes with some suggestions of references for both linear type theory and dependent
type theory, as having the right point of view on both subjects greatly simplifies understanding our
presentation of the new material.
0.1
(Non-Linear) Intuitionistic Dependent Type Theory (DTT)
Syntax
For an introductory but thorough treatment of the syntax, we refer the reader to [5].
Categorical Semantics
There are many (more or less) equivalent kinds of categorical model of a dependent type theory. Perhaps the best definition the have in mind while reading this text is either the notion of a full (split)
comprehension category of [6] or the following rephrasing of a full split comprehension category with 1and ×-types1 .
Definition (Strict Indexed Cartesian Monoidal Category with Comprehension). By a strict2 indexed3
Cartesian monoidal category with comprehension, we will mean the following data.
1. A category of intuitionistic contexts C with a terminal object ⋅.
2. A strict indexed Cartesian monoidal category I over C, i.e. a contravariant functor I into the
category CMCat of (small) Cartesian monoidal categories and strong cartesian monoidal functors
I
C op Ð→ CMCat.
We will also write −{f } ∶= I(f ) for the action of I on a morphism f of C.
3. A comprehension schema, i.e. for each ∆ ∈ ob(C) and A ∈ ob(I(∆)) a representation for the functor
x ↦ I(dom(x))(1, A{x}) ∶ (C/∆)op Ð→ Set,
p∆,A
which induces a fully faithful comprehension functor : if we write the representing object ∆.A Ð→
∆ ∈ ob(C/∆), write a ↦ ⟨f, a⟩ for the isomorphism I(∆′ )(1, A{f }) ≅ C/∆(f, p∆,A ), and write
v∆,A ∈ I(∆.A)(1, A{p∆,A }) for the universal element, the comprehension functor I Ð→ C/− is
defined as
✲ C/∆
I(∆)
f
A Ð→ B
✲
p∆,A
⟨p∆,A , I(p∆,A )(f ) ○ v∆,A ⟩
✲ p∆,B .
Note that the comprehension schema says that we build up the morphisms into objects that arise as
lists of types in our category of contexts C as lists of closed terms. It will be the crucial aspect that makes
sure we are in the world of internal quantification. The fact that we are demanding the comprehension
functor to be fully faithful means that also the non-closed terms in I(∆)(A, B) correspond precisely
with the closed terms I(∆.A)(I, B). The equivalent condition on a comprehension category is the
notion of fullness of [6]. This is essential to get a precise fit with the syntax. However, we shall drop this
restriction in modelling linear dependent types, as, there, having A as a linear assumption (which will
be an assumption in the fibre) can really be different from having A as an intuitionistic assumption (an
assumption in the base).
1
This restriction avoids us having to talk about multicategories. We are sure the reader familiar with multicategories
can see the obvious generalisation to the case without 1- and ×-types.
2
The strictness, here, refers to the fact that I is a functor, rather than a pseudofunctor. This reflects the fact that term
substitution in types is a strict operation in type theory, even though the non-strict version might seem more natural from
the point of view of category theory. It should be noted, though, that every indexed category is equivalent to a strict one.
[17]
3
Of course, using the Grothendieck construction, we could, equivalently, use split fibred categories, here, as is in fact
done in certain accounts of semantics for DTT. However, indexed categories seem closer to the syntax and are technically
less involved in the strict case.
5
0.2
Intuitionistic Linear (Non-Dependent) Type Theory (ILTT)
The most suitable flavour of ILTT to have in mind while reading these notes is the so-called dual
intuitionistic linear logic (DILL) of [4]. This will be our principal reference for both syntax and semantics.
Both the syntax and semantics given in this reference are very close in intuition to our syntax and
semantics of linear dependent types. For the semantics, more background is provided in [18].
6
1
Intuitionistic Linear Dependent Type Theory?
Although it is a priori not entirely clear what linear dependent type theory should be, one can easily
come up with some guiding criteria. My gut reaction was the following.
What we want to arrive at is a version of dependent type theory without weakening and contraction
rules. Moreover, we would want an exponential co-modality on the type theory that gives us back the
missing structural rules.
When one first tries to write down such a type theory, however, one will run into the following
discrepancy.
• The lack of weakening and contraction rules in linear type theory forces us to refer to each declared
variable precisely once: for a sequent x ∶ A ⊢ t ∶ B, we know that x has a unique occurence in t.
• In dependent type theory, types can have free (term) variables: x ∶ A ⊢ B type, where x is a free
variable in B. Crucially, we can then talk about terms of B: x ∶ A ⊢ b ∶ B, where generally x may
also be free in b. For almost all interesting applications we will need multiple occurences of x to
construct b ∶ B, at least one for B and one for b.
The question now is what it means to refer to a declared variable only once.
Do we not count occurences in types? This point of view seems incompatible with universes, however, which play an important role in dependent type theory. If we do, however, the language seems to
lose much of its expressive power. In particular, it prevents us from talking about constant types, it seems.
In this paper, we will circumvent the issue by restricting to type dependency on terms of intuitionistic types. In this case, there is no conflict, as those terms can be copied and deleted freely.
An argument for this system, that is of a semantic rather than a syntactic nature, is the following.
We shall see that if we start out with a model of linear type theory with external quantification (i.e.
strict indexed symmetric monoidal category) and demand the quantification to be internal (i.e. demand
it to be a linear dependent type theory) in the natural way (i.e. impose the comprehension axiom), we
will see that our base category becomes a cartesian category (i.e. a model of intuitionistic type theory).
Our best attempt to give meaning on type dependency on linear types, will be through dependency
on a (co-Kleisli or co-Eilenberg-Moore) category of co-algbras for !. We will briefly return to the issue
later, but most of the discussion is beyond the scope of this paper.
In a linear dependent type theory, one could initially conceive of the option of an additive - denote
⊗
with Σ&
x∶X A - and multiplicative - denote with Σx∶X A - Σ-type to generalise the additive and multiplicative conjunction of intuitionistic linear logic. The modality would relate the additive Σ-type of linear type
families to the multiplicative Σ-type of intuitionistic type families, through a Seely-like isomorphism:
⊗
!(Σ&
x∶X A) = Σx∶!X !A.
The idea is that the linear Σ-types should simultaneously generalise the conjunctions from propositional
linear logic and the Σ-types from intuitionistic dependent type theory.
Generally, though, since we are restricting to type dependency on !-ed types, the additive Σ-type (as
far as one can make sense of the notion) would become effectively intuitionistic (as both terms in the
introduction rule would have to depend on the same context, one of which is forced to be intuitionistic).
This also ruins the symmetry in the Seely isomorphism. Moreover, we will see that the multiplicative
Σ-types (and Π-types) arise naturally in the categorical semantics. We will later briefly return to the
possibility of a really linear additive Σ-type and Seely isomorphisms.
For the Π-type, one would also expect to obtain two linear cousins: an additive and a multiplicative
version. However, given the overwhelming preference in the linear logic literature of the multiplicative
implication, we can probably safely restrict our attention to multiplicative Π-types for now.
7
2
Syntax of ILDTT
We assume the reader has some familiarity with the formal syntax of dependent type theory and linear
type theory. In particular, we will not go into syntactic details like α-conversion, name binding, capturefree substitution of a for x in t (write t[a/x]), and pre-syntax. The reader can find details on all of these
topics in [5].
We next present the formal syntax of ILDTT. We start with a presentation of the judgements that
will represent the propositions in the language and then discuss its rules of inference: first its structural
core, then the logical rules for a series of optional type formers. We conclude this section with a few
basic results about the syntax.
Judgements
We adopt a notation ∆; Ξ for contexts, where ∆ is ‘an intuitionistic region’ and Ξ is ‘a linear region’,
similarly to [4]. The idea will be that we have an empty context and can extend an existing context ∆; Ξ
with both intuitionistic and linear types that are allowed to depend on ∆.
Our language will express judgements of the following six forms.
ILDTT judgement
⊢ ∆; Ξ ctxt
∆; ⋅ ⊢ A type
∆; Ξ ⊢ a ∶ A
⊢ ∆; Ξ ≡ ∆′ ; Ξ′ ctxt
∆; ⋅ ⊢ A ≡ A′ type
∆; Ξ ⊢ a ≡ a′ ∶ A
Intended meaning
∆; Ξ is a valid context
A is a type in (intuitionistic) context ∆
a is a term of type A in context ∆; Ξ
∆; Ξ and ∆′ ; Ξ′ are judgementally equal contexts
A and A′ are judgementally equal types in (intuitionistic) context ∆
a and a′ are judgementally equal terms of type A in context ∆; Ξ
Figure 1: Judgements of ILDTT.
Structural Rules
We will use the following structural rules, which are essentially the structural rules of dependent type
theory where some rules appear in both an intuitionistic and a linear form. We present the rules per
group, with their names, from left-to-right, top-to-bottom.
⋅; ⋅ ctxt
C-Emp
⊢ ∆; Ξ ctxt
∆; ⋅ ⊢ A type
Int-C-Ext
⊢ ∆, x ∶ A; Ξ ctxt
∆; Ξ ≡ ∆′ ; Ξ′ ctxt
⊢ ∆; Ξ ctxt
∆; ⋅ ⊢ A type
Lin-C-Ext
⊢ ∆; Ξ, x ∶ A ctxt
∆; Ξ ≡ ∆′ ; Ξ′ ctxt
∆, x ∶ A, ∆′ ; ⋅ ctxt
∆, x ∶ A, ∆′ ; ⋅ ⊢ x ∶ A
∆; ⋅ ⊢ A ≡ B type
⊢ ∆, x ∶ A; Ξ ≡ ∆′ , y ∶ B; Ξ′ ctxt
∆; ⋅ ⊢ A ≡ B type
⊢ ∆; Ξ, x ∶ A ≡ ∆′ ; Ξ′ , y ∶ B ctxt
Int-C-Ext-Eq
Lin-C-Ext-Eq
∆; x ∶ A ctxt
Lin-Var
∆; x ∶ A ⊢ x ∶ A
Int-Var
Figure 2: Context formation and variable declaration rules.
8
⊢ ∆; Ξ ≡ ∆′ ; Ξ′ ctxt
⊢ ∆; Ξ ctxt
C-Eq-R
⊢ ∆; Ξ ≡ ∆; Ξ ctxt
⊢ ∆; Ξ ≡ ∆′ ; Ξ′ ctxt
⊢ ∆′ ; Ξ′ ≡ ∆; Ξ ctxt
⊢ ∆′ ; Ξ′ ≡ ∆′′ ; Ξ′′ ctxt
∆′′ ; Ξ′′
⊢ ∆; Ξ ≡
ctxt
C-Eq-T
∆; Ξ ⊢ A ≡ A′ type
∆; Ξ ⊢ A type
Ty-Eq-R
∆; Ξ ⊢ A ≡ A type
∆; Ξ ⊢ A ≡ A′ type
∆; Ξ ⊢ A′ ≡ A type
∆; Ξ ⊢ A′ ≡ A′′ type
∆; Ξ ⊢ A ≡ A′′ type
∆; Ξ ⊢ a ≡ a′ ∶ A
∆; Ξ ⊢ a′ ≡ a ∶ A
∆; Ξ ⊢ a′ ≡ a′′ ∶ A
∆; Ξ ⊢ a ≡ a′′ ∶ A
⊢ ∆; Ξ ≡ ∆; Ξ′ ctxt
∆; Ξ ⊢ a ∶ A
′
′
∆ ;Ξ ⊢ a ∶ A
∆′ ; ⋅ ⊢ A type
∆; ⋅ ⊢ A ≡ A′ Type
⊢ ∆; ⋅ ≡ ∆′ ; ⋅ ctxt
∆ ; ⋅ ⊢ A type
Tm-Eq-S
Tm-Eq-T
′
′
Ty-Eq-S
Ty-Eq-T
∆; Ξ ⊢ a ∶ A
Tm-Eq-R
∆; Ξ ⊢ a ≡ a ∶ A
∆; Ξ ⊢ a ≡ a′ ∶ A
C-Eq-S
Tm-Conv
Ty-Conv
Figure 3: A few standard rules for judgemental equality, saying that it is an equivalence relation and is
compatible with typing.
∆, ∆′ ; Ξ ⊢ J
∆; ⋅ ⊢ A type
∆, x ∶ A, ∆′ ; Ξ ⊢ J
∆, x ∶ A, x′ ∶ A′ , ∆′ ; Ξ ⊢ J
∆, x′ ∶ A′ , x ∶ A, ∆′ ; Ξ ⊢ J
Int-Weak
∆; Ξ, x ∶ A, x′ ∶ A′ , Ξ′ ⊢ J
Int-Exch
∆; Ξ, x′ ∶ A′ , x ∶ A, Ξ′ ⊢ J
Lin-Exch
(if x is not free in A′ )
∆, x ∶ A, ∆′ ; ⋅ ⊢ B type
∆, ∆′ [a/x]; ⋅
∆; ⋅ ⊢ a ∶ A
⊢ B[a/x] type
∆, x ∶ A, ∆′ ; Ξ ⊢ b ∶ B
∆; ⋅ ⊢ a ∶ A
′
∆, ∆ [a/x]; Ξ[a/x] ⊢ b[a/x] ∶ B[a/x]
∆; Ξ, x ∶ A ⊢ b ∶ B
∆; Ξ′ ⊢ a ∶ A
∆; Ξ, Ξ′ ⊢ b[a/x] ∶ B
Int-Ty-Subst
∆, x ∶ A, ∆′ ; ⋅ ⊢ B ≡ B ′ type
∆; ⋅ ⊢ a ∶ A
∆, ∆′ [a/x]; ⋅ ⊢ B[a/x] ≡ B ′ [a/x] type
∆, x ∶ A, ∆′ ; Ξ ⊢ b ≡ b′ ∶ B
Int-Tm-Subst
∆; ⋅ ⊢ a ∶ A
∆, ∆′ [a/x]; Ξ ⊢ b[a/x] ≡ b′ [a/x] ∶ B[a/x]
∆; Ξ, x ∶ A ⊢ b ≡ b′ ∶ B
Lin-Tm-Subst
∆; Ξ′ ⊢ a ∶ A
∆; Ξ, Ξ′ ⊢ b[a/x] ≡ b′ [a/x] ∶ B
Int-Ty-Subst-Eq
Int-Tm-Subst-Eq
Lin-Tm-Subst-Eq
Figure 4: Exchange, weakening, and substitution rules. Here, J represents a statement of the form
B type, B ≡ B ′ , b ∶ B, or b ≡ b′ ∶ B, such that all judgements are well-formed.
9
Logical Rules
We introduce some basic (optional) type and term formers, for which we will give type formation (denoted
-F), term introduction (-I), term elimination (-E), term computation rules (-C), and (judgemental) term
uniqueness principles (-U). We also assume the obvious rules to hold that state that the new type
formers and term formers respect judgemental equality. Moreover, Σ!x∶!A , Π!x∶!A , λ!x∶!A , and λx∶A are
name binding operators, binding free occurences of x within their scope. Preempting some theorems of
the calculus, we overload some of the notation for -I and -E various type formers, in order to avoid a
lot of syntactic clutter. Needless to say, uniqueness of typing can easily be restored by carrying around
enough type information on the notations corresponding to the various -I and -E rules.
We demand -U-rules for the various type formers in this paper, as this allows us to give a natural
categorical semantics. In practice, when building a computational implementation of a type theory like
ours, one would probably drop these rules to make the system decidable, which would correspond to
switching to weak equivalents of the categorical constructions presented here.4
∆, x ∶ A; ⋅ ⊢ B type
Σ-F
∆; ⋅ ⊢ Σ!x∶!A B type
∆; ⋅ ⊢ C type
∆; Ξ ⊢ t ∶ Σ!x∶!AB
∆; ⋅ ⊢ a ∶ A
∆; Ξ ⊢ b ∶ B[a/x]
Σ-I
∆; Ξ ⊢!a ⊗ b ∶ Σ!x∶!A B
∆, x ∶ A; Ξ′ , y ∶ B ⊢ c ∶ C
∆; Ξ, Ξ′ ⊢ let t be !x ⊗ y in c ∶ C
∆; Ξ ⊢ let !a ⊗ b be !x ⊗ y in c ∶ C
Σ-C
∆; Ξ ⊢ let !a ⊗ b be !x ⊗ y in c ≡ c[a/x, b/y] ∶ C
Σ-E
∆; Ξ ⊢ let t be !x ⊗ y in !x ⊗ y ∶ Σ!x∶!A B
Σ-U
∆; Ξ ⊢ let t be !x ⊗ y in !x ⊗ y ≡ t ∶ Σ!x∶!A B
∆, x ∶ A; ⋅ ⊢ B type
Π-F
∆; ⋅ ⊢ Π!x∶!A B type
∆; ⋅ ⊢ a ∶ A
⊢ ∆; Ξ ctxt
∆, x ∶ A; Ξ ⊢ b ∶ B
Π-I
∆; Ξ ⊢ λ!x∶!A b ∶ Π!x∶!A B
∆; Ξ ⊢ (λ!x∶!A b)(!a) ∶ B
∆; Ξ ⊢ (λ!x∶!A b)(!a) ≡ b[a/x] ∶ B[a/x]
∆; ⋅ ⊢ a ∶ A
∆; ⋅ ⊢ a′ ∶ A
∆; ⋅ ⊢ Id!A (a, a′ ) type
∆; ⋅ ⊢ a ∶ A
∆; ⋅ ⊢ refl!a ∶ Id!A (a, a)
∆; Ξ ⊢ f ∶ Π!x∶!A B
∆; Ξ ⊢ f (!a) ∶ B[a/x]
∆; Ξ ⊢ λ!x∶!A f (!x) ∶ Π!x∶!A B
Π-C
∆; Ξ ⊢ f ≡ λ!x∶!A f (!x) ∶ Π!x∶!A B
∆, x ∶ A, x′ ∶ A; ⋅ ⊢ D type
∆, z ∶ A; Ξ ⊢ d ∶ D[z/x, z/x′ ]
∆; ⋅ ⊢ a ∶ A
∆; ⋅ ⊢ a′ ∶ A
∆; Ξ′ ⊢ p ∶ Id!A (a, a′ )
Id-F
Id-I
∆; Ξ[a/z], Ξ′ ⊢ let (a, a′ , p) be (z, z, refl!z ) in d ∶ D[a/x, a′ /x′ ]
∆; Ξ ⊢ let (a, a, refl!a ) be (z, z, refl!z ) in d ∶ D[a/x, a/x′ ]
∆; Ξ ⊢ let (a, a, refl!a ) be (z, z, refl!z ) in d ≡ d[a/z] ∶ D[a/x, a/x′ ]
Π-E
Π-U
Id-E
Id-C
∆, x ∶ A, x′ ∶ A; Ξ, z ∶ Id!A (x, x′ ) ⊢ let (x, x′ , z) be (x, x, refl!x ) in c[x/x′ , refl!x /z] ∶ C
∆, x ∶ A, x′ ∶ A; Ξ, z ∶ Id!A (x, x′ ) ⊢ let (x, x′ , z) be (x, x, refl!x ) in c[x/x′ , refl!x /z] ≡ c ∶ C
Id-U
Figure 5: Rules for linear equivalents of some of the usual type formers from DTT: Σ-, Π-, and Id-types.
4
In that case in DTT, one would usually demand some stronger ‘dependent’ elimination rules, which would make propositional equivalents of the -U-rules provable, adding some extensionality to the system, while preserving its computational
properties. Such rules are problematic in ILDTT, however, both from a syntactic and semantic point of view and a further
investigation is warranted here.
10
I-F
∆; ⋅ ⊢ I type
I-I
∆; ⋅ ⊢ ∗ ∶ I
∆; Ξ′ ⊢ t ∶ I
∆; Ξ ⊢ a ∶ A
∆; Ξ, Ξ′ ⊢ let t be ∗ in a ∶ A
I-E
∆; Ξ ⊢ let t be ∗ in ∗ ∶ I
I-U
∆; Ξ ⊢ let t be ∗ in ∗ ≡ t ∶ I
∆; Ξ ⊢ let ∗ be ∗ in a ∶ A
I-C
∆; Ξ ⊢ let ∗ be ∗ in a ≡ a ∶ A
∆; ⋅ ⊢ A type
∆; ⋅ ⊢ B type
⊗-F
∆; ⋅ ⊢ A ⊗ B type
∆; Ξ′ ⊢ b ∶ B
∆; Ξ ⊢ a ∶ A
′
∆; Ξ, Ξ ⊢ a ⊗ b ∶ A ⊗ B
∆; Ξ ⊢ t ∶ A ⊗ B
⊗-I
∆; Ξ′ , x ∶ A, y ∶ B ⊢ c ∶ C
′
∆; Ξ, Ξ ⊢ let t be x ⊗ y in c ∶ C
∆; Ξ ⊢ let a ⊗ b be x ⊗ y in c ∶ C
⊗-C
∆; Ξ ⊢ let a ⊗ b be x ⊗ y in c ≡ c[a/x, b/y] ∶ C
⊗-E
∆; Ξ ⊢ let t be x ⊗ y in x ⊗ y ∶ A ⊗ B
⊗-U
∆; Ξ ⊢ let t be x ⊗ y in x ⊗ y ≡ t ∶ A ⊗ B
∆; ⋅ ⊢ A type
∆; ⋅ ⊢ B type
⊸-F
∆; ⋅ ⊢ A ⊸ B type
∆; Ξ ⊢ f ∶ A ⊸ B
∆; Ξ, x ∶ A ⊢ b ∶ B
⊸-I
∆; Ξ ⊢ λx∶A b ∶ A ⊸ B
∆; Ξ ⊢ (λx∶A b)(a) ∶ B
∆; Ξ ⊢ (λx∶A b)(a) ≡ b[a/x] ∶ B
∆; ⋅ ⊢ ⊺ type
∆; Ξ, Ξ′
∆; Ξ ctxt
⊺-I
∆; Ξ ⊢ ⟨⟩ ∶ ⊺
⊸-E
∆; Ξ ⊢ t ∶ ⊺
⊺-U
∆; Ξ ⊢ t ≡ ⟨⟩ ∶ ⊺
∆; Ξ ⊢ a ∶ A
∆; Ξ ⊢ b ∶ B
&-I
∆; Ξ ⊢ ⟨a, b⟩ ∶ A&B
∆; ⋅ ⊢ A type
∆; ⋅ ⊢ B type
&-F
∆; ⋅ ⊢ A&B type
∆; Ξ ⊢ t ∶ A&B
&-E1
∆; Ξ ⊢ fst(t) ∶ A
∆; Ξ ⊢ t ∶ A&B
&-E2
∆; Ξ ⊢ snd(t) ∶ B
∆; Ξ ⊢ fst(⟨a, b⟩) ∶ A
∆; Ξ ⊢ fst(⟨a, b⟩) ≡ a ∶ A
∆; Ξ ⊢ snd(⟨a, b⟩) ∶ B
&-C1
∆; Ξ ⊢ ⟨fst(t), snd(t)⟩ ∶ A&B
∆; Ξ ⊢ ⟨fst(t), snd(t)⟩ ≡ t ∶ A&B
∆; ⋅ ⊢ 0 type
⊢ f (a) ∶ B
∆; Ξ ⊢ λx∶A f x ∶ A ⊸ B
⊸-U
∆; Ξ ⊢ λx∶A f x ≡ f ∶ A ⊸ B
⊸-C
⊺-F
∆; Ξ′ ⊢ a ∶ A
∆; Ξ ⊢ snd(⟨a, b⟩) ≡ b ∶ B
&-C2
&-U
∆; Ξ ⊢ t ∶ 0
0-F
∆; Ξ, Ξ′ ⊢ false(t) ∶ B
0-E
∆; Ξ ⊢ t ∶ 0
0-U
∆; Ξ ⊢ false(t) ≡ t ∶ 0
∆; ⋅ ⊢ A type
∆; ⋅ ⊢ B type
⊕-F
∆; ⋅ ⊢ A ⊕ B type
∆; Ξ ⊢ b ∶ B
∆; Ξ ⊢ a ∶ A
⊕-I1
∆; Ξ ⊢ inl(a) ∶ A ⊕ B
∆; Ξ, x ∶ A ⊢ c ∶ C
∆; Ξ, y ∶ B ⊢ d ∶ C
∆; Ξ ⊢ inr(b) ∶ A ⊕ B
∆; Ξ′ ⊢ t ∶ A ⊕ B
′
∆; Ξ, Ξ ⊢ case t of inl(x) → c ∣∣ inr(y) → d ∶ C
∆; Ξ, Ξ′ ⊢ case inl(a) of inl(x) → c ∣∣ inr(y) → d ∶ C
∆; Ξ, Ξ′ ⊢ case inl(a) of inl(x) → c ∣∣ inr(y) → d ≡ c[a/x] ∶ C
∆; Ξ, Ξ′ ⊢ case inr(b) of inl(x) → c ∣∣ inr(y) → d ∶ C
∆; Ξ, Ξ′ ⊢ case inr(b) of inl(x) → c ∣∣ inr(y) → d ≡ d[b/y] ∶ C
⊕-C1
⊕-C2
∆; Ξ, Ξ′ ⊢ case t of inl(x) → inl(x) ∣∣ inr(y) → inr(y) ∶ A ⊕ B
∆; Ξ, Ξ′ ⊢ case t of inl(x) → inl(x) ∣∣ inr(y) → inr(y) ≡ t ∶ A ⊕ B
11
⊕-E
⊕-U
⊕-I2
∆; ⋅ ⊢ A type
!-F
∆; ⋅ ⊢!A type
∆; Ξ ⊢ t ∶!A
∆; ⋅ ⊢ a ∶ A
!-I
∆; ⋅ ⊢!a ∶!A
′
∆, x ∶ A; Ξ′ ⊢ b ∶ B
∆; Ξ, Ξ ⊢ let t be !x in b ∶ B
∆; Ξ ⊢ let !a be !x in b ∶ B
∆; Ξ ⊢ let !a be !x in b ≡ b[a/x] ∶ B
!-E
∆; Ξ ⊢ let t be !x in !x ∶!A
!-U
∆; Ξ ⊢ let t be !x in !x ≡ t ∶!A
!-C
Figure 6: Rules for the usual linear type formers in each context: I-, ⊗-, ⊸-, ⊺-, &-, 0-, ⊕-, and !-types.
Finally, we add rules that say we have all the possible commuting conversions, which from a syntactic
point of view restore the subformula property and from a semantic point of view say that our rules are
natural transformations (between hom-functors), which simplifies the categorical semantics significantly.
We represent these schematically, following [4]. That is, if C[−] is a linear (program rather than typing)
context, then (abusing notation and dealing with all the let be in -constructors in one go).
∆; Ξ ⊢ C[let a be b in c] ∶ D
∆; Ξ ⊢ C[let a be b in c] ≡ let a be b in C[c] ∶ D
if C[−] does not bind any free variables in a or b;
∆; Ξ ⊢ C[false(t)] ∶ D
∆; Ξ ⊢ C[false(t)] ≡ false(t) ∶ D
if C[−] does not bind any free variables in t;
∆; Ξ ⊢ C[case t of inl(x) → c ∣∣ inr(y) → d] ∶ D
∆; Ξ ⊢ C[case t of inl(x) → c ∣∣ inr(y) → d] ≡ case t of inl(x) → C[c] ∣∣ inr(y) → C[d] ∶ D
if C[−] does not bind any free variables in t or x or y.
Figure 7: Commuting conversions.
Remark 1. Note that all type formers that are defined context-wise (I, ⊗, ⊸, ⊺, &, 0, ⊕, and !)
are automatically preserved under the substitutions from Int-Ty-Subst (up to canonical isomorphism5 ),
in the sense that F (A1 , . . . , An )[a/x] is isomorphic to F (A1 [a/x], . . . , An [a/x]) for an n-ary type former F . Similarly, for T = Σ or Π, we have that (T!y∶!B C)[a/x] is isomorphic to T!y∶!B[a/x] C[a/x] and
(Id!B (b, b′ ))[a/x] is isomorphic to Id!B[a/x] (b[a/x], b′ [a/x]). (This gives us Beck-Chevalley conditions
in the categorical semantics.) These are the remaining naturality conditions for rules.
Remark 2. The reader can note that the usual formulation of universes for DTT transfers very naturally
to ILDTT, giving us a notion of universes for linear types. This allows us to write rules for forming types
as rules for forming terms, as usual. We do not choose this approach and define the various type formers
in the setting without universes, as this will give a cleaner categorical semantics.
Some Basic Results
As the focus of this paper is the syntax-semantics correspondence, we will only briefly state a few syntactic
results. For some standard metatheoretic properties for the ⊸, Π, ⊺, &-fragment of our syntax, we refer
the reader to [1]. Standard techniques and some small adaptations of the system should be enough to
extend the results to all of ILDTT.
Theorem 1 (Consistency). ILDTT with all its type formers is consistent.
Proof. We discuss a class of models in section 4.
5
By an isomorphism of types ∆; ⋅ ⊢ A type and ∆; ⋅ ⊢ B type in context ∆, we here mean a pair of terms ∆; x ∶ A ⊢ f ∶ B
and ∆; y ∶ B ⊢ g ∶ A together with a pair of judgemental equalities ∆; x ∶ A ⊢ g[f /y] ≡ x ∶ A and ∆; y ∶ B ⊢ f [g/x] ≡ y ∶ B.
12
To give the reader some intuition for these linear Π- and Σ-types, we suggest the following two
interpretations.
Theorem 2 (Π and Σ as Dependent !(−) ⊸ (−) and !(−) ⊗ (−)). Suppose we have !-types. Let ∆, x ∶
A; ⋅ ⊢ B type, where x does not occur freely in B. Then, for the purposes of the type theory,
1. Π!x∶!A B is isomorphic to !A ⊸ B, if we have Π-types and ⊸-types;
2. Σ!x∶!A B is isomorphic to !A ⊗ B, if we have Σ-types and ⊗-types.
Proof.
1. We will construct terms
and ∆; y ′ ∶!A ⊸ B ⊢ g ∶ Π!x∶!A B
∆; y ∶ Π!x∶!A B ⊢ f ∶!A ⊸ B
s.t.
∆; y ∶ Π!x∶!A B ⊢ g[f /y ′] ≡ y ∶ Π!x∶!A B
and ∆; y ′ ∶!A ⊸ B ⊢ f [g/y] ≡ y ′ ∶!A ⊸ B.
First, we construct f .
Int-Var
Lin-Var
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆, x ∶ A; y ∶ Π!x∶!A B ⊢ y ∶ Π!x∶!A B
Π-E
Lin-Var
∆, x ∶ A; y ∶ Π!x∶!A B ⊢ y(!x) ∶ B
∆; x′ ∶!A ⊢ x′ ∶!A
!-E
∆; y ∶ Π!x∶!A B, x′ ∶!A ⊢ let x′ be !x in y(!x) ∶ B
⊸-I
∆; y ∶ Π!x∶!A B ⊢ f ∶!A ⊸ B
Then, we construct g.
Int-Var
∆, x ∶ A; ⋅ ⊢ x ∶ A
Lin-Var
!-I
∆, x ∶ A; ⋅ ⊢!x ∶!A
∆, x ∶ A; y ′ ∶!A ⊸ B ⊢ y ′ ∶!A ⊸ B
⊸-E
∆, x ∶ A; y ′ ∶!A ⊸ B ⊢ y ′ (!x) ∶ B
Π-I
∆; y ′ ∶!A ⊸ B ⊢ g ∶ Π!x∶!A B
It is easily verified that ⊸-C, !-C, and Π-U imply the first judgemental equality:
g[f /y ′] ≡ λ!x∶!A (λx′ ∶!A let x′ be !x in y(!x))(!x) ≡ λ!x∶!A let !x be !x in y(!x) ≡ λ!x∶!A y(!x) ≡ y.
Similarly, Π-C, commuting conversions, !-U, and ⊸-U imply the second judgemental equality:
f [g/y] ≡ λx′ ∶!A let x′ be !x in (λ!x∶!A y ′ (!x))(!x) ≡ λx′ ∶!A let x′ be !x in y ′ (!x) ≡ λx′ ∶!A y ′ (let x′ be !x in !x)
≡ λx′ ∶!A y ′ (x′ ) ≡ y ′ .
2. We will construct terms
and ∆; y ′ ∶!A ⊗ B ⊢ g ∶ Σ!x∶!A B
∆; y ∶ Σ!x∶!A B ⊢ f ∶!A ⊗ B
s.t.
∆; y ∶ Σ!x∶!A B ⊢ g[f /y ′] ≡ y ∶ Σ!x∶!A B
and ∆; y ′ ∶!A ⊗ B ⊢ f [g/y] ≡ y ′ ∶!A ⊗ B.
First, we construct f .
∆; x′ ∶!A ⊢ x′ ∶!A
Lin-Var
′
∆; z ∶ B ⊢ z ∶ B
′
∆; x ∶!A, z ∶ B ⊢ x ⊗ z ∶!A ⊗ B
′
∆; y ∶ Σ!x∶!A B ⊢ y ∶ Σ!x∶!A B
Lin-Var
′
∆, x ∶ A; x ∶!A, z ∶ B ⊢ x ⊗ z ∶!A ⊗ B
Lin-Var
⊗-I
Int-Weak
∆, x ∶ A; z ∶ B ⊢!x ⊗ z ∶!A ⊗ B
∆; y ∶ Σ!x∶!A B ⊢ f ∶!A ⊗ B
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆, x ∶ A; ⋅ ⊢!x ∶!A
Σ-E
Then, we construct g.
Int-Var
Lin-Var
∆, x ∶ A; y ∶ B ⊢ y ∶ B
Σ-I
∆; x′ ∶!A ⊢ x′ ∶!A
∆, x ∶ A; y ∶ B ⊢!x ⊗ y ∶ Σ!x∶!A B
∆; x′ ∶!A, y ∶ B ⊢ let x′ be !x in !x ⊗ y ∶ Σ!x∶!A B
⊗-E
∆; y′ ∶!A ⊗ B ⊢ g ∶ Σ!x∶!A B
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆; y′ ∶!A ⊗ B ⊢ y′ ∶!A ⊗ B
Lin-Var
Lin-Var
!-E
Here, the first judgemental equality follows from commuting conversions, ⊗-C, !-C, and Σ-U:
13
Int-Var
!-I
Lin-Subst
g[f /y ′] ≡ let (let y be !x ⊗ z in !x ⊗ z) be x′ ⊗ y in (let x′ be !x in !x ⊗ y) ≡ let y be !x ⊗ z in let !x ⊗
z be x′ ⊗ y in (let x′ be !x in !x ⊗ y) ≡ let y be !x ⊗ z in (let x′ be !x in !x ⊗ y)[!x/x′ ][z/y] ≡
let y be !x ⊗ z in (let !x be !x in !x ⊗ z) ≡ let y be !x ⊗ z in !x ⊗ z ≡ y.
The second judgemental equality follows from commuting conversions, Σ-C, !-U, and ⊗-U:
f [g/y] ≡ let (let y ′ be x′ ⊗ y in (let x′ be !x in !x ⊗ y)) be !x ⊗ z in !x ⊗ z ≡ let y ′ be x′ ⊗
y in let x′ be !x in let !x ⊗ y be !x ⊗ z in !x ⊗ z ≡ let y ′ be x′ ⊗ y in let x′ be !x in (!x ⊗ y) ≡
let y ′ be x′ ⊗ y in (let x′ be !x in !x) ⊗ y ≡ let y ′ be x′ ⊗ y in x′ ⊗ y ≡ y ′ .
In particular, we have the following stronger version of a special case.
Theorem 3 (! as ΣI). Suppose we have Σ- and I-types. Let ∆; ⋅ ⊢ A type. Then, Σ!x∶!A I satisfies the
rules for !A. Conversely, if we have !- and I-types, then !A satisfies the rules for Σ!x∶!A I.
Proof. We obtain the !-rule as follows.
I-I
∆; ⋅ ⊢ a ∶ A
∆, x ∶ A; ⋅ ⊢ ∗ ∶ I
Σ-I
∆; ⋅ ⊢!a ⊗ ∗ ∶ Σ!x∶!A I
We obtain the !-E rule as follows.
Lin-Var
∆, x ∶ A; Ξ′ ⊢ c ∶ C
∆; y ∶ I ⊢ y ∶ I
I-E
∆; Ξ ⊢ t ∶ Σ!x∶!A I
∆, x ∶ A; Ξ′ , y ∶ I ⊢ let y be ∗ in c ∶ C
Σ-E
∆; Ξ, Ξ′ ⊢ let t be !x ⊗ y in let y be ∗ in c ∶ C.
It is easily seen that Σ-C and I-C imply !-C (let !a ⊗ ∗ be !x ⊗ y in let y be ∗ in c ≡ (let y be ∗
in c)[a/x][∗/y] ≡ let ∗ be ∗ in c[a/x] ≡ c[a/x]) and that I-U and Σ-U (and commuting conversions) imply !-U (let t be !x⊗y in let y be ∗ in x⊗∗ ≡ let t be !x⊗y in !x⊗let y be ∗ in ∗ ≡ let t be !x⊗y in !x⊗y ≡ t).
The converse statement follows through a similarly trivial argument, noting that I[a/x] is isomorphic to
I.
A second interpretation is that Π and Σ generalise & and ⊕. Indeed, the idea is that (or their
infinitary equivalents) is what they reduce to when taken over discrete types. The subtlety in this result
will be the definition of a discrete type. The same phenomenon is observed in a different context in
section 4.
For our purposes, a discrete type is a strong sum of I (a sum with a dependent -E-rule). Let us for
simplicity limit ourselves to the binary case. For us, the discrete type with two elements will be 2 = I ⊕ I,
where ⊕ has a strong/dependent -E-rule (note that this is not our ⊕-E). Explicitly, 2 is a type with the
following rules:
∆; ⋅ ⊢ 2 type
2-F
∆, x ∶ 2; ⋅ ⊢ A type
∆; ⋅ ⊢ tt ∶ 2
∆; ⋅ ⊢ t ∶ 2
∆; Ξ ⊢ att ∶ A[tt/x]
2-I1
∆; ⋅ ⊢ ff ∶ 2
∆; Ξ ⊢ aff ∶ A[ff/x]
∆; Ξ ⊢ if t then att else aff ∶ A[t/x]
∆; Ξ ⊢ if tt then att else aff ∶ A[tt/x]
∆; Ξ ⊢ if tt then att else aff ≡ att ∶ A[tt/x]
2-E
∆; Ξ ⊢ if ff then att else aff ∶ A[ff/x]
2-C1
2-I2
∆; Ξ ⊢ if ff then att else aff ≡ aff ∶ A[ff/x]
2-C2
∆; Ξ ⊢ if t then tt else ff ∶ 2
2-U
∆; Ξ ⊢ if t then tt else ff ≡ t ∶ 2
Figure 8: Rules for a discrete type 2.
Theorem 4 (Π and Σ as Infinitary Non-Discrete & and ⊕). If we have a discrete type 2 and a type
family ∆, x ∶ 2; ⋅ ⊢ A, then
1. Π!x∶!2 A satisfies the rules for A[tt/x]&A[ff/x];
14
2. Σ!x∶!2 A satisfies the rules for A[tt/x] ⊕ A[ff/x].
Proof.
1. We obtain &-I as follows.
∆, x ∶ 2; Ξ ⊢ a ∶ A[tt/x]
Int-Var
∆, x ∶ 2; Ξ ⊢ b ∶ A[ff/x]
∆, x ∶ 2; ⋅ ⊢ x ∶ 2
∆, x ∶ 2; Ξ ⊢ if x then a else b ∶ A
Π-I
∆; Ξ ⊢ λ!x∶!2 if x then a else b ∶ Π!x∶!2 A
∆, x ∶ 2; ⋅ ⊢ A type
Assumption
2-E-dep
Moreover, we obtain &-E1 as follows (similarly, we obtain &-E2).
2-I1
∆; ⋅ ⊢ tt ∶ 2
∆; Ξ ⊢ t ∶ Π!x∶!2 A
Π-E
t(!tt)
The &-C-rules follow from Π-C and 2-C, e.g.
fst⟨a, b⟩ ∶≡ (λ!x∶!2 if x then a else b)(!tt) ≡ if tt then a else b ≡ a.
The &-U-rules follow from Π-U and 2-U (and commuting conversions):
⟨fst(t), snd(t)⟩ ∶≡ λ!x∶!2 if x then t(!tt) else t(!ff) ≡ λ!x∶!2 t(!x) ≡ t.
2. We obtain ⊕-I1 as follows (and similarly, we obtain ⊕-I2):
2-I1
∆; ⋅ ⊢ tt ∶ 2
∆; Ξ ⊢ a ∶ A[tt/x]
Σ-I
∆; Ξ ⊢ !tt ⊗ a ∶ Σ!x∶!2 A
Moreover, we obtain ⊕-E as follows.
∆; Ξ′ ⊢ t ∶ Σ!x∶!2 A
∆; Ξ, z ∶ A[tt/x] ⊢ c ∶ C
∆; Ξ, w ∶ A[ff/x] ⊢ d ∶ C
∆, x ∶ 2; Ξ, y ∶ A ⊢ c[y/z] ∶ C
∆, x ∶ 2; Ξ, y ∶ A ⊢ d[y/w] ∶ C
∆, x ∶ 2; ⋅ ⊢ x ∶ 2
∆, x ∶ 2; ⋅ ⊢ A type
∆, x ∶ 2; Ξ, y ∶ A ⊢ if x then c[y/z] else d[y/w] ∶ C
∆; Ξ, Ξ′ ⊢ let t be !x ⊗ y in if x then c[y/z] else d[y/w] ∶ C
Assump
2-E-dep
Σ-E
The ⊕-C-rules follow from Σ-C and 2-C, e.g.
case inl(a) of inl(z) → c∣∣ inr(w) → d ∶ ≡ let !tt ⊗ a be !x ⊗ y in if x then c[y/z] else d[y/w]
≡ if tt then c[a/z] else d[a/w]
= c[a/z].
The ⊕-U-rules follow from Σ-U and 2-U (and commuting conversions):
case t of inl(z) → inl(z)∣∣ inr(w) → inr(w) ∶ ≡ let t be !x ⊗ y in if x then inl(z)[y/z] else inr(w)[y/w]
≡ let t be !x ⊗ y in if x then !tt ⊗ z[y/z] else !ff ⊗ w[y/w]
≡ let t be !x ⊗ y in if x then !tt ⊗ y else !ff ⊗ y
≡ let t be !x ⊗ y in !(if x then tt else ff) ⊗ y
≡ let t be !x ⊗ y in !x ⊗ y
≡ t.
We see that we can view Π and Σ as generalisations of & and ⊕, respectively.
15
3
Semantics of ILDTT
The idea behind the categorical semantics we present for the structural core of our syntax (with I- and
⊗-types) will be to take our suggested categorical semantics for the structural core of DTT (with ⊺- and
∧-types) and relax the assumption of the Cartesian character of its fibres to them only being (possibly
non-Cartesian) symmetric. This entirely reflects the relation between the conventional semantics of nondependent intuitionistic and linear type systems. The structure we obtain is that of a strict indexed
symmetric monoidal6 category with comprehension.
The Σ- and Π-types arise as left and right adjoints of substitution functors along projections in
the base-category and the Id-types arise as left adjoints to substitution along diagonals, all satisfying
Beck-Chevalley (and Frobenius) conditions, as is the case in the semantics for DTT. The !-types boil
down to having a left adjoint to the comprehension (which can be made a functor), giving a linear-nonlinear adjunction as in the conventional semantics for linear logic. Finally, additive connectives arise as
compatible Cartesian and distributive co-Cartesian structures on the fibres, as would be expected from
the semantics of linear logic.
3.1
Tautological models of ILDTT
First, we translate the structural core of our syntax to the tautological notion of model. We will later
prove this to be equivalent to the more intuitive notion of categorical model we referred to above.
̃ of ILDTT, we shall mean
Definition (Tautological model of ILDTT). With a (tautological) model T
the following.
1. We have a set ICtxt, of which the elements will be interpreted as (dependent) contexts consisting
of intuitionistic types. Then, for all ∆ ∈ ICtxt, we have a set LType(∆) of linear types and a set
LCtxt(∆) of linear contexts (multisets of linear types) in the context ∆. For each ∆, Ξ ∈ LCtxt(∆),
and each A ∈ LType(∆), we have a set LTerm(∆, Ξ, A) of (linear) terms of A in context ∆; Ξ. On
all these sets, judgemental equality is interpreted by equality of elements, taking into account αconversion for terms. (The means that, if we construct a model from the syntax, we devide out
judgemental equality on the syntactic objects.) This takes care of the -Eq rules, the rules expressing
that judgemental equality is an equivalence relation, and the rules relating typing and judgemental
equality.
2. C-Emp says that ICtxt has a distinguished element ⋅ and that LType(⋅) has a distinguished element
⋅.
3. Int-C-Ext says that for all ∆ ∈ ICtxt and A ∈ LType(∆), we can form a context ∆.A ∈ ICtxt, and
that we have a function LCtxt(∆) Ð→ LCtxt(∆.A), introducing fake dependencies of the linear
types on A. Int-Exch says that for two types in the same context, the order in which we append
them to a context does not matter.
4. Lin-C-Ext says that for all Ξ ∈ LCtxt(∆) and A ∈ LType(∆), we can form a context Ξ.A ∈
LCtxt(∆). Lin-Exch says that the order in which we do this does not matter.
5. Int-Var says that for a context ∆.A.∆′ , we have a term der∆,A,∆′ ∈ LTerm(∆.A.∆′ , ⋅, A), that this is implicitly present in the syntax - acts as a diagonal morphism through the substitution
operations of Int-Ty-Subst and Int-Tm-Subst, equating the values of two variables of type A.
6. Lin-Var says that for a context ∆ and a type A ∈ LType(∆), we have a term idA ∈ LTerm(∆, A, A),
that - this is implicitly present in the syntax - acts as the identity for the substitution operation
of Lin-Tm-Subst.
7. Int-Weak says that we have functions (abusing notation) LType(∆.∆′ )
weak∆,A,∆′
weak∆,A,∆′
Ð→
LType(∆.A.∆′ )
Ð→
LTerm(∆.A.∆′ , Ξ, B). We think of this as projecting away a
and LTerm(∆.∆′ , Ξ, B)
variable x ∶ A to introduce a fake dependency.
6
As said before, we can easily obtain a sound and complete semantics for only the structural core, possibly without Iand ⊗-types, by considering strict indexed symmetric multicategories with comprehension.
16
8. Int-Ty-Subst says that for B ∈ LType(∆.A.∆′ ) and a ∈ LTerm(∆, ⋅, A), we have a context ∆.∆′ [a/x] ∈
ICtxt and a B[a/x] ∈ LType(∆.∆′ [a/x]).
9. Int-Tm-Subst says that for b ∈ LTerm(∆.A, ∆′ , Ξ, B) and a ∈ LTerm(∆, ⋅, A), we have b[a/x] ∈
LTerm(∆.∆′ [a/x], Ξ[a/x], B[a/x]).
10. Lin-Tm-Subst says that for b ∈ LTerm(∆, Ξ.A, B) and a ∈ LTerm(∆, Ξ′ , A), we have b[a/x] ∈
LTerm(∆, Ξ.Ξ′ , B).
For these last three substitution operations it is implicit in the syntax that they are associative.
Finally, the remarks in Int-Var and Int-Weak about diagonals and projections in formal terms mean
that the morphisms coming from these rules work together to form a generalised comonoid.
It is tautological that there is a one to one correspondence between theories T in ILDTT and models
̃ of this sort.
T
We now define what it means for the model to support various type formers.
̃ supports I-types, if for all ∆ ∈ ICtxt, we have
Definition (Semantic I- and ⊗-types). We say a model T
an I ∈ LType(∆) and ∗ ∈ LTerm(∆, ⋅, I) and whenever t ∈ LTerm(∆, Ξ, I) and a ∈ LTerm(∆, Ξ, A), we
have let t be ∗ in a ∈ LTerm(∆, Ξ.Ξ′ , A), such that let ∗ be ∗ in a = a and let t be ∗ in ∗ = t.
Similarly, we say it admits ⊗-types, if for all A, B ∈ LType(∆), we have a A ⊗ B ∈ LType(∆),
for all a ∈ LTerm(∆, Ξ, A), b ∈ LTerm(∆, Ξ′ , B), we have a ⊗ b ∈ LType(∆, Ξ.Ξ′ , A ⊗ B), and if t ∈
LTerm(∆, Ξ, A ⊗ B) and c ∈ LTerm(∆, Ξ′ .A.B, C), we have let t be x ⊗ y in c ∈ LTerm(∆, Ξ.Ξ′ , C), such
that let a ⊗ b be x ⊗ y in c = c and let t be x ⊗ y in x ⊗ y = t.
⊗
Note that this defines a function LCtxt(∆) Ð→ LType. The C-rule precisely says that from the
point of view of the (terms of the) type theory this map is an injection, while the U-rule says it is a
surjection7 . We conclude that in presence of I- and ⊗-types, we can faithfully describe the type theory
without mentioning linear contexts, replacing them by the linear type that is their ⊗-product.
We will henceforth assume that our type theory has I- and ⊗-types, as this simplifies the categorical
semantics8 and is appropriate for the examples we are interested in.
For the other type formers, one can give a similar, almost tautological, translation from the syntax
into a tautological model. We trust this to the reader when we discuss the semantic equivalent of various
type formers in the categorical semantics we present next.
3.2
Categorical Semantics of ILDTT
Strict Indexed Symmetric Monoidal Categories with Comprehension
We now introduce a notion of categorical model for which soundness and completeness results hold with
respect to the syntax of ILDTT in presence of I- and ⊗-types9 . This notion of model will prove to be
particularly useful when thinking about various (extensional) type formers.
Definition 1. By a strict indexed symmetric monoidal category with comprehension, we will mean the
following data.
1. A category C with a terminal object ⋅.
2. A strict indexed symmetric monoidal category L over C, i.e. a contravariant functor L into the
L
category SMCat of (small) symmetric monoidal categories and strong monoidal functors C op Ð→
SMCat. We will also write −{f } ∶= L(f ) for the action of L on a morphism f of C.
7
The precise statement that we are alluding to here would be that the multicategory of linear contexts is equivalent
to the (monoidal) multicategory of linear types. Really, ⊗ is only part of an equivalence of categories rather than an
isomorphism, i.e. it is injective on objects up to isomorphism rather than on the nose.
8
To be precise, it allows us to give a categorical semantics in terms of monoidal categories rather than multicategories.
9
In case we are interested in the case without I- and ⊗-types, the semantics easily generalises to strict indexed symmetric
multicategories with comprehension.
17
3. A comprehension schema, i.e. for each ∆ ∈ ob(C) and A ∈ ob(L(∆)) a representation for the
functor
x ↦ L(dom(x))(I, A{x}) ∶ (C/∆)op Ð→ Set.
p∆,A
We will write its representing object10 ∆.A Ð→ ∆ ∈ ob(C/∆) and universal element v∆,A ∈
L(∆.A)(I, A{p∆,A }). We will write a ↦ ⟨f, a⟩ for the isomorphism L(∆′ )(I, A{f }) ≅ C/∆(f, p∆,A ).
Again, the comprehension schema means that the morphisms in our category of contexts C, into a
context built by adjoining types, arise as lists of closed linear terms. Here, there is the crucial identification with intuitionistic terms of linear terms without linear assumptions: they can be freely copied and
discarded.
Remark 3. Note that this notion of model reduces to a standard notion of model for intuitionistic
dependent type theory in the case that the monoidal structures on the fibre categories are Cartesian:
a strict indexed Cartesian monoidal category with comprehension. Indeed, these are easily seen to be
equivalent to, for instance, the split comprehension categories of [6] with terminal object and Cartesian
products. However, it turns we have to impose the extra requirement of fullness on the comprehension
category to get an exact match with the syntax. The corresponding condition in our framework is to ask
for the comprehension functor to be full and faithful. See [6] for more discussion.
Theorem 5 (Comprehension functor). A comprehension schema (p, v) on a strict indexed symmetric
monoidal category (C, L) defines a morphism L Ð→ I of indexed categories, which lax-ly sends the
monoidal structure of L to products in I (where they exist), where I is the full subindexed11 category of
C/− on the objects of the form p∆,A .
M
Proof. First note that a morphism M of indexed symmetric monoidal categories consists of lax monoidal
functors in each context ∆ ∈ C such that
M∆
✲ I(∆)
L(∆)
L(f )
I(f ) = ”pullback along f ”
≅
❄
✲ I(∆′ ).
❄ M ′
∆
L(∆′ )
We define
M∆ (A Ð→ B) ∶= p∆,A
a
⟨p∆,A , a{p∆,A } ○ v∆,A ⟩
✲ p∆,B .
Functoriality follows from the uniqueness property of ⟨id∆ , a⟩.
We define the lax monoidal structure
mI∆
id∆
p∆,A ○ p∆.A,B{p∆,A} = M∆ (A) × M∆ (B)
✲ M∆ (I) = p∆,I
mA,B
∆✲
M∆ (A ⊗ B) = p∆,A⊗B ,
where mI∆ ∶= ⟨id∆ , idI ⟩ and mA,B
∶= ⟨p∆,A ○ p∆.A,B{p∆,A} , v∆,A {p∆.A,B{p∆,A} } ⊗ v∆.A,B{p∆,A } ⟩.
∆
Finally, we verify that I(f )M∆ = M∆′ L(f ). This follows directly from the fact that the following
square is a pullback square:
∆′ .A{f }
qf,A
✲ ∆.A
p∆,A
p∆′ ,A{f }
❄
∆′
10 Really,
f
p∆,M A
❄
✲ ∆,
∆.M A Ð→ ∆ would be a better notation, where we think of L ⊣ M as an adjunction inducing !, but it would
be very verbose.
11
Here, we use the axiom of choice to make a choice of pullback and make I really into an indexed category (or cloven
fibration). Alternatively, we can avoid the axiom of choice and treat it as a more general fibration.
18
where qf,A ∶= ⟨f p∆′ ,A{f } , v∆′ ,A{f } ⟩. We leave this verification to the reader as an exercise. Alternatively,
a proof for this fact in DTT, that will transfer to our setting in its entirety, can be found in [5].
Remark 4. Note that I is a display map category (or, less specifically, a full comprehension category).
Hence, it is a model of intuitionistic type theory. We will see that, in many ways, we can regard it as
the intuitionistic content of L.
Remark 5. We will see that this functor will give us a unique candidate for !-types: ! ∶= LM , where
L ⊣ M . We conclude that, in ILDTT, the !-modality is uniquely determined by the indexing. This is
worth noting, because, in propositional linear type theory, we might have many different candidates for
!-types.
Moreover, it explains why we don’t demand M to be fully faithful in the case of linear types. Indeed,
although we have a map L(∆)(A, B) Ð→ I(∆)(p∆,A , p∆,B ) ≅ L(∆.A)(I, B{p∆,A }), this is not generally
an isomorphism. In fact, in presence of !-types, we will see that the right hand side is precisely isomorphic
to L(∆)(!A, B) and the map is precomposition with dereliction.
M∆
Next, we prove that we have a sound interpretation of ILDTT in such categories.
Theorem 6 (Soundness). A strict indexed symmetric monoidal category with comprehension (C, L, p, v)
̃(C,L,p,v) of ILDTT with I- and ⊗-types.
defines a model T
Proof. We define
1. ICtxt ∶= ob(C)
LType(∆) ∶= ob(L(∆))
LCtxt(∆) ∶= free − comm. − monoid(LType(∆)) (where we will write 0 and + for the operations)
LTerm(∆, Ξ, A) ∶= L(∆)(⊗ Ξ, A)
2. C-Emp: ⋅ICtxt ∶= ⋅C and ⋅LCtxt(∆) ∶= 0LCtxt(∆) .
3. Int-C-Ext: ∆.ICtxt A ∶= ∆.C A and LType(∆) Ð→ LType(∆.A) ∶= −{p∆,A} inducing the obvious
function LCtxt(∆) Ð→ LCtxt(∆.A). We have seen how
∆.A.B{p∆,A }
qp∆,A ,B
✲ ∆.B
p∆.A,B{p∆,A }
p∆,B
❄
∆.A
p∆,A
❄
✲ ∆
is a product in C/∆ which interprets the double context extension ∆.A.B where A, B ∈ LType(∆).
Being a Cartesian monoidal structure, this is, in particular, symmetric, so validates Int-Exch.
4. Lin-C-Ext: Ξ.A ∶= Ξ + A
5. Int-Var: der∆,A,∆′ ∈ LTerm(∆.A.∆′ , ⋅, A) is defined as
v∆,A {p∆.A,∆′ } ∶ I Ð→ A{p∆,A ○ p∆.A,∆′ } ∈ L(∆.A.∆′ )
Note that der∆,A,∆′ defines a morphism
∆.A.∆′ Ð→ ∆.A.∆′ .A{p∆,A ○ p∆.A,∆′ } ∶= ⟨id∆.A.∆′ , der∆,A,∆′ ⟩.
We will later show that this in fact behaves as a diagonal morphism on A.
diag∆,A,∆′
6. idA ∈ LTerm(∆, A, A) is taken to be idA ∈ L(∆)(A, A). Note that this is indeed the neutral element
for our semantic linear term substitution operation that we will define shortly.
19
7. The required morphisms in Int-Weak are interpreted as follows. Suppose we are given A, ∆′ ∈
ob(L(∆)). We will define a functor
L(∆.∆′ {p∆,A })
L(⟨f, a⟩)
✲
L(∆.A.∆′ ),
where f and a are defined as follows.
f ∶= p∆,A ○ p∆.A,∆′ {p∆,A }
✲ ∆
∆.A.∆′ {p∆,A}
and
I
a = v∆,A.∆′ {p∆,A } ′
✲ ∆ {f } = ∆′ {p∆,A.∆′{p } } ∈ L(∆.A.∆′ {p∆,A }).
∆,A
8.&9. Int-Ty-Subst and Int-Tm-Subst, along a term ∆; ⋅ ⊢ a ∶ A, are interpreted by the functors L(⟨id∆ , a⟩) =
−{⟨id∆ , a⟩}. Indeed, let B ∈ L(∆.A.∆′ ) and a ∈ L(∆)(I, A). Then, we define the context ∆.∆′ [a/x]
as ∆.(∆′ {⟨id∆ , a⟩}) and the type B[a/x] as B{⟨f, a′ ⟩}, where
∆.∆′ {⟨id∆ , a⟩}.I
′
⟨f, a✲
⟩
∆.A.∆′
is defined from
∆.∆′ {⟨id∆ , a⟩}
p∆,∆′ {⟨id∆ ,a⟩}
✲ ∆
f
✲
⟨id∆ , a⟩
❄
∆.A
and
I
a′ ∶= v∆,∆′ {⟨id∆ ,a⟩} ′
✲ ∆ {f } = (∆′ {⟨id∆ , a⟩}){p∆,∆′{⟨id ,a⟩} }.
∆
10. Lin-Tm-Subst is interpreted by composition in L(∆). To be precise, given b ∈ L(∆)(⊗ Ξ ⊗ A, B)
and a ∈ L(∆)(⊗ Ξ′ , A), we define b[a/x] ∈ L(∆)(⊗ Ξ ⊗ ⊗ Ξ′ , B) as b ○ id⊗ Ξ ⊗ a.
Note that all our substitution rules are interpreted by functors and are therefore clearly associative.
The fact that Int-Var and Int-Weak define compatible (generalised) diagonals and projections is
reflected in the fact that diag and p obey generalised comonoid laws:
∆.A.∆′
diag∆,A,∆′
✲ ∆.A.∆′ .A{p∆,A.∆′ }
id
∆.
∆.A.∆′
A.
p∆.A.∆′,A{p∆,!A.∆′ }
∆′
✲ ❄
∆.A.∆′
diag∆,A,∆′
✲ ∆.A.∆′ .A{p∆,A.∆′ }
diag∆,A,∆′
❄
∆.A.∆′ .A{p∆,A.∆′ }
diag∆,A,∆′ .A{p∆,A.∆′ }
❄
✲ ∆.A.∆′ .A{p∆,A.∆′ }.A{p∆,A.∆′.A{p
},
∆,A.∆′ }
diag∆.A.∆′,A{p∆,A.∆′ },⋅
where we use p∆,A.∆′ as a shorthand notation for p∆,A ○ p∆.A,∆′ .
Finally, the model clearly supports I- and ⊗-types. We interpret I ∈ LType(∆) as the unit object in
L(∆) while its term ∗ is interpreted as the identity morphism. Similarly, we interpret ⊗ by the monoidal
product on the fibres: ∗ ∶= idI ∈ L(∆), let t be ∗ in a ∶= t ⊗ a, a ⊗ b is defined as the tensor product of
morphisms in L(∆), and let t be x ⊗ y in c ∶= c ○ (idΞ′ ⊗ t) (ignoring associatiators and unitors, here).
The C- and U-rules are immediate.
20
In fact, the converse is also true: we can build a category of this sort from the syntax of ILDTT.
̃ of ILDTT with I and ⊗-types defines a strict indexed
Theorem 7 (co-Soundness). A tautological model T
symmetric monoidal category with comprehension (C T , LT , pT , vT ).
Proof. The main technical difficulty in this proof will be that our category of contexts has context
morphisms as morphisms (corresponding to lists of terms of the type theory) while the type theory only
talks about individual terms. This exact difficulty is also encountered when proving completeness of the
categories with families semantics for ordinary DTT. It is sometimes fixed by (conservatively) extending
the the type theory to also talk about context morphisms explicitly. See e.g. [19].
1. We define ob(C T ) ∶= ICtxt, modulo α-equivalence, and write ∆.A ∶= ∆, x ∶ A. The designated
object will be ⋅ (from C-Emp), which will automatically become a terminal object because of our
definition of a morphism of C T (context morphism). Indeed, we define morphisms in C T , as follows,
by induction.
Start out defining C T (∆′ , ⋅) ∶= {⟨⟩} and for ∆ ∈ ICtxt that are not of the form ∆′′ .A, define
C T (∆′ , ∆) = {id∆ } if ∆′ = ∆ and C T (∆′ , ∆) = ∅ otherwise.
Then, by induction on the length n of ∆ = x1 ∶ A1 , . . . , xn ∶ An , we define
C T (∆′ , ∆.An+1 ) ∶= Σf ∈C T (∆′ ,∆) LTerm(∆′ , ⋅, An+1 [f /x]),
where An+1 [f /x] is defined, using Int-Ty-Subst, to be the (syntactic operation of) parallel substitution (see [5], section 2.4) of the list f1 , . . . , fn of linear terms ∆′ ; ⋅ ⊢ fi ∶ Ai [f1 /x1 , . . . , fi−1 /xi−1 ]
that f is made up out of, for the variables x1 , . . . , xn in ∆.
Note that, in particular, according to Int-Var, LTerm(A1 . . . . .An ., ⋅, Ai ) contains a term
derA1 .....Ai−1 ,Ai ,Ai+1 .....An , which allows us to define, inductively,
pA1 .....An ,⋅ ∶= ⟨⟩ ∈ C T (A1 . . . . .An , ⋅)
pA1 .....An ,A1 .....Ai ∶= pA1 .....An ,A1 .....Ai−1 , derA1 .....Ai−1 ,Ai ,Ai+1 .....An ∈ C T (A1 . . . . .An , A1 . . . . .Ai )
In particular, we define identities in C T from these: idA1 .....An ∶= pA1 .....An ,A1 .....An . We will also
use these ’projections’ in 3. to define the comprehension schema.
f =f1 ,...,fn
We define composition in C T by induction. Let B1 . . . . .Bm = ∆′
Ð→
∆ = A1 . . . . .An and
′′ g=g1 ,...,gm
′
∆
Ð→ ∆ . Then, we define, (f1 , . . . , fn−1 , fn ) ○ g ∶= (f1 , . . . , fn−1 ) ○ g, fn [g/x], where fn [g/x]
denotes the parallel substitution of g = g1 , . . . , gm for the free variables x1 , . . . , xm in fn , using IntTm-Subst. Note that associativity of composition comes from the associativity of substitution that
is implicit in the syntax while the identity morphism we defined clearly acts as a neutral element
for our composition.
2. Define ob(LT (∆)) ∶= LCtxt(∆) and LT (∆)(Ξ, Ξ′ ) ∶= LTerm(∆, Ξ, ⊗ Ξ′ ). Composition is defined
through Lin-Tm-Subst and ⊗-E. Identities are given by Lin-Var. The monoidal unit is given by
⋅ ∈ LCtxt(∆), while the monoidal product ⊗ on objects is given by context concatenation. The
monoidal product ⊗ on morphisms is given by ⊗-I. Note that the associators and unitors follow
from the associative and unital laws for the commutative monoid of contexts together with ⊗-C
and ⊗-U. (Note that the rules for ⊗ give us an isomorphism between an arbitrary context Ξ and
the one-type-context ⊗ Ξ, while the rules for I do the same for ⋅ and I.)
We define LT (f ) on objects by parallel substitution in each type in a linear context, via Int-TySubst, and on morphisms by parallel substitution, via Int-Tm-Subst. Note that functoriality is
given by implicit properties of the syntax like associativity of substitution. Note that this defines
a strong symmetric monoidal functor. We conclude that LT is a functor C Top Ð→ SMCat.
3. We define following comprehension schema on LT . Suppose ∆ ∈ C T and A ∈ LT (∆).
Define ∆.A Ð→ ∆ as p∆,A from 1. and I Ð→ A{p∆,A } (through Int-Var) as
derA ∈ LTerm(∆.A, ⋅, A) = LT (∆.A)(I, A{pT ∆,A }).
pT ∆,A
vT ∆,A
21
Suppose we are given ∆′ Ð→ ∆ and a ∈ LT (∆′ )(I, A{f }) = LTerm(∆′ , ⋅, A[f /c]). Then, by definition of the morphisms in C T , there is a unique morphism
⟨f, a⟩ ∶= f, a ∈ C T (∆′ , ∆.A) ∶= Σf ∈C T (∆′ ,∆) LTerm(∆′ , ⋅, A[f /x]) such that pT∆,A ○ ⟨f, a⟩ = f and
T
v∆,A
{⟨f, a⟩} = a. The uniqueness follows from the fact that pT∆,A ○ − and vT ∆,A {−} are the two
(dependent) projections of the Σ-type (in Set) that defines this homset.
f
Theorem 8 (Completeness). The construction described in ’co-Soundness’ followed by the one described
in ’Soundness’ is the identity (up to categorical equivalence12 ): i.e. strict indexed symmetric monoidal
categories with comprehension provide a complete semantics for ILDTT with I- and ⊗-types13 .
Proof. This is a trivial exercise.
Theorem 9 (Failure of co-Completeness). The construction described in ’Soundness’ followed by the
one described in ’co-Soundness’ is not equivalent to the identity: i.e. co-Completeness fails (as for the
categories with families semantics for DTT).
Proof. Indeed, if we start with a strict indexed symmetric monoidal category with comprehension, coñ and then construct its syntactic category, we effectively
struct the corresponding tautological model T
have thrown away all the non-trivial morphisms into objects that are not of the form ∆.A.
Of course, we can easily obtain a co-complete model theory by putting this extra restriction on our
models. Alternatively - this may be nicer from a categorical point of view -, we can take the obvious (see
e.g. [19]) conservative extension of our syntax by also talking about context morphisms (corresponding
to morphisms in our base category). In that case, we would obtain an actual internal language for strict
indexed symmetric monoidal categories with comprehension. This also has the advantage that we can
easily obtain an internal language for strict indexed monoidal categories by dropping the axioms Int-CExt, Int-C-Ext-Eq, and Int-Var, which correspond to the comprehension schema. We have not chosen
this route as it would mean that the syntax does not fit as well with what has been considered so far in
the syntactic tradition.
Corollary 1 (Relation with DTT and ILTT). A model (C, L, p, v) of ILDTT with I- and ⊗-types defines
a model I of DTT, that should be thought of the intuitionistic content of the linear type theory. This will
become even more clear through our treatment of !-types and in the examples we treat.
Moreover, it clearly defines a model of ILTT with I- and ⊗-types (i.e. a symmetric monoidal category)
in every context.
Conversely, it is easily seen that every syntactic model14 of DTT can be obtained this way, up to
equivalence, from a syntactic model of ILDTT (we take the same constants and axioms: effectively the
same theory but in a system without contraction and weakening) and that every model of ILTT can be
embedded in a model of ILDTT. (As we will see later, we can cofreely add type dependency on Set.)
Semantic Type Formers
Theorem 10 (Semantic type formers). For the other type formers, we have the following. A model of
ILDTT with I- and ⊗-types (a strict indexed symmetric monoidal category with comprehension)...
1. ...supports Σ-types iff all the pullback functors L(p∆,A ) have left adjoints Σ!A that satisfy the
12
The correct formal statement here would be that co-soundness followed by soundness (both of which define 2-functors
between the 2-category of tautological models of ILDTT and the 2-category of strict indexed symmetric monoidal categories
with comprehension) is 2-equivalent to the identity.
13 It is easy to see that, similarly, indexed symmetric multicategories with comprehension form a complete semantics for
ILDTT, possibly without I- and ⊗-types.
14
i.e. a model where we do not have any non-trivial morphisms into contexts that are not built from ⋅ by appending
types.
22
Beck-Chevalley condition1516 for pullback squares in C of the following form,17 .
∆′ .B{f }
qf,B
✲ ∆.B
(∗)
p∆′ ,B{f }
❄
∆′
f
p∆,B
❄
✲ ∆,
and that satisfy Frobenius reciprocity18 in the sense that the canonical morphism
Σ!A (Ξ′ {p∆,A} ⊗ B) Ð→ Ξ′ ⊗ Σ!A B
is an isomorphism , for all Ξ′ ∈ L(∆), B ∈ L(∆.A) .
2. ...supports Π-types iff all the pullback functors L(p∆,A ) have right adjoints Π!A that satisfy the
dual Beck-Chevalley condition for pullbacks of the form (∗).
3. ...supports ⊸-types iff L factors over SMCCat.
4. ...supports ⊺-types and &-types iff L factors over the category SMCCat of Cartesian categories with
a symmetric monoidal structure and their homomorphisms.
5. ...supports 0-types and ⊕-types iff L factors over the category dSMcCCat of co-Cartesian categories
with a distributive19 symmetric monoidal structure and their homomorphisms.
M∆
6. ...that supports ⊸-types20 , supports !-types iff all the comprehension functors L(∆) Ð→ I(∆)
L∆
have a left adjoint I(∆) Ð→ L(∆) in the 2-category SMCat of symmetric monoidal categories,
lax symmetric monoidal functors, and monoidal natural transformations21 and (compatibility with
15
Remember that the Beck-Chevalley condition for a pullback square
A
h
✲
f
B
k
❄
C
✲
g
❄
D
corresponds to the statement that the obvious morphism (from commuting of pullback square, unit, and counit) f! h∗ Ð→
≅
f! h∗ k ∗ k! Ð→ f! f ∗ g ∗ k! Ð→ g ∗ k! is an isomorphism, where we write f ∗ for L(f ) and f! for its left adjoint. In case f ∗
has a right adjoint (as for the Π-type), f∗ , we mean by the dual Beck-Chevalley condition that the obvious morphism
≅
g ∗ k∗ Ð→ g ∗ k∗ h∗ h∗ Ð→ g ∗ g∗ f∗ h∗ Ð→ f∗ h∗ is an iso.
16
The reader is encouraged to think of this condition as the equivalent for Σ- and Π-types of the condition on the
subtitution functors preserving the appropriate categorical structure for other type formers. It says that, in a sense, Σand Π-types are preserved under substitution.
17
Recall that qf,B ∶= ⟨f ○ p∆′ ,B{f } , v∆′ ,B{f } ⟩ and that this square is indeed a pullback.
18
Frobenius reciprocity expresses compatibility of Σ and ⊗, which is reasonable if we want a reading of Σ as a generalisation of ⊗. If one wants to drop Frobenius reciprocity in the semantics, it is easy to see that the equivalent in the syntax
is setting Ξ′ ≡ ⋅ in the Σ-E-rule.
19
Note that in the light of theorem 4, the demand of distributivity here is essentially the same phenomenon as the
demand of Frobenius reciprocity for Σ-types.
20 Actually, we only need this for the ’if’. The ’only if’ always holds. To make the ’if’ work, as well, in absence of ⊸-types,
we have to restrict !-E to the case where Ξ′ ≡ ⋅.
21
i.e. a symmetric lax monoidal left adjoint functor L∆ such that an inverse for its lax structure is given by the oplax
structure on L∆ coming from the lax structure on M∆ . Put differently, L∆ is a left adjoint functor to M∆ and is a strong
monoidal functor in a way that is compatible with the lax structure on M∆ .
23
f
substitution) for all ∆′ Ð→ ∆ ∈ C which makes L− into a morphism of indexed categories.:
L(f )
L(∆)
✻
✲ L(∆′ )
✻
L∆′
L∆
I(∆)
✲ I(∆′ ).
I(f ) = f ∗ (=pullback along f )
Then the linear exponential comonad !∆ ∶= L∆ ○ M∆ ∶ L(∆) Ð→ L(∆) will be our interpretation of
the comodality ! in the context ∆.
7. ... that supports ⊸-types, supports Id-types iff for all A ∈ ob L(∆), we have left adjoints Id!A ⊣
−{diag∆,A } that satisfy a Beck-Chevalley condition: Id!A{f } ○ L(qf,A ) Ð→ L(qqf,A ,A{p∆,A } ) ○ Id!A
is an iso.
diag∆,A ∶= ⟨id∆.A , v∆,A ⟩
✲ ∆.A.A{p∆,A }.
Here, ∆.A
Proof.
1. Assume our model supports Σ-types. We will show the claimed adjunction. The morphism
from left to right is provided by Σ-I. The morphism from right to left is provided by Σ-E. Σ-C and
Σ-U say exactly that these are mutually inverse. Naturality corresponds to the compatibility of
Σ-I and Σ-E with substitution.
c′
✲ c′ {p∆,A } ○ ⟨diag∆,A,⋅ , idB ⟩
L(∆)(Σ!A B, C) ✛≅
✲
L(∆.A)(B, C{p∆,A })
let z be !x ⊗ y in c ✛
c
We show how the morphism from left to right arises from Σ-I.
Int-Var
Lin-Var
∆; z ∶ Σ!x∶!A B ⊢ c′ ∶ C
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆, x ∶ A; w ∶ B ⊢ w ∶ B
Int-Weak
Σ-I
∆, x ∶ A; z ∶ Σ!x∶!A B ⊢ c′ ∶ C
∆, x ∶ A; w ∶ B ⊢!x ⊗ w ∶ Σ!x∶!A B
Lin-Tm-Subst
∆, x ∶ A; w ∶ B ⊢ c′ [!x ⊗ w/z] ∶ C
We show how the morphism from right to left is exactly Σ-E (with Ξ′ ≡ ⋅, Ξ ≡ z ∶ Σ!x∶!A B, t ≡ z).
Lin-Var
∆; z ∶ Σ!x∶!A B ⊢ z ∶ Σ!x∶!A B
∆, x ∶ A; y ∶ B ⊢ c ∶ C
Σ-E
∆; z ∶ Σ!x∶!A B ⊢ let z be !x ⊗ y in c ∶ C
We show how Frobenius reciprocity can be proved in our type system (particularly relying on
the form of the Σ-E-rule22 ).
∆; ⋅ ⊢ C type
Lemma 1 (Frobenius reciprocity). The canonical morphism
Σ!A (Ξ′ {p∆,A} ⊗ B) Ð→ Ξ′ ⊗ Σ!A B
f
is an isomorphism, for all Ξ′ ∈ L(∆), B ∈ L(∆.A).
Proof. We first show how to construct the morphism f we mean.
22
To be precise, we will see Frobenius reciprocity is validated because we allow dependency on Ξ′ in the Σ-E-rule.
Conversely, it is easy to see we can prove Frobenius reciprocity holds in our model if we have (semantic) ⊸-types, as this
allows us to remove the dependency on Ξ′ in Σ-E.
24
∆, x ∶ A; z ∶ Ξ′ ⊢ z ∶ Ξ′
∆; x′ ∶ Σ!x∶!A (Ξ′ ⊗ B) ⊢ x′ ∶ Σ!x∶!A (Ξ′ ⊗ B)
Lin-Var
∆, x ∶ A; ⋅ ⊢ x ∶ A
Int-Var
∆; y ∶ B ⊢ y ∶ B
∆, x ∶ A; y ∶ B ⊢!x ⊗ y ∶ Σ!x∶!A B
∆, x ∶ A; z ∶ Ξ′ , y ∶ B ⊢ z⊗!x ⊗ y ∶ Ξ′ ⊗ Σ!x∶!A B
Lin-Var
Lin-Var
Σ-I
⊗-I
⊗-E
∆, x ∶ A; w ∶ Ξ′ ⊗ B ⊢ let w be z ⊗ y in z⊗!x ⊗ y ∶ Ξ′ ⊗ Σ!x∶!A B
Σ-E
∆; x′ ∶ Σ!x∶!A (Ξ′ ⊗ B) ⊢ f ∶ Ξ′ ⊗ Σ!x∶!A B
We now construct its inverse. Call it g.
∆; y1 ∶ Ξ′ ⊢ y1 ∶ Ξ′
Lin-Var
Lin-Var
∆; y ∶ B ⊢ y ∶ B
⊗-I
Int-Var
∆, x ∶ A; y1 ∶ Ξ′ , y ∶ B ⊢ y1 ⊗ y ∶ Ξ′ ⊗ B
∆, x ∶ A; ⋅ ⊢ x ∶ A
Lin-Var
∆, x ∶ A; y1 ∶ Ξ′ , y ∶ B ⊢!x ⊗ y1 ⊗ y ∶ Σ!x∶!A (Ξ′ ⊗ B)
∆; y2 ∶ Σ!x∶!A B ⊢ y2 ∶ Σ!x∶!A B
Σ-E
∆; y1 ∶ Ξ′ , y2 ∶ Σ!x∶!A B ⊢ let y2 be !x ⊗ y in !x ⊗ y1 ⊗ y ∶ Σ!x∶!A (Ξ′ ⊗ B)
⊗-E
′
′
′
∆; y ∶ Ξ ⊗ Σ!x∶!A B ⊢ g ∶ Σ!x∶!A (Ξ ⊗ B)
Σ-I
23
We leave it to the reader to verify that these morphisms are mutually inverse in the sense that
∆; x′ ∶ Σ!x∶!A (Ξ′ ⊗B) ⊢ g[f /y ′ ] ≡ x′ ∶ Σ!x∶!A (Ξ′ ⊗B) and ∆; y ′ ∶ Ξ′ ⊗Σ!x∶!A B ⊢ f [g/x′ ] ≡ y ′ ∶ ⊗Σ!x∶!A B.
For the converse, we show how to obtain Σ-I from our morphism from left to right:
Lin-Var
∆; z ∶ Σ!x∶!AB ⊢ z ∶ Σ!x∶!A B
”left to right”
∆, x ∶ A; w ∶ B ⊢!x ⊗ w ∶ Σ!x∶!A B
∆; w ∶ B ⊢!a ⊗ w ∶ Σ!x∶!A B
∆; ⋅ ⊢ a ∶ A
Int-Tm-Subst
∆; Ξ ⊢!a ⊗ b ∶ Σ!x∶!A B
∆; Ξ ⊢ b ∶ B[a/x]
Lin-Tm-Subst
We show how to obtain Σ-E from our morphism from right to left, using Frobenius reciprocity. In
particular, note that we don’t need Frobius reciprocity.
∆, x ∶ A; y ∶ Ξ′ , B ⊢ c ∶ C
∆; ⋅ ⊢ C type
∆, x ∶ A; y ∶ Ξ′ ⊗ B ⊢ c ∶ C
∆; z ∶ Σ!x∶!A (Ξ′ ⊗ B) ⊢ let z be !x ⊗ y in c ∶ C
∆; z ∶ (Ξ′ ⊗ Σ!x∶!A B) ⊢ let z be !x ⊗ y in c ∶ C
⊗-E
”right to left”
Frobenius reciprocity
∆; z1 ∶ Ξ′ , z2 ∶ Σ!x∶!A B ⊢ let z1 ⊗ z2 be !x ⊗ y in c ∶ C
Lin-Tm-Subst,⊗-I,2×Lin-Var
∆; Ξ ⊢ t ∶ Σ!x∶!A B
∆; z1 ∶ Ξ′ , Ξ ⊢ (let z1 ⊗ z2 be !x ⊗ y in c)[t/z2 ] ∶ C
Lin-Tm-Subst
As usual, the Beck-Chevalley condition says precisely that Σ-types commute with substitution,
as dictated by the type theory.
2. Assume our model supports Π-types. We will show the claimed adjunction. The morphism from
left to right is provided by Π-I - in fact, it is exactly the I-rule - and the one from right to left
by Π-E. Π-C and Π-U say exactly that these are mutually inverse. Naturality corresponds to the
compatibility of Π-I and Π-E with substitution.
✲ λ!x∶!A b
b
✲
L(∆.A)(Ξ{p∆.A }, B) ✛
≅ L(∆)(Ξ, Π!x∶!A B)
f (!x) ✛
f.
We show how we obtain the definition of f (!x) from Π-E.
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆; Ξ ⊢ f ∶ Π!x∶!A B
Int-Weak
∆, x ∶ A; Ξ ⊢ f ∶ (Π!x∶!A B)
Π-E
∆, x ∶ A; Ξ ⊢ f (!x) ∶ B
Int-Var
For the converse, we have to show that we can recover Π-E from the definition of f (!x).
23
At the use of Σ-E, is really where the Frobenius reciprocity comes in, because of the factor Ξ′ in the Σ-E-rule.
25
∆; Ξ ⊢ f ∶ Π!x∶!A B
Definition f (x)
∆; ⋅ ⊢ a ∶ A
∆, x ∶ A; Ξ ⊢ f (!x) ∶ B
Int-Tm-Subst
∆; Ξ ⊢ f (!x)[a/x] ∶ B[a/x]
∆; Ξ ⊢ f (!a) ∶ B[a/x]
This shows that individual Π-types correspond to right adjoint functors to substitution along projections. The type theory dictates that Π-types interact well with substitution. This corresponds
to the dual Beck-Chevalley condition, as usual.
3. From the categorical semantics of (non-dependent) linear type theory (see e.g. [20] for a very complete account) we know that ⊸-types correspond to monoidal closure of the category of contexts.
The extra feature in dependent linear type theory is that the syntax dictates that the type formers
are compatible with substitution. This means that we also have to restrict the functors L(f ) to
preserve the relevant categorical structure.
4. Idem.
5. Idem.
6. Assume that we have !-types. We will define a left adjoint L∆ ⊣ M∆ as L∆ p∆,A ∶=!A (this is
easily seen to be well-defined up to isomorphism, so we can use AC for a definition on the nose)
and, noting that every morphism p∆,A Ð→ p∆,B in C/∆ is of the form ⟨p∆,A , b⟩ for some unique
I Ð→ B{p∆,A } ∈ L(∆.A), we define L∆ as acting on b as the map obtained from
∆, x ∶ A; ⋅ ⊢ b ∶ B
Lin-Var
!-I
∆; y ∶!A ⊢ y ∶!a
∆, x ∶ A; ⋅ ⊢!b ∶!B
!-E
∆; y ∶!A ⊢ let y be !x in !b ∶!B
which indeed gives us L∆ (⟨p∆,A , b⟩) ∈ L(∆)(!A, !B). Note that L∆ is strong monoidal, as the rules
for ! define a natural bijection between terms ∆, x ∶ A, y ∶ B; ⋅ ⊢ t ∶ C and ∆; x′ ∶!A, y ′ ∶!B ⊢ t′ ∶ C.
In terms of the model, this gives a natural bijection L(∆)(L∆ (M∆ A × M∆B), C) ≅ L(∆)(!D, C) ≅
L(∆.D)(I, C) ≅ L(∆.A.B)(I, C) ≅ L(∆)(!A⊗!B, C), where we write D for an object such that
M∆ D = M∆ A × M∆ B (which exists if the product exists), so strong monoidality follows by the
Yoneda lemma. (A keen reader can verify that the oplax structure on L∆ corresponds with the lax
structure on M∆ .)
b
We exhibit the adjunction by the following isomorphism of hom-sets, where the morphism from
left to right comes from !-I and the one from right to left comes from !-E.
✲ b[!x/x′ ]
b
✲
L(∆)(L∆ p∆,A , B) = L(∆)(!A, B) ✛
≅ L(∆.A)(I, B{p∆,A }) ≅ C/∆(p∆,A , p∆,B ) = I(∆)(p∆,A , M∆ B)
let y be !x in b′
✛
b′
We show how to construct the morphism from left to right, using !-I.
Int-Var
∆; x′ ∶!A ⊢ b ∶ B
∆, x ∶ A; ⋅ ⊢ x ∶ A
Int-Weak
!-I
′
∆, x ∶ A; x ∶!A ⊢ b ∶ B
∆, x ∶ A; ⋅ ⊢!x ∶!A
Lin-Tm-Subst
∆, x ∶ A; ⋅ ⊢ b[!x/x′ ] ∶ B
We show to construct the morphism from right to left, using !-E. Suppose we’re given b′ ∈
L(∆.A)(I, B{p∆,A }). From this, we produce a morphism in L(∆)(!A, B) as follows.
Lin-Var
∆; y ∶!A ⊢ y ∶!A
∆, x ∶ A; ⋅ ⊢ b′ ∶ B
!E
∆; y ∶!A ⊢ let y be !x in b′ ∶ B
We leave it up to the reader to verify that these morphisms are mutually inverse, according to
!-C and !-U.
Conversely, suppose we have a strong monoidal left adjoint L∆ ⊣ M∆ . We define, for A ∈ ob(L(∆)),
!A ∶= L∆ M∆ (A).
26
We verify that !-I can be derived from the homset morphism from left to right:
Lin-Var
∆; x′ ∶!A ⊢ x′ ∶!A
”left to right”
∆, x ∶ A; ⋅ ⊢!x ∶!A
∆; ⋅ ⊢!x[a/x] ∶!A
∆; ⋅ ⊢ a ∶ A
Int-Tm-Subst
We verify that, in presence of ⊸-types, !-E can be derived from the homset morphism from right
to left:
∆; w ∶ Ξ′ ⊢ w ∶ Ξ′
∆, x ∶ A; y ∶ Ξ′ ⊢ b ∶ B
⊸-I
∆, x ∶ A; ⋅ ⊢ λy∶Ξ′ b ∶ Ξ′ ⊸ B
Lin-Var
∆; z ∶!A, Ξ′ ⊢ let z be !x in b[w/y] ∶ B
∆; Ξ ⊢ t ∶!A
”right to left”
∆; z ∶!A ⊢ let z be !x in λy∶Ξ′ b ∶ Ξ′ ⊸ B
∆; Ξ, Ξ′ ⊢ let t be !x in b[w/y] ∶ B
⊸-E
Lin-Tm-Subst
Note that the !-C- and !-U-rules correspond precisely to the fact that our morphisms from left
to right and from right to left define a homset isomorphism.
Finally, it is easily verified that the condition that L(f ) ○ L∆ ≅ L∆′ ○ L(f ) corresponds exactly
to the compatibility of ! with substitution.
7. Suppose we have Id!A ⊣ −{diag∆,A }, i.e. we have a (natural) homset isomorphism
✲
L(∆.A.A{p∆,A })(Id!A (B), C) ✛
≅ L(∆.A)(B, C{diag∆,A }).
The claim is that Id!A (I) satisfies the rules for the Id-type of A (or maybe it would be more appropriate to say of !A). Indeed, we have Id-I as follows.
∆, x ∶ A, x′ ∶ A; w ∶ Id!A (I)(x, x′ ) ⊢ w ∶ Id!A (I)(x, x′ )
∆, x ∶ A; y ∶ I ⊢
refly
!x
∶ Id!A (I)(x, x)
Int-Tm-Subst
Lin-Var
”left to right”
I-I
∆, x ∶ A; ⋅ ⊢ ∗ ∶ I
Lin-Tm-Subst
∆, x ∶ A; ⋅ ⊢ refl!x ∶ Id!A (I)(x, x)
∆; ⋅ ⊢ a ∶ A
∆; ⋅ ⊢ refl!x ∶ Id!A (I)(a, a)
We obtain Id-E as follows. Let ∆, x ∶ A, x′ ∶ A; ⋅ ⊢ C type.
∆, x ∶ A; B ⊢ c ∶ C[x/x′ ]
∆, x ∶ A, x′ ∶ A; Id!A (B) ⊢ c′ ∶ C
Lin-Tm-Subst
”right to left”
∆; ⋅ ⊢ a′ ∶ A
∆; ⋅ ⊢ a ∶ A
Int-Tm-Subst
∆; Id!A (B)[a/x, a′ /x′ ] ⊢ c′ [a/x, a′ /x′ ] ∶ C[a/x, a′ /x′ ]
(*)
∆; B ′ ⊢ p ∶ Id!A (I)[a/x, a′ /x′ ]
∆; B[a/x], B ′ ⊢ p′ ∶ Id!A (B)[a/x, a′ /x′ ]
∆; B[a/x], B ′ ⊢ let (a, a′ , p) be (z, z, refl !z ) in c ∶ C[a/x, a′ /x]
Here, (∗) is a slightly non-trivial step that follows immediately when we note that Id!A (B) ≅
Id!A (I) ⊗ B{p∆,A } (by tensoring with idB ). Indeed,
L(∆.A.A{p∆,A })(Id!A (B), C) ≅ L(∆.A)(B, C{diag∆,A })
≅ L(∆.A)(I, (B{p∆,A } ⊸ C){diag∆,A })
≅ L(∆.A.A{p∆,A })(Id!A (I), B{p∆,A } ⊸ C)
≅ L(∆.A.A{p∆,A })(Id!A (I) ⊗ B{p∆,A }, C).
Since all these isomorphisms are natural in C, the Yoneda lemma says that Id!A (B) ≅ Id!A (I) ⊗
B{p∆,A }.
Conversely, suppose we have Id-types. Then, define Id!A (B) ∶= Id!A ⊗ B{p∆,A }, with the obvious extension on morphisms. Then, we obtain the morphism ”left to right” as follows.
∆, x ∶ A, x′ ∶ A; z ∶ Id!A , y ∶ B ⊢ c ∶ C
′
∆, x ∶ A; ⋅ ⊢ x ∶ A
′
′
∆, x ∶ A; z ∶ Id!A [x/x ], y ∶ B ⊢ c[x/x ] ∶ C[x/x ]
Int-Var
Int-Tm-Subst
∆, x ∶ A; y ∶ B ⊢ c′ ∶ C[x/x′ ]
Int-Var
Id-I
∶ Id!A (x, x)
Lin-Tm-Subst
∆, x ∶ A; ⋅ ⊢ x ∶ A
∆, x ∶ A; ⋅ ⊢ refl!x
The morphism ”right to left” is obtained as follows.
∆, x0 ∶ A; y ∶ B ⊢ c ∶ C[x0 /x1 ]
Lin-Var
∆, x0 ∶ A, x1 ∶ A; w ∶ Id!A ⊢ w ∶ Id!A
∆, x0 ∶ A, x1 ∶ A; w ∶ Id!A , y ∶ B ⊢ c′ ∶ C
27
∆, x0 ∶ A, x1 ∶ A; ⋅ ⊢ xi ∶ A
Int-Var
Id-E
We leave it to the reader to verify that the Id-C- and Id-U-rules translate precisely into the ”right
to left” and ”left to right” morphisms being inverse.
The semantics of ! suggests an alternative definition for the notion of a comprehension: if we Σ-types
in a strong sense, it is a derived notion!
Theorem 11 (Lawvere Comprehension). Given a strict indexed monoidal category (C, L) with left ad-
joints ΣLf to L(f ) for arbitrary ∆′ Ð→ ∆ ∈ C, then we can define C/∆ Ð→ L(∆) by
f
L∆
L∆ (−) ∶= ΣL− I.
In that case, (C, L) has a comprehension schema iff L∆ has a right adjoint M∆ (which then automatically
satisfies M∆′ ○ L(f ) = L(f ) ○ M∆ for all ∆′ Ð→ ∆ ∈ C). That is, our notion of comprehension generalises
that of [21].
Finally, if ΣLf are demanded to satisfy the Beck-Chevalley condition and Frobenius reciprocity, then
(C, L) satisfies the comprehension schema iff it admits !-types.
f
Proof. Suppose that we have said right adjoints M∆ . We will construct a comprehension schema.
We define p∆,A ∶= M∆ (A) and
≅
≅
L(∆′ )(I, A{f }) ✲ L(∆)(ΣLf I∆′ , A) = L(∆)(L∆ f, A) ✲ C/∆(f, M∆ A)
✲ ⟨f, a⟩,
✲ af
a
where the first natural isomorphism comes from the adjunction ΣLf ⊣ −{f } and the second one comes
from the adjunction L∆ ⊣ M∆ . Note that, by definition, p∆,A ⟨f, a⟩ = f .
In particular, we obtain a unique v∆,A ∈ L(∆.A)(I, A{p∆,A }) inducing idM∆ as ⟨p∆,A , v∆,A ⟩. Finally, the Yoneda lemma (i.e. naturality of these isomorphisms) says that v∆,A {⟨f, a⟩} = a.
Conversely, suppose we’re given a comprehension schema. Then, we know, by theorem 5, that we
can define a comprehension functor M∆ such that M∆′ ○ L(f ) = L(f ) ○ M∆ . Then we have the following:
≅
≅
C/∆(f, M∆ A) ✲ L(∆′ )(I, A{f }) ✲ L(∆)(ΣLf I∆′ , A) = L(∆)(L∆ f, A)
⟨f, a⟩
✲ a
✲ af ,
where the first isomorphism is precisely the representation defined by our comprehension and the second
isomorphism comes from the fact that ΣLf ⊣ −{f }.
Finally, the following calculation shows that it follows from Frobenius reciprocity and Beck-Chevalley
that L∆ is strong monoidal:
L∆ (f ) ⊗ L∆ (g) = (ΣLf Idomf ) ⊗ (ΣLg Idomg )
= ΣLg ((ΣLf Idomf ){g} ⊗ Idomg ) (Frobenius reciprocity)
= ΣLg ((ΣLf Idomf ){g})
= ΣLg ΣLf ∗ g Idomf {g ∗ f } (Beck-Chevalley)
= ΣLg ΣLf ∗ g Idomf ×g
= ΣL(f ×g) Idomf ×g
= L∆ (f × g).
Theorem 12 (Type Formers in I). I supports Σ-types iff ob(I) is closed under compositions (as
morphisms in C). It supports Id-types iff ob(I) is closed under post-composition with maps diag∆,A . If
L supports !- and Π-types, then I supports Π-types. Moreover, we have that
Σ!A !B ≅ L(ΣMA M B)
Id!A (!B) ≅ LIdMA (M B)
28
M Π!B C ≅ ΠMB M C.
Proof. We write out the adjointness condition
I(∆)(Σp∆,B f, p∆,D ) ≅ I(∆.B)(f, p∆,D {p∆,B })
!
≅ I(∆.B)(f, p∆,D{p∆,B } )
≅ L(∆.B.C)(I, D{p∆,B }{f })
≅ L(∆.B.C)(I, D{p∆,B ○ f })
≅ I(∆)(p∆,B ○ f, p∆,D ).
Now, the Yoneda lemma gives us that Σp∆,B f = p∆,B ○ f .
Similarly,
I(∆.A.A)(Idp∆,A (f ), p∆.A.A,C ) ≅ I(∆.A)(f, p∆.A.A,C {diag∆,A })
!
≅ L(∆.A.B)(I, C{diag∆,A }{f })
≅ L(∆.A.B)(I, C{diag∆,A ○ f })
≅ I(∆.A.A)(diag∆,A ○ f, p∆.A.A,C ),
so diag∆,A ○ f models Idp∆,A (f ).
Finally,
I(∆)(M∆ D, Πp∆,B p∆.B,C ) ≅ I(∆.B)((M∆ D){p∆,B }, p∆.B,C )
!
≅ I(∆.B)((M∆ D){p∆,B }, M∆.B C)
≅ L(∆.B)(L∆.B ((M∆ D){p∆,B }), C)
≅ L(∆.B)((L∆ M∆ D){p∆,B }, C)
≅ L(∆)(L∆ M∆ D, Π!B C)
≅ I(∆)(M∆ D, M∆ Π!B C)
Again, using the Yoneda lemma, we conclude that M∆ Π!B C models ΠM∆ B M∆.B C.
In all cases, we have not worried about Beck-Chevalley (and Frobenius reciprocity for Σ-types) as they
are trivially seen to hold.
Note that if L has !- and Σ-types, then
L(∆)(L∆ (ΣM∆ A M∆.A B), C) ≅ I(∆)(ΣM∆ A M∆.A B, M∆ C)
≅ I(∆.A)(M∆.A B, (M∆ C){p∆,A })
≅ I(∆.A)(M∆.A B, M∆ (C{p∆,A }))
≅ L(∆.A)(!B, C{p∆,A })
≅ L(∆)(Σ!A !B, C).
By the Yoneda lemma, conclude that Σ!A !B ≅ L∆ (ΣM∆ A M∆.A B).
Note that, in case L admits !- and Id-types,
L(∆.A.A)(Id!A (!B), C) ≅ L(∆.A)(!B, C{diag ∆,A })
≅ I(∆.A)(M∆.A B, M∆.A (C{diag∆,A }))
≅ I(∆.A)(M∆.A B, M∆.A.A (C){diag∆,A })
≅ I(∆.A.A)(diag ∆,A ○ M∆.A B, M∆.A (C))
≅ I(∆.A.A)(IdM∆ A (M∆.A B), M∆.A (C))
≅ L(∆.A.A)(L∆.A.A IdM∆ A (M∆.A B), C).
29
We conclude that Id!A (!B) = L∆.A.A IdM∆ A (M∆.A B) and in particular Id!A (I) ≅ L∆.A.AIdM∆ A (id∆.A ).
The last statement is easily seen to also be valid in absence of ⊺-types.
Remark 6 (Dependent Seely Isomorphisms?). Note that, in our setup, we have a version of the normal
Seely isomorphisms in each fibre. Indeed, suppose L supports ⊺-, &, and !-types. Then, M∆ (⊺) = id∆
and M∆ (A&B) = M∆ (A) × M∆ (B), as M∆ has a left adjoint and therefore preserves products. Now,
L∆ is strong monoidal and !∆ = L∆ M∆ , so it follows that !∆ ⊺ = I and !∆ (A&B) =!∆ A⊗!∆ B.
Now, theorem 12 suggests the possibility of similar Seely isomorphisms on us for Σ-types and Idtypes. Indeed, I supports Σ-types iff we have additive Σ-types in L in the sense of objects Σ&
A B such
that
&
⊗
M Σ&
A B ≅ ΣMA M B and hence !ΣA B ≅ Σ!A !B,
where we suggestively write Σ⊗ for the usual multiplicative Σ-type in L. In an ideal world, one would
hope that Σ&
A B generalise A&B in the same way that Σ!A B is a dependent generalisation of !A ⊗ B.
Similarly, we get a notion of additive Id-types: I supports Id-types iff we have objects Id&
A (B) in L
such that
&
⊗
M Id&
A (B) ≅ IdMA (M B) and hence !IdA (B) ≅ Id!A (!B),
writing Id⊗ for the usual (multiplicative) Id-type in L. Note that this suggests that, in the same way
⊗
&
that Id⊗
!A (B) ≅ Id!A (I) ⊗ B (a sense in which usual Id-types are multiplicative connectives), IdA (B) ≅
&
&
IdA (⊺)&B. In fact, if we have ⊺- and &-types, we only have to give IdA (⊺) and can then define
&
Id&
A (B) ∶= IdA (⊺)&B to obtain additive Id-types in generality.
A fortiori, if some M∆ is essentially surjective, we obtain such additive Σ- and Id-types in the fibre
over ∆. In particular, we are in this situation if L⋅ ⊣ M⋅ is the usual co-Kleisli adjunction of !⋅ , where
I(⋅) = C. This shows that if we are hoping to obtain a model of ILDTT indexed over the co-Kleisli
category, in the natural way, we need to support these additive connectives.
It still remains to be seen if an understanding of these “additive connectives” can be found from
a syntactic point of view. Similarly, it seems like the natural models of ILDTT do not support them.
Finally, it is difficult to come up with an intuitive interpretation of the meaning of such connectives, in
the sense of a resource interpretation. Clearly, further investigation is necessary, here.
30
4
Some Discrete Models: Monoidal Families
We discuss a simple class of models in terms of families with values in a symmetric monoidal category. On
a logical level, what the construction boils down to is starting with a model V of a linear propositional
logic and taking the cofree linear predicate logic on Set with values in this propositional logic. This
important example illustrates how Σ- and Π-types can represent infinitary additive disjunctions and
conjunctions. The model is discrete in nature, however, and in that respect not representative for the
type theory.
Suppose V is a symmetric monoidal category. We can then consider a strict Set-indexed category,
defined through the following enriched Yoneda embedding Fam(V) ∶= V − ∶= SMCat(−, V):
Setop
Fam(V)
✲ SMCat
′
−○f
✲ V S ←Ð
VS .
f
S Ð→ S ′
Note that this definition naturally extends to a functor Fam.
Theorem 13 (Families Model ILDTT). The construction Fam adds type dependency on Set cofreely in
the sense that it is right adjoint to the forgetful functor ev1 that evaluates a model of linear dependent
Setop
type theory at the empty context to obtain a model of linear propositional type theory (where SMCatcompr
op
is the full subcategory of SMCatSet on the objects with comprehension):
SMCat
ev1
✛
op
✲ SMCatSet
compr .
⊂
Fam
Proof. Fam(V) admits a comprehension, by the following isomorphism
Fam(V)(S)(I, B{f }) = V S (I, B ○ f )
= Πs∈S V(I, B(f (s)))
≅ Set/S(S Ð→ S, Σs∈S V(I, B(f (s))) Ð→ S)
idS
fst
≅ Set/S ′ (S Ð→ S ′ , Σs′ ∈S ′ V(I, B(s′ )) Ð→ S ′ )
f
fst
= Set/S ′ (f, pS ′ ,B ),
where pS ′ ,B ∶= Σs′ ∈S ′ V(I, B(s′ )) Ð→ S ′ . (vS ′ ,B is obtained as the image of idS ′ ∈ Set/S ′ under this
s
isomorphism.) To see that ev1 ⊣ Fam, note that the following naturality diagrams for elements 1 Ð→ S
fst
1
ev1 (L) = L(1)
φ1 ✲
V = Fam(V)(1)
✻
✻
s
−○s
−{s}
❄
S
L(S)
φS
✲ V S = Fam(V)(S)
s
together with the fact that all 1 Ð→ S are jointly surjectiveopand hence the fact that − ○ s are jointly
injective means that a natural transformation φ ∈ SMCatSet (L, Fam(V)) is uniquely determined by
φ1 ∈ Cat(ev1 (L), V).
We have the following results for type formers24 .
Theorem 14 (Type Formers for Families). V has small coproducts that distribute over ⊗ iff Fam(V)
supports Σ-types. In that case, Fam(V) also supports 0- and ⊕-types (which correspond precisely to finite
distributive coproducts).
24
We do not examine Id-types here, as they will precisely correspond to intuitionistic identity type in I, which probably
is not very interesting seeing that I is a submodel of the normal set-based model of dependent types (i.e. fibred sets, which
is equivalent to indexed sets: Set-valued families).
31
V has small products iff Fam(V) supports Π-types. In that case, Fam(V) also supports ⊺- and &-types
(which correspond precisely to finite products).
Fam(V) supports ⊸-types iff V is monoidal closed.
Fam(V) supports !-types iff V has small coproducts of I that are preserved by ⊗ in the sense that
the canonical morphism coprodS (Ξ′ ⊗ I) Ð→ Ξ′ ⊗ coprodS I is an isomorphism for any Ξ′ ∈ ob V and
S ∈ ob Set. In particular, if Fam(V) supports Σ-types, then it also supports !-types.
Fam(V) supports Id-types if V has an initial object. Supposing that V has a terminal object, the only
if also holds.
Proof. The statement about 0-, ⊕-, ⊺-, and &-types should be clear from the previous sections, as
products and coproducts in V S are pointwise (and hence automatically preserved under substitution).
We will denote coproducts in V with ⊕. Then,
Πs′ ∈S ′ V(
⊕
s∈f −1 (s′ )
A(s), B(s′ )) ≅ Πs′ ∈S ′ Πs∈f −1 (s′ ) V(A(s), B(s′ ))
≅ Πs∈Σs′ ∈S′ f −1 (s′ ) V(A(s), B(f (s)))
≅ Πs∈S V(A(s), B(f (s))
= V S (A, B ○ f ).
So, we see that we can define ΣLf (A)(s′ ) ∶= ⊕s∈f −1 (s′ ) A(s) to get a left adjoint ΣLf ⊣ −{f }, if we have
coproducts. (With the obvious definition on morphisms coming from the coCartesian monoidal structure
on V.) Conversely, we can clearly use ΣLf to define any coproduct by using, for instance, an identity
function for f on the set we want to take a coproduct over and a family A that denotes the objects
we want to comultiply. The Beck-Chevalley condition is taken care of by the fact that our substitution
morphisms are given by precomposition. Frobenius reciprocity precisely corresponds to distributivity of
the coproducts over ⊗.
˘
Similarly, if V has products, we will denote them ˘
with
to suggest the connections with linear type
theory. In that case, we can define ΠLf (A)(s′ ) ∶= s∈f −1 (s′ ) A(s) to get a right adjoint −{f } ⊣ ΠLf .
(With the obvious definition on morphisms coming from the Cartesian monoidal structure on V.) Indeed,
¯
A(s)) ≅ Πs′ ∈S ′ Πs∈f −1 (s′ ) V(B(s′ ), A(s))
Πs′ ∈S ′ V(B(s′ ),
s∈f −1 (s′ )
≅ Πs∈Σs′ ∈S′ f −1 (s′ ) V(B(f (s)), A(s))
≅ Πs∈S V(B(f (s)), A(s))
= V S (B ○ f, A).
Again, in the same way as before, we can construct any product using ΠLf . The dual Beck-Chevalley
condition comes for free as our substitution morphisms are precomposition.
The claim about ⊸-types follows immediately from the previous section: Fam(V) supports ⊸-types
iff all its fibres have a monoidal closed structure that is preserved by the substitution functors. Seeing
that our monoidal structure is pointwise, the same will hold for any monoidal closed structure. Seeing
that substitution is given by precomposition the preservation requirement comes for free.
The characterisation of !-types is given by 3, which tells us we can define !A ∶= ΣpS′ ,A I = s′ ↦
⊕V(I,A(s′ )) I and conversely.
Finally, for Id-types, note that the adjointness condition Id!A ⊣ −{diag∆,A } boils down to the requirement (*)
Πs∈S Πa∈A(s) V(B(s, a), C(s, a, a)) ≅ V Σs∈S A(s) (B, C{diagS,A })
≅ V Σs∈S A(s)×A(s) (Id!A (B), C)
!
≅ Πs∈S Πa∈A(s) Πa′ ∈A(s) V(Id!A (B)(s, a, a′ ), C(s, a, a′ )).
We see that if we have an initial object 0 ∈ ob(V), we can define
Id!A (B)(s, a, a′ ) ∶= {
32
B(s, a) if a = a′
0 else
For a partial converse, suppose we have a terminal object ⊺ ∈ V. Let V ∈ ob(V). Let S ∶= {∗}, A ∶= {0, 1}
and C s.t. C(0, 0) = C(1, 1) = C(0, 1) = ⊺ and C(1, 0) = V . Then, (*) becomes the condition that
{∗} ≅ V(Id!A (B)(1, 0), V ). We conclude that Id!A (B)(1, 0) is initial in V.
Remark 7. Note that an obvious way to guarantee distributivity of coproducts over ⊗ is by demanding
that V is monoidal closed.
Remark 8. It is easily seen that Σ-types in I, or additive Σ-types in L = Fam(V), boil down to having
an object ors∈S C(s) ∈ ob(V) for a family (C(s) ∈ ob(V))s∈S such that Σs∈S V(I, C(s)) ≅ V(I, ors∈S C(s)).
Similarly, Id-types in I, or additive Id-types in L, boil down to having objects one, zero ∈ ob(V) such
that V(I, one) ≅ 1 and V(I, zero) = 0.
Two particularly simple concrete examples of V come to mind that can accomodate all type formers
and form a nice illustration: a category V = VectF of vector spaces over a field F , with the tensor product,
and the category V = Set∗ of pointed sets, with the smash product. All type formers get their obvious
interpretation, but let us stop to think about ! for a second as it is a novelty of ILDTT that it gets uniquely
determined by the indexing, while in propositional linear type theory we might have several different
choices. In the first example, ! boils down to the following: (!B)(s′ ) = coprodVectF (F,B(s′ )) F ≅ ⊕B(s′ ) F ,
i.e. taking the vector space freely spanned by all the vectors. In the second example, (!B)(s′ ) =
coprodSet∗ (2∗ ,B(s′ )) 2∗ = ⋁B(s′ ) 2∗ = B(s′ ) + {∗}, i.e. ! freely adds a new basepoint. These models show
the following.
Theorem 15 (DTT, DILL⊊ILDTT). ILDTT is a proper generalisation of DTT and DILL: we have
inclusions of the classes of models DTT,DILL⊊ILDTT.
Proof. By the Grothendieck construction, every split fibration can equivalently be seen as the category
of elements of the corresponding strict indexed category defined by the fibres. Under this equivalence
split full comprehension categories with finite fibrewise products (i.e. models of DTT with 1- and ×types) correspond precisely to strict indexed Cartesian monoidal categories with comprehension where
the comprehension functor is full and faithful. Clearly, these are a special case of our notion of model of
ILDTT. Moreover, in such cases clearly !A ≅ A. From their categorical description, it is also clear that
the other connectives of ILDTT reduce to those of DTT. This proves the inclusion DDT⊂ILDTT.
The models described above are clearly more general than those of DTT, as we are dealing with a
non-Cartesian monoidal structure on the fibre categories. This proves that the inclusion is proper.
We have seen that the Fam-construction realises the category of models of DILL as a reflective
subcategory of the category of models of ILDTT. Moreover, from various non-trivial models of DTT
indexed over other categories than Set it is clear that this inclusion is proper as well.
Finally, we note that these inclusions still remain valid in the sub-algebraic setting where we do not
have I- and ⊗-types. A simple variation of the argument using multicategories rather than monoidal
categories does the trick.
Although this class of families models is important, it is clear that it only represents a very limited
part of the generality of ILDTT: not every model of ILDTT is either a model of DTT or of DILL.
Hence, we are in need of models that are less discrete in nature but still linear, if we are hoping to
observe interesting new phenomena arising from the connectives of linear dependent type theory. Some
suggestions and work in progress will be discussed in the next section.
33
5
Conclusions and Future Work
We hope to have convinced the reader that linear dependent types fit very naturally in the landscape
of existing type theories and that they admit a rich theory rather than being limited to the specific
examples that had been considered so far. There is a larger story connecting these examples!
On a syntactic level our system is a very natural blend between (intuitionistic) dependent type theory
and dual intuitionistic linear logic. On a semantic level, if one starts with the right notion of model for
dependent types, the linear generalisation is obtained through the usual philosophy of passing from
Cartesian to symmetric monoidal structures. The resulting notion of a model forms a natural blend
between comprehension categories, modelling DTT, and linear-non-linear models, modelling DILL.
It is very pleasing to see that all the syntactically natural rules for type formers are equivalent to their
semantic counterparts that would be expected based on the traditions of categorical logic of dependent
types and linear types. In particular, from the point of view of logic, it is interesting to see that the
categorical semantics seems to have a preference for multiplicative quantifiers.
Finally, have shown that, as in the intuitionistic case, we can represent infinitary (additive) disjunctions and conjunctions in linear type theory, through cofree Σ- and Π-types, indexed over Set. In
particular, this construction exhibits a family of non-trivial truly linear models of dependent types, providing an essential reality check for our system.
Despite what might be expected from this paper, much of this work has been very semantically motivated, by specific models. In joint work with Samson Abramsky, a model of linear dependent types
with comprehension has been constructed in a category of coherence spaces. Apart from the usual type
constructors from linear logic, it also supports Σ-, Π-, and Id-types. A detailed account of this model
will be made available soon.
In addition to providing (what as far as we are aware is) the first non-trivial, semantically motivated
model of such a type system, this work serves as a stepping stone for a model that we are currently
developing in a category of games, together with Samson Abramsky and Radha Jagadeesan. This, in
particular, should provide a game semantics for dependent type theory.
An indexed category of spectra up to homotopy over topological spaces has been studied in e.g. [15, 3]
as a setting for stable homotopy theory. It has been shown to admit I-, ⊗-, ⊸-, and Σ-types. The natural
candidate for a comprehension adjunction, here, is that between the infinite suspension spectrum and
the infinite loop space: L ⊣ M = Σ∞ ⊣ Ω∞ . A detailed examination of the situation and an explanation
of the relation with the Goodwillie calculus would be very desirable here, though. This might fit in with
our related objective of giving a linear analysis of homotopy type theory.
Another fascinating possibility is that of models related to quantum mechanics. Non-dependent linear
type theory has found very interesting interpretations in quantum computation, e.g. [22]. The question
rises if the extension to dependent linear types has a natural counterpart in physics. In [16], Urs Schreiber
has recently sketched how linear dependent types can serve as a language to talk about quantum field
theory and quantisation in particular. There are plenty of interesting open questions here.
Finally, there are still plenty of theoretical questions within the type theory. Can we expect to have
interesting models with type dependency on the co-Kleisli category of ! and can we make sense of additive
Σ- and Id-types, e.g. from a syntactic point of view? Is there an equivalent of strong/dependent E-rules
for ILDTT? Does the Curry-Howard correspondence extend in its full glory: do we have a propositionsas-types interpretation of linear predicate logic in ILDTT? These questions need to be addressed by a
combination of research into the formal system and study of specific models. We hope that the general
framework we sketched will play its part in connecting all the different sides of the story: from syntax
to semantics; from computer science and logic to geometry and physics.
34
References
[1] Cervesato, I., Pfenning, F.: A linear logical framework. In: LICS’96. Proceedings., IEEE (1996)
264–275
[2] Shulman, M.: Enriched indexed categories. Theory and Applications of Categories 28(21) (2013)
616–695
[3] Ponto, K., Shulman, M.: Duality and traces for indexed monoidal categories. Theory and Applications of Categories 26(23) (2012) 582–659
[4] Barber, A.: Dual intuitionistic linear logic. Technical Report ECS-LFCS-96-347, University of
Edinburgh, Edinburgh (1996)
[5] Hofmann, M.: Syntax and semantics of dependent types. In: Extensional Constructs in Intensional
Type Theory. Springer (1997) 13–54
[6] Jacobs, B.: Comprehension categories and the semantics of type dependency. Theoretical Computer
Science 107(2) (1993) 169–207
[7] Church, A.: A formulation of the simple theory of types. The journal of symbolic logic 5(02) (1940)
56–68
[8] Howard, W.A.: The formulae-as-types notion of construction. (1995)
[9] Lambek, J.: Deductive systems and categories iii. cartesian closed categories, intuitionist propositional calculus, and combinatory logic. In: Toposes, algebraic geometry and logic. Springer (1972)
57–82
[10] Martin-Löf, P.: An intuitionistic theory of types. Twenty-five years of constructive type theory 36
(1998) 127–172
[11] Girard, J.Y.: Linear logic. Theoretical computer science 50(1) (1987) 1–101
[12] Dal Lago, U., Gaboardi, M.: Linear dependent types and relative completeness. In: LiCS 2011.
Proceedings., IEEE (2011) 133–142
[13] Petit, B., et al.: Linear dependent types in a call-by-value scenario. In: Proceedings of the 14th
symposium on Principles and practice of declarative programming, ACM (2012) 115–126
[14] Gaboardi, M., Haeberlen, A., Hsu, J., Narayan, A., Pierce, B.C.: Linear dependent types for
differential privacy. In: ACM SIGPLAN Notices. Volume 48., ACM (2013) 357–370
[15] May, J.P., Sigurdsson, J.: Parametrized homotopy theory. Number 132. American Mathematical
Soc. (2006)
[16] Schreiber, U.: Quantization via linear homotopy types. arXiv preprint arXiv:1402.7041 (2014)
[17] Power, A.J.: A general coherence result. Journal of Pure and Applied Algebra 57(2) (1989) 165–173
[18] Melliès, P.A.: Categorical semantics of linear logic. Panoramas et Syntheses 27 (2009) 15–215
[19] Pitts, A.M.: Categorical logic. In Abramsky, S., Gabbay, D.M., Maibaum, T.S.E., eds.: Handbook
of Logic in Computer Science, Volume 5. Algebraic and Logical Structures. Oxford University Press
(2000) 39–128
[20] Bierman, G.M.: On intuitionistic linear logic. PhD thesis, Citeseer (1994)
[21] Lawvere, F.W.: Equality in hyperdoctrines and comprehension schema as an adjoint functor.
Applications of Categorical Algebra 17 (1970) 1–14
[22] Abramsky, S., Duncan, R.: A categorical quantum logic. Mathematical Structures in Computer
Science 16(3) (2006) 469–489 Preprint available at http://arxiv.org/abs/quant-ph/0512114.
35
| 6cs.PL
|
Noname manuscript No.
(will be inserted by the editor)
Improved Power Decoding of Interleaved One-Point
Hermitian Codes
Sven Puchinger · Johan Rosenkilde · Irene
Bouw
arXiv:1801.07006v1 [cs.IT] 22 Jan 2018
Received: date / Accepted: date
Abstract We propose a new partial decoding algorithm for h-interleaved one-point
Hermitian codes that can decode—under certain assumptions—an error of relative
g hh
+1 , where k is the dimension, n the length, and g the genus
weight up to 1 − ( k+
n )
of the code. Simulation results for various parameters indicate that the new decoder
achieves this maximal decoding radius with high probability. The algorithm is based
on a recent generalization of Rosenkilde’s improved power decoder to interleaved Reed–
Solomon codes, does not require an expensive root-finding step, and improves upon
the previous best decoding radius by Kampf at all rates. In the special case h = 1, we
obtain an adaption of the improved power decoding algorithm to one-point Hermitian
codes, which for all simulated parameters achieves a similar observed failure probability
as the Guruswami–Sudan decoder above the latter’s guaranteed decoding radius.
Keywords Interleaved One-Point Hermitian Codes · Power Decoding · Collaborative
Decoding · 94B35 · 14G50
1 Introduction
One-point Hermitian (1-H) codes are algebraic geometry codes that can be decoded
beyond half the minimum Goppa distance. Most of their decoders are conceptually
similar to their Reed–Solomon (RS) code analogs, such as the Guruswami–Sudan (GS)
Sven Puchinger
Institute of Communications Engineering, Ulm University, Germany
E-mail: [email protected]
Johan Rosenkilde
Department of Applied Mathematics and Computer Science, Technical University of Denmark
E-mail: [email protected]
Irene Bouw
Institute of Pure Mathematics, Ulm University, Germany
E-mail: [email protected]
2
Sven Puchinger, Johan Rosenkilde, Irene Bouw
algorithm [1] and power decoding (PD) [2–4]. For both RS and 1-H codes, PD is only
able to correct as many errors as the Sudan algorithm, which is a special case of the
GS algorithm. Recently [5], PD for RS codes was improved to correct as many errors as
the GS algorithm.
An h-interleaved 1-H code is a direct sum of h many 1-H codes. By assuming that errors
occur at the same positions in the constituent codewords (burst errors), it is possible to
decode far beyond half the minimum distance [6], which is inspired by decoding methods
for interleaved RS codes [2, 7]. In the RS case, there have been many improvements on
the decoding radius in the last two decades [2, 7–12], which have not all been adapted
to 1-H codes. The currently best-known decoding radius for interleaved RS codes is
achieved by both the interpolation-based technique in [8, 10] and the method based on
improved PD in [12], where the latter has a smaller complexity since it does not rely on
an expensive root-finding step.
In this paper, we adapt the decoder in [12], which is based on improved PD, to hinterleaved 1-H codes using the description of PD for 1-H as in [4]. Similar to the RS
case, we derive a larger system of non-linear key equations (cf. Section 3) and reduce
the decoding problem to a linear problem whose solution—under certain assumptions—
agrees with the solution of the system of key equations (cf. Section 4).
Using a linear-algebraic argument, we derive an upper bound on the maximum number
of errors which can yield a unique solution of the linear problem (cf. Section 5). This
decoding radius improves upon the previous best, [6], at all rates. In Section 6, we
present simulation results for various code and decoder parameters which indicate that
the new algorithm achieves the maximal decoding radius with high probability. The
complexity of solving the linear problem is shown to be sub-quadratic in the code
length in Section 7. Finally, we compare the decoding radii of RS, interleaved RS and
interleaved 1-H codes for the same overall field size and length in Section 8.
In the special case h = 1, we obtain an 1-H analogue of the improved PD for RS
codes [5]. This improves the decoding radius of the 1-H PD decoder of [4] at a similar
cost, sub-quadratic in the code length, and similar to the best known cost of the 1-H
GS algorithm [4]. Simulation results suggest that the decoder has a similar failure
probability as the GS algorithm for the same parameters when the decoding radius is
beyond the guaranteed radius of the GS algorithm (cf. Section 6).
The decoder is described for codes of full length n = q 3 ; the approach works for any
n < q 3 , but to obtain the good complexities, certain restrictions to how the evaluation
points are chosen should be kept. For notational convenience, we restrict ourselves to
homogeneous interleaved 1-H codes, i.e., where the constituent codes have the same
rate. The generalization to inhomogeneous codes is straightforward.
The results of this article were partly presented at the International Workshop on
Coding and Cryptography, Saint-Petersburg, Russia, 2017, where we only considered
the case h = 1 [13]. While writing this extension, we discovered some slightly improved
key equations, which are presented here. In the previous paper we sought Λs where
Λ is a usual notion of error-locator; now we instead define and seek Λs , which is an
“error-locator of multiplicity s”.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
3
2 Preliminaries
Let q be a prime power. We follow the notation of [4]. The Hermitian curve H/Fq2
is the smooth projective plane curve defined by the affine equation Y q + Y = X q+1 .
The curve H(Fq2 ) has genus g = 12 q (q − 1) and q 3 + 1 many Fq2 -rational points
P = {P1 , . . . , Pq3 , P∞ }, where P∞ denotes the point at infinity. We define R :=
∪mH ≥0 L(mH P∞ ) = Fq2 [X, Y ]/(Y q + Y − X q+1 ), which has an Fq2 -basis of the form
{X i Y j : 0 ≤ i, 0 ≤ j < q}. The order function degH : R → Z≥0 ∪{−∞}, f 7→ −vP∞ (f )
is defined by the valuation vP∞ at P∞ . As a result, we have degH (X i Y j ) = iq + j (q +1).
We will think often operate with elements of R as bivariate polynomials in X and Y ,
represented as Fq2 -linear combinations of the aforementioned basis. In this paper, when
we say “degree” of an element in R, we mean its degH . A non-zero element of R is
called monic if its monomial of largest degH has coefficient 1.
Let n = q 3 and mH ∈ N with 2(g − 1) < mH < n. The one-point Hermitian code of
length n and parameter mH over Fq2 is defined by
CH (n, mH ) = {(f (P1 ), . . . , f (Pn )) : f ∈ L(mH P∞ )} .
The dimension of CH is given by k = mH − g + 1 and the minimum distance d is
lower-bounded by the designed minimum distance d∗ := n − mH .
The (homogeneous) h-interleaved one-point Hermitian code of length n and parameter
mH over Fq2 is the direct sum of h one-point Hermitian codes CH (n, mH ), i.e.,
c1
..
h×n
CH (n, mH ; h) = . ∈ Fq2 : ci ∈ CH (n, mH ) .
ch
As a metric for errors, we consider burst errors: If r = c + e ∈ Fh×n
is received for a
q2
codeword c ∈ CH (n, mH ; h), then the error positions E ⊆ {1, . . . , n} are given by the
non-zero columns of e, i.e.,
E :=
Sh
j =1
i : ej,i 6= 0 .
For a vector i = [i1 , . . . , im ] ∈ Zh
≥0 , we define its size as |i| :=
P
µ iµ .
We denote by
the product partial order on Zh
≥0 , i.e. i j if iµ ≤ jµ for all µ. The number of vectors
h+µ−1
∈Zm
). We use the following relations, which follow
≥0 of size |i| = µ is given by (
µ
from properties of the binomial coefficient.
Lemma 1 Let m, t ∈ Z>0 . Then,
Pt
m+µ−1
)
µ=0 (
µ
= (mm+t), and
Pt−1
µ=0 µ(
m+µ−1
)
µ
Note that the Lemma 1 means e.g.
X
+t−1
|i| = t(hh
+1 ) .
i∈Zh
≥0 ,|i|<t
We also introduce the following notational short-hands:
t−1
= t(mm++1
).
4
Sven Puchinger, Johan Rosenkilde, Irene Bouw
Definition 1 For a ∈ Rh , and i, j ∈ Zh
≥0 , we define
ai :=
iµ
µ=1 aµ ,
Qh
(ji ) :=
Qm
jµ
µ=1 ( iµ ).
By extending the binomial theorem to this notation, we obtain the following lemma.
Lemma 2 Let a, b ∈ Rh , and j ∈ Zm
≥0 . Then,
(a + b)j =
j i j−i
.
ij ( i )a b
P
For computational complexities, we use the soft-O notation O∼ , which omits log factors.
3 System of Key Equations
In this section, we derive the system of key equations that we need for decoding, using
the same trick as [12] for interleaved Reed–Solomon codes. We use the description of
power decoding for one-point Hermitian codes as in [4]. Suppose that the received word
is r = c + e ∈ Fh×n
q 2 , consisting of an error e with corresponding (burst) error positions
E and a codeword c ∈ CH (n, mH ; h), which is obtained from the message polynomials
f = [f1 , . . . , fh ] ∈ L(mP∞ )h .
In the following sections we show how to retrieve the message polynomials f from the
received word r if the number of errors |E| does not exceed a certain decoding radius,
which depends on the parameters of the decoding algorithm. Similar to [4], we define
the following polynomials.
Definition 2 Let
locator polynomial Λs of multiplicity s is the
P s ∈ N. The error
element in L − i∈E sPi + ∞P∞ of minimal degree that is monic.
Theorem 1 The error locator polynomial of multiplicity s is unique and has degree
s|E| ≤ degH Λs ≤ s|E| + g.
clear since if there were
Proof: The proof is similar to [4, Lemma 23]. Uniqueness is P
two such polynomials, their difference
would
also
be
in
L
−
i∈E sPi + ∞P∞ , but
P
of smaller degH . Being in L − i∈E sPi + ∞P∞ specifies s|E| homogeneous linear
equations
P in the coefficients of Λs , since for any i ∈ E, we can expand Λs into a power
series j≥s γi,j φji for a local parameter φi of Pi (e.g., take φi = X − αi if Pi = (αi , βi )).
By requiring degH Λs ≤ s|E| + g, we have more variables than equations, so there is a
non-zero Λs of the sought form with degree at most s|E| + g. The lower bound works
exactly as in [4, Lemma 23].
Lemma 3 For each i = 1, . . . , h, there is a polynomial Ri ∈ R with degH (Ri ) <
n + 2g that satisfies R(Pj ) = ri,j for all Pj ∈ P ∗ . Each Ri can be computed in O∼ (n)
operations over Fq2 .
Improved Power Decoding of Interleaved One-Point Hermitian Codes
5
Proof: Apply [4, Lemma 6] to each row of the received word.
In the following, let R = [R1 , . . . , Rh ] ∈ Rh be as in Lemma 3 and G ∈ R be defined as
G=
Q
α∈Fq2 (X
2
− α) = X q − X.
Lemma 4 For each i ∈ Z≥0 with |i| ≤ s, there is a unique Ωs,i ∈ R of degree
degH Ωs,i ≤ degH Λs + |i|(2g − 1) such that
Λs (f − R)i = G|i| Ωs,i .
Proof: Consider vPj (Λs (f − R)i ) for j = 1, . . . , n: if j ∈ E then vPj (Λs (f − R)i ) =
vPj (Λs ) ≥ s ≥ |i|. If j ∈
/ E then vPj (Λs (f − R)i ) ≥ vPj ((f − R)i ) ≥ |i|. We conclude
Λs (f − R)i ∈ L − |i|
n
X
Pj + ∞P∞ .
j =1
Since the divisor in that L-space is exactly div(G|i| ) + ∞P∞ , then Λs (f − R)i must
be divisible by G|i| (see e.g., [4, Lemma 3]) with quotient in R. The degree is given by
taking degH on both sides and using degH (Ri ) < n + 2g − 1.
The following theorem states the system of key equations that we will use for decoding in
the next sections. Note that the formulation is similar to its interleaved Reed–Solomon
analog [12], with the difference that all involved polynomials are elements of the ring R.
Theorem 2 (System of Key Equations) Let `, s ∈ Z>0 be such that s ≤ ` and Λs ,
f , R, G, and Ωs,i as above. Then, for all j ∈ Zh
≥0 of size 1 ≤ |j| ≤ `, we have
"
j
Λs f =
X
Ωs,i
ij
"
j
Λs f ≡
X
Ωs,i
ij
|i|<s
!
#
j
Rj−i G|i| ,
i
!
j
Rj−i G|i|
i
1 ≤ |j| < s
(1)
s ≤ |j| ≤ `,
(2)
i
(3)
#
mod Gs ,
as congruences over R.
Proof: Using Lemma 2, we obtain
j
Λs f j = Λs (R + (f − R)) =
j
ij ( i )Λs (f
P
− R) Rj−i .
In all summands with |i| < s, we can rewrite, using Lemma 4,
Λs (f − R) = G|i| Ωs,i .
i
(4)
If |i| ≥ s, we can write i = i0 + i00 , for some i0 , i00 ∈ Z≥0 with |i0 | = s, and
i
0
00
00
Λs (f − R) = Λs (f − R)i (f − R)i = Gs Ωs,i0 (f − R)i ,
so all those terms are divisible by Gs . For |j| < s, all summands of (3) have |i| ≤ |j| < s
and are of the form (4). We therefore obtain (1). For |j| ≥ s, all summands of (3) with
|i| ≥ s are divisible by Gs , so we get (2).
6
Sven Puchinger, Johan Rosenkilde, Irene Bouw
4 Solving the System of Key Equations
The key equations in Theorem 2 are non-linear relations between the unknown polynomials Λs , f , and Ω. We therefore relax them into—at the first glance much weaker—linear
problem and hope that their solutions agree. The resulting problem is a heavy generalisation of multi-sequence linear shift register synthesis [14, 15], which is very related to
simultaneous Hermite Padé approximations [16].
Problem 1 Consider a code C = CH (n, mH ; h) and a decoding instance with received
word r = c + e ∈ Fh×n
q 2 , where c ∈ C is unknown and is obtained from the unknown
message polynomials f ∈ L(mP∞ )h . Let R and G be as in Section 3. Given positive
integers s ≤ `, let
Ai,j = (ji )Rj−i G|i| ∈ R
h
for all i ∈ I := {i ∈ Nh
0 : 0 ≤ |i| < s} and j ∈ J := {j ∈ N0 : 1 ≤ |j| ≤ `}. Find
λi , ψj ∈ R for i ∈ I and j ∈ J with monic λ0 , such that
X
j ∈ J and |j| < s,
(5)
j ∈ J and |j| ≥ s,
(6)
degH λ0 ≥ degH λi − |i|(2g − 1)
i ∈ I,
(7)
degH λ0 ≥ degH ψj − |j|mH
j ∈ J.
(8)
ψj =
λi Ai,j
i∈I
ψj ≡
X
λi Ai,j
mod Gs
i∈I
Definition 3 Consider an instance of Problem 1. We say that a solution (λi )i∈I , (ψj )j∈J ,
has degree τ ∈ Z≥0 if degH λ0 = τ . Furthermore, we call a solution minimal if its
degree is minimal among all solutions.
Problem 1 is connected to the key equations through the following statement.
Theorem 3 Consider an instance of Problem 1. Then,
λi = Λi := Ωs,i ,
i ∈ I,
ψj = Ψj := Λs f j ,
j ∈ J,
is a solution to the problem of degree τ = degH Λs , where s · |E| ≤ τ ≤ s · |E| + g.
Proof: Note Ωs,0 = Λs . The equalities and congruences are now clear from the key
equations. As for the degree restrictions, we have
degH Λi ≤ degH Λ0 + |i|(2g − 1),
degH Ψj = degH (Λs ) + degH (f j ) ≤ degH Λ0 + |j|mH ,
which proves the claim.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
7
Remark 1 Most received words will satisfy degH Ri = n + 2g − 1 for all i = 1, . . . , h. In
such a case, the solution of Problem 1 given in Theorem 3 fulfills all degree restrictions
of the problem with equality. These relative upper bounds on the degrees of λi and ψj
are therefore the minimal choice among all such bounds for which Theorem 3 holds.
Theorem 3 motivates a decoding strategy, which is outlined in Algorithm 1: To every
codeword c0 ∈ CH (n, mH ; h) corresponds a solution to Problem 1 whose degree is
roughly s · |E 0 |, where |E 0 | is the number of errors (i.e., non-zero columns) of r − c0 .
Among those solutions, we want to find the one of smallest degree, i.e., the one for
the closest codeword. There will also be other solutions to Problem 1, which do not
correspond to codewords, but the idea is that in most cases, and when the number of
errors is not too large, the minimal solution will correspond to the closest codeword.
Algorithm 1: Improved Power Decoder for h-Interleaved 1-Point Hermitian Codes
Input: Received word r ∈ Fh×n
q m and positive integers s ≤ `
1
Output: f ∈ L(mH P∞ )h such that ci = [fi (P1 ), . . . , fi (Pn )] for all i = 1, . . . , h is
the codeword with a corresponding minimal degH Λs ; or “decoding failure”.
Compute R and G as in Section 3
8
Ai,j ← (ji )Rj−i G|i| for all i j
λi , ψj ← Minimal solution to Problem 1 with input s, `, Ai,j , and G
if λ0 divides all ψui over R for i = 1, . . . , h, where ui is the ith unit vector then
f ← [ψu1 /λ0 , . . . , ψuh /λ0 ]
E ← Error setPcorresponding toei = r i − [fi (α1 ), . . . , fi (αn )] for i = 1, . . . , h
if λ0 ∈ L − i∈E sPi + ∞P∞ and s · |E| ≤ degH λ0 ≤ s(|E| + g ) then
return f
9
return “decoding failure”
2
3
4
5
6
7
In the cases for which this does not happen, the decoder will fail; we will return to this
in Section 5. If the algorithm finds a solution that corresponds to a codeword, then we
have λ0 = Λs and ψui = Λs fi for i = 1, . . . , h, where ui = [0, . . . , 1, . . . , 0] is the ith
unit vector. Hence, we obtain the ith message polynomial fi by division of ψui by λ0 .
Note that Algorithm 1 does not exactly promise to find the closest codeword: it finds
the codewords whose corresponding Λs has minimal degH . When the number of errors
is very small, we will often or always have degH Λs < s|E| + g; but in this case all
other codewords are much farther away from r. On the other hand, when the number
of errors is large, most error vectors will satisfy degH Λs = s|E| + g. In both these cases
Algorithm 1 will find the closest codewords. It seems reasonable to expect, however,
that there exist some rare received words for which a farther codeword will have an
associated Λs of lower degH than the closest codeword.
We will see in Section 7 that we can find a minimal solution of Problem 1 efficiently.
8
Sven Puchinger, Johan Rosenkilde, Irene Bouw
5 Decoding Radius and Failure Behavior
In this section, we derive an upper bound on the maximal degree of the error locator
polynomial Λs for which there can be a unique minimal solution of Problem 1. Since the
degree of Λs is related to the number of errors, this implies an estimate of the maximal
decoding radius of our decoder. We also briefly discuss in which cases the decoder fails
below this bound.
Lemma 5 Let τ, `, s ∈ N such that s ≤ ` and τ + `mH < sn. All polynomials λi , ψj ∈
R for i ∈ I and j ∈ J that fulfill (5), (6), and the absolute degree restrictions
degH λi − |i|(2g − 1) ≤ τ,
(9)
degH ψj − |j|mH ≤ τ,
(10)
can be computed by a homogeneous linear system of equations over Fq2 with at least
h
i
`
h+s−1
h+`
h+s−1
h+`
h+`
δ (τ ) = (τ + 1)(h+
h ) − n h( h+1 ) + s( h ) − s( h ) + mH h(h+1) − g ( h )
more variables than equations, whenever δ (τ ) ≥ 0. If τ ≥ 2g − 1, there are received
words for which the difference is exactly δ (τ ).
Proof: We have degH Ai,j ≤ (n + 2g − 1)|j| − (2g − 1)|i|, so we get
degH
X
λi Ai,j
≤ τ + |j|(n + 2g − 1)
∀j ∈ J.
i∈I
P
Thus, for most j the polynomial ψj has lower degree than the terms in i∈I λi Ai,j in
the case |j| < s and less than the degree of the modulus Gs in the other case.
P Consider
i j
2 . Since the Fq 2 -coefficients of
functions
in
R
over
the
basis
{X
Y
}
over
F
q
i∈I λi Ai,j
P
and ( i∈I λi Ai,j mod Gs ) are known linear combinations of the unknown coefficients
of the λi , the restrictions of the lemma on the degrees of ψj ∈ R can be described by
an Fq2 homogeneous linear system of equations that specify that the top coefficients of
ψj be zero (τ + |j|mH + 1 and higher).
For non-negative integers a and b, there are between b − a − g and b − a many monomials
xi y j ∈ R with j < q of degree at least a and less than b. The lower bound is due to the
Riemann–Roch theorem and the upper bound follows from the injectivity of degH on
the set of monomials.
Due to the degrees of the involved polynomials, the number of Fq2 -linear restrictions
for each |j| < s becomes
Nj = (τ + |j|(n + 2g − 1)) − (τ + |j|mH ) = |j|(n + 2g − 1 − mH ).
For |j| ≥ s, the analysis is a bit more involved: Since G is a polynomial only in X with
s
degX (Gs ) = sq 2 , the congruence
the X-degree of all monomials
Pmodulo G reduces
2
below sq , i.e., the polynomial ( i∈I λi Ai,j mod Gs ) can be written as
!
X
i∈I
s
λi Ai,j mod G
=
2
q−
−1
X1 sqX
j =0 i=0
aij X i Y j ,
Improved Power Decoding of Interleaved One-Point Hermitian Codes
9
where aij ∈ Fq2 are linear expressions of the coefficients of the λi . By the degree
restriction of ψj , we must have that the coefficients aij with degH (X i Y j ) > τ + |j|mH
are zero. Thus, we get at most
q−
X1
Nj =
j =0
sq 2 − |{(i, j ) : qi + (q + 1)j ≤ τ + |j|mH }|
|
{z
}
= τ +|j|mH −g +1
= sn − τ − |j|mH + g − 1
linear equations. Note that the condition τ + `mH < sn guarantees that there is no
monomial X i Y j of degH (X i Y j ) > τ + |j|mH with degX ≥ sq 2 . In total, and using
Lemma 1 repeatedly:
NE =
X
Nj =
X
1≤|j|<s
j∈J
= (n + 2g
X
|j|(n + 2g − 1 − mH ) +
sn − τ − |j|mH + g − 1
s≤|j|≤`
s−1
− 1)h(h+
h+1 ) + (sn − τ
`
h+s−1
h+`
− 1 + g ) (h+
h ) − ( h ) − mH h(h+1).
The number of variables, i.e., the number of Fq2 -coefficients of the λi is at least
NV =
X
s−1
(τ + |i|(2g − 1) + 1 − g ) = (τ + 1 − g )(h+hs−1) + (2g − 1)h(h+
h+1 ).
i∈I
The claim follows by subtracting NV − NE.
In the case τ ≥ 2g − 1, all Weierstraß gaps are below the degree bounds of the λi and
ψj . Hence, the number of variables and equations is equal to the derived NE and NV,
respectively, as long as the maximal possible degree of degH
P
i∈I
λi Ai,j , i.e., for
some choice of the λi , is equal to τ + |j|(n + 2g − 1). There are received
words for which
degH Ri = n + 2g − 1 for all i. In these cases, we can have degH
P
i∈I
=
λi Ai,j
τ + |j|(n + 2g − 1) for some values of λi , so (if τ ≥ 2g − 1), the number of variables
minus the number of equations is exactly δ (τ ).
Lemma 6 If Problem 1 has a solution of degree τ , it has at least (q 2 )δ(τ )−1 many
such solutions.
Proof: Solutions of degree τ of Problem 1 are exactly the solutions of the homogeneous
linear system in Lemma 5 with degH λ0 = τ and monic λ0 . Thus, we set the τ th
coefficient of λ0 to 1 and obtain an inhomogeneous linear system of equations with at
least δ (τ ) − 1 more variables than equations. If Problem 1 has a solution of degree τ ,
then this system has at least (q 2 )ker(A) solutions, where A is the system’s matrix. The
claim follows by dim(ker(A)) ≥ δ (τ ) − 1.
Lemma 6 implies the following statement.
Theorem 4 Let τ = degH Λs and s, ` ∈ N such that s ≤ ` and τ + `mH < sn and
τ > τmax := sn 1 −
s−1
s−1
s(h+h
)−h(h+h+1
)
+`
s(hh
)
−
h
h+1 `mH +
Then, Problem 1 has at least two solutions of degree τ .
1
(hh+`)
−1
+ g.
(11)
10
Sven Puchinger, Johan Rosenkilde, Irene Bouw
Proof: Condition (11) is fulfilled if and only if δ (τ ) > 1. Due to τ = degH Λs , Problem 1
has a solution of degree τ and the claim follows by Lemma 6.
Theorem 4 can be interpreted as follows: If degH Λs > τmax , then either Λs does not
correspond to a minimal solution of Problem 1, or it is a minimal solution but there
are many other minimal solutions as well. There is no reason to think that our solver
for Problem 1 will find Λs among all those solutions, so decoding will likely fail.
Since we have |E| ≤ degH Λs ≤ s|E| + g, and often degH Λs = s|E| + g, we usually have
no unique solution whenever s|E| + g > τmax , i.e.,
|E| >
and for sure if |E| >
tnew =
τmax −g
s
τmax
s .
τmax −g
s
We therefore call
s−1
s−1
s(h+h
)−h(h+h+1
)
=n 1−
−
+`
s(hh
)
h
`
h+1 s mH
+
1
s
1
(hh+`)
−1
(12)
the decoding radius of Algorithm 1.
Remark 2 For τ ≥ 2g − 1, by Lemma 5, there are received words (in fact most of them)
such that the difference of numbers of variables and equations of the inhomogeneous
system for computing all degree-τ solutions of Problem 1 is exactly δ (τ ) − 1. Thus, if
there are sufficiently many linearly independent equations1 , there is no other solution
of the problem, besides the error locator, of degree τ whenever δ (τ ) < 0.
For τ < 2g − 1, the degree bounds of λ0 and ψ0 are smaller than 2g − 1, but those of all
other λi and ψj are bigger (note that mH ≥ 2g − 1). Thus, there can be up to g fewer
equations and up to g more variables than predicted by δ (τ ) for any received word. The
value of τmax as in Theorem 4 can in this case therefore be smaller by a value up to
0
τmax
= τmax − 2g
1
(hh+`)
,
`
which reduces the decoding radius by at most 2g/[s(h+
h )].
In the case τ + `mH ≥ sn, the number of equations is also smaller than predicted by
τmax . However, we will see in Section 5.1 that the best choice of s for a given ` yields
τ + `mH < sn for τ ≤ τmax .
Since we cannot guarantee that the linear equations of the system in Lemma 5 are
linearly independent for τ ≤ degH Λs , Algorithm 1 can fail to return the sent codeword
c for some errors of weight less than the maximal decoding radius. In these cases, we
have one of the following.
– There is a solution of Problem 1 of degree < degH Λs .
– There is more than one solution of Problem 1 of degree = degH Λs and the decoder
picks the wrong one.
1 This linear-algebraic condition resembles, but seems weaker than, the “(non-linear) algebraic
independence assumption” in [10] for decoding interleaved RS codes.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
11
However, the simulation results for various code and decoder parameters, presented
in the following section, indicate that the new decoder is able to decode most error
patterns up to the derived decoding radius tnew . Sometimes, decoding succeeds even
beyond tnew . In these cases, we usually have degH (Λs ) < s|E| + g.
In all previous power decoding algorithms for Reed–Solomon [2, 5], one-point Hermitian
[4, 6], and interleaved Reed–Solomon codes [12], simulation results indicate that the
failure probability for a number of errors below the maximal decoding radius is small
and decreases exponentially in the difference of maximal decoding radius and number
of errors.
As for these other variants, except for a few parameters of theirs (e.g., ` ≤ 3 and s ≤ 2
for a single Reed–Solomon code in [5]), it remains an open problem to prove an analytic
upper bound on the failure probability of Algorithm 1.
5.1 Asymptotic Analysis and Parameter Choice
We study the asymptotic behavior of the decoding radius τmax and give explicit parameters to achieve the given limit. The analysis is based on the following lemma.
Lemma 7 ([12, Lemma 14]) Let γ ∈ (0, 1) and h ∈ N be fixed. Then, we have
(h+bγic
)
h
= γ h + O( 1i )
(hh+i)
for (i → ∞).
Theorem 5 Let (`i , si ) = (i, bγic + 1) for i ∈ N, where γ =
tnew (`i , si ) = n 1 −
mH
n
hh+1
− O( 1i )
pm
h+1
H
n
. Then,
for (i → ∞).
Proof: We have
tnew
n
h
h i (h+bγic
)
h
= 1 − 1 + 1 − s1i h+1
−
(hh+i)
| {z }
|
{z
}
= 1−O ( 1i )
= γ h +O ( 1 )
h
h+1
m
m+1
γ h − γ −1 mnH
{z
}
|
=0
− γh − O
1
i
mH
n
1
+
−1
=γ
+O ( 1i )
=1−
mH
n
s
|{z}
i
=1+
`i
si
i
(
|{z}
|
1
=O i
hh+1
−O
1
h+`i
h
)
{z
−1
}
= −1+O ( 1i )
1
i
,
which proves the claim.
Note that the choice of `i and si in Theorem 5 ensures that τ + `i mH < si n for all
τ ≤ s · tnew (`i , si ).
12
Sven Puchinger, Johan Rosenkilde, Irene Bouw
6 Numerical Results
In this section, we present simulation results. We have conducted Monte-Carlo simulations for estimating the failure probability of the new decoding algorithm in a channel
that randomly adds t = |E| errors, using N ∈ {103 , 104 } samples. The decoder was
implemented in SageMath v7.5 [17], based on the power decoder implementation of [4].
All simulated examples fulfill degH Λs ≥ st ≥ 2g − 1. If this condition is not fulfilled,
the simulation results might differ from the expected decoding radius, cf. Remark 2.
6.1 Case h = 1
We first compare the new improved power (P̂fail,IPD ) with the Guruswami–Sudan
(P̂fail,GS ) decoder. The used implementation of the Guruswami–Sudan decoder is the
publicly available one from [4]. Table 1 presents the simulation results for various code
(q, m, n, k, d∗ ), decoder (`, s), and channel (t) parameters.
Table 1 Observed failure rate of the improved power (P̂fail,IPD ) and Guruswami–Sudan
(P̂fail,GS ) decoder for h = 1. Code parameters q, m, n, k, d∗ . Decoder parameters `, s. Number
of errors t (+ t = tnew decoding radius as in (12)). Number of experiments N .
q
m
n
k
d∗
`
s
4
15
64
10
49
4
2
5
55
125
46
70
3
2
5
20
125
11
105
5
2
7
70
343
50
273
3
2
7
70
343
50
273
4
2
7
55
343
35
288
4
2
t
28
29+
30
35
36+
37
67
68+
69
160
161+
162
168
169+
170
183
184+
185
P̂fail,IPD
P̂fail,GS
N
0
0
9.93 · 10−1
0
0
9.57 · 10−1
0
0
9.91 · 10−1
0
0
9.78 · 10−1
0
0
9.79 · 10−1
0
0
9.82 · 10−1
0
3.30 · 10−3
9.39 · 10−1
0
4.00 · 10−4
9.60 · 10−1
0
7.00 · 10−3
9.60 · 10−1
0
0
9.86 · 10−1
0
0
2.2 · 10−2
0
0
1.9 · 10−2
104
104
104
104
104
104
103
103
103
103
103
103
103
103
103
103
103
103
It can be observed that both algorithms can almost always correct tnew many errors,
improving upon classical power decoding. Also, neither of the two algorithms is generally
superior in terms of failure probability.
When comparing the two algorithms, one has to keep in mind that the GS algorithm is
g
g
`
`
guaranteed to work only up to n[1 − 2(s`+1
+1) ] − 2s mH − s = tnew − s + s(`+1) errors.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
13
6.2 General Case
We now turn to the general case of h > 1, where the previous best relative decoding
h
radius is tK = h+1
(n − mH ) [6]. The simulations results for various code and decoder
parameters are given in Table 2.
Table 2 Observed failure rate of Algorithm 1 (P̂fail,IPD ) for h > 1. Code parameters
q, mH , n, k, d∗ , h. Decoder parameters `, s. Number of errors t (+ t = tnew decoding radius
as in (12)). Number of experiments N . Previous best decoding radius tK [6].
q
4
mH
15
n
64
k
10
d∗
49
h
2
`
3
s
2
t
P̂fail,IPD
N
tK
36
36+
37
38+
39
39+
40
29+
30
0
9.18 · 10−1
0
9.31 · 10−1
0
9.42 · 10−1
0
1
0
9.44 · 10−1
105
103
103
103
105
103
102
102
103
103
32
32
32
32
36
36
36
36
28
28
79+
80
81+
82
86+
87
0
9.37 · 10−1
0
9.93 · 10−1
0
9.94 · 10−1
105
103
103
103
103
103
70
70
70
70
78
78
35+
4
15
64
10
49
2
5
3
4
15
64
10
49
3
3
2
4
15
64
10
49
3
4
3
4
22
64
17
42
2
4
3
5
20
125
11
105
2
3
2
5
20
125
11
105
2
4
2
5
20
125
11
105
3
3
2
5
55
125
46
70
2
4
3
48+
49
0
9.86 · 10−1
103
103
46
46
7
90
343
70
253
2
3
2
183+
184
0
9.72 · 10−1
103
103
168
168
8
128
512
101
384
2
3
2
281+
282
0
1
102
102
256
256
In all tested cases, Algorithm 1 corrected all decoding trials up to tnew many errors and
failed with large observed probability one error beyond this radius.
7 Efficiently Finding a Minimal Solution of Problem 1
We use the Fq2 [X ]-vector representation of an element of R (cf. [4]) to reformulate
Pq−1
i
Problem 1 over Fq2 [X ]. Recall that for a ∈ R, we can write a =
i=0 ai Y ∈ R
with unique ai ∈ Fq2 [X ]. Then, the vector representation [4] of a is defined by ν (a) =
(a0 , . . . , aq−1 ) ∈ Fq2 [X ]q . Note that q deg(ai ) + i(q + 1) ≤ degH (a). For a, b ∈ R it
can be shown that
ν (a + b) = ν (a) + ν (b),
ν (ab) = ν (a)µ(b)Ξ,
14
Sven Puchinger, Johan Rosenkilde, Irene Bouw
where µ(b) ∈ Fq2 [X ]q×(2q−1) and Ξ ∈ Fq2 [X ](2q−1)×q are defined by
b0 b1 b2
b0 b1
µ(b) :=
..
.
. . . bq−1
. . . bq−2 bq−1
,
..
..
..
. ...
.
.
b0 b1 . . . bq−2 bq−1
1
1
Ξ := q+1
−1
X
X q+1
..
.
..
.
−1
..
.
X q+1
1
.
−1
Note further that for c ∈ Fq2 [X ] we have simply µ(ac) = µ(a)c. Using the notation
above, we can reformulate Problem 1 into the following problem over Fq2 [X ]. In the
following, let [q ) denote {0, . . . , q − 1}.
Problem 2 Given positive integers s ≤ `, R and G as in Section 3, and
A(i,j ) := µ(Ai,j )Ξ = µ (ji )Rj−i G|i| Ξ ∈ Fq2 [X ]q×q
h
for all i ∈ I := {i ∈ Nh
0 : 0 ≤ |i| < s} and j ∈ J := {j ∈ N0 : 1 ≤ |j| ≤ `}. Find
λi,ι , ψj,κ ∈ Fq2 [X ] for i ∈ I, j ∈ J , ι, κ ∈ [q ), not all zero, such that
ψj,κ =
X q−
X1
(i,j )
1 ≤ |j| < s,
λi,ι Aι,κ
i∈I ι=0
ψj,κ ≡
X q−
X1
(i,j )
λi,ι Aι,κ
mod Gs
1 ≤ |j| < s,
i∈I ι=0
max q deg λ0,ι + ι(q + 1) ≤ q deg λi,ι + ι(q + 1) − |i|(2g − 1)
ι∈[q )
max q deg λ0,ι + ι(q + 1) ≤ q deg ψj,κ + κ(q + 1) − |j|mH
ι∈[q )
0 ≤ |i| < s, ι ∈ [q )
1 ≤ |j| ≤ `, κ ∈ [q )
Similar toits R-equivalent, we define the degree of a solution of the above problem to be
maxι∈[q) q deg λ0,ι + ι(q + 1) and call the solution monic if the leading coefficient of
the λ0,ι that maximizes maxι∈[q) q deg λ0,ι + ι(q + 1) is 1. The following statement
establishes the connection between Problem 1 and Problem 2.
Theorem 6 Let τ ∈ Z≥0 . Then, λi , ψj ∈ R for i ∈ I and j ∈ J is a solution of
degree τ of Problem 1 if and only if
[λi,0 , . . . , λi,q−1 ] := ν (λi )
[ψj,0 , . . . , ψj,q−1 ] := ν (ψj )
is monic solution of degree τ of Problem 2.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
15
Proof: If λi and ψj forms a solution to Problem 1 this means for |j| ≥ s that there is
some uj ∈ R such that:
ψj =
X
λi Ai,j + uj Gs
⇐⇒
ij
|i|<s
ν (ψj ) =
X
ν (λi )µ(Ai,j )Ξ + µ(uj )Gs .
ij
|i|<s
since Gs ∈ Fq2 [X ]. This implies element-wise the congruence of Problem 2. The opposite
direction is analogous, as is the case |j| < s. The degree restrictions follow immediately
from degH (a) = maxι∈[q) {q deg aι + ι(q +1)} for any a ∈ R and ν (a) = (a1 , . . . , aq−1 ).
Problem 2 is of the same instance as the problem discussed in [4, Section V.B],2 which
can be solved by transforming an Fq2 [X ]-module basis that depends on the entries of
A(i,j ) and the relative degree bounds in Problem 2, into a reduced polynomial matrix
form (weak Popov form). Using this approach, finding a minimal solution of Problem 1
can be implemented in
ω
`
O∼ (h+
h ) sn
ω +2
3
⊆ O∼ `hω sn
ω +2
3
operations over Fq2 , where 2 ≤ ω ≤ 3 is the matrix multiplication exponent. In [5],
a similar kind of problem was reduced to finding solution bases of so-called Padé
approximation problems. In this way, the complexity can be slightly reduced to
`
O∼ (h+hs−1)(h+
h )
ω−1
sn
ω +2
3
⊆ O∼ `h(ω−1) sh+1 n
ω +2
3
operations over Fq2 . In order to achieve the asymptotic decoding radius, the code
parameters must be chosen as in Section 5.1. In this case, the two asymptotic complexity
statements above coincide and we get the following result.
Theorem 7 For a fixed code of rate R =
k
n
s, ` ∈ O(1/ε) such that tnew ≥ n(1 − (R +
∼
costs O ((1/ε)
hω +1
n
ω +2
3
and any constant ε > 0, we can choose
g−1 hh
+1
n )
− ε). In this case, Algorithm 1
).
Proof: The first statement directly follows from Theorem 5. The pre- and post-computations
in Algorithm 1 are negligible compared to Line 3 by similar arguments as in [4]. The
complexity thus follows by the arguments above.
2
Using this approach, it is necessary to reformulate the equations for 1 ≤ |j| < s into
(i,j)
congruences modulo xξ , where ξ is greater than the largest possible degree of the λi,ι Aι,κ for
the maximal number of errors that we expect to corrected.
16
Sven Puchinger, Johan Rosenkilde, Irene Bouw
8 Comparison to Interleaved Reed–Solomon Codes
An h-interleaved code over some field FQ over the burst error channel can equivalently
be considered as a code over FQh considered over the Qh -ary channel. This allows
comparing the decoding capability of interleaved 1-H codes with other constructions
of short codes over large fields, most notably RS codes and interleaved RS codes, see
Figure 1 for the case Qh = q 6 .
More precisely, for any h ∈ Z>0 , we have several ways of obtaining [n, k] codes over
Fq6h for n = q 3 and some dimension k < n. We will compare the following relative
decoding radii:
tRS : an RS code over Fq6h decoding up to the Johnson radius using one of [1, 5, 18].3
tIRS : 2h-interleaved RS code over Fq3 using one of [10, 12].
t1H : 3h-interleaved 1-H code over Fq2 using the proposed algorithm.
RS Codes
Interleaved RS Codes
Interleaved 1-H Codes
n = q3
n = q3
n = q3
(q 3 )2
q6
1
(q 2 )3
2
1 − R2
1 − R3
1 − (R +
g 3
)4
n
Fig. 1 Example: Comparison of interleaved Reed–Solomon and one-point Hermitian codes of
length n = q 3 over an overall field size of q 6 .
These values are as follows:
1
1 2
tRS = 1 − ( k−
n )
2h
1 2h+1
tIRS = 1 − ( k−
,
n )
1
tIH = 1 − ( k−
n −
g 3h3h
+1 ,
n)
g
→ 0 for n → ∞, we can asymptotically
The asymptotics are already clear: since n
achieve larger decoding radii with interleaved 1-H codes than with interleaved RS codes,
when considering comparable overall field size. Below follows some concrete parameter
examples.
q = 13 is the smallest prime power for which tIH > tIRS for rate 1/2, i.e. both interleaved
codes can be considered as [2197, 1098] codes over F136 , and the decoding radii are
tRS = 644, tIRS = 814 and tIH = 823. A list of decoding radii of rate 1/2 codes with
even q is given in Table 3.
9 Conclusion
We have presented a new decoding algorithm for h-interleaved one-point Hermitian codes
based on the improved power decoder for Reed–Solomon codes in [5], its generalization
3 As pointed out in [19], an RS code over F
q 6h whose evaluation points all lie in Fq 3 are
equivalent to a 2h-interleaved RS codes over Fq3 , i.e. can be decoded up to tIRS . In our
comparison here we therefore consider RS codes with arbitrary evaluation points for which this
equivalence doesn’t hold.
Improved Power Decoding of Interleaved One-Point Hermitian Codes
17
Table 3 Examples for tIH and tIRS for rate 1/2 codes of several lengths for hRS = 2 and
hH = 3.
q
23
24
25
26
27
n = q3
512
4096
32768
262144
2097152
k= n
2
256
2048
16384
131072
1048576
tRS
150
1200
9598
76780
614242
tIRS
190
1516
12126
97004
776029
tIH
183
1555
12844
104478
842936
tIH /tRS ≈
1.22
1.30
1.34
1.36
1.37
tIH /tIRS ≈
0.96
1.03
1.06
1.08
1.09
to h-interleaved Reed–Solomon codes in [12], and the power decoder for one-point
Hermitian codes in [4, 6].
The maximal decoding radius of the new algorithm is n(1 − (R +
ω +2
3
g−1 hh
+1
n )
− ε) at
a cost of O∼ ((1/ε)hω+1 n
) operations over Fq2 , where 2 ≤ ω ≤ 3 is the matrix
multiplication exponent, and improves upon previous best decoding radii at all rates.
Experimental results indicate that the algorithm achieves this maximal decoding radius
with large probability.
For large n, interleaved one-point Hermitian codes achieve larger maximal decoding
radii than interleaved Reed–Solomon codes when compared for the same length and
overall field size.
In the case h = 1, we obtain a one-point Hermitian codes equivalent of the improved
power decoder for Reed–Solomon codes in [5], which achieves a similar decoding radius
as the Guruswami–Sudan list decoder. Simulation results indicate that the new decoder
has a similar failure probability for numbers of errors beyond the latter’s guaranteed
decoding radius.
As for any other power decoding algorithm, both for Reed–Solomon and one-point
Hermitian codes, deriving analytic bounds on the failure probability remains an open
problem. So far, the only parameters for which such an expression is known are h = 1,
` ≤ 3, and s ≤ 2, cf. [2, 5].
References
1. V. Guruswami and M. Sudan, “Improved Decoding of Reed–Solomon and AlgebraicGeometric Codes,” in IEEE Annual Symposium on Foundations of Computer Science,
1998, pp. 28–37.
2. G. Schmidt, V. R. Sidorenko, and M. Bossert, “Syndrome Decoding of Reed–Solomon Codes
Beyond Half the Minimum Distance Based on Shift-Register Synthesis,” IEEE Transactions
on Information Theory, vol. 56, no. 10, pp. 5245–5252, 2010.
3. S. Kampf, “Decoding Hermitian Codes - An Engineering Approach,” Ph.D. dissertation,
Universität Ulm, 2012.
4. J. S. R. Nielsen and P. Beelen, “Sub-Quadratic Decoding of One-Point Hermitian Codes,”
IEEE Transactions on Information Theory, vol. 61, no. 6, pp. 3225–3240, 2015.
5. J. Rosenkilde, “Power Decoding Reed–Solomon Codes up to the Johnson Radius,” Accepted
for: Advances in Mathematics of Communications, (2018), arXiv preprint arXiv:1505.02111.
6. S. Kampf, “Bounds on Collaborative Decoding of Interleaved Hermitian Codes and Virtual
Extension,” Designs, codes and cryptography, vol. 70, no. 1-2, pp. 9–25, 2014.
7. V. Y. Krachkovsky and Y. X. Lee, “Decoding for Iterative Reed–Solomon Coding Schemes,”
IEEE Trans. Magn., vol. 33, no. 5, pp. 2740–2742, 1997.
18
Sven Puchinger, Johan Rosenkilde, Irene Bouw
8. F. Parvaresh and A. Vardy, “Multivariate Interpolation Decoding Beyond the Guruswami–
Sudan Radius,” in Proceedings of the 42nd Allerton Conference on Communication, Control
and Computing, 2004.
9. G. Schmidt, V. Sidorenko, and M. Bossert, “Enhancing the Correcting Radius of Interleaved
Reed-Solomon Decoding using Syndrome Extension Techniques,” in IEEE ISIT, 2007, pp.
1341–1345.
10. H. Cohn and N. Heninger, “Approximate Common Divisors via Lattices,” The Open Book
Series, vol. 1, no. 1, pp. 271–293, 2013.
11. A. Wachter-Zeh, A. Zeh, and M. Bossert, “Decoding Interleaved Reed–Solomon Codes
Beyond Their Joint Error-Correcting Capability,” Designs, Codes and Cryptography, vol. 71,
no. 2, pp. 261–281, 2014.
12. S. Puchinger and J. Rosenkilde né Nielsen, “Decoding of Interleaved Reed-Solomon Codes
Using Improved Power Decoding,” IEEE International Symposium on Information Theory,
2017.
13. S. Puchinger, I. Bouw, and J. Rosenkilde né Nielsen, “Improved Power Decoding of OnePoint Hermitian Codes,” in International Workshop on Coding and Cryptography, 2017,
arXiv:1703.07982.
14. G.-L. Feng and K. K. Tzeng, “A Generalized Euclidean Algorithm for Multisequence ShiftRegister Synthesis,” IEEE Transactions on Information Theory, vol. 35, no. 3, pp. 584–594,
1989.
15. J. S. R. Nielsen, “Generalised Multi-sequence Shift-Register Synthesis using Module
Minimisation,” in IEEE International Symposium on Information Theory, 2013, pp. 882–886.
[Online]. Available: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6620353
16. B. Beckermann and G. Labahn, “A Uniform Approach for the Fast Computation
of Matrix-Type Padé Approximants,” SIAM Journal on Matrix Analysis and
Applications, vol. 15, no. 3, pp. 804–823, Jul. 1994. [Online]. Available: http:
//epubs.siam.org/doi/abs/10.1137/S0895479892230031
17. W. A. Stein et al., “SageMath Software,” http://www.sagemath.org.
18. Y. Wu, “New List Decoding Algorithms for Reed-Solomon and BCH Codes,” IEEE Transactions on Information Theory, vol. 54, no. 8, pp. 3611–3630, 2008.
19. V. Sidorenko, G. Schmidt, and M. Bossert, “Decoding Punctured Reed–Solomon Codes up
to the Singleton Bound,” in International ITG Conference on Source and Channel Coding,
2008, pp. 1–6.
| 7cs.IT
|
Bayesian Conditional Cointegration
Chris Bracegirdle
[email protected]
David Barber
[email protected]
Centre for Computational Statistics and Machine Learning, University College London, Gower Street, London
Abstract
Cointegration is an important topic for timeseries, and describes a relationship between
two series in which a linear combination is
stationary. Classically, the test for cointegration is based on a two stage process in
which first the linear relation between the series is estimated by Ordinary Least Squares.
Subsequently a unit root test is performed
on the residuals. A well-known deficiency of
this classical approach is that it can lead to
erroneous conclusions about the presence of
cointegration. As an alternative, we present a
framework for estimating whether cointegration exists using Bayesian inference which is
empirically superior to the classical approach.
Finally, we apply our technique to model segmented cointegration in which cointegration
may exist only for limited time. In contrast
to previous approaches our model makes no
restriction on the number of possible cointegration segments.
1. Introduction
Cointegration is an important concept in time-series
analysis and relates to the fact that the differences
between two time-series may be more predictable than
any individual time-series itself. We make two contributions to this area—first, we formulate determining
whether two series are cointegrated as an instance of
Bayesian inference in an associated probabilistic model.
This enables us to bring modern statistical and machine
learning techniques to the analysis of cointegration and
also provides a useful conceptual framework to describe
cointegration. The resulting regression estimation algorithm is shown to be more robust than least-squares
estimation to spurious results. This is we believe an
Appearing in Proceedings of the 29 th International Conference on Machine Learning, Edinburgh, Scotland, UK, 2012.
Copyright 2012 by the author(s)/owner(s).
important result that establishes a new and coherent
approach to the long-standing problem of inconsistency
in establishing cointegration. Second, in practice two
series may only be intermittently cointegrated—that
is, they are only cointegrated over shorter segments of
time. To date the identification of these segments has
been attempted with rather limiting assumptions such
as the presence of either only two or three segments
(Gregory & Hansen, 1996; Hatemi-J, 2008). To address
this we phrase the problem as inference in a corresponding changepoint model, placing no limitation on the
number of segments. We develop an efficient specialised
inference scheme for this model and demonstrate its
practical applicability on real-world problems.
1.1. Cointegration
Whilst individual time-series may not be predictable,
the relationships between two time-series may be more
predictable. For example, series x1:T , y1:T formed by
xt+1 = xt + t+1 ,
yt+1 = yt + t+1 ,
t ∼ N (0, 1)
both follow an unpredictable random walk. However,
the future difference xt+1 − yt+1 = xt − yt , is perfectly
predictable given knowledge of the current difference.
In the economics and finance literature such instances
are common; for example underlying mechanisms such
as limited money supply may forge dependencies between time-series (Dickey et al., 1991). It is of significant interest in finance to find pairs of asset price series
that are cointegrated—such estimation underpins one
of the classical statistical arbitrage strategies known as
‘pairs trading’.
From an individual series x1:T we can form a new
series x12:T by taking the difference x1t = xt − xt−1 .
Repeating this differencing process d times, a series is
order d integrated, written I(d), if the series xdt formed
from repeatedly taking the difference d times yields
a stationary series I(0). In our example above, both
x and y are each I(1) and hence integrated. As we
showed, there is a linear combination that is I(0). More
generally, two series x1:T and y1:T are cointegrated
Bayesian Conditional Cointegration
if they are each individually integrated and a linear
combination of the two is integrated with a lower order.
For our purposes, we define xt and yt to be cointegrated
within a specified time segment if there is a linear
regression relationship between the two variables that
forms a stationary process. For the most simple case
of a single regressor, we write such a relationship as
we form a generative model on observations1 y1:T and
latent variables 1:T
Y
p(yt xt , t ) p(t t−1 ) , 0 = ∅
p(y1:T , 1:T x1:T ) =
t
where2
p(yt xt , t ) = δ(yt − α − βxt − t )
yt = α + βxt + t
t = φt−1 + ηt ,
ηt ∼ N 0, σ 2 ,
|φ| < 1
and the transition for t is given as
p(t t−1 ) = N t φt−1 , σ 2 .
where α represents a constant, and β the regression coefficient. Here 1:T forms a mean-reverting, stationary
process. For brevity, we focus attention on estimation
in this model, however more flexible cases can easily be
considered, including the case with a trend coefficient
γ in which the regression is written
yt = α + βxt + γt + t
with the same autoregression for t . Adapting our
technique to more complex models with exogenous
variables, and to the vector case, is also possible.
Testing for and estimating a cointegration relationship is classically a two-step process (Granger, 1986).
Firstly, the regression equation is estimated based on
a
least squares (OLS) fit to minimise
Psimple ordinary
2
t (yt −α−βxt ) (which we show is equivalent to maximum likelihood (ML) parameter optimisation assuming
that φ = 0 in a corresponding model). Subsequently,
a test for a unit root in the residuals is performed using the Dickey-Fuller test, see for example Harris &
Sollis (2003), which tests the hypothesis that φ = 1
against the alternative |φ| < 1 (for |φ| < 1 the series is
stationary and non-stationary otherwise). In the case
that φ = 1, it is well-known that OLS can deliver a
spurious regression (Granger & Newbold, 1974), and
this problem is not limited to cointegration. The classical approach is conceptually undesirable since it makes
strong, potentially conflicting assumptions about the
data: in the regression part, the residuals are effectively assumed uncorrelated whilst in the subsequent
unit root test, they may be determined not to be (under the null hypothesis of the test, the regression was
spurious). In this work we relax the assumptions placed
on φ during parameter estimation to reduce the impact
of any conflicting assumptions.
2. Modelling Cointegration
Our approach is to form a generative
model of observations p(y1:T x1:T , θ), where θ = α, β, σ 2 , φ . First,
The belief network (see Barber (2012) for an introduction) for this model is given in figure 1. The likelihood
on the observations y1:T is given by
Z
p(y1:T x1:T ) =
p(y1:T , 1:T x1:T ) .
1:T
The integration distributes,
TY
−1
Z
p(y1:T x1:T ) =
p(yt xt , t ) p(t t−1 )
1:T −1 t=1
Z
×
p(yT xT , T ) p(T T −1 )
T
and since we have a delta function, the
final term
evaluates to N yT − α − βxT φT −1 , σ 2 . Iterating
this yields a product of Gaussian terms
T
Y
N yt − α − βxt φ (yt−1 − α − βxt−1 ) , σ 2 .
t=2
Finally, applying labels according to t = yt − α − βxt ,
p(y1:T x1:T ) = p(1 )
T
Y
N t φt−1 , σ 2 = p(1:T ) .
t=2
Hence the likelihood p(y1:T x1:T ) is equivalent to the
likelihood on the Markov chain with ‘observations’ t
shown in figure 2. The log likelihood log p(1:T ) is
log p(1 ) −
1
T
1 X
T −1
2
(t − φt−1 ) −
log 2πσ 2
2
2σ t=2
2
Note that we make no assumption about the underlying process x; we work with the conditional relationship
p(y1:T x1:T ) not the joint p(x1:T , y1:T ). As far as we are
aware previous Bayesian approaches to cointegration make
stronger assumptions about the underlying process, see for
example Koop et al. (2006).
2
The Dirac delta distribution δ(x − a) represents a deRgenerate probability density function with the property that
δ(x − a) f (x) = f (a) .
x
Bayesian Conditional Cointegration
1
2
3
...
T
x1
x2
x3
...
xT
y1
y2
y3
...
yT
Figure 1. Belief network for the natural model for cointegration. Diamonds represent delta functions, shaded variables
are in the conditioning set.
1
2
3
...
T
Figure 2. Belief network for the Markov chain model for t .
3.1. Inference of φ
The process 1:T is stationary when each3 ht i = 0
and 2t is constant, independent of t. According to
the recurrence relation for t , the variance satisfies
2t+1 = φ2 2t + σ 2 and we can therefore
ensure
2
2
2
stationary of 1:T if 1 = σ / 1 −
φ . We then
choose p(1 φ) = N 1 0, σ 2 / 1 − φ2 . The posterior
p(φ 1:T ) is proportional to
"
#
T
X
p
−1
2
p(φ) 1 − φ2 exp 2 21 1 − φ2 +
(t − φt−1 )
2σ
t=2
and by completing the square we see that the posterior
is given by a truncated Gaussian distribution with a
prefactor (subject to normalisation),
p
ê12 σ 2
,
p(φ) 1 − φ2 N φ
ê1 ê1
where
and when φ = 0, maximising this degenerates to minimising the sum of squared residual terms
ê12 ≡
T
X
ê1 ≡
t t−1 ,
t=2
T
X
t=2
2t =
T
X
2
(yt − α − βxt ) .
T
X
2t−1 .
t=3
For the data likelihood, we write for p(1:T ),
t=2
OLS estimation for the regression parameters α and β
therefore corresponds to the ML solution of this model,
on the assumption that φ = 0. In the case that φ is
non-zero, a different solution may be optimal in the
ML sense. This is a source of potential inconsistency
in the classical approach to cointegration testing.
2πσ
2
1−T
2
!
T
2
X
(ê12 )
1
2
t −
exp − 2
2σ
ê1
t=1
Z 1 p
1
ê
σ2
1
2 N φ 12 ,
×√
1
−
φ
.
ê1 ê12
ê1 −1 2
φ
3.2. Estimating α, β, σ 2
3. Bayesian Cointegration
We can construct a model for estimating both φ and
the parameters α, β, σ 2 by considering φ to be a latent
variable and providing a prior distribution. In this case,
for cointegration we require |φ| < 1, and we can encode
this with a uniform distribution p(φ) = U(φ (−1, 1)) =
1
2 [φ ∈ (−1, 1)]. We complete the model specification
with a distribution for 1 and form the joint model
p(y1:T , 1:T , φ x1:T ) = p(y1:T 1:T , x1:T ) p(1:T φ) p(φ)
from which the marginal model
φ
Since φ is a latent variable, it is convenient to approach
this using the Expectation-Maximisation algorithm.
Writing v for the observations and h for the latent
variables EM is based on the Kullback-Leibler bound
q(h v)
0 ≤ KL(p q) ≡ log
p(h v) q(h v)
log p(v) ≥ hlog p(h, v)iq(h v) − hlog q(h v)iq(h v)
|
{z
} |
{z
}
Z
p(y1:T , φ x1:T ) = p(φ)
Our interest is to set the parameters θ = α, β, σ 2
based on maximising the likelihood
Z
p(y1:T x1:T , θ) =
p(φ) p(1:T φ) .
energy
p(y1:T 1:T , x1:T ) p(1:T φ)
1:T
is formed by integration. The posterior p(φ x1:T , y1:T )
is then equivalent to p(φ 1:T ) ∝ p(φ) p(1:T φ).
entropy
where p is a variational distribution and q a fixed posterior. EM corresponds to iteratively maximising this
3
Angled brackets represent expectation, hxi = E [x].
Bayesian Conditional Cointegration
lower bound with respect to the variational distribution
p, and using this as the next q following inference. For
this model, we replace h → φ, v → 1:T .
The energy term4 is the log of the model joint, and
since we are interested in maximising with respect to
the regression parameters α and β the only relevant
terms are given as
hlog p(1 φ)iq(φ
1:T ) +
T
X
hlog p(t t−1 , φ)iq(φ
Algorithm 1 Bayesian cointegration testing
α, β, σ 2 ← LinearRegression (x1:T , y1:T )
repeat
1:T ← y1:T − α − βx1:T
2
2
l
,
hφi
,
φ
←
CointInference
,
σ
C
1:T
α, β, σ 2 ← EM x1:T , y1:T , hφi , φ2
until convergence
QT
b2
lRW ← t=2 N t t−1 , σ
return cointegrated ← lRW /lC < threshold
1:T )
t=2
and up to a constant, this equals
"
#
T D
E
X
−1 2
2
2
1−φ +
(t − φt−1 )
−
2σ 2 1
t=2
T
2
log 2πσ 2
where q(φ 1:T ) = p φ 1:T , θold is the posterior from
the previous iteration. The energy can be optimised
by finding the stationary point by differentiating by α
and β. Since
D
E
2
(t − φt−1 ) = 2t − 2t t−1 hφi + 2t−1 φ2
the result is a system of linear equations for α and β
involving the first and second (non-central) moments
of φ from the posterior q(φ 1:T ).
By differentiating the above energy term with respect
to σ 2 , we find that optimally
"
#
T D
E
X
1 2
2
2
2
σ
b =
1−φ +
(t − φt−1 )
T 1
t=2
so the variance can be estimated once the new regression estimates α and β have been found.
4. The Random Walk model
A primary concern in the economics literature is to test
the hypothesis of cointegration for series. The analogue
with our Bayesian generative model is to compare the
likelihoods of the above cointegration model and a
random walk (RW) model. The likelihood for 1:T a
RW is calculated as
p(1:T φ = 1) = p(1 )
T
Y
N t t−1 , σ 2
t=2
For p(1 ) we choose a wide-interval uniform distribution.
The cointegration model and random walk model can
then be compared according to the Bayes factor,
p(y1:T x1:T ; φ = 1)
p(1:T φ = 1)
lRW
=
≡
p(y1:T x1:T ; |φ| < 1)
p(1:T |φ| < 1)
lC
4
Also called the expected completed data log likelihood.
where the numerator represents the ‘null hypothesis’
of a unit root in the residuals process. The denominator is simply the marginal likelihood for the Bayesian
cointegration model given in section 3.1.
Whilst a fully-Bayesian approach to cointegration testing may be considered by placing prior distributions
and seeking to integrate over the parameters α, β, σ 2
separately for each model, here we restrict our analysis
to point estimation for reasons of both simplicity and
computational speed. Furthermore we wish to relate
closely to the classical point estimate approach to cointegration estimation and testing. The purpose of this
study is demonstrate the potential advantage of using
a ‘partial’ Bayesian approach for φ alone.
4.1. Point Estimation
Both the numerator and denominator of the Bayes
factor are functions in the variance σ 2 , and we set σ
for each of the two models by ML. The estimate for σ 2
for the RW model is given according to
T
σ
b2 =
1 X
2
(t − t−1 ) .
T − 1 t=2
Note that here t depends on α and β, and we now
discuss possible ways to determine point estimates for
these parameters. There are three primary options: (i)
find the parameters by OLS and use these values for
both the cointegration and RW models; (ii) find the
parameters by ML in the cointegration model and use
the same parameters for the RW model; and (iii) find
the parameters for each of the cointegration and RW
models by ML. Conceptually, our interest is to first
estimate a cointegration relationship, and second to test
whether this relationship is more likely to be a random
walk—and for this reason, we consider the third option
undesirable. Whilst the first option is computationally
the simplest it suffers from the implicit assumption
that φ = 0. For this reason the second is our preferred
method.
Algorithm 1 shows the calculation steps for the
Bayesian cointegration test. The results in figure 3
Bayesian Conditional Cointegration
1
0.7
DF FP
DF FN
Bayes FP
Bayes FN
0.99
0.6
0.98
True Positive Rate
Error rate
0.5
0.4
0.3
0.2
0.97
0.96
0.95
0.94
0.93
0.92
0.1
OLS/DF
OLS/Likelihood
EM/DF
EM/Likelihood
0.91
0.9
0
0
50
100
150
200 250 300
Series length T
350
400
450
500
Figure 3. Plot of false positive (FP) and false negative (FN)
rates (calculated as number of FP/FN outcomes divided
by number of true negatives/positives respectively) for the
classical OLS/DF approach (5% significance) and Bayesian
EM/likelihood method (threshold C = exp 2) for generated
series. 5000 series of each length were simulated according
to the generative model, uniformly split between |φ| < 1
in the cointegrated case and φ = 1 in the random walk
case. The high rate of false positives for the classical test
is striking: here cointegration was detected when the data
were generated with φ = 1. We attribute this to the fact
that OLS is known to produce invalid estimates when the
process is non-stationary: so called ‘spurious regression’.
show that, compared with OLS estimation and unit
root testing, our Bayesian technique is less likely to
result in a spurious relationship for series of length
T > 20. In figure 4 we show comparisons for differing
decision boundaries5 .
5. Intermittent Cointegration
For some series, cointegration may only apply for certain segments. A good example is the Interconnector
gas pipeline between Bacton, UK and Zeebrugge, Belgium, which allows gas to flow between the two countries, providing a direct link in the gas price at each end.
When the pipeline closes each year for maintenance,
the link between the prices is temporarily broken. It
was shown by Kim (2003) that classical tests for cointegration may fail to detect a relationship even though,
for a part of the series, there is such a relationship. The
following model seeks to detect a cointegration relationship between two series, while allowing for regions
when the relationship is in fact a random walk.
5
Differing decision thresholds correspond to different
widths of uniform prior for 1 in the RW model since
˜
lRW
lRW
<C⇔
< 1,
lC
lC
lRW
˜
lRW ≡
.
C
0
0.1
0.2
0.3
0.4 0.5 0.6 0.7
False Positive Rate
0.8
0.9
1
Figure 4. Plot of the receiver operating characteristic (ROC)
curve for the cointegration estimation/testing combinations,
for T = 100. 10,000 series were generated; we plot the true
positive rate against the false positive rate for different
values of the decision threshold. EM refers to estimation
in the Bayesian cointegration model by ML; DF refers to
a Dickey-Fuller test; Likelihood refers to a comparison of
the likelihood in the cointegration model with the random
walk model φ = 1. Optimal classification occurs with true
positive rate 100%, false positive rate 0%—the results show
that OLS/DF is dominated by EM/Likelihood.
Previous works in this area typically limit the number
of regimes in which cointegration can occur to either
only two or three segments (Gregory & Hansen, 1996;
Hatemi-J, 2008). In contrast, for our model we allow
time-varying φt arranged to be piecewise-constant with
no a priori restriction on the number of piecewiseconstant regions. The model will switch between the
cointegrated case |φt | < 1 and the alternative φt = 1,
corresponding to a random walk.
We use the binary state switch it to denote whether
there is a cointegration relationship at time t: it =
1 denotes regions of t corresponding to a random
walk; alternatively when it = 0, t follows a stationary
cointegration relation. In the random walk regions, we
have φt = 1, which is encoded in the prior distribution
p1 (φt ) = δ(φt − 1). We therefore specify
(
p1 (φ2 ) = δ(φ2 − 1)
i2 = 1
p(φ2 i2 = 0) =
0
p (φ2 ) = U(φ2 (−1, 1)) i2 = 0
and for the piecewise-constant transition,
1
it = 1
p (φt )
p(φt φt−1 , it , it−1 ) = δ(φt − φt−1 ) it = 0, it−1 = 0
0
p (φt )
it = 0, it−1 = 1.
Whilst the state transition p(it it−1 ) and p(i2 ) can
also in principle be learned on the basis of ML, we
Bayesian Conditional Cointegration
1
i2
i3
...
iT
φ2
φ3
...
φT
2
3
...
T
Figure 5. Belief network for the intermittent cointegration
model. Dashed edges to φt are absent when it = 1.
leave these quantities to be user specified. Note that
whilst φt can change with time, for computational
tractability we assume that the remaining parameters
α, β, σ 2 are fixed throughout time. The belief network
representation is given in figure 5.
The switching inference routine described in the following section relies on the recursive updates given in
appendix A, in which we use the uniform prior for t
at the start of each cointegrated segment.
5.1. Inference of i2:T , φ2:T
The regime-switching model with piecewise-constant φt
is an example of a changepoint model, in which a change
of regime drops the dependency of the continuous latent
variable φt on the past. The model is therefore an
example of the ‘reset model’ as set out by Bracegirdle
& Barber (2011), and we apply such results here. Exact
inference is then possible by applying recursions for
filtering (calculating each p(φt , it 1:t )) and correctionsmoothing (for p(φt , it 1:T )). For brevity, we show here
only some of the key considerations; a full derivation
is given in the supplementary material.
It is easy to show that both the filtered and smoothed
posteriors for φt in the case that it = 1 degenerate in
an intuitive way to
p(φt it = 1, 1:t ) = p(φt it = 1, 1:T ) = p1 (φt ) .
For the case it = 0, the posterior for φt is more complex
because of the temporal dependency. For filtering, the
posterior depends on the previous regime: in the event
that it−1 = 1, a new cointegration regime has begun,
and a Gaussian term is contributed to the posterior
in accordance with (2); otherwise, each component is
updated as for the simple cointegration case (1). We
therefore see that the posterior p(φt it = 0, 1:t ) is given
as a truncated mixture of Gaussian components.
For correction smoothing, it can be difficult to derive a
recursion analytically because the backwards-recursive
step requires a ‘dynamics reversal’ term that is calculated with the filtered posterior in the denominator. In
the case that the filtered posterior is a mixture distribution, the algebra is intractable. However, we appeal
to the result of Bracegirdle & Barber (2011) that by
indexing the components in the filtered posterior according to the number of observations since a switch
to cointegration, the smoothing recursion can be written exactly. The resulting recursion shows that, as we
may expect, a subset of the Gaussian components from
p(φt+1 it+1 = 0, 1:T ) are contributed to the posterior
p(φt it = 0, 1:T ) without change, along with the components from filtering p(φt it = 0, 1:t ) in the case that
the cointegration regime ends at t.
The discrete filtering and smoothing components,
p(it 1:t ) and p(it 1:T ), are also calculated as part of
the recursions. The result is an algorithm for exact
inference in this switching model that scales as T 2 .
5.2. Likelihood
It is useful to evaluate the likelihood in order to (i)
ensure that the likelihood is maximised, and (ii) provide
a convergence criterion. Fortunately the likelihood
is easy to calculate since the value is given by the
product of normalisation constants from each step of
the filtering recursion,
p(1:T ) = p(1 )
T
Y
p(t 1:t−1 ) = p(1 )
t=2
T
Y
Zt
t=2
where each Zt is calculated when filtering, see the
supplementary material.
5.3. Learning
EM is again used for parameter estimation in this
model. For this regime-switching problem, the latent
variables are h → {i2:T , φ2:T }, and the observations
remain as before, v → 1:T .
Terms relevant to the optimal solution are the sum of
quadratic forms derived in section 3.2 with varying φt ,
D
E
2
(t − φt t−1 ) = 2t − 2t t−1 hφt i + 2t−1 φ2t .
By retaining the first and second (non-central) moments
of each component found while filtering, the linear
system of equations for the regression parameters can
be solved exactly in this switching model, and the
variance estimate can be updated accordingly.
5.4. Point Estimate of φt
Whilst our model gives a distribution over φt for each
timepoint, in order to be able to compare our approach
Bayesian Conditional Cointegration
2.5
1
5.5
0.8
1
0.8
5
2
0.6
0.6
4.5
0.4
0.4
1.5
4
0.2
1
0.2
0
Sep 2010
Sep 2011
3.5
Sep 2010
0.4
1.2
Mar 2008
Mar 2008
0.8
1.6
0.6
1.4
0.6
0.4
1.2
0.4
0.2
1
0
0.8
1
0.2
0
Sep 2011
0.8
0
−0.2
0.2
−0.4
−0.6
0
Sep 2010
−0.2
Sep 2011
Sep 2010
Sep 2011
Figure 6. Results of learning for the Interconnector gas
prices—maintenance occurs each September. We show
(NW) the price series xt , yt ; (SW) residuals t calculated
with the final estimates for α, β; (NE) filtered (blue) and
smoothed (red) posterior for random walk it = 1; and (SE)
the maximum posterior marginal result for φt .
to other intermittent cointegration methods that only
give point estimates, we may take each
bit = arg max p(it 1:T )
it
and then calculate the posteriors p φt bi2:T , 1:T to
give the point estimates
φbt = arg max p φt bi2:T , 1:T .
φt
6. Experiments
We consider real-world applications for the intermittent
cointegration model. For these experiments, we ran the
inference and learning algorithm for the parameters θ =
{α, β, σ} initialised by OLS, and use the state transition
distributions fixed at values designed to match our a
priori belief about the cointegration regimes.
6.1. Gas Prices
The Interconnector, as noted in section 5, is a sub-sea
natural gas pipeline: we took for y UK gas prices6 and
for x the Zeebrugge prices7 . There are approximately
6
UK SAP natural gas prices were downloaded from
www.nationalgrid.com/uk/Gas/Data.
7
Prices for Zeebrugge are from www.apxendex.com and
were converted into kWh using 1 kWh=29.3072222 therm.
−0.2
Mar 2008
2
4
6
8
10
Figure 7. Results of learning for the Euro-area bond yields.
We show (NW) the yield series xt , yt ; (SW) residuals t
calculated with the final estimates for α, β; (NE) filtered
(blue) and smoothed (red) posterior for random walk it = 1;
and (SE) the convergence of parameters α, β—note the
relative change in the estimate for intercept α (green).
245 pricing days per year, and the pipeline closes annually for two weeks, so we specified the state transition
distribution accordingly. The algorithm reaches convergence within four minutes with parameters α, β that
show cointegration between the annual maintenance
period, see figure 6.
6.2. Euro-area Bond Yields
We consider possible cointegration between Greek and
German 10-year benchmark bond yields8 prior to the
Euro-area financial crisis. We use the cointegration
switching model to find regions of cointegration, which
we expect to hold until the yield on Greek debt spiralled as Greece’s debt burden took toll. The results
are shown in figure 7; inference and learning takes
roughly five minutes. This is an example for which,
over the time window shown in the figure, the data
do not show cointegration according to the classical
test (Dickey-Fuller test shows p-value 0.927241, and
does not reject the null hypothesis of random walk),
but the intermittent cointegration model does show a
segment of cointegration—the classical test passes in
the detected region, and fails for the remainder.
Whilst our approach is both flexible and simple, there
are limitations. In particular, the switching model
8
The bond yield data were obtained from Reuters.
Bayesian Conditional Cointegration
requires that the regions of cointegration are governed
by a time-invariant linear relationship. For example,
in the case that deviations from the relationship in
the random-walk segments are permanent, the model
presented here would struggle to properly estimate
a relationship since piecewise-constant values of the
constant term α would be required. Such considerations
provide opportunities for future research.
Gregory, A. W. and Hansen, B. E. Residual-based tests
for cointegration in models with regime shifts. Journal of Econometrics, 70(1):99–126, January 1996.
7. Discussion
Hatemi-J, A. Tests for cointegration with two unknown
regime shifts with an application to financial market
integration. Emp. Economics, 35(3):497–505, 2008.
We presented two novel techniques for estimation of a
linear cointegration relationship for time-series. First,
the method in section 3 allows estimation of a linear
relationship between variables and we showed that our
resulting algorithm is less likely to deliver a spurious
relationship than the classical OLS–unit root testing
approach. The benefit of our approach is that it results
in a fast O(T ) algorithm for testing cointegration. A
natural extension would be to consider a more complete
Bayesian analysis and marginalise over all parameters
in both the cointegration and random walk models.
Second, we developed a switching model to detect a
relationship and regimes in which that relationship
is a random walk. Whilst the model has an obvious
limitation (shared by other existing intermittent cointegration models) it has the benefit of fast exact inference,
scaling O(T 2 ). A further natural extension would be
to consider regimes in which the linear relationship can
also vary. A full derivation of inference for the switching
model is given in the supplementary material.
Acknowledgements
We thank Ricardo Silva for useful discussions.
References
Barber, D. Bayesian Reasoning and Machine Learning.
Cambridge University Press, 2012.
in econometrics. Journal of Econometrics, 2(2):111–
120, 1974.
Harris, R. and Sollis, R. Applied Time Series Modelling
and Forecasting. Wiley, 2003.
Kim, J. Y. Inference on Segmented Cointegration.
Econometric Theory, 19(4):620–639, 2003.
Koop, G. M., Strachan, R. W., Van Dijk, H., and
Villani, M. Bayesian approaches to cointegration. In
The Palgrave Handbook of Theoretical Econometrics,
pp. 871–898. Palgrave Macmillan, 2006.
A. Sequential inference of φ
Given a sequence of ‘observations’ 1:T we describe
here a method to sequentially update the posterior
distribution p(φ 1:T ). An extension of this inference
routine is used in the intermittent cointegration model
in section 5. We place an improper uniform prior
p(1 ) = U(1 R), representing our belief that 1 may
take any real value, and use a recursive update filtering
routine—see for example Barber (2012).
Initially, we begin with the prior p(φ) = U(φ (−1, 1)),
and then update the distribution by finding the
posterior after observing each t . For 1 , there
is no update to make since p(φ 1 ) ∝ p(φ) p(1 ).
Thereafter, p(φ 1:t ) ∝ p(t t−1 , φ) p(φ 1:t−1 ).
This can be calculated analytically since, if
p(φ 1:t−1 ) ∝ p(φ) N (φ ft−1 , Ft−1 ), then9 p(φ 1:t ) ∝
p(φ) N (φ ft , Ft ) where
ft =
Bracegirdle, C. and Barber, D. Switch-Reset Models :
Exact and Approximate Inference. In AISTATS,
volume 15. JMLR, 2011.
Dickey, D. A., Jansen, D. W., and Thornton, D. L.
A primer on cointegration with an application to
money and income. Federal Reserve Bank of St.
Louis Review, (Mar):58–78, 1991.
Granger, C. W. J. Developments in the study of cointegrated economic variables. Oxford Bulletin of Economics and Statistics, 48(3):213–228, 1986.
Granger, C. W. J. and Newbold, P. Spurious regressions
ft−1 σ 2 + t t−1 Ft−1
,
σ 2 + 2t−1 Ft−1
Ft =
σ2
σ 2 Ft−1
(1)
+ 2t−1 Ft−1
which on setting the initial10 f2 , F2 from the emission
1
2 σ 2
2
N 2 φ1 , σ =
N φ , 2
(2)
|1 |
1 1
defines a recursion for the parameters ft , Ft . After completing the updates, the required posterior p(φ 1:T ) is
given by a Gaussian distribution with mean fT and
covariance FT truncated to the interval (−1, 1).
9
Details of the derivation are omitted; recursions correspond to Kalman updates—see for example Barber (2012).
10
In the unlikely event that 1 = 0, the Gaussian term
arises from the first non-zero t−1 .
| 5cs.CE
|
A Note on the Practicality of
Maximal Planar Subgraph Algorithms
Markus Chimani1 , Karsten Klein2 , Tilo Wiedera1
arXiv:1608.07505v1 [cs.DS] 26 Aug 2016
1
University Osnabrück, Germany
{markus.chimani,tilo.wiedera}@uni-osnabrueck.de
2
Uni Konstanz, Germany
[email protected]
Abstract. Given a graph G, the NP-hard Maximum Planar Subgraph
problem (MPS ) asks for a planar subgraph of G with the maximum
number of edges. There are several heuristic, approximative, and exact
algorithms to tackle the problem, but—to the best of our knowledge—
they have never been compared competitively in practice.
We report on an exploratory study on the relative merits of the diverse
approaches, focusing on practical runtime, solution quality, and implementation complexity. Surprisingly, a seemingly only theoretically strong
approximation forms the building block of the strongest choice.
1
Introduction
We consider the problem of finding a large planar subgraph in a given non-planar
graph G = (V, E); n := |V |, m := |E|. We distinguish between algorithms that
find a large, maximal, or maximum such graph: while the latter (MPS) is one with
largest edge cardinality and NP-hard to find [18], a subgraph is inclusionwise
maximal if it cannot be enlarged by adding any further edge of G. Sometimes,
the inverse question—the skewness of G—is asked: find the smallest number
skew (G) of edges to remove, such that the remaining graph is planar.
The problem is a natural non-trivial graph property, and the probably best
known non-planarity measure next to crossing number. This already may be
reason enough to investigate its computation. Moreover, MPS/skewness arises
at the core of several other applications: E.g., the practically strongest heuristic
to draw G with few crossings—the planarization method [2, 7]3 —starts with a
large planar subgraph, and then adds the other edges into it.
Recognizing graphs of small skewness also plays a crucial role in parameterized complexity: Many problems become easier when considering a planar graph;
e.g., maximum flow can be computed in O(n log n) time, the Steiner tree problem allows a PTAS, the maximum cut can be found in polynomial time, etc. It
hence can be a good idea to (in a nutshell) remove a couple of edges to obtain
a planar graph, solve the problem on this subgraph, and then consider suitable
3
In contrast to this meaning, the MPS problem itself is also sometimes called planarization. We refrain from the latter use to avoid confusion.
2
Markus Chimani, Karsten Klein, Tilo Wiedera
modifications to the solution to accommodate for the previously ignored edges.
E.g., we can compute a maximum flow in time O(skew (G)3 · n log n) [13].
While solving MPS is NP-hard, there are diverse polynomial-time approaches
to compute a large or maximal planar subgraph, ranging from very trivial to
sophisticated. By Euler’s formula we know that already a spanning tree gives a
1/3-approximation for MPS. Hence all reasonable algorithms achieve this ratio.
Only the cactus algorithms (see below) are known to exhibit better ratios. We will
also consider an exact MPS algorithm based on integer linear programs (ILPs).
All algorithms considered in this paper are known (for quite some time, in
fact), and are theory-wise well understood both in terms of worst case solution
quality and running time. To our knowledge, however, they have never been
practically compared. In this paper we are in particular interested in the following quality measures, and their interplay:
– What is the practical difference in terms of running time?
– What is the practical difference in solution quality (i.e., subgraph density)?
– What is the implementation effort of the various approaches?
Overall, understanding these different quality measures as a multi-criteria setting, we can argue for each of the considered algorithms that it is pareto-optimal.
We are in particular interested in studying a somewhat “blurred” notion of
pareto-optimality: We want to investigate, e.g., in which situations the additional sophistication of algorithms gives “significant enough” improvements.4
Also the measure of “implementation complexity” is surprisingly hard to concisely define, and even in the field of software-engineers there is no prevailing
notion; “lines of code” are, for example, very unrelated to the intricacies of algorithm implementation. We will hence only argue in terms of direct comparisons
between pairs of algorithms, based on our experience when implementing them.5
As we will see in the next section, there are certain building blocks all algorithms require, e.g., a graph data structure and (except for C, see below) an
efficient planarity test. When discussing implementation complexity, it seems
safe to assume that a programmer will already start off with some kind of graph
library for her basic datastructure needs.6 In the context of the ILP-based approach, we assume that the programmer uses one of the various freely available
(or commercial) frameworks. Writing a competitive branch-and-cut framework
from ground up would require a staggering amount of knowledge, experience,
time, and finesse. The ILP method is simply not an option if the programmer
may not use a preexisting framework.
In the following section, we discuss our considered algorithms and their implementation complexity. In Section 3, we present our experimental study. We
4
5
6
Clearly, there is a systematic weakness, as it may be highly application-dependent
what one considers “significant enough”. We hence cannot universally answer this
question, but aim to give a guideline with which one can come to her own conclusions.
This measure is still intrinsically subjective (although we feel that the situation is
quite clear in most cases), and opinions may vary depending on the implementor’s
knowledge, experience, etc. We discuss these issues when they become relevant.
Many freely available graph libraries contain linear-time planarity tests. They usually
lack sophisticated algorithms for finding large planar subgraphs.
A Note on the Practicality of Maximal Planar Subgraph Algorithms
3
first consider the pure problem of obtaining a planar subgraph. Thereafter, we
investigate the algorithm choices when solving MPS as a subproblem in a typical
graph drawing setting—the planarization heuristic.
2
Algorithms
Naı̈ve Approach (Nı̈). The algorithmically simplest way to find a maximal planar
subgraph is to start with the empty graph and to insert each edge (in random
order) unless the planarity test fails. Given an O(n) time planarity test (we use
the algorithm by Boyer and Myrvold [3], which is also supposed to be among
the practically fastest), this approach requires O(nm) overall time.7
In our study, we consider a trivial multi-start variant that picks the best solution after several runs of the algorithm, each with a different randomized order.
The obvious benefit of this approach is the fact that it is trivial to understand
and implement—once one has any planarity test as a black box.
Augmented Planarity Test (BM, BM+). Planarity tests can be modified to allow
the construction of large planar subgraphs. We will briefly sketch these modifications in the context of the above mentioned O(n) planarity test by Boyer and
Myrvold [3]: In the original test, we start with a DFS tree and build the original
graph bottom-up; we incrementally add a vertex together with its DFS edges to
its children and the backedges from its decendents. The test fails if at least one
backedge cannot be embedded.
We can obtain a large (though in general not maximal) planar subgraph by
ignoring whether some backedges have not been embedded, and continuing with
the algorithm (BM). If we require maximality, we can use Nı̈ as a prostprocessing
to grow the obtained graph further (BM+). While this voids the linear runtime,
it will be faster than the pure naı̈ve approach. Given an implementation of the
planarity testing algorithm, the required modifications are relatively simple per
se—however, they are potentially hard to get right as the implementor needs to
understand side effects within the complex planarity testing implementation.
Alternatively, Hsu [14] showed how to overcome the lack of maximality directly within the planarity testing algorithm [19] (which is essentially equivalent
to [3]), retaining linear runtime. While this approach is the most promising in
terms of running time, it would require the most demanding implementation of
all approaches discussed in this paper (including the next subsection)—it means
to implement a full planarity testing algorithm plus intricate additional procedures. We know of no implementation of this algorithm.
Cactus Algorithm (C, C+). The only non-trivial approximation ratios are achieved
by two cactus-based algorithms [4]. Thereby, one starts with the disconnected
vertices of G. To obtain a ratio of 7/18 (C), we iteratively add triangles connecting formerly disconnected components. This process leaves a forest F of tree-like
7
We could speed-up the process in practice by starting with a spanning tree plus three
edges. However, there are instances where the initial inclusion of a spanning tree
prohibits the optimal solution and restricts one to approximation ratios ≤ 2/3 [8].
4
Markus Chimani, Karsten Klein, Tilo Wiedera
structures made out of triangles—cactusses. Finally, we make F connected by
adding arbitrary edges of E between disconnected components. Since this subgraph will not be maximal in general, we can use Nı̈ to grow it further (C+).
From the implementation point of view, this algorithm is very trivial and—
unless one requires maximality—does not even involve any planarity test. While
a bit more complex than the naı̈ve approach, it does not require modifications
to complex and potentially hard-to-understand planarity testing code like BM.
For the best approximation ratio of 4/9 one seeks not a maximal but a
maximum cactus forest. However, this approach is of mostly theoretical interest
as it requires non-trivial polynomial time matroid algorithms.
ILP Approach (ILP). Finally, we use an integer linear program (ILP) to solve
MPS exactly in reasonable (but formally non-polynomial) time, see [15]. With
binary variables for each edge, specifying whether it is in the solution, we have
nX
o
X
max
xe
xe ≤ |K|−1 for all Kuratowski subdivisions K ⊆ G .
e∈E
e∈K
Kuratowski’s theorem [17] states that a graph is planar if and only if it does
not contain a K5 or a K3,3 as a subdivision—Kuratowski subdivisions. Hence
we guarantee a planar solution by requiring to remove at least one edge in each
such subgraph K. While the set of these constraints is exponential in size, we
can separate them heuristically within a branch-and-cut framework, see [15]:
after each LP relaxation, we round the fractional solution and try to identify a
Kuratowski subdivision that leads to a violated constraint.
This separation in fact constitutes the central implementation effort. Typical planarity testing algorithms initially only answer yes or no. In the latter
case, however, all known linear-time algorithms can be extended to extract a
witness of non-planarity in the form of a Kuratowski subdivision in O(n) time.
If the available implementation does not support this additional query, it can
be simulated using O(n) calls to the planarity testing algorithm, by incrementally removing edges whenever the graph stays non-planar after the removal.
Both methods result in a straight-forward implementation (assuming some familiarity with ILP frameworks), but an additional tuning step to decide, e.g., on
rounding thresholds, is necessary. The overall complexity is probably somewhere
in-between C and BM. In our study, we decided to use the effective extraction
scheme described in [10] which gives several Kuratowski subdivions via a single
call. We propose, however, to use this feature only if it is already available in
the library: its implementation effort would otherwise be comparable to a full
planarity test, and in particular for harder instances its benefit is not significant.
3
Experiments
For an exploratory study we conducted experiments on several benchmark sets.
We summarize the results as follows—observe the inversion between F1 and F2.
F1. C+ yields the best solutions. Choosing a “well-growable” initial subgraph—
in our case a good cactus—is practically important. The better solution of
BM is a weak starting point for BM+; even Nı̈ gives clearly better solutions.
A Note on the Practicality of Maximal Planar Subgraph Algorithms
5
F2. BM gives better solutions than C; both are the by far fastest approaches.
Thus, if runtime is more crucial than maximality, we suggest to use BM.
F3. ILP only works for small graphs. Expander graphs (they are sparse but
well-connected) seem to be among the hardest instances for the approach.
F4. Larger planar subgraphs lead to fewer crossings for the planarization method.
However, this is much less pronounced with modern insertion methods.
Setup & Instances. All considered algorithms are implemented in C++ (g++ 5.3.1,
64bit, -O3) as part of OGDF [5], the ILP is solved via CPLEX 12.6. We use an
Intel Xeon E5-2430 v2, 2.50 GHz running Debian 8; algorithms use singles cores
out of twelve, with a memory limit of 4GB per process.
We use the non-planar graphs of the established benchmark sets North [12]
(423 instances), Rome [11] (8249), and SteinLib [16] (586), all of which include
real-world instances. In our plots, we group instances according to |V | rounded to
the nearest multiple of 10; for Rome we only consider graph with ≥ 25 vertices.
Additionally, we consider two artificial sets: BaAl [1] are scale-free graphs,
and Regular [20] (implemented as part of the OGDF) are random regular
graphs; they are expander graphs w.h.p. [folklore]. Both sets contain 20 instances
for each combination of |V | ∈ {102, 103 , 104 } and |E|/|V | ∈ {2, 3, 5, 10, 20}.
Evaluation. Our results confirm the need for heuristic approaches, as ILP solves
less than 25% of the larger graphs of the (comparably simple) Rome set within
10min. Even deploying strong preprocessing [6] (+PP) and doubling the computation time does not help significantly, cf. Fig. 1(d). Already 30-vertex graphs with
density 3, generated like Regular, cannot be solved within 48 hours (→F3).
We measure solution quality by the density (edges per vertices) of the computed planar subgraph. Independently of the benchmark set, C+ always achieves
the best solutions, cf. Fig. 1(a)(b)(table) (→F1). We know instances where Nı̈
is only a 1/3 approximation whereas the worst ratio known for BM+ is 2/3 [8].
Surprisingly, Nı̈ yields distinctly better solutions than BM+ in practice (→F1).
On SteinLib, BaAl, and Regular, both C and BM behave similar w.r.t.
solution quality. For Rome and North, however, BM yields solutions that are 20–
30% better, respectively (→F2). This discrepancy seems to be due the fact that
the found subgraphs are generally very sparse for both algorithms on BaAl and
Regular (average density of 1.1 and 1.2, respectively, for the largest graphs).
Both C and BM are extremly (and similarly) fast; Fig. 1(c)(table) (→F2).
For BM+ and C+, the Nı̈-postprocessing dominates the running time: Nı̈ is worst,
followed by C+ and BM+—a larger initial solution leads to fewer trys for growing.
Nonetheless, we observe that the (weaker) solution of C allows for significantly
more successful growing steps that BM (→F1).
Finally, we investigate the importance of the subgraph selection for the planarization method, cf. Fig. 1(e)(f). For the simplest insertion algorithms (iterative edge insertions, fixed embedding, no postprocessing, [2]), a strong subgraph
method (C+) is important; C leads to very bad solutions. For state-of-the-art insertion routines (simultaneous edge insertions, variable embedding, strong postprocessing, [7, 9]) the subgraph selection is less important; even C is feasible.
Markus Chimani, Karsten Klein, Tilo Wiedera
S (= constr. sparse): PUC, SP;
V (= VLSI): ALUE, ALUT, LIN, TAQ, DIW, DMXA, GAP, MSM, 1R, 2R
(b) North, solution quality relative to ILP
(over instances solved by ILP),
C = 0.76[0.64, 0.85]
density
20
15
10
1.2
5
1.15
40
60
80
relative density
BM
BM+
C+
Nı̈
# instances (×100)
(a) Rome, solution quality,
C = 1.02[1.01, 1.06]
1.25
0
100
1
80
0.98
60
BM
BM+
C+
Nı̈
0.96
0.94
0.92
50
# vertices
3
15
2
10
1
5
60
80
100
success-rate
runtime [ms]
20
# instances (×100)
25
BM+
C+
Nı̈
40
20
0.75
15
0.5
0.25
0
0
10
ILP, 10min
ILP+PP, 10min
ILP+PP, 20min
40
10
5
5
40
60
80
# vertices
100
0
no. of crossings, relative to best (∆%)
15
# instances (×100)
no. of crossings, relative to best (∆%)
20
BM
BM+
C+
Nı̈
10
0
60
5
80
100
0
# vertices
(e) Rome, simple planarization,
C = 102[61, 173]
15
0
1
# vertices
20
20
(d) Rome, success-rate of ILP
5
0
100
40
# vertices
(c) Rome, running time,
BM ≈ 0, C ≈ 0
4
Regular
2
3
4
.00 .02
2
.03 4.60 236
.00 .01
1
.04 5.34 217
.04 4.85 252
# instances
†
relative to best
runtime [s]
BaAl
Regular
SteinLib
BaAl
2 3 4 2 3 4
B-E
I*
S
V
2
3
4
.72 .64 .66 .82 .84 .95
.06
.07 .01
.11 .00
.02
1
.85 .74 .73 .89 .89 .97 47.34 8.85 2.70 90.97 .06 10.57 143
.60 .68 .75 .74 .88 .95
.04
.04 .01
.07 .00
.01
0
1.0 1.0 1.0 1.0 1.0 1.0 9.19 16.22 3.13 16.35 .06 12.32 152
.96 .94 .92 .92 .91 .97 49.38 28.59 2.43 95.08 .05 7.92 239
# instances (×100)
density,
SteinLib
B-E I* S† V†
BM .86 .85 .82 .84
BM+ .90 .90 .88 .86
C
.84 .67 .81 .87
C+ 1.0 1.0 1.0 1.0
Nı̈ .92 .98 .91 .91
(f ) Rome, state-of-the-art planarization
BM
BM+
C
C+
Nı̈
10
7.5
5
15
10
2.5
0
20
5
40
60
80
100
# instances (×100)
6
0
# vertices
Fig. 1. We may omit algorithms whose values are unsuitable for a plot; instead we
give their average[min, max] in the caption.
A Note on the Practicality of Maximal Planar Subgraph Algorithms
7
References
1. BarabasiAlbertGenerator of the java universal network/graph framework,
jung.sourceforge.net
2. Batini, C., Talamo, M., Tamassia, R.: Computer aided layout of entity relationship
diagrams. Journal of Systems and Software 4(2-3), 163–173 (1984)
3. Boyer, J.M., Myrvold, W.J.: On the cutting edge: Simplified O(n) planarity by
edge addition. J. Graph Algorithms Appl. 8(2), 241–273 (2004)
4. Călinescu, G., Fernandes, C., Finkler, U., Karloff, H.: A better approximation
algorithm for finding planar subgraphs. J. Algorithms 27, 269–302 (1998)
5. Chimani, M., Gutwenger, C., Jünger, M., Klau, G.W., Klein, K., Mutzel, P.:
The open graph drawing framework (OGDF). In: Tamassia, R. (ed.) Handbook
of Graph Drawing and Visualization, chap. 17, pp. 543–569. CRC Press (2014),
www.ogdf.net
6. Chimani, M., Gutwenger, C.: Non-planar core reduction of graphs. Discrete Mathematics 309(7), 1838–1855 (2009)
7. Chimani, M., Gutwenger, C.: Advances in the planarization method: Effective multiple edge insertions. J. Graph Algorithms Appl. 16(3), 729–757 (2012)
8. Chimani, M., Hedtke, I., Wiedera, T.: Limits of greedy approximation algorithms
for the maximum planar subgraph problem. In: Proc. Int’l Workshop on Combinatorial Algorithms (IWOCA) 2016. LNCS (2016), to appear
9. Chimani, M., Hlinený, P.: Inserting multiple edges into a planar graph. CoRR
abs/1509.07952 (2015), http://arxiv.org/abs/1509.07952
10. Chimani, M., Mutzel, P., Schmidt, J.M.: Efficient extraction of multiple kuratowski
subdivisions. In: Graph Drawing, 15th Int’l Symposium, GD 2007, Sydney, Australia, September 24-26, 2007. Revised Papers. pp. 159–170 (2007)
11. Di Battista, G., Garg, A., Liotta, G., Tamassia, R., Tassinari, E., Vargiu, F.: An
experimental comparison of four graph drawing algorithms. Computational Geometry 7(56), 303–325 (1997)
12. Di Battista, G., Garg, A., Liotta, G., Parise, A., Tassinari, R., Tassinari, E., Vargiu,
F., Vismara, L.: Drawing directed acyclic graphs: An experimental study. Int’l J.
of Computational Geometry & Applications 10(06), 623–648 (2000)
13. Hochstein, J.M., Weihe, K.: Maximum s-t-flow with k crossings in O(k3 n log n).
In: Proc. SODA 2007. pp. 843–847. ACM-SIAM (2007)
14. Hsu, W.: A linear time algorithm for finding a maximal planar subgraph based on
PC-trees. LNCS, vol. 3595, pp. 787–797. Springer (2005)
15. Jünger, M., Mutzel, P.: Maximum planar subgraphs and nice embeddings: Practical
layout tools. Algorithmica 16(1), 33–59 (1996)
16. Koch, T., Martin, A., Voß, S.: SteinLib: An updated library on steiner tree problems in graphs. Tech. Rep. ZIB-Report 00-37, Konrad-Zuse-Zentrum für Informationstechnik Berlin (2000), http://elib.zib.de/steinlib
17. Kuratowski, C.: Sur le problme des courbes gauches en topologie. Fundamenta
Mathematicae 15(1), 271–283 (1930)
18. Liu, P.C., Geldmacher, R.C.: On the deletion of nonplanar edges of a graph. In:
Proc. 10th Southeastern Conf. on Combinat., Graph Theory and Comput. pp.
727–738. Congress. Numer., XXIII–XXIV, Utilitas Math., Winnipeg, Man. (1979)
19. Shih, W., Hsu, W.: A new planarity test. Theor. Comput. Sci. 223(1-2), 179–191
(1999)
20. Steger, A., Wormald, N.C.: Generating random regular graphs quickly. Combinatorics, Probability & Computing 8(4), 377–396 (1999)
| 8cs.DS
|
arXiv:1801.08350v1 [cs.LO] 25 Jan 2018
THEORY OF HIGHER ORDER INTERPRETATIONS AND APPLICATION TO BASIC
FEASIBLE FUNCTIONS
EMMANUEL HAINRY AND ROMAIN PÉCHOUX
Inria project Mocqua, CNRS, and Université de Lorraine, LORIA
e-mail address: [email protected]
Inria project Mocqua, CNRS, and Université de Lorraine, LORIA
e-mail address: [email protected]
A BSTRACT. Interpretation methods and their restrictions to polynomials have been deeply used to
control the termination and complexity of first-order term rewrite systems. This paper extends interpretation methods to a pure higher order functional language. We develop a theory of higher order
functions that is well-suited for the complexity analysis of this programming language. The interpretation domain is a complete lattice and, consequently, we express program interpretation in terms
of a least fixpoint. As an application, by bounding interpretations by higher order polynomials, we
characterize Basic Feasible Functions at any order.
1. INTRODUCTION
1.1. Higher order interpretations. This paper introduces a theory of higher order interpretations.
These interpretations are an extension of usual (polynomial) interpretation methods introduced in [MN70,
Lan79] and used to show the termination of (first order) term rewrite systems [CMPU05, CL92] or
to study their complexity [BMM11].
This theory is a novel and uniform extension to higher order functional programs: the definition works at any order on a simple programming language, where interpretations can be elegantly
expressed in terms of a least fixpoint, and no extra constraints are required.
The language has only one semantics restriction: its reduction strategy is enforced to be leftmost outermost as interpretations are non decreasing functions. Similarly to first order interpretations, higher order interpretations ensure that each reduction step corresponds to a strict decrease.
Consequently, some of the system properties could be lost if a reduction occurs under a context.
This work has been partially supported by ANR Project ELICA ANR-14-CE25-0005.
LOGICAL METHODS
IN COMPUTER SCIENCE
DOI:10.2168/LMCS-???
1
© E. Hainry and R. Péchoux
Creative Commons
2
E. HAINRY AND R. PÉCHOUX
1.2. Application to higher order polynomial time. As Church-Turing’s thesis does not hold at
higher order, distinct and mostly pairwise uncomparable complexity classes are candidates as a natural equivalent of the notion of polynomial time computation for higher order.
The class of polynomially computable real functions by Ko [Ko91] and the class of Basic Feasible Functional at order 𝑖 (BFF𝑖 ) by Irwin, Kapron and Royer [IKR02] belong to the most popular
definitions for such classes. In [Ko91] polynomially computable real functions are defined in terms
of first order functions over real numbers. They consist in order 2 functions over natural numbers
and an extension at any order is proposed by Kawamura and Cook in [KC10]. The main distinctions
between these two models are the following:
∙ the paper [Ko91] deals with representation of real numbers as input while the paper [IKR02]
deals with general functions as input,
∙ the paper [Ko91] deals with the number of steps needed to produce an output at a given precision while the paper [IKR02] deals with the number of reduction steps needed to evaluate
the program.
Moreover, it was shown in [IKR02] and [Fér14] that the classes BFF𝑖 cannot capture some functions that could be naturally considered to be polynomial time computable because they do not take
into account the size of their higher order arguments. However they have been demonstrated to be
robust, they characterize exactly the well-known class FPTIME and BFF at order 1 and order 2, and
have already been characterized in various ways, e.g. [CK89].
The current paper provides a characterization of the BFF𝑖 classes as they deal with discrete
data as input and they are consequently more suited to be studied with respect to usual functional
languages. This result was expected to hold as it is known for a long time that (first order) polynomial
interpretations characterize polynomial time (that is BFF1 ) and as it is shown in [FHHP15] that (first
order) polynomial interpretations on stream programs characterize BFF2 .
1.3. Related works. The present paper is an extended version of the results in [HP17]: more proofs
and examples have been provided. An erratum has been provided: the interpretation of the case
construct has been slightly modified so that we can consider non decreasing functions (and not only
strictly increasing functions).
There are two lines of work that are related to our approach. In [VdP93], Van De Pol introduced higher order interpretation for showing the termination of higher order term rewrite systems.
In [BL12, BL16], Baillot and Dal Lago introduce a higher order interpretations for complexity analysis of term rewrite systems. While the first work only deals with termination properties, the second
work is restricted to a family of higher order term rewrite systems called simply typed term rewrite
systems.
1.4. Outline. In Section 2, the syntax and semantics of the functional language are introduced. The
new notion of higher order interpretation and its properties are described in Section 3. Next, in Section 4, we briefly recall the BFF𝑖 classes and their main characterizations, including a characterization
based on the BTLP programming language of [IKR02]. Section 5 is devoted to the characterization
of these classes using higher order polynomials. The soundness relies on interpretation properties: the reduction length is bounded by the interpretation of the initial term. The completeness is
demonstrated by simulating a BTLP procedure: compiling procedures to terms after applying some
program transformations. In Section 6, we briefly discuss the open issues and future works related
to higher order interpretations.
INSTRUCTIONS
2. FUNCTIONAL
3
LANGUAGE
2.1. Syntax. The considered programming language consists in a unpure lambda calculus with constructors, primitive operators, a 𝚌𝚊𝚜𝚎 construct for pattern matching and a 𝚕𝚎𝚝𝚁𝚎𝚌 instruction for
function definitions that can be recursive. It is as an extension of PCF to inductive data types.
The set of terms of the language is generated by the following grammar:
𝙼, 𝙽 ∶∶= 𝚡 | 𝚌 | 𝚘𝚙 | 𝚌𝚊𝚜𝚎 𝙼 𝚘𝚏 𝚌1 (⃖⃖⃖
𝚡⃗1 ) → 𝙼1 , ..., 𝚌𝑛(⃖⃖⃖
𝚡⃗𝑛 ) → 𝙼𝑛 | 𝙼 𝙽 | 𝜆𝚡.𝙼 | 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼,
where 𝚌, 𝚌1 , ⋯ , 𝚌𝑛 are constructor symbols of fixed arity and 𝚘𝚙 is an operator of fixed arity.
Given a constructor or operator symbol 𝑏, we write 𝑎𝑟(𝑏) = 𝑛 whenever 𝑏 is of arity 𝑛. 𝚡, 𝚏 are
variables in and 𝚡⃖⃖⃗𝑖 is a sequence of 𝑎𝑟(𝚌𝑖 ) variables.
The free variables 𝐹 𝑉 (𝙼) of a term 𝙼 are defined as usual. Bounded variables are assumed to
have distinct names in order to avoid name clashes. A closed term is a term 𝙼 with no free variables,
𝐹 𝑉 (𝙼) = ∅.
A substitution {𝙽1 ∕𝚡1 , ⋯ , 𝙽𝑛 ∕𝚡𝑛 } is a partial function mapping variables 𝚡1 , ⋯ , 𝚡𝑛 to terms 𝙽1 , ⋯ , 𝙽𝑛 .
The result of applying the substitution {𝙽1 ∕𝚡1 , ⋯ , 𝙽𝑛 ∕𝚡𝑛 } to a term 𝙼 is noted 𝙼{𝙽1 ∕𝚡1 , ⋯ , 𝙽𝑛 ∕𝚡𝑛 }
⃖⃗ 𝚡⃗} when the substituting terms are clear from the context.
or 𝙼{𝙽∕⃖
2.2. Semantics. Each primitive operator 𝚘𝚙 of arity 𝑛 has a corresponding semantics J𝚘𝚙K fixed by
the language implementation, a total function from 𝑛 to .
We define the following relations between two terms of the language:
∙ 𝛽-reduction: 𝜆𝚡.𝙼 𝙽 →𝛽 𝙼{𝙽∕𝚡},
⃖⃖⃖⃗𝑗 ) of … 𝚌𝑗 (⃖⃖⃖
⃖⃖⃖⃗𝑗 ∕⃖⃖⃖
∙ pattern matching: case 𝚌𝑗 (𝙽
𝚡⃗𝑗 ) → 𝙼𝑗 … →𝚌𝚊𝚜𝚎 𝙼𝑗 {𝙽
𝚡⃗𝑗 },
∙ operator evaluation: 𝚘𝚙 𝙼1 … 𝙼𝑛 →𝚘𝚙 J𝚘𝚙K(𝙼1 , … , 𝑀𝑛 ),
∙ fixpoint evaluation: 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼 →𝚕𝚎𝚝𝚁𝚎𝚌 𝙼{𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼∕𝚏},
We extend these relations by: if 𝙼 →𝛼 𝙽 then case 𝙼 of … →𝛼 case 𝙽 of … , with 𝛼 ∈
{𝛽, 𝚌𝚊𝚜𝚎, 𝚕𝚎𝚝𝚁𝚎𝚌, 𝚘𝚙}.
Let ⇒ be the left-most outermost evaluation strategy with respect to, →{𝛽,𝚌𝚊𝚜𝚎,𝚕𝚎𝚝𝚁𝚎𝚌,𝚘𝚙} . Let ⇒𝑘
be the k-fold self composition of the relation ⇒. Moreover, let |𝙼 ⇒𝑘 𝙽| be the number of reductions
distinct from →𝚘𝚙 in a given a derivation 𝙼 ⇒𝑘 𝙽. |𝙼 ⇒𝑘 𝙽| ≤ 𝑘 always holds. J𝙼K is a notation for
the term computed by 𝙼 (if it exists), i.e. ∃𝑘, 𝙼 ⇒𝑘 J𝙼K and ∄𝙽, J𝙼K ⇒ 𝙽.
A (first order) value 𝚟 is defined inductively by either 𝚟 = 𝚌, if 𝑎𝑟(𝚌) = 0, or 𝚟 = 𝚌 𝚟⃖⃗, for
𝑎𝑟(𝚌) > 0 values 𝚟⃖⃗, otherwise.
2.3. Type system. Let B be a set of basic inductive types b described by their constructor symbol
set b .
The set of simple types is defined by:
𝚃 ∶∶= b | 𝚃 ⟶ 𝚃,
with b ∈ B.
As usual ⟶ associates to the right.
Example 2.1. The type of unary numbers 𝙽𝚊𝚝 can be described by 𝙽𝚊𝚝 = {0, +𝟷}, 0 being a
constructor symbol of 0-arity and +𝟷 being a constructor symbol of 1-arity.
For any type 𝑇 , [𝑇 ] is the base type for lists of elements of type 𝚃 and has constructor symbol
set [𝚃] = {𝚗𝚒𝚕, 𝚌}, 𝚗𝚒𝚕 being a constructor symbol of 0-arity and 𝚌 being a constructor symbol of
2-arity.
4
E. HAINRY AND R. PÉCHOUX
Γ(𝚡) = 𝚃
(Var)
Γ; Δ ⊢ 𝚡 ∶∶ 𝚃
Δ(𝚌) = 𝚃
(Cons)
Γ; Δ ⊢ 𝚌 ∶∶ 𝚃
Δ(𝚘𝚙) = 𝚃
(Op)
Γ; Δ ⊢ 𝚘𝚙 ∶∶ 𝚃
Γ; Δ ⊢ 𝙼 ∶∶ 𝚃1 ⟶ 𝚃2 Γ; Δ ⊢ 𝙽 ∶∶ 𝚃1
(App)
Γ; Δ ⊢ 𝙼 𝙽 ∶∶ 𝚃2
Γ, 𝚡 ∶∶ 𝚃1 ; Δ ⊢ 𝙼 ∶∶ 𝚃2
(Abs)
Γ; Δ ⊢ 𝜆𝚡.𝙼 ∶∶ 𝚃1 → 𝚃2
Γ, 𝚏 ∶∶ 𝚃; Δ ⊢ 𝙼 ∶∶ 𝚃
(Let)
Γ; Δ ⊢ 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼 ∶∶ 𝚃
Γ; Δ ⊢ 𝙼 ∶∶ b
Γ; Δ ⊢ 𝚌𝑖 ∶∶ b⃖⃖⃗𝑖 ⟶ b Γ, 𝚡⃖⃖⃗𝑖 ∶∶ b⃖⃖⃗𝑖 ; Δ ⊢ 𝙼𝑖 ∶∶ 𝚃 (1 ≤ 𝑖 ≤ 𝑚)
(Case)
Γ; Δ ⊢ 𝚌𝚊𝚜𝚎 𝙼 𝚘𝚏 𝚌1 (⃖⃖⃖
𝚡⃗1 ) → 𝙼1 , ..., 𝚌𝑛 (⃖⃖⃖
𝚡⃗𝑛 ) → 𝙼𝑛 ∶∶ 𝚃
Figure 1: Type system
The type system is described in Figure 1 and proves judgments of the shape Γ; Δ ⊢ 𝙼 ∶∶ 𝚃
meaning that the term 𝙼 has type 𝚃 under the variable and constructor symbol contexts Γ and Δ
respectively ; a variable (a constructor, respectively) context being a partial function that assigns
types to variables (constructors and operators, respectively).
As usual, the input type and output type of constructors and operators of arity 𝑛 will be restricted
to basic types. Consequently, their types are of the shape b1 ⟶ … ⟶ b𝑛 ⟶ b. A well-typed
term will consist in a term 𝙼 such that ∅; Δ ⊢ 𝙼 ∶∶ 𝚃 (Consequently, it is mandatory for a term to be
closed in order to be well-typed).
In what follows, we will consider only well-typed terms. The type system assigns types to all
the syntactic constructions of the language and ensures that a program does not go wrong. Notice
that the typing discipline does not prevent a program from diverging.
Definition 2.2 (Order). Given a term 𝙼 of type 𝚃, i.e. ∅; Δ ⊢ 𝙼 ∶∶ 𝚃, the order of 𝙼, noted 𝚘𝚛𝚍(𝙼),
is equal to the order of 𝚃, noted 𝚘𝚛𝚍(𝚃) and defined inductively by:
if b ∈ B,
𝚘𝚛𝚍(b) = 0,
′
′
𝚘𝚛𝚍(𝚃 ⟶ 𝚃 ) = max(𝚘𝚛𝚍(𝚃) + 1, 𝚘𝚛𝚍(𝚃 ))
otherwise.
Example 2.3. Consider the following term 𝙼 that maps a function to a list given as inputs:
𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝜆g.𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌 (g 𝚢) (𝚏 g 𝚣),
𝚗𝚒𝚕 → 𝚗𝚒𝚕
Let [𝙽𝚊𝚝] is the base type for lists of natural numbers of constructor symbol set [𝙽𝚊𝚝] = {𝚗𝚒𝚕, 𝚌}.
The term 𝙼 can be typed by ∅; Δ ⊢ 𝙼 ∶∶ (𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝) ⟶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝], , as illustrated by the
following typing derivation:
INSTRUCTIONS
5
⋯
(App)
⋯
Γ′ ; Δ ⊢ (g 𝚢) ∶∶ 𝙽𝚊𝚝
⋯
(App)
(App)
′
′
⋯
Γ ; Δ ⊢ 𝚌 (g 𝚢) ∶∶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝]
Γ ; Δ ⊢ (𝚏 g 𝚣) ∶∶ [𝙽𝚊𝚝]
(App)
⋯
Γ, 𝚢 ∶∶ 𝙽𝚊𝚝, 𝚣 ∶∶ [𝙽𝚊𝚝]; Δ ⊢ 𝚌 (g 𝚢) (𝚏 g 𝚣) ∶∶ [𝙽𝚊𝚝]
(Case)
Γ, 𝚡 ∶∶ [𝙽𝚊𝚝]; Δ ⊢ 𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌 (g 𝚢) (𝚏 g 𝚣), 𝚗𝚒𝚕 → 𝚗𝚒𝚕 ∶∶ [𝙽𝚊𝚝]
(Abs)
Γ; Δ ⊢ 𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌 (g 𝚢) (𝚏 g 𝚣), 𝚗𝚒𝚕 → 𝚗𝚒𝚕 ∶∶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝]
(Abs)
𝚏 ∶∶ 𝚃; Δ ⊢ 𝜆g.𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌 (g 𝚢) (𝚏 g 𝚣), 𝚗𝚒𝚕 → 𝚗𝚒𝚕 ∶∶ 𝚃
(Let)
∅; Δ ⊢ 𝙼 ∶∶ (𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝) ⟶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝]
where with 𝚃 is a shorthand notation for (𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝) ⟶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝], where the derivation
of the base case 𝚗𝚒𝚕 has been omitted for readability and where the contexts Δ, Γ, Γ′ are such that
Δ(𝚌) = 𝙽𝚊𝚝 ⟶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝], Δ(𝚗𝚒𝚕) = [𝙽𝚊𝚝] and Γ = 𝚏 ∶∶ 𝚃, g ∶∶ 𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝 and
Γ′ = Γ, 𝚢 ∶∶ 𝙽𝚊𝚝, 𝚣 ∶∶ [𝙽𝚊𝚝]. Consequently, the order of 𝙼 is equal to 2, i.e. 𝚘𝚛𝚍(𝙼) = 2.
3. INTERPRETATIONS
3.1. Interpretations of types. We briefly recall some basic definitions that are very close from the
notions used in denotational semantics (See [Win93]) since, as we shall see later, interpretations can
be defined in terms of fixpoints. Let (ℕ, ≤, ⊔, ⊓) be the set of natural numbers equipped with the
usual ordering ≤, a max operator ⊔ and min operator ⊓ and let ℕ be ℕ ∪ {⊤}, where ⊤ is the greatest
element satisfying ∀𝑛 ∈ ℕ, 𝑛 ≤ ⊤, 𝑛 ⊔ ⊤ = ⊤ ⊔ 𝑛 = ⊤ and 𝑛 ⊓ ⊤ = ⊤ ⊓ 𝑛 = 𝑛.
The interpretation of a type is defined inductively by:
if b is a basic type,
⦇b⦈ = ℕ,
′
↑
′
⦇𝚃 ⟶ 𝚃 ⦈ = ⦇𝚃⦈ ⟶ ⦇𝚃 ⦈,
otherwise,
where ⦇𝚃⦈ ⟶↑ ⦇𝚃′ ⦈ denotes the set of total non decreasing functions from ⦇𝚃⦈ to ⦇𝚃′ ⦈. A function
𝐹 from the set 𝐴 to the set 𝐵 being non decreasing if for each 𝑋, 𝑌 ∈ 𝐴, 𝑋 ≤𝐴 𝑌 implies 𝐹 (𝑋) ≤𝐵
𝐹 (𝑌 ), where ≤𝐴 is the usual pointwise ordering induced by ≤ and defined by:
𝑛 ≤ℕ 𝑚 iff 𝑛 ≤ 𝑚
𝐹 ≤𝐴⟶↑ 𝐵 𝐺 iff ∀𝑋 ∈ 𝐴, 𝐹 (𝑋) ≤𝐵 𝐺(𝑋)
Example 3.1. The type 𝚃 = (𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝) ⟶ [𝙽𝚊𝚝] ⟶ [𝙽𝚊𝚝] of the term 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼 in
Example 2.3 is interpreted by:
⦇𝚃⦈ = (ℕ ⟶↑ ℕ) ⟶↑ (ℕ ⟶↑ ℕ).
⃖⃖⃗ of 𝑚 terms in the interpretation domain and a sequence 𝚃⃖⃗
In what follows, given a sequence 𝐹
⃖⃖⃖⃖⃖
⃗
⃖⃖
⃗
of 𝑘 types, the notation 𝐹 ∈ ⦇𝚃⦈ means that both 𝑘 = 𝑚 and ∀𝑖 ∈ [1, 𝑚], 𝐹𝑖 ∈ ⦇𝚃𝑖 ⦈.
6
E. HAINRY AND R. PÉCHOUX
3.2. Interpretations of terms. Each closed term of type 𝚃 will be interpreted by a function in ⦇𝚃⦈.
The application is denoted as usual whereas we use the notation Λ for abstraction on this function
space in order to avoid confusion between terms of our calculus and objects of the interpretation
domain. Variables of the interpretation domain will be denoted using upper case letters. When
needed, Church typing discipline will be used in order to highlight the type of the bound variable in
a lambda abstraction.
An important distinction between the terms of the language and the objects of the interpretation
domain lies in the fact that beta-reduction is considered as an equivalence relation on (closed terms
of) the interpretation domain, i.e. Λ𝑋.𝐹 𝐺 = 𝐹 {𝐺∕𝑋} underlying that Λ𝑋.𝐹 𝐺 and 𝐹 {𝐺∕𝑋}
are distinct notations that represent the same higher order function. The same property holds for
𝜂-reduction, i.e. Λ𝑋.𝐹 𝑋 and 𝐹 denote the same function.
In order to obtain complete lattices, each type ⦇𝚃⦈ has to be completed by a lower bound ⊥⦇𝚃⦈
and an upper bound ⊤⦇𝚃⦈ as follows:
⊥ℕ = 0
⊤ℕ = ⊤
⊥⦇𝚃⟶𝚃′ ⦈ = Λ𝑋 ⦇𝚃⦈ .⊥⦇𝚃′ ⦈
⊤⦇𝚃⟶𝚃′ ⦈ = Λ𝑋 ⦇𝚃⦈ .⊤⦇𝚃′ ⦈
Lemma 3.2. For each 𝚃 and for each 𝐹 ∈ ⦇𝚃⦈, ⊥⦇𝚃⦈ ≤⦇𝚃⦈ 𝐹 ≤⦇𝚃⦈ ⊤⦇𝚃⦈ .
Proof. By induction on types.
Notice that for each type 𝚃 it also holds that ⊤⦇𝚃⦈ ≤⦇𝚃⦈ ⊤⦇𝚃⦈ , by an easy induction.
In the same spirit, max and min operators ⊔ (and ⊓) over ℕ can be extended to higher order
functions 𝐹 , 𝐺 of any arbitrary type ⦇𝚃⦈ ⟶↑ ⦇𝚃′ ⦈ by:
⊔⦇𝚃⦈⟶
↑ ⦇𝚃′ ⦈
(𝐹 , 𝐺) = Λ𝑋 ⦇𝚃⦈ . ⊔⦇𝚃 ⦈ (𝐹 (𝑋), 𝐺(𝑋))
′
⊓⦇𝚃⦈⟶
↑ ⦇𝚃′ ⦈
(𝐹 , 𝐺) = Λ𝑋 ⦇𝚃⦈ . ⊓⦇𝚃 ⦈ (𝐹 (𝑋), 𝐺(𝑋))
′
In the following, we use the notations ⊥, ⊤, ≤, <, ⊔ and ⊓ instead of ⊥⦇𝚃⦈ , ⊤⦇𝚃⦈ , ≤⦇𝚃⦈ , <⦇𝚃⦈ , ⊔⦇𝚃⦈ and
⊓⦇𝚃⦈ , respectively, when ⦇𝚃⦈ is clear from the typing context. Moreover, given a boolean predicate
𝑃 on functions, we will use the notation ⊔𝑃 {𝐹 } as a shorthand notation for ⊔{𝐹 | 𝑃 }.
Lemma 3.3. For each type 𝚃, (⦇𝚃⦈, ≤, ⊔, ⊓, ⊤, ⊥) is a complete lattice.
Proof. Consider a subset 𝑆 of elements in ⦇𝚃⦈ and define ⊔𝑆 = ⊔𝐹 ∈𝑆 𝐹 . By definition, we have
𝐹 ≤ ⊔𝑆, for any 𝐹 ∈ 𝑆. Now consider some 𝐺 such that for all 𝐹 ∈ 𝑆, 𝐹 ≤ 𝐺. We have ∀𝐹 ∈ 𝑆,
∀𝑋, 𝐹 (𝑋) ≤ 𝐺(𝑋). Consequently, ∀𝑋, 𝑆(𝑋) = ⊔𝐹 ∈𝑆 𝐹 (𝑋) ≤ 𝐺(𝑋) and 𝑆 is a supremum. The
same holds for the infimum.
Now we need to define a unit (or constant) cost function for any interpretation of type 𝚃 in order
to take the cost of recursive calls into account. For that purpose, let + denote natural number addition
extended to ℕ by ∀𝑛, ⊤ + 𝑛 = 𝑛 + ⊤ = ⊤. For each type ⦇𝚃⦈, we define inductively a dyadic sum
function ⊕⦇𝚃⦈ by:
𝑋 ℕ ⊕ℕ 𝑌 ℕ = 𝑋 + 𝑌
𝐹 ⊕⦇𝚃⟶𝚃′ ⦈ 𝐺 = Λ𝑋 ⦇𝚃⦈ .(𝐹 (𝑋) ⊕⦇𝚃′ ⦈ 𝐺(𝑋))
INSTRUCTIONS
7
Let us also define the constant function 𝑛⦇𝚃⦈ , for each type 𝚃 and each integer 𝑛 ≥ 1, by:
𝑛ℕ = 𝑛
𝑛⦇𝚃⟶𝚃′ ⦈ = Λ𝑋 ⦇𝚃⦈ .𝑛⦇𝚃′ ⦈
Once again, we will omit the type when it is unambiguous using the notation 𝑛⊕ to denote the
function 𝑛⦇𝚃⦈ ⊕⦇𝚃⦈ when ⦇𝚃⦈ is clear from the typing context.
For each type ⦇𝚃⦈, we can define a strict ordering < by: 𝐹 < 𝐺 whenever 1 ⊕ 𝐹 ≤ 𝐺.
Definition 3.4. A variable assignment, denoted 𝜌, is a map associating to each 𝚏 ∈ of type 𝚃 a
variable 𝐹 of type ⦇𝚃⦈.
Now we are ready to define the notions of variable assignment and interpretation of a term 𝙼:
Definition 3.5 (Interpretation). Given a variable assignment 𝜌, an interpretation is the extension of
𝜌 to well-typed terms, mapping each term of type 𝚃 to an object in ⦇𝚃⦈ and defined by:
∙ ⦇𝚏⦈𝜌 = 𝜌(𝚏), if 𝚏 ∈ ,
∙ ⦇𝚌⦈𝜌 = 1 ⊕ (Λ𝑋1 . … .Λ𝑋𝑛 .
∙ ⦇𝙼𝙽⦈𝜌 = ⦇𝙼⦈𝜌 ⦇𝙽⦈𝜌 ,
∑𝑛
𝑖=1 𝑋𝑖 ),
if 𝑎𝑟(𝚌) = 𝑛,
∙ ⦇𝜆𝚡.𝙼⦈𝜌 = 1 ⊕ (Λ⦇𝚡⦈𝜌 .⦇𝙼⦈𝜌 ),
∙ ⦇case 𝙼 of … 𝚌𝑗 (⃖⃖⃖
𝚡⃗𝑗 ) → 𝙼𝑗 …⦈𝜌 = 1 ⊕ ⊔1≤𝑖≤𝑚 ⊔⦇𝙼⦈
⃖⃖⃖⃗ ⃖⃖⃖⃖⃖⃖⃖⃖⃖⃗
⃖⃖⃗ {⦇𝙼⦈𝜌 ⊕ ⦇𝙼𝑖 ⦈𝜌 {𝐹𝑖 ∕⦇𝚡𝑖 ⦈𝜌 }},
𝜌 ≥⦇𝚌𝑖 ⦈𝜌 𝐹𝑖
∙ ⦇𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼⦈𝜌 = ⊓{𝐹 ∈ ⦇𝚃⦈ | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙼⦈𝜌 𝐹 }.
where ⦇𝚘𝚙⦈𝜌 is a sup-interpretation, i.e. a non decreasing total function such that:
∀𝙼1 , … , ∀𝙼𝑛 , ⦇𝚘𝚙 𝙼1 … 𝙼𝑛 ⦈𝜌 ≥ ⦇J𝚘𝚙K(𝙼1 , … , 𝙼𝑛 )⦈𝜌 .
See [Péc13] for more details about sup-interpretations.
As operator sup-interpretations are fixed, an interpretations should also be indexed by some
mapping 𝑚 assigning a sup-interpretation to each operator of the language. To simplify the formalism, we will omit this mapping in what follows.
3.3. Existence of an interpretation. The interpretation of a term is always defined. Indeed, in Definition 3.5, ⦇𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼⦈𝜌 is defined in terms of the least fixpoint of the function Λ𝑋 ⦇𝚃⦈ .1 ⊕⦇𝚃⦈
(Λ⦇𝚏⦈𝜌 .⦇𝙼⦈𝜌 𝑋) and, consequently, we obtain the following result as a direct consequence of KnasterTarski [Tar55, KS01] Fixpoint Theorem:
Proposition 3.6. Each term 𝙼 of type 𝚃 has an interpretation.
Proof. By Lemma 3.3, 𝐿 = (⦇𝚃⦈, ≤, ⊔, ⊓, ⊤, ⊥) is a complete lattice. The function 𝐹 = Λ𝑋 ⦇𝚃⦈ .1⊕⦇𝚃⦈
(Λ⦇𝚏⦈𝜌 .⦇𝙼⦈𝜌 𝑋) ∶ 𝐿 → 𝐿 is monotonic. Indeed, both constructor terms and letrec terms of type
⦇𝚃⦈ are interpreted over a space of monotonic functions ⦇𝚃⦈. Moreover monotonicity is preserved
by application, abstraction and the ⊓ and ⊔ operators. Applying Knaster-Tarski, we obtain that 𝐹
admits a least fixpoint, which is exactly ⊓{𝑋 ∈ ⦇𝚃⦈ | 𝑋 ≥ 𝐹 𝑋}.
8
E. HAINRY AND R. PÉCHOUX
3.4. Properties of interpretations. We now show intermediate lemmata. The following Lemma
can be shown by structural induction on terms:
Lemma 3.7. For all 𝙼, 𝙽, 𝚡 such that 𝚡 ∶∶ 𝚃; Δ ⊢ 𝙼 ∶∶ 𝚃′ , ∅; Δ ⊢ 𝙽 ∶∶ 𝚃, we have
⦇𝙼⦈𝜌 {⦇𝙽⦈𝜌 ∕⦇𝚡⦈𝜌 } = ⦇𝙼{𝙽∕𝚡}⦈𝜌 .
Lemma 3.8. For all 𝙼, 𝙽, 𝚡 such that 𝚡 ∶∶ 𝚃; Δ ⊢ 𝙼 ∶∶ 𝚃′ , ∅; Δ ⊢ 𝙽 ∶∶ 𝚃, we have ⦇𝜆𝚡.𝙼 𝙽⦈𝜌 >
⦇𝙼{𝙽∕𝚡}⦈𝜌.
Proof.
⦇𝜆𝚡.𝙼 𝙽⦈𝜌 = ⦇𝜆𝚡.𝙼⦈𝜌 ⦇𝙽⦈𝜌
= Λ⦇𝚡⦈𝜌 .1 ⊕ ⦇𝙼⦈𝜌 ⦇𝙽⦈𝜌
= 1 ⊕ ⦇𝙼⦈𝜌 {⦇𝙽⦈𝜌 ∕⦇𝚡⦈𝜌 }
= 1 ⊕ ⦇𝙼{𝙽∕𝚡}⦈𝜌
> ⦇𝙼{𝙽∕𝚡}⦈𝜌
By Definition 3.5
By Definition 3.5
By definition of =
By Lemma 3.7
By definition of >
and so the conclusion.
Lemma 3.9. For all 𝙼, we have: if 𝙼 ⇒ 𝙽 then ⦇𝙼⦈𝜌 ≥ ⦇𝙽⦈𝜌 . Moreover if |𝙼 ⇒ 𝙽| = 1 then
⦇𝙼⦈𝜌 > ⦇𝙽⦈𝜌 .
Proof. If |𝙼 ⇒ 𝙽| = 0 then 𝙼 = 𝚘𝚙 𝙼1 … 𝙼𝑛 →𝚘𝚙 J𝚘𝚙K(𝙼1 , … , 𝙼𝑛 ) = 𝙽, for some operator 𝚘𝚙
and terms 𝙼1 , … , 𝙼𝑛 and consequently, by Definition of interpretations we have ⦇𝚘𝚙 𝙼1 … 𝙼𝑛 ⦈𝜌 ≥
⦇J𝚘𝚙K(𝙼1 , … , 𝙼𝑛 )⦈𝜌 .
If |𝙼 ⇒ 𝙽| = 1 then the reduction does not involve a →𝚘𝚙 (i.e. an operator evaluation). By
Lemma 3.8, in the case of a 𝛽-reduction and, by induction, by Lemma 3.7 and Definition 3.5 for the
other cases. e.g. For a letrec reduction, we have: if 𝙼 = 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼′ →𝚕𝚎𝚝𝚁𝚎𝚌 𝙼′ {𝙼∕𝚏} = 𝙽 then:
⦇𝙼⦈𝜌 = ⊓{𝐹 ∈ ⦇𝚃⦈ | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 𝐹 }
≥ ⊓{1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 𝐹 | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 𝐹 }
≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 ⊓ {𝐹 | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 𝐹 }
≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 ⦇𝙼⦈𝜌
≥ 1 ⊕ ⦇𝙽⦈𝜌 {⦇𝙼⦈𝜌 ∕⦇𝚏⦈𝜌 }
≥ 1 ⊕ ⦇𝙽{𝙼∕𝚏}⦈𝜌, by Lemma 3.7
> ⦇𝙽{𝙼∕𝚏}⦈𝜌, By definition of >
The first inequality holds since we are only considering higher order functions 𝐹 satisfying 𝐹 ≥
1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 𝐹 . The second inequality holds because Λ⦇𝚏⦈𝜌 .⦇𝙽⦈𝜌 is a non decreasing function
(as the interpretation domain only consists in such functions).
As each reduction distinct from an operator evaluation corresponds to a strict decrease, the
following corollary can be obtained:
⃖⃗ such that ∅; Δ ⊢ 𝙼 𝙽⃖⃗ ∶∶ 𝚃, if 𝙼 𝙽⃖⃗ ⇒𝑘 𝙼′ then ⦇𝙼⦈𝜌 ⦇𝙽⦈
⃖⃗ 𝜌 ≥
Corollary 3.10. For all terms, 𝙼, 𝙽,
|𝙼 𝙽⃖⃗ ⇒𝑘 𝙼′ | ⊕ ⦇𝙼′ ⦈𝜌 .
As basic operators can be considered as constant time computable objects the following Corollary also holds:
INSTRUCTIONS
9
⃖⃗ such that ∅; Δ ⊢ 𝙼 𝙽⃖⃗ ∶∶ b, with b ∈ B, if ⦇𝙼 𝙽⦈
⃖⃗ 𝜌 ≠ ⊤ then 𝙼 𝙽⃖⃗
Corollary 3.11. For all terms, 𝙼, 𝙽,
⃖⃗ 𝜌 ).
terminates in a number of reduction steps in 𝑂(⦇𝙼 𝙽⦈
The size |𝚟| of a value 𝚟 is defined by |𝚌| = 1 and |𝙼 𝙽| = |𝙼| + |𝙽|.
Lemma 3.12. For any value 𝚟, such that ∅; Δ ⊢ 𝚟 ∶∶ b, with b ∈ B, we have ⦇𝚟⦈ = |𝚟|.
Example 3.13. Consider the following term 𝙼 ∶∶ 𝙽𝚊𝚝 ⟶ 𝙽𝚊𝚝 computing the double of a unary
number given as input:
𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 +𝟷(𝚢) → +𝟷(+𝟷(𝚏 𝚢)),
0→0
We can see below how the interpretation rules of Definition 3.5 are applied on such a term.
⦇𝙼⦈𝜌 = ⊓{𝐹 | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 +𝟷(𝚢) → +𝟷(+𝟷(𝚏 𝚢))| 0 → 0⦈𝜌 𝐹 }
= ⊓{𝐹 | 𝐹 ≥ 2 ⊕ (Λ⦇𝚡⦈𝜌 .⦇𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 +𝟷(𝚢) → +𝟷(+𝟷(𝚏 𝚢))| 0 → 0⦈𝜌 {𝐹 ∕⦇𝚏⦈𝜌 })}
= ⊓{𝐹 | 𝐹 ≥ 3 ⊕ (Λ𝑋.𝑋 ⊕ ((⊔𝑋≥⦇+𝟷(𝚢)⦈𝜌 ⦇+𝟷(+𝟷(𝚏 𝚢))⦈𝜌 ) ⊔ (⊔𝑋≥⦇0⦈𝜌 ⦇0⦈𝜌 ){𝐹 ∕⦇𝚏⦈𝜌 })}
= ⊓{𝐹 | 𝐹 ≥ 3 ⊕ (Λ𝑋.𝑋 ⊕ ((⊔𝑋≥1⊕⦇𝚢⦈𝜌 2 ⊕ (⦇𝚏⦈𝜌 ⦇𝚢⦈𝜌 )) ⊔ (⊔𝑋≥1 1){𝐹 ∕⦇𝚏⦈𝜌 })}
= ⊓{𝐹 | 𝐹 ≥ 3 ⊕ (Λ𝑋.𝑋 ⊕ ((⊔𝑋≥1⊕⦇𝚢⦈𝜌 2 ⊕ (𝐹 ⦇𝚢⦈𝜌 )) ⊔ (1))}
= ⊓{𝐹 | 𝐹 ≥ 3 ⊕ (Λ𝑋.𝑋 ⊕ ((2 ⊕ (𝐹 (𝑋 − 1))) ⊔ (1))},
= ⊓{𝐹 | 𝐹 ≥ Λ𝑋.(5 ⊕ 𝑋 ⊕ (𝐹 (𝑋 − 1))) ⊔ (4 ⊕ 𝑋)}
𝑋 −1≥ 0
≤ Λ𝑋.6𝑋 2 + 5
At the end we search for the minimal non decreasing function 𝐹 greater than Λ𝑋.(5 ⊕ 𝑋 ⊕
(𝐹 (𝑋 − 1))) ⊔ (4 ⊕ 𝑋), for 𝑋 > 1. As the function Λ𝑋.6𝑋 2 ⊕ 5 is solution of such an inequality,
the fixpoint is smaller than this function. Notice that such an interpretation is not tight as one should
have expected the interpretation of such a program to be Λ𝑋.2𝑋. This interpretation underlies that:
∙ the iteration steps, distinct from the base case, count for 5 ⊕ 𝑋: 1 for the letRec call, 1 for
the application, 1 ⊕ 𝑋 for pattern matching and 2 for the extra-constructors added,
∙ the base case counts for 4 ⊕ 𝑋: 1 for recursive call, 1 for application, 1 ⊕ 𝑋 for pattern
matching and 1 for the constructor.
Consequently, we have a bound on both size of terms and reduction length though this upper bound
is not that accurate. This is not that surprising as this technique suffers from the same issues as
first-order interpretation based methods.
3.5. Relaxing interpretations. For a given program it is somewhat difficult to find an interpretation
that can be expressed in an easy way. This difficulty lies in the homogeneous definition of the
considered interpretations using a max (for the case construct) and a min (for the letrec construct).
Indeed, it is sometimes difficult to eliminate the constraint (parameters) of a max generated by the
interpretation of a case. Moreover, it is a hard task to find the fixpoint of the interpretation of a
letRec. All this can be relaxed as follows:
∙ finding an upper-bound of the max by eliminating constraint in the case construct interpretation,
∙ taking a particular function satisfying the inequality in the letrec construct interpretation.
In both cases, we will no longer compute an exact interpretation of the term but rather an upper
bound of the interpretation.
10
E. HAINRY AND R. PÉCHOUX
Lemma 3.14. Given a set of functions 𝑆 and a function 𝐹 ∈ 𝑆, the following inequality always
holds 𝐹 ≥ ⊓{𝐺|𝐺 ∈ 𝑆}.
This relaxation is highly desirable in order to find “lighter” upper-bounds on the interpretation
of a term. Moreover, it is a reasonable approximation as we are interested in worst case complexity.
Obviously, it is possible by relaxing too much to attain the interpretation ⊤⦇𝚃⦈ , which gives
nothing interesting. Of course these approximation have to be performed with moderation as taking
too big intermediate upper bounds might lead to an uninteresting upper bound on the interpretation
of the term (i.e. ⊤).
Example 3.15. Consider the term 𝙼 of Example 2.3:
𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝜆g.𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌 (g 𝚢) (𝚏 g 𝚣),
𝚗𝚒𝚕 → 𝚗𝚒𝚕
Its interpretation ⦇𝙼⦈𝜌 is equal to:
= ⊓{𝐹 | 𝐹 ≥ 1 ⊕ Λ⦇𝚏⦈𝜌 .⦇𝜆g.𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌(g 𝚢)(𝚏 g 𝚣)| 𝚗𝚒𝚕 → 𝚗𝚒𝚕⦈𝜌 𝐹 }
= ⊓{𝐹 | 𝐹 ≥ 3 ⊕ (Λ⦇𝚏⦈𝜌 .Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → 𝚌(g 𝚢)(𝚏 g 𝚣)| 𝚗𝚒𝚕 → 𝚗𝚒𝚕⦈𝜌 𝐹 )}
= ⊓{𝐹 | 𝐹 ≥ 4 ⊕ (Λ⦇𝚏⦈𝜌 .Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚡⦈𝜌 ⊕ (⊔(⦇𝚗𝚒𝚕⦈𝜌 , ⊔⦇𝚡⦈𝜌 ≥⦇𝚌(𝚢,𝚣)⦈𝜌 (⦇𝚌(g 𝚢)(𝚏 g 𝚣)⦈𝜌 ))𝐹 ))}
= ⊓{𝐹 | 𝐹 ≥ 4 ⊕ (Λ⦇𝚏⦈𝜌 .Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚡⦈𝜌 ⊕ (⊔(1, ⊔⦇𝚡⦈𝜌 ≥1⊕⦇𝚢⦈𝜌 +⦇𝚣)⦈𝜌 (1 ⊕ (⦇g⦈𝜌 ⦇𝚢⦈𝜌 )
+ (⦇𝚏⦈𝜌 ⦇g⦈𝜌 ⦇𝚣⦈𝜌 )))𝐹 ))}
= ⊓{𝐹 | 𝐹 ≥ 5 ⊕ (Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚡⦈𝜌 ⊕ (⊔⦇𝚡⦈𝜌 ≥1⊕⦇𝚢⦈𝜌 +⦇𝚣)⦈𝜌 ((⦇g⦈𝜌 ⦇𝚢⦈𝜌 ) + (𝐹 ⦇g⦈𝜌 ⦇𝚣⦈𝜌 ))))}
≤ ⊓{𝐹 | 𝐹 ≥ 5 ⊕ (Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚡⦈𝜌 ⊕ (((⦇g⦈𝜌 (⦇𝚡⦈𝜌 − 1)) + (𝐹 ⦇g⦈𝜌 (⦇𝚡⦈𝜌 − 1)))))}
≤ Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .(5 ⊕ (⦇g⦈𝜌 ⦇𝚡⦈𝜌 )) × (2 × ⦇𝚡⦈𝜌 )2
In the penultimate line, we obtain an upper-bound on the interpretation by approximating the
case interpretation, substituting ⦇𝑥⦈𝜌 − 1 to both ⦇𝑦⦈𝜌 and ⦇𝑧⦈𝜌 . This is a first step of relaxation
where we find an upper bound for the max. The below inequality holds for any non decreasing
function 𝐹 :
𝑎 = ⊔⦇𝚡⦈𝜌 ≥1⊕⦇𝚢⦈𝜌 +⦇𝚣)⦈𝜌 𝐹 (⦇𝚢⦈𝜌 , ⦇𝚣⦈𝜌 ) ≤ ⊔⦇𝚡⦈𝜌 ≥1⊕⦇𝚢⦈𝜌 ,⦇𝚡⦈𝜌 ≥1⊕⦇𝚣)⦈𝜌 𝐹 (⦇𝚢⦈𝜌 , ⦇𝚣⦈𝜌 ) = 𝑏.
Consequently, ⊔{𝐹 | 𝐹 ≥ 𝑏} ≤ ⊔{𝐹 | 𝐹 ≥ 𝑎}.
In the last line, we obtain an upper-bound on the interpretation by approximating the letrec
interpretation, just checking that the function 𝐹 = Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .(5 ⊕ (⦇g⦈𝜌 ⦇𝚡⦈𝜌 )) × (2 × ⦇𝚡⦈𝜌 )2 ,
where × is the usual multiplication symbol over natural numbers, satisfies the inequality:
𝐹 ≥ 5 ⊕ (Λ⦇g⦈𝜌 .Λ⦇𝚡⦈𝜌 .⦇𝚡⦈𝜌 ⊕ (((⦇g⦈𝜌 (⦇𝚡⦈𝜌 − 1)) + (𝐹 ⦇g⦈𝜌 (⦇𝚡⦈𝜌 − 1))))).
3.6. Higher Order Polynomial Interpretations. At the present time, the interpretation of a term
of type 𝚃 can be any total functional over ⦇𝚃⦈. In the next section, we will concentrate our efforts to
study polynomial time at higher order. Consequently, we need to restrict the shape of the admissible
interpretations to higher order polynomials which are the higher order counterpart to polynomials in
this theory of complexity.
INSTRUCTIONS
11
Definition 3.16. Let 𝑃𝑖 denote a polynomial of order 𝑖 and let 𝑋𝑖 denote an order 𝑖 variable. Higher
order (order 1 and order i+1) polynomials can be defined by the following grammar:
𝑃1 ∶∶= 𝑐|𝑋0 |𝑃1 + 𝑃1 |𝑃1 × 𝑃1
𝑃𝑖+1 ∶∶= 𝑃𝑖 |𝑃𝑖+1 + 𝑃𝑖+1 |𝑃𝑖+1 × 𝑃𝑖+1 |𝑋𝑖 (𝑃𝑖+1 )
where 𝑐 represents constants in ℕ.
Clearly, the set of order 𝑖 polynomials is strictly included in the set of order 𝑖 + 1 polynomials
by the above definition. Moreover, by definition, a higher order polynomial 𝑃𝑖+1 has variables of
⃖⃖⃗ is the sequence of such variables ordered by decreasing order, we will treat the
order at most 𝑖. If 𝑋
⃖⃖⃗ 𝑖+1 (𝑋).
⃖⃖⃗
polynomial 𝑃𝑖+1 as total functions Λ𝑋.𝑃
We are now ready to define the class of functions computed by terms admitting an interpretation
that is (higher order) polynomially bounded:
Definition 3.17. Let FP𝑖 , 𝑖 ≥ 0, be the class of polynomial functions at order 𝑖 that consist in functions
computed by terms 𝙼 over the basic type 𝙽𝚊𝚝 and such that:
∙ 𝚘𝚛𝚍(𝙼) = 𝑖
∙ ⦇𝙼⦈𝜌 is bounded by an order 𝑖 polynomial (i.e. ∃𝑃𝑖 , ⦇𝙼⦈𝜌 ≤ 𝑃𝑖 ).
Example 3.18. The term 𝙼 of Example 3.15 is such that 𝚘𝚛𝚍(𝙼) = 1 and admits an interpretation
bounded by 𝑃1 = Λ𝑋0 .6𝑋02 + 5. Consequently, J𝙼K ∈ FP1 .
4. BASIC FEASIBLE FUNCTIONALS
The class of tractable type 2 functionals has been introduced by Constable and Mehlhorn [Con73,
Meh74]. It was later named BFF for the class of Basic Feasible Functionals and characterized in
terms of function algebra [CK89, IRK01]. We choose to define the class through a characterization
by Bounded Typed Loop Programs from [IKR02] which extends the original BFF to any order.
4.1. Bounded Typed Loop Programs.
Definition 4.1 (BTLP). A Bounded Typed Loop Program (BTLP) is a non-recursive and wellformed procedure defined by the grammar of Figure 2.
The well-formedness assumption is given by the following constraints: Each procedure is supposed to be well-typed with respect to simple types over 𝙳, the set of natural numbers of dyadic
representation over {0, 1} (0 ≡ 𝜖, 1 ≡ 0, 2 ≡ 1, 3 ≡ 00, …). When needed, types are explicitly mentioned in variables’ superscript. Each variable of a BTLP procedure is bound by either the procedure
declaration parameter list, a local variable declaration or a lambda abstraction. In a loop statement,
the guard variables 𝑣0 and 𝑣1 cannot be assigned to within 𝐼 ∗ . In what follows 𝑣1 will be called the
loop bound.
The operational semantics of BTLP procedures is standard: parameters are passed by call-byvalue. +, − and # denote addition, proper subtraction and smash function (i.e. 𝑥#𝑦 = 2|𝑥|×|𝑦| , the size
|𝑥| of the number 𝑥 being the size of its dyadic representation), respectively. Each loop statement
is evaluated by iterating |𝑣0 |-many times the loop body instruction under the following restriction:
if an assignment 𝑣 ∶= 𝐸 is to be executed within the loop body, we check if the value obtained by
evaluating 𝐸 is of size smaller than the size of the loop bound |𝑣1 |. If not then the result of evaluating
this assignment is to assign 0 to 𝑣.
12
E. HAINRY AND R. PÉCHOUX
(Procedures) ∋ 𝑃
(Declarations) ∋ 𝑉
𝜏
𝜏
∶∶= Procedure 𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝑣11 , … , 𝑣𝑛𝑛 ) 𝑃 ∗ 𝑉 𝐼 ∗ Return 𝑣𝙳𝑟
∶∶= Var 𝑣𝙳1 , … , 𝑣𝙳𝑛 ;
(Instructions) ∋ 𝐼
(Expressions) ∋ 𝐸
(Arguments) ∋ 𝐴
∶∶= 𝑣𝙳 ∶= 𝐸; | Loop 𝑣𝙳0 with 𝑣𝙳1 do {𝐼 ∗ }
𝜏
𝜏
∶∶= 1 | 𝑣𝙳 | 𝑣𝙳0 + 𝑣𝙳1 | 𝑣𝙳0 − 𝑣𝙳1 | 𝑣𝙳0 #𝑣𝙳1 | 𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝐴11 , … , 𝐴𝑛𝑛 )
∶∶= 𝑣 | 𝜆𝑣1 , … , 𝑣𝑛 .𝑣(𝑣′1 … , 𝑣′𝑚 ) with 𝑣 ∉ {𝑣1 , … , 𝑣𝑛 }
Figure 2: BTLP grammar
Let the size of an argument be the number of syntactic elements in it. The size of input arguments
is the sum of the size of the arguments in the input.
𝜏
𝜏
Definition 4.2 (Time complexity). For a given procedure 𝑃 of parameters (𝑣11 , … , 𝑣𝑛𝑛 ), we define
its time complexity 𝑡(𝑃 ) to be a function of type ℕ → ℕ that, given inputs of type 𝜏1 × … × 𝜏𝑛 and
size bounded by 𝑛 returns the maximal number of assignments executed during the evaluation of the
procedure.
We are now ready to provide a definition of Basic Feasible Functionals at any order:
Definition 4.3 (BFF𝑖 ). For any 𝑖 ≥ 1, BFF𝑖 is the class of order 𝑖 functionals computable by a BTLP
procedure1 .
It is straightforward that BFF1 = FPTIME and BFF2 =BFF.
4.2. Safe Feasible Functionals. Now we restrict the domain of BFF𝑖 classes to inputs in BFF𝑘 for
𝑘 < 𝑖, the obtained classes are named SFF for Safe Feasible Functionals.
Definition 4.4 (SFF𝑖 ). SFF1 is defined to be the class of order 1 functionals computable by BTLP a
procedure and, for any 𝑖 ≥ 1, SFF𝑖+1 is the class of order 𝑖 + 1 functionals computable by BTLP a
procedure on the input domain SFF𝑖 . In other words,
SFF 1
SFF 𝑖+1
=BFF1 ,
=BFF𝑖+1↾SFF 𝑖
∀𝑖, 𝑖 ≥ 1
This is not a huge restriction since we want an arbitrary term of a given complexity class at
order 𝑖 to compute over terms that are already in classes of the same family at order 𝑘, for 𝑘 < 𝑖.
Consequently, programs can be built in a constructive way component by component. Another strong
argument in favor of this domain restriction is that the partial evaluation of a functional at order 𝑖
will, at the end, provide a function in ℕ ⟶ ℕ that is shown to be in BFF1 (=FPTIME).
5. A CHARACTERIZATION
OF SAFE FEASIBLE FUNCTIONALS OF ANY ORDER
In this section, we show our main characterization of safe feasible functionals:
Theorem 5.1. For any order 𝑖 ≥ 0, the class of functions in FP𝑖+1 over
class of functions in SFF𝑖+1 , up to an isomorphism. In other words,
SFF 𝑖+1
FP 𝑘 ,
𝑘 ≤ 𝑖, is exactly the
≡ FP𝑖+1↾(∪𝑘≤𝑖 FP𝑘 ) ,
for all 𝑖 ≥ 0, up to an isomorphism.
1As demonstrated in [IKR02], all types in the procedure can be restricted to be of order at most 𝑖 without any distinction.
INSTRUCTIONS
13
Proof. For a fixed 𝑖, the theorem is proved in two steps: Soundness, Theorem 5.2, and Completeness,
Theorem 5.14. Soundness consists in showing that any term 𝙼 whose interpretation is bounded by an
order 𝑖 polynomial, computes a function in SFF𝑖 . Completeness consists in showing that any BTLP
procedure 𝑃 of order 𝑖 can be encoded by a term 𝙼 computing the same function and admitting a
polynomial interpretation of order 𝑖.
Notice that functions in SFF𝑖+1 return the dyadic representation of a natural number. Consequently, the isomorphism is used on functions in FP𝑖 to illustrate that a function of type ℕ → (ℕ → ℕ)
and order 1 is isomorphic to a functional of type ℕ → ℕ and of the same order using decurryfication
and pair encoding over ℕ. In order to simplify the treatment we will restrict ourselves to functional
terms computing functionals that are terms of type 𝚃 ⟶ b, with b ∈ B, in the remaining of the
paper.
5.1. Soundness. The soundness means that any term whose interpretation is bounded by an order
𝑖 polynomial belongs to SFF𝑖 . For that, we will show that the interpretation allows us to bound the
computing time with an higher order polynomial.
Theorem 5.2. Any functional term 𝙼 whose interpretation is bounded by an order 𝑖 polynomial,
computes a functional in SFF𝑖 .
Proof. For order 1, consider that the term 𝙼 has an interpretation bounded by a polynomial 𝑃1 . For
any value 𝑣, we have, by Corollary 3.10, that the computing time of 𝙼 on input 𝑣 bounded by ⦇𝙼 𝑣⦈𝜌 .
Consequently, using Lemma 3.12, we have that:
⦇𝙼 𝑣⦈𝜌 = ⦇𝙼⦈𝜌 ⦇𝑣⦈𝜌 = ⦇𝙼⦈𝜌 (|𝑣|) ≤ 𝑃1 (|𝑣|).
Hence 𝙼 belongs to FPTIME = SFF 1 .
Now, for higher order, let 𝙼 be an order 𝑖 + 1 term of interpretation ⦇𝙼⦈𝜌 . There exists an
order 𝑖 + 1 polynomial 𝑃𝑖+1 such that ⦇𝙼⦈𝜌 ≤ 𝑃𝑖+1 . We know that on input 𝙽, 𝙼 normalizes in
𝑂(⦇𝙼 𝙽⦈𝜌 ), by Corollary 3.11. Since 𝙽 computes a functional J𝙽K ∈ SFF𝑖 there is a polynomial 𝑃𝑖
such that ⦇𝙽⦈𝜌 ≤ 𝑃𝑖 , by induction on the order 𝑖. Consequently, we obtain that the maximal number
of reduction steps is bounded polynomially in the input size by:
⦇𝙼 𝙽⦈𝜌 = ⦇𝙼⦈𝜌 ⦇𝙽⦈𝜌 ≤ 𝑃𝑖+1 ◦𝑃𝑖
that is, by a polynomial 𝑄𝑖+1 of order 𝑖 + 1 defined by 𝑄𝑖+1 = 𝑃𝑖+1 ◦𝑃𝑖 .
The above result holds for terms over arbitrary basic inductive types, by considering that each
value on such a type encodes an integer value.
5.2. Completeness. To prove that all functions computable by a BTLP program of order 𝑖 can be
defined as terms admitting a polynomial interpretation, we proceed in several steps:
(1) We show that it is possible to encode each BTLP procedure 𝑃 into an intermediate procedure
❲𝑃 ❳ of a language called IBTLP (See Figure 3) for If-Then-Else Bounded Typed Loop Program such that ❲𝑃 ❳ computes the same function as 𝑃 using the same number of assignments
(i.e. with the same time complexity).
(2) We show that we can translate each IBTLP procedure ❲𝑃 ❳ into a flat IBTLP procedure ❲𝑃 ❳,
i.e. a procedure with no nested loops and no procedure calls.
14
E. HAINRY AND R. PÉCHOUX
𝜏
𝜏
𝐼𝑃 ∶∶= 𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝑣11 , … , 𝑣𝑛𝑛 ){𝐼𝑃 ∗ 𝐼𝑉 𝐼𝐼 ∗ } ret 𝑣𝙳𝑟
𝐼𝑉 ∶∶= var 𝑣𝙳1 , … , 𝑣𝙳𝑛 ;
𝐼𝐼 ∶∶= 𝑣𝙳 ∶= 𝐼𝐸 𝙳 ; | loop 𝑣𝙳0 {𝐼𝐼 ∗ } | 𝚒𝚏 𝑣𝙳 { 𝐼𝐼 ∗ } [else { 𝐼𝐼 ∗ }] | (𝐼𝐼 ∗ )𝑣𝙳
𝜏
𝜏
𝐼𝐸 ∶∶= 1 | 𝑣𝙳 | 𝑣𝙳0 + 𝑣𝙳1 | 𝑣𝙳0 − 𝑣𝙳1 | 𝑣𝙳0 #𝑣𝙳1 | 𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝐼𝐴11 , … , 𝐼𝐴𝑛𝑛 ) | 𝑣𝙳0 × 𝑣𝙳1 | 𝚌𝚞𝚝(𝑣𝙳 )
| 𝚌𝚑𝚔𝚋𝚍(𝐸, 𝑋)
𝐼𝐴 ∶∶= 𝑣 | 𝜆𝑣1 , … , 𝑣𝑛 .𝑣(𝑣′1 … , 𝑣′𝑚 ) with 𝑣 ∉ {𝑣1 , … , 𝑣𝑛 }
Figure 3: IBTLP grammar
(3) Then we transform the flat IBTLP procedure ❲𝑃 ❳ into a “local” and flat IBTLP procedure
[❲𝑃 ❳]∅ checking bounds locally in each assignment instead of checking it globally in each
loop. For that purpose, we use a polynomial time computable operator of the IBTLP language called 𝚌𝚑𝚔𝚋𝚍. The time complexity is then asymptotically preserved.
(4) Finally, we compile the IBTLP procedure [❲𝑃 ❳]∅ into a term of our language and we use
completeness for first order function to show that there is a term computing the same function and admitting a polynomial interpretation. This latter transformation does not change
the program behavior in terms of computability and complexity, up to a O, but it makes
the simulation by a functional program easier as each local assignment can be simulated
independently of the context.
The 3 first steps just consist in transforming a BTLP procedure into a IBTLP procedure in order to
simplify the compilation procedure of the last step. These steps can be subsumed as follows:
Language
BTLP
Program
𝑃
IBTLP
❲❳
❲𝑃 ❳
IBTLP
flat
❲𝑃 ❳
local
IBTLP
[❲𝑃 ❳]∅
compile
𝚌𝚘𝚖𝚙([❲𝑃 ❳]∅ )
For each step, we check that the complexity in terms of reduction steps is preserved and that the
transformed program computes the same function.
5.2.1. From BTLP to IBTLP.
Definition 5.3 (IBTLP). A If-Then-Else Bounded Typed Loop Program (IBTLP) is a non-recursive
and well-formed procedure defined by the grammar of Figure 3.
The well-formedness assumption and variable bounds are the same as for BTLP. In a loop statement, the guard variable 𝑣0 still cannot be assigned to within 𝐼𝐼 ∗ . A IBTLP procedure 𝐼𝑃 has also
a time complexity 𝑡(𝐼𝑃 ) defined similarly to the one of BTLP procedures.
The main distinctions between an IBTLP procedure and a BTLP procedure are the following:
∙ there are no loop bounds in IBTLP loops. Instead loop bounds are written as instruction annotations: a IBTPL loop (loop 𝑣𝙳0 {𝐼𝐼 ∗ })𝑣𝙳 corresponds to a BTLP instruction Loop 𝑣𝙳0 with 𝑣𝙳 do {𝐼𝐼 ∗ }.
∙ IBTLP includes a conditional statement if 𝑣𝙳 { 𝐼𝐼1∗ } else { 𝐼𝐼2∗ } evaluated in a standard way:
if variable 𝑣 is 𝜖 (the dyadic encoding of 0) then it is evaluated to 𝐼𝐼2∗ . In all other cases, it
is evaluated to 𝐼𝐼1∗ , the else branching being optional.
∙ IBTLP includes a basic operator × such that 𝑥 × 𝑦 = 2|𝑥|+|𝑦| .
INSTRUCTIONS
15
∙ IBTLP includes a unary operator 𝚌𝚞𝚝 which removes the first character of a number (i.e.
𝚌𝚞𝚝(𝜖) = 𝜖, 𝚌𝚞𝚝(𝑖𝑣) = 𝑣).
∙ IBTLP includes an operator 𝚌𝚑𝚔𝚋𝚍 computing the following function:
{
J𝐸K if |J𝐸K| ≤ |𝑥|, 𝑥 ∈ 𝑋
𝚌𝚑𝚔𝚋𝚍(𝐸, 𝑋) =
0
otherwise
where J𝐸K is the dyadic number obtained after the evaluation of expression 𝐸 and 𝑋 is a
finite set of variables.
Notice that 𝚌𝚑𝚔𝚋𝚍 is in SFF1 provided that the input 𝐸 is computable in polynomial time and both
× and cut are in SFF1 . The semantics of an IBTLP procedure is also similar to the one of a BTLP
procedure: during the execution of an assignment, the bound check is performed on instruction
annotations instead of being performed on loop bounds. However IBTLP is strictly more expressive
than BTLP from an extensional perspective: a loop can be unbounded. This is the reason why
only IBTLP procedures obtained by well-behaved transformation from BTLP procedures will be
considered in the remainder of the paper.
Now we define a program transformation ❲.❳ from BTLP to IBTLP. For each loop of a BTLP
program, this transformation just consists in recording the variable appearing in the with argument
of a contextual loop and putting it into an instruction annotation as follows:
❲Loop 𝑣𝙳0 with 𝑣𝙳1 do {𝐼 ∗ }❳ = (loop 𝑣𝙳0 {❲𝐼❳∗ })𝑣𝙳
1
Any assignment is left unchanged and this transformation is propagated inductively on procedure
instructions so that any inner loop is transformed. We denote by ❲𝑃 ❳ the IBTLP procedure obtained
from the BTLP procedure 𝑃 . Hence the following lemma straightforwardly holds:
Lemma 5.4. Given a procedure 𝑃 , let J𝑃 K denote the function computed by 𝑃 . For any BTLP
procedure 𝑃 , we have J𝑃 K = J❲𝑃 ❳K and 𝑡(𝑃 ) = 𝑡(❲𝑃 ❳).
Proof. The transformation is semantics-preserving (the computed function is the same). Any assignment in 𝑃 corresponds to exactly one assignment in ❲𝑃 ❳ and the number of iterations of Loop 𝑣 with …
and loop 𝑣 are both in |𝑣|.
5.2.2. From IBTLP to Flat IBTLP.
Definition 5.5 (Flat IBTLP). A If-Then-Else Bounded Typed Loop Program (IBTLP) is flat if it
does not contain nested loops.
We will show that it is possible to translate any IBTLP procedure into a Flat IBTLP procedure
using the if construct.
There are only three patterns of transformation:
(1) one pattern for nested loops, called Unnest,
(2) one pattern for sequential loops, called Unseq,
(3) and one for procedure calls inside a loop, called Unfold
that we describe below:
(1) The first transformation Unnest consists in removing a nested loop of a given procedure.
Assume we have a IBTLP procedure with two nested loops:
( loop x { 𝐼𝐼1∗ ( loop y {𝐼𝐼2∗ }) 𝑧 𝐼𝐼3∗ }) w
we can translate it to a IBTLP procedure as:
16
E. HAINRY AND R. PÉCHOUX
total := w×y ; dx := 1; gt := total ; gy := y;
lb := x#total ;
loop lb {
if dx {(𝐼𝐼1∗ ) w }
if gy {((𝐼𝐼2∗ ) z ) w gy := cut ( gy );}
if dx { dx := 𝜀; (𝐼𝐼3∗ ) w }
if gt { gt := cut ( gt );}
else { gt := total ; gy := y ; dx := 1;}
}
where all the variables dx, total, gt, gy and lb are fresh local variables and 𝐼𝐼1∗ , 𝐼𝐼2∗ and
𝐼𝐼3∗ are instructions with no loop. The intuition is that variables dx and gy tell whether the
loop should execute 𝐼𝐼1∗ and 𝐼𝐼2∗ respectively. Variable gt counts the number of times to
execute the subloop.
(2) The second transformation Unseq consists in removing two sequential loops of a given procedure. Assume we have a IBTLP procedure with two sequential loops,
( loop x 1 {𝐼𝐼1∗ }) w1 𝐼𝐼2∗ ( loop x 3 {𝐼𝐼3∗ }) w3
We can translate it to a single loop IBTLP procedure as:
gx := x1 ; dy := 1; lb := x 1 ×x 3 ;
loop lb {
if gx { gx := cut gx ; (𝐼𝐼1∗ ) w1 }
else {
if dy {𝐼𝐼2∗ }
else { dy := 𝜀; (𝐼𝐼3∗ ) w3 }
}
}
where gx, dy and lb are fresh local variables and 𝐼𝐼1∗ , 𝐼𝐼2∗ and 𝐼𝐼3∗ are instructions with no
loop.
(3) The last transformation Unfold consists in removing one procedure call of a given procedure.
Assume we have a IBTLP procedure with a call to procedure 𝑃 of arity 𝑛: 𝑋 := 𝑃 (𝐼𝐴1 , ..., 𝐼𝐴𝑛 );
We can translate it to a computationally equivalent IBTLP procedure after removing the
call to procedure 𝑃 . For that purpose, we carefully alpha-rename all the variables of the procedure declaration (parameters and local variables) to obtain the procedure 𝑃 (𝑣1 , … , 𝑣𝑛 ){𝐼𝑃 ∗ 𝐼𝑉 𝐼𝐼 ∗ }
ret 𝑣𝑟 then we add the procedure declarations 𝐼𝑃 ∗ to the caller procedure list of procedure
declarations, and the local variables 𝐼𝑉 and parameters 𝑣1 , … , 𝑣𝑛 to the caller procedure
list of local variable declarations and then we generate the following code:
𝑣1 := 𝐼𝐴1 ; … 𝑣𝑛 := 𝐼𝐴𝑛 ; 𝐼𝐼 ∗ 𝑋 := 𝑣𝑟 ;
This program transformation can be extended straightforwardly to the case where the procedure call is performed in a general expression. Notice that unfolding a procedure is necessary
as nested loops may appear because of procedure calls.
These three patterns can be iterated on a IBTLP procedure (from top to bottom) to obtain a Flat
IBTLP procedure in the following way:
INSTRUCTIONS
𝜏
𝜏
17
𝜏
𝜏
[𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝑣11 , … , 𝑣𝑛𝑛 )𝐼𝑃 ∗ 𝐼𝑉 𝐼𝐼 ∗ ret 𝑣𝙳𝑟 ]𝑋 = 𝑣𝜏1 ×…×𝜏𝑛 →𝙳 (𝑣11 , … , 𝑣𝑛𝑛 )𝐼𝑃 ∗ 𝐼𝑉 [𝐼𝐼]∗𝑋 ret 𝑣𝑟
[𝐼𝐼𝑣 ]𝑋 = [𝐼𝐼]𝑋∪{𝑣}
[if 𝑣𝙳 { 𝐼𝐼1∗ } else { 𝐼𝐼2∗ }]𝑋 = if 𝑣𝙳 { [𝐼𝐼1 ]∗𝑋 } else { [𝐼𝐼2 ]∗𝑋 }
[loop 𝑣𝙳0 {𝐼𝐼 ∗ }; ]𝑋 = loop 𝑣𝙳0 {[𝐼𝐼]∗𝑋 }
[𝑣𝙳 ∶= 𝐼𝐸; ]𝑋 = 𝑣𝙳 ∶= 𝚌𝚑𝚔𝚋𝚍(𝐼𝐸 𝙳 , 𝑋);
Figure 4: From Flat IBTLP to Local and Flat IBTLP
Definition 5.6. The transformation 𝐼𝑃 is a mapping from IBTLP to IBTLP defined by:
𝐼𝑃 = (Unnest! ◦Unseq! )! ◦Unfold! (𝐼𝑃 ),
where, for given function f, 𝑓 ! is the least fixpoint of 𝑓 on a given input and ◦ is the usual mapping
composition.
Notice that this procedure is polynomial time computable as each application of an Unfold call
consumes a procedure call (and procedures are non recursive) and each application of an Unnest or
Unseq call consumes one loop. Consequently, fixpoints always exist.
Lemma 5.7. For any BTLP procedure 𝑃 , ❲𝑃 ❳ is a flat IBTLP procedure.
Proof. First notice that repeated application of the Unfold pattern may only introduce a constant
number of new loops as procedures are non-recursive. Consequently, the fixpoint Unfold! is defined
and reached after a constant number (in the size of the procedure) of applications. Each application of a Unseq pattern or Unnest pattern decreases by one the number of loop within a procedure.
Consequently, a fixpoint is reached (modulo 𝛼-equivalence) and the only programs for which such
patterns cannot apply are programs with a number of loops smaller than 1.
Lemma 5.8. For any BTLP procedure 𝑃 , we have J❲𝑃 ❳K = J𝑃 K and 𝑡(❲𝑃 ❳) = 𝑂(𝑡(𝑃 )).
Proof. In the first equality, the computed functions are the same since the program transformation
preserves the extensional properties. For the second equality, the general case can be treated by
a structural induction on the procedure ❲𝑃 ❳. For simplicity, we consider the case of a procedure
❲𝑃 ❳ only involving 𝑛 nested loops over guard variables 𝑥1 , … , 𝑥𝑛 and loop bounds 𝑥𝑛+1 , … , 𝑥2𝑛 ,
respectively over one single basic instruction (e.g. one assignment with no procedure call). With
inputs of size 𝑚, this procedure will have a worst case complexity of 𝑚𝑛 (when the loop bounds are
not reached). Consequently, 𝑡(𝑃 )(𝑚) = 𝑚𝑛 . This procedure will be transformed into a flat IBTLP
procedure using the Unnest transformation 𝑛 − 1 times over a variable 𝑧 containing the result of the
computation 𝑥1 #((𝑥2 ×𝑥𝑛+1 )#(… 𝑥3 ×𝑥𝑛+2 )#(𝑥𝑛 ×𝑥2𝑛 ) …) as initial value. Consequently, on an input
of size 𝑚, we have 𝑡(❲𝑃 ❳)(𝑚) ≤ |𝑧| = |𝑥1 |×(|𝑥2 |+|𝑥𝑛+1 |)×…×(|𝑥𝑛 |+|𝑥2𝑛 |) ≤ 2𝑛−1 ×𝑚𝑛 = 𝑂(𝑚𝑛 ),
as for each 𝑖 |𝑥𝑖 | ≤ 𝑚 and by definition of the # and × operators. We conclude using Lemma 5.4.
5.2.3. From Flat IBTLP to Local and Flat IBTLP. Now we describe the last program transformation
[−]𝑋 that makes the check bound performed on instruction annotations explicit. For that purpose,
[−]𝑋 makes use of the operator 𝚌𝚑𝚔𝚋𝚍 and records a set of variables 𝑋 (the annotations enclosing the
considered instruction). The procedure is described in Figure 4. Notice that the semantics condition
ensuring that no assignment must be performed on a computed value of size greater than the size
18
E. HAINRY AND R. PÉCHOUX
𝚌𝚘𝚖𝚙(𝑣(𝑣1 , … , 𝑣𝑚 ) 𝐼𝑃 ∗ var 𝑣𝑚+1 , … , 𝑣𝑛 ; 𝐼𝐼 ∗ ret 𝑣𝑟 ) = 𝜆𝑠.𝜋𝑟𝑛 (𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼 ∗ ) 𝑠)
𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼1 … 𝐼𝐼𝑘 ) = 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼𝑘 ) … 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼1 )
𝚌𝚘𝚖𝚒𝑛 (𝑣𝑖 ∶= 𝐼𝐸; ) = 𝜆⟨𝑣1 , … , 𝑣𝑛 ⟩.⟨𝑣1 , … , 𝑣𝑖−1 , 𝚌𝚘𝚖𝚎(𝐼𝐸), 𝑣𝑖+1 , … , 𝑣𝑛 ⟩
(
𝚕𝚎𝚝𝚁𝚎𝚌 𝑓 = 𝜆𝑡̃.𝜆𝑠.𝚌𝚊𝚜𝚎 𝑡̃ 𝚘𝚏 𝜖
𝑛
∗
𝚌𝚘𝚖𝚒 (loop 𝑣𝑖 {𝐼𝐼 }) = 𝜆⟨𝑣1 , … , 𝑣𝑛 ⟩.(
𝑗(𝑡)
→
→
𝑠
𝑓 𝑡 (𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼 ∗ ) 𝑠)
)
𝑣𝑖 )
𝚌𝚘𝚖𝚒𝑛 (if 𝑣𝙳𝑖 { 𝐼𝐼1∗ } else { 𝐼𝐼2∗ }) = 𝜆⟨𝑣1 , … , 𝑣𝑛 ⟩.𝚌𝚊𝚜𝚎 𝑣𝑖 𝚘𝚏 𝜖 → 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼2∗ ) | 𝑗(𝑡) → 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼1∗ )
with 𝑗 ∈ {0, 1}
𝚌𝚘𝚖𝚎(𝑐) = 𝑐,
𝑐 ∈ {1, 𝑣𝙳 }
𝚌𝚘𝚖𝚎(𝑣𝙳0 𝚘𝚙 𝑣𝙳1 ) = 𝚘𝚙 𝑣0 𝑣1 ,
𝚘𝚙 ∈ {+, −, #, ×}
𝚌𝚘𝚖𝚎(𝚌𝚞𝚝(𝑣𝙳 )) = 𝚌𝚞𝚝 𝑣
𝚌𝚘𝚖𝚎(𝚌𝚑𝚔𝚋𝚍(𝐼𝐸, {𝑣𝑗1 , … , 𝑣𝑗𝑟 })) = 𝚌𝚑𝚔𝚋𝚍 𝚌𝚘𝚖𝚎(𝐼𝐸) (⊓𝙳𝑟 𝑣𝑗1 … 𝑣𝑗𝑟 )
𝚌𝚘𝚖𝚎(𝑣(𝐼𝐴1 , … , 𝐼𝐴𝑛 )) = 𝑣 𝚌𝚘𝚖𝚎(𝐼𝐴1 ) … 𝚌𝚘𝚖𝚎(𝐼𝐴𝑛 )
𝚌𝚘𝚖𝚎(𝜆𝑣1 , … , 𝑣𝑛 .𝑣(𝑣′1 … , 𝑣′𝑚 )) = 𝜆𝑣1 . … 𝜆𝑣𝑛 .(… (𝑣 𝑣′1 ) … 𝑣′𝑚 ) …)
Figure 5: Compiler from Local and Flat IBTLP to terms
of the loop bounds for a BTLP procedure or than the size of the loop annotations for a IBTLP
procedure has been replaced by a local computation performing exactly the same check using the
operator 𝚌𝚑𝚔𝚋𝚍. Consequently, we have:
Lemma 5.9. For any BTLP procedure 𝑃 , we have J[❲𝑃 ❳]∅ K = J𝑃 K and 𝑡([❲𝑃 ❳]∅ ) = 𝑂(𝑡(𝑃 )).
Proof. We use Lemma 5.8 together with the fact that extensionality and complexity are both preserved.
5.2.4. From Local and Flat IBTLP to terms. We then encode Flat and Local IBTLP in our functional
language. For this, we define a procedure 𝚌𝚘𝚖𝚙 that will “compile” IBTLP procedures into terms.
For that purpose, we suppose that the target term language includes constructors for dyadic numbers
(𝜖, 0 and 1), a constructor for tuples ⟨…⟩, all the operators of IBTLP as basic prefix operators (+, -,
#, …), min operators ⊓𝙳𝑛 computing the min of the sizes of 𝑛 dyadic numbers and a 𝚌𝚑𝚔𝚋𝚍 operator
of arity 2 such that J𝚌𝚑𝚔𝚋𝚍(𝙼, 𝙽)K = J𝙼K if |J𝙼K| ≤ |J𝙽K| (and 𝜖 otherwise). All these operators
are extended to be total functions in the term language: they return 𝜖 on input terms for which they
are undefined. Moreover, we also require that the Flat and Local IBTLP procedures given as input
are alpha renamed so that all parameters and local variables of a given procedure have the same
name and are indexed by natural numbers. The compiling process is described in Figure 5, defining
𝚌𝚘𝚖𝚙, 𝚌𝚘𝚖𝚒𝑛 , 𝚌𝚘𝚖𝚎 that respectively compile procedures, instructions and expressions. The 𝚌𝚘𝚖𝚒
compiling process is indexed by the number of variables in the program 𝑛.
For convenience, let 𝜆⟨𝑣1 , … , 𝑣𝑛 ⟩. be a shorthand notation for 𝜆𝑠.𝚌𝚊𝚜𝚎 𝑠 𝚘𝚏 ⟨𝑣1 , … , 𝑣𝑛 ⟩ → and
let 𝜋𝑟𝑛 be a shorthand notation for the r-th projection 𝜆⟨𝑣1 , … , 𝑣𝑛 ⟩.𝑣𝑟 .
The compilation procedure works as follows: any Local and Flat IBTLP procedure of the shape
𝑣(𝑣1 , … , 𝑣𝑚 ) 𝐼𝑃 ∗ var 𝑣𝑚+1 , … , 𝑣𝑛 ; 𝐼𝐼 ∗ ret 𝑣𝑟 will be transformed into a term of type 𝜏1 × … ×
𝜏𝑛 → 𝜏𝑟 , provided that 𝜏𝑖 is the type of 𝑣𝑖 and that 𝜏1 × … × 𝜏𝑛 is the type for n-ary tuples of the
shape ⟨𝑣1 , … , 𝑣𝑛 ⟩. Each instruction within a procedure of type 𝜏1 × … × 𝜏𝑛 → 𝜏𝑟 will have type
INSTRUCTIONS
19
𝜏1 × … × 𝜏𝑛 → 𝜏1 × … × 𝜏𝑛 . Consequently, two sequential instructions 𝐼𝐼1 𝐼𝐼2 within a procedure
of 𝑛 variables will be compiled into a term application of the shape 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼2 ) 𝚌𝚘𝚖𝚒𝑛 (𝐼𝐼1 ) and
instruction type is preserved by composition. Each assignment of the shape 𝑣𝑖 ∶= 𝐼𝐸; is compiled
into a term that takes a tuple as input and returns the identity but on the i-th component. The ith component is replaced by the term obtained after compilation of 𝐼𝐸 on which a checkbound
is performed. The min operator applied for this checkbound is ⊓𝙳𝑛 whenever 𝑋 is of cardinality
𝑛. The compilation process for expressions is quite standard: each construct is replaced by the
corresponding construct in the target language. Notice that the compiling procedure is very simple
for procedures as it only applies to Flat IBTLP procedures on which any procedure call has been
removed by unfolding. The only difficulty to face is for loop compilation: we make use of a letrec
of type 𝙳 → 𝜏1 × … × 𝜏𝑛 → 𝜏1 × … × 𝜏𝑛 . The first argument is a counter and is fed with a copy of
the loop counter 𝑣𝑖 so that the obtained term has the expected type 𝜏1 × … × 𝜏𝑛 → 𝜏1 × … × 𝜏𝑛 .
Again, for a given term 𝙼 of type 𝜏1 … ⟶ … 𝜏𝑛 ⟶ 𝜏, we define its time complexity 𝑡(𝙼)
to be a function of type ℕ → ℕ that, for any inputs of type 𝜏𝑖 𝙽⃗ and size bounded by 𝑚 returns the
maximmal value of |𝙼 𝙽⃗ ⇒ J𝙼 𝙽⃗ K|. In other words, 𝑡(𝙼)(𝑚) = max|𝙽⃗ |≤𝑚 {|𝙼 𝙽⃗ ⇒ J𝙼 𝙽⃗ K|}.
Lemma 5.10. For any BTLP procedure 𝑃 , we have J𝚌𝚘𝚖𝚙([❲𝑃 ❳]∅ )K = J𝑃 K and 𝑡(𝚌𝚘𝚖𝚙([❲𝑃 ❳]∅ )) =
𝑂(𝑡(𝑃 )).
Proof. The term obtained by the compilation process is designed to compute the same function
as the one computed by the initial procedure. It remains to check that for a given instruction the
number of reductions remains in 𝑂(𝑡(𝑃 )). This is clearly the case for an assignment as it consists in
performing one beta-reduction, one case reduction, and the evaluation of a fixed number of symbols
within the expression to evaluate. For an iteration the complexity of the letrec call is in |𝑣𝑖 | as for
an IBTLP loop. Indeed each recursive call consists in removing one symbol. Finally, it remains to
apply lemma 5.9.
Example 5.11. Let us take a simple BTLP procedure:
Procedure mult (x ,y)
Var z ,b; z := 𝜖; b := x #y;
Loop x with b do {z := z+ y ;}
Return z
It will be translated in Local and Flat IBTLP as:
mult (x ,y ){
var z ,b ; z := 𝜖; b := x#y ;
loop x {z := chkbd (z +y ,{ b });}
ret z}
And be compiled modulo 𝛼-equivalence in a term as:
(
)
⎛
⎞
𝜆𝑡̃.𝜆̃𝑠.case 𝑡̃ of 𝜀 → 𝑠
𝜆⟨x, y, z, b⟩. 𝚕𝚎𝚝𝚁𝚎𝚌 𝑓 =
x ⎟
𝑗(𝑡) → 𝑓 𝑡 (𝙼 𝑠̃))
𝜆𝑠.𝜋34 ⎜
⎜
⎟
⎝ (𝜆⟨x, y, z, b⟩.⟨x, y, z, # x y⟩(𝜆⟨x, y, z, b⟩.⟨x, y, 𝜖, 𝑏⟩ 𝑠))
⎠
where 𝙼 = 𝜆⟨x, y, z, b⟩.⟨x, y, 𝚌𝚑𝚔𝚋𝚍 (+ z y) (⊓𝙳1 b), b⟩
Now it remains to check that any term obtained during the compilation procedure has a polynomial interpretation. In order to show this result, we first demonstrate some intermediate results:
20
E. HAINRY AND R. PÉCHOUX
Lemma 5.12. Given an assignment 𝜌, the operators +, −, #, ×, 𝚌𝚞𝚝, ⊓𝙳𝑛 and 𝚌𝚑𝚔𝚋𝚍 admit a polynomial sup-interpretation.
Proof. We can check that the following are polynomial sup-interpretations.
⦇+⦈𝜌
⦇−⦈𝜌
⦇#⦈𝜌
⦇×⦈𝜌
⦇𝚌𝚞𝚝⦈𝜌
= Λ𝑋.Λ𝑌 .𝑋 + 𝑌 ,
= Λ𝑋.Λ𝑌 .𝑋,
= Λ𝑋.Λ𝑌 .𝑋 × 𝑌 ,
= Λ𝑋.Λ𝑌 .𝑋 + 𝑌 ,
= Λ𝑋. ⊔ (𝑋 − 1, 0),
⦇⊓𝙳𝑛 ⦈𝜌 = Λ𝑋1 . … .Λ𝑋𝑛 . ⊓ℕ (𝑋1 , … , 𝑋𝑛 ),
⦇𝚌𝚑𝚔𝚋𝚍⦈𝜌 = Λ𝑋.Λ𝑌 .𝑌 .
The inequalities are straightforward for the basic operators +, −, #, …. For ⦇𝚌𝚑𝚔𝚋𝚍⦈𝜌 , to be a supinterpretation, we have to check that:
∀𝙼, ∀𝙽, ⦇𝚌𝚑𝚔𝚋𝚍 𝙼 𝙽⦈𝜌 ≥ ⦇J𝚌𝚑𝚔𝚋𝚍(𝙼, 𝙽)K⦈𝜌 .
⦇𝚌𝚑𝚔𝚋𝚍 𝙼 𝙽⦈𝜌 = ⦇𝚌𝚑𝚔𝚋𝚍⦈𝜌 ⦇𝙼⦈𝜌 ⦇𝙽⦈𝜌
= Λ𝑋.Λ𝑌 .𝑌 ⦇𝙼⦈𝜌 ⦇𝙽⦈𝜌
= ⦇𝙽⦈𝜌
≥ ⦇J𝙽K⦈𝜌
by Corollary 3.10
≥ |J𝙽K|
by Lemma 3.12
≥ |J𝚌𝚑𝚔𝚋𝚍(𝙼, 𝙽)K|
≥ ⦇J𝚌𝚑𝚔𝚋𝚍(𝙼, 𝙽)K⦈𝜌
by Definition of 𝚌𝚑𝚔𝚋𝚍
by Lemma 3.12
and so the conclusion.
In the particular case where J𝙽K is not a value, we have ⦇J𝚌𝚑𝚔𝚋𝚍𝙼, 𝙽)K⦈𝜌 = ⦇𝜖⦈𝜌 = 1 so the
inequality is preserved.
Now we are able to provide the interpretation of each term encoding an expression:
Corollary 5.13. Given a BTLP procedure 𝑃 , any term 𝙼 obtained by compiling an expression 𝐼𝐸
(i.e. 𝙼 = 𝚌𝚘𝚖𝚎(𝐼𝐸)) of the flat and local IBTLP procedure [❲𝑃 ❳]∅ admits a polynomial interpretation ⦇𝙼⦈𝜌 .
Proof. By unfolding, there is no procedure call in the expressions of procedure [❲𝑃 ❳]∅ . By Lemma 5.12
and by Definition 3.5 , any symbol of an expression has a polynomial interpretation, obtained by finite composition of polynomials.
Theorem 5.14. Any BTLP procedure 𝑃 can be encoded by a term 𝙼 computing the same function
and admitting a polynomial interpretation.
Proof. We have demonstrated that any loop can be encoded by a first order term whose runtime is
polynomial in the input size. Each higher order expression in a tuple can be encoded by a first order
term using defunctionalization. Consequently, by completeness of first-order polynomial interpretations there exists a term computing the same function and admitting a polynomial interpretation.
INSTRUCTIONS
21
Example 5.15. We provide a last example to illustrate the expressivity of the presented methodology.
Define the term 𝙼 by:
𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝜆g.𝜆𝚡.𝚌𝚊𝚜𝚎 𝚡 𝚘𝚏 𝚌(𝚢, 𝚣) → g (𝚏 g 𝚣),
𝚗𝚒𝚕 → 𝚗𝚒𝚕
The interpretation of 𝙼 has to satisfy the following inequality:
⊓{𝐹 | 𝐹 ≥ Λ𝐺.Λ𝑋.4 ⊕ (1 ⊔ (⊔𝑋≥1 (𝐺(𝐹 𝐺 (𝑋 − 1))))}
Clearly, this inequality does not admit any polynomial interpretation as it is at least exponential
in 𝑋. Now consider the term 𝙼 (𝜆𝚡.1 + (1 + (𝚡)). The term 𝜆𝚡.1 + (1 + (𝚡)) can be given the
interpretation Λ𝑋.𝑋 ⊕ 3. We have to find a function 𝐹 such that 𝐹 (Λ𝑋.𝑋 ⊕ 3) ≥ Λ𝑌 .4 ⊕
(1 ⊔ (⊔𝑌 ≥1 (𝐹 (Λ𝑋.𝑋 ⊕ 3) (𝑌 − 1)) ⊕ 3)). This inequality is satisfied by the function 𝐹 such that
𝐹 (Λ𝑋.𝑋 ⊕ 3) 𝑌 = (7 × 𝑌 ) ⊕ 4 and consequently 𝙼 has an interpretation. This highlights the fact
that a term may have an interpretation even if some of its subterms might not have any. As expected,
any term admitting an interpretation of the shape Λ𝑋.𝑋 ⊕ 𝛽, for some constant 𝛽, will have an
interpretation when applied as first operand of this fold function.
6. CONCLUSION
This paper introduces a theory for higher-order interpretations that can be used to deal with higherorder complexity classes. This is a novel approach but there are still some important issues to discuss.
∙ Synthesis: it is well-known for a long time that the synthesis problem that consists in finding
the sup-interpretation of a given term is undecidable in general for first order terms using
interpretations over natural numbers (see [Péc13] for a survey). As a consequence this problem is also undecidable for higher order. However it is also well-known that it will become
decidable if programs are defunctionalized (up to a first order term) using polynomial functions over real numbers.
∙ Expressivity: the expressivity of interpretations is as usual their main drawback: As for first
order interpretations, a lot of interesting terms computing polynomial time functions will
not have any polynomial interpretation, i.e. its interpretation will sometimes be ⊤, although
the function will be computed by another algorithm (term) admitting a finite interpretation.
It should be possible to relax the tool and to extend it to more complex data structures such
as streams, infinite trees. At least, the presented paper has shown that the tool expressivity
can be extended to higher order.
∙ Space: space issues were not discussed in this paper as there is no theory for higher order polynomial space. In analogy with the usual first order theory, a suitable definition for
higher order space complexity classes could be to consider terminating terms admitting a
non strict polynomial interpretation. By non strict, we mean, for example, that the last rule
of Definition 3.5 can be replaced by:
⦇𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝙼⦈𝜌 = 1 ⊕ ⊓{𝐹 ∈ ⦇𝚃⦈ | 𝐹 ≥ Λ⦇𝚏⦈𝜌 .⦇𝙼⦈𝜌 (𝐹 )}
Thus termination is lost as the term 𝚕𝚎𝚝𝚁𝚎𝚌 𝚏 = 𝚏 could be interpreted by 1 ⊕ Λ𝐹 .𝐹 .
This is the reason why we need to restrict the class to terminating terms. However a result
equivalent to Lemma 3.9 holds: we still keep a non-strict upper bound on the interpretation
of any derived term.
22
E. HAINRY AND R. PÉCHOUX
R EFERENCES
[BL12]
Patrick Baillot and Ugo Dal Lago. Higher-order interpretations and program complexity. In Computer Science
Logic (CSL’12), volume 16 of LIPIcs, pages 62–76, 2012.
[BL16] Patrick Baillot and Ugo Dal Lago. Higher-order interpretations and program complexity. Inf. Comput., volume 248, pages 56–81, 2016.
[BMM11] Guillaume Bonfante, Jean-Yves Marion and Jean-Yves Moyen. Quasi-interpretations a way to control resources. In Theoretical Computer Science, volume 412, pages 2776–2796, Elsevier, 2011.
[CL92] Adam Cichon and Pierre Lescanne. Polynomial interpretations and the complexity of algorithms. In Journal of
Automated Reasoning, volume 34(4), pages 325–363, 2005.
[CMPU05] Evelyne Contejean, Claude Marché, Ana Paula Tomás and Xavier Urbain. Higher-order interpretations and
program complexity. International Conference on Automated Deduction, Springer, pages 139–147, 1992.
[CK89] Stephen A. Cook and Bruce M. Kapron. Characterizations of the basic feasible functionals of finite type. In
Foundations of Computer Science, pages 154–159. IEEE Computer Society, 1989.
[Con73] Robert L. Constable. Type two computational complexity. In Proc. 5th annual ACM Symposium on Theory of
Computing, pages 108–121. ACM, 1973.
[Fér14] Hugo Férée. Complexité d’ordre supérieur et analyse récursive. (Higher order complexity and computable analysis). PhD thesis, University of Lorraine, Nancy, France, 2014.
[FHHP15] Hugo Férée, Emmanuel Hainry, Mathieu Hoyrup and Romain Péchoux. Characterizing polynomial time complexity of stream programs using interpretations In Theor. Comput. Sci., volume 585, pages 41–54, 2015.
[HP17] Emmanuel Hainry and Romain Péchoux. Higher order interpretation for higher order complexity In 21st International Conference on Logic for Programming, Artificial Intelligence and Reasoning, pages 269–285, 2017.
[IKR02] Robert J. Irwin, Bruce M. Kapron, and James S. Royer. On characterizations of the basic feasible functionals
(part II). Technical report, Syracuse University, 2002.
[IRK01] Robert J. Irwin, James S. Royer, and Bruce M. Kapron. On characterizations of the basic feasible functionals
(part I). Journal of Functional Programming, 11(1):117–153, 2001.
[KC10] Akitoshi Kawamura and Stephen A. Cook. Complexity theory for operators in analysis. In Leonard J. Schulman,
editor, Proc. 42nd ACM Symposium on Theory of Computing, pages 495–502. ACM, 2010.
[Ko91] Ker-I Ko. Complexity Theory of Real Functions. Birkhäuser, 1991.
[KS01] William A Kirk and Brailey Sims. Handbook of metric fixed point theory. Springer, 2001.
[Lan79] D.S. Lankford. On proving term rewriting systems are noetherian. Technical report, 1979.
[Meh74] Kurt Mehlhorn. Polynomial and abstract subrecursive classes. In Proc. 6th annual ACM Symposium on Theory
of Computing, pages 96–109. ACM, 1974.
[MN70] Zohar Manna and Steven Ness. On the termination of Markov algorithms. In Third Hawaii International Conference on System Science, pages 789–792, 1970.
[Péc13] Romain Péchoux. Synthesis of sup-interpretations: A survey. Theor. Comput. Sci., 467:30–52, 2013.
[Tar55] Alfred Tarski. A lattice-theoretical fixpoint theorem and its applications. Pacific journal of Mathematics,
5(2):285–309, 1955.
[VdP93] Jaco Van de Pol. Termination proofs for higher-order rewrite systems. In International Workshop on HigherOrder Algebra, Logic, and Term Rewriting, volume 816 of LNCS, pages 305–325. Springer, 1993.
[Win93] Glynn Winskel. The formal semantics of programming languages: an introduction. MIT press, 1993.
| 6cs.PL
|
Statistical mechanics of unsupervised feature learning in a
restricted Boltzmann machine with binary synapses
Haiping Huang∗
RIKEN Brain Science Institute, Wako-shi, Saitama 351-0198, Japan
(Dated: March 8, 2017)
arXiv:1612.01717v2 [cs.LG] 7 Mar 2017
Revealing hidden features in unlabeled data is called unsupervised feature learning, which plays an important role in pretraining a deep neural network. Here we
provide a statistical mechanics analysis of the unsupervised learning in a restricted
Boltzmann machine with binary synapses. A message passing equation to infer the
hidden feature is derived, and furthermore, variants of this equation are analyzed. A
statistical analysis by replica theory describes the thermodynamic properties of the
model. Our analysis confirms an entropy crisis preceding the non-convergence of the
message passing equation, suggesting a discontinuous phase transition as a key characteristic of the restricted Boltzmann machine. Continuous phase transition is also
confirmed depending on the embedded feature strength in the data. The mean-field
result under the replica symmetric assumption agrees with that obtained by running
message passing algorithms on single instances of finite sizes. Interestingly, in an
approximate Hopfield model, the entropy crisis is absent, and a continuous phase
transition is observed instead. We also develop an iterative equation to infer the
hyper-parameter (temperature) hidden in the data, which in physics corresponds
to iteratively imposing Nishimori condition. Our study provides insights towards
understanding the thermodynamic properties of the restricted Boltzmann machine
learning, and moreover important theoretical basis to build simplified deep networks.
Keywords: Neuronal networks, Cavity and replica method, Learning theory, Statistical inference
I.
INTRODUCTION
Standard machine learning algorithms require a huge number of training examples to
uncover hidden features, while humans and other animals can learn new concepts from
only a few examples without any supervision signal [1]. Learning hidden features in unlabeled training examples is called unsupervised learning. Understanding how the number of
examples confines the learning process is of fundamental importance in both cognitive neuroscience and machine learning [2, 3]. As already observed in training of deep neural networks,
unsupervised pretraining can significantly enhance the final performance, because the unsupervised pretraining provides a good initial region in parameter space from which the final
fine-tuning starts [4]. However, there are few theoretical works addressing how unsupervised
learning extracts hidden features. One potential reason is that the unsupervised learning
process in a deep neural network is typically very complicated. Hence, understanding the
∗
Electronic address: [email protected]
2
mechanism of unsupervised learning in simple models is of significant importance.
This topic is recently studied based on the Bayesian inference framework [5]. In this
recent work, the authors treated each example (data) as a constraint on the factor graph, and
reformulated the learning of model parameters as a Bayesian inference problem on graphical
models, and derived the message passing equations to infer the hidden feature from a finite
amount of data. They observed an entropy crisis in a simple restricted Boltzmann machine
(RBM) model, and predicted a discontinuous phase transition. However, in an approximate
Hopfield model obtained by a high-temperature expansion of the RBM model, the entropy
crisis is absent, and instead, a continuous transition is observed. These properties observed
in studies of single instances capture key characteristics of the unsupervised feature learning.
Here, we further demonstrate that the message passing equation derived in the recent
work [5] agrees with the statistical analysis of the system in the thermodynamic limit via
replica theory, a standard theoretical tool in spin glass theory of disordered systems [6].
The replica computation predicts the location of phase transition separating an impossibleto-infer regime to inferable regime. This transition can be continuous depending on the
embedded feature strength. A discontinuous phase transition always exists in a restricted
Boltzmann machine learning, but absent in an approximate Hopfield model where only continuous phase transition is observed. We also develop an iterative equation to infer the
hyper-parameter (temperature) hidden in the data, which in physics corresponds to iteratively imposing Nishimori condition. This iterative scheme can even quantitatively predict
how apparent features embedded in a real dataset are. Our analysis gives a thorough understanding of novel properties of the restricted Boltzmann machine within replica symmetric
approximation.
This paper is structured as follows. In Sec. II, we introduce a simple RBM model for
unsupervised feature learning, and propose the Bayesian inference framework to derive the
message passing equation on factor graph representation of the learning process, and this
equation is then statistically analyzed and compared with replica computation under the
replica symmetric assumption. A more efficient approximate message passing equation is also
derived. We also derive an iterative equation based on Bayes rule to predict the unknown
temperature (feature strength) in the data. In Sec. III, we approximate the RBM with
the Hopfield model where the stored pattern is interpreted as the feature vector. Similar
statistical analysis is carried out, and its physical implications are discussed. We end the
paper with a summary in Sec. IV.
II.
SIMPLE RESTRICTED BOLTZMANN MACHINE LEARNING AND ITS
STATISTICAL MECHANICS PROPERTIES
A.
Simple restricted Boltzmann machine learning with binary synapses
Restricted Boltzmann machine is a basic unit widely used in building a deep belief network [4, 7]. It consists of two layers of neurons. The visible layer receives the input examples
while the other hidden layer builds an internal representation of the input. No lateral connections exist within each layer for computational efficiency. The symmetric connections
(synapses) between visible and hidden neurons are considered as features the network tries
to learn from a large number of training examples.
It is a common strategy to use sampling-based gradient-decent method to learn features
in the data [4], however, the gradient-decent learning is complicated and not amenable for
3
analytical studies. Recent work showed that learning features can also be studied within a
Bayesian learning framework [5], which has the advantage of accounting for the uncertainty
(about the features) caused by noises in the data [8], and furthermore can be analytically
studied on probabilistic graphical models.
We focus on an unsupervised learning of finite samplings generated by a simple RBM,
where a single hidden neuron is considered. The task is to uncover an unknown rule embedded in the unlabeled data. The rule is represented by a binary feature vector defined
as {ξi } where i goes from 1 to N, the number of neurons in the visible layer. We assume
components of the true hidden feature vector connecting the visible neurons and the hidden
neuron can take only two values, i.e., +1 or −1, with equal probabilities. Then this feature
vector is used to generate
independent random samples according to the joint probability
√
P
−βE(σ,h)/ N
P (σ, h) ∝ e
, where E(σ, h) = − i hξi σi , and σ is the visible configuration and
h is the hidden neuron’s state. Both h and components of σ take binary values (±1) as
well. ξi represents the connection between visible neuron σi and the unique hidden neuron
h. Moreover, we assume ξ is a binary feature vector in the current setting of unsupervised
learning. A rescaled feature factor by the system size is assumed as well. The feature vector
is also multiplied by an inverse-temperature parameter β to investigate effects of the feature
strength on the unsupervised learning. For simplicity, we consider the case of neurons without any external biases (fields). Generalization to the case of neurons with external fields is
straightforward (Appendix F).
The distribution of σ can be obtained by marginalization of h on the joint distribution
P (σ, h), resulting in
cosh √βN ξ T σ
,
P (σ|ξ) = P
(1)
√β ξ T σ
cosh
σ
N
P
√β ξ T σ
where the normalization is in fact independent of ξ, since
cosh
=
σ
N
iN
h
2 cosh √βN . Suppose we have M independent samples or examples {σ a }M
a=1 to learn the
true hidden feature vector ξ, using the Bayes’ formula, we have the posterior distribution of
the feature vector as
Q
P (σ a |ξ)
β T a
1Y
a
a
P (ξ|{σ }) = P Q
(2)
cosh √ ξ σ ,
=
a
Z a
N
ξ
a P (σ |ξ)
where Z is the partition function of the model, a goes over all examples and T denotes a
vector transpose operation. A uniform prior probability for the feature vector is assumed for
simplicity. A large β indicates the feature in the data is strong, and expected to be revealed
by a few examples, while a weak feature vector may not be revealed by a huge number of
examples. Each example serves as a constraint to the learning process. Once M > 1, the
model becomes non-trivial as the partition function could not be computed exactly for a
large number of visible neurons. This M can be finite or proportional to the system size, and
in the latter case we define a data density as α = M/N. Hereafter, we omit the conditional
dependence of P (ξ|{σ a}) on {σ a }.
4
FIG. 1: (Color online) Schematic illustration of factor graph representation and message passing.
Left panel: circle nodes indicate feature components to be inferred. Square nodes indicate data
constraints. The strength of each connection is given by the data, e.g., σia indicates the strength
(±1) with which the feature component ξi is related to a-th example. Right panel: the top
panel shows constraint a collects information from its neighboring feature nodes other than i and
produces an output message to node i. The bottom panel shows node i collects information from
its neighboring constraints other than b and produces an output message to node b. The figure is
taken from Ref [5].
B.
Bayesian learning via message passing
We call optimizing the marginal posterior probability of feature vectors given the data
as Bayesian learning in the current unsupervised learning context. That is, we compute the
maximizer of the posterior marginals (MPM) estimator ξˆi = arg maxξi Pi (ξi ) [9]. We define
P
the overlap between the inferred feature vector and the true one as q = N1 i ξitrue ξˆi , where
ξ̂ is the inferred feature vector. The MPM estimator maximizes the overlap. If q = 0, the
examples do not give any information about the feature vector. If q = 1, the feature vector is
perfectly determined. In the numerical simulations, q is evaluated by an average over many
random instances (different true feature
D P vectors).
D EE To compare with replica computation
1
true ˆ
(Sec. II D), the average value q = N i ξi
ξi
is used. Note that the inner average is
the thermal average, and the outer average is taken with respect to different choices of true
features.
The statistical inference now is simplified to the computation of marginal probabilities,
e.g., Pi (ξi ), which is still a hard problem due to the interaction among example constraints.
However, by mapping the model (Eq. (2)) onto a factor graph [10, 11], the marginal probability can be estimated by message passing (Fig. 1). For simplicity, we give the final simplified
message passing equations (sMP) as follows (detailed derivations are given in Ref [5], also
in Appendix A):
X
mi→a = tanh
ub→i ,
(3a)
b∈∂i\a
√
ub→i = tanh−1 tanh(βGb→i ) tanh(βσib / N ) ,
(3b)
P
where Gb→i = √1N j∈∂b\i σjb mj→b. The cavity magnetization is defined as mj→b =
P
ξj ξj Pj→b (ξj ). mi→a can be interpreted as the message passing from feature i to the
5
data constraint a, while ub→i can be interpreted as the message passing from data constraint b to its feature i. If the weak correlation assumption (also named Bethe approximation [12]) is self-consistent, the sMP would converge to a fixed point corresponding to
a stationary point of the Bethe free energy function with respect to the cavity messages
{mi→a , ua→i } [10]. From this fixed point, one can extract useful information about the
1+mi ξi
true feature vector of the data,
by calculating the marginal probability as Pi (ξi ) = 2
P
where mi = tanh
b∈∂i ub→i . Note that we perform the feature inference using the same
inverse-temperature as used to generate the data, thus the inference is Bayes-optimal and
satisfies the Nishimori condition [9]. In the simulation section, we also perform the Bayesian
inference using slightly different temperatures.
P
Next,P
we compute the Bethe free energy, which is obtained by −βNfRS = i ln Zi −
(N − 1) a ln Za . The free energy contribution of a feature node reads
Y
Xh
√ i
Ga→i ,
(4)
β 2 Ξ2a→i /2 + ln cosh βGa→i + βσia / N + ln 1 +
ln Zi =
a∈∂i
a∈∂i
and the free energy contribution of a data node reads
ln Za = β 2 Ξ2a /2 + ln cosh βGa ,
(5)
P
P
where we define Ga→i = e−2ua→i , Ξ2a→i ≃ N1 j∈∂a\i (1 − m2j→a ), Ga = √1N i∈∂a σia mi→a , and
P
Ξ2a = N1 i∈∂a (1 − m2i→a ).
Another important quantity is the number of feature vectors consistent
Pwith the presented
1
random samplings, characterized by the entropy per neuron s = − N ξ P (ξ) ln P (ξ). In
the presence of a larger dataset, the generative machine should have less uncertainty about
the underlying feature, corresponding to small or vanishing entropy. The entropy can be
∂
derived by using the standard thermodynamic formula s = (1 − β ∂β
)(−βfRS ). Under the
Bethe approximation, s is evaluated
as summing
P
P up contributions from single feature nodes
and example nodes: Ns = i ∆Si −(N −1) a ∆Sa , where single feature node contribution
is expressed as
!
i
Xh
Y
√
β 2 Ξ2a→i /2 + ln cosh(βGa→i + βσia / N ) + ln 1 +
∆Si =
Ga→i
a∈∂i
−
a∈∂i
"
X
a∈∂i
Ha→i (+1) +
Y
a∈∂i
Ga→i
X
a∈∂i
#
Ha→i (−1) / 1 +
Y
a∈∂i
Ga→i
!
(6)
,
and single example contribution reads
∆Sa = ln cosh(βGa ) − β 2 Ξ2a /2 − βGa tanh(βGa ),
(7)
√
√
where we define Ha→i (ξi ) = β 2 Ξ2a→i + (βGa→i + βσia ξi / N) tanh(βGa→i + βσia ξi / N ).
C.
Approximate message passing equations
One iteration of the sMP equation (Eq. (3)) requires the time complexity of the order
O(MN) and memory of the order O(MN). The sMP equation can be further simplified
6
by reducing computational complexity. The final equation in physics is called ThoulessAnderson-Palmer (TAP) equation [13], and in information theory is named approximate
message passing (AMP) equation [14]. One strategy is to use large-N limit. We first get
βσb
the cavity bias in this limit as ub→i ≃ √Ni tanh βGb→i . Then by applying the same large-N
βσa
expansion, we obtain mi→a ≃ mi − (1 − m2i ) √Ni tanh βGa→i . Therefore, we get the first AMP
equation as follows:
1 X a
σi mi − β(1 − Q) tanh βGa ,
(8)
Ga = √
N i∈∂a
P
P
σib
tanh βGb→i , and note
where Q ≡ N1 i m2i . Then we define the local field Hi = b∈∂i √N
that Gb→i = Gb −
σb
√ i mi→b ,
N
Hi ≃
we can obtain an approximate Hi in the large-N expansion,
X σb
βmi X
√ i tanh βGb −
(1 − tanh2 βGb ).
N
N
b∈∂i
b∈∂i
(9)
The last term in the expression of Hi serves as an Onsager reaction term in a standard TAP
equation. Finally we arrive at the second AMP equation:
!
2
X
X βσ b
β
m
i
√ i tanh βGb −
(1 − tanh2 βGb ) .
(10)
mi ≃ tanh
N
N
b∈∂i
b∈∂i
Now we have only N + M equations to solve rather than 2NM equations in the sMP
equation (Eq. (3)). To make AMP equations converge in a parallel iteration, the time
indexes for the variables are important [15, 16]. Here we write down the closed form of
AMP equation with correct time indexes:
1 X a t−1
Gt−1
=√
σi mi − β(1 − Qt−1 ) tanh βGt−2
(11a)
a
a ,
N i∈∂a
!
2 t−1 X
X βσ b
β
m
2
i
i
√ tanh βGt−1
−
mti ≃ tanh
(1 − tanh βGt−1
(11b)
b
b ) ,
N
N
b∈∂i
b∈∂i
where t denotes the time index for iteration. These time indexes just follow the temporal
order when we derive the AMP equation from the sMP equation.
D.
Statistical analysis of sMP equations and replica computation
Next, we P
give a statistical analysis of the sMP equation. We first define the cavity field
hi→a = √1N b∈∂i\a σib tanh βGb→i . Under the replica symmetric assumption, hi→a follows
a Gaussian distribution with mean zero and variance αQ̂ in the large-N limit. We define
Q̂ ≡ tanh2 βGb→i . Similarly, Gb→i also follows a Gaussian distribution with mean zero
but variance Q. Therefore, we arrive at the following thermodynamic equation:
Z
q
2
(12a)
Q = Dz tanh β αQ̂z,
Z
p
Q̂ = Dz tanh2 β Qz,
(12b)
7
−z 2 /2
where Dz = dze√2π . One can expect that when α is small, only one solution of Q = 0
exists for the above equation, however, at some critical αc , there is a nontrivial solution of
Q 6= 0, which signals the fixed point of sMP or AMP starts to contain information about
the underlying true feature vector. αc can be determined by expanding the above equation
around Q = 0. The expansion leads to αc = β −4, which implies that when α < αc , Q = 0
is the stable solution of the thermodynamic equation, but as long as α > αc , the Q = 0 is
not the stable solution any more. However, as we compare this solution with sMP result
on single instances, Q 6= 0 solution does not match the numerical simulation very well. To
explain this, a replica computation is required.
Now, we perform a replica computation of the free energy function. Instead of calculating
a disorder average of ln Z, the replica trick computes the disorder average of an integer power
of Z, then the free energy density (multiplied by −β) can be obtained as [9]
− βf =
ln hZ n i
,
n→0,N →∞ nN
lim
(13)
where the limit N → ∞ should be taken first since we can apply the saddle-point analysis [9], and the disorder average is taken over all possible samplings (data) and the random
realizations of true feature vector. The explicit form of hZ n i reads
γ a
X
XY
βξ σ
1
a
true
n
√
P ({σ }|ξ )
cosh
hZ i = N
,
(14)
2
N
{σa ,ξtrue }
{ξγ } a,γ
where γ indicates the replica index. We leave the technical details to the appendix D, and
give the final result here. The free energy function reads,
Z
√
r̂(r − 1) αβ 2
+
(1 − r) + Dz ln 2 cosh(q̂ + r̂z)
−βfRS = −q q̂ +
2
2
Z
Z
(15)
p
−β 2 /2
2
Dy Dt cosh βt ln cosh β(qt + r − q y).
+αe
and the associated saddle-point equations are expressed as
Z
√
q = Dz tanh(q̂ + r̂z),
Z
√
r = Dz tanh2 (q̂ + r̂z),
Z
Z
p
2 −β 2 /2
Dt Dy sinh βt tanh β(qt + r − q 2 y),
q̂ = αβ e
Z
Z
p
2 −β 2 /2
Dt Dy cosh βt tanh2 β(qt + r − q 2 y).
r̂ = αβ e
(16a)
(16b)
(16c)
(16d)
We make some remarks about the above saddle-point equations. q indicates the typical
value of the overlap between the true feature vector and the estimated one, while r indicates
the typical value of the overlap between two estimated feature vectors selected from the
posterior probability (Eq. (2)). According to the Nishimori condition, q = r, implying
that the embedded true feature vector follows the same posterior distribution in Bayesian
inference. We verify this point later in numerical solution of the saddle point equations.
8
Assuming q and r are both small values close to zero, Eq. (16) in this limit implies that a
critical αc = β14 , above which q = 0 is not a stable solution any more. By expanding Eq. (16)
around q = 0 up to the second order O(q 2 ), we find q ≃ β 2 (α − αc ) when α approaches αc
from above.
Note that by assuming q = 0 in Eq. (16), we obtain Eq. (12); this is the reason why
Eq. (12) can predict the correct threshold for transition but could not describe the property
of q 6= 0. This may be because the Gaussian assumption for messages does not generally
hold when the messages start to have partial (even full) alignment with the true feature
vector (i.e., q 6= 0), as also observed in a similar study of the retrieval phase in the Hopfield
model by message passing methods [17].
The entropy can be derived from the free energy, and the result (Appendix D) is given
by
Z
√
r̂(r − 1) αβ 2
s = −2q q̂ + rr̂ +
−
(r + 1) + Dz ln 2 cosh(q̂ + r̂z)
2
2
Z
Z
(17)
p
2
−β /2
2
Dy Dt cosh βt ln cosh β(qt + r − q y).
+αe
2
When q = 0, s = ln 2 − αβ2 , which coincides with that obtained with cavity method
(Appendix B). Given q = 0, that is the data still do not contain information about the
hidden feature, the entropy will become negative once α > 2βln2 2 ≡ αs=0 . This suggests that
the entropy crisis can even occur within
√
pq α= 0 regime. Alternatively, at a fixed α, the entropy
crisis occurs at a temperature Tc = 2 ln 2 . Setting αc = αs=0 , one obtains Te = 2 ln 2,
which distinguishes two cases: (i) for T > Te , the transition of q from zero to non-zero value
takes place after the entropy crisis; (ii) for T < Te , the transition takes place before the
crisis. This has clear physical implications. If the transition occurs before the entropy crisis,
the location of transition identified by the replica-symmetric theory is correct. However,
the transition after the crisis is incorrect under the replica-symmetric assumption, because
although the replica-symmetric solution is stable, the entropy is negative, violating the fact
that for a system with discrete degrees of freedom, the entropy should be non-negative.
According to arguments in Refs. [18, 19], there exists a discontinuous transition before the
entropy crisis takes place, since the transition can not continuously emerge from a stable
replica-symmetric solution. Without further solving complex more-steps replica symmetric
breaking equations, we adopt an alternative explanation of this entropy crisis. Under the
current context, if the data size is large enough, the data would shrink the feature space to
a sub-exponential regime where the number of candidate features is not exponential with
N any more. In this case, we encounter the entropy crisis. Therefore, the entropy crisis
separates an exponential regime from a sub-exponential regime. The transition for q occurs
within the exponential regime if the feature strength is strong enough (large β).
E.
Simulations on single instances compared with theory
We use the above mean field theory to analyze single realizations (instances) of the
unsupervised learning model. Random
samplings are first generated according to a RBM
√
distribution P (σ, h) ∝ e−βE(σ,h)/ N [11, 20], where the energy is rescaled by the system size
and the inverse temperature, which tunes difficulty level of the learning task. These random
samplings then serve as the quenched disorder specifying the interaction between the feature
9
(a)
0.6
=1.0
entropy per neuron
=0.8
0.4
0.2
0.0
-0.2
-0.4
0.0
0.5
0.4
r,
q,
r,
0.3
q, r
q,
1.0
1.5
2.0
2.5
3.0
3.5
(b)
=0.8
=0.8
=1.0
=1.0
0.2
0.1
0.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
FIG. 2: (Color online) Mean-field result obtained by replica theory compared with simulations
carried out by running sMP on single instances of RBM. In simulations, we consider 20 instances
of size N = 400. (a) Entropy per neuron versus data density α. The lines are replica result, while
the symbols are results obtained on single instances. The error bars are smaller than the symbol
size. (b) Order parameters (q, r) versus α. q = r as expected. The numerical simulations (q) on
single instances are indicated by symbols (solid circles for β = 1.0, and solid squares for β = 0.8),
while the theoretical predictions of replica computation are indicated by lines.
vector and the example constraint (Fig. 1). Finally, by initializing the message on each link
of the factor graph (Fig. 1), we run the sMP equation (Eq. (3)) until it converges within
a prefixed precision. From the fixed point, we compute the entropy of consistent feature
vectors and the overlap between the inferred feature vector and the true one.
We first compare results of the message passing algorithm with those obtained by replica
computation. In Fig. 2 (a), we show the entropy density versus the data density α. The
entropy characterizes how the number of candidate feature vectors compatible with the given
data changes with the network size N. The replica result predicts an entropy crisis, i.e., the
entropy becomes negative at some data size, but the negativity of the entropy is not allowed
in a system with discrete degrees of freedom (ξi = ±1). This implies that, a discontinuous
phase transition should be present before the crisis [18, 19]. The results obtained by running
sMP coincide perfectly with the replica result. The entropy density decreases more rapidly
10
1.0
sMP
0.8
AMP
q
0.6
0.4
0.2
0.0
0
1
2
3
4
5
FIG. 3: (Color online) Performance comparison between sMP and AMP for RBM. The model
parameters are N = 100, β = 1. 30 random instances are considered.
with α at larger β. This is expected, because large β indicates strong feature, thus to shrink
the feature space to the same size, less data is required compared to the case of detecting
weak feature (small β).
In Fig. 2 (b), we show how order parameters change with α. The simulation results agree
with the replica prediction, despite slightly large deviations observed around the transition
point. For β > Te−1 , a first continuous transition occurs at αc = 1, which should be correct
since the replica computation is stable and the entropy is positive there. At α = 1, the
system starts to have information about the embedded feature, and therefore, the overlap
q starts to increase even for a finite-size system. The asymptotic behavior of q at a slightly
larger α (> αc ) is captured by β 2 (α−αc ), as already derived in the theory section (Sec. II D).
As predicted by replica computation, at a larger value of α = αs=0 , the entropy becomes
negative. Equivalently, at this α, the entropy vanishes at a critical temperature, and thus the
equilibrium is dominated by a finite number of lowest energy states (so-called condensation
phenomenon [10, 21]). But the sMP is still stable, therefore a one-step replica symmetry
breaking solution should grow discontinuously from the replica symmetric solution, and
this second discontinuous glass transition is expected before or at the crisis α to resolve the
entropy crisis. Intuitively, we expect that the increasing data will freeze the value of synapses,
thus vanishing entropy indicates that the feature space develops isolated configurations: each
configuration forms a single valley in the free energy profile, while the number of these valleys
is not exponential any more (but sub-exponential). To prove this picture, one needs to go
beyond the replica symmetric assumption.
For β < Te−1 , the continuous transition takes place after the entropy crisis, which is
incorrect. Therefore, according to the above argument, a discontinuous transition should be
expected at or before the crisis data size. Our simulation result also confirms the Nishimori
condition (q = r), that is, when the temperature used to generate data is equal to that used
to infer the true feature, the true feature follows the posterior distribution as well. As a
consequence, the overlap between a typical feature configuration and the embedded one is
equal to the overlap between two typical configurations.
11
0.8
0.7
1.2
0.6
0.9
q
0.5
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
5
6
FIG. 4: (Color online) Inference performance with different inference temperatures for RBM. The
model parameters are N = 100, β ∗ = 0.8. 30 random instances are considered.
Secondly, the inference can also be carried out by using AMP with less requirements of
computer memory and time. The result is compared with that obtained by sMP, which is
shown in Fig. 3.
We also study the effects of temperature deviation. If the inference is carried out in a
different temperature from that used to generate the data, is the performance degraded?
We address this question by considering two different temperatures: one is slightly larger
than the data temperature (0.9β ∗ ); the other is slightly below the data temperature (1.2β ∗),
where β ∗ denotes the inverse data temperature. As shown in Fig. 4, when β = β ∗ , the
performance is optimal in the inferable regime, compared to other inference temperatures,
as expected from the Nishimori condition [9]. Large fluctuations around the transition point
may be caused by finite size effects.
Finally, we explore the effect of network size keeping the identical feature strength (Fig. 5).
At a given number of examples, larger network size yields better performance in terms of
prediction overlap. However, the performance seems to get saturated when N ≃ 1000 at
a relatively large M in the current context. Using a larger network seems to make the
unsupervised learning better, but further increasing the network size has a little effect on
the performance.
F.
Learning temperature parameters from a dataset: how cold is a dataset?
In Fig. 4, we have showed the inference performance with slightly different inference
temperatures. Is it possible to infer the true temperatures used to generate the data itself?
If we can learn the temperature parameters, we can know the typical properties of phase
transitions intrinsic in the system. This is possible by applying the Bayesian rule once again.
12
1.0
0.8
0.6
q
N=100
N=500
0.4
N=1000
0.2
0.0
0
100
200
300
400
500
600
M
FIG. 5: (Color online) Inference √
overlap versus the number of examples for RBM with different
values of N . Feature strength β/ N is kept constant (0.1).
The posterior probability of β given the data {σ a }M
a=1 is given by
P ({σ a }|ξ, β)P0(ξ, β)
P
dβ ξ P ({σ a }|ξ, β)P0(ξ, β)
ξ
√
X
−N M ln 2 cosh(β/ N ) Y
β T a
1
e
cosh √ ξ σ
=
Z({σ a }) ξ
N
a
P (β|{σ a}) =
X
P (β, ξ|{σ a}) = R
∝ e−M
β2
2
(18)
Z(β, {σ a}),
where we used the uniform prior probability P0 for the hyper-parameters. Note that
Z(β, {σ a }) is the same partition function as in Eq. (2). We maximize the posterior probability with respect to β, and obtain the self-consistent equation β should satisfy:
∂ ln Z(β, {σ a})
= Nαβ.
∂β
(19)
The left hand side of the above equation is exactly the negative energy (−Nǫ), which can
be evaluated by sMP equation (Eq. (3)).
When
N is not very large, the equation deter
√
−1
− α√ǫ N , from which Eq. (19) is recovered in large
mining β is given by β = N tanh
N limit.PUnder the BethePapproximation, the energy per neuron ǫ can be computed by
Nǫ = − i ∆ǫi + (N − 1) a ∆ǫa , where ∆ǫi and ∆ǫa are given respectively by
!
"
#
X
X
Y
Y
∆ǫi =
Ha→i (+1) +
Ga→i
Ha→i (−1) / β + β
Ga→i ,
(20a)
∆ǫa =
a∈∂i
βΞ2a +
a∈∂i
Ga tanh(βGa ).
a∈∂i
a∈∂i
(20b)
Starting from some initial value of β, one can iteratively update the value of β until
convergence within some precision. After one updating, the messages in sMP equation
13
are also updated. To avoid numerical instability, we used the damping technique, i.e.,
β(t) = ηβ(t) + (1 − η)β(t − 1), where t denotes the iteration step and η ∈ [0, 1] is a damping
factor. It is not guaranteed that there exists unique maximum of the posterior (Eq. (18)) [16],
but if necessary, one can choose the hyper-parameter corresponding to the global maximum
of the posterior by running the sMP from different initial conditions.
In statistics, this iterative scheme is named Expectation-Maximization algorithm [22],
where the message updates are called E-step, and the temperature update is called M-step.
In physics, Eq. (19) corresponds to the Nishimori condition (q = r, see also Appendix D
for derivation of the energy function). This means that, in principle, the hyper-parameter
can be learned by iteratively imposing the Nishimori condition [16, 23]. On the Nishimori
condition, the state space of the model is simple [24], and thus sMP yields informative
information about the dominant feature vector.
We know that the temperature parameter is related to the feature strength embedded in
the data. Once we learn the temperature, we are able to know how apparent the hidden
feature is in a dataset, and determine the critical data size for unsupervised feature learning.
We first test our method in synthetic dataset as already studied in Sec. II E, where the true
value of hyper-parameter is known. We then infer the embedded feature strength in the real
dataset (MNIST handwritten digit dataset [25]), where we do not have any knowledge about
the true feature strength. Results are shown in Fig. 6. For the synthetic data at βtrue = 1,
as the data size grows, inferred value of β gets closer to the true value as expected (Fig. 6
(a)). As shown in the inset, the time (iteration steps) dependent inferred value first drops to
a lower value, and then gradually approaches the true value. With a larger data size (e.g.,
M = 800), β increases more rapidly after a sudden drop. For the real dataset (Fig. 6 (b)),
we observe that the final fixed point of β is quite large, implying that the feature strength
in the handwritten digits is very strong (β ∗ ≃ 21.5).
We also test effects of β on learned features. As shown in the feature map (Fig. 6 (c)), it
turns out that the learning works quite well for a broad range of values for β except for very
small values (e.g., 0.005), even when the data is scarce (M = 50). When β = 0.005, learning
fails to identify the meaningful feature. However, a relatively large value of β results in
well-structured feature, which has been shown to have the discriminative power for image
classification [5]. We conjecture that given the data, these values of β have non-vanishing
posterior probability (Eq. (18)). In fact, the value of β = 21.5 has a maximal posterior
probability.
III.
THE HOPFIELD MODEL AND SIMPLIFIED MESSAGE PASSING
EQUATIONS
A.
Message passing equations for an approximate Hopfield model
It is interesting to show that one can also perform the same unsupervised learning task
by using an associative memory (Hopfield) model defined by
2
β̃
T
a
Y 2N
ξ σ
P̂ (ξ) ∝
e
,
(21)
a
where β̃ = β 2 . This posterior distribution of feature vectors given the input examples can
be obtained by a small-β expansion of Eq. (2) [26]. This relationship implies that one can
14
100
0.16
N=100,
(a)
=1.0
(b)
MNIST(0&1) dataset
80
true
M=200
M=400
1.0
0.12
60
0.08
M=600
0.6
M=800
0.4
|
(t)
true
|
(t)
0.8
true
0
200
400
20
t
0.04
40
600
0
0.00
2
3
4
5
6
7
8
0
200
400
600
800
t
(c)
FIG. 6: (Color online) Inference of hyper-parameter β in synthetic (RBM) and real (MNIST, digits
0 and 1) dataset. (a) Deviation of inferred β from the true value decreases with the data size. In
simulations, we consider 10 instances of size N = 100, and use η = 0.02 and initial value of β0 = 0.8.
Two examplar trajectories of β(t) are shown in the inset. (b) Examplar trajectories of β(t) are
shown for the real dataset. We use β0 = 1.0 and η = 0.01. The fixed point does not change when
we use β0 = 1.5. (c) Feature vector (ξ) organized as a 28 × 28 matrix. They are learned from
M = 50 training images with corresponding β = 0.005, 0.5, 5.0 and 21.5 (from the left to the right,
and from the top to the bottom). The color black and white indicate active (ξi = +1) and inactive
(ξi = −1) feature components, respectively.
infer the feature vector of a RBM by an approximate Hopfield model, and the feature vector
is interpreted as the stored pattern in the Hopfield model, encoding memory characteristics
of the input data. Note that, {σ a } are still governed by a RBM distribution, whereas, by
applying the associative memory framework, we show many similar interesting properties of
the unsupervised learning model.
In analogous to the derivation of Eq. (3), we have the sMP corresponding to the posterior
probability (Eq. (21)):
X
β̃
mi→a = tanh √
σib G̃b→i Fb→i ,
(22)
N b∈∂i\a
P
where G̃b→i = √1N k∈∂b\i σkb mk→b , Fb→i = 1+ 1−β̃Cβ̃Cb→i in which Cb→i =
b→i
Details to derive Eq. (22) are given in Appendix C.
1
N
P
2
k∈∂b\i (1−mk→b ).
15
In the approximate
Hopfield model,
P
P the Bethe free energy can be constructed similarly,
i.e., −βNfRS = i ln Zi − (N − 1) a ln Za , where
#
"
X β̃
1
(1/N + G̃2a→i )Fa→i − ln(1 − β̃Ca→i ) + ln 2 cosh β̃Hi ,
(23a)
ln Zi =
2
2
a∈∂i
β̃ 2
1
G̃a Fa − ln(1 − β̃Ca ),
(23b)
2
2
P
P
where we define G̃a = √1N k∈∂a σka mk→a , Ca = N1 k∈∂a (1 − m2k→a ), Fa = 1 + 1−β̃Cβ̃Ca , and
a
P
Hi = √1N b∈∂i σib G̃b→i Fb→i .
Similar
RBM, the entropy for the approximate model can be evaluated as
P to the case in P
Ns = i ∆Si − (N − 1) a ∆Sa , where single feature node contribution reads
"
#
X 1
β̃
β̃Ca→i
′
+ (1/N + G̃2a→i )Fa→i
∆Si = −
ln(1 − β̃Ca→i ) +
2
2
2(1
−
β̃C
)
a→i
(24)
a∈∂i
+ ln 2 cosh(β̃Hi ) − (β̃Hi + β̃Hi′ ) tanh(β̃Hi ),
ln Za =
and single example contribution reads
β̃
β̃Ca
1
− G̃2a Fa′ ,
∆Sa = − ln(1 − β̃Ca ) −
(25)
2
2(1 − β̃Ca ) 2
P
β̃Ca
b
′
′
′
′
a→i
√1
,
F
=
,
and
H
=
where Fa→i
= (1−β̃C
2
2
a
i
b∈∂i σi G̃b→i Fb→i .
β̃Ca→i )
(1−β̃Ca )
N
We also derive AMP equations for the Hopfield model. Note that mi→a ≃ mi − (1 −
a
P
1
2 β̃σ
mi ) √Ni G̃a→i Fa→i . Therefore G̃a ≃ √1N i∈∂a σia mi − β̃(1 − Q)G̃a 1−β̃(1−Q)
, we thus derive the
first AMP equation for the Hopfield model as
G̃a =
where Q ≡ N1
P
1
1
√
1−β̃(1−Q) N
1 − β̃(1 − Q) X a
√
σi mi ,
N
i∈∂a
(26)
2
From the definition of the local field Hi ,
i mi .
α
a
a∈∂i σi G̃a − 1−β̃(1−Q) mi in the large-N expansion. Finally, we
AMP equation:
P
mi ≃ tanh
1 X a
αβ̃
β̃
√
mi
σi G̃a −
1 − β̃(1 − Q) N a∈∂i
1 − β̃(1 − Q)
we have Hi ≃
derive the second
!
.
(27)
Taking the time index into account, the AMP equations can be summarized in a parallel
update scheme as follows:
1 X a t−1
1
√
G̃t−1
≃
σi mi − β̃(1 − Qt−1 )G̃t−2
,
a
a
N i∈∂a
1 − β̃(1 − Qt−2 )
mti ≃ tanh
1 X a t−1
β̃
αβ̃
√
mt−1
σi G̃a −
i
t−1
t−1
1 − β̃(1 − Q ) N a∈∂i
1 − β̃(1 − Q )
(28a)
!
.
(28b)
16
B.
Thermodynamic equations for the approximate Hopfield model
In this section,
we derive the thermodynamic equation. Similarly, the local field defined as
P
1
√
hi→a = N b∈∂i\a σib G̃b→i Fb→i follows a Gaussian distribution with mean zero and variance
D
E
αQ
2
. The variance can be derived by noting that G̃b→i ≃ Q. Therefore, we have
(1−β̃(1−Q))2
the following thermodynamic equation for the Hopfield model:
!
Z
p
β̃
Q = Dz tanh2
αQz .
(29)
1 − β̃(1 − Q)
i2
h
Q = 0 is a solution of Eq. (29), however, it is stable only when α ≤ αc = 1−β̃ β̃ . This threshold can be derived by expanding Eq. (29) around Q = 0 to the first order. Interestingly, this
equation matches the mean-field equation without ferromagnetic part (related to retrieval
phase) derived in standard Hopfield model [27]. Gaussian assumption for messages does not
generally hold, particularly for those messages related to the memorized patterns [17].
We also perform replica computation for the approximate Hopfield model. The free
energy function is given in the appendix E. The associated saddle-point equations are given
as follows:
Z
√
q = Dz tanh(q̂ + r̂z),
(30a)
Z
√
(30b)
r = Dz tanh2 (q̂ + r̂z),
αβ̃ 2 q
,
q̂ =
1 − β̃(1 − r)
(30c)
αβ̃ 2 (β̃q 2 + r)
r̂ =
.
(1 − β̃(1 − r))2
(30d)
Note that, the data is generated by the RBM, but the inference is carried out in an approximate Hopfield model. Although we use the same temperature, the model mismatching leads
to q 6= r. The threshold for the transition ofhr can
i be determined by studying the linear
stability around r = 0, and the result is αcr =
1−β̃
β̃
2
, consistent with the cavity prediction.
2
β̃
When α approaches αcr from above, r behaves like r ≃ 2(1−
(α − αcr ). The transition for
β̃)
q can only be determined√ numerically, since r could not be assumed a small value, and it
R
αβ̃ 2 r
The recursive equation for r in the
follows r = Dz tanh2 ( r̂z) where r̂ = (1−β̃(1−r))
2.
regime of q = 0 is exactly the equation derived from the sMP equation (Eq. (29)).
The entropy of the model can also be similarly computed, and reads as follows:
s=
Z
√
αβ̃((1 − β̃(1 − r))(1 − 3r) + 2r + 2q 2 β̃)
α
Dz ln 2 cosh(q̂ + r̂z) − ln(1 − β̃(1 − r)) −
.
2
2(1 − β̃(1 − r))2
(31)
17
entropy per neuron or overlap
0.7
s
0.6
r
q
0.5
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
FIG. 7: (Color online) Thermodynamic properties of the approximate Hopfield model compared
with the inference performed on single instances. The lines are replica result, compared with
symbols indicating the sMP results. In simulations, we consider 20 instances of size N = 400. The
feature strength β = 0.8.
C.
Numerical simulations compared with theory
The thermodynamic properties of the approximate Hopfield model are shown in Fig. 7.
First, we show that the entropy crisis is absent in the Hopfield model, although the inference
is carried out by the sMP equation of Hopfield model. This is quite interesting, because
within the associative memory framework, the inference is improved smoothly and there
does not exist condensation in the feature space. Secondly, the replica computation predicts
q 6= r, as expected from the fact that by applying Hopfield model approximation, the
Nishimori condition does not hold. Thirdly, the simulation results obtained by running
sMP equation agree with the theoretical predictions for the entropy, in spite of observed
fluctuations of order parameters caused by finite size effects. The asymptotic behavior of r
near αcr can be analytically determined by small-r expansion of the saddle-point equation
β̃ 2
(α − αcr ) when α
(Eq. (30)). As already derived in the theory section (Sec. III B), r ≃ 2(1−
β̃)
tends to αcr from above. The transition for q can only be determined by numerically solving
the saddle-point equation. It seems that q changes smoothly to a non-zero value at the same
data size as that of the RBM. Due to the model mismatching, transition for r takes place
much more earlier than that for q.
Finally, the inference can also be carried out by using AMP with less requirements of
memory storage and computer time. The result is compared with that obtained by sMP,
which is shown in Fig. 8. We also study the effects of temperature deviation. As shown
in Fig. 9, even when β = β ∗ , it is not guaranteed that the performance is optimal in the
inferable regime, compared to other inference temperatures.
IV.
CONCLUSION
In conclusion, we build a physical model of unsupervised learning from a finite number of examples in the framework of RBM. Here, we consider binary features rather than
18
0.7
0.6
sMP
AMP
0.5
q
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
5
FIG. 8: (Color online) Performance comparison between sMP and AMP for Hopfield model. The
model parameters are N = 100, β = 0.8. 30 random instances are considered.
0.8
0.7
1.2
0.6
0.9
q
0.5
0.4
0.3
0.2
0.1
0.0
0
1
2
3
4
5
6
FIG. 9: (Color online) Inference performance with different inference temperatures for Hopfield
model. The model parameters are N = 100, β ∗ = 0.8. 30 random instances are considered.
real-valued ones; this is because binary features are more robust and efficient in large-scale
neuromorphic applications [28], yet it remains open to figure out an efficient algorithm.
We show that physics method can inspire an efficient (fully-distributed) message passing
procedure not only to infer the hidden feature embedded in a noisy data, but also to estimate the entropy of candidate features. Distinct from conventional slow sampling-based
methods, each example in this work is treated as a constraint on the factor graph, and the
message passing carries out a direct Bayesian inference of the hidden feature, which marks
an important step implementing unsupervised learning in neural networks. In particular,
the approximate message passing equation has low requirements of computer space and time
in practical applications.
19
We show that, the results obtained by the cavity method are consistent with the statistical
analysis by replica theory. The replica theory describes the thermodynamic properties of the
unsupervised learning system. It first predicts a discontinuous phase transition in a restricted
Boltzmann machine, signaled by the entropy crisis before the message passing equation loses
its stability. However, if the feature strength is strong enough, there exists another phase
transition which is continuous, i.e., the order parameter (the overlap between the true feature
vector and the inferred one) smoothly changes from zero to non-zero value. This continuous
transition will be followed by an additional discontinuous transition at a larger data size.
Interestingly, in an approximate Hopfield model, the entropy crisis is absent, and the entropy
decreases much more slowly towards zero. Therefore, there exists a continuous transition
from impossible-to-infer to inferable regime. Unlike the RBM, inference in the Hopfield
model does not satisfy the Nishimori condition, and thus the statistics of metastable states
would be very interesting, and its relationship with the dynamics of inference deserves further
investigation.
Our work not only derives in detail various kinds of message passing algorithms in a
Bayesian framework for practical applications, but also statistically characterizes the thermodynamic properties of the restricted Boltzmann machine learning with binary synapses,
and its connection with associative memory networks. Many interesting properties related to
phase transitions are also revealed. In addition, we derive an iterative equation to infer the
unknown temperature in the data, providing a quantitative measure of how cold a dataset
is. This method corresponds to Expectation-Maximization algorithm in statistics [22], and
in physics iteratively imposing Nishimori condition [9, 23]. Therefore, our study forms a
theoretical basis of unsupervised feature learning in a single simple RBM, and are expected
to be helpful in constructing a deep architecture for hierarchical information processing,
which is currently under way.
Appendix A: Derivation of simplified mesage passing equations for RBM
We first assume feature components on the factor graph are weakly correlated, then by
using the cavity method [10], we define a cavity probability Pi→a (ξi ) of ξi on a modified factor
graph with example node a removed. Due to the weak correlation assumption, Pi→a (ξi )
satisfies a recursive equation (namely belief propagation (BP) in computer science [29]):
Y
Pi→a (ξi ) ∝
µb→i (ξi ),
(A1a)
b∈∂i\a
µb→i (ξi ) =
X
{ξj |j∈∂b\i}
cosh
β
√ ξTσb
N
Y
Pj→b(ξj ),
(A1b)
j∈∂b\i
where the symbol ∝ indicates a normalization constant, ∂i\a defines the neighbors of node
i except constraint a, ∂b\i defines the neighbors of constraint b except visible node i, and
the auxiliary quantity µb→i (ξi ) represents the contribution from constraint b to visible node
i given the value of ξi [10]. Eq. (A1) has been similarly derived to understand RBM in a
recent paper [11]. Here, we exchange the role of the observed data {σ a } and that of synaptic
interaction (feature vector here), and predict feature vector given the data. Therefore the
data (random samplings), rather than the synaptic interaction in the previous work [11],
becomes a quenched disorder.
20
Note that in Eq. (A1b), the sum inside the hyperbolic cosine function with the idependent term excluded is a random variable following
a normal distributionP
with mean
P
1
1
b
2
2
Gb→i and variance Ξb→i [11], where Gb→i = √N j∈∂b\i σj mj→b and Ξb→i ≃ N j∈∂b\i (1 −
P
m2j→b ). The cavity magnetization is defined as mj→b = ξj ξj Pj→b(ξj ). Thus the intractable
sum over all ξj (j 6= i) can be replaced by an integral over the normal distribution. Using
the magnetization representation [11], the BP equation (Eq. (A1)) could be reduced to the
simplified message passing equations (see Eq. (3) in the main text).
In physics, the contribution fromPa single
Q feature node to the partition function, Zi is
obtained via cavity method as Zi = ξi =±1 b∈∂i µb→i (ξi ); the contribution of a single data
Q
P
node reads Za = {ξj |j∈∂a} cosh √βN ξ T σ a
j∈∂a Pj→a (ξj ), which can be further computed
by applying the central-limit theorem as well. This calculation is exact only when the
underlying factor graph is a tree. However, it is approximately correct when correlations
among feature components are weak. It needs to be compared with numerical simulations
and replica computations.
Appendix B: Simplification of entropy formula for RBM in the limit of q = 0
√
First, we compute ∆Si . By noting that Ξ2a→i = 1 − N1 , ln cosh(βσia / N ) ≃
√
√
2
βσia ξi / N tanh(βσia ξi / N) ≃ βN , we have
X β2
αβ 2
Ξ2a→i − αβ 2 +
+ ln 2,
∆Si = −
2
2
a∈∂i
β2
,
2N
and
(B1)
where we have used the fact that Ga→i = 0 and Ga→i = 1. Analogously, ∆Sa is simplified to
be
β2
∆Sa = − .
(B2)
2
2
Collecting the above results, we arrive at the final simplified entropy as s = ln 2 − αβ2 .
Appendix C: Derivation of simplified mesage passing equations for the approximate
Hopfield model
For the approximate Hopfield model, we similarly define the auxiliary quantity µb→i (ξi )
as
X
µb→i (ξi ) =
exp
{ξj |j∈∂b\i}
≃
Z
=q
z 2 /2
Dz exp
1
β̃ T b 2
(ξ σ )
2N
!
Y
Pj→b(ξj )
j∈∂b\i
!
p
ξi σib i2
β̃ h
G̃b→i + Cb→i z + √
2
N
!
β̃Fb→i 1
exp
+ G̃2b→i
exp
2
N
1 − β̃Cb→i
(C1)
β̃ G̃b→i ξi σib Fb→i
√
N
!
,
where Dz ≡ e √2πdz . Pi→a (ξi ) is the same as that in RBM. Using Eq. (C1), the cavity
magnetization mi→a can thus be derived as Eq. (22).
21
Appendix D: Replica computation for the RBM model
true
γ
a
a
We first define ua = ξ √Nσ , and v γa = ξ√σN . Both ua and v γa are random variables
subject to the covariance structure: hui = 0, hu2 i = 1, hv γ i = 0, h(v γ )2 i = 1, huv γ i = q γ ,
′
′
v γ v γ = r γγ , where we have dropped off the data index a because of independence among
data samples,
defined the overlap between true feature vector and the estimated one
P and
′
γ true
1
γ
as q = N i ξi ξi , and the overlap between two estimated feature vectors as r γγ =
P γ γ′
1
γ
γγ ′
= r, after introducing
i ξi ξi . Under the replica symmetric assumption, q = q and r
N
γ
γγ ′
n
the definition of q (and r ) as a delta function, hZ i can be estimated as
Z
Z
Z
√
dqdq̂
drdr̂
n(n − 1)
r̂
n
n
hZ i =
exp −Nnq q̂ − Nrr̂
− Nn + N ln Dz(2 cosh(q̂ + r̂z))
2πi/N
2πi/N
2
2
Z
Z
p
n
−β 2 /2
2
,
Dy Dt cosh βt(cosh β(qt + r − q y))
× exp αN ln e
(D1)
p
where we have written u = t, v γ = qt + 1 − rxγ + r − q 2 y (t, xγ and y are standard
Gaussian random variables). Finally, we arrived at the following free energy function:
Z
√
r̂(r − 1) αβ 2
−βfRS = −q q̂ +
+
(1 − r) + Dz ln 2 cosh(q̂ + r̂z)
2
2
Z
Z
(D2)
p
−β 2 /2
2
Dy Dt cosh βt ln cosh β(qt + r − q y).
+αe
√
The saddle-point equation for the order parameters {q, q̂, r, r̂} can be derived from
RS )
0, ∂(−βf
∂r
∂(−βfRS )
∂ q̂
∂(−βfRS )
∂q
=
∂(−βfRS )
∂ r̂
= 0,
= 0, and
= 0.
∂
) hln Zi (β ′ , β)|β ′=β , where
Note that to derive the entropy formula, we used s = (1 − β ∂β
β ′ is the inverse temperature at which the data is generated, and β the temperature at which
the Bayesian inference is carried out.
Appendix E: Replica computation for the Hopfield model
Q
γ a
βξ
√σ
N
Q
β̃
2
γ σa
ξ√
N
2
For the approximate Hopfield model, we replace a,γ cosh
with a,γ e
in Eq. (14). The subsequent calculation proceeds similarly to the appendix D. Analogously,
we have
Z
Z
Z
√
drdr̂
n(n − 1)
r̂
dqdq̂
n
n
exp −Nnq q̂ − Nrr̂
− Nn + N ln Dz(2 cosh(q̂ + r̂z))
hZ i =
2πi/N
2πi/N
2
2
√
Z
Z
2
2
β̃(qt+ r−q y)
n
1
2
.
× exp αN ln e−β /2 Dy Dt cosh βt q
e 2(1−β̃(1−r))
1 − β̃(1 − r)
(E1)
Using the replica trick defined in Eq. (13), we obtain the free energy function
Z
√
α
αβ̃(r + β̃q 2 )
r̂
, (E2)
−β̃fRS = −q q̂+ (r−1)+ Dz ln 2 cosh(q̂+ r̂z)− ln(1−β̃(1−r))+
2
2
2(1 − β̃(1 − r))
22
p
R
R
where we used the identity Dy Dt[qt + r − q 2 y]2 cosh βt = eβ̃/2 (β̃q 2 + r). The saddlepoint equations can be derived similarly.
Appendix F: Generalization to the case of the hidden neuron with an external field
in RBM
The sMP for RBM can be easily generalized to take into account external fields of visible
neurons and the hidden neuron. Here, for simplicity, we consider only the case of hidden
neuron with external field. The external field has binary values as well, defined by Bh φ0
(Bh = ±1). The only modification to the factor graph in Fig. 1 is to add one additional
variable node named by λ for the unknown external field. The additional node λ is connected
to all data nodes. Following the similar procedure as in Appendix A, we obtain the following
four kinds of messages:
X
mλ→a = tanh
ub→λ ,
(F1a)
b∈∂λ\a
−1
ub→λ = tanh
mi→a
(tanh(βGb→λ ) tanh(βφ0 )) ,
X
= tanh
ub→i ,
(F1b)
(F1c)
b∈∂i\a
ub→i
1
= ln
2
1+mλ→b
2
1+mλ→b
2
√
cosh β(Gb→i + σib / N + φ0 ) +
√
cosh β(Gb→i − σib / N + φ0 ) +
1−mλ→b
2
1−mλ→b
2
√
cosh β(Gb→i + σib / N − φ0 )
√
.
cosh β(Gb→i − σib / N − φ0 )
(F1d)
Once φ0 = 0, the original sMP for the RBM in the main text is recovered. To derive sMP for
the case of visible neurons with external fields, an additional central limit theorem applies to
the interacting external fields ({Bi φ0 }), which leads to introducing a joint cavity probability
Pi→a (ξi , Bi ) as well as µb→i (ξi , Bi ) where Bi = ±1.
Acknowledgments
I am very grateful to Taro Toyoizumi, Lukasz Kusmierz, Alireza Goudarzi and Roberto
Legaspi for attending a series of lectures about this work and their useful feedback. I thank
Lukasz Kusmierz for a careful reading of the manuscript and his useful feedback. This
work was supported by the program for Brain Mapping by Integrated Neurotechnologies for
Disease Studies (Brain/MINDS) from Japan Agency for Medical Research and development,
AMED.
[1] Brenden M. Lake, Ruslan Salakhutdinov, and Joshua B. Tenenbaum. Human-level concept
learning through probabilistic program induction. Science, 350:1332–1338, 2015.
[2] Daniel Kersten, Pascal Mamassian, and Alan Yuille. Object Perception as Bayesian Inference.
Annu. Rev. Psychol., 55:271–304, 2004.
23
[3] Geoffrey E. Hinton. Learning multiple layers of representation. Trends in Cognitive Sciences,
11:428–434, 2007.
[4] Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 35:1798–1828,
2013.
[5] Haiping Huang and Taro Toyoizumi. Unsupervised feature learning from finite data by message
passing: Discontinuous versus continuous phase transition. Phys. Rev. E, 94:062310, 2016.
[6] M. Mézard, G. Parisi, and M. A. Virasoro. Spin Glass Theory and Beyond. World Scientific,
Singapore, 1987.
[7] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural
networks. Science, 313(5786):504–507, 2006.
[8] C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight Uncertainty in Neural
Networks. arXiv: 1505.05424, 2015. In Proceedings of the 32nd International Conference on
Machine Learning (ICML 2015).
[9] H. Nishimori. Statistical Physics of Spin Glasses and Information Processing: An Introduction.
Oxford University Press, Oxford, 2001.
[10] M. Mézard and A. Montanari. Information, Physics, and Computation. Oxford University
Press, Oxford, 2009.
[11] Haiping Huang and Taro Toyoizumi. Advanced mean-field theory of the restricted boltzmann
machine. Phys. Rev. E, 91:050101, 2015.
[12] M. Mézard and G. Parisi. The bethe lattice spin glass revisited. Eur. Phys. J. B, 20:217,
2001.
[13] D. J. Thouless, P. W. Anderson, and R. G. Palmer. Solution of ’Solvable model of a spin
glass’. Philosophical Magazine, 35(3):593–601, 1977.
[14] David L. Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for
compressed sensing. Proceedings of the National Academy of Sciences, 106(45):18914–18919,
2009.
[15] Erwin Bolthausen. An Iterative Construction of Solutions of the TAP Equations for the
SherringtonKirkpatrick Model. Communications in Mathematical Physics, 325:333–366, 2014.
[16] Lenka Zdeborová and Florent Krzakala. Statistical physics of inference: thresholds and algorithms. Advances in Physics, 65:453–552, 2016.
[17] Marc Mézard. Mean-field message-passing equations in the hopfield model and its generalizations. Phys. Rev. E, 95:022117, 2017.
[18] D.J. Gross and M. Mezard. The simplest spin glass. Nuclear Physics B, 240:431–452, 1984.
[19] E. Gardner. Spin glasses with p-spin interactions. Nuclear Physics B, 257:747–765, 1985.
[20] G Hinton, S Osindero, and Y Teh. A fast learning algorithm for deep belief nets. Neural
Computation, 18:1527–1554, 2006.
[21] Tommaso Castellani and Andrea Cavagna. Spin-glass theory for pedestrians. Journal of
Statistical Mechanics: Theory and Experiment, 2005(05):P05012, 2005.
[22] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data
via the em algorithm. Journal of the royal statistical society. Series B, 39:1–38, 1977.
[23] Florent Krzakala, Marc Mezard, Francois Sausset, Yifan Sun, and Lenka Zdeborova.
Probabilistic reconstruction in compressed sensing: algorithms, phase diagrams, and
threshold achieving matrices. Journal of Statistical Mechanics: Theory and Experiment,
2012(08):P08009, 2012.
[24] Hidetoshi Nishimori and David Sherrington. Absence of replica symmetry breaking in a region
24
[25]
[26]
[27]
[28]
[29]
of the phase diagram of the Ising spin glass. AIP Conference Proceedings, 553:67–72, 2001.
Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86:2278–2324, 1998.
Elena Agliari, Adriano Barra, Andrea Galluzzi, Francesco Guerra, and Francesco Moauro.
Multitasking associative networks. Phys. Rev. Lett., 109:268101, 2012.
Daniel J. Amit, Hanoch Gutfreund, and H. Sompolinsky. Storing infinite numbers of patterns
in a spin-glass model of neural networks. Phys. Rev. Lett., 55:1530–1533, 1985.
Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training
deep neural networks with binary weights during propagations. In C. Cortes, N. D. Lawrence,
D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing
Systems 28, pages 3105–3113. Curran Associates, Inc., 2015.
J. S. Yedidia, W. T. Freeman, and Y. Weiss. Constructing free energy approximations and
generalized belief propagation algorithms. IEEE Trans Inf Theory, 51:2282–2312, 2005.
| 9cs.NE
|
Khalid Raza / Indian Journal of Computer Science and Engineering
Vol 1 No 2, 114-118
APPLICATION OF DATA MINING IN BIOINFORMATICS
KHALID RAZA
Centre for Theoretical Physics,
Jamia Millia Islamia, New Delhi-110025, India
Abstract
This article highlights some of the basic concepts of bioinformatics and data mining. The major research areas of bioinformatics are highlighted.
The application of data mining in the domain of bioinformatics is explained. It also highlights some of the current challenges and opportunities of
data mining in bioinformatics.
Keywords: Data Mining, Bioinformatics, Protein Sequences Analysis, Bioinformatics Tools.
1. Introduction
In recent years, rapid developments in genomics and proteomics have generated a large amount of biological data.
Drawing conclusions from these data requires sophisticated computational analyses. Bioinformatics, or
computational biology, is the interdisciplinary science of interpreting biological data using information technology
and computer science. The importance of this new field of inquiry will grow as we continue to generate and
integrate large quantities of genomic, proteomic, and other data.
A particular active area of research in bioinformatics is the application and development of data mining
techniques to solve biological problems. Analyzing large biological data sets requires making sense of the data by
inferring structure or generalizations from the data. Examples of this type of analysis include protein structure
prediction, gene classification, cancer classification based on microarray data, clustering of gene expression data,
statistical modeling of protein-protein interaction, etc. Therefore, we see a great potential to increase the interaction
between data mining and bioinformatics.
2. Bioinformatics
The term bioinformatics was coined by Paulien Hogeweg in 1979 for the study of informatic processes in biotic
systems. It was primary used since late 1980s has been in genomics and genetics, particularly in those areas of
genomics involving large-scale DNA sequencing.
Bioinformatics can be defined as the application of computer technology to the management of biological
information. Bioinformatics is the science of storing, extracting, organizing, analyzing, interpreting and utilizing
information from biological sequences and molecules. It has been mainly fueled by advances in DNA sequencing
and mapping techniques. Over the past few decades rapid developments in genomic and other molecular research
technologies and developments in information technologies have combined to produce a tremendous amount of
information related to molecular biology. The primary goal of bioinformatics is to increase the understanding of
biological processes.
Some of the grand area of research in bioinformatics includes:
2.1. Sequence analysis
Sequence analysis is the most primitive operation in computational biology. This operation consists of finding which
part of the biological sequences are alike and which part differs during medical analysis and genome mapping
processes. The sequence analysis implies subjecting a DNA or peptide sequence to sequence alignment, sequence
databases, repeated sequence searches, or other bioinformatics methods on a computer.
ISSN : 0976-5166
114
Khalid Raza / Indian Journal of Computer Science and Engineering
Vol 1 No 2, 114-118
2.2. Genome annotation
In the context of genomics, annotation is the process of marking the genes and other biological features in a DNA
sequence. The first genome annotation software system was designed in 1995 by Dr. Owen White.
2.3. Analysis of gene expression
The expression of many genes can be determined by measuring mRNA levels with various techniques such as
microarrays, expressed cDNA sequence tag (EST) sequencing, serial analysis of gene expression (SAGE) tag
sequencing, massively parallel signature sequencing (MPSS), or various applications of multiplexed in-situ
hybridization etc. All of these techniques are extremely noise-prone and subject to bias in the biological
measurement. Here the major research area involves developing statistical tools to separate signal from noise in
high-throughput gene expression studies.
2.4. Analysis of protein expression
Gene expression is measured in many ways including mRNA and protein expression, however protein expression is
one of the best clues of actual gene activity since proteins are usually final catalysts of cell activity. Protein
microarrays and high throughput (HT) mass spectrometry (MS) can provide a snapshot of the proteins present in a
biological sample. Bioinformatics is very much involved in making sense of protein microarray and HT MS data.
2.5. Analysis of mutations in cancer
In cancer, the genomes of affected cells are rearranged in complex or even unpredictable ways. Massive sequencing
efforts are used to identify previously unknown point mutations in a variety of genes in cancer. Bioinformaticians
continue to produce specialized automated systems to manage the sheer volume of sequence data produced, and they
create new algorithms and software to compare the sequencing results to the growing collection of human genome
sequences and germline polymorphisms. New physical detection technologies are employed, such as oligonucleotide
microarrays to identify chromosomal gains and losses and single-nucleotide polymorphism arrays to detect known
point mutations. Another type of data that requires novel informatics development is the analysis of lesions found to
be recurrent among many tumors.
2.6. Protein structure prediction
The amino acid sequence of a protein (so-called, primary structure) can be easily determined from the sequence on
the gene that codes for it. In most of the cases, this primary structure uniquely determines a structure in its native
environment. Knowledge of this structure is vital in understanding the function of the protein. For lack of better
terms, structural information is usually classified as secondary, tertiary and quaternary structure. Protein structure
prediction is one of the most important for drug design and the design of novel enzymes. A general solution to such
predictions remains an open problem for the researchers.
2.7. Comparative genomics
Comparative genomics is the study of the relationship of genome structure and function across different biological
species. Gene finding is an important application of comparative genomics, as is discovery of new, non-coding
functional elements of the genome. Comparative genomics exploits both similarities and differences in the proteins,
RNA, and regulatory regions of different organisms. Computational approaches to genome comparison have
recently become a common research topic in computer science.
2.8. Modeling biological systems
Modeling biological systems is a significant task of systems biology and mathematical biology. Computational
systems biology aims to develop and use efficient algorithms, data structures, visualization and communication tools
for the integration of large quantities of biological data with the goal of computer modeling. It involves the use of
ISSN : 0976-5166
115
Khalid Raza / Indian Journal of Computer Science and Engineering
Vol 1 No 2, 114-118
computer simulations of biological systems, like cellular subsystems such as the networks of metabolites and
enzymes, signal transduction pathways and gene regulatory networks to both analyze and visualize the complex
connections of these cellular processes. Artificial life is an attempt to understand evolutionary processes via the
computer simulation of simple life forms
2.9. High-throughput image analysis
Computational technologies are used to accelerate or fully automate the processing, quantification and analysis of
large amounts of high-information-content biomedical images. Modern image analysis systems augment an
observer's ability to make measurements from a large or complex set of images. A fully developed analysis system
may completely replace the observer. Biomedical imaging is becoming more important for both diagnostics and
research. Some of the examples of research in this area are: clinical image analysis and visualization, inferring clone
overlaps in DNA mapping, Bioimage informatics, etc.
2.10. Protein-protein docking
In the last two decades, tens of thousands of protein three-dimensional structures have been determined by X-ray
crystallography and Protein nuclear magnetic resonance spectroscopy (protein NMR). One central question for the
biological scientist is whether it is practical to predict possible protein-protein interactions only based on these 3D
shapes, without doing protein-protein interaction experiments. A variety of methods have been developed to tackle
the Protein-protein docking problem, though it seems that there is still much work to be done in this field.
3. Bioinformatics Tools
Following are the some of the important tools for bioinformatics (Table 1)
Table 1: Some of the tools for bioinformatics
Bioinformatics Research Area
Sequence alignment
Multiple sequence alignment
Gene Finding
Protein Domain Analysis
Pattern Identification
Genomic Analysis
Motif finding
Tool (Application)
BLAST
CS-BLAST
HMMER
FASTA
MSAProbs
DNA Alignment
MultAlin
DiAlign
GenScan
GenomeScan
GeneMark
Pfam
BLOCKS
ProDom
Gibbs Sampler
AlignACE
MEME
SLAM
Multiz
MEME/MAST
eMOTIF
References
http://blast.ncbi.nlm.nih.gov/Blast.cgi
ftp://toolkit.lmb.uni-muenchen.de/csblast/
http://hmmer.janelia.org/
www.ebi.ac.uk/fasta33
http://msaprobs.sourceforge.net/
http://www.fluxus-engineering.com/align.htm
http://multalin.toulouse.inra.fr/multalin/multalin.html
http://bibiserv.techfak.uni-bielefeld.de/dialign/
genes.mit.edu/GENSCAN.html
http://genes.mit.edu/genomescan.html
http://exon.biology.gatech.edu/
http://pfam.sanger.ac.uk/
http://blocks.fhcrc.org/
http://prodom.prabi.fr/prodom/current/html/home.php
http://bayesweb.wadsworth.org/gibbs/gibbs.html
http://atlas.med.harvard.edu/
http://meme.sdsc.edu/
http://bio.math.berkeley.edu/slam/
http://www.bx.psu.edu/miller_lab/
http://meme.sdsc.edu
http://motif.stanford.edu
4. Data Mining
Data mining refers to extracting or “mining” knowledge from large amounts of data. Data Mining (DM) is the
science of finding new interesting patterns and relationship in huge amount of data. It is defined as “the process of
ISSN : 0976-5166
116
Khalid Raza / Indian Journal of Computer Science and Engineering
Vol 1 No 2, 114-118
discovering meaningful new correlations, patterns, and trends by digging into large amounts of data stored in
warehouses”. Data mining is also sometimes called Knowledge Discovery in Databases (KDD). Data mining is not
specific to any industry. It requires intelligent technologies and the willingness to explore the possibility of hidden
knowledge that resides in the data.
Data Mining approaches seem ideally suited for Bioinformatics, since it is data-rich, but lacks a comprehensive
theory of life’s organization at the molecular level. The extensive databases of biological information create both
challenges and opportunities for development of novel KDD methods. Mining biological data helps to extract useful
knowledge from massive datasets gathered in biology, and in other related life sciences areas such as medicine and
neuroscience.
4.1. Data mining tasks
The two "high-level" primary goals of data mining, in practice, are prediction and description. The main tasks wellsuited for data mining, all of which involves mining meaningful new patterns from the data, are:
Classification: Classification is learning a function that maps (classifies) a data item into one of several predefined
classes.
Estimation: Given some input data, coming up with a value for some unknown continuous variable.
Prediction: Same as classification & estimation except that the records are classified according to some future
behaviour or estimated future value).
Association rules: Determining which things go together, also called dependency modeling.
Clustering: Segmenting a population into a number of subgroups or clusters.
Description & visualization: Representing the data using visualization techniques.
Learning from data falls into two categories: directed (“supervised”) and undirected (“unsupervised”) learning.
The first three tasks – classification, estimation and prediction – are examples of supervised learning. The next three
tasks – association rules, clustering and description & visualization – are examples of unsupervised learning. In
unsupervised learning, no variable is singled out as the target; the goal is to establish some relationship among all
the variables. Unsupervised learning attempts to find patterns without the use of a particular target field.
The development of new data mining and knowledge discovery tools is a subject of active research. One
motivation behind the development of these tools is their potential application in modern biology.
5. Application of Data Mining in Bioinformatics
Applications of data mining to bioinformatics include gene finding, protein function domain detection, function
motif detection, protein function inference, disease diagnosis, disease prognosis, disease treatment optimization,
protein and gene interaction network reconstruction, data cleansing, and protein sub-cellular location prediction.
For example, microarray technologies are used to predict a patient’s outcome. On the basis of patients’ genotypic
microarray data, their survival time and risk of tumor metastasis or recurrence can be estimated. Machine learning
can be used for peptide identification through mass spectroscopy. Correlation among fragment ions in a tandem
mass spectrum is crucial in reducing stochastic mismatches for peptide identification by database searching. An
efficient scoring algorithm that considers the correlative information in a tunable and comprehensive manner is
highly desirable.
6. Conclusion and challenges
Bioinformatics and data mining are developing as interdisciplinary science. Data mining approaches seem ideally
suited for bioinformatics, since bioinformatics is data-rich but lacks a comprehensive theory of life’s organization at
the molecular level.
However, data mining in bioinformatics is hampered by many facets of biological databases, including their size,
number, diversity and the lack of a standard ontology to aid the querying of them as well as the heterogeneous data
of the quality and provenance information they contain. Another problem is the range of levels the domains of
expertise present amongst potential users, so it can be difficult for the database curators to provide access
mechanism appropriate to all. The integration of biological databases is also a problem. Data mining and
ISSN : 0976-5166
117
Khalid Raza / Indian Journal of Computer Science and Engineering
Vol 1 No 2, 114-118
bioinformatics are fast growing research area today. It is important to examine what are the important research
issues in bioinformatics and develop new data mining methods for scalable and effective analysis.
7. References
[1] Zaki , J.; Wang , T.L. and Toivonen, T.T. (2001). BIOKDD01: Workshop on Data Mining in Bioinformatics”.
[2] Li, J.; Wong, L. and Yang, Q. (2005 ). Data Mining in Bioinformatics, IEEE Intelligent System, IEEE Computer Society.
[3] Liu, H.; Li, J. and Wong, L. (2005). Use of Extreme Patient Samples for Outcome Prediction from Gene Expression Data,
Bioinformatics, vol. 21, no. 16, pp. 3377–3384
[4] Yang, Qiang. Data Mining and Bioinformatics: Some Challenges, http://www.cse.ust.hk/~qyang
[5] Berson, Alex, Smith, Stephen and Threaling, Kurt, “Building Data Mining Application for CRM”, Tata McGraw Hill.
[6] Zhang, Yanqing; C., Jagath, Rajapakse, Machine Learning in Bioinformatics, Wiley, ISBN: 978-0-470-11662-3
[7] Kuonen, Diego. Challenges in Bioinformatics for Statistical Data Miner, Bulletin of the Swiss Statistical Society, 46; 10-17.
[8] Richard, R.J. A. and Sriraam, N. (2005). A Feasibility Study of Challenges and Opportunities in Computational Biology: A
Malaysian Perspective, American Journal of Applied Sciences 2 (9): 1296-1300.
[9] Tang, Haixu and Kim, Sun. Bioinformatics: mining the massive data from high throughput genomics experiments, analysis
of biological data: a soft computing approach, edited by Sanghamitra Bandyopadhyay, Indian Statistical Institute, India
[10] Nayeem, Akbar; Sitkoff, Doree, and Krystek, Jr., Stanley. (2006) A comparative study of available software for highaccuracy homology modeling: From sequence alignments to structural models, Protein Sci. April; 15(4): 808–824
[11] N., Cristianini and M., Hahn. (2006) Introduction to Computational Genomics, Cambridge University Press. ISBN 0-52167191-4.
[12] SJ, Wodak and Janin, J. (1978). Computer Analysis of Protein-Protein Interactions. Journal of Molecular Biology 124 (2):
323–42.
[13] Mewes, H.W.; Frishman, D.; X.Mayer, K. F.; Munsterkotter, M., Noubibou , O.; Pagel, P. and Rattei, T. (2006) Nucleic
Acids Research, 34, D169.
[14] Lee, Kyoungrim. (2008). Computational Study for Protein-Protein Docking Using Global Optimization and Empirical
Potentials, Int. J. Mol. Sci. 9, 65-77.
[15] Hirschman, Lynette; C. Park, Jong; T., Junichi, Wong, L. and H. Wu., Cathy (2002). Accomplishments and challenges in
literature data mining for biology, BIOINFORMATICS REVIEW, Vol. 18 no. 12, 1553–1561
[16] Luis, T.; Chitta; B. and Kim, S. (2008). Fuzzy c-means clustering with prior biological knowledge, Journal of Biomedical
Informatics.
[17] Han and Kamber (2006). Data Mining concepts and techniques, Morgan Kaufmann Publishers.
[18] Hand, D. J.; Mannila, H. and Smyth, P. Principles of Data Mining, MIT Press.
[19] Aluru, S., ed. (2006). Handbook of Computational Molecular Biology. Chapman & Hall/Crc,
[20] Baxevanis, A.D.; Petsko, G.A.; Stein, L.D. and Stormo, G.D., eds. (2007). Current Protocols in Bioinformatics. Wiley.
[21] Mount, D. W. (2002). Bioinformatics: Sequence and Genome Analysis Spring Harbor Press.
[22] Gilbert, D. (2004). Bioinformatics software resources. Briefings in Bioinformatics, Briefings in Bioinformatics.
[23] Jiong, Lei Liu; Yang, A. and Tung, K. H. Data Mining Techniques for Microarray Datasets, Proceedings of the 21st
International Conference on Data Engineering (ICDE 2005).
[24] Pevzner, P. A. (200). Computational Molecular Biology: An Algorithmic Approach The MIT Press.
[25] Soinov, L. (2006). Bioinformatics and Pattern Recognition Come Together. Journal of Pattern Recognition Research
(JPRR), Vol 1 (1) p.37-41
ISSN : 0976-5166
118
| 5cs.CE
|
arXiv:1803.04639v1 [cs.IT] 13 Mar 2018
Correct Undetected Errors with List Decoding in
ARQ Error-control Systems
Jingzhao Wang
Yuan Luo
Department of Computer Science and
Engineering
Shanghai Jiao Tong University
Shanghai 200240, China
Email: [email protected]
Department of Computer Science and
Engineering
Shanghai Jiao Tong University
Shanghai 200240, China
Email: [email protected]
Abstract—Undetected errors are important for linear codes,
which are the only type of errors after hard decision and
automatic-repeat-request (ARQ), but do not receive much attention on their correction. In concatenated channel coding, suboptimal source coding and joint source-channel coding, constrains
among successive codewords may be utilized to improve decoding
performance. In this paper, list decoding is used to correct the
undetected errors. The benefit proportion of the correction is
obviously improved especially on Hamming codes and ReedMuller codes, which achieves about 40% in some cases. But
this improvement is significant only after the selection of final
codewords from the lists based on the constrains among the
successive transmitted codewords. The selection algorithm is
investigated here to complete the list decoding program in the
application of Markov context model. The performance of the
algorithm is analysed and a lower bound of the correctly selected
probability is derived to determine the proper context length.
I. I NTRODUCTION
Error detection using block codes combined with repeat
requests is a widely used method of error control in communication systems[6], which is called automatic-repeat-request
(ARQ) scheme.
Undetected errors are the only type of errors after hard
decision and automatic-repeat-request (ARQ), where the sent
codeword is transmitted into another codeword. List decoding
was introduced independently by Elias [1] and Wozencraft
[11] in the late 1950’s. For a received vector, the decoder
doesn’t output a single codeword, but outputs a list of possible
codewords. It is considered to be list decoded correctly if the
sent codeword is in the list. This paper uses list decoding
to correct undetected errors of ARQ systems and give a
list of possible codewords for each received codeword after
retransmissions.
In some cases, there may be constraints among successive codewords, i.e. the input codewords are dependent. For
example, when concatenated codes [2] are used in channel
coding, the codewords of the inner code must be constrained
by the outer code. In joint source-channel coding, there also
be memory structure inherent with in the sequence output
from the source [12]. Moreover, in the model for storage or
communication, many files are not compressed or compressed
sub-optimally in source coding [4],[5],[10]. In this situation,
there may also be constraints among the successive transmitted
codewords , where redundancy mining can be used for error
correction, which is considered in [4],[5]. Inspired by these
works, we assume the code context model is a Markov chain
which is an elementary stochastic model used in natural
language processing models, and investigate an algorithm to
select codewords from the lists.
This paper is organized as follows. Section II presents
some necessary preliminaries and notations. In Section III
the error correcting probability by list decoding of linear
codes is analysed. And Hamming codes and Reed-Muller
codes are investigated as numerical examples where the benefit
proportion of list decoding can achieve about 40% in some
cases. In Section IV, an algorithm based on context is investigated to complete list decoding program, with which the most
probable codewords are selected from the lists. The algorithm
performance is analysed in Section V. Finally, we conclude
this paper in Section VI.
II. P RELIMINARIES AND N OTATIONS
The communication channel used in this paper is binary
symmetric channel (BSC) with crossover probability p. The
encoder is a one-to-one map function from the message set
M = {0, 1, ..., M − 1} to the codebook C, where C is a
subset of {0, 1}n and n is known as the blocklength of code
C .
wH (x) is the Hamming weight of a vector x and dH (x, y)
is the Hamming distance between two vectors x, y ∈ {0, 1}n .
For a binary [n, k, d] linear code C with minimum distance
d = min{wH (x)|x ∈ C, x 6= 0}, the weight distribution is
defined as
w
Aw
i = Ai (C) = #{x ∈ C|wH (x) = i}
and the weight distribution function is defined as
Aw
C (x, y) =
n
X
n−i i
Aw
y.
i x
i=0
III. T HE ERROR PROBABILITY OF ARQ SYSTEMS WITH
LIST DECODING
In this section, we first analyse the error correcting probability by list decoding of linear codes, which can be calculated
by Proposition 1. Then the error probability of ARQ systems
is calculated in Proposition 2. In subsection III-A and III-B,
the benefit proportion of error probability after list decoding
is illustrated in Fig. 1 and Fig. 2, where Hamming codes and
Reed-Muller codes are investigated as numerical examples.
Assume X = (X1 , X2 , ..., Xn ) ∈ {0, 1}n is sent and Y =
(Y1 , Y2 , ..., Yn ) ∈ {0, 1}n is received, where X and Y are
vectors of random variables. Let B(y, r) = {z ∈ {0, 1}n :
y ∈ {0, 1}n , dH (z, y) ≤ r} denote the sphere of radius r
about y and consider the conditional probability as follows:
Pr{Y ∈ C and Y 6= 0n , 0n ∈ B(Y, r)|X = 0n }
= Pr{Y ∈ C and Y 6= 0n , Y ∈ B(0n , r)|X = 0n },
After list decoding in ARQ system, the Pue is reduced to
List
= Pue − P List and the error probability changes from
Pue
(3) to
PeARQ,List =
The benefit proportion is defined by
Pb =
PeARQ − PeARQ,List
PeARQ
.
(5)
The results (3), (4) and (5) form the following proposition.
Proposition 2. If the crossover probability of BSC is p, it is
w
clear that Pc = (1 − p)n , Pue = Aw
C (1 − p, p) − Pc = AC (1 −
n
p, p) − (1 − p) . The probability of undetected error after list
List
decoding is Pue
= Pue − P List . After ARQ retransmissions,
assume all the received vectors are codewords, then the error
probability is
Pue
Pue + Pc
Aw (1 − p, p) − (1 − p)n
,
= C
Aw
C (1 − p, p)
PeARQ =
c∈C,c6=0n
where r ≥ d. P List (r) is called the error correcting probability
by list decoding, which is also the reduced probability of
undetected error, see Proposition 1.
Remark. As is shown in [3], the list size should not be too
large and be at most a polynomial in the blocklength. So this
paper considers the cases that the decoding radius r equals to
d, the minimum distance of code C. Note that, the completing
of list decoding program will be provided in Section IV by
using context.
Proposition 1. Assume the channel is BSC with crossover
probability p. Then, the error correcting probability by list
decoding with radius r = d is
P List , P List (d)
X
=
P r{Y = c, c ∈ B(0n , d)|X = 0n }
=
(4)
(1)
where r ≥ d. 0n ∈ B(Y, r) and Y 6= 0n mean that an
undetected error occurs when X = 0n is sent and the correct
codeword 0n is in the list with decoding radius r (r ≥ d). In
this case, it is defined to be list decoded correctly.
For all the codewords in C, the probability in (1) can be
described as follows:
X
P List (r) =
Pr{Y = c, c ∈ B(0n , r)|X = 0n }, (2)
c∈C,c6=0n
d
Aw
d · p · (1
List
Pue
.
Pue + Pc
the error probability after list decoding is
List
Pue
Pue + Pc
d
n−d
Aw (1 − p, p) − (1 − p)n − Aw
d · p · (1 − p)
,
= C
w
AC (1 − p, p)
PeARQ,List =
and the benefit proportion is
Pb =
PeARQ − PeARQ,List
PeARQ
Aw · pd · (1 − p)n−d
= wd
.
AC (1 − p, p) − (1 − p)n
Two numerical examples of Proportion 2 in ARQ systems
are given in Subsection III-A and III-B, respectively.
− p)n−d .
A. Hamming Codes
The list size is L = Aw
d + 1.
For ordinary linear codes, the computation of Aw
i (0 ≤ i ≤
n) is an NP-hard problem. But for some special codes, the
weight distributions are known and can be used to calculate
P List .
In ARQ systems, the error probability is defined in [7], see
PeARQ =
Pue
,
Pue + Pc
(3)
where Pc denotes the probability that a received vector contains no error and Pue denotes the probability that a received
vector contains an undetected error pattern.
Next we illustrate the benefits of list decoding on the error
probability PeARQ of [n = 2m −1, n−m] Hamming code. The
weight distribution function of Hamming code is known [8]
and the corresponding list size in Proposition 1 is L = Aw
3 +1.
ARQ
ARQ,List
The parameters of Proposition 2, i.e. Pe
, Pe
and
Pb , are illustrated in Fig. 1, where the messages are transmitted
through BSC with crossover probability p. As is shown by
the figure, PeARQ increases with block length. For Hamming
codes of m = 4, 5, 6 (p = 0.1) or m = 3, 4, 5 (p = 0.2), the
benefit of list decoding is obvious. Especially for m = 5 (p =
0.1) and m = 4 (p = 0.2), PeARQ is large and the benefit
proportion Pb is about 40%, a meaningful amount.
Assume the process is a Markov chain and let Pij denote
the one-step transition probability from state i to state j, then
Pij = Pr{Xt+1 = j|Xt = i},
(6)
for all states i, j ∈ M and all t ≥ 0. Let P denote the the
matrix of Pij , then
P = (Pij )M ×M (0 ≤ i, j ≤ M − 1).
Fig. 1. The Numerical Results of Hamming Codes
B. Reed-Muller Codes
As is shown in [8], the weight distribution of RM codes
is partly known. In most cases, the higher the code rate is,
the higher the undetected error probability is, and correcting
undetected errors is significant when the probability is high.
Let RM (r, m) denote the rth order binary Reed-Muller code
with block length n = 2m .
The numerical results of PeARQ , PeARQ,List and Pb in
Proposition 2 are illustrated in Fig. 2, where we can see
that the benefit proportion is decreasing with the crossover
probability p. And for Reed-Muller codes of high rate, the
benefit of list decoding is evident for small p, which is
supported by Fig. 2 (a) and (d)-(f). When PeARQ is less than
0.4, the benefit proportion is more than 40% shown in Fig. 2
(a)-(e).
IV. L IST- DECODING BASED ON C ONTEXT
In this section, an algorithm based on the context is investigated and dubbed context list decoding algorithm (CLDA).
The code context is described with Markov chain satisfying
equation (6). The performance of CLDA depends on the
context structure. The process of CLDA is shown in Algorithm
1 and Example 1, with which the most probable codewords
are selected from the lists.
The appearance of a codeword may be affected by former
or latter codewords, which is called the context of a code.
For instance, the appearance of each English letter in a word
depends on its former and latter letters. At the same time,
an English word is also affected by its adjacent words in
a sentence. In natural language processing models, Markov
chain is primary. Therefore, we assume the code context model
is Markov.
Then the context of all the received codewords can be seen
as a stochastic process {Xt , t = 0, 1, 2, ...} that takes values
on M = {0, 1, ..., M − 1}. Each codeword is bijectively
mapped to a state of the process. After retransmissions in
ARQ systems, all the received vectors are codewords. So
in the following part, we use the state in M to represent
the corresponding codeword and give a list of all possible
codewords for each received codeword. The proper ones are
selected from the lists based on the context. The selection of
codewords from lists is equivalent to the determination of the
states, see the following part.
The context of a code can be represented by P, with which
Algorithm 1 (CLDA) will select the most probable queue of
codewords. To describe CLDA, some definitions are given as
follows.
Definition 1. A queue of successive codewords with context
constraint is called a codeword sentence. If the queue length
is N , it is said to be an N -codeword sentence. The appearance probability of a codeword sentence is dubbed codeword
sentence weight.
Then, the weight of N -codeword sentence (i0 , i1 , ..., iN −1 )
is
Pr{X0 = i0 , ..., XN −1 = iN −1 }
= Pr{i0 } · Pr{i1 |i0 } · ... · Pr{iN −2 |iN −1 }
= πi0 · Pi1 ,i0 · ... · PiN −2 ,iN −1 ,
(7)
where let Pr{i0 } , Pr{X0 = i0 } , πi0 denote the priori probability of i0 and let Pr{ik+1 |ik } , Pr{Xk+1 = ik+1 |Xk =
ik }.
For linear codes, every given list contains the same number
of codewords and L is the list size. Let r = (r0 , r1 , ..., rN −1 )
denote the received codeword sentence and N is the codeword
sentence length. Let l = (l0 , l1 , ..., lN −1 ) denote the codeword
sentence selected from the successive lists, then the problem
of finding the most probable N -codeword sentence in the lists
is equivalent to compute
arg max
Pr{(l0 , l1 , ..., lN −1 )},
(l0 ,l1 ,...,lN −1 )∈L0 ×L1 ×...×LN −1
where L0 , L1 , ..., LN −1 are successive lists given by r as
follows:
Li = L(ri ), 0 ≤ i ≤ N − 1,
L(ri ) = {c ∈ C : dH (c, ri ) ≤ d}.
(8)
Let function T (li ) compute the maximal codeword sentence
weight when some codeword li is selected form List Li
and appended to the previously selected codeword sentence
(l0 , l1 , ..., li−1 ). For i ∈ [2, N − 1], we have
T (li ) ,
max
(l0 ,l1 ,...,li−1 )∈L0 ×L1 ×...×Li−1
Pr{(l0 , l1 , ..., li )}.
Then
T (li ) =
=
=
max
(l0 ,l1 ,...,li−1 )∈
L0 ×L1 ×...×Li−1
max
li−1 ∈Li−1
max
li−1 ∈Li−1
Pr{(l0 , l1 , ..., li−1 )}Pr{li |li−1 }
Pr{li |li−1 } ·
max
(l0 ,l1 ,...,li−2 )∈
L0 ×L1 ×...×Li−2
Pr{li |li−1 }T (li−1 )
Pr{(l0 , l1 , ..., li−1 )}
(a) RM(2,4)
(b) RM(2,5)
(c) RM(3,5)
(d) RM(4,6)
(e) RM(5,7)
(f) RM(6,8)
Fig. 2.
The Numerical Results of Reed-Muller Codes
and T (l0 ) = Pr{l0 } = πl0 .
The derived recurrence suggests that the optimization problem can be solved with dynamic programming algorithm
specified in Algorithm 1 (CLDA), which is adapted from the
Viterbi decoding [9]. The final solution is computed iteratively,
starting from T (l1 ) according to the recurrence. When the
last iteration is finished, we trace back along the path with
the maximal codeword sentence weight, selecting the most
probable codeword sentence.
Let L = (lij )L×N (0 ≤ i ≤ L − 1, 0 ≤ j ≤ N − 1) store the
values of each node. Let P = (Pij )M ×M (0 ≤ i, j ≤ M − 1)
denote the codeword transition probability matrix and
Π = (π0 , π1 , ..., πM −1 )T denote the priori probability of
each codeword, which are decided by the context of the
codewords. The CLDA is as follows:
Algorithm 1(CLDA).
In the following four steps, first, P, L and Π defined above
can be used to calculate a weight matrix T and a jumping
point matrix D. Then, matrix R, which stores the jumping
points of the path with length N and the largest weight, is
determined by T and D. At last, the final codewords after
context list decoding are stored in F calculated by R and L.
1. Initialize matrix T = (tij )L×N where tij stores the
maximum codeword sentence weight from the nodes in List
0 to the ith node in the current List j. The first column
of T is initialized with (πl0,0 , πl1,0 , ..., πlL−1,0 )T , namely,
ti,0 = πli,0 (0 ≤ i ≤ L − 1) and others with zeroes.
2. From List 0 to List N −1, compute the codeword sentence
weight:
tij = max (tk,j−1 · Plk,j−1 ,lij ),
(9)
0≤k≤L−1
where 0 ≤ i ≤ L − 1, 1 ≤ j ≤ N − 1.
The corresponding k which achieves the maximum value is
stored in matrix D = (xij )L×N :
xij = arg
max (tk,j−1 · Plk,j−1 ,lij ),
0≤k≤L−1
(10)
where 0 ≤ i ≤ L − 1, 1 ≤ j ≤ N − 1.
Remark.
• k is the row (node) index and j is the column (list) index;
• lij is the value (codeword) of the ith node in List j;
• xij is the jumping point from List j −1 to the ith node in
List j, which reaches the largest weight to current node.
• we select one codeword randomly if there are more than
one achieving the maximum value.
3. Calculate R = (r0 , r1 , ..., rN −1 ), where rN −1 =
arg max ti,N −1 and rk−1 = xrk ,k (k is from N − 1 to
0≤i≤L−1
0).
4. Calculate F = (f0 , f1 , ..., fN −1 ), where fk = lrk ,k (k is
from 0 to N − 1). Return F.
The complexity of CLDA is O(L2 N ), where L is the list
size and N is the length of codeword sentence. An example
of CLDA is illustrated as follows.
Example 1. Let the codeword
be
1/3
P = 2/9
0
transition probability matrix P
2/3
0
5/9 2/9
2/3 1/3
Note that, P is from the context and has no relation with
BSC. Meanwhile, let the priori probability of each codeword
be Π = (1/5, 3/5, 1/5)T . There are 3 successive lists with list
size 3, which is stored in L.
0 0 0
L= 1 1 1
2 2 2
1. Initialize the first column of T with Π =
(1/5, 3/5, 1/5)T .
2. For each 2-codeword sentence from List 0 to List 1, (9)
and (10) are calculated as follows: π0 ∗P00 = 1/15, π1 ∗P10 =
2/15, π2 ∗ P20 = 0. Then, record the biggest value 2/15 in T
and corresponding jumping point x01 : t01 = 2/15, x01 = 1.
Similarly, calculate π0 ∗ P01 = 2/15, π1 ∗ P11 = 1/3, π2 ∗
P21 = 2/15. Then t11 = 1/3, x11 = 1.
And calculate π0 ∗ P02 = 0, π1 ∗ P12 = 2/15, π2 ∗ P22 =
1/15. Then t21 = 2/15, x21 = 1.
For each 3-codeword sentence from List 0 to List 3,
compute t01 ∗ P00 = 2/45, t11 ∗ P10 = 2/27, t21 ∗ P20 = 0,
and record the biggest value 2/27 and corresponding index 1.
Then t02 = 2/27, x02 = 1.
Similarly, t12 = 5/27, x12 = 1, t22 = 2/27, x22 = 1.
3. We can find the most probable sentence using T and A,
as shown in Algorithm 1. t12 is the maximum among t02 , t12
and t22 . Then r2 = 1 and r1 = x12 = 1, r0 = x11 = 1. Using
R = (1, 1, 1), we get F = (l10 , l11 , l12 ) = (1, 1, 1) which is
the final codeword sentence after context list decoding.
V. P ERFORMANCE OF CLDA
As the ARQ system transmits messages through BSC with
crossover probability p, the probability of retransmission each
time is
Pretrans = 1 − Aw
C (1 − p, p),
where we assume the sent codeword is (0, 0, ..., 0)n . In the
following paragraphs, α(α ≥ 0) denotes the number of
retransmissions for receiving a legal codeword.
Let s = (s0 , s1 , ..., sN −1 ) be the sent codeword sentence.
To guarantee that s can be selected from the N lists in the
decoding, it is obvious that Pr{s} should be the biggest among
all the sentences in L0 × L1 × ... × LN −1 . This constraint
depends not only on the codeword sentence but also on the
transition probability matrix P .
Assumption. Assume that Pr{s} is the biggest among all
the codeword sentence in L(s0 ) × L(s1 ) × ... × L(sN −1 ).
Then if received sentence r equals to s, the CLDA selects
s obviously. The following paragraphs try to consider the
situation that r 6= s.
Assume si 6= ri for some 0 ≤ i ≤ N − 1. If the list given
by si is the same with the list given by ri , s will be selected
from the lists correctly, otherwise s may be selected wrongly.
Let S(si ) denote the set of codewords which give the same
list with si , namely
S(si ) = {c ∈ C : L(c) = L(si )}.
(11)
We use Ec to denote the event that s is selected from the
lists correctly. Meanwhile, let Ec (i)(0 ≤ i ≤ N − 1) denote
the event that si is selected correctly from Li . Then a lower
bound of Pr{Ec } is as follows:
Pr{Ec } = Pr{
N\
−1
Ec (i)} =
i=0
≥
N
−1
Y
N
−1
Y
Pr{Ec (i)}
i=0
Pr{L(ri ) = L{si }}
i=0
=
N
−1 X
∞
Y
i=0
=
α
Pretrans
Pr{ri ∈ S(si )}
α=0
N
−1
Y
1
(1 − Pretrans
)N
(Pc + (|S(si )| − 1)pd (1 − p)n−d ). (12)
i=0
Note that Pr{ri ∈ S(si )} is calculated based on the definition
of L(·) and S(·), see (8) and (11).
Numerical results for (12) are shown in Fig. 3, where the
lower bound of PEc is evaluated with RM(3,5) code. In Fig. 3
(a), the lower bound of PEc decreases when the crossover
probability p of BSC varies from 0 to 0.1 for codeword
sentence length N = 10, 20. It can be seen that the bound is
better for small p, i.e. p ∈ (0, 0.06), which is partly show in
Fig. 3 (b). The lower bound is about 0.98 for p ∈ (0.04, 0.06)
which may be practical. Fig. 3 (c) shows that the lower bound
of PEc decreases when the codeword sentence N varying from
5 to 30 and p = 0.06. For small N , PEc may be relatively
large, however, the context may not work well in CLDA.
When p and PEc is fixed, the largest N can be calculated
out with (12), which is meaningful for the selection of the
proper codeword sentence length.
When p is fixed, the lower bound can be used to calculated
the proper N needed to achieve required correctly selected
probability PEc .
In the following part, we try to analyse the average correctly
selected probability denoted by Paverage .
The probability of all the M N codeword sentences can
be calculated out with P. It is convenient to order all the
codeword sentences by the probability from large to small.
Let si = (si0 , si1 , ..., si,N −1 )(1 ≤ i ≤ M N ) denote the
ordered codeword sentences, where Pr{si } is decreasing by
i from 1 to M N , namely, Pr{s1 } is largest and Pr{sM N } is
the smallest.
For each sent codeword sentence si , the probability that si is
selected correctly varies with Pr{si }. Consider si and Pr{si }
for different 1 ≤ i ≤ M N respectively as follows:
1. s1 and Pr{s1 }:
(a)
(b)
(c)
Fig. 3. The lower bound of correctly selected probability Pr{Ec }
Pr{s1 } is the largest among all the codeword sentences, so
if s1 in the given lists, it will be selected correctly.
Pr{s1 in the lists|s1 is sent}
1
=
(Pc + Ad · pd (1 − p)n−d )N
(1 − Pretrans )N
2. si and Pr{s1 }:
If sj (1 ≤ j ≤ i) in the lists, si will not be selected out and
errors happen. When si is sent, the given lists should contain
si and not contain sj (1 ≤ j ≤ i).
Pr{si is selected correctly|si is sent}
i−1
[
L(sk ) |si is sent}
= Pr{r ∈ L(si )\
k=1
=
1
·
(1 − Pretrans )N
N
−1
X
Y
r∈L(si )\(
Si−1
k=1
pdH (rk ,si,k ) (1 − p)n−dH (rk ,si,k ) ,
L(sk )) k=0
where L(si ) = {L(si0 ) × L(si1 ) × L(si,N −1 )} and L(ri ) =
{c ∈ C : dH (c, ri ) ≤ d}, defined in (19).
Then, the average correctly selected probability is
Paverage
=
N
M
X
(Pr{si is sent} · Pr{si is selected correctly|si is sent})
i=1
=
1
(1 − Pretrans )N
X
N
M
X
i=1
N
−1
Y
Pr{si }·
pdH (rk ,si,k ) (1 − p)n−dH (rk ,si,k ) .
S
k=0
r∈L(si )\( i−1
k=1 L(sk ))
S0
To be convenient, we assume k=1 L(sk ) = ∅.
Paverage can be calculated when the transition probability
matrix P is known. However, the calculation complexity is
high. But when P is sparse or with other special structures,
the evaluation maybe useful. Besides, when M and N are
relatively small, Paverage can be calculated and used to
construct possible codes with better performance.
VI. C ONCLUSION
In this paper, we use list decoding to correct undetected
errors. The benefit is illustrated with numerical results of
Hamming codes and Reed-Muller codes in ARQ systems,
where the benefit proportion with list decoding can achieve
about 40% in some cases. We assume the code context model
is a Markov chain and an algorithm based on context, named
CLDA, is investigated to select the most probable codewords
from lists. The performance of CLDA is analysed and a lower
bound on the probability that the CLDA selects the correct
codeword sentence is derived.
R EFERENCES
[1] P. Elias, “List decoding for noisy channels,” Technical Report 335,
Research Laboratory of Electronics, 1957.
[2] G. D. Forney and G. D. Forney, Concatenated codes. Citeseer, 1966,
vol. 11.
[3] V. Guruswami and M. Sudan, “List decoding of error-correcting codes,”
Ph.D. dissertation, Springer, 2001.
[4] A. Jiang, Y. Li, and J. Bruck, “Enhanced error correction via language
processing,” in Proc. Non-Volatile Memories Workshop (NMVW), 2015.
[5] A. Jiang, Y. Li, and J. Bruck, “Error correction through language
processing,” in Proc. IEEE Information Theory Workshop (ITW), 2015.
[6] T. Klove, Codes for Error Detection. River Edge, NJ, USA: World
Scientific Publishing Co., Inc., 2007.
[7] S. Lin and D. J. Costello, Error Control Coding, Second Edition. Upper
Saddle River, NJ, USA: Prentice-Hall, Inc., 2004.
[8] F. MacWilliams and N. Sloane, The Theory of Error-Correcting Codes,
2nd ed. North-holland Publishing Company, 1978.
[9] A. J. Viterbi, “Error bounds for convolutional codes and an asymptotically
optimum decoding algorithm,” Information Theory, IEEE Transactions
on, vol. 13, no. 2, pp. 260–269, 1967.
[10] T. Weissman, E. Ordentlich, G. Seroussi, S. Verdú, and M. J. Weinberger,
“Universal discrete denoising: Known channel,” Information Theory,
IEEE Transactions on, vol. 51, no. 1, pp. 5–28, 2005
[11] J. M. Wozencraft, “List decoding,” Quarterly Progress Report, Research
Laboratory of Electronics, vol. 48, pp. 90–95, Jan 1958.
[12] G.-C. Zhu and F. Alajaji, “Joint source-channel turbo coding for binary
markov sources,” IEEE Transactions on Wireless Communications, vol. 5,
no. 5, pp. 1065–1075, May 2006.
| 7cs.IT
|
1
Generalised Precoded Spatial Modulation for
Integrated Wireless Information and Power Transfer
arXiv:1705.11076v1 [cs.IT] 31 May 2017
Rong Zhang, Lie-Liang Yang and Lajos Hanzo
Communications, Signal Processing and Control, School of ECS, University of Southampton, SO17 1BJ, UK
Email: rz,lly,[email protected], http://www-mobile.ecs.soton.ac.uk
Abstract—Conventional wireless information transfer by modulating the amplitude, phase or frequency leads to an inevitable
Rate-Energy (RE) trade-off in the presence of simultaneous
Wireless Power Transfer (WPT). In echoing Varshney’s seminal
concept of jointly transmitting both information and energy,
we propose the so-called Generalised Precoded Spatial Modulation (GPSM) aided Integrated Wireless Information and Power
Transfer (IWIPT) concept employing a power-split receiver. The
principle of GPSM is that a particular subset of Receive Antennas
(RA) is activated and the activation pattern itself conveys useful
information. Hence, the novelty of our GPSM aided IWIPT
concept is that RA pattern-based information transfer is used in
addition to the conventional waveform-based information carried
by the classic M -ary PSK/QAM modulation. Following the Radio
Frequency (RF) to Direct Current (DC) power conversion invoked
for WPT at the power-split receiver, the non-coherent detector
simply compares the remaining received power accumulated by
each legitimate RA pattern for the sake of identifying the most
likely RA. This operation is then followed by down-conversion
and conventional Base Band (BB) M -ary PSK/QAM detection.
Both our analysis and simulations show that the RA patternbased information transfer represented in the Spatial Domain
(SD) exhibits a beneficial immunity to any potential powerconversion induced performance degradation and hence improves
the overall RE trade-off when additionally the waveform-based
information transfer is also taken into account. Moreover, we
investigate the impact of realistic imperfect Channel State Information at the Transmitter (CSIT) as well as that of the antenna
correlations encountered. Finally, the system’s asymptotic performance is characterised in the context of large-scale Multiple
Input Multiple Output (MIMO) systems.
I. I NTRODUCTION
1) Motivation: In thermal and statistical physics, the earliest and most famous thought experiment regarding information
and energy was conceived by Maxwell in 1867, which is
referred to as ”Maxwell’s Demon” [1], where the second
law of thermo-dynamics was hypothetically violated by the
bold hypothesis of information to energy conversion. This
stimulated further intriguing research in the mid-20th century
as to whether information processing itself dissipates energy,
which subsequently led to ”Landauer’s principle” [2], [3] suggesting that thermo-dynamically reversible manipulations of
information in computing, measurement and communications
do not necessarily dissipate energy, since no energy is required
to perform mathematical calculations. This principle has thus
The financial support of the EPSRC under the India-UK Advanced Technology Centre (IU-ATC), that of the EU under the Concerto project as well
as that of the European Research Council’s (ERC) Advanced Fellow Grant is
gratefully acknowledged.
led to further research in the area of reversible computing and
reversible electronic circuits [4]. Although elusive, a so-called
”information heat engine” was experimentally demonstrated
in [5] for showcasing information to energy conversion with
the aid of feedback control.
Despite the fact that information to energy conversion is
in its infancy, it was suggested from a fundamental perspective [6] that the currently separate treatment of information and
energy has to be challenged in the practical design of engineering systems. Naturally, information is carried by attaching
itself to a physical medium, such as waves or particles. In
molecular and nano communications, information is delivered
by conveying encoded particles from the source to destination [7], [8]. Similarly, in optical communications, information
is delivered by photons having information-dependent intensities, which may be detected by a photon-counting process [9],
[10]. Given the nature of the process, the system is capable
of providing a heating/illumination/propulsion function. Both
of the above examples suggest that the underlying matter that
carries information can be effectively reused for diverse applications. The explicit concept of transporting both information
and energy simultaneously was raised by the authors of [11],
[12], [13] in the context of power-line communications, where
information is conveyed by wires that carry electricity. This
topic was recently further extended to wireless communications, since it is desirable that a mobile device is free from
being tethered in any way, as prophesied a century ago [14].
2) Background: Before delving into the topic of Simultaneous Wireless Information and Power Transfer (SWIPT), a
brief historical portrayal of Wireless Power Transfer (WPT)
is warranted [15]. The earliest experiments on WPT were
conducted by Tesla with the ultimate goal of creating a
worldwide wireless power distribution system [16]. In the mid20th century, one of the most influential projects was the
microwave-powered helicopter [17]. Similarly, the National
Aeronautics and Space Administration’s (NASA) feasibility
study of the space-solar program [18] attracted global attention. The need for small to moderate power requirements (upto
hundreds of watts) and near-field (upto a few meters) WPT
increased substantially owing to the development of electronic
devices in the late 20th century. Hence the wireless power
consortium was established for promoting the standardization
for wireless charging, known as Qi 1 . WPT can be carried out
in two basic ways, namely based on either electromagnetic
1 see
websites: http://www.wirelesspowerconsortium.com/
induction in the form of inductive coupling and resonant
coupling relying on coils or with the aid of electromagnetic
radiation using microwave frequencies by relying on so-called
rectennas, which will be discussed in Section II-C. At the
time of writing there are already numerous WPT applications, ranging from wirelessly charged mobile phones [19] to
wirelessly powered medical implants [20] and to ’immortal’
sensor networks [21] in addition to passive Radio-frequency
IDentification (RFID) devices [22]. Furthermore, along with
successful start-up companies like WiTricity and Powercast 2 ,
it is thus the maturing WPT and wireless communications
fields that make the SWIPT an interesting emerging research
topic.
However, the current research of SWIPT is still in its
infancy and hence there is a paucity of literature, which
may be classified as follows. The first set of contributions
considered one-way SWIPT, where the transmitter simultaneously conveys both information and power to the receiver
that may be operated in two different modes [23], namely
either on a time-division basis or on a power-split basis.
To be more specific, in the time-division mode, the receiver
alternatively and opportunistically acts as an information detector and power converter [24]. On the other hand, in the
power-split mode, a portion of the received power is used
for powering the receiver and the remaining received power
is used for retrieving information [25], [26]. The second set
of contributions considered two-way SWIPT. Specifically, the
authors of [27], [28] investigated the scenario, where the
transmitter conveys power to the receiver, which is converted
to Direct Current (DC) power and reused for the destination’s
information transmission in the reverse direction. This mode
of operation is similar to that of the passive RFID devices
and hence it may be referred to as being two-way half-duplex.
Distinctively, [29] investigated a two-way full-duplex operational mode, where a pair of nodes interactively communicates
and exchanges power. The final set of treatises considered a
range of energy-transfer-aided systems, such as multi-carrier
systems [30], relay-assisted systems [31], interference-limited
arrangements [32], beamforming-assisted regimes [33], [34],
multiple access [35] and unicast/multicast scenarios [36].
3) Novelty: A close inspection of the existing literature
reveals that most of the current designs are centered on the
information transfer strategy in the presence of power transfer,
while satisfying a specific received energy constraint. It would
be however more beneficial to aim for the full ”integration” of
information and power transfer as the ultimate objective in the
spirit of Varshney’s concept [12], who proposed that energy
and information transfer should be innately inter-linked.
To be more explicit, in one-way SWIPT equipped with a
power-split receiver 3 , an inevitable Rate-Energy (RE) tradeoff has to be obeyed, which simply suggests that the more
power is transferred for powering the receiver, the lower the
communications rate becomes. This observation is intuitive,
since from an information theoretic point of view, less power
is left for information transfer, which inevitably leads to a
2 see
websites: http://www.witricity.com/ and http://www.powercastco.com/
time-division receiver is conceptually simple but it has been shown
to be sub-optimum [12]. Hence it is not considered in this paper.
3 The
reduced information rate. Naturally, having a reduced power
for information transfer implies that the information becomes
more prone to errors. One of the fundamental reasons for the
RE trade-off is that information is conventionally transferred
by modulating the amplitude, phase or frequency, which are
naturally subjected to channel-induced distortion, noise and
power-splitting. In order to eliminate these impediments, we
propose a completely different modulation type for Integrated
Wireless Information and Power Transfer (IWIPT), where
information is carried not by waveforms but by patterns.
Explicitly, a multi-antenna aided system is conceived for
IWIPT, where the specific Receive Antenna (RA) patterns
are used for conveying information, while simultaneously
transferring power.
• We achieve this with the aid of our recently proposed Generalised Pre-coding aided Spatial Modulation
(GPSM) scheme [37], where information is conveyed by
appropriately selecting the RA indices for information
transfer in addition to the information carried by the
conventional M -ary PSK/QAM modulation 4 .
• We show that the RA pattern-based information transfer represented in the Spatial Domain (SD) exhibits
a beneficial immunity to any performance degradation
imposed by the Radio Frequency (RF) to DC power
conversion in a power-split receiver and that our GPSM
scheme is capable of improving the overall RE tradeoff, when additionally the waveform-based information
transfer represented in the classic time domain is also
taken into account.
4) Organisation: The rest of our paper is organised as
follows. In Section II, we introduce the underlying concept as
well as the transceiver architecture of the GPSM aided IWIPT
system employing a power-split receiver. This is followed
by our analytical study in Section III, where we demonstrate that the proposed RA pattern-based information transfer
represented in the SD exhibits a beneficial immunity to RF
to DC power conversion induced performance erosion. We
then continue by charactering the RE trade-off of the overall
system. Our simulation results are provided in Section IV,
while we conclude in Section V.
II. S YSTEM M ODEL
Our proposed GPSM aided IWIPT system relies on the
availability of Channel State Information at the Transmitter
(CSIT) for supporting Transmitter Pre-Coding (TPC). This
scenario has been widely considered in the SWIPT literature [23], [27], [32], [33], [34], since it is desirable to shift
most/all signal processing demands from the less powerful
Down-Link (DL) receiver that may be passive or semi-passive
to the more powerful DL transmitter that may have access to
the mains power.
A. Conceptual Description
Consider a Multiple Input Multiple Output (MIMO) system
equipped with Nt Transmit Antennas (TA) and Nr RAs, where
4 Spatial Modulation [38], [39] conveys extra information by appropriately
selecting the transmit antenna indices, but our GPSM scheme conveys extra
information by appropriately selecting the receive antenna indices [37].
Booster
σa2 ρ Rectifier
RF
DC
k
DC
DC
σb2
bm 1
bm N a
inactive
(1)
BB
RF
active
s km = Ω kb m ,
bm̂1
RF
P
we have the relationship
Battery
Nt
1−ρ
Nr
BB
k̂
bm̂Na
Fig. 1.
Transceiver architecture of GPSM aided IWIPT. For example, activating Na = 2 RAs out of Nr = 4 RAs results in a total
of |Ct | = 6 legitimate activation patterns, i.e. the patterns of Ct =
{[1, 2], [1, 3], [1, 4], [2, 3], [2, 4], [3, 4]}. This configuration delivers kant =
2 bits of information because ⌊log2 (|6|)⌋ = 2. Upon selecting for example
C = {Ct (1), Ct (2), Ct (3), Ct (4)}, we have the following mapping between
the SD symbol and the activation patterns k = 1 7→ C(1) = [1, 2], k = 2 7→
C(2) = [1, 3], k = 3 7→ C(5) = [1, 4] and k = 4 7→ C(6) = [2, 3].
we assume Nt ≥ Nr . In this MIMO set-up, a maximum
of Nr = min(Nt , Nr ) parallel data streams may be supported by conventional waveform-based information transfer,
conveying a total of kef f = Nr kmod bits altogether, where
kmod = log2 (M ) denotes the number of bits per symbol of
a conventional M -ary PSK/QAM scheme and its alphabet is
denoted by A. The TPC matrix of P ∈ CNt ×Nr may be used
for pro-actively mitigating the Inter-Channel Interference (ICI)
at the transmitter by pre-processing the source signal before
its transmission upon exploiting the knowledge of the CSIT.
As a benefit, a low-complexity single-stream detection scheme
may be used by the receiver, because the ICI is eliminated at
the transmitter.
In contrast to the above-mentioned classic waveform-based
multiplexing of Nr data streams, in our GPSM scheme a total
of Na < Nr RAs are activated so as to facilitate the simultaneous transmission of Na data streams, where the particular
pattern of the Na RAs activated conveys information in form
of so-called SD symbols in addition to the information carried
by the conventional M -ary PSK/QAM modulated symbols as
seen in Fig 1 5 . Hence, the number of bits in GPSM conveyed
by a SD symbol becomes kant = ⌊log2 (|Ct |)⌋, where the set Ct
contains all the combinations associated with choosing Na activated RAs out of Nr RAs. As a result, the total number of bits
transmitted by the GPSM scheme is kef f = kant +Na kmod . To
assist our further discussions, let us define the set of selected
activation patterns as C ⊂ Ct . We also let C(k) and C(k, i)
denote the kth RA activation pattern and the ith activated RA
in the kth activation pattern, respectively. Finally, it is plausible
that the conventional MIMO scheme using purely waveformbased information transfer obeys Na = Nr .
B. IWIPT Transmitter
More specifically, let s km be an explicit representation of
a so-called super-symbol s ∈ CNr ×1 , indicating that the RA
pattern k is activated and Na conventional modulated symbols
b m = [bm1 , . . . , bmNa ]T ∈ CNa ×1 are transmitted, where we
have bmi ∈ A and E[|bmi |2 ] = 1, ∀i ∈ [1, Na ]. In other words,
5 Note that in our GPSM scheme, it is not necessary to convey any additional
information by a conventional M -ary PSK/QAM scheme, we may consider
a pure spatial modulation-aided system.
where Ω k = I [:, C(k)] is constituted by the specifically
selected columns determined by C(k) of an identity matrix of
I Nr . Following TPC, the resultant transmit signal x ∈ CNt ×1
may be written as
p
(2)
x = βs /NaP s km .
In order to avoid any power fluctuation during the preprocessing, we introduce the scaling factor of βs designed for
x ||2 = 1. As a natural
maintaining the power-constraint of ||x
requirement, the TPC matrix has to ensure that no energy leaks
into the RA-elements of the unintended RA patterns. Hence,
the classic linear Channel Inversion (CI)-based TPC [40], [41]
may be used, which is formulated as
H H H )−1 ,
P = H H (H
(3)
where H ∈ CNr ×Nt represents the MIMO channel involved.
We assume furthermore that each entry of H undergoes
frequency-flat Rayleigh fading and it is uncorrelated between
different super-symbol transmissions, while remains constant
within the duration of a super-symbol’s transmission. The
power-normalisation factor of the output power after preprocessing is given by
H H H )−1s .
βs = Na /ssH (H
(4)
The Base Band (BB) signal x is then up converted to x RF
and sent through Nt TAs.
C. IWIPT Receiver
1) Architecture: The power-split receiver of Fig 1 is employed [23], where we have a power transfer tunnel and
information transfer tunnel. As seen from Fig 1, the received
signal will be first subject to power conversion at each RA
using a power-splitting ratio of ρ ∈ (0, 1) at the RF stage,
where the power transfer tunnel consists of a rectifier used
for converting the received RF power to DC power, which is
followed by a (multi-stage) DC to DC booster. Note that from
a pure WPT point of view, the RAs, the rectifier and the DCbooster are jointly known as a rectenna. After power-splitting,
the remaining signal power will be used for information
transfer, which relies on determining the SD symbol for RA
pattern-based information transfer at the RF-stage and on
detecting the conventional modulated symbols for waveformbased information transfer at the BB-stage. Note that an
additional benefit of our GPSM scheme is that a reduced
number of RF to BB conversion chains are required.
2) Power Conversion: The RF signal observed at the Nr
RAs may be written as
r RF = H xRF + w RF
a ,
w RF
a
(5)
where
is the RF receiver’s noise. The equivalent BB
representation of the noise w a ∈ CNr ×1 may be modelled by
a circularly symmetric complex-valued Gaussian noise vector
with each entry having a zero mean and a variance of σa2 , i.e.
w a ||2 ] = σa2 Nr .
we have E[||w
At the RF-stage of Fig 1, a portion ρ of the received power
is converted to DC, where the average energy transferred by
the transmitter per unit time and gleaned from all the Nr RAs
is represented as
#
" N
r
X
RF 2
|r̃i | ,
(6)
Q = ξE ρ
i=1
where r̃ RF = H x RF is the noiseless part of r RF , representing
the power transferred by the transmitter, while the additive
noise of w RF
in (5) is not transferred by the transmitter, it
a
is imposed by the RF receiver. For the time being, designing
adaptive RA specific power-splitting ratios ρi , i ∈ [1, Nr ] is
set aside for our future investigations and we also set the RF
to DC conversion efficiency to ξ = 1 in this paper.
3) RA Pattern-based Information Transfer: After RF to DC
power conversion of each RA at a ratio of ρ, the remaining
RF signal observed at the Nr RAs invoked for information
transfer may be written as
p
= 1 − ρ H x RF + w RF
y RF
,
(7)
a
a
where the SD symbol may be determined by simply sorting
the remaining received power accumulated by each legitimate
RA pattern, which is mathematically represented as
(N
)
a
X
RF
2
|ya,C(ℓ,i) | .
(8)
k̂ = arg max
ℓ∈[1,|C|]
m̂i = arg min
ni ∈[1,M]
|yb,v̂i −
s
(1 − ρ)β
h v̂i p v̂i bni |2 , (11)
Na
where v̂i = C(k̂, i). Thus, correct detection is declared, when
we have m̂i = mi , ∀i.
III. S YSTEM A NALYSIS
In this section, we will first show that the RA pattern-based
information transfer of (8) is immune to power conversion,
since the Symbol Error Ratio (SER) of the SD symbols
remains the same as in the absence of power transfer. This
is then followed by our RE trade-off analysis of the overall
GPSM scheme, which is constituted by both the SD symbol
used for RA pattern-based information transfer and by the
conventional modulated symbols employed for waveformbased information transfer.
A. Analytical SER of RA Pattern-based Information Transfer
We commence our discussion by directly stating:
Lemma III.1. The RA pattern-based information transfer of
(8) is immune to the power conversion operation of (6) for
ρ ∈ (0, 1).
i=1
Thus, correct detection is declared when we have k̂ = k. We
will later show that the factor (1 − ρ) does not affect the detection of SD symbols and our RA pattern-based information
transfer exhibits a beneficial immunity to power conversion.
4) Waveform-based Information Transfer: After determining the RA pattern using the non-coherent detection process
of (8), we proceed with the conventional waveform-based
information transfer. After RF to BB down conversion, the
BB signal may be written as
p
p
βs /NaH P s km + w a + w b ,
(9)
yb = 1 − ρ
where w b ∈ CNr ×1 is the down-conversion induced noise,
which is modelled similarly to the RF receiver’s noise w a ,
but with a variance of σb2 . We also let σa2 = ασ 2 and σb2 =
(1 − α)σ 2 , where σ 2 represents the total noise variance at the
power-split receiver of Fig 1.
Assuming that the vi = C(k, i)th RA is activated, the
Signal-to-Interference-plus-Noise-Ratio (SINR) γvi of the
conventional modulated symbols after power conversion is
given by
γvi =
as
hvi p vi ||2
||h
,
2
[σa2 + σb /(1 − ρ)]Na /βs
(10)
where h vi is the vi th row of H representing the channel
between the vi th RA and the transmitter, while p vi is the vi th
column of P representing the vi th TPC vector.
As a subsequent stage after the RA pattern determination,
the detection of conventional modulated symbols is formulated
Proof: Firstly we note that (8) relies on the principle of
energy detection. Hence our analysis may be performed using
the equivalent BB representation of (7), as suggested in [42],
which is represented as
p
p
βs /NaH P s km + w a .
(12)
ya = 1 − ρ
Hence (8) becomes
k̂ = arg max
ℓ∈[1,|C|]
(N
a
X
i=1
|ya,C(ℓ,i) |
2
)
.
(13)
Given the values of Nr as well as Na and assuming that the
RA pattern C(k) was activated, after substituting (3) into (12),
we have:
p
p
βs /Na bmi + wa,vi ,∀vi ∈ C(k), (14)
ya,vi = (1 − ρ)
p
¯
ya,ui = (1 − ρ)wa,ui ,
∀ui ∈ C(k),
(15)
¯ denotes the complementary set of the activated RA
where C(k)
2
pattern C(k) in C. Furthermore, upon introducing σa,0
= σa2 /2,
we have:
|ya,vi |2 = R(ya,vi )2 + I(ya,vi )2
hp
i
2
∼N
(1 − ρ)βs /Na R(bmi ), (1 − ρ)σa,0
i
hp
2
, (16)
(1 − ρ)βs /Na I(bmi ), (1 − ρ)σa,0
+N
|ya,ui |2 = R(wa,ui )2 + I(wa,ui )2
2
2
,
+ N 0, (1 − ρ)σa,0
∼ N 0, (1 − ρ)σa,0
(17)
where R(·) and I(·) represent the real and imaginary operators, respectively. As a result, by normalisation with respect
2
to (1 − ρ)σa,0
, we have the following observations:
|ya,vi |2 ∼ χ22 (g; λvi ),
2
|ya,ui | ∼
χ22 (g),
∀vi ∈ C(k),
¯
∀ui ∈ C(k),
(18)
(19)
where the non-centrality parameter is given by
2
.
λvi = βs |bmi |2 /Na σa,0
(20)
Exploiting the fact that E[|bmi |2 ] = 1, ∀i or |bmi |2 = 1, ∀i for
PSK modulation, we have
2
,
λ = λvi = βs /Na σa,0
∀vi
(21)
Note that λ is also a random variable obeying the distribution
of fλ (λ), whose analytical expression will be given in Lemma
III.2.
Recall from (13) that theP
correct decision concerning the
Na
2
SD symbol occurs, when
is the maximum.
i=1 |ya,vi |
Since the non-centrality parameter given by (20) is not a
function of ρ, P
power conversion operation does not affect
a
2
the quantity of N
i=1 |ya,vi | . Hence, it becomes explicit that
the power-splitting ratio ρ does not affect the SER esant of
the SD symbol of our GPSM scheme for RA pattern-based
information transfer expressed in Lemma III.3.
Lemma III.2. (Proof see Appendix C of [43]) The distribution
fλ (λ) of the non-centrality parameter λ is given by:
fλ (λ) =
NaNt −Nr +1 σa2 /2 −λNa σa2 /2
e
(λσa2 /2)Nt −Nr .
(Nt − Nr )!
(1 − ρ)σa2 + σb2 . Thus the SER esmod may be upper bounded
by [44]:
Z ∞
p
(25)
esmod ≤ Nmin
Q dmin γ/2 fγ (γ)dγ,
0
where fγ (γ) is a scaled version of fλ (λ) given in Lemma
III.2, i.e. we have fγ (γ) = κfλ (κγ), where
κ = 2[σa2 + σb2 /(1 − ρ)]/σa2 .
(26)
Moreover, dmin is the minimum Euclidean distance in the
conventional modulated symbol constellation, Nmin is the
average number of the nearest neighbours separated by dmin
in the constellation and Q(·) denotes the Gaussian Q-function.
We then introduce ẽsmod for representing the SER of the
conventional modulated symbols in the presence of SD symbol
errors due to the detection of (8), which is formally given by
Lemma III.4. (Proof see Appendix A of [43]) Given the kth
activated RA patten, the SER of the conventional modulated
symbols for waveform-based information transfer in the presence of SD symbol errors can be calculated as:
X Nc es + Nd es
esant
o
mod
,
ẽsmod = (1 − esant )esmod + kant
− 1)
(2
Na
ℓ6=k |
{z
}
B
(27)
where Nc and Nd = (Na − Nc ) represent the number of
RAs that are common and different between C(ℓ) and C(k),
respectively. Mathematically we have:
(22)
Nc =
Na
X
I [C(ℓ, i) ∈ C(k)] .
(28)
i=1
Lemma III.3. (Proof see Theorem III.1 of [43]) The analytical
Moreover, eso is defined as the modulation-type-dependent SER.
SER esant of the SD symbol of our GPSM scheme relying on
CI TPC may be formulated as:
Finally, when CI based TPC is introduced, only random
Na
Z ∞ Z ∞
s
Nr −Na
noise
may be received by the Nd RAs in C(ℓ), thus eso simply
eant ≈ 1 −
[Fχ22 (g)]
fχ22 (g; λ)dg
fλ (λ)dλ,
represents the SER of the conventional modulated symbols as
0
0
(23) a result of a random guess. For example, for QPSK, we have
eso = 3/4. Note that, when the SER of the SD symbols esant
where Fχ22 (g) represents the Cumulative Distribution Function
is small, we have ẽsmod ≈ esmod .
(CDF) of a chi-square distribution having two degrees of
freedom, while fχ22 (g; λ) represents the Probability Distribution Function (PDF) of a non-central chi-square distribution C. Rate Energy Trade-off
having two degrees of freedom and non-centrality given by λ
1) DCMC Capacity Expression: The RE trade-off may be
with its PDF of fλ (λ).
formulated using the power conversion of (6) and comparing it directly against the Discrete-input Continuous-output
Memoryless Channel (DCMC) capacity of our GPSM scheme.
B. Analytical SER of Waveform-based Information Transfer
This is because in our GPSM scheme the SD symbols convey
We first introduce the SER esmod of the conventional modinteger values constituted by the RA pattern index, which does
ulated symbols bmi ∈ A for waveform-based information
not obey the shaping requirements of Gaussian signalling [45].
transfer in the absence of SD symbol errors. For a specific
Hence, for simplicity’s sake, we discuss the DCMC capacity of
activated RA pattern in C(k), the SINR of (10) becomes
our GPSM scheme in the context of discrete-input signalling
2
2
∀vi
(24) for both the SD symbol and for the conventional modulated
γ = γvi = βs /Na [σa + σb /(1 − ρ)],
symbols mapped to it.
¯
and for the remaining deactivated RAs in C(k)
we have only
The Mutual Information per Bit (MIB) I(z; ẑ) of our GPSM
random noise contributions of zero mean and of a variance of scheme measured between the input bits z ∈ [0, 1] and the
corresponding demodulated output bits ẑ ∈ [0, 1] is given by:
"
#
X
X
X
I(z; ẑ) = −
Pz log Pz −
Pz|ẑ log Pz|ẑ
Pẑ
z
=1+
ẑ
ebef f
log ebef f
+ (1 −
z
b
eef f ) log(1
− ebef f ), (29)
where Pz is the Probability Mass Function (PMF) of z, where
we adopt the common assumption of equi-probability bits of
Pz=0 = Pz=1 = 1/2. On the other hand, the conditional
entropy, representing the average uncertainty about z after
observing ẑ, is a function of the average Bit Error Ratio (BER)
of our GPSM scheme, denoted as ebef f . As a result, the socalled DCMC capacity becomes
R = kef f I(z; ẑ).
(30)
2) Average BER Expression: Let us now discuss the expression of ebef f for the sake of calculating (29) and (30), which
may be derived from its accurate SER expression of both
the SD symbol for the RA pattern-based information transfer
given in (23) and for the conventional modulated symbols for
waveform-based information transfer given in (27).
Let ebant and ẽbmod represent the BER of the SD symbol for
RA pattern-based information transfer and of the conventional
modulated symbols for waveform-based information transfer.
Hence, the accurate expression of the average BER of our
GPSM scheme may be written as:
ebef f
=
=
(kant ebant + Na kmod ẽbmod )/kef f
(δant esant + Na ẽsmod )/kef f ,
(31)
where the second equation of (31) obeys from the relationship
of
ẽbmod = ẽsmod /kmod ,
ebant = δkant esant /kant ,
(32)
(33)
where we formulate Lemma III.5 for the expression of δkant
acting as a correction factor in (31).
Lemma III.5. (Proof see Appendix B of [43]) The generic
expression of the correction factor δkant for kant bits of
information is given by:
2kant −1 − δkant −1
,
(34)
2kant − 1
where given δ0 = 0, we can recursively determine δkant .
δkant = δkant −1 +
D. Further Discussions
1) Robustness: The above RE trade-off is based on the
assumption of having perfect CSIT. Let us now consider the
system’s robustness to both CSIT-errors and antenna correlations, where the expression ebef f of our GPSM scheme will
be determined empirically for calculating (29) and (30).
Like in all TPC schemes, an important related aspect is
its resilience to CSIT inaccuracies. In this paper, we let
H = H t + H e , where H t represents the matrix hosting the
average CSI, with each entry obeying the complex Gaussian
distribution of ht ∼ CN (0, σt2 ) and H e is the instantaneous
CSI error matrix obeying the complex Gaussian distribution of
he ∼ CN (0, σe2 ), where we have σt2 +σe2 = 1. As a result, only
H t is available at the transmitter for pre-processing. There is
a plethora of CSIT inaccuracy counter-measures conceived for
general transmitter pre-coding schemes in the literature [46],
[47], but they are beyond the scope of our discussions in this
paper.
Another practical phenomenon is the presence of correlation
amongst the signals at the transmit and receive antennas.
The correlated MIMO channel is modelled by the widelyused Kronecker model [48], which is formulated as H =
T
R 1/2
G (R
R 1/2
(R
r ) , with G representing the original MIMO
t )G
channel imposing no correlation, while R t and R r represents
the correlations at the transmitter and receiver side, respec|i−j|
tively, with the correlation entries given by Rt (i, j) = ρt
|i−j|
and Rr (i, j) = ρr .
2) Asymptotic Insights: It is also interesting to investigate
the impact of power conversion on both the MIB as well as on
the DCMC capacity of our GPSM scheme and to contrast it to
that of the conventional CI pre-coding based MIMO scheme
for large-scale MIMO settings of Nt → ∞, Nr → ∞ and for
a fixed system load ratio of Nr /Nt .
In order to provide these asymptotic insights for these largescale MIMO settings, we have to replace the constraint of
x||2 = 1 with its relaxed version of E[||x
x ||2 ] = 1. Hence
||x
instead of (4), we have the relaxed normalisation factor of
H H H )−1 ].
βl = Nr /Tr[(H
(35)
As shown in Corollary 4 of [49] and also in Chapter 14 of [50],
under asymptotic settings, (35) tends to
βl 7−→ (Nt /Nr − 1).
(36)
By exploiting (36) and by replacing (4) with (35) in the expressions of (21) and (24), we have the following deterministic
quantity
2
λ 7−→ λd = (Nt /Nr − 1)/Na σa,0
,
γ 7−→ γd = (Nt /Nr −
1)/Na [σa2
+
(37)
σb2 /(1
− ρ)].
(38)
As a result, we have
Z ∞
Na
s
Nr −Na
eant,d = 1 −
[Fχ22 (g)]
fχ22 (g; λd )dg
, (39)
0
p
(40)
esmod,d = Nmin Q dmin γd /2 .
Noting that when esant,d is small, we arrive at the approximation ẽsmod,d ≈ esmod,d as suggested by (27), thus we may
obtain ebef f,d by replacing esant and ẽsmod in (31) with their
deterministic expressions in (39) and (40). Finally, by inserting
ebef f,d in (29) and (30) instead of ebef f , we arrive at the MIB
and DCMC capacity for both our GPSM scheme as well
as for the conventional CI pre-coding based MIMO scheme
(Na = Nr ) under the asymptotic settings of Nt → ∞ and
Nr → ∞.
IV. N UMERICAL R ESULTS
Let us now consider our numerical results for characterizing
the GPSM aided IWIPT for both RA pattern-based information
transfer and for waveform-based information transfer relying
Nt
16
16
16
16
16
16
•16
Nr
8
8
8
8
8
8
8
Na
1
2
3
4
5
6
8
M
2
2
2
2
2
2
2
kef f
5 = 3+1×2
8 = 4+2×2
11 = 5+3×2
14 = 6+4×2
15 = 5+5×2
16 = 4+6×2
16 = 0+8×2
on the CI-based precoding of (3) under {Nt , Nr } = {16, 8}.
Since our focus is on RA pattern-based information transfer,
simple QPSK was used for the conventional data modulation component of our waveform-based information transfer. Furthermore, we investigate the scenarios correspond to
Na = [1, . . . , 6] in our GPSM scheme since the setting of
Na = 7 results into higher complexity than the conventional
CI pre-coding based MIMO bench-maker as discussed in [37].
Maximum achievable throughput under MIMO settings of
{Nt , Nr } = {16, 8} and employing QPSK modulation is
shown in Table I, which may be used to compare our DCMC
capacity related investigations. Finally, the Signal-to-NoiseRatio (SNR) per bit is defined as SN Rb = 1/σ 2 (kef f /Na ).
0
Spatial SER
10
Na = 1
−2
10
Na = 2
Na = 3
−4
10
N =4
a
−10 −9
−8
−7
−6
−5
−4
−3 −2 −1
SNR [dB]
0
1
2
3
4
5
−3 −2 −1
SNR [dB]
0
1
2
3
4
5
b
0
10
GPSM BER
TABLE I
M AXIMUM ACHIEVABLE THROUGHPUT kef f OF GPSM AND OF THE
CONVENTIONAL MIMO CORRESPONDING TO Na = Nr ( IN •) UNDER
SETTINGS OF {Nt , Nr } = {16, 8} AND EMPLOYING QPSK MODULATION
M = 2, WHERE kef f = kant + Na kmod FOR GPSM AND
kef f = Nr kmod FOR CONVENTIONAL MIMO.
−2
10
N =5
a
Na = 6
−4
10
analytical
−10 −9
−8
−7
−6
−5
−4
b
Fig. 2.
Spatial SER of (23) (upper) and GPSM overall BER of (31)
(lower) using CI based TPC and a power-normalisation factor of (4) under
{Nt , Nr } = {16, 8}, where we set the power-splitting ratio to ρ = 0 and
the noise ratio to α = 1.
0
10
with WPT
no WPT
N =1
a
−1
10
N =2
a
N =4
a
N =6
a
−2
10
Fig 2 shows the spatial SER (upper) as well as the GPSM
scheme’s overall BER (lower) under the power-normalisation
factor of (4) for {Nt , Nr } = {16, 8}. Without loss of generality, we set the power-splitting ratio to ρ = 0 and the noise
ratio to α = 1.
Observe in the upper subplot of Fig 2 that our empirical
simulation based SER results recorded for the SD symbol of
the RA pattern-based information transfer accurately match
the analytical results given in (23) and form tight upper
bounds. Similarly, when the GPSM scheme’s overall BER is
considered in the lower subplot of both figures, our empirical
results accurately match the analytical results given in (31).
Fig 3 shows the GPSM scheme’s overall BER under the
power-normalisation factor of (4) for {Nt , Nr } = {16, 8},
where we set the power-splitting ratio to ρ = 0.5 and the
noise ratio to α = 0.4. As suggested by (24), the total noise at
the BB for waveform-based information transfer now becomes
1.6 times that with no WPT. Hence, we may expect to see a
BER performance degradation of 10 log10 (1.6) ≈ 2dB. This is
clearly demonstrated in Fig 3 at BER of 10−5 , which implies
that the RA pattern-based information transfer in our GPSM
scheme is immune to the RF to DC power conversion induced
performance erosion as otherwise the performance degradation
would be more than 2dB.
B. Investigation on MIB
1) Immunity to Power Conversion: Fig 4 investigates the
MIB versus energy trade-off of purely RA pattern-based information transfer for {Nt , Nr } = {16, 8} and SN Rb = 0dB,
BER
A. Investigation on Error Probability
Na = 8
−3
10
−4
10
−5
10
−10
−8
−6
−4
−2
0
2
4
SNRb [dB]
6
8
10
12
14
Fig. 3. BER of conventional CI pre-coding based MIMO bench-maker as
well as GPSM using CI based TPC and a power-normalisation factor of (4)
under {Nt , Nr } = {16, 8}, where we set the power-splitting ratio to ρ = 0.5
and the noise ratio to α = 0.4.
where we set α = 0.4 in the left subplot and α = 0.6 in the
right subplot. The results of MIB were evaluated from (29).
Observe from both subplots of Fig 4 that for all values of
Na , apart from setting ρ = 1, where the receiver acts purely
as a WPT and the normalised converted energy is unity, the
remaining settings of the power-splitting ratio ρ do not affect
the MIB of the SD symbols in our GPSM scheme, as suggested
by the horizontal lines. More explicitly, for all values of Na
seen in Fig 4, the MIB of SD symbols vs the power conversion
ratio of ρ ∈ [0, 0.1, . . . , 0.9] remains identical to that under
ρ = 0. This implies that the RA pattern-based information
transfer relying on SD symbols operates as if no WPT was
provided, which justifies our Lemma III.1. Furthermore, by
comparing the left and right subplot of Fig 4, it can be seen that
the MIB of SD symbols becomes lower, when an increased
RF receiver noise is encountered in conjunction with α = 0.6.
1
0.9
0.9
0.9
0.9
0.8
0.8
0.8
0.8
0.7
0.7
0.7
0.7
0.6
0.6
0.6
0.6
0.5
0.4
0.3
0.5
MIB
1
MIB
1
MIB
0.5
0.4
0.4
0.3
0.3
0.3
Na = 1
Na = 1
0.2
Na = 2
Na = 3
0.1
0.2
Na = 5
0.1
N =4
N =6
a
0
0
0.2 0.4 0.6 0.8
Normalised Energy
Na = 2
Na = 3
0.1
N =4
a
1
0
0
0.2 0.4 0.6 0.8
Normalised Energy
0.2
N =5
0.1
Na = 6
a
N =8
a
0
1
Fig. 4. MIB versus energy trade-off curve of RA pattern-based information
transfer under {Nt , Nr } = {16, 8} and SN Rb = 0dB with α = 0.4 (left)
and α = 0.6 (right). The MIB results were evaluated from (29).
Note that our RA pattern-based information transfer is immune
to power conversion, but naturally not to the RF receiver noise.
2) MIB versus Energy Trade-off: Fig 5 investigates the MIB
versus energy trade-off of our GPSM aided IWIPT system
supporting both RA pattern-based information transfer and
waveform-based information transfer for {Nt , Nr } = {16, 8}
and SN Rb = 0dB, where we set α = 0.4 in the left subplot
and α = 0.6 in the right subplot. The results of MIB were
evaluated from (29).
Observe from both subplots of Fig 5 that the near-constant
MIB curves of our GPSM schemes shown in Fig 4 can no
longer be maintained upon increasing the number of activated
RAs, although they exhibit a less eroded MIB than the
conventional CI pre-coding based MIMO scheme associated
with Na = 8 relying on pure waveform-based information
transfer. This is because when waveform-based information
transfer is also invoked, the resultant MIB of our GPSM
scheme becomes more sensitive to the power conversion ratio
than the purely RA pattern-based information transfer. This
is evidenced in Fig 5, where the higher the value of Na , the
more conventional modulated symbols are conveyed. When
comparing the left subplot to the right subplot, the MIB of our
GPSM scheme appears more sensitive to the value of α. In
other words, when the RF receiver’s noise is more dominant
(α = 0.6), the MIB of our GPSM schemes is reduced. On
the other hand, the conventional CI pre-coding based MIMO
scheme is less sensitive to the value of α.
C. Investigation on DCMC Capacity
1) DCMC Capacity versus Energy Trade-off: Fig 6 investigates the DCMC capacity versus energy trade-off of our
GPSM aided IWIPT system, where the rest of the settings
were kept the same as in Fig 5. We can see that although Fig
5 suggests that the MIB of our GPSM schemes is higher than
that of the conventional CI pre-coding based MIMO scheme
for all power-splitting ratios for all values of Na , when the
DCMC capacity is considered in Fig 6, the GPSM curves
0
0.2 0.4 0.6 0.8
Normalised Energy
a
0
1
0
0.2 0.4 0.6 0.8
Normalised Energy
1
Fig. 5. MIB versus energy trade-off curve of our GPSM aided IWIPT system
relying on both RA pattern-based information transfer and on waveform-based
information transfer under {Nt , Nr } = {16, 8} and SN Rb = 0dB using
α = 0.4 (left) and α = 0.6 (right). The MIB results were evaluated from
(29).
DCMC Capacity [bits/supersymbol]
0.2
0.5
0.4
14
14
12
12
DCMC Capacity [bits/supersymbol]
MIB
1
10
8
6
4
Na = 1
Na = 2
2
10
8
6
4
N =5
a
2
Na = 3
Na = 6
Na = 4
0
0
0.2 0.4 0.6 0.8
Normalised Energy
Na = 8
1
0
0
0.2 0.4 0.6 0.8
Normalised Energy
1
Fig. 6.
DCMC capacity versus energy trade-off curve of our GPSM
aided IWIPT system relying on both RA pattern-based information transfer
and waveform-based information transfer under {Nt , Nr } = {16, 8} and
SN Rb = 0dB for α = 0.4 (left) and α = 0.6 (right). The DCMC capacity
results were evaluated from (30) and these DCMC capacity results may be
compared to the maximum achievable throughput values listed in Table I.
associated with Na = 1 and Na = 2 exhibit intersections
with the curve of the conventional CI pre-coding based MIMO
scheme. By contrast, the GPSM schemes using Na ≤ 3 offer
a DCMC capacity, which is higher than that exhibited by the
conventional CI pre-coding based MIMO scheme. Finally, all
of our GPSM schemes exhibit a higher degree of immunity to
power conversion than the conventional CI pre-coding based
MIMO scheme.
2) DCMC Capacity versus SN Rb : It is thus interesting to
observe in detail the behaviour of our GPSM scheme associated with Na = 1 and Na = 2. Fig 7 compares the DCMC
capacity versus SNR per bit curve of the conventional CI precoding based MIMO scheme relying on Na = 8 to our GPSM
14
14
a
DCMC Capacity [bits/supersymbol]
DCMC Capacity [bits/supersymbol]
N =2
a
12
Na = 8
10
8
6
4
2
Na = 8
DCMC Capacity [bits/supersymbol]
N =2
12
14
14
12
12
Na = 1
10
8
6
4
2
DCMC Capacity [bits/supersymbol]
Na = 1
10
8
6
4
2
10
8
6
4
Na = 2
Na = 4
2
N =6
a
no impairments
with impairments
0
−10
−8
−6
−4
SNR [dB]
−2
b
0
0
−10
−8
−6
−4
SNR [dB]
−2
0
b
0
0
0.2 0.4 0.6 0.8
Normalised Energy
N =8
a
1
0
0
0.2 0.4 0.6 0.8
Normalised Energy
1
Fig. 7. DCMC capacity versus SNR per bit curve of our GPSM aided
IWIPT system relying on both RA pattern-based information transfer and
waveform-based information transfer under {Nt , Nr } = {16, 8} with α =
0.4 (left) and α = 0.6 (right). For each group of curves (circle, square
and no legend), the curves spanning from top to bottom correspond to ρ =
0.2, 0.4, 0.6, 0.8 in both plots. The DCMC capacity results were evaluated
from (30) and these DCMC capacity results may be compared to the maximum
achievable throughput values listed in Table I.
Fig. 8.
DCMC capacity versus energy trade-off curve for our GPSM
aided IWIPT system relying on both RA pattern-based information transfer
and waveform-based information transfer under {Nt , Nr } = {16, 8} and
SN Rb = 0dB with α = 0.4 suffering from imperfect CSIT having
a variance of σe = 0.2 (left) and antenna correlation associated with
ρt = ρr = 0.4 (right). The DCMC capacity results were evaluated from
(30) and these DCMC capacity results may be compared to the maximum
achievable throughput values listed in Table I.
schemes associated with Na = 1 and Na = 2 supporting both
RA pattern-based information transfer and waveform-based
information transfer under {Nt , Nr } = {16, 8} using α = 0.4
(left) and α = 0.6 (right). Observe for each group of curves
(circle Na = 1, square Na = 2 and no legend Na = 8) in both
subplots of Fig 7 that the curves spanning from top to bottom
correspond to power-splitting ratios of ρ = 0.2, 0.4, 0.6, 0.8.
In can also be seen in Fig 7 that our GPSM schemes using
Na = 1 and Na = 2 offer a higher DCMC capacity than that
facilitated by the conventional CI pre-coding based MIMO
scheme of Na = 8 in the low SN Rb range. Furthermore,
observe in both subplots of Fig 7 that the DCMC capacity
curves of our GPSM schemes using Na = 1 and Na = 2 are
only slightly affected upon increasing the power-splitting ratio
ρ, while we observe a dramatic difference in the DCMC capacity curves of the conventional CI pre-coding based MIMO
scheme of Na = 8. Thus, our GPSM schemes using Na = 1
and Na = 2 is significantly more immune to the performance
erosion of power conversion for a wide range of SN Rb values.
our GPSM scheme is more immune to any power conversion
induced performance erosion, which is an explicit benefit of
our RA pattern-based information transfer, as evidenced in Fig
8.
2) Asymptotic Investigation: Fig 9 characterizes the impact
of power conversion both on the MIB (left) and on the DCMC
capacity (right) for our GPSM scheme associated with Na = 1
and Na = 2 as well as for the conventional CI pre-coding
based MIMO scheme using Na = Nr under asymptotic
settings of Nt = 2048 and Nr ∈ [64, 2046]. Furthermore,
SN Rb = 0dB and α = 0.4 were set in Fig 9. For each group
of curves (Na = 1, Na = 2 and Na = Nr ) seen in both
subplots of Fig 9, the curves spanning from top to bottom
correspond to ρ = 0.2, 0.4, 0.6, 0.8.
It can be seen from the left subplot of Fig 9 that for all
values of Nr /Nt , the MIB of both of our GPSM schemes
is higher than that of the conventional CI pre-coding based
MIMO scheme. Furthermore, the common MIB trend for all
schemes is monotonically decreasing upon increasing Nr /Nt .
However, we observe in the right subplot of Fig 9 that upon
increasing Nr /Nt , the DCMC capacity of the conventional CI
pre-coding based MIMO scheme is monotonically decreasing,
while the best value of Nr /Nt may be deemed to be the one
that attains the highest DCMC capacity in our GPSM schemes.
Specifically, the DCMC capacity of the conventional CI precoding based MIMO scheme of Na = Nr is seen to be higher
than that of our GPSM schemes for small values of Nr /Nt .
Furthermore, our GPSM arrangement using Nr = 2 provides
a higher DCMC capacity than the GPSM scheme relying on
Nr = 1, when Nr /Nt is lower than about 0.5, but this trend
is reversed otherwise.
Finally, a substantial DCMC capacity penalty is observed
for the conventional CI pre-coding based MIMO regime of
Na = Nr upon increasing the power-splitting ratio according
D. Further Investigations
1) Robustness Investigation: Fig 8 investigates the DCMC
capacity versus energy trade-off curves of our GPSM aided
IWIPT system under {Nt , Nr } = {16, 8} and SN Rb = 0dB
using α = 0.4 and suffering from imperfect CSIT associated
with a variance of σe = 0.2 (left) or when experiencing an
antenna correlation associated with ρt = ρr = 0.4 (right).
For reasons of space-economy and to avoid crowded figures,
only the results of Na = {2, 4, 6, 8} are shown here. Observe
from both subplots of Fig 8 that as expected, a DCMC capacity
degradation is imposed both on our GPSM schemes and on the
conventional CI pre-coding based MIMO scheme of Na = 8.
Nonetheless, our previous conclusions still hold, namely that
1
45
0.9
40
0.8
DCMC Capacity [bits/supersymbol]
Na = 1
Na = 2
0.7
MIB
0.6
0.5
0.4
0.3
Na = Nr
0.2
Na = Nr
30
N =2
25
a
B
20
15
N =1
A
a
10
5
0.1
0
35
0
0.2
0.4 0.6
N /N
r
0.8
1
0
0
t
0.2
0.4 0.6
N /N
r
0.8
1
t
Fig. 9. The impact of power conversion on the MIB (left) and DCMC
capacity (right) of both our GPSM scheme using Na = 1 and Na = 2 as well
as the conventional CI pre-coding based MIMO scheme of Na = Nr under
the asymptotic settings of Nt = 2048 and Nr ∈ [64, 2046]. Furthermore,
we set SN Rb = 0dB and α = 0.4. For each group of curves (Na = 1,
Na = 2 and Na = Nr ), the curves spanning from top to bottom correspond
to ρ = 0.2, 0.4, 0.6, 0.8 in both subplots. The MIB and DCMC capacity
results were evaluated from (29) and (30).
to ρ = 0.2, 0.4, 0.6, 0.8, while this trend is less obvious in
both of our GPSM schemes. For example, both A and B
in the right subplot of Fig 9 show that an obvious DCMC
capacity penalty may only be imposed for ρ = 0.8. This again
implies that our GPSM aided IWIPT system is less sensitive
to the power conversion-induced performance erosion in the
asymptotic regime.
V. C ONCLUSIONS
A novel GPSM aided IWIPT system was proposed,
where both RA pattern-based information transfer as well
as waveform-based information transfer using conventional
PSK/QAM modulation were employed. The analytical SER
of the SD symbols and the BER of the overall GPSM scheme
were derived. Furthermore, the MIB vs energy trade-off and
the DCMC capacity vs energy trade-offs were discussed. Both
the analytical and simulation results demonstrate that the RA
pattern-based information transfer represented in the SD exhibits a beneficial immunity to power conversion and improves
the overall RE trade-off, when the additional waveform-based
QPSK information transfer is also taken into account. This
conclusion holds true also both in the presence of inaccurate
CSIT and of antenna correlations, even for large-scale MIMO
settings. Hence, our proposed GPSM aided IWIPT system may
be viewed as an instantiation of joint information and energy
transfer in the spirit of Varshney’s seminal concept .
VI. L IST
OF
P UBLICATIONS
R EFERENCES
[1] K. Maruyama, F. Nori, and V. Vedral, “The physics of Maxwell’s
demon and information,” Reviews of Modern Physics, vol. 81, pp. 1–23,
Jan 2009.
[2] R. Landauer, “Irreversibility and heat generation in the computing
process,” IBM Journal of Research and Development, vol. 5, no. 3,
pp. 183–191, 1961.
[3] ——, “Dissipation and noise immunity in computation and communication,” Nature, vol. 335, pp. 779–784, 1988.
[4] M. Saeedi and I. L. Markov, “Synthesis and optimization of reversible
circuits - a survey,” ACM Computing Surveys, vol. 45, no. 2, pp. 1–34,
Mar. 2013.
[5] S. Toyabe, T. Sagawa, M. Ueda, E. Muneyuki, and M. Sano, “Experimental demonstration of information-to-energy conversion and validation of the generalized Jarzynski equality,” Nature Physics, vol. 6, pp.
988–992, 2010.
[6] O. L. de Weck, D. Roos, and C. L. Magee, Engineering Systems:
Meeting Human Needs in a Complex Technological World. MIT Press,
2011.
[7] T. Nakano, M. Moore, F. Wei, A. Vasilakos, and J. Shuai, “Molecular
communication and networking: opportunities and challenges,” IEEE
Transactions on NanoBioscience, vol. 11, no. 2, pp. 135–148, 2012.
[8] I. F. Akyildiz, J. M. Jornet, and M. Pierobon, “Nanonetworks: a new
frontier in communications,” Communications of the ACM, vol. 54,
no. 11, pp. 84–89, Nov. 2011.
[9] X. Zhou, X. Zheng, R. Zhang, and L. Hanzo, “Chip-interleaved optical
code division multiple access relying on a photon-counting iterative
successive interference canceller for free-space optical channels,” Optics Express, vol. 21, no. 13, pp. 15 926–15 937, Jul 2013.
[10] L. Li, X. Zhou, R. Zhang, and L. Hanzo, “Performance and capacity
analysis of poisson photon-counting based Iter-PIC OCDMA systems,”
Optics Express, vol. 21, no. 22, pp. 25 954–25 967, Nov 2013.
[11] L. Varshney, “Transporting information and energy simultaneously,” in
2008 IEEE International Symposium on Information Theory, 2008, pp.
1612–1616.
[12] ——, “On energy/information cross-layer architectures,” in 2012 IEEE
International Symposium on Information Theory, 2012, pp. 1356–1360.
[13] P. Grover and A. Sahai, “Shannon meets Tesla: wireless information
and power transfer,” in 2010 IEEE International Symposium on Information Theory, 2010, pp. 2363–2367.
[14] N. Tesla, “The future of the wireless art,” Wireless Telegraphy &
Telephony, pp. 67–71, 1908.
[15] J. Garnica, R. Chinga, and J. Lin, “Wireless power transmission: From
far-field to near-field,” Proceedings of the IEEE, vol. 101, no. 6, pp.
1321–1331, 2013.
[16] N. Tesla, “Apparatus for transmitting electrical energy,” U.S. Patent
Patent 1 119 732, Dec, 1914.
[17] W. C. Brown, “The history of power transmission by radio waves,”
IEEE Transactions on Microwave Theory and Techniques, vol. 32,
no. 9, pp. 1230–1242, 1984.
[18] J. McSpadden and J. Mankins, “Space solar power programs and
microwave wireless power transmission technology,” IEEE Microwave
Magazine, vol. 3, no. 4, pp. 46–57, 2002.
[19] J. Lin, “Wireless power transfer for mobile applications and health
effects,” IEEE Antennas and Propagation Magazine, vol. 55, no. 2,
pp. 250–253, 2013.
[20] A. Yakovlev, S. Kim, and A. Poon, “Implantable biomedical devices:
wireless powering and communication,” IEEE Communications Magazine, vol. 50, no. 4, pp. 152–159, 2012.
[21] L. Xie, Y. Shi, Y. Hou, and A. Lou, “Wireless power transfer and
applications to sensor networks,” IEEE Wireless Communications,
vol. 20, no. 4, pp. 140–145, 2013.
[22] V. Chawla and D.-S. Ha, “An overview of passive RFID,” IEEE
Communications Magazine, vol. 45, no. 9, pp. 11–17, 2007.
[23] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless information and power transfer,” IEEE Transactions on Wireless
Communications, vol. 12, no. 5, pp. 1989–2001, 2013.
[24] L. Liu, R. Zhang, and K.-C. Chua, “Wireless information transfer
with opportunistic energy harvesting,” IEEE Transactions on Wireless
Communications, vol. 12, no. 1, pp. 288–300, 2013.
[25] ——, “Wireless information and power transfer: a dynamic power
splitting approach,” IEEE Transactions on Communications, vol. 61,
no. 9, pp. 3990–4001, 2013.
[26] X. Zhou, R. Zhang, and C. K. Ho, “Wireless information and power
transfer: architecture design and rate-energy tradeoff,” accepted by
IEEE Transactions on Communications.
[27] K. Huang and E. G. Larsson, “Simultaneous information-and-power
transfer for broadband downlink systems,” accepted by IEEE Transactions on Signal Processing.
[28] H. Ju and R. Zhang, “Throughput maximization for wireless powered
communication networks,” accepted by IEEE Transactions on Wireless
Communications.
[29] P. Popovski, A. Fouladgar, and O. Simeone, “Interactive joint transfer
of energy and information,” IEEE Transactions on Communications,
vol. 61, no. 5, pp. 2086–2097, 2013.
[30] D. W. K. Ng, E. S. Lo, and R. Schober, “Wireless information and
power transfer: energy efficiency optimization in OFDMA systems,”
accepted by IEEE Transactions on Wireless Communications.
[31] I. Krikidis, S. Timotheou, and S. Sasaki, “RF energy transfer for
cooperative networks: data relaying or energy harvesting?” IEEE
Communications Letters, vol. 16, no. 11, pp. 1772–1775, 2012.
[32] J. Park and B. Clerckx, “Joint wireless information and energy transfer
in a two-user MIMO interference channel,” IEEE Transactions on
Wireless Communications, vol. 12, no. 8, pp. 4210–4221, 2013.
[33] Z. Xiang and M. Tao, “Robust beamforming for wireless information
and power transmission,” IEEE Wireless Communications Letters,
vol. 1, no. 4, pp. 372–375, 2012.
[34] X. Chen, C. Yuen, and Z. Zhang, “Wireless energy and information
transfer tradeoff for limited feedback multi-antenna systems with
energy beamforming,” accepted by IEEE Transactions on Vehicular
Technology.
[35] A. Fouladgar and O. Simeone, “On the transfer of information and
energy in multi-user systems,” IEEE Communications Letters, vol. 16,
no. 11, pp. 1733–1736, 2012.
[36] ——, “Information and energy flows in graphical networks with energy
transfer and reuse,” IEEE Wireless Communications Letters, vol. 2,
no. 4, pp. 371–374, 2013.
[37] R. Zhang, L.-L. Yang, and L. Hanzo, “Generalised pre-coding aided
spatial modulation,” IEEE Transactions on Wireless Communications,
vol. 12, no. 11, pp. 5434–5443, 2013.
[38] R. Mesleh, H. Haas, S. Sinanovic, C. W. Ahn, and S. Yun, “Spatial
modulation,” IEEE Transactions on Vehicular Technology, vol. 57,
no. 4, pp. 2228–2241, 2008.
[39] M. D. Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and L. Hanzo,
“Spatial modulation for generalized MIMO: Challenges, opportunities
and implementation,” Proceedings of IEEE. [Online]. Available:
http://eprints.soton.ac.uk/354175/
[40] C. Peel, B. Hochwald, and A. Swindlehurst, “A vector-perturbation
technique for near-capacity multiantenna multiuser communicationPart I: channel inversion and regularization,” IEEE Transactions on
Communications, vol. 53, no. 1, pp. 195–202, 2005.
[41] Q. Spencer, A. Swindlehurst, and M. Haardt, “Zero-forcing methods
for downlink spatial multiplexing in multiuser MIMO channels,” IEEE
Transactions on Signal Processing, vol. 52, no. 2, pp. 461–471, 2004.
[42] H. Urkowitz, “Energy detection of unknown deterministic signals,”
Proceedings of the IEEE, vol. 55, no. 4, pp. 523–531, 1967.
[43] R. Zhang, L.-L. Yang, and L. Hanzo, “Error probability and
capacity analysis of generalised pre-coding aided spatial modulation,”
submitted to IEEE Transactions on Wireless Communications. [Online].
Available: http://eprints.soton.ac.uk/id/eprint/360422
[44] J. Proakis and M. Salehi, Digital Communications, 5th ed. McGrawHill Higher Education, 2008.
[45] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed.
Wiley-Interscience, 2006.
[46] D. Gesbert, “Robust linear MIMO receivers: a minimum error-rate
approach,” IEEE Transactions on Signal Processing, vol. 51, no. 11,
pp. 2863–2871, 2003.
[47] X. Zhang, D. Palomar, and B. Ottersten, “Statistically robust design of
linear MIMO transceivers,” IEEE Transactions on Signal Processing,
vol. 56, no. 8, pp. 3678–3689, 2008.
[48] J. Kermoal, L. Schumacher, K. Pedersen, P. Mogensen, and F. Frederiksen, “A stochastic MIMO radio channel model with experimental
validation,” IEEE Journal on Selected Areas in Communications,
vol. 20, no. 6, pp. 1211–1226, 2002.
[49] S. Wagner, R. Couillet, M. Debbah, and D. Slock, “Large system
analysis of linear precoding in correlated MISO broadcast channels
under limited feedback,” IEEE Transactions on Information Theory,
vol. 58, no. 7, pp. 4509–4537, 2012.
[50] R. Couillet and M. Debbah, Random Matrix Methods for Wireless
Communications. New York, NY, USA: Cambridge University Press,
2011.
[51] R. Zhang, Y. Cui, H. Claussen, H. Haas, and L. Hanzo, “Anticipatory
association for indoor visible light communications: Light, follow me
!” IEEE Transactions on Wireless Communications, submitted.
[52] R. Zhang, M. Biagi, L. Lampe, T. D. Little, S. Mangold, and Z. Xu,
“Localisation, communication and networking with VLC: Challenges
and opportunities,” IEEE Journal on Selected Areas in Communications, submitted.
[53] R. Zhang, A. F. A. Rawi, L. D. Humphrey, and L. Hanzo, “Expanded
constellation mapping for enhanced far-end-cross-talk cancellation in
G.fast,” IEEE Communications Letters, vol. 21, no. 1, pp. 56–59, Jan
2017.
[54] R. Zhang, L. L. Yang, and L. Hanzo, “Performance analysis of
non-linear generalized pre-coding aided spatial modulation,” IEEE
Transactions on Wireless Communications, vol. 15, no. 10, pp. 6731–
6741, Oct 2016.
[55] R. Zhang, H. Claussen, H. Haas, and L. Hanzo, “Energy efficient visible
light communications relying on amorphous cells,” IEEE Journal on
Selected Areas in Communications, vol. 34, no. 4, pp. 894–906, April
2016.
[56] R. Zhang, J. Wang, Z. Wang, Z. Xu, C. Zhao, and L. Hanzo, “Visible
light communications in heterogeneous networks: Paving the way for
user-centric design,” IEEE Wireless Communications, vol. 22, no. 2,
pp. 8–16, April 2015.
[57] R. Zhang, L. L. Yang, and L. Hanzo, “Energy pattern aided simultaneous wireless information and power transfer,” IEEE Journal on Selected
Areas in Communications, vol. 33, no. 8, pp. 1492–1504, Aug 2015.
[58] R. Zhang, R. G. Maunder, and L. Hanzo, “Wireless information and
power transfer: from scientific hypothesis to engineering practice,”
IEEE Communications Magazine, vol. 53, no. 8, pp. 99–105, August
2015.
[59] R. Zhang, L. L. Yang, and L. Hanzo, “Error probability and capacity
analysis of generalised pre-coding aided spatial modulation,” IEEE
Transactions on Wireless Communications, vol. 14, no. 1, pp. 364–
375, Jan 2015.
[60] R. Zhang and L. Hanzo, “Multi-layer modulation for intensitymodulated direct-detection optical OFDM,” IEEE/OSA Journal of
Optical Communications and Networking, vol. 5, no. 12, pp. 1402–
1412, Dec 2013.
[61] ——, “Advances in base- and mobile-station aided cooperative wireless
communications: An overview,” IEEE Vehicular Technology Magazine,
vol. 8, no. 1, pp. 57–69, March 2013.
[62] ——, “Cooperative downlink multicell preprocessing relying on
reduced-rate back-haul data exchange,” IEEE Transactions on Vehicular Technology, vol. 60, no. 2, pp. 539–545, 2011.
[63] ——, “Multiple-source cooperation: From code-division multiplexing
to variable-rate network coding,” IEEE Transactions on Vehicular
Technology, vol. 60, no. 3, pp. 1005–1015, March 2011.
[64] ——, “A unified treatment of superposition coding aided communications: Theory and practice,” IEEE Communications Surveys Tutorials,
vol. 13, no. 3, pp. 503–520, Third 2011.
[65] ——, “Superposition-aided delay-constrained hybrid automatic repeat
request,” IEEE Transactions on Vehicular Technology, vol. 59, no. 4,
pp. 2109–2115, May 2010.
[66] ——, “Wireless cellular networks,” IEEE Vehicular Technology Magazine, vol. 5, no. 4, pp. 31–39, 2010.
[67] R. Zhang, L. Xu, S. Chen, and L. Hanzo, “EXIT-chart-aided hybrid
multiuser detector for multicarrier interleave-division multiple access,”
IEEE Transactions on Vehicular Technology, vol. 59, no. 3, pp. 1563–
1567, March 2010.
[68] R. Zhang and L. Hanzo, “Interleaved random space time coding for
multisource cooperation,” IEEE Transactions on Vehicular Technology,
vol. 58, no. 4, pp. 2120–2125, May 2009.
[69] ——, “Superposition-coding-aided multiplexed hybrid ARQ scheme
for improved end-to-end transmission efficiency,” IEEE Transactions
on Vehicular Technology, vol. 58, no. 8, pp. 4681–4686, Oct 2009.
[70] ——, “Coding schemes for energy efficient multi-source cooperation
aided uplink transmission,” IEEE Signal Processing Letters, vol. 16,
no. 5, pp. 438–441, May 2009.
[71] ——, “Iterative multiuser detection and channel decoding for DSCDMA using harmony search,” IEEE Signal Processing Letters,
vol. 16, no. 10, pp. 917–920, Oct 2009.
[72] ——, “Three design aspects of multicarrier interleave division multiple
access,” IEEE Transactions on Vehicular Technology, vol. 57, no. 6,
pp. 3607–3617, Nov 2008.
[73] ——, “Space-time coding for high-throughput interleave division multiplexing aided multi-source co-operation,” Electronics Letters, vol. 44,
no. 5, pp. 367–368, Feb 2008.
[74] J. Zhang, R. Zhang, R. G. Maunder, S. Chen, and L. Hanzo, “Adaptive
coding and modulation for large-scale antenna array based aeronautical
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
[94]
[95]
[96]
communications in the presence of co-channel interference,” IEEE
Transactions on Wireless Communications, submitted.
Q. Wang, R. Zhang, L. L. Yang, and L. Hanzo, “Non-orthogonal multiple access: A unified perspective,” IEEE Wireless Communications,
submitted.
S. Gupta, R. Zhang, and L. Hanzo, “Energy harvesting aided device-todevice communication in the heterogeneous two-tier downlink,” IEEE
Transactions on Communications, submitted.
Y. Wang, Y. Xu, Y. Yang, X. Gao, B. Zhu, W. Cai, J. Yuan, R. Zhang,
and H. Zhu, “Simultaneous light emission and detection of InGaN/GaN
multiple quantum well diodes for in-plane visible light communication,” Optics Communications, (early access), 2016.
X. Li, Y. Huo, R. Zhang, and L. Hanzo, “User-centric visible light
communications for energy-efficient scalable video streaming,” IEEE
Transactions on Green Communications and Networking, vol. 1, no. 1,
pp. 59–73, March 2017.
S. Gupta, R. Zhang, and L. Hanzo, “Energy harvesting aided deviceto-device communication underlaying the cellular downlink,” IEEE
Access, (early access), 2016.
F. Jin, X. Li, R. Zhang, C. Dong, and L. Hanzo, “Resource allocation
under delay-guarantee constraints for visible-light communication,”
IEEE Access, vol. 4, pp. 7301–7312, 2016.
J. Wang, Y. Xu, X. Ling, R. Zhang, Z. Ding, and C. Zhao, “PAPR
analysis for OFDM visible light communication,” Optics Express,
vol. 24, no. 24, pp. 27 457–27 474, Nov 2016.
B. Li, R. Zhang, W. Xu, C. Zhao, and L. Hanzo, “Joint dimming control
and transceiver design for MIMO-aided visible light communication,”
IEEE Communications Letters, vol. 20, no. 11, pp. 2193–2196, Nov
2016.
S. Gupta, S. Kumar, R. Zhang, S. Kalyani, K. Giridhar, and L. Hanzo,
“Resource allocation for D2D links in the FFR and SFR aided cellular
downlink,” IEEE Transactions on Communications, vol. 64, no. 10, pp.
4434–4448, Oct 2016.
S. Gupta, R. Zhang, and L. Hanzo, “Throughput maximization for a
buffer-aided successive relaying network employing energy harvesting,” IEEE Transactions on Vehicular Technology, vol. 65, no. 8, pp.
6758–6765, Aug 2016.
J. Jiang, P. Zhang, R. Zhang, S. Chen, and L. Hanzo, “Aperture
selection for ACO-OFDM in free-space optical turbulence channel,”
IEEE Transactions on Vehicular Technology, vol. 65, no. 8, pp. 6089–
6100, Aug 2016.
S. Feng, X. Li, R. Zhang, M. Jiang, and L. Hanzo, “Hybrid positioning aided amorphous-cell assisted user-centric visible light downlink
techniques,” IEEE Access, vol. 4, pp. 2705–2713, 2016.
C. Zhu, Y. Huo, J. Jiang, H. Sun, C. Dong, R. Zhang, and L. Hanzo,
“Hierarchical colour-shift-keying aided layered video streaming for the
visible light downlink,” IEEE Access, vol. 4, pp. 3127–3152, 2016.
C. Zhu, Y. Huo, B. Zhang, R. Zhang, M. El-Hajjar, and L. Hanzo,
“Adaptive-truncated-HARQ-aided layered video streaming relying on
interlayer fec coding,” IEEE Transactions on Vehicular Technology,
vol. 65, no. 3, pp. 1506–1521, March 2016.
X. Li, F. Jin, R. Zhang, J. Wang, Z. Xu, and L. Hanzo, “Users first:
User-centric cluster formation for interference-mitigation in visiblelight networks,” IEEE Transactions on Wireless Communications,
vol. 15, no. 1, pp. 39–53, Jan 2016.
X. Li, R. Zhang, and L. Hanzo, “Cooperative load balancing in
hybrid visible light communications and WiFi,” IEEE Transactions on
Communications, vol. 63, no. 4, pp. 1319–1329, April 2015.
F. Jin, R. Zhang, and L. Hanzo, “Resource allocation under delayguarantee constraints for heterogeneous visible-light and RF femtocell,”
IEEE Transactions on Wireless Communications, vol. 14, no. 2, pp.
1020–1034, Feb 2015.
B. Li, J. Wang, R. Zhang, H. Shen, C. Zhao, and L. Hanzo, “Multiuser
MISO transceiver design for indoor downlink visible light communication under per-LED optical power constraints,” IEEE Photonics
Journal, vol. 7, no. 4, pp. 1–15, Aug 2015.
J. Jiang, R. Zhang, and L. Hanzo, “Analysis and design of threestage concatenated color-shift keying,” IEEE Transactions on Vehicular
Technology, vol. 64, no. 11, pp. 5126–5136, Nov 2015.
F. Wu, R. Zhang, L. L. Yang, and W. Wang, “Transmitter precodingaided spatial modulation for secrecy communications,” IEEE Transactions on Vehicular Technology, vol. 65, no. 1, pp. 467–471, Jan 2016.
J. Feng, R. Zhang, L. Hanzo, and S. X. Ng, “Cooperative medium
access control based on spectrum leasing,” IEEE Transactions on
Vehicular Technology, vol. 63, no. 1, pp. 297–307, Jan 2014.
J. Zhang, F. Jin, R. Zhang, G. Li, and L. Hanzo, “Analysis and design
of distributed antenna-aided twin-layer femto- and macrocell networks
[97]
[98]
[99]
[100]
[101]
[102]
[103]
[104]
[105]
[106]
[107]
[108]
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
[117]
relying on fractional frequency reuse,” IEEE Transactions on Vehicular
Technology, vol. 63, no. 2, pp. 763–774, Feb 2014.
F. Jin, R. Zhang, and L. Hanzo, “Fractional frequency reuse aided twinlayer femtocell networks: Analysis, design and optimization,” IEEE
Transactions on Communications, vol. 61, no. 5, pp. 2074–2085, May
2013.
J. Zhang, R. Zhang, G. Li, and L. Hanzo, “Distributed antenna
systems in fractional-frequency-reuse-aided cellular networks,” IEEE
Transactions on Vehicular Technology, vol. 62, no. 3, pp. 1340–1349,
March 2013.
L. Li, X. Zhou, R. Zhang, D. Zhang, and L. Hanzo, “Performance and
capacity analysis of poisson photon-counting based iter-PIC OCDMA
systems,” Optics Express, vol. 21, no. 22, pp. 25 954–25 967, Nov 2013.
X. Zhou, X. Zheng, R. Zhang, and L. Hanzo, “Chip-interleaved optical
code division multiple access relying on a photon-counting iterative
successive interference canceller for free-space optical channels,” Optics Express, vol. 21, no. 13, pp. 15 926–15 937, Jul 2013.
X. Zhou, D. Zhang, R. Zhang, and L. Hanzo, “A photon-counting
spatial-diversity-and-multiplexing MIMO scheme for poisson atmospheric channels relying on Q-ary PPM,” Optics Express, vol. 20,
no. 24, pp. 26 379–26 393, Nov 2012.
J. Zhang, R. Zhang, G. Li, and L. Hanzo, “Remote coalition network
elements for base station cooperation aided multicell processing,” IEEE
Transactions on Vehicular Technology, vol. 61, no. 3, pp. 1406–1415,
March 2012.
J. Feng, R. Zhang, and L. Hanzo, “A spectrum leasing cooperative
medium access protocol and its stability analysis,” IEEE Transactions
on Vehicular Technology, vol. 61, no. 8, pp. 3718–3730, Oct 2012.
X. Xu, R. Zhang, S. Ghafoor, and L. Hanzo, “Imperfect digitalfiber-optic-link-based cooperative distributed antennas with fractional
frequency reuse in multicell multiuser networks,” IEEE Transactions
on Vehicular Technology, vol. 60, no. 9, pp. 4439–4449, Nov 2011.
N. Bonello, R. Zhang, S. Chen, and L. Hanzo, “Channel code-division
multiple access and its multilevel-structured LDPC-based instantiation,” IEEE Transactions on Vehicular Technology, vol. 58, no. 5, pp.
2549–2553, Jun 2009.
——, “Reconfigurable rateless codes,” IEEE Transactions on Wireless
Communications, vol. 8, no. 11, pp. 5592–5600, November 2009.
X. Li, F. Jin, R. Zhang, and L. Hanzo, “Joint cluster formation and
user association under delay guarantees in visible-light networks,” in
2016 IEEE Global Telecommunications Conference, Dec (early access)
2016.
X. Li, R. Zhang, J. Wang, and L. Hanzo, “Cell-centric and user-centric
multi-user scheduling in visible light communication aided networks,”
in 2015 IEEE International Conference on Communications, June
2015, pp. 5120–5125.
J. Zhang, R. Zhang, G. Li, and L. Hanzo, “Coalition network elements
for base station cooperation,” in 2012 IEEE Vehicular Technology
Conference Fall, Sept 2012, pp. 1–5.
F. Jin, R. Zhang, and L. Hanzo, “Frequency-swapping aided femtocells
in twin-layer cellular networks relying on fractional frequency reuse,”
in 2012 IEEE Wireless Communications and Networking Conference,
April 2012, pp. 3097–3101.
J. Zhang, R. Zhang, X. Xu, G. Li, and L. Hanzo, “Effects of practical
impairments on cooperative distributed antennas combined with fractional frequency reuse,” in 2012 IEEE Wireless Communications and
Networking Conference, April 2012, pp. 1088–1092.
X. Xu, R. Zhang, and L. Hanzo, “Digital RoF aided cooperative
distributed antennas with FFR in multicell multiuser networks,” in 2011
IEEE Vehicular Technology Conference Fall, Sept 2011, pp. 1–5.
J. Feng, R. Zhang, and L. Hanzo, “Auction-style cooperative medium
access control,” in 2011 IEEE Vehicular Technology Conference Fall,
Sept 2011, pp. 1–5.
J. Feng, R. Zhang, S. X. Ng, and L. Hanzo, “Relay selection for energyefficient cooperative media access control,” in 2011 IEEE Wireless
Communications and Networking Conference, March 2011, pp. 287–
292.
R. Zhang and L. Hanzo, “Variable-rate network coding for multi-source
cooperation,” in 2011 IEEE Wireless Communications and Networking
Conference, March 2011, pp. 1517–1522.
R. Zhang, K. Giridhar, and L. Hanzo, “Distributed downlink multi-cell
processing requiring reduced-rate back-haul data exchange,” in 2011
IEEE Wireless Communications and Networking Conference, March
2011, pp. 1277–1281.
R. Zhang, X. Xu, and L. Hanzo, “Co-channel interference mitigation
capability of fixed relays connected by optical fibre,” in 2010 IEEE
Vehicular Technology Conference Fall, Sept 2010, pp. 1–5.
[118] X. Xu, R. Zhang, and L. Hanzo, “Imperfect radio over fibre aided
distributed antennas with fractional frequency reuse,” in 2010 IEEE
Vehicular Technology Conference Fall, Sept 2010, pp. 1–5.
[119] R. Zhang and L. Hanzo, “Joint and distributed linear precoding for
centralised and decentralised multicell processing,” in 2010 IEEE
Vehicular Technology Conference Fall, Sept 2010, pp. 1–5.
[120] ——, “Harmony search aided iterative channel estimation, multiuser
detection and channel decoding for DS-CDMA,” in 2010 IEEE Vehicular Technology Conference Fall, Sept 2010, pp. 1–5.
[121] ——, “Multiplexed hybrid ARQ for energy efficient transmissions
under delay constraints,” in 2010 IEEE International Conference on
Communications, May 2010, pp. 1–5.
[122] M. F. U. Butt, R. Zhang, S. X. Ng, and L. Hanzo, “Superposition coding
aided bi-directional relay transmission employing iteratively decoded
self-concatenated convolutional codes,” in 2010 IEEE Vehicular Technology Conference Spring, May 2010, pp. 1–5.
[123] R. Zhang and L. Hanzo, “Superposition-coding aided multiplexed
hybrid ARQ scheme for improved link-layer transmission efficiency,” in
2009 IEEE International Conference on Communications, June 2009,
pp. 1–5.
[124] N. Bonello, R. Zhang, S. Chen, and L. Hanzo, “Reconfigurable rateless
codes,” in 2009 IEEE Vehicular Technology Conference Spring, April
2009, pp. 1–5.
[125] R. Zhang and L. Hanzo, “Physical-layer algebraic network coding and
superposition coding for the multi-source cooperation aided uplink,” in
2009 IEEE Vehicular Technology Conference Spring, April 2009, pp.
1–5.
[126] ——, “High-throughput non-orthogonal interleaved random space-time
coding for multi-source cooperation,” in 2008 IEEE Global Telecommunications Conference, Nov 2008, pp. 1–5.
[127] N. Bonello, R. Zhang, S. Chen, and L. Hanzo, “Channel code division
multiple access and its multilevel structured LDPC based instantiation,”
in 2008 IEEE Vehicular Technology Conference Fall, Sept 2008, pp.
1–5.
[128] R. Zhang, L. Xu, S. Chen, and L. Hanzo, “Repeat accumulate code
division multiple access and its hybrid detection,” in 2008 IEEE
International Conference on Communications, May 2008, pp. 4790–
4794.
[129] L. Xu, R. Zhang, S. Chen, and L. Hanzo, “EXIT-chart aided hybrid
multiuser detector design for frequency-domain-spread chip-interleaved
MC-CDMA,” in 2008 IEEE Vehicular Technology Conference Spring,
May 2008, pp. 1816–1820.
[130] R. Zhang and L. Hanzo, “Interleave division multiplexing aided spacetime coding for high-throughput uplink cooperative communications,”
in 2008 IEEE Wireless Communications and Networking Conference,
March 2008, pp. 465–469.
[131] ——, “EXIT chart based joint code-rate and spreading-factor optimisation of single-carrier interleave division multiple access,” in 2007
IEEE Wireless Communications and Networking Conference, March
2007, pp. 735–739.
| 7cs.IT
|
Densely Entangled Financial Systems
arXiv:1402.5208v1 [q-fin.RM] 21 Feb 2014
Bhaskar DasGupta and Lakshmi Kaligounder
Abstract In [1] Zawadoski introduces a banking network model in which the asset
and counter-party risks are treated separately and the banks hedge their assets risks
by appropriate OTC contracts. In his model, each bank has only two counter-party
neighbors, a bank fails due to the counter-party risk only if at least one of its two
neighbors default, and such a counter-party risk is a low probability event. Informally, the author shows that the banks will hedge their asset risks by appropriate
OTC contracts, and, though it may be socially optimal to insure against counterparty risk, in equilibrium banks will not choose to insure this low probability event.
In this paper, we consider the above model for more general network topologies,
namely when each node has exactly 2r counter-party neighbors for some integer
r > 0. We extend the analysis of [1] to show that as the number of counter-party
neighbors increase the probability of counter-party risk also increases, and in particular the socially optimal solution becomes privately sustainable when each bank
hedges its risk to at least n/2 banks, where n is the number of banks in the network,
i.e., when 2r is at least n/2, banks not only hedge their asset risk but also hedge its
counter-party risk.
1 Introduction
Economic stability has received special attention during the past several years
mainly because of the economic downturn experienced globally in the recent past.
This attention has generated renewed interest in evaluating how important the ecoBhaskar DasGupta
Department of Computer Science, University of Illinois at Chicago, Chicago, IL 60607-7053, email: [email protected]
Lakshmi Kaligounder
Department of Computer Science, University of Illinois at Chicago, Chicago, IL 60607-7053, email: [email protected]
1
2
Bhaskar DasGupta and Lakshmi Kaligounder
nomic stability is as having an unstable economy can pave way to an economic crisis
each time when the global markets sees a downward trend. Financial instability and
its effects on the economy can be very costly due to its contagion or spillover effects to other parts of the economy and it is fundamental to have a sound, stable and
healthy financial system to support the efficient allocation of resources and distribution of risks across the economy. Financial risk management is a critical component
of maintaining economic stability. Hedging is a risk management option that protects the owner of an asset from loss. It is the process of shifting risks to the futures
market. The risks in the market must be first identified in order to manage the risk.
To identify the risk one must examine both the immediate risk (asset risk) as well as
the risk due to indirect effects (counter-party risk). Though hedging will minimize
overall profit when markets are moving positive, it also helps in reducing risk during
undesirable market conditions. However, as the owner hedges his/her asset risk to
protect against defaults, the owner also gets exposed to the counter-party risk.
In [1] Zawadoski introduces a banking network model in which the asset and
counter-party risks are treated separately, and showed that, under certain situations,
banks do not manage counter-party risk in equilibrium. In his model, each bank has
only two counter-party neighbors, a bank fails due to the counter-party risk only if
at least one of its two neighbors default, and such a counter-party risk is an event
with low probability. Informally, the author shows that the banks will hedge their
asset risks by appropriate OTC contracts, and, though it may be socially optimal
for banks to insure against counter-party risk, in equilibrium banks will not choose
to insure this low probability event. The OTC contract not only creates a contagion
but also creates externalities which undermines the incentives of the banks to avert
contagion. The model uses short term debt to finance their real asset. The failure in
this model is from the liability side, where the investors run on the banks when they
do not trust the bank, i.e., the investors do not roll over the debts of the banks. Hence
the contagion can be avoided only by increasing the equity and not by providing
liquidity.
In this paper, we consider the above model for more general network topologies,
namely when each node has exactly 2r counter-party neighbors for some integer
r > 0. We extend the analysis of [1] to show that as the number of counter-party
neighbors increase the probability of counter-party risk also increases1 , and in particular, the socially optimal solution becomes privately sustainable when each bank
hedges its risk to a sufficiently large number of other banks. The counter-party risk
can be hedged by holding more equity, buying default insurance on their counterparties or collateralizing OTC contracts. Since holding excess capital or collateralizing OTC contracts is a wasteful use of scarce capital [1], when the banks choose
to hedge their counter-party risk they buy the default insurance on their counterparties. More precisely, our conclusions for the general case of 2r neighbors are as
follows:
• All the banks will still decide to hedge their asset risks.
1 Thus, the owners will decide to hedge their counter-party risk thereby helping to contain an
economic crisis.
Densely Entangled Financial Systems
3
• If the number of counter-party neighbors is at least n/2, then all banks will decide to insure their counter-parties, and socially optimal solution in case of two
counter-parties for each bank now becomes privately optimal solution.
• In the limit when the number of banks n in the network tend to ∞, as the number
of counter-party neighbors approach n − 1, failure of very few of its counter-party
banks will not affect a bank.
2 Related Prior Research Works
As we have already mentioned, Zawadowski [1] introduced a banking model in
which asset risk and counter-party risk are treated separately, showed that banks
always prefer to hedge their asset risk using OTC contracts and also showed that
banks do not hedge their counter-party risk even though hedging counter-party risk
is possible and socially desirable. Allen and Gale [23] showed that interbank deposits help banks share liquidity risk but expose them to asset losses if their counterparty defaults. Their model cannot be used to understand the contractual choices in
case of OTC derivatives as they modeled the liquidity risk. Babus [18] proposed
a model in which links are formed between banks which serves as an insurance
mechanism to reduce the risk of contagion. Allen and Babus [12] pointed out that
graph-theoretic concepts provide a conceptual framework used to describe and analyze the banking network, and showed that more interbank links provide banks with
a form of coinsurance against uncertain liquidity flows. Gai and Kapadi [5] showed
that more interbank links increase the opportunity for spreading failures to other
banks during crisis. Several prior researchers such as [11, 12, 17, 21] commented
that graph-theoretic frameworks may provide a powerful tool for analyzing stability
of banking and other financial networks. Kleindorfer et al. [11] argued that network
analyses can play a crucial role in understanding many important phenomena in finance. Freixas et al. [24] explored the case of banks that face liquidity fluctuations
due to the uncertainty about consumers withdrawing funds. Iazzetta and Manna [13]
analysed the monthly data on deposit exchange to understand the spread of liquidity
crisis using network topology.
Babus [20] studied how the trade-off between the benefits and the costs of being linked changes depending on the network structure and observed that, when the
network is maximal, liquidity can be redistributed in the system to make the risk of
contagion minimal. Corbo and Demange [6] explored the relationship of the structure of interbank connections to the contagion risk of defaults given the exogenous
default of set of banks. Nier et al. [17] explored the dependency of systemic risks
on the structure of the banking system via network theoretic approach and the resilience of such a system to contagious defaults. Haldane [13] suggested that contagion should be measured based on the interconnectedness of each institution within
the financial system. Liedorp et al. [8] investigated if interconnectedness in the interbank market is a channel through which banks affect each others riskiness, and
argued that both large lending and borrowing shares in interbank markets increase
4
Bhaskar DasGupta and Lakshmi Kaligounder
the riskiness of banks active in the dutch banking market. Kiyotaki and Moore [26]
studied the chain reaction caused by the shock in one firm and the financial difficulties in other firms due to this chain reaction. Acharya and Bisin [14] compared
centralized markets to OTC markets and showed that counter-party risk externalities can lead to excessive default and production of aggregate risk. Caballero and
Simsek [16] concluded that OTC derivatives are not the sole reason for the inefficiency of financial networks. Pirrong [15] argued that central counter-parties (CCP)
can also increase the systemic risk under certain circumstances and hence the introduction of CCP will not guarantee to mitigate the systemic risk. Zawadowski [7]
showed that complicated interwoven financial intermediation can be a reason for
inefficient financial networks, and hence OTC are not the only reason for financial
instability. Stulz [9] showed that exchange trading has both benefits and costs compared to OTC trading, and argued that credit default swaps (CDS) did not cause
the crisis since they worked well during much of the first year of the crisis. Zhu
and Pykhtin [19] showed that modeling credit exposures is vital for risk management application, while modelling credit value adjustment (CVA) is necessary step
for pricing and hedging counter-party credit risk. Corbo and Demange [6] showed
that introduction of central clearing house for credit default swaps will mitigate the
counter-party risk. Gandhi et al. [4] paralleled and complemented the conclusion
of [6], i.e., the creation of central clearing house for CDS contracts may not reduce
the counter-party risk.
3 The basic model
The model has n > 3 banks and three time periods t = 0, 1, 2 termed as initial, interim
and final, respectively. Each bank has exactly 2r counter-party neighbors for some
integer r > 0 (see Fig. 1 for an illustration). The unit investment of each bank in
|X
i−2r |
|X
i−r |
εk at t = 2, where R =
εk −
the long term real asset yields a return of R +
(
k = | i−1 |
k = | i−(r+1)|
RH ,
if the project succeeds
, and each εk is realized at t = 2 taking values of u
RL < RH , if the project fails
or −u each with probability 1/2. For each unit investment made by the bank at t = 0,
the investor lends D ≥ 0 as short term debt and equity 1 − D ≥ 0 is the bank’s share.
The short term debt has to be rolled over at time period t = 1 for the banks to operate
successfully. Thus the debt holders have an option to withdraw funding and force
the bank to liquidate the real project.
Let e ∈ {0, 1} be the unobservable effort choice such that a bank needs to exert
an effort of e = 1 at both time period t = 0 and t = 1 for the project to be successful
(i.e., R = RH ). At t = 1 the project can be in one of the two states: a “bad” state
with probability p or a “good” state with probability 1 − p, irrespective of the effort
exerted by the bank. At a “bad” state the project of one of the randomly chosen bank
fails and delivers RL , even if e = 1 at both time periods t = 0 and t = 1. Unless the
Densely Entangled Financial Systems
5
˜b 4
˜b 3
˜b 2
˜b 5
˜b 1
˜b 6
˜b 0
˜b 15
˜b 7
˜b 8
˜b 14
˜b 9
˜b 13
˜b 10
˜b 11
˜b 12
Fig. 1 An illustration of a network of 16 banks ˜b0 , ˜b1 , . . ., ˜b15 with r = 3. Each ˜bi is connected to
the banks ˜b i−3 (mod 16) , ˜b i−2 (mod 16) , ˜b i−1 (mod 16) , ˜b i+1 (mod 16) , ˜b i+2 (mod 16) and ˜b i+3 (mod 16) .
bank demand collateral from its counter-parties, if the bank defaults at t = 1 then all
the hedging liabilities of the defaulted bank gets cancelled, the investors liquidate
the bank and take equal share of L (the value of the bank when it is liquidated). If
the bank survives till t = 2 and the counter-party risk gets realized then, the bank has
to settle the counter-party hedging contract before paying its debt.
We use the following notations for four specific values of the probability of bad
state p:
psoc : if p < psoc , then irrespective of the number of counter-party neighbors there
is no need for counter-party insurance even from social perspective.
pind : if p > pind then the banks will not buy counter-party insurance as the private
benefits of insuring exceeds the cost.
pterm : if p < pterm then the banks will continue to prefer short term debt.
paut : if p < paut then no bank will have an incentive to hold more equity and borrow less.
Parameter Restrictions and Assumptions
The following parameter restriction are adopted from [1] to make them consistent
for a network model with 2r counter-parties. B is the banks’ private benefit with the
6
Bhaskar DasGupta and Lakshmi Kaligounder
subscript representing the specific time period and X denotes the additional nonpledgable payoff. Inequality (1) ensures that the investors will choose to roll over
the debt at t = 1 when the project is expected to succeed (i.e., R = RH ), and the
investors will decide to liquidate the bank at t = 1 if the bank’s project is expected
to fail (i.e., R = RL ). Inequality (2) implies that it is socially optimal to exert effort.
Inequality (3) ensures that banks have to keep positive equity to overcome moral
hazard. Inequality (4) ensures that, counter-party risk of the bank is large enough
to lead to contagion but small enough that the bank does not want to engage in
risk-shifting.
p
R H + X − B1
(1)
RL < L < 1 −
n
2B1
1 − np
(2)
B1 ≥ R H − 1 + X
(3)
RH − RL >
B1 − X < u <
r
X
B1 22r −
2r!
K! 2r − K !
K=0
r
X
2r − 2K
2r
K!
2r − K !
K=0
1 − 1 − p R H + X − B1 − p L
β > 1/2 and β >
1 − p B1
2u ≤ B0 < 1 − p B1
(4)
(5)
(6)
4 Our Results
Our results imply that when the number of counter-party neighbors is at least n/2,
the socially optimal outcome become privately sustainable.
n
o
Theorem 1. If the probability of bad state is p ∈ [ 0, p⋆ , where p⋆ = min pind , paut , pterm
then the followings hold.
(a) Banks endogenously enter into OTC contracts as shown by Zawadowski in [1].
(b) Banks borrow D < 1 for short term at t = 0 at an interest rate Rı,0 > 1 and Rı,1 = 1
as shown by Zawadowski in [1].
(c) In a bad state, failure of a single bank leads to run on all banks in the system
only when 2r < n/2.
(d) If a bank loses at least r counter-parties, it needs a debt reduction of I = ru −
B1 + X > 0.
(e) The contagious equilibrium stated by Zawadowski in [1] exist only if 2r < n/2.
When 2r ≥ n/2 banks insure against counter-party risk using default insurance.
Densely Entangled Financial Systems
7
(f) If p ∈ psoc , pind then the socially optimal outcome is sustainable in equilibrium.
Theorem 1(a) is proved in Lemma 5, whereas Theorem 1(c),(e),(f) are showed
in Lemma 6. Theorem 1(d) follows from the derivations of parameter values as described in Section 4.1; these derivations follow from the work in [1] and are provided
in the appendix. Theorem 1(b) uses the same proof as that in [1] and is therefore
omitted.
4.1 Proofs of Theorem 1
The derivations of the parameters and their values described in items (I) – (VIII)
below follow from the work in [1] since they are not affected by changing the number of counter-party neighbors from 2 to 2r. For the sake of completeness, these
derivations are provided in the appendix.
(I) The maximum amount of borrowing at t = 0 that can be roll over at t = 1 is
given by:
R H − B1 + X
<1
Dmax Rı,0 =
Rı,0
and the expected payoff of the above bank borrowing Dmax is B1 .
(II) In a stable system where all banks buy counter-party insurance, the price per
unit of default insurance is
ssafe =
1−β
p
−β
n
n
(where the superscript “safe” denotes the insured system).
(III) In a stable system where all banks buy counter-party insurance, the amount
borrowed by the bank for unit investment in the project is
p
p
Dsafe = 1 −
R H − B1 + X + L
n
n
(IV) In a contagious system where banks decide not to buy insurance, the amount
borrowed by the bank for unit investment in the project is
D ⋆ = 1 − p R H − B1 + X + p L
(V) In a contagious system where banks decide not to buy insurance, the interest
rate for the amount borrowed is
⋆
Rı,0
=
1
L
1− p 1−
R H − B1 + X
!
8
Bhaskar DasGupta and Lakshmi Kaligounder
n
term
(VI) p
= 1−β
n − 1
B0
.
RL + βB1 − L
R H + X − L − B1 1 − β +
n−1
1 − β R H + X − B1 − R L
f.aut
.
(VII) p
=
β
R H + X − L − B1 1 − β − R H − R L
n
n
o
(VIII) paut = min ps.aut , pr.aut , pf.aut , where the superscripts s.aut, r.aut and f.aut stand for
safe autarky, risky autarky and full autarky, respectively.
I 1−β
Lemma 1. pind =
β B1
(Informally,
Lemma 1 states that the probability of bad state must be at least pind =
I 1−β
β B1 for a bank’s benefit to outweigh the cost of its counter-party insurance).
Proof. The private incentive of a single bank to stay insured instead of deviating
from counter-party insurance is when the following holds (in the left-hand side of
the inequality, the first term is the payoff if no bank defaults, the second term is
the payoff if any of the counter-parties of the bank default, the third term is money
invested to buy the insurance on the bank’s counter-parties and the fourth term is the
bank’s equity):
≡
≡
≡
≡
benefits from all banks being insured ≥ benefits from all but one bank not being insured
!
1 + 2r p
2rp
B1 + β
B1 + I − ssafe 2rI − 1 − Dsafe
β 1−
n
!n
(1 + 2r).p
≥ β 1−
B1 − 1 − Dsafe
n
!
safe
1−β
2rp
2rp
p
β
B1 + I − s
2rI ≥ 0 ≡ β
B1 + I −
2rI ≥ 0
+β
n
n
n
n
2βrpB1 2βrpI 2rI 2rβI 2βrpI
+
−
+
−
≥ 0 ≡ βrpB1 − rI + rβI ≥ 0
n
n
n
n
n
I 1−β
I 1−β
βrpB1 ≥ rI − rβI ≡ βpB1 ≥ I − βI ≡ p ≥
=⇒ pind =
βB1
βB1
⊔
⊓
Lemma 2. psoc =
2Ir 1 − β
.
n − 1 R H + X − L − B1 1 − β
(Informally, Lemma 2 states that the probability of bad state must be at most psoc =
2Ir (1−β)
(n−1) (RH +X−L−B1 (1−β)) for the social benefits to outweigh the social cost of its counterparty insurance).
Proof. The social benefits outweigh the social cost of the counter-party insurance
of the system if and only if the following holds:
Densely Entangled Financial Systems
9
benefits from all banks being insured ≥ benefits from no bank being insured
!
1 + 2r p
2rp
≡ β 1−
B1 + β
B1 + I − ssafe 2rI − 1 − Dsafe ≥ β 1 − p B1 − 1 − D ⋆
n
n
!
pL
p
1−β βp
βB1 p 2βB1 p r 2βB1 p r 2βI p r
2rI − 1 + 1 −
R H + X − B1 +
−
+
+
−
+
≡ βB1 −
n
n
n
n
n
n
n
n
≥ β B1 1 − p − 1 + 1 − p (RH + X − B1 ) + p L
≡ βB1 −
βB1 p 2βI p r 2rI 2βrI 2βI p r
pX pB1 p L
+
−
+
−
− 1 + RH + X − B1 − pRn H −
+
+
n
n
n
n
n
n
n
n
≥ βB1 − pβB1 − 1 + RH + X − B1 − pRH − pX + pB1 + p L
βB1 p 2rI 2β r I p RH p X p B1 p L
−
+
−
−
+
+
≥ −p βB1 − p RH − p X + p B1 + p L
n
n
n
n
n
n
n
2RI 2βRI
βB1 p p RH p X p B1 p L
−
−
+
+
+ p βB1 + p RH + p X − p B1 − p L ≥
−
−
n
n
n
n
n
n
n
p − βB1 − RH − X + B1 + L − nβB1 + nRH + nX − nB1 − nL ≥ 2Ir 1 − β
p βB1 (n − 1) + RH (n − 1) + X(n − 1) − B1(n − 1) − L(n − 1) ≥ 2Ir 1 − β
2Ir 1 − β
2Ir 1 − β
p ≥
≡ p ≥
n − 1 βB1 + RH + X − B1 − L
n − 1 RH + X − L − B1(1 − β)
≡−
≡
≡
≡
≡
and thus p
soc
2Ir 1 − β
=
.
n − 1 RH + X − L − B1(1 − β)
⊔
⊓
Corollary 1. If each bank has only two counter-parties then r = 1 and thus psoc =
2I(1 − β)
.
n − 1 RH + X − L − B1(1 − β)
Lemma 3 (Probability of failure in risky autarky).
r
1
2r! X
2r
2 K=0 K! (2r − K)!
n
pr.aut = (1 − β)
X
r
n−1
(nβ − 1)X
1
2r! X
RH + X − L − B1(1 − β) −
1 − 2r
n−1
2 K=0 K! (2r − K)!
1−
(An informal explanation of Lemma 3 is as follows. If a bank has enough equity to survive even if all of its counter-parties collapse then the bank may be
forced to borrow less and this is not profitable. But, if the bank decides to deviate to risky autarky then it cannot pay back its debt if the real project delivers
RH − 2ru. The probability of failure
if the bank chooses risky autarky is given by
P
n
X
pr.aut = (1 − β) n−1
r
1
K=0 K! (2r−K)!
Pr
(nβ−1)X
1
RH +X−L−B1 (1−β)− n−1 1− 22r!
2r
K=0 K! (2r−K)!
1− 2r!
2r
2
).
Proof. For a bank to be in risky autarky, they borrow less such that it can roll over
the debt at t = 1 if its neighbors collapse but need not survive at t = 2. The relevant
incentive constraint is:
10
Bhaskar DasGupta and Lakshmi Kaligounder
RH +
r
X
1
2r! X
r.aut
− Rr.aut
≥ B1
ı,0 D
22 r K=0 K! (2r − K)!
r.aut
≡ Rr.aut
≤ R H − B1 +
ı,0 D
r
1
2r! X X
22 r K=0 K! (2r − K)!
The break-even condition for investors at t = 0 is:
p
p
Dr.aut = L + 1 − Rr.aut
Dr.aut
n
n ı,0
r
p
p
p 2r! X X
p
1
≡ Dr.aut = L + RH − B1 − RH + B1 + 1 −
n
n
n
n 22r K=0 K! (2r − K)!
A bank decides not to deviate to risky autarky in a contagious system if and only if
the following holds:
payoff in contagious system ≥ payoff in risky autarky
n−1
r.aut r.aut
≡ β (1 − p)B1 − 1 − D ⋆ ≥ β (1 − p) RH + X − Rı,0
D
+β
p B1 − 1 − Dr.aut
n
≡ βB1 − p βB1 − 1 + RH + X − B1 − p RH − p X + p B1
!
2r! X Pr
n−1
p
1
≥ β − p β RH + X − RH + B1 − 2r
+β
p B1 − 1 + L
K=0
K!
(2r
−
K)!
n
n
2
1
p
p
p 2r! X Pr
+ R H − B1 − R H + B1 + 1 −
K=0
2r
n
n
K! (2r − K)!
!n 2
2r! Pr
1
n
(1 − β)X 1 − 2r K=0
≡
n−1
K! (2r − K)!
2
"
!#
(nβ − 1)X
2r! P
1
≥ p RH + X − L − B1(1 − β) −
1 − 2r rK=0
n−1
K! (2r − K)!
2
which implies
r
1
2r! X
1 − 2r
2
K!
(2r
− K)!
n
K=0
pr.aut = (1 − β)
X
r
n−1
(nβ − 1)X
1
2r! X
RH + X − L − B1(1 − β) −
1 − 2r
n−1
2 K=0 K! (2r − K)!
⊔
⊓
Corollary 2. If the number of neighbors is 2 then r = 1 and thus
!
2 1 1
+
1−
4 2 1
n
!!
X
pr.aut = (1 − β)
n−1
2 1 1
nβ − 1
X 1−
+
RH + X − L − B1(1 − β) −
n−1
4 2 1
Densely Entangled Financial Systems
n
X
≡ pr.aut = (1 − β)
n−1
11
1−
3
4
nβ − 1
X 1 − 43
RH + X − L − B1(1 − β) −
n−1
X
n
4
!
≡ pr.aut = (1 − β)
n−1
nβ − 1 X
RH + X − L − B1(1 − β) −
n−1
4
Lemma 4 (Probability of failure in safe autarky).
n
2ru + X − B1
ps.aut = (1 − β)
1−β
n−1
RH + X − L − B1 + βB1 +
X + 2ru − B1
n−1
(An informal explanation of Lemma 4 is as follows. If a bank chooses to deviate to
safe autarky, it will survive unless it is directly affected by low return RL . It can pay
back its debt even if the real project delivers RH −2ru.
of failure if the
probability
The
2ru+X−B1
n
).
bank chooses safe autarky is given by ps.aut = (1−β) n−1
1−β
RH +X−L−B1 +βB1 + n−1 X+2ru−B1
Proof. Suppose that a bank survives at t = 2 even if all of its counter-party risks
are realized to ensure that the payoff of the real project is RH − 2ru. Since the bank
needs enough equity to survive even if real project yields RH − 2ru, we have
s.aut
s.aut
RH − 2ru − Rs.aut
≥ 0 ≡ Rs.aut
≤ RH − 2ru
ı,0 D
ı,0 D
The break-even condition for investors at t = 0 is:
p
p s.aut s.aut p
p
Rı,0 D + L ≡ Ds.aut = 1 −
Ds.aut = 1 −
RH − 2ru + L
n
n
n
n
Banks do not deviate to safe autarky from contagious system if and only if the
following holds:
payoff in contagious system ≥ payoff in safe autarky
p
≡ β 1 − p B1 − 1 − D ⋆ ≥ β 1 −
RH + X − Rs.aut
Ds.aut − 1 − Ds.aut
ı,0
n
≡ βB1 − pβB1 − 1 + RH + X − B1 − p RH − p X + p B1 + p L
p
pβ
p
p
≥ β−
RH + X − RH + 2ru − 1 + RH − 2ru − RH + 2ru + L
n
n
n
n
≡ βB1 − pβB1 + X − B1 − pRH − p X + p B1 + p L
p
pβRH pβX pβRH 2pβru
p
p
≥ βRH + βX − βRH + 2βru −
−
+
−
− 2ru − RH + 2ru + L
n
n
n
n
n
n
n
!
p
nβB1 + nRH + nX − nB1 − nL − βX − 2rβu − RH + 2ru + L
≡ βB1 + X − B1 − βX − 2βru + 2ru ≥
n
≡ (1 − β)X − (1 − β)B1 + 2ru(1 − β)
12
Bhaskar DasGupta and Lakshmi Kaligounder
!
p
nβB1 + RH (n − 1) + nX − nB1 − L(n − 1) − βX + 2ru(1 − β) + X − X + B1 − B1 + βB1 − βB1
n
≡ (1 − β) n 2ru + X − B1
!
≥ p βB1 (n − 1) + RH(n − 1) + X(n − 1) − B1(n − 1) − L(n − 1) + X(1 − β) + 2ru(1 − β) − B1(1 − β)
!
n
1−β
≡ (1 − β)
(2Ru + X − B1 ) ≥ p βB1 + RH + X − B1 − L +
(X + 2ru − B1)
n−1
n−1
n
2ru + X − B1
≡ p ≤ (1 − β)
1−β
n−1
X + 2ru − B1
RH + X − L − B1 + βB1 +
n−1
n
2ru + X − B1
This implies ps.aut = (1 − β)
.
1−β
n−1
RH + X − L − B1 + βB1 +
X + 2ru − B1
n−1
⊔
⊓
≥
Corollary 3. If the number of neighbors is 2 then r = 1 and thus
2u + X − B1
n
ps.aut = 1 − β
1−β
n−1
RH + X − L − B1 + βB1 +
X + 2u − B1
n−1
Lemma 5. In equilibrium banks hedge all of its counter-party risks i.e. banks endogenously enter into OTC contracts.
Proof. The banks hedge all of its counter-party risks if and only if
payoff from hedging in contagious equilibrium > payoff from not hedging
≡ β 1 − p B1
2r! B1 + 2ru
B
+
(2r
−
2)u
B
+
(2r
−
4)u
B
+
(2r
−
2r)u
1
1
1
!
+
+
+ ···+
> β 1 − p 2r
1! (2r − 1)!
2! (2r − 2)!
2 0! (2r − 0)!
2r
2r
!
! 2r −
2
2
B1 + (2r − 2)u B1 + (2r − 4)u
B1 + (2r − 2r)u
2r! B1 + 2ru
!
+
+
+ ···+
≡ B1 > 2r
1! (2r − 1)!
2! (2r − 2)!
2 0! (2r − 0)!
2r
2r
!
! 2r −
2
2
r
X
2r − 2K
2ur
K! (2r − K)!
K=0
≡ B1 >
r
X
2r!
22r −
K! (2r − K)!
K=0
which satisfies inequality (4), and thus banks hedge all of its counter-party risks.
Densely Entangled Financial Systems
13
Let the returns from the successful project be RH +
|X
i−r |
k = | i−1 |
|X
i−2r |
εk .
k = | i−(r+1)|
εk −
Assume
that the bank goes bankrupt at t = 2 if the counter-party realization of its unhedged
risk is −u. This is true if the bank cannot repay its debt at t = 2, i.e.,
⋆
D ⋆ ≡ R H − u < R H − B1 + X ≡ u > B1 − X
RH − u < Rı,0
if a bank fails when it loses −u on its unhedged counter-party exposure, it will fail
when the loss is greater than −u on its unhedged counter-party exposure.
⊔
⊓
Corollary 4. If number of counter-party
neighbors is 2 (i.e., r = 1), then
!
2−2
2−0
+
2u
0! (2 − 0)! 1! (2 − 1)!
! ≡ B1 > 2u
B1 >
and
u > B1 − X.
2!
2!
22 −
+
0! (2 − 0)! 1! (2 − 1)!
Lemma 6. When the number of neighboring counter-parties is less than n/2, a bank
chooses to shirk if one or more of its counter-parties default by leaving the bank
unhedged resulting in its debt not being rolled over at t = 1.
Proof. If a bank borrows Dmax (Rl,0 ) at t = 0 and if the bank has a low expected
realization of RL , then the debt financing is not rolled over at t = 1. Since RL < L,
the creditors will want to terminate the project. The bank goes bankrupt if its debt
financing is not rolled over at t = 1.
If 2r, the number of counter-party neighbors, is less than n/2, then probability of
n
failure due to counter-party risk is less than /n2 , i.e., the probability of counter-party
risk is less than 1/2. Since we assume that banks will consider a counter party risk
probability of at least 1/2 to insure against counter-party risk, banks do not insure
against counter-party risk using default insurance when 2r < n/2. When 2r ≥ n/2, the
probability of counter-party risk becomes at least 1/2 and hence banks will hedge the
counter-party risk. The counter-party insurance payoff happens with probability 2rp
n
in case of social perspecin case of private perspective and with probability (n−1)p
n
tive. Thus, as 2r increases to n − 1, the counter-party insurance payoff probability in
private perspective becomes the same as that in social perspective.
When 2r ≥ n/2, the individual banks will hedge the counter-party risk by taking
counter-party insurance. When 2r < n/2, banks will not hedge the counter-party risk
and hence failure of a counter-party will lead to the violation of its incentive constraint, thus the bank shirks and the project delivers RL . Let D1 be the amount of
debt to be rolled over at t = 1. The investors will demand higher interest rate Rı,1 in
order to break even. Let P s be the probability of a bank that do not default, P f be
the probability of a bank that defaults, and nd be the number of neighbors of any
2r − nd
nd
bank that default. Thus, P s =
and P f = . By the break even condition of
2r
2r
investors, we get
P s Rı,1 D1 + P f RL = D1 ≡ Rı,1 D1 =
D1 − P f R L
Ps
14
Bhaskar DasGupta and Lakshmi Kaligounder
The incentive constraint is
β(RH − Rı,1 D1 + X) ≥ βB1
!
D1 − P f R L
+ X ≥ βB1
≡ β RH −
Ps
≡ P s R H − D1 + P f R L + P s X ≥ B 1 P s
≡ P s R H − D1 + P f R L + P s X ≥ B 1 P s
≡ P s R H + X − B 1 + P f R L ≥ D1
If the bank had originally borrowed D0 , the amount it has to roll over at t = 1 is
Rı,0 D0 = RH + X − B1 , but the amount that is actually getting rolled over is only
P s (RH + X − B1) + P f RL .
⊔
⊓
Corollary 5. If number of counter-party neighbors is 2 (i.e., 2r = 2) and number of
nd
d
1
1
banks defaulted is 1 (i.e., nd = 1), then P s = 2r−n
2r = /2, P f = 2r = /2, and thus
1
1
P s R H + X − B 1 + P f R L ≥ D1 ≡
R H + X − B 1 + R L ≥ D1
2
2
Appendix 1: Remaining Proofs
(I) Proof of Dmax =
RH −B1+X
Rı,0
<1
The incentive constraint of a bank surviving at t = 1 and holding no risk can be
written as follows:
payoff if bank exerts effort ≥ payoff if bank does not exert effort
R H − B1 + X
R H − B1 + X
≡ β[RH − Rı,0 D + X] ≥ βB1 ≡
≥ D ≡ Dmax =
<1
Rı,0
Rı,0
|
{z
}
since Rı,0 ≥ 1
and, by inequality (3), RH + X − B1 ≤ 1
(II) Proof of ssafe =
1−β
n
− β np
Assume all counter parties are insured. The insurance company has to hold a capital
of 2RI. If all the banks insure against the failure of their counter-parties, the insurance fund breaks even if the price per unit of the default insurance s is determined
by the break even condition. Using the superscript safe to denote the insured system,
we get
Densely Entangled Financial Systems
15
expected amount that remains in the insurance fund at t = 2
= amount that insurance fund has to set aside at t = 0
1−β
p
≡ β(2rI − 2prI) = 2rI − 2nrI s ≡ β 1 − p = 1 − ns ≡ ssafe =
−β
n
n
(III) Proof of Dsafe = 1 − np (RH − B1 + X) + np L
p
p
Since the investors break even, we have 1 − Rsafe Dsafe + L = Dsafe . Combining
n
n
with the maximum amount that can be borrowed, we get
R H − B1 + X
=⇒ Rsafe Dsafe = RH − B1 + X
Rı,0
Dmax =
which proves our claim.
(IV) Proof of D⋆ = (1 − p)(RH − B1 + X) + p L
If all the banks decide not to buy insurance and hold the minimum amount of equity
to roll over debt, the system is contagious. The investors will always anticipate the
equilibrium and break even, implying
(1 − p) R⋆ı,0 D⋆ + p L = D⋆ ≡ D⋆ = (1 − p)(RH − B1 + X) + p L
|
{z
}
since by (I) R⋆ı,0 D⋆ = RH − B1 + X
(V) Proof of R⋆
ı,0 =
1
1−p 1− R
L
H −B1 +X
R H − B1 + X
R H − B1 + X
≡ R⋆ı,0 =
R⋆ı,0 D⋆ = RH − B1 + X ≡ R⋆ı,0 =
⋆
D
(1 − p)(RH − B1 + X) + pL
|
{z
}
|
{z
}
by (I)
⋆
since, by (IV), D = (1 − p)(RH − B1 + X) + p L
1
≡ R⋆ı,0 =
1 − p 1 − RH −BL 1 +X
‘
16
Bhaskar DasGupta and Lakshmi Kaligounder
(VI) Proof of
pterm
= (1 − β)
n
n−1
B0
R +βB −L
RH +X−L−B1(1−βB1)+ L n−11
!
If a bank borrows long term, the banker can borrow up to the point to make him/her
indifferent between long-term borrowing and shirking in both periods to collect
B0 + B1 . In long-term lending, investors cannot liquidate the bank under any circumstances. Since shirking in only the second period is clearly suboptimal, the banker
shirks in both periods and the expected payoff as of t = 0 is B0 + B1, and thus the
payoff in the good state must be at least B0 + B1. This implies:
p
p
RH − R term D term + X + B1 ≥ B0 + B1
1−
n
n
Considering the worst case scenario
1−
p
p
RH − R term D term + X + B1 = B0 + B1
n
n
p
B0 + B1 − B1
n
≡ R term D term = RH + X −
p
1−
n
(7)
Since the lenders break even in expectation at t = 0, we have
p
p
p
p
p
D term = 1 − R term D term + RL =⇒ D term = 1 − (X + RH ) − B0 − B1 + B1 + RL
n
n
n
n
n
|
{z
}
using (7)
p
p
p
p
≡ D term = X + RH − X − RH − B0 − B1 + B1 + RL
n
n
n
n
Banks will not decide to deviate from contagious equilibrium provided the following
holds:
payoff in contagious system ≥ payoff in long-term borrowing
≡
β(1 − p).B1 − (1 − D∗) ≥ β(B0 + B1 ) − (1 − D term)
≡
βB1 − pβB1 − 1 + RH + X − B1 − pRH − pX + pB1 + pL
p
p
p
p
≥ βB0 + βB1 − 1 + X + RH − X − RH − B0 − B1 + B1 + RL
n
n
n
n
p
B0 − βB0 ≥ nβB1 + nRH + nX − nB1 − nL − X − RH + B1 + RL + L − L + βB1 − βB1
n
nB0(1 − β) ≥ p(βB1(n − 1) + RH (n − 1) + X(n − 1) − B1(n − 1) − L(n − 1) + RL + βB1 − L)
n
RL + βB1 − L
B0 (1 − β) ≥ p βB1 + RH + X − B1 − L +
n−1
n−1
≡
≡
≡
Densely Entangled Financial Systems
17
n
p ≤ (1 − β)
n − 1
B0
≡
RL + βB1 − L
RH + X − L − B1(1 − βB1) +
n−1
n
B0
term
=⇒ p
= (1 − β)
R
+
βB
−
L
n − 1
L
1
RH + X − L − B1(1 − βB1) +
n−1
(VII) Proof of p f.aut =
(1−β)(RH +X−B1−RL )
RH +X−L−B1(1−β)− nβ (RH −RL )
Since a bank survives at t = 2 even if its real project only delivers RL , we have
payoff in contagious system ≥ payoff in full autarky
p
p
≡ β(1 − p)B1 − (1 − D⋆) ≥ β 1 − (RH + X − RL ) + β X − 1 − RL
n
n
≡ βB1 − pβB1 − 1 + RH + X − B1 − pRH − pX + pB1 + pL ≥
p
p
p
p
βRH + βX − βRL − βRH − βX + βRL + β X − 1 − RL
n
n
n
n
βRH βRL
≡ βB1 + RH + X − B1 − βX + βRL − RL − βRH ≥ p βB1 + RH + X − B1 − L −
+
n
n
β
≡ RH (1 − β) + X(1 − β) − B1(1 − β) − RL(1 − β) ≥ p RH + X − L − B1(1 − β) − (RH − RL )
n
(1 − β)(RH + X − B1 − RL )
(1
−
β)(R
+
X
−
B1 − R L )
H
≡ p≤
≡ p f.aut =
β
β
RH + X − L − B1(1 − β) − (RH − RL )
RH + X − L − B1(1 − β) − (RH − RL )
n
n
(VIII) Proof of I = ru + X − B1 > 0
The amount of debt reduction needed at time t = 1 to stabilize a bank that lost at
least r counter-parties is I = ru + X − B1 > 0. If less than r counter-party neighbors
fail, then the fraction of failed counter-parties of the bank is less than 1/2, hence bank
can survive this loss. If a bank looses at least r hedges at t = 1, it can only roll over
its debt if its incentive constraint is not violated. This can be insured by injecting
enough equity into the bank to make sure it can pay back its debt even if it loses ru
at t = 2 due to the lost hedge. Thus, we have
max
max
RH − ru − Rmax
+ I ≥ 0 ≡ Rmax
= RH + X − B1 =⇒ I = ru + X − B1
ı,1 D
ı,1 D
18
Bhaskar DasGupta and Lakshmi Kaligounder
Appendix 2: Glossary of financial terminologies
Risk: Risk is a chance that an investment’s actual return will be less than expected.
Asset: An asset is anything that is owned by an individual or business that has a
monetary value.
Counter-party Risk: Risk that one party in an agreement defaults on its obligation to repay or return securities.
Hedging: A strategy to reduce the risk by making a transaction in one market to
protect against the loss in another market.
Over The Counter (OTC): A market that is conducted between dealers by telephone and computer and not on a listed exchange. OTC stocks tend to be of those
companies that do not meet the listing requirements of an exchange, although
some companies that do meet the listing requirements choose to remain as OTC
stocks. The deals and instruments are generally not standardized and there is no
public record of the price associated with any transaction.
Equity: The amount that shareholders own, in the form of common or preferred
stock, in a publicly quoted company. Equity is the risk-bearing part of the company’s capital and contrasts with debt capital which is usually secured in some
way and which has priority over shareholders if the company becomes insolvent
and its assets are distributed.
Liquidity: The degree to which an asset or security can be bought or sold in the
market without affecting the asset’s price, i.e., the ability to convert an asset to
cash quickly.
Credit Default Swap (CDS): A specific type of counter-party agreement that allows the transfer of third party credit risk from one party to the other. One party
in the swap is a lender and faces a credit risk from a third party, and the counterparty in the swap agrees to insure this risk in exchange for regular periodic payments.
Appendix 3: List of notations and variables
n : total number of banks in the network.
t : the time variable. Three time periods t = 0, 1, 2 are considered (initial, interim
and final time period, respectively).
β : the discount factor (we assume that β < 1).
R : the borrowing rate. R0 and R1 are the borrowing rates at t = 0 and at t = 1,
respectively.
ε : an independent random variable, realized at t = 2, that takes a value of +u or
−u, each with probability 1/2.
Densely Entangled Financial Systems
19
L : the return if the investments are liquidated early at t = 1 L(< RH ).
X : Non-pledgable payoff.
r : each bank is assumed to have 2r counter-party neighbors.
D : investors share of investment in the bank at t = 0; 1 − D is the bankers share
of investment in the bank at t = 0 (D, 1 − D ≥ 1).
e : the unobservable effort choice made by the bank (e ∈ {0, 1}).
Bi : Bankers private benefit at the time period i.
p : the probability of bad state.
p
soc
: if p < psoc , then irrespective of the number of counter-party neighbors there
is no need for any counter-party insurance even from a social perspective.
pind : if p > pind then the banks will not buy any counter-party insurance as the
private benefits of insuring exceed the cost.
pterm : if p < pterm then the banks will continue to prefer short-term debt.
paut : if p < paut then no bank will have
n an incentive oto hold more equity and boraut
row less. p is equal to min ps.aut , pr.aut , pf.aut where the superscripts s.aut,
r.aut and f.aut stand for safe autarky, risky autarky and full autarky, respectively.
Dmax : The maximum amount of borrowing at t = 0 that can be rolled over at t = 1.
ssafe : The price per unit of default insurance in a stable system where all banks
buy counter-party insurance.
Dsafe : The amount borrowed by the bank for unit investment in the project in a
stable system where all banks buy counter-party insurance.
D ⋆ : The amount borrowed by the bank for unit investment in the project in a
contagious system where banks decide not to buy insurance.
References
1. A. Zawadowski, “Entangled Financial Systems” (Review of Financial Studies, 26 (5), 12911323, 2013).
2. F. Castiglionesi and N. Navarro, “Fragile Financial Networks” (Tilburg University Working
Paper, 2011).
3. H. Amini, R. Cont and A. Minca, “Resilience to contagion in financial networks”
(arXiv:1112.5687v1, December 2011).
4. P. Gandhi, F. A. Longstaff and N. Arora, “Counterparty Credit Risk and the Credit Default
Swap Market” (2012 Social Science Electronic Publishing, Inc. March 2011).
5. P. Gai and S. Kapadia, “Contagion in financial networks” (Proc. R. Soc. A, 466 (2120), 24012423, 2010).
6. J. Corbo and G. Demange, “Contagion in Financial Networks: A Random Graph Model” (3rd
financial Risks International Forum, Paris, France, 2010).
7. A. Zawadowski, “Interwoven Lending, Uncertainty, and Liquidity Hoarding” (Boston University, Working Paper, 2010).
20
Bhaskar DasGupta and Lakshmi Kaligounder
8. F. R. Liedorp, L. Medema, M. Koetter, R. H. Koning and I. van Lelyveld, “Peer monitoring or
contagion? Interbank market exposure and bank risk” (DNB working paper No. 248, 2010).
9. R. M. Stulz, “Credit Default Swaps and the Credit Crisis” (Journal of Economic Perspectives,
24 (1), 73-92, 2010).
10. D. Duffie and H. Zhu, “Does a Central Clearing Counterparty Reduce Counterparty Risk?”
(Stanford University, Working Paper, 2009).
11. P. R. Kleindorfer, Y. Wind and R. E. Gunther, “The Network Challenge: Strategy, Profit and
Risk in an Interlinked World” (Pearson Prentice Hall, 2009).
12. F. Allen and A. Babus, “Networks in Finance” (Network-based Strategies and Competencies,
P. Kleindorfer and J. Wind (eds.), 367-382, Wharton School Publishing, 2009).
13. A. Haldane, “Why Banks Failed the Stress Test” (speech given at the Marcus-Evans Conference on Stress Testing, 9-10 February 2009, available at http://www.bankofengland.co.
uk/publications/speeches/2009/speech374.pdf).
14. V. Acharya and A. Bisin, “Counterparty Risk Externality: Centralized versus Over-TheCounter Markets” (New York University, Working Paper, 2009).
15. C. Pirrong, “The Economics of Clearing in Derivatives Markets: Netting, Asymmetric Information, and the Sharing of Default Risks Through a Central Counterparty” (University of
Houston, Working Paper, 2009).
16. R. J. Caballero and A. Simsek, “Complexity and Financial Panics” (NBER Working Paper
No. 14997, 2009).
17. E. Nier, J. Yang, T. Yorulmazer and A. Alentorn, “Network models and financial stability”
(Journal of Economics Dynamics and Control, 31, 2033-2060, 2007).
18. A. Babus, “The Formation of Financial Networks” (Tinbergen Institute Discussion Paper No.
2006-093/2; FEEM Working Paper No. 69.2007, 2007).
19. S. H. Zhu and M. Pykhtin, “A Guide to Modeling Counterparty Credit Risk” (GARP Risk
Review, 16-22, July 2007).
20. A. Babus, “Contagion Risk in Financial Networks” (K. Liebscher (ed.), Financial Development, Integration and Stability, 423-440, 2006).
21. M. Eboli, “Systemic Risk in Financial Network: a Graph Theoretical Approach” (Mimeo,
Univ. di Chieti, Pescara, 2004).
22. A. Bisin and D. Guaitoli, “Moral Hazard and Nonexclusive Contracts” (The RAND Journal
of Economics, 2004, 35(2), 306-328).
23. F. Allen and D. Gale, “Financial Contagion” (Journal of Political Economy, 108 (1), 1-33,
2000).
24. X. Freixas, B. M. Parigi and J.-C. Rochet, “Systemic risk, interbank relations, and liquidity
provision by the central bank” (Journal of Money, Credit and Banking, 32(3), 611-638, 2000).
25. C. W. Calomiris, “U.S. Bank Deregulation in Historical Perspective” (Cambridge University
Press, 2000).
26. N. Kiyotaki and J. Moore, “Credit Chains” (Princeton University, Working Paper, 1997).
27. D. W. Diamond and P. H. Dybvig, “Bank runs, deposit insurance and liquidity” (Journal of
Political Economy, 91, 401-419, 1983).
| 5cs.CE
|
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
Optimizing run-length
algorithm using octonary repetition tree
Kaveh Geyratmand Haghighi1, Mir Kamal Mirnia*2, Ahmad Habibizad Navin3
1Department
of Computer ,East Azarbaijan Science and Research Branch,
Islamic Azad University,Tabriz, Iran
1 Department of Computer Engineering, Tabriz Branch,
Islamic Azad University, Tabriz, Iran.
2Department
of Computer Engineering,
University Of Tabriz, Tabriz, Iran
3Department
of Computer ,East Azarbaijan Science and Research Branch,
Islamic Azad University,Tabriz, Iran
Abstract-Compression is beneficial because it helps detract resource usage. It reduces data storage space as well as
transmission traffic and improves web pages loading. Run-length coding (RLC) is a lossless data compression algorithm.
Data are stored as a data value and counts. This is useful on data that contains many consecutive runs. This paper proposes
a compression algorithm using octonary repetition tree (ORT), based on RLC. ORT is used to overcome the duplication
problem in primary RLC algorithms, instead of using flag or codeword. It's the first method of run-length encoding which
has the compression ratio greater than one in all tested cases. Experimental results, show average improvement of roughly
3 times, 3 times and 2 times in compression ratio field of study comparing to PRLC1, PRLC2, DF-RLC respectively. By
using this approach of run-length encoding we can compress wider types of data, such as multimedia, document, executive
files, etc.
I.
INTRODUCTION
Digital image compression has become an important tool because image capturing devices, display units and social
networks are widely used. When a raw image is stored as an array, usually the volume of file will be large and variable
due to image resolution and color channel numbers. So, image compression may offer lower file size for storage and
transmission. Usually there are two main approaches to compress data, such as lossy and lossless. Lossy image
compression is irreversible, in the sense that it will lead missing details depending on the level of compression. Thus,
this approach cannot be used in medical and satellite images and high level graphic design, since the original image
should be remained with no changes. But, the original data can be reconstructed from the compressed data obtained
from compression algorithm in lossless image compression.
Run-Length Coding (RLC) or run-length encoding [1], being part of JPEG [2], is a common lossless compression
method which is used in International Telecommunication Union too. Run-length Coding is often used in other
compression algorithms. In RLC sequence of bytes with the same value, encoded to number of repetition beside of
byte value. For example, this sequence of values 338888888 is encoded to (3 2 8 7), where, 3 and 8 stands for the
value of bytes, 2 and 7 shows the repetition correspondingly. The RLC is very efficient in encoding images when long
265
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
runs of bytes appear with the same intensity value (eg., black and white images). In the other hand it is inefficient in
encoding image with different values in sequence [5]. For example, 5 9 6 1 2 is a sequence of different bytes and will
be encoded as (5 1 9 1 6 1 1 1 2 1) by RLC. In this case, adding number of repetition will increase data size (double
in this case) and this is named as duplication problem in Run-Length Coding.
A new optimizing run-length algorithm is proposed here by using ORT. An octonary tree is added to point the
sequence of bytes with the same value. The tree features is used to locate sequence of the same bytes with less
overhead. Hence, in the proposed algorithm, there is no need to use extra bytes for determining the number of the
same bytes, nor reserving a bit as a flag. The proposed approach has no duplication problem, and there is no limitation
to encode number of bytes with the same value in a sequence. Also existence of an unused byte in image is
unnecessary. To the best of our knowledge, the compression rate just in this method is greater than 1 in all tested
images. Thus it is possible to use the compressing method for wider types of data, such as multimedia, documents,
executive files, etc.
This paper is organized in five sections: Section 2 describes related works in RLC compression algorithm. Proposed
method explains in section 3. Section 4 compares the new proposed method with recent works and section 5 concludes
the paper and suggests future works.
II.
RELATED WORKS
As a solution to solve the duplication problem, [2], [5], [6] offered that RLC encoding occurs only when number of
bytes with the same value is three or more and code other cases without using RLC. In this method, it is necessary to
use unique flag bits for pointing byte intensity levels from number of occurrence in a stream of RLC encoded bytes.
Two main techniques were suggested to implement the flag bits in RLC. The first technique uses a reserved and fixed
codeword. This codeword treated as flag. The flag, locates a byte value with its count. In [6], flag is associated to a
byte intensity value that does not exist. However, this technique is not acceptable if all of 256 bytes brought in the
image. A similar approach that uses an unmatched codeword of one byte was applied in [7]. We named this technique
as primary Run-Length 1 (PRLC1). Since a byte including 8-bits has range of [0, 255], count of consecutive bytes
(with the same intensity value) that can be encoded with one run is at most 256. The main problem of PRLC1 is that,
adding extra byte for separation run from byte value causes to duplication problem when different values in sequence
appears. This approach could not Assurances that compression ratio is always greater than 1.
The second technique distinguishes count from value by defining MSB bit as a flag [8]. When MSB is set to 0, rest
of bits (7-bits) indicates byte value and MSB=1, indicates following 7-bits is the count. This technique solves the
duplication problem, but it is only acceptable when the image contains less than 128 unique values. Also, the count
of consecutive bytes that can be encoded is at most 128. We named this technique as primary Run-Length 2 (PRLC2).
Same as PRLC1, this method could not Assurances that compression ratio is always greater than 1.
Another approach in [10], named duplication free run-length coding (DF-RLC), uses rule-based generative coding
to deriving codeword. This method solves duplication problem and the count of consecutive bytes that can be encoded
is infinitive. This technique achieves better compression ratio than PRLC1 and PRLC2. But looks like other methods
in this section, compression ratio is not always greater than 1.
266
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
a
b
j
k
c
…
…
…
l
e
d
f
…
a: Root
b: 1st child (a)
c: 2nd child (a)
d: 3nd child (a)
e: 4th child (a)
f: 5th child (a)
g
h
i
g: 6th child (a)
h: 7th child (a)
i: 8th child (a)
j: 1st child (b)
k: 2nd child (b)
l: 3nd child (b)
d: 3nd child (a)
Figure 1: ORT structure
III.
PROPOSED METHOD
In the proposed method, a tree used to display duplicate bytes instead of using flag or codeword. This tree
generates dynamically while reviewing the data. Fig.1 has shown the generated tree structure Each node of this tree
includes one byte (containing 8 bit). We named this tree "octonary repetition tree". This tree has 8 childes for each
node, named k= {1,2, … ,8}.
The formula for calculating the array structure indices of the different relatives of a node are as follows:
Parent(r) =⌊(r−1)/8⌋ if r≠0.
(1)
Kth child(r) =8r+k if 8r+k≤n.
(2)
The total count of nodes in the tree is n in (1), (2). The index of the node named as r, which is in the range of 0 to
n−1. Each node in this tree has been ranged in [0-255] that represents a byte (containing 8-bits) in binary. The root
of this tree has eight childes. As the lower layers has been reached, the data is divided into eight parts for each layer.
This trend continues till leaves has reached. This method represents exact place of the duplicated byte. Each 8-bits in
non-leaves, represents existence of equality of pair bytes in sub-branches. If there is no equality in sub-branches,
Table 1
calculating the array structure indices of the different relatives of a node
Position
Parent
1st Child
2nd Child
3nd Child
4th Child
5th Child
6th Child
7th Child
8th Child
0
1
2
3
4
5
6
7
8
1
0
9
10
11
12
13
14
15
16
2
0
17
18
19
20
21
22
23
24
3
0
25
26
27
28
29
30
31
32
4
0
33
34
35
36
37
38
39
40
5
0
41
42
43
44
45
46
47
48
267
6
0
49
50
51
52
53
54
55
56
7
0
57
58
59
60
61
62
63
64
8
0
65
66
67
68
69
70
71
72
9
1
73
74
75
76
77
78
79
80
10
1
81
82
83
84
85
86
87
88
11
1
89
90
91
92
93
94
95
96
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
Figure 2. Pruned tree for assumptive file
algorithm prunes it. Fig.2 has shown pruned tree for assumptive file including 64 bytes of data. Bytes with the same
value are {(3,4,5),(7,8),(13,14,15),(51,52,53,54)}.
Actually the non-leaf nodes are in charge of dividing the file. This division continues till the eight byte blocks has
reached in data. Each 8-bits in leaves, shows equality of pair bytes in sequence. If bit value is set to 0, indicates pair
bytes is not the same; when bit value is 1, means that pair bytes are the same. So, one byte in leaf of tree, shows
similarity of 8 bytes in data. Also we defined limitation of similarity to construct leaves of tree. For example,
limitation of 2, only allows to have leaves with at least 2 bits with value=1. Levels of the tree is determined by the
size of the file. If N is the total counts of bytes in file, then we calculate the number of levels as:
Level=⌈𝑙𝑜𝑔8 (𝑁)⌉
(3)
A fixed table defined to determine the depth of the tree and as it is fixed, it is embedded in the application and
would not lead to further process. Proposed algorithm eliminates the same bytes in sequence while passes data and
locates their place in tree. When algorithm reaches to end of file, attaches generated tree to end of it. This algorithm
runs more than one time recursively to gain better results. In each run, we treat the output of algorithm as input and
encode it again. In each iteration, comparison of bytes increases by one. Fig.3 has shown these steps in flowchart. If
n is count of bytes and k is number of runs. comparison of bytes at each run is as follow:
10
10
(4)
∑ 𝑏𝑦𝑡𝑒[𝑛] = ∑ 𝑏𝑦𝑡𝑒[𝑛 + 𝑘 + 1]
𝑘=0
𝑘=0
IV. EXPERIMENTAL RESULTS
The ORT method is implemented in C# programing language software. Hardware used to run implemented program
has an intel processor with 3.2 GHz frequency containing 8 thread and 8 GB ram with 1600 MHz frequency and SSD
268
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
Figure 3. Flowchart of proposed algorithm
storage drive with reading and writing speed of 500 m/s. Windows 10 is also used as the operating system. and tested
with 12 standard images [9]. Some of the test images is shown in Fig4.
Compared with the methods used codeword for encoding, the proposed method in this article outs down the need
for a codeword table. Compared with the methods used flag bit, this method has eliminated the limits of coding
consecutive bytes (this problem exist in PRLC1, PRLC2). The proposed method in this article just as the PRLC1
does not have the duplication issue. The use of tree to display the duplicated bytes is an innovative method in this
branch of encoding and has not be seen in any other article. One of the features of this tree is pruning any extra subbranches to reach the least occupied byte. Generating individual tree (at the end of the file) also simplifies the
uncompressing task. One of the important challenges in compressing is compression ratio. Data compression ratio
(CR) is the ratio between the uncompressed and compressed size.
𝐶𝑅 =
𝑈𝑛𝑐𝑜𝑚𝑝𝑟𝑒𝑠𝑠𝑒𝑑 𝑆𝑖𝑧𝑒
𝑐𝑜𝑚𝑝𝑟𝑒𝑠𝑠𝑒𝑑 𝑆𝑖𝑧𝑒
(5)
The results of the proposed method shown in Fig.5 and Table 2. Note that this results written after 10 times of
running proposed algorithm recursively and limitation of the same bytes is >=3. In this table the comparison criterion
269
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
Figure 4. Test images
800
)CR(
600
400
200
0
General Image
TRLC1
Gray-21
TRLC2
DF-RLC
Ruler
Proposed method
Figure 5. CR comparison of recent methods with proposed method
is compression ratio. These results comprised with the PRLC1, PRLC2 and DF-RLC. If the compression ratio>1 it
means the compression resulting data has gained free spaces after encoding. if compression ratio<1 indicates the
increased size of data after encoding. Our approach has reached the CR>1 in any tested images. While in other
methods, CR<1 has shown in some of cases.
Table 2
Comparison of proposed algorithm compression ratio with other approaches
Image
Proposed method
DF-RLC
TRLC2
TRLC1
1
General Image
9.467
8.3
5.4
5.5
2
Gray-21
625.961
42.7
38.0
49.0
3
Ruler
69.866
11.1
4.6
4.4
4
Clock
1.113
1.1
0.9
0.9
5
Tank
1.009
1.1
0.9
0.9
6
APC
1.000
1.4
0.9
0.9
7
Lenna
1.000
0.9
0.9
0.9
8
Numbers
1.171
0.9
1.1
1.1
9
Airplane
1.084
2.1
1.1
1.1
10
F-16
1.021
1.0
0.9
0.9
11
Boat
1.001
0.9
0.9
0.9
12
House
1.050
0.9
0.9
0.9
270
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 14, No. 8, August 2016
New approach improved the compression ratio in all cases. The previous approaches which used the RLC for
compression were limited to black and white or simple colored images due to their ineffectiveness. Hence these
methods were limited to geographical and satellite and medical image. But the proposed approach in this article, is
able to encode and compress other types of data such as multimedia files, applications, documents. In this method
instead of using flag or codeword, a tree has been used to indicate the duplicated bytes. This tree is generated while
passing the data and avoids reviewing the data repeatedly.
V.
CONCLUSION & FUTURE WORKS
In this paper run-length compression algorithm using ORT has been proposed. In this method, a tree used to display
duplicate bytes instead of using flag or codeword. This method has eliminated the limits of coding consecutive bytes.
In comparison to the traditional RLC approaches, this method overcomes to duplication problem. Our approach has
reached the CR>1 in any tested cases and considered to the test results of compressed photos, improvements have
been made in this regard. While in other methods, CR<1 has shown in some of images. It makes our algorithm
suitable to use in compression of other types of data. The generated tree is pruned which leads to decreasing the
number of bytes. We have a fixed table to determine the depth of the tree and since it is fixed it is embedded in the
application and would not lead to further process. Parallelizing the algorithm by dividing the data and pipelining the
recursive runs, speeds up compressing.
References
[1] Bentley, J. L., et al. (1986). "A locally adaptive data compression scheme." Commun. ACM 29(4): 320-330.
[2] Acharya, T. (2006). "VLSI Algorithms and architectures for JPEG2000." Ubiquity 2006(September): 1-42.
[3] Togneri, R. and J. Christopher (2003). Fundamentals of information theory and coding design, CRC Press.
[4] Zhan, W. and A. El-Maleh (2009). A new collaborative scheme of test vector compression based on equal-runlength coding (ERLC). Proceedings of the 2009 13th International Conference on Computer Supported
Cooperative Work in Design, IEEE Computer Society: 21-25.
[5] Luse, M. (1993). Bitmapped graphics programming in C++, Addison-Wesley.
[6] (2008). "Data Compression: The Complete Reference (by D. Salomon; 2007) [Book review]." IEEE Signal
Processing Magazine 25(2): 147-149.
[7] Hughes, J. F., et al. (2014). Computer graphics: principles and practice, Pearson Education.
[8] Glassner, A. (1991). "Adaptive Run-Length Encoding." Graphics Gems II 2: 89.\
[9] “USC-SIPI image
database.”
[Online].
Available http://sipi.usc.edu/services/database/
[10] Al-Wahaib, M. S. and K. Wong (2010). A Lossless Image Compression Algorithm Using Duplication Free RunLength Coding. Network Applications Protocols and Services (NETAPPS), 2010 Second International
Conference on.
271
https://sites.google.com/site/ijcsis/
ISSN 1947-5500
| 8cs.DS
|
arXiv:1611.02187v3 [math.AT] 9 Feb 2018
Classifying spaces for the family of virtually cyclic
subgroups of braid groups
Ramón Flores and Juan González-Meneses
∗
February 7th, 2018
Abstract
We prove that, for n ≥ 3, the minimal dimension of a model of the classifying
space of the braid group Bn , and of the pure braid group Pn , with respect to the
family of virtually cyclic groups is n.
1
Introduction
In the last years, there has been a growing interest in the description of the classifying
space of a group with respect to the family of its virtually cyclic subgroups, usually
denoted EG. The main reason for this interest is that this space is the target of the
Farrell-Jones conjecture [16], which given a group G, intends to compute the (algebraic)
K-theory of the group ring ZG through an assembly map that uses as a source nonconnective (topological) K-theory of the classifying space. The conjecture has been proved
to be true for a number of groups, see [38] for an excellent survey.
The knowledge of the algebraic K-theory of ZG is based in part on finding manageable
and “small” models for EG, and in particular in understanding which is the minimal
possible dimension of such a model. Hence, this problem has been widely studied in the
last years, and computations are available for an important number of classes of groups, as
for example locally finite groups [13], polycyclic groups [39] and more generally elementary
amenable groups [14], [19], CAT(0)-groups [37], [15], linear groups [12], hyperbolic groups
[31], [35] or mapping class groups [15], [33].
We are interested in the (full) braid groups Bn , which in fact can be seen as a particular
instance of mapping class groups. These objects, aside from being interesting by their
own right, have appeared in very different fields of Mathematics, as for example algebra,
topology, physics or geometry. In our context, it has been proved by Juan-Sánchez [32]
that Farrell-Jones isomorphism holds for them, and moreover their lower algebraic Ktheory groups have been computed [26]. However, not a lot is known about concrete
The first author was supported by MEC-FEDER grant MTM2013-42293-P and the second author
was supported by MEC-FEDER grant MTM2013-44233-P. Both authors supported by MEC-FEDER
grant MTM2016-76453-C2-1-P
∗
1
models for EBn , not even the minimal dimension of these spaces. In his PhD Thesis [17],
M. Fluch (Proposition 4.16), estimated this dimension between 3 and 5 in the case of B3 .
More recently, Juan-Trujillo [33] bounded it for a general Bn with a factorial bound (more
information at the end of Section 3).
The main result of our paper (Theorem 5.1) is that the minimal dimension of a model
for EBn (and also of the pure version EPn ), for every n ≥ 3, is n. In this way, it is proved
that the equality in Question 1.2 of [37] holds for these groups, and we expect that our
result is helpful for computations of Bredon (co)homology and/or algebraic K-theory. Of
course, next challenge in this context will be to find explicit models of EBn that realize
the minimal dimension, and in this sense, a promising line of research concerns the actions
of braid groups on CAT(0)-spaces (see [27] and references there). On the other hand, our
methods strongly depend on the rich internal structure of the (full) braid groups, so it
is likely they can be applied to braid groups over other surfaces for which the virtually
cyclic subgroups are understood (see for example [20] for the case of the sphere).
The structure of the paper is the following. In Section 2 we provide the background
about classifying spaces and geometric dimensions that will be needed later, and in particular, we describe Lück-Weiermann model for the case of virtually cyclic subgroups.
Section 3 is devoted to braid groups, and in it we recall general facts about these objects,
and prove some new results that will be useful for our computations. In Section 4 we
describe the structure of the commensurators of the cyclic subgroups. Section 5 contains
the proof of our main theorem. The most difficult point of the proof is to bound the
dimensions in the reducible non-periodic case, and this goal is achieved by proving that
the quotient of certain normalizers in Bn are virtually torsion-free, and then applying an
appropriate result of C. Martı́nez (Theorem 2.5 in [42]) that only works in this case.
2
2.1
Background
Classifying space for families
In this section we review the notion of classifying space for a family of subgroups of a
group, which will be the main object of study in this note, with special emphasis in the
case of the family of virtually cyclic subgroups. We intend to give a brief survey based
on the treatment of [14], the reader interested in a more thorough approach is referred to
Lück excellent monography [36]. Moreover, we assume that the reader is familiar with the
notion of G-CW-complex and other fundamental concepts of equivariant G-homotopy; in
any case, all the necessary information can be found in [43], Part I, section 2.
Definition 2.1 Let G be a discrete group, and F a family of subgroups of G which is
closed under conjugation and passing to subgroups. Then a G-CW-complex X is a classifying space for the family F if the fixed point set X H is contractible for every H ∈ F and
empty for every H ∈
/ F.
The classifying space for the family F is always unique up to G-homotopy equivalence,
and is denoted by EF G. Observe that the definition implies that the isotropy groups of
2
the action of G over any model for EF G are in F . Moreover, as the family is closed under
passing to subgroups, the trivial group is always a member of F , and this implies that
EF G is always a contractible G-space. Remark also that if G ∈ F , the point is a model
for EF G.
Remark 2.2 Given a family F of subgroups of G closed under conjugation and subgroups
and a subgroup H < G, the family F ∩ H whose elements are the intersections of H with
the elements of F is a family of subgroups of H for which the same holds. Actually, F ∩H
is precisely the set of subgroups of H which belong to F . If we consider the action of H
over EF G by restriction, EF G becomes a model also for EF ∩H H.
Example 2.3 The following are the most important instances of classifying spaces for
families. Each one deserves its own notation:
• If F{1} is the family whose only element is the trivial subgroup, EF{1} G = EG, the
universal space for principal G-bundles.
• If FF in is the family of finite subgroups of G, then EFF in G is usually denoted EG
and called the classifying space for proper actions of G.
• If Fvc is the family of virtually cyclic subgroups of G, the space EFvc G is denoted
EG.
Observe that in a torsion-free group G, F{1} = FF in (G). In particular, in this case,
EG = EG. Also, it is well-known that in a torsion-free group G, every virtually cyclic
subgroup is cyclic ([40], Lemma 3.2), so Fvc (G) is the set of cyclic subgroups of G.
The classifying space for proper actions has been widely studied from the nineties,
as it is the target of the Baum-Connes conjecture [6], which remains unsolved in full
generality. However, although the classifying space EG is the object of study of the
Farrell-Jones conjecture [16] about the algebraic K-theory and L-theory of the group ring
of G, also unsolved in general, much less is known about this classifying space, because
a construction of a manageable model for it is usually much more difficult than the
construction of a good model for EG. In this sense, a big step forward has been the work
of Lück-Weiermann, which builds a model for EG using as building pieces models EF H,
with H ≤ G and F families of groups which are contained in the family Fvc (G). Because
of its importance in our development, we will describe now this model; details can be
found in [39].
Let G be a group. An equivalence relation in the set Fvc \FF in of infinite virtually
cyclic subgroups of G is defined by establishing that H ∼ K if and only if H ∩ K is
infinite. We will denote by [H] the equivalence class of a subgroup H by this relation.
Notice that conjugation in G preserves Fvc , FF in and the relation ∼, hence we can define
g −1 [H]g as [g −1 Hg] for any g ∈ G and any H in Fvc \FF in .
Definition 2.4 If [H] is an equivalence class by ∼, the normalizer of [H] is defined as
the subgroup
NG [H] := {g ∈ G | g −1 [H]g = [H]}.
3
It is important to observe that NG [H] contains the normalizer NG (H) for every representative H of the class [H], but in general is not equal to any of these normalizers
(Example 2.6 in [39]). However, in the groups we will be working with, there always will
be a group H0 ∈ [H] such that NG [H] = NG (H0 ). In general, we will see in Section 4
that for every group G, one has NG [H] = CommG (H), the commensurator of H in G (see
Definition 4.1).
Now a family of subgroups of NG [H] can be defined as follows:
F [H] := {K < NG [H] | K ∼ H or |K| < ∞}.
Observe that this family is closed under conjugation (in NG [H]) and taking subgroups.
Using these subgroups and families, an appropriate model for EG can be defined:
Theorem 2.5 ([39], Theorem 2.3) Let I be a complete set of representatives of the Gorbits (under conjugation) of equivalence classes [H] of infinite virtually cyclic subgroups
of G. For every [H] ∈ I, choose models for ENG [H] and EF [H] NG [H]. Choose also a
model for EG. Now consider the G-pushout:
`
[H]∈I
G ×NG [H] ENG [H]
`
[H]∈I
i
/
EG
idG ×NG [H] f[H]
`
[H]∈I
G ×NG [H] EF [H] NG [H]
/
X
where f[H] is a cellular NG [H]-map for every [H] ∈ I and i is the inclusion, or i is a
cellular G-map and f[H] is an inclusion for every [H] ∈ I. Then X is a model for EG.
In practice, this theorem implies that the existence of good models for the proper
classifying space of the commensurators and G, and also of the classifying spaces with
respect to the families F [H] will lead to the knowledge of good models for EG. Moreover,
the push-out implies the existence of a long exact sequence in Bredon homology, and
dimensional consequences that we will analyze in next section.
Although we have described here only the particular case of finite and virtually cyclic
subgroups (as it is the only one that will be needed), it is worth to point out that LückWeiermann model is defined for any two families F and G such that F ⊆ G and an
equivalence relation subject to some conditions is defined in G\F (see [39], Section 2).
2.2
Minimal dimensions for EG
In this section we will recall the results about geometric dimensions of classifying spaces
that will be necessary in the remaining of this note.
Definition 2.6 Let G a group, F a family closed under conjugation and taking subgroups.
The geometric dimension of G with respect to the family F is the minimal dimension gdF G
of a model for EF G.
4
According to the classical notation, we denote gdF G by gd G when F = F{1} ; by gd G
when F = FF in ; and by gd G when F = Fvc . We are mainly interested in the latter.
The study of the dimension of a group is a classic topic of research where group theory,
homological algebra and geometry overlap via the different geometric and (co)homological
versions of the dimension. Usually, these different versions are related, and small values
of them allow a sharp knowledge of the structure of the group. Of course, in this paper we are mainly interested in the geometric dimension with respect to the family of
virtually cyclic groups, but in our arguments we will also need to compute some proper
geometric dimensions, as well as some algebraic counterparts of it. Among them, it will
be particularly important the next one:
Definition 2.7 Given a group G and a family F of subgroups of G closed under conjugation and taking subgroups, the (Bredon) cohomological dimension of G with respect to the
family F , denoted by cdF G, is the maximal dimension of a nonzero Bredon cohomology
group with respect to F .
The main definitions about Bredon (co)homology can be found in [43] or [17]. In
particular, the following fact is described in the second reference, and it will be relevant
for our purposes:
Proposition 2.8 In the notation of the previous definition, for every G and F the equality cdF G = gdF G holds whenever gdF G 6= 3. If gdF G = 3, then 2 ≤ cdF G ≤ 3.
The question about the equality when gdF G = 3 is known as Eilenberg-Ganea conjecture, and it is yet unsolved in the case of the trivial family. The paper [10] contains an
interesting discussion about this conjecture, including examples of the strict inequality
for the family of finite groups; examples for the family of virtually cyclic groups can be
found in [18].
The key result that will lead us to identify gd Bn is a bound implied by the previously
described model of Lück-Weiermann:
Corollary 2.9 (Remark 2.5 in [39]) With the above notations, suppose there exists a
natural number k such that:
• gd G ≤ k.
• gd NG [H] ≤ k − 1, for every [H] ∈ I.
• gdF [H] NG [H] ≤ k, for every [H] ∈ I.
Then gd G ≤ k.
When applying this corollary, it is interesting to realize that according to Remark
2.2, for every H < G and every family of subgroups of G closed under conjugation and
subgroups, gdF ∩H H ≤ gdF G. Then, gd G and gd G work as upper bounds for the
corresponding minimal dimension of any of their subgroups.
We should also mention here Proposition 5.1 in [39], which establishes that an upper
bound for gd automatically gives a bound for gd:
5
Proposition 2.10 For any discrete group G, gd G ≤ gd G + 1.
Note that the equality is possible: for example, as Z is virtually cyclic, gd Z = 0,
while gd Z = 1. Usually, however, gd G ≤ gd G, and it is a question of Lück [36] to
identify families for which the inequality gd G ≤ gd G + 1 holds, and also for which this
inequality is an equality. For example, Degrijse-Petrosyan (Corollary 4.4 and Example
6.5 in [14]), prove that the inequality holds for elementary amenable groups, and also
construct Bestvina-Brady-like counterexamples for which it does not hold. While it is
easy to check that the inequality holds for braid groups (see Proposition 3.7 and the
previous discussion), to prove that it is in fact an equality has been one of the main
motivations of our paper, and is a beautiful and immediate consequence of Theorem 5.1.
3
3.1
Braid groups
A little survey
In this section we collect the basic definitions and facts about braid groups. For further
information the interested reader is referred to the more detailed treatments of [26] or [23]
Given an integer n ≥ 2, the braid group Bn on n strands can be defined in several
different ways. Algebraically, it is given by the following presentation [4, 5]:
σi σj = σj σi ,
if |i − j| > 1
Bn = σ1 , . . . , σn−1
σi σj σi = σj σi σj , if |i − j| = 1
Also, if Mn is the configuration space of n distinct points in the plane C,
Mn = {(x1 , . . . , xn ) ∈ Cn | xi 6= xj for i 6= j},
and Σn is the symmetric group on n elements, which acts on Mn by permuting coordinates,
then
Bn = π1 (Mn /Σn ).
This can be visualized by choosing n base points, say {1, . . . , n} ⊂ C, and considering a
braid as a collection of n disjoint paths in C × [0, 1], called strands, where the i-th strand
starts at (i, 0), moves monotonically on the second coordinate, and ends at (j, 1) for some
j ∈ {1, . . . , n}. This collection of strands (i.e. this braid) is considered up to isotopy
of C × [0, 1], fixing the boundary pointwise, and the multiplication of braids is given by
stacking and rescaling. The generator σj corresponds to a braid in which the strands j
and j + 1 cross, while the other n − 2 strands are constant. More precisely, σj can be
given by the paths
1 1 (1−t)iπ
1 1 −tiπ
(1, t) , . . . , (j − 1, t) , j + + e
,t , j + + e
, t , (j + 2, t) , . . . , (n, t)
2 2
2 2
for t ∈ [0, 1].
6
Each braid has a corresponding permutation, induced by the endpoints of its strands,
so there is a surjective map Bn → Σn . The kernel of this map, called the pure braid group
and denoted Pn , is the subgroup of braids in which the i-th strand starts at (i, 0) and
ends at (i, 1) for i = 1, . . . , n. Actually, Pn = π1 (Mn ). [7]
There is a third definition of Bn that will be important for us in this paper. Let Dn
be a n-times punctured disc (for instance, Dn = D\{1, . . . , n} where D is the disc in C
with diameter [0, n + 1] ⊂ R). Then a braid can be seen as an orientable homeomorphism
of Dn to itself, fixing ∂(Dn ) pointwise, up to isotopy relative to ∂(Dn ). That is, Bn is the
mapping class group of Dn [30, 7]. We remark that an automorphism corresponding to a
braid can be obtained from viewing the family of strands as a motion of the n punctures in
D, which can be extended to a continuous motion of all points of D, fixing the boundary.
The final position of the points yields the homeomorphism, which is unique up to isotopy
of Dn relative to ∂(Dn ). Conversely, if an automorphism f of Dn is given which fixes
the boundary ∂(Dn ) pointwise, it can be uniquely extended to an automorphism f of the
whole disk D. As every automorphism of D fixing the boundary is isotopic to the trivial
automorphism (Alexander’s trick), we can continuously deform idD into f . The trace of
the points 1, . . . , n under this deformation yields a family of n paths in D × [0, 1], which
can be extended to C × [0, 1] by the identity outside the cylinder. These n paths form the
braid corresponding to the isotopy class of f .
The braid group Bn is torsion-free [11]. Its center is infinite cyclic, generated by
∆2 = (σ1 (σ2 σ1 ) · · · (σn−1 σn−2 · · · σ1 ))2
This element corresponds to a Dehn-twist along a curve parallel to the the boundary of
Dn (roughly speaking, a rotation of ∂(Dn ) by 360 degrees). Hence, to quotient Bn by
its center corresponds to collapsing the boundary to a new puncture, so Bn /h∆2 i is a
subgroup (of index n + 1) of the mapping class group of the (n + 1)-times punctured
sphere Sn+1 .
This way of viewing braids as mapping classes allows to use the powerful theory of
Nielsen-Thurston [45], and to classify braids into three geometric types. In this way, a
braid α is said to be:
• Periodic, if some nontrivial power of α belongs to h∆2 i.
• Pseudo-Anosov, if there is a pair of transverse measured foliations of Dn preserved
by α, such that the action of a homeomorphism f (representing α) on one of them
scales the measure by some real number λ > 1 (called dilatation factor), and the
measure of the other one by λ−1 .
• Reducible non-periodic, if α preserves a family of isotopy classes of disjoint, essential
simple closed curves in Dn . Here essential means enclosing more than one and less
than n punctures.
Essentially, reducible braids are those which can be reduced into simpler ones. More
precisely, suppose that α is a reducible, non-periodic braid. Up to conjugacy in Bn , we can
7
assume that the family of isotopy classes of curves preserved by α can be represented by
a family of circles. Up to replacing α by a power, we can assume that α is represented by
an automorphism f which sends every circle to itself, so we can restrict the automorphism
f to the connected components obtained by removing the family of circles from Dn . As
each of these connected components is again homeomorphic to a punctured disc, each
restriction corresponds to a braid with fewer strands. Actually, there is a particular
family of isotopy classes of curves, called the canonical reduction system of α, CRS(α),
such that each of the mentioned restrictions is either periodic or pseudo-Anosov (see [8]).
The decomposition of a reducible braid into simpler ones will be used several times
in this paper, so we will make it more precise. Given a reducible, non-periodic braid α,
let Cα be the set of outermost isotopy classes of curves in the canonical reduction system
CRS(α). Notice that α preserves Cα (up to isotopy). Up to conjugacy, we can assume
that Cα can be represented by a disjoint union of unnested circles. Now recall that an
automorphism representing α can be extended to an automorphism f of D, and that there
is an isotopy from idD to f . The trace of the punctures under this isotopy represent the
strands of α, and the trace of the circles representing Cα looks like a family of tubes, each
one enclosing more than one and less than n punctures. See Figure 1.
Figure 1: A reducible braid α ∈ B8 , and the external braid αext . In this case, Cα is represented by a family of three circles. The internal braids inside each tube are, respectively,
σ1 ∈ B2 , σ1 σ2 ∈ B3 and the trivial braid in B2 . The external braid is σ1 σ2 σ3−1 σ1 σ3−1 ∈ B4 .
The decomposition of α along Cα yields an external braid αext , and some internal
braids. The external braid αext corresponds to forgetting what happens inside the tubes:
one gets a braid made of tubes and (possibly some) strands, which is transformed into a
braid by shrinking each tube to form a single strand. The internal braids are those which
are contained into each tube. Actually, as a tube does not necessarily end at the same
place it started, there is some shifting of the final points of the strands that should be
made in order to define the internal braids properly, but it is clear from the picture (see
Figure 1). See [24] for details.
A further simplification can still be done. Denote C1 , . . . , Cm the circles representing
Cα . The action of α on Cα induces a permutation on these circles. If Ci1 , . . . , Cir is a cycle
under this permutation, then we can conjugate α in such a way that all internal braids
in the tubes corresponding to Ci2 , . . . , Cir are trivial, and the only possibly nontrivial one
8
is inside the tube corresponding to Ci1 . We can do the same with all cycles, so we can
assume (up to conjugating α) that there is just one internal braid αi for each cycle of
tubes determined by α. These internal braids α1 , . . . , αk are unique up to conjugacy (see
[25]), and can be used to describe the centralizer of α:
Theorem 3.1 (Theorem 1.1 in [25]) Let α ∈ Bn . If n = 2 the centralizer Z(α) = B2 ≃ Z.
If n ≥ 3, the centralizer Z(α) is as follows:
• If α is periodic, Z(α) either is equal to Bn or is isomorphic to a subgroup of Bm
for some m < n. (Actually, in the latter case it is the braid group of an annulus on
less than n − 1 strands, which can be embedded into Bm ).
• If α is pseudo-Anosov, Z(α) ≃ Z2 , consisting of all elements having a representative
which preserves the same two foliations as α.
• If α is reducible, there is a split short exact sequence
pα
1 −→ Z(α1 ) × · · · × Z(αk ) −→ Z(α) −→ Z0 (αext ) → 1,
where α1 , . . . , αk and αext are defined as above, and Z0 (αext ) is a finite index subgroup
of Z(αext ).
The map pα : Z(α) → Z(αext ) mentioned above goes as follows: given β ∈ Z(α), one
has β −1 αβ = α. The canonical reduction system of a braid behaves in a natural way with
respect to conjugations [8], meaning that CRS(β −1αβ) is the image of CRS(α) under
the mapping class β. But CRS(β −1 αβ) = CRS(α), hence β sends CRS(α) to itself, and
then it sends Cα to itself, too. Therefore, β can be decomposed along Cα , and pα (β) is just
the external braid corresponding to this decomposition. Notice that we do not necessarily
have Cβ = Cα (actually, β could even be periodic and Cβ could be empty), so pα (β) is
not necessarily βext (which could be not defined). pα (β) is the tubular braid described by
the action of β on Dn , by looking at the motion of the circles representing Cα and of the
punctures not enclosed by Cα .
3.2
Centralizers and roots
In this section we will state some further properties of centralizers and roots in the pure
braid group Pn , that will be useful later.
Proposition 3.2 Roots in Pn are unique. That is, if α, β ∈ Pn are such that αm = β m
for some m 6= 0, then α = β.
Proof. This can be obtained from the results in [22], and can also be found in [9].
Corollary 3.3 If α ∈ Pn is a pure braid, ZBn (α) = ZBn (αm ) for every m 6= 0.
9
✷
Proof. If γ ∈ ZBn (αm ) then γαm γ −1 = αm , so γαγ −1 and α are two pure m-th roots of
the same pure braid, hence they are equal, so γ ∈ ZBn (α).
✷
Corollary 3.4 If x ∈ Pn and y ∈ Bn are such that xk = y m for some nonzero integers k
and m, then the subgroup hx, yi ⊂ Bn is cyclic.
Proof. Let z = xk = y m . Notice that y ∈ ZBn (z) = ZBn (xk ) = ZBn (x) as x is pure,
hence x and y commute. As Bn is torsion-free, then hx, yi is isomorphic to either Z2 or
Z. But since xk = y m , it must be isomorphic to Z.
✷
Proposition 3.5 Every cyclic subgroup hxi ⊂ Pn is contained in a unique maximal cyclic
subgroup.
Proof. Suppose first that x is periodic. The set of periodic pure braids is precisely the
cyclic group h∆2 i. Therefore, h∆2 i is the desired maximal cyclic subgroup. Now suppose
that x is pseudo-Anosov, with √
dilatation factor λ. A k-th root of x will then be pseudok
Anosov, with dilatation factor λ. As the set of possible dilatation factors is discrete [3],
the set of positive integers k for which x admits a k-th root in Pn is finite. The k-th root
of x for which k is maximal is then a generator for a maximal cyclic subgroup containing
hxi. Finally, suppose that x is reducible non-periodic. Then xext is either periodic or
pseudo-Anosov, and it is also a pure braid (on fewer strands). If α is a pure k-th root
of x, then αext is a pure k-th root of xext (the canonical reduction system is preserved
by taking powers, so CRS(α) = CRS(x), and the external braids are just obtained by
emptying the tubes). Hence, as there is a maximal k such that a k-th root of xext exists,
there is also a maximal k for which a k-th root of x exists. Such k-th root is a generator
of the desired cyclic subgroup containing hxi.
The uniqueness of the maximal cyclic subgroup follows from Corollary 3.4.
✷
Proposition 3.6 Let α ∈ Pn with n ≥ 2. The centralizer ZPn (α) is as follows:
• If α is periodic, ZPn (α) = ZPn (h∆2 i) = Pn .
• If α is pseudo-Anosov, ZPn (α) ≃ Z2 , consisting of all pure braids admitting a representative which preserves the same two foliations as α.
• If α is reducible, ZPn (α) ≃ ZPm1 (α1 ) × · · · × ZPmk (αk ) × ZPm (αext ), where α1 , . . . , αk
and αext are the internal and external braids, respectively, defined for α. Moreover,
m1 + · · · + mk + m = n + k.
Therefore, in any case, ZPn (α) ≃ Pt1 ×· · ·×Ptr ×Zt for some r ≥ 0, where t1 +· · ·+tr +t ≤
n + r − 1.
10
Proof. The periodic case is clear, as the only periodic pure braids are the powers of
∆2 . If α is pseudo-Anosov, it follows from Theorem 3.1 that ZPn (α) is the subgroup of
ZBn (α) ≃ Z2 consisting of pure braids admitting a representative which preserves the same
two foliations as α. Now ZPn (α) cannot be cyclic, as it contains a pseudo-Anosov braid
(α) and a periodic braid (∆2 ), which cannot have a common power. Hence ZPn (α) ≃ Z2
in this case.
Suppose finally that α is reducible. By Theorem 3.1, we have that
ZBn (α) = ZBm1 (α1 ) × · · · × ZBmk (αk ) ⋊ Z0 (αext ).
Notice that α1 , . . . , αk and αext are pure braids, and that mi is the number of strands of
the braid αi . Recall also that Z0 (αext ) ⊂ ZBm (αext ), where m is the number of strands of
αext .
Now ZPn (α) = ZBn (α) ∩ Pn , hence we need to describe the pure elements in the above
semidirect product. The left factor just becomes ZPm1 (α1 ) × · · · × ZPmk (αk ). For the
second factor, we need to describe the subgroup Z0 (αext ) with more detail, as it is done in
[25]. The group ZBm (αext ) (whose elements can be seen in Bn as braids made of strands
and tubes, with trivial braids inside the tubes) acts on ZBm1 (α1 ) × · · · × ZBmk (αk ) by
permuting coordinates (conjugating in Bn by the tubular braid, just permutes the tubes).
Then Z0 (αext ) is the subgroup formed by the elements whose induced permutation acts
as follows: if i is sent to j, then αi is conjugate to αj .
Now take a pure element in ZBm (αext ). Its induced permutation is trivial, hence it
belongs to Z0 (αext ). This means that Z0 (αext ) ∩ Pm = ZBm (αext ) ∩ Pm = ZPm (αext ).
Moreover, the action of this group on ZPm1 (α1 ) × · · · × ZPmk (αk ) is trivial. Hence the
semidirect product becomes a direct product when restricted to Pn , and hence:
ZPn (α) = ZPm1 (α1 ) × · · · × ZPmk (αk ) × ZPm (αext ).
Now recall that m1 , . . . , mk and m are, respectively, the number of strands of α1 , . . . , αk
and αext . By the decomposition of α into the internal braids α1 , . . . , αk and the external
braid αext , it follows that the punctured disc Dn is obtained from the punctured disc Dm
by replacing k punctures with the k punctured discs Dm1 , . . . , Dmk . At each replacement,
a puncture is replaced by mi punctures. Hence n = m + (m1 − 1) + · · · + (mk − 1), so
m1 + · · · + mk + m = n + k as we wanted to show. We remark that if α is not pure, this
equality becomes an inequality, m1 + · · · + mk + m ≤ n + k, as the permutation induced
by α on the tubes may contain nontrivial cycles, and there is just one internal braid for
each cycle. But in our case, as α is pure, the equality holds.
Notice that, if we define r ∗ = r − 1 for every positive integer r, the above equality
reads m∗1 + · · · + m∗k + m∗ = n∗ .
Now if some αi is reducible an not periodic, we can decompose ZPmi (αi ) into a direct
product of pure centralizers. We can keep doing this until all braids involved are either
periodic or pseudo-Anosov (recall that at every iteration the braids involved have fewer
strands, and in B2 all braids are periodic). Hence we will obtain
ZPn (α) = ZPt1 (β1 ) × · · · × ZPts (βs ),
11
where we can assume that β1 , . . . , βr are periodic, and βr+1 , . . . , βs are pseudo-Anosov.
Also, by induction on the number of strands it follows that t∗1 + · · · + t∗s = n∗ , that is,
t1 + · · · + ts = n + s − 1. But then
ZPn (α) = Pt1 × · · · × Ptr × Z2(s−r) .
Let t = 2(s − r). As every pseudo-Anosov braid has at least 3 strands, one has 3(s − r) ≤
tr+1 + · · · + ts , so t1 + · · · + tr + t = t1 + · · · + tr + 3(s − r) − (s − r) ≤ t1 + · · · + ts − (s − r) =
n + s − 1 − (s − r) = n + r − 1.
✷
Concerning the geometric dimension of Bn , the case of proper actions is well known
for mapping class groups in general: given a mapping class group Γbg,n of a connected
orientable surface of genus g, with n punctures and b boundary components, it is known
that gd Γbg,n = vcd Γbg,n , where vcd G is the virtual cohomological dimension of a group G.
This follows from Harer [28] for b > 0, has been showed by Hensel, Osajda, and Przytycki
[29] when b = 0 and n > 0, and by Aramayona and Martı́nez [1] when b = 0 and n = 0.
The virtual cohomological dimension of Γbg,n was computed by Harer in the mentioned
paper. The case we are interested in is Bn = Γ10,n , and by Harer’s formula we obtain:
gd Bn = vcd Bn = n − 1.
This equality is also consequence of the results in ([2], section 3).
In this paper we are interested in gd Bn . We can easily find a lower bound thanks to
the following result.
Proposition 3.7 The pure braid group Pn (and hence Bn ) contains a subgroup isomorphic to Zn−1 .
Proof. We define explicit generators, a1 , . . . , an−1 , as follows:
ai = (σi σi+1 · · · σn−1 )(σn−1 · · · σi+1 σi ).
One can check that these elements commute with each other just by drawing the corresponding braids, or by applying the braid relations. As Pn is torsion-free, these elements
generate a free abelian subgroup of Pn . To check that its rank is n − 1, just define a
projection π : Pn → Zn−1 which sends a pure braid α to π(α) = (x1 , . . . , xn−1 ), where xi
is the algebraic number of crossings of the strands i and n in the braid α, divided by 2.
This projection is a well defined homomorphism, and its restriction to ha1 , . . . , an−1 i is
surjective as π(ai ) = (0, . . . , 0, 1, 0, . . . , 0) with the 1 in position i.
✷
Corollary 3.8 gd B2 = gd P2 = 0, and gd Bn ≥ gd Pn ≥ n for n ≥ 3.
Proof. First, B2 ≃ P2 ≃ Z, so gd B2 = gd P2 = 0. Assume n ≥ 3. The first inequality
holds as Pn is a subgroup of Bn . Finally, as gd Zn−1 = n for n ≥ 3 (Example 5.21 in
[39]), the second inequality follows from Proposition 3.7.
✷
12
Some upper bounds for gd Bn are already known. In [33] such a bound is given for
any mapping class group Γbg,n of a compact orientable surface with punctures and negative
Euler characteristic. In the particular case of braid groups with n ≥ 2 (in order to have
negative Euler characteristic) and m ≥ 3 this bound is:
gd Bn ≤ [Bn : Bn [m]] · (vcd Bn + 1) = [Bn : Bn [m]] · n,
where Bn [m] is the m-th congruence subgroup of Bn , that is the subgroup of elements
which act trivially on H1 (Dn , Z/m).
Notice that the fundamental group of Dn is the free group Fn of rank n, hence the first
homology group is H1 (Dn , Z/m) ≃ (Z/m)n . It is well known [7] that the action of the
braid group (by automorphisms) on the fundamental group of Dn coincides with the Artin
representation Bn → Aut(Fn ), (see [5]). The action induced by this homomorphism on
the abelianization of Fn sends a braid α to the automorphism of Zn which just permutes
the coordinates in the same way as α permutes its strands. Therefore, the action of
α ∈ Bn on H1 (Dn , Z/m) is just given by permuting the coordinates in the same way as
α permutes its strands. This means that the elements which induce the trivial action are
precisely the pure braids. Hence Bn [m] = Pn for every m ≥ 2.
It follows that [Bn : Bn [m]] = [Bn : Pn ] = n! Hence, the above upper bound for gd Bn
is n! n. The goal of this paper is to show that gd Bn = n.
4
Commensurators
The main tool that we are going to use to compute geometric dimensions with respect
to the family of virtually cyclic groups is Lück-Weiermann’s Theorem 2.5, and a necessary step in our calculation will be to describe the commensurators of the virtually cyclic
subgroups generated by periodic or pseudo-Anosov braids. This task will be undertaken
in this section, while in the last one we will use different tools to compute the dimensions in the reducible non-periodic case. We start by reviewing some facts concerning
commensurators that will be needed in the rest of the paper.
Definition 4.1 Given an inclusion H < G of groups, the commensurator of H in G is
the group:
CommG (H) = {x ∈ G| [H : H ∩ H x ] < ∞ and [H x : H ∩ H x ] < ∞}.
We will be mainly interested in the case in which H is a virtually cyclic subgroup of
G. In this case, the previous definition is equivalent to:
CommG (H) = {x ∈ G| |H ∩ H x | = ∞}.
Note that this definition implies that, given the equivalence relation defined just before
Definition 2.4 over the virtually cyclic infinite subgroups of a group, and [H] an equivalence
class, NG [H] = CommG (H).
When computing the geometric dimension of Bn , we will make frequent use of the
following result of Degrijse-Petrosyan:
13
Theorem 4.2 ([14], 4.2) If H < G is virtually cyclic, and such that H ≃ Z and H ✂
CommG (H), then gdF [H] (CommG (H)) = gd (CommG (H)/H).
After these preliminaries we are ready for computing the commensurators of the virtually cyclic subgroups of Bn . Recall that, as Bn is torsion-free, any non-trivial virtually
cyclic group of Bn is infinite cyclic. The following lemma will be very useful, as it identifies
the commensurators with normalizers of subgroups, in the classic sense:
Lemma 4.3 Let H = hyi < Bn be a cyclic subgroup. Then there exists a braid x such
that:
• CommBn (H) = NBn (hxi).
• The braids x and y belong to the same cyclic group.
• The braid x is pure, and no proper root of x is pure.
Proof. This result is similar to Proposition 4.8 in [33], but we provide a proof adapted
to braid groups.
Let m be the smallest positive integer such that z = y m is a pure braid, and let x be a
generator of the maximal cyclic subgroup in Pn containing hzi (see Proposition 3.5). We
will see that x satisfies the properties in the statement. Clearly, it satisfies the third one
by construction.
Notice that xk = z = y m for some nonzero integers k and m. Hence by Corollary 3.4
hx, yi is cyclic, so the second property also holds.
Let us finally show that CommBn (hyi) = NBn (hxi). Let γ ∈ CommBn (hyi). This
means that |hγyγ −1i ∩ hyi| = ∞. As every nontrivial braid has a pure nontrivial power,
this intersection must contain an infinite number of pure braids, so |hγxγ −1 i ∩ hxi| = ∞,
hence γxp γ −1 = xq for some nonzero integers p and q. We will now show that p = ±q.
If x is periodic (and pure), xp is a nontrivial power of ∆2 , which is central. Hence
p −1
γx γ = xp , so xp = xq and, as Bn is torsion-free, p = q in this case. If x is pseudoAnosov and its dilatation factor is λ, the dilatation factor of xp is λ|p|, and the dilatation
factor of xq is λ|q| . As dilatation factors are preserved by conjugations, it follows that
|p| = |q|, as we wanted to show. (It is worth mentioning that a conjugation may transform
x into its inverse. For instance, in B3 we have ∆(σ1 σ2−1 )3 ∆−1 = (σ1 σ2−1 )−3 .) Now suppose
that x is reducible. Then xext is either periodic or pseudo-Anosov, and γ sends CRS(x)
to itself. Hence px (γ)(xext )p px (γ)−1 = (xext )q . Applying the claim to xext , it follows that
p = ±q, as desired.
If p = q then γxq γ −1 = xq , so γ ∈ ZBn (xq ) = ZBn (x) and then γ ∈ ZBn (hxi) ⊂
NBn (hxi). If p = −q then γx−q γ −1 = xq , so γx−1 γ −1 and x are two pure q-th roots of the
same pure braid, hence they coincide, and then γ ∈ NBn (hxi).
Conversely, suppose that γ ∈ NBn (hxi). This means that either γxγ −1 = x or γxγ −1 =
x−1 . Recall that y k = z = xm for some nonzero integers k and m. Then γy k γ −1 =
γxm γ −1 = x±m = y ±k ∈ hγyγ −1i ∩ hyi = H γ ∩ H, so |H γ ∩ H| = ∞. This shows that
γ ∈ CommBn (H), finishing the proof.
✷
14
We shall use the following results later. They can be deduced from the previous lemma,
but the first one can also be showed independently.
Lemma 4.4 If y ∈ Bn is periodic and nontrivial, then CommBn (hyi) = Bn .
Proof. If y is periodic, y m is central for some nonzero integer m, that is, y m = ∆2k
for some nonzero integer k. Then, for every α ∈ Bn , the conjugate αyα−1 is such that
(αyα−1)m = α∆2k α−1 = ∆2k . Hence hyi ∩ hαyα−1i ⊃ h∆2k i, so α ∈ CommBn (hyi).
✷
Corollary 4.5 If y ∈ Bn is pseudo-Anosov, then CommBn (hyi) is isomorphic to either
Z2 or an extension of Z2 by Z/2.
Proof. By Lemma 4.3, there is some x ∈ Pn such that CommBn (hyi) = NBn (hxi) As
x is a root of a power of y, it is also pseudo-Anosov, hence ZBn (hxi) ≃ Z2 . Now notice
that for any α ∈ NBn (hxi), either α−1 xα = x or α−1 xα = x−1 . If all elements are of the
first kind, then NBn (hxi) = ZBn (hxi) ≃ Z2 . Otherwise, we have the following short exact
sequence:
p
1 −→ ZBn (hxi) −→ NBn (hxi) −→ Z/2 −→ 1
where p(α) = 0 if α−1 xα = x, and p(α) = 1 if α−1 xα = x−1 . As ZBn (hxi) ≃ Z2 the result
holds.
✷
5
Classifying space for the family of virtually cyclic
groups
In this section we will compute gd Bn , which is the major achievement of this paper. Our
result is:
Theorem 5.1 The geometric dimension of Bn with respect to the family of virtually cyclic
subgroups is n, for n ≥ 3.
We devote this whole section to prove this result, and assume from now on that n ≥ 3.
According to Corollary 3.8, gd Bn ≥ n. Then, we must prove that n is also an upper
bound for this geometric dimension. As said above, we base our arguments on the main
theorem 2.5 of Lück-Weiermann, and we need to bound the three geometric dimensions
that appear Corollary 2.9. The first two bounds are quite straightforward, and we will
deal with them later. So let us concentrate in gdF [H] NBn [H], which has to be upper
bounded by n.
Recall from Lemma 4.3 that we can assume that H is generated by a pure braid x
which admits no proper roots, and that NBn [H] = CommBn (H) = NBn (H). But every
subgroup is normal in its normalizer, hence by Theorem 4.2 we have gdF [H] (NBn [H]) =
gd (NBn [H]/H). Hence we need to upper bound gd (NBn [H]/H) by n.
15
We invoke again Nielsen-Thurston theory to divide the computation in three cases.
Notice that if the original generator of y was periodic (respectively pseudo-Anosov or
reducible), then so is the pure element x given by Lemma 4.3, as it is a root of a power
of y.
1. Periodic case. Suppose that H = hxi where x is periodic. According to
Lemma 4.4, CommBn (H) = Bn , and as x is pure, periodic and has no proper roots,
x = ∆±2 . Therefore gd (NBn [H]/H) = gd (Bn /Z(Bn )), being Z(Bn ) = h∆2 i the center
of Bn . It is known that Bn /h∆2 i is a subgroup of Γ0,n+1, the mapping class group of the
sphere with n + 1 punctures, and it is a consequence of work of [29] and [28] that the
proper geometric dimension of Γ0,n+1 is n − 2. Hence, gd (NBn [H]/H) ≤ n − 2 in this
case.
2. Pseudo-Anosov case. Suppose now that x is pseudo-Anosov. By Corollary 4.5,
we know that CommBn (H) is either Z2 or an index 2 extension of Z2 . Without loss of
generality, we may assume that x is a power of one of the two generators of Z2 . Hence,
CommBn (H)/H is isomorphic to a finite extension of Z ⊕ Z/m (m = 1 is allowed). Applying Theorem 5.26 of [36], we obtain that gd CommBn (H)/H = 1, the Hirsch number,
and hence gd (NBn [H]/H) = 1 in this case.
3. Reducible, non-periodic case. Suppose finally that x is reducible and not periodic. We will find an upper bound gd (NBn (H)/H) basing our strategy in the following
result of C. Martı́nez (Theorem 2.5 in [42]):
Theorem 5.2 For any discrete group Γ, the following inequality holds:
cd Γ ≤
max {pdZWΓ F B(WΓ F ) + rk(WΓ F )}.
F ∈FFin (Γ)
Here pd is the projective dimension over the corresponding group ring, WΓ F = NΓ F/F
is the Weyl group, and B(WΓ F ) is the module of bounded functions over WΓ F (see the
mentioned paper for details about these concepts). Moreover, rk(WΓ F ) is the rank of a
maximal elementary abelian subgroup of WΓ F , as usual. Recall from Proposition 2.8 that
for any group G, cd G = gd G except possibly in the case gd G = 3. From now on we
assume Γ = NBn (H)/H, and we should bound the previous sum.
We first deal with the term pdZWΓ F B(WΓ F ). By Lemma 7.3 in [34], pdZWΓ F B(ZWΓ F ) ≤
pdZΓ B(ZΓ) for every finite F < Γ. Moreover, if we can prove that Γ is virtually torsionfree, Lemma 3.9 in [41] implies that the latter coincides with vcd(Γ), the virtual cohomological dimension of Γ, and we would have:
cdΓ ≤
max {vcd(Γ) + rk(WΓ F )}.
F ∈FFin (Γ)
Let us then show that this is true:
Lemma 5.3 The group Γ is virtually torsion-free.
p
Proof. We will use the short exact sequence 1 → Pn → Bn → Σn → 1, which we can
restrict to NBn (H):
p
1 −→ NBn (H) ∩ Pn −→ NBn (H) −→ G −→ 1,
16
where G = p(NBn (H)) ⊂ Σn is a finite group.
Recall that H = hxi where x is pure. Then H ✂ (NBn (H) ∩ Pn ) and H ✂ NBn (H), so
we can quotient the above injection by H and we obtain:
p
1 −→ (NBn (H) ∩ Pn )/H −→ Γ −→ G −→ 1.
We need to show that (NBn (H) ∩ Pn )/H is torsion-free.
Let αH ∈ (NBn (H) ∩ Pn )/H be a torsion element. That is, α ∈ NBn (H) ∩ Pn is
such that αm ∈ H for some nonzero integer m. This means that αm = xk for some
nonzero integers m and k. By Corollary 3.4, hα, xi is a cyclic subgroup, which belongs
to Pn as it is generated by pure braids. As hxi is a maximal cyclic subgroup in Pn ,
we get hα, xi = hxi = H, so α ∈ H and αH is trivial in (NBn (H) ∩ Pn )/H. Hence
(NBn (H) ∩ Pn )/H is torsion-free.
✷
Now we can obtain our bound for the projective dimension.
Lemma 5.4 The virtual cohomological dimension of Γ is bounded above by n − 2.
Proof. We saw in the previous lemma that (NBn (H) ∩ Pn )/H is a torsion-free, finite
index subgroup of Γ. Hence, we must show that cd ((NBn (H) ∩ Pn )/H) ≤ n − 2.
Notice that NBn (H)∩Pn = NPn (H). Now we point out that the pure braid group Pn is
bi-orderable [44]. This means that there is a total order of its elements which is invariant
under multiplication on the left and also on the right. In such an order, if an element is
positive (greater than the neutral element) its inverse is negative, and viceversa. Recall
that H = hxi. If x > 1, then ax > a and hence axa−1 > 1 for every a ∈ Pn . Analogously,
if x < 1 then axa−1 < 1 for every a ∈ Pn . This means that, in Pn , x cannot be conjugated
to its inverse. Therefore NPn (H) = ZPn (H).
Now recall from Proposition 3.6 that ZPn (H) = Pt1 × · · · × Ptr × Zt , where r, t ≥ 0
and t1 + · · · + tr + t ≤ n + r − 1.
Let C be the center of ZPn (H). As the center of each Pti is cyclic (generated by ∆2ti ),
it follows that C ≃ Zr+t . Denote by Gi the quotient of Pti by its center, which is a (finite
index) subgroup of the mapping class group Γ0,ti +1 of the (ti + 1)-times punctured sphere.
Then ZPn (H)/C = G1 × · · · × Gr .
We then have the following short exact sequence, which is a central extension:
1 −→ C −→ ZPn (H) −→ G1 × · · · × Gr −→ 1.
Moreover, as H is central in ZPn (H), we can quotient by H and obtain:
1 −→ C/H −→ ZPn (H)/H −→ G1 × · · · × Gr −→ 1.
Recall that we want to bound the cohomological dimension of the middle group. It is
important to remark that all groups involved in the above exact sequence are torsion-free:
First, we have shown in the previous lemma that ZPn (H)/H is torsion-free, hence so is
its subgroup C/H. On the other hand, a torsion element in Gi = Pti /h∆2ti i comes from a
17
periodic element in Pti . But the only pure braids which are periodic in Pti are the powers
of ∆2ti , hence Gi is torsion-free for every i.
Now it follows from Theorem 5.15 in [36] that gd G = gd G is subadditive for extensions of torsion-free groups. Hence
gd(ZPn (H)/H) ≤ gd(C/H) + gd(G1 ) + · · · + gd(Gr ).
Now C ≃ Zr+t , and H is a cyclic subgroup which is maximal. Hence C/H ≃ Zr+t−1 , and
gd(C/H) = r + t − 1. Also, each Gi is a subgroup of the mapping class group Γ0,ti +1 , so
gd(Gi ) = gd (Gi ) ≤ gd (Γ0,ti +1 ) = ti − 2. We then have:
gd(ZPn (H)/H) ≤ (r + t − 1) + t1 + · · · + tr − 2r ≤ r − 1 + (n + r − 1) − 2r = n − 2.
Therefore gd(ZPn (H)/H) ≤ n − 2. As cd G = gd G except in the case cd G = 2 and
gd G = 3, it follows that cd(ZPn (H)/H) ≤ n − 2.
✷
Remark 5.5 A different proof of the previous lemma follows from a direct application of
the Gysin sequence (see for example [46], 5.12), but we prefer the one we provide because
it includes an explicit description of the normalizer.
Now, in order to bound the second term of the sum in Theorem 5.2, we need to
understand the torsion of Γ.
Lemma 5.6 Any finite subgroup of Γ is cyclic.
Proof. Consider the natural projection p : NBn (H) → NBn (H)/H = Γ. Let F be a
finite subgroup of Γ, and let G = p−1 (F ). Consider the restriction of p to G. Its image is
finite, and its kernel is a subgroup of H, so is cyclic. Hence G is a virtually cyclic group.
As G ⊂ Bn , which is torsion-free, G is cyclic. Hence p(G) = F is also cyclic.
✷
We are now in a good position to deal with the rank of Weyl groups, and our bound
will be a consequence of the following lemma:
Lemma 5.7 For every finite F < Γ, every finite group of the Weyl group WΓ F is cyclic,
and in particular, rk(WΓ F ) = 1.
Proof. Let F < Γ be finite. Consider the short exact sequence:
p
1 −→ F −→ NΓ F −→ NΓ F/F −→ 1,
where p is the natural projection.
Let X be a finite subgroup of NΓ F/F = WΓ F , and let Y = p−1 (X). Then Y is an
extension of Y ∩F by X, which are both finite groups. Hence Y is finite. As Y ⊂ NΓ F ⊂ Γ,
by Lemma 5.6 Y is cyclic. Hence X = p(Y ) is also cyclic.
Observe, in particular, that this implies that rk(WΓ F ) = 1 for every finite F in Γ.
✷
Now we can join the previous information to obtain:
18
Proposition 5.8 The proper geometric dimension of Γ is smaller or equal to n−1, except
possibly in the case when gd Γ = 3, in which it can be at most n.
Proof.
By Theorem 5.2, Lemma 5.4 and Lemma 5.7, we have cd Γ ≤ n − 1. Hence by
Proposition 2.8, cd Γ = gd Γ ≤ n − 1, except possibly when cd Γ = 2 and gd Γ = 3.
✷
Proof of Theorem 5.1.
Let Bn be the full braid group, for n ≥ 3. First, as Pn < Bn , Corollary 3.8 implies
that gd Bn ≥ n.
To check the other equality, we make use of Lück-Weiermann model, and in particular
the dimensional conditions of Corollary 2.9:
• We know (by [2] or [28], see Section 3 above) that the geometric dimension gd Bn
is always n − 1, so it is smaller than n.
• As NBn [H] ⊆ Bn for every cyclic H < Bn , gd NBn [H] ≤ gd Bn ≤ n − 1.
• Let n ≥ 3 and H < Bn cyclic. If H is generated by a periodic or a pseudo-Anosov
braid, it was stated above that gdF [H] NBn [H] ≤ n − 2, and in particular smaller
than n. If H is generated by a reducible non-periodic braid, the previous proposition
states that gdF [H] NBn [H] ≤ n.
Now appealing to Corollary 2.9, gd Bn ≤ n, and thus gd Bn = n. So we are done.
✷
The same is true for pure subgroups of Bn :
Corollary 5.9 We have gd Pn = n for n ≥ 3.
Proof. The bound gd Pn ≥ n was proved in Corollary 3.8, while gd Pn ≤ n is a
consequence of Theorem 5.1, taking into account Remark 2.2.
✷
References
[1] J. Aramayona and C. Marı́nez-Pérez, The proper geometric dimension of the mapping
class group, Algebr. Geom. Topol. 14 (2014), 217–227.
[2] V. Arnold, On some topological invariants of algebraic functions, Trans. Moscow
Math. Soc. 21 (1970), 33–52.
[3] P. Arnoux, J. Yoccoz, Construction de diffeomorphismes pseudo-Anosov, C. R. Acad.
Sci. Paris 292 (1981), 75-–78.
[4] E. Artin, Theorie der Zp̈fe, Abh. Math. Sem. Univ. Hamburg 4 (1925), no. 1, 47—72.
19
[5] E. Artin, Theory of braids, Ann. of Math. (2) 48 (1947), 101-–126.
[6] P. Baum, A. Connes and N. Higson, Classifying space for proper actions and Ktheory of group C ∗ -algebras, in C ∗ -algebras: 1943-1993, 240–291, Contemp. Math.,
167, Amer. Math. Soc., Providence, RI, 1994.
[7] J. S. Birman, Braids, links, and mapping class groups, Annals of Mathematics Studies, No. 82. Princeton University Press, Princeton, N.J.; University of Tokyo Press,
Tokyo, 1974.
[8] J. S. Birman, A. Lubotzky and J. McCarthy, Abelian and solvable subgroups of the
mapping class groups, Duke Math. J. 50 (1983), 1107-–1120.
[9] C. Bonatti and L. Parı́s, Roots in the mapping class groups, Proc. Lond. Math. Soc.
98 (2009), 471–503.
[10] N. Brady, I. Leary and B. Nucinkis, On algebraic and geometric dimensions for groups
with torsion, J. London Math. Soc. 64 (2001), 489-–500.
[11] W.-L. Chow. On the algebraical braid group, Ann. of Math. (2), 49 (1948), 654-–658.
[12] D. Degrijse, R. Köhl and N. Petrosyan, Classifying spaces with virtually cyclic stabilizers for linear groups, Transformation Groups 20 (2015), 381–394.
[13] D. Degrijse and N. Petrosyan, Commensurators and classifying spaces with virtually
cyclic stabilizers, Groups Geom. Dyn. 7 (2013) 543—555.
[14] D. Degrijse and N. Petrosyan, Geometric dimension of groups for the family of virtually cyclic subgroups, J. Topol. 7 (2014) 697—724.
[15] D. Degrijse and N. Petrosyan, Bredon cohomological dimensions for groups acting on
CAT(0)-spaces, Groups Geom. Dyn. 9 (2015), 1231—1265.
[16] T. Farrell and L. Jones, Isomorphism conjectures in algebraic K-theory, J. Amer.
Math. Soc., 6 (1993), 249—297.
[17] M. Fluch, On Bredon (co-)homological dimensions of groups PhD Thesis,
arXiv:1009.4633.
[18] M. Fluch and I. Leary, An Eilenberg-Ganea phenomenon for actions with virtually
cyclic stabilisers Groups Geom. Dyn. 8 (2014), 135-–142.
[19] M. Fluch and B. Nucinkis, On the classifying space for the family of virtually cyclic
subgroups for elementary amenable groups, Proc. Amer. Math. Soc. 141 (2013), 3755–3769.
[20] D. Gonçalves and J. Guaschi, The classification of the virtually cyclic subgroups of
the sphere braid groups SpringerBriefs in Mathematics, Springer, 2013.
20
[21] J. González-Meneses, Ordering pure braid groups on compact, connected surfaces,
Pacific J. of Math. 203 (2002) 369–378.
[22] J. González-Meneses, The n-th root of a braid is unique up to conjugacy, Alg. Geom.
Top. 3 (2003), 1103–1118.
[23] J. González-Meneses, Basic results on braid groups, Annales Mathématiques Blaise
Pascal 18 (2011) 15–59.
[24] J. González-Meneses, On reduction curves and Garside properties of braids, Contemp.
Math. 538 (2011), 227–244.
[25] J. González-Meneses and B. Wiest, On the structure of the centralizer of a braid,
Ann. Sci. Éc. Norm. Sup. 37 (2004), 729–757.
[26] J. Guaschi and D. Juan-Pineda, A survey of surface braid groups and the lower algebraic K-theory of their group rings, In Handbook of Group Actions 23–76, Advanced
Lectures in Mathematics 32, International Press of Boston Inc., 2015.
[27] T. Haettel, D. Kielak and P. Schwer, The 6-strand braid group is CAT(0) Geom.
Dedicata 182 (2016), 263-–286.
[28] J. Harer, The virtual cohomological dimension of the mapping class group of an orientable surface, Invent. Math, 84 (1986), 157–176.
[29] S. Hensel, D. Osjada and P. Przytycki, Realisation and dismantlability, Preprint
arXiv:1205.0513.
[30] A. Hurwitz, Uber Riemannsche Flächen mit gegebenen Verzweigungspunkten, Math.
Ann. 39, no. 1 (1891), 1–60.
[31] D. Juan-Pineda and I. Leary, On classifying spaces for the family of virtually cyclic
subgroups Contemp. Math. 407 (2006), 135–145.
[32] D. Juan-Pineda and J. Sánchez-Saldaña, The K and L theoretic Farrell-Jones isomorphism conjecture for braid groups Preprint arXiv:1511.02737.
[33] D. Juan-Pineda and A. Trujillo-Negrete, On classifying spaces for the family of virtually cyclic subgroups in mapping class groups, Preprint arXiv:1606.00306.
[34] P. Kropholler and G. Mislin, Groups acting on finite-dimensional spaces with finite
stabilizers, Comment. Math. Helv. 73 (1998), 122—136.
[35] J. Lafont and I. Ortiz, Relative hyperbolicity, classifying spaces, and lower algebraic
K-theory Topology 46 (2007), 527–553.
[36] W. Lück, Survey on classifying spaces for families of subgroups, in Infinite groups:
geometric, combinatorial and dynamical aspects, 269—322, Progr. Math., 248,
Birkhäuser, Basel, 2005.
21
[37] W. Lück, On the classifying space of the family of finite and of virtually cyclic subgroups for CAT(0)-groups, Münster J. of Math. 2 (2009), 201–214.
[38] W. Lück and H. Reich, The Baum-Connes and the Farrell-Jones conjectures in Kand L-theory In Handbook of K-theory, 703–842, Springer.
[39] W. Lück and M. Weiermann, On the classifying space of the family of virtually cyclic
subgroups, Pure Appl. Math. Q. 8, (2012), 497–555.
[40] D. Macpherson, Permutation groups whose subgroups have just finitely many orbits,
in Ordered groups and infinite permutation groups, 221–230, Kluwer Academic Publishers, 1996.
[41] C. Martı́nez-Pérez, A bound for the Bredon cohomological dimension for groups, J.
of Group Theory 10 (2007), 731–747.
[42] C. Martı́nez-Pérez, Euler classes and Bredon cohomology for groups with restricted
families of finite subgroups, Math. Z. 275 (2013), 761—780.
[43] G. Mislin and A. Valette, Proper group actions and the Baum-Connes conjecture, in
Advanced courses in Mathematics, no. 55, CRM Barcelona, 2003.
[44] D. Rolfsen, J. Zhu, Braids, Orderings and zero divisors, J. Knot Theory Ramif. 7
(1998), no. 6, 837–ñ841.
[45] W. P. Thurston. On the geometry and dynamics of diffeomorphisms of surfaces, Bull.
Amer. Math. Soc., 19 No. 2 (1988), 417-–431.
[46] G.W. Whitehead, Elements of homotopy theory, Graduate Texts in Math., vol 61,
Springer-Verlag, 1978.
Ramón Flores:
Dpto. Geometrı́a y Topologı́a.
Facultad de Matemáticas
Instituto de Matemáticas (IMUS)
Universidad de Sevilla
Av. Reina Mercedes s/n
41012 Sevilla (SPAIN)
E-mail: [email protected]
URL: www.personal.us.es/ramonjflores
Juan González-Meneses:
Dpto. Álgebra.
Facultad de Matemáticas
Instituto de Matemáticas (IMUS)
Universidad de Sevilla.
Av. Reina Mercedes s/n
41012 Sevilla (SPAIN)
E-mail: [email protected]
URL: www.personal.us.es/meneses
22
| 4math.GR
|
Multi-Objective Event-triggered Consensus of
Linear Multi-agent Systems
Amir Amini, Student Member, IEEE, Arash Mohammadi, Member, IEEE and Amir Asif, Senior
arXiv:1702.06285v1 [cs.SY] 21 Feb 2017
Member, IEEE
Abstract
This paper proposes a distributed consensus algorithm for linear event-based heterogeneous multi-agent
systems (MAS). The proposed scheme is event-triggered in the sense that an agent selectively transmits
its information within its local neighbourhood based on a directed network topology under the fulfillment
of certain conditions. Using the Lyapunov stability theorem, the system constraints and event-triggering
condition are expressed in terms of several linear matrix inequalities (LMIs) to derive the consensus parameters. The objective is to design the transmission threshold and minimum-norm heterogeneous control gains
which collectively ensure an exponential consensus convergence rate for the closed-loop systems. The LMI
computed control gains are robust to uncertainty with some deviation from their nominal values allowed.
The practicability of the proposed event-based framework is further studied by proving the Zeno behaviour
exclusion. Numerical simulations quantify the advantages of our event-triggered consensus approach in
second-order, linear and heterogeneous multi-agent systems.
Index Terms— Multi-agent Systems, Event-based Consensus, Multi-objective Control Design,
Linear Matrix Inequalities.
I. Introduction
Among many cooperative behaviors in multi-agent systems (MAS), consensus has attracted considerable
attention due to its vast application in sensor networks, unmanned aerial vehicles (UAV), and mobile
robotic teams. Related works in this area primarily deal with the problem of distributed consensus, where
all agents constantly transmit their information within the network [1]–[4]. Although various interesting
practical features have previously been studied to solve consensus in MASs, continuous data transmission
and continuous control input update are critical restrictions in practice. Therefore, in order to preserve
the limited and valuable energy allocated to local microprocessors installed on each agent, strategies which
decrease data transmission and control input updates are of great interest [5], [6]. In this regard, periodic
time-triggered communication and control scheme were proposed in [7] and [8] to cope with consensus in
first and second-order integrators. Event-triggered broadcasting strategies, however, offer superior performance as they allow agents to transmit information, irrespective of time interval, and only if a predefined
triggering condition is satisfied. Motivated by early results in event-triggered control methods such as [9],
the authors in [10] extended the procedure to address the consensus problem of first-order integrators. More
2
recently, researchers have studied the event-based consensus problem from several aspects. For instance,
in [11] effective triggering rules are proposed to guarantee consensus of second-order agents in undirected
networks. In [12], event-based output consensus problem in heterogeneous linear MASs is investigated again
in undirected topologies. Moreover, the consensus problem of linear and nonlinear second-order MASs is
addressed in [13]. In [14], an edge event hybrid-driven rule is proposed to ensure second-order leader-following
consensus. The aforementioned methods are limited to integrators and are not capable of achieving eventbased consensus in more general linear agents with directed topologies. At the same time, most existing
event-based schemes addressing consensus, overlook control objectives and the closed-loop system is only
guaranteed to be stable [15], [16]. The use of multi-objective optimization is essential in practical applications
with multiple performance requirements [17]. Furthermore, a vast majority of the relevant works only design
a common control gain and share it among all agents to acquire consensus [15], [16], [18]. Such a design
approach is not completely distributed. In a fully distributed structure, each agent should be able to choose
a specific control gain, according to its own dynamics and connectivity within the communication network.
In addition, the aforementioned single-value control design methods are not capable of handling possible
heterogeneity among multi-agent system dynamics. As fully discussed in [19], control design problem for
consensus of heterogeneous MASs is nontrivial, and non-selection of control gains often lead to unstable
system behavior under directed networks.
As a powerful design method, linear matrix inequality (LMI) optimization guarantees system stability for
desired design objectives through convex optimization [20]. For a multi-objective problem such as minimum
gain resilient heterogeneous control design in an event-based communication scheme as is being considered
in this paper, an analytical solution to compute the design parameters is difficult (if not impossible) to
derive. As discussed earlier, analytical solutions proposed in the literature need strong assumptions on agent
dynamics or/and network topology, and no specific closed-loop performance is guaranteed. Formulating the
problem within an LMI framework is a practically feasible solution to pursue in many control applications
[21]–[25]. We note that deriving optimization matrix inequalities in a linear form is a non-trivial effort. Some
suggested consensus approaches result in bilinear matrix inequalities (BMIs) that are even more difficult to
solve [26], [27].
To address the aforementioned limitations, the paper investigates the problem of objective-based control
design for event-triggered consensus in heterogeneous linear MASs. The main contributions of the paper
are listed as follows: 1) To guarantee consensus, we couple the control gain design and event-triggering
function to benefit from multi-objective optimization. The proposed algorithm has an exponential consensus
convergence rate with robust minimum-norm control gains; 2) The proposed approach provides additional
degrees of freedom by designing heterogeneous control gains for event-triggered multi-agent networks. This is
a unique advancement over most of the existing works in event-triggered MASs where a common control gain
is used by the agents; 3) To the best of our knowledge, the event-triggered consensus problem has not been
considered previously in the context of heterogeneous second-order MASs with directed network topologies.
This is the first instance of incorporating multi-objective LMI optimization in consensus problems for second
3
order event-triggered heterogeneous MASs.
The remaining paper is organized as follows. Section 2 introduces required preliminary concepts. The
problem is stated in Section3. In Section 4, we proceed to formulate the multi-objective event-based consensus
problem. The algorithm to derive unknown design parameters within the LMI optimization is proposed in
Section 4. We provide simulation examples in Section 5 to evaluate the capability of the algorithm. Finally,
Section 6 concludes the paper.
II. Preliminaries and Graph Theory
Throughout the paper, we use normal alphabets to denote matrices or scalars, and bold letters to specify vectors. Notation Rm×n refers to (m × n) real-valued matrices. In what follows we present necessary
matrix notation and commonly-used operations for matrix A = {aij } ∈ Rm×n with real entries aij , and
B = {bij } ∈ Rm×n . |A|: Matrix with component-wise absolute value of A, i.e., { |aij | } ∈ Rm×n ; kAk: Frobenius
norm of A; AT : Transpose of A; A† : Pseudo inverse of A; λmin(max) (A): minimum (maximum) eigenvalue of A;
A > 0: A is symmetric positive definite, i.e., xT Ax > 0, ∀ x ∈ Rn ; A ≥ 0: A is symmetric semi-positive definite,
i.e., xT Ax ≥ 0, ∀ x ∈ Rn ; null(A): Null space of A, i.e., {x ∈ Rn | Ax = 0}; A⊗B: Kronecker product of A and
B; A ◦ B: Hadamard product of A and B; a(i,•) : The i-th row of matrix A, i.e., [ ai1 , . . . , ain ]. Similarly, for
vectors u ∈ Rn and v ∈ Rn , the term u ≤ v defines the component-wise inequality, i.e., ui ≤ vi , i = 1, . . . , N .
Moreover, kuk is the Euclidean norm of u, and 1n defines the (n×1) column vector with all elements equal
to one.
In symmetric block matrices, the asterisk ∗ represents the lower triangle block which is induced by
symmetry.
Lemma 1. Schur Complement [28].
Considering matrices R, Q, and S with appropriate dimensions, the following two statements are equivalent.
Q
S
> 0.
R > 0, Q − SR−1 S T > 0 ⇔
∗ R
Lemma 2. S-procedure [28].
Let T0 , T1 ∈ Rn×n be symmetric matrices. If there exists a scalar τ ≥ 0 such that T0 − τ T1 > 0, then the
following inequalities on T0 and T1 are satisfied
xT T0 x > 0,
∀x 6= 0
such that xT T1 x ≥ 0.
(1)
The communication network of a MAS consisting of N agents is modeled using a graph G = (V, E, A), where
V = {1, 2, ..., N } denotes the agent set, i.e., the i-th vertex indicates the i-th agent. The edge set E is defined
as the Cartesian product of the two sets, i.e., E ⊆ V × V. If agent j communicates its information to agent
i, then the pair (j, i) is an element of E denoted by j → i in graph representation. In a directed graph,
(j, i) ∈ E is not equivalent to (i, j) ∈ E. Term A = {aij } ∈ RN ×N denotes the weighted adjacency matrix for
graph G, where aii = 0, aij 6= 0 if (i, j) ∈ E, and aij = 0 if (i, j) ∈
/ E. The neighbor set of agent i is defined
4
by Ni = {j ∈ V | (i, j) ∈ E}. A directed graph contains a directed spanning tree if there exists a node in
the graph which has directed paths to all other nodes. The Laplacian matrix corresponding to G is defined
PN
as L = {lij } = D − A, where D = diag (deg1 , ..., degN ), with degi = j=1 aij . The Laplacian matrix has an
eigenvalue of zero if and only if the directed network contains a directed spanning tree. Under this condition,
all other eigenvalues have positive real components [29].
III. Problem Statement
Consider a multi-agent network system comprising of N agents with the general linear dynamics given by
ẋi (t) = Axi (t) + Bi ui (t),
1 ≤ i ≤ N.
(2)
where xi (t) ∈ Rn is the state vector at time instant t, and ui (t) ∈ Rm is the control input vector. Matrices
A and Bi , with appropriate dimensions, represent the system matrix and control input matrix, respectively.
Despite the common practice in event-based strategies ( [30], [16]) that assumes identical agents across the
network, we consider the input matrix Bi to be different among agents. The agent model defined in (2)
satisfies the controllability assumption for any pairs (A, Bi ). Moreover, the network configuration contains
a directed spanning tree.
Remark 1. As a large class of mechanical systems, we note that the heterogeneous second-order MASs can
be represented by (2).
A proposed distributed protocol ui (t) is said to solve the consensus problem if the following condition is
fulfilled.
Definition 1. Given any initial condition, the consensus problem is solved if and only if the state disagreement norm of any two agents in the network asymptotically converges to zero [31], i.e.,
lim k xi (t) − xj (t) k= 0,
t→∞
∀i, j ∈ V.
(3)
The agents share their information with the neighbors through a directed network to reach a common
state value. However, in order to decrease the number of transmissions in the distributed scheme, an efficient
event-triggering mechanism is of great interest. In the desired event-based strategy for data communication,
agent i observes its own state vector constantly. If a certain proposed condition with a designed threshold is
violated, it transmits the state vector to its neighboring agents. Upon receiving the data, node j, a neighbor
of agent i, updates its information regarding agent i by incorporating the newest received data. Denoting
ti0 , ti1 , . . . as the triggering time sequence of agent i, we define the most recently broadcasted information of
agent i for any interval between two consecutive triggering instants as follows
x̂i (t) = xi (tik ),
t ∈ [tik , tik+1 ).
(4)
5
In order to reach the consensus condition specified in Definition 1, the following distributed protocol for
agent i is proposed
ui (t) = ( Ki + ∆K i (t) )
X
( x̂i (t) − x̂j (t) ).
(5)
j∈Ni
Matrix Ki ∈ Rm×n is the nominal control gain to be designed for agent i, and ∆K i (t) is the additive unknown
norm-bounded, structured uncertainty in the controller parameter [32]. The control law proposed in (5)
depends only on the last transmitted states. Such a control structure leads to an event-based control input
update mechanism, meaning that the actuators receive input signals only on certain instants. Therefore,
the number of control input updates are lower [30]. Another specific characteristic of the proposed eventtriggered protocol (5) is assigning different control gains to individual agents, leading to a heterogeneous
controller design. Performance degradation due to the sensitivity of the closed-loop system to inaccuracies
in the control’s coefficients at the implementation stage is inevitable. Therefore, it is required to provide
some level of robustness in the design of the control parameters. Noting that the perturbation in control
parameters mainly happens due to inaccurate system modeling or round off errors, a practical solution
considered here is to develop resilient control design techniques. Since a rapid non-conservative convergence
rate is desirable in most applications, the control gain Ki ’s, here, are supposed to assure a sufficient fast
consensus with an exponential rate [33]. Unlike most work where convergence rate is neglected in the control
design, we are therefore interested to incorporate conditions for exponential stabilizing with other design
features. In the design of the proposed control law, we also minimize the norm of Ki ’s in order to avoid
large undesirable control inputs.
To summarize, we will design the event-triggered control law defined in (5) to reach consensus by assuring
the stability of the closed-loop system as the primitive focus while satisfying these objectives: I) Minimizing
the number of transmissions among the agents, leading to a lower number of control input updates; II)
Efficient performance by designing heterogeneous control parameters; III) Robustness to a predefined level
of uncertainty in the obtained control parameters; IV) Exponential rate of convergence, and; V) Minimizing
control gains to decrease control input effort.
IV. problem formulation
Let ei (t) = x̂i (t) − xi (t) denote the measurement error between the most recently transmitted state and
its instantaneous value for agent i. For the benefit of analyzing the MAS in a collective manner, we define
x(t) = [ xT1 (t) , . . . , xTN (t) ]T as the stacked state vector, and x̂(t) = [ x̂T1 (t) , . . . , x̂TN (t) ]T as the stacked vector
T
for the last transmitted states. We also define e(t) = [ eT1 (t) , . . . , eTN (t) ] as the stacked measurement error
vector, which is equivalent to
e(t) = x̂(t) − x(t).
(6)
Now, we combine (2) with the proposed controller (5) to obtain the following augmented closed-loop system
ẋ(t) = AhNi + B (K + ∆K (t)) Lhni x(t) + B ( K + ∆K (t) )Lhni e(t),
(7)
6
where Lhni = L ⊗ In . The new variable AhNi = IN ⊗ A is the global system matrix, and block-diagonal matrix
B = diag ( B1 , . . . , BN ) is the global control input matrix. The unknown control gains are accumulated in
matrix K = diag (K1 , . . . , KN ) and their corresponding, possibly time-variant, perturbation are denoted by
∆K (t) = diag ( ∆K 1 (t) , . . . , ∆K N (t) ). The latter matrix satisfies the following assumption.
Assumption 1. The following predefined upper bound (threshold) holds for ∆K (t) in all time instants t.
k ∆K (t) k ≤ δ.
(8)
A. System Transformation
Before we proceed to present the event-triggering scheme and control design procedure, it is necessary to
note that if A in (2) is unstable, then system matrix in (7) will be unstable since Lhni always contains a
zero eigenvalue [34]. Therefore, the goal is to reach consensus by establishing the stability of a transformed
version of the system. Another reason for using this approach is to utilize the well-developed results in the
Lyapunov method, which also provides a variety of performance indices beside ensuring the stability of the
system. Hence, we convert the consensus problem at hand, i.e., the closed-loop system defined in (7), into
an equivalent stability problem using an appropriate transformation. A proper state transformation, e.g.,
x̄(t) = T x(t), for achieving this objective needs to satisfy the following two conditions: (i) First, stability of
such a transformed system, x̄(t), must be equivalent to the consensus problem for the closed-loop system
defined by (7), and; (ii) All heterogeneous parameters Bi , and Ki should be involved in the transformed
system. It is worth mentioning that the transformations suggested in a majority of related works, such as [15]
and [34], are incapable of meeting the latter challenge and, thus, heterogeneous control gain design is not
applicable in such approaches.
In order to achieve the consensus condition defined in Definition 1, the Laplacian matrix is first nominated
as the transformation matrix, i.e., x̄(t) = L ⊗ In x(t). Using Laplacian L as the transformation matrix,
however, would result in a singular system, since L is not full ranked and thus the set of state disagreement
emerging in x̄(t) would be linearly dependent. Since dealing with singular systems brings several technical and
analytical difficulties [35], we remove one row of L to design the transformation matrix. The proposed solution
will eliminate system redundancy and provides a reduced full-rank system. Therefore, we let L̂ ∈ R(N −1)×N
denote a matrix which is obtained by removing one arbitrary row of the Laplacian matrix. The proposed
state transformation is, therefore, given by
xr (t) = L̂hni x(t),
(9)
where L̂hni = L̂ ⊗ In .
Lemma 3. It follows from (9) that xr (t) = 0 if and only if x1 (t) = · · · = xN (t). The consensus condition in
Definition 1 is satisfied when xr (t) = 0.
Proof. If xr (t) = 0, then according to (9) we have L̂hni x(t) = 0, which means x(t) belongs to the null space
of L̂hni , i.e., x(t) ∈ null (L̂hni ). Since the row sum of L, and similarly L̂, is zero, the null space of L̂hni is
7
given by 1N ⊗ xcns (t), i.e., x(t) ∈ null ⊗ (L̂hni ) = 1N ⊗ xcns (t), where xcns (t) is the consensus vector to which
all xi (t), (1 ≤ i ≤ N ), converge. Therefore, it is concluded that x1 (t) = · · · = xN (t) = xcns (t). Accordingly,
the consensus equation defined in (3) is satisfied. The statements are bidirectional in the sense that if
x1 (t) = · · · = xN (t) = xcns (t), then xr (t) = 0 holds.
According to Lemma 3, the consensus problem for system (7) is equivalent to the stability problem of the
system expressed in terms of transformation (9). It is worth mentioning that the consensus vector xcns (t)
may be constant or time-varying depending on the dynamics of the MAS.
Lemma 4. L̂hni AhNi = AhN−1i L̂hni , where AhN−1i = IN −1 ⊗ A.
Proof. L̂hni AhNi = (L̂ ⊗ In )(IN ⊗ A) = (L̂hni IN )⊗ (In A) =(IN −1 L̂)⊗ (AIn ) = (IN −1 ⊗ A)(L̂ ⊗ In ) = AhN−1i L̂hni .
Using Lemma 4, the closed-loop system given in (7) is transformed to the following reduced order structure,
ẋr (t) = ( AhN−1i + A + ∆A ) xr (t) + ( A + ∆A ) er (t),
(10)
†
where A = L̂hni BKL, ∆A = L̂hni B∆K (t)L, and er (t) = L̂hni e(t), with L = Lhni L̂hni
. The reduced measurement
error for the closed-loop system given in (10) is er (t) = x̂r (t) − xr (t), where x̂r (t) = L̂hni x̂(t).
Remark 2. Without loss of generality and for the sake of brevity in notation, we remove row N from the
Laplacian matrix L, to derive L̂.
The exponential stability for system (10) is defined below.
Definition 2. Given damping coefficient ζ > 0, system (10) is ζ-exponentially stable if there exists a positive
scalar c such that xr (t) satisfies the following condition [36]
k xr (t) k≤ ce−ζt k xr (0) k,
t ≥ 0.
(11)
In the following section, we proceed to introduce and formulate the event-triggering mechanism.
B. Event-triggering scheme
hni
hni
We define the disagreement vector for agent i as X̂i (t) = l(i,•) x̂(t), with l(i,•) = l(i,•) ⊗ In . In fact, X̂i (t)
provides the instantaneous disagreement between the last transmitted state corresponding to agent i and
T
T
the last received states from its neighbors. Let X̂(t) = [ X̂1 (t), . . . , X̂N (t) ]T denote the stacked disagreement
vector. Given tik , the next triggering instant for agent i is, therefore, determined from the following condition
tik+1 = inf { t > tik : h ( ei (t), X̂i (t), φ ) ≥ 0},
(12)
where h ( ei (t), X̂i (t), φ ) = kei (t)k−φkX̂i (t)k, and real-valued scalar φ > 0 is the transmission threshold to be
determined. Note that the triggering function given in (12) is asynchronous, i.e., each agent independently
decides on its own triggering time. The primary goal here is to determine the maximum stable value for φ,
which provides the minimum number of transmissions for a particular network configuration with guaranteed
8
control performances. Between two consecutive events for agent i, the triggering function is non-positive,
i.e., hi ≤ 0. Thus, we consider the following component-wise inequality derived based on (12)
e[Nr] ≤ φX̂
where e[Nr] = [ ke1 (t)k , . . . , keN (t)k ]T , and X̂
[Nr]
[Nr]
,
(13)
= [ kX̂1 (t)k , . . . , kX̂N (t)k ]T . In order to merge the design of
maximum possible transmission φ with desired control objectives, the event-triggering condition (13) needs
to be expressed as a function of the system’s state variables, i.e., xr (t) and er (t). In this regard, the following
two Lemmas are introduced to transform (13) into the required structure.
Lemma 5. If a certain value φ satisfies (13), the following entry-wise inequality is also satisfied
L̂e[Nr] ≤ φ |L̂| X̂
[Nr]
.
(14)
Proof. Component i in L̂e[Nr] is computed as l(i,•) e[Nr] , (1 ≤ i ≤ N − 1). Therefore, the entries of e[Nr] are
multiplied by exactly one positive and at least one negative value in a network containing a directed spanning
tree. According to the Euclidean normed-space properties, the absolute value of row vector l(i,•) lies within
an upper bound, i.e.,
l(i,•) e[Nr] ≤ φ |l(i,•) | X̂
[Nr]
1 ≤ i ≤ N − 1.
,
(15)
Expanding (15) for all rows of L̂ results in (14).
Lemma 6. Denote matrix M = { mij } = { lij + αj liN }, 1 ≤ i ≤ N −1 and 1 ≤ j ≤ N −1, as the correlation
matrix, with α = [ α1 , . . . , αN −1 ] = l(N ,•) L̂† . If a certain value φ satisfies the following entry-wise inequality
ψe ≤ ψx̂ ,
(16)
iT
h
hni
hni
e(t)k ,
it also satisfies inequality (14). The undefined vectors in (16) are ψe = kl(1,•) e(t)k , . . . , kl(N
−1,•)
iT
h
hni
hni
hni
x̂r (t)k , with m(i,•) = m(i,•) ⊗ In .
and ψx̂ = k φm(1,•) x̂r (t)k , . . . , kφm(N
−1,•)
Proof. Based on the reverse triangle inequality in the Euclidean normed-space, we conclude that
hni
l(i,•) e[Nr] ≤ k l(i,•) e(t) k,
1 ≤ i ≤ N − 1,
(17)
which is equivalent to L̂e[Nr] ≤ ψe considering all rows. On the other hand, the sub-additivity property in
the Euclidean normed-space proves that
hni
φl(i,•) X̂(t) ≤ φ |l(i,•) | X̂
[Nr]
1 ≤ i ≤ N − 1.
,
(18)
Since X̂(t) is formed by the row space of Laplacian matrix, a certain component in X̂(t), e.g., X̂N (t), is always
dependent on the other components; meaning that X̂N (t) can be written as a linear combination of X̂1 (t) to
X̂N −1 (t), i.e., X̂N (t) = α1 X̂1 (t) + · · · + αN −1 X̂N −1 (t). Thus, the coefficients αi are calculated as α = l(N ,•) L̂† . By
substituting X̂N (t) with its linear equivalent value, inequality (18) reduces to
hni
φm(i,•) x̂r (t) ≤ φ |l(i,•) | X̂
[Nr]
,
1 ≤ i ≤ N − 1.
(19)
9
We conclude from (19) that ψx̂ ≤ φ|L̂|X̂
[Nr]
. Under the assumption outlined in inequality (16), the outcome
of (17) and (19) is the following sequence of inequalities
L̂e[Nr] ≤ ψe ≤ ψx̂ ≤ φ |L̂| X̂
[Nr]
,
(20)
which complete the proof for Lemma 6.
In conclusion, Lemma 6 states that any obtained value for φ which satisfies (16), also satisfies (14). Furthermore, inequality (14) is equivalent to the triggering condition defined in (13) according to Lemma 5.
Inequality (16) is favorable in the sense that it can be expressed as a global quadratic constraint in the
T
form of eTr (t)er (t) ≤ x̂Tr (t)Mhni
Φ2 Mhni x̂r (t), where Φ = φI(N −1 )n , and Mhni =M ⊗ In . Replacing x̂r (t) with
er (t) + xr (t), the equivalent condition is given below.
T
T
eTr (t)er (t) ≤ ( er (t) + xr (t) ) Mhni
Φ2 Mhni ( er (t) + xr (t) ).
(21)
Inequality (21) represents the event-triggering constraint that is expected to appear in the convex optimization framework. Once the feasible transmission threshold φ is obtained through optimization, the desired
event condition defined in (12) is exploited to determine the triggering moments for each agent. Inequality
(21) is, therefore, guaranteed according to Lemma 5 and Lemma 6.
Remark 3. Although transmission threshold φ seems to depend only on xr (t), er (t), and Mhni from (21),
we will see that φ is also affected by other design parameters emerging in the optimization problem.
C. Main Result
The following theorem computes the minimum-norm resilient heterogeneous control gain Ki ’s and maximum
transmission threshold φ used in our event-based consensus algorithm.
Theorem 1. The optimum values for the transmission threshold φ and control gains Ki ’s (1 ≤ i ≤ N ) are
computed from
φ=
p
τ3 γ −1 ,
and Ki = Bi† P−1 Θi ,
(22)
which are conditioned on the existence of matrices Θi ∈ Rn×n (1 ≤ i ≤ N ), symmetric positive definite matrix
P ∈ Rn×n , and positive scalars τj (1 ≤ j ≤ 3). Under such conditions, the following minimization derives the
minimum-valued positive scalars γ, µ, and υi for (1 ≤ i ≤ N )
min
Θi ,γ,τj ,P,υi ,µ
subject to:
Π1 Π2
< 0,
∗ Π3
γ+µ+
N
X
υi ,
for (1 ≤ i ≤ N ),
i=1
µI
I
∗
P
> 0,
−Υ ΘT
∗
−I
< 0,
(23)
10
where
Π1 =
π11
ΞL
∗
τ2 δ 2 LT L − τ3 I
Θ = diag ( Θ1 , . . . , ΘN ),
,
Π2 =
P L̂hni B
P L̂hni B
T
τ3 Mhni
0
0
T
τ3 Mhni
Υ = diag (υ1 In , . . . , υN In ) ,
,
Π3 = diag (−τ1 I, −τ2 I, −Γ) ,
(24)
with
T
π11 = AhN−1i
P + P AhN−1i + 2ζP + τ1 δ 2 LT L + ΞL + LT ΞT ,
Ξ = L̂ ⊗ 1n 1Tn ◦ 1N −1 ⊗ [ Θ1 , . . . , ΘN ] ,
Γ = γIn(N −1 ) ,
and P = IN −1 ⊗ P.
(25)
In the above terms, parameters {δ, ζ} are constants with know positive values. The designed parameters stabiq
lizes the system defined in (10) at the ζ-exponential rate kxr (t)k < ce−ζt kxr (0)k, where c = λmax (P)λ−1
min (P)
. The control objectives defined in section III are simultaneously guaranteed.
Proof. To derive the stability conditions for the closed-loop system defined in (10), we consider the Lyapunov
function candidate
V (t) = xT
r (t)P xr (t).
(26)
V̇ (t) + 2ζV (t) < 0,
(27)
Now, consider the following inequality
where the time derivative of V (t) is defined as V̇ (t). The condition defined in (27) is equivalent to V (t) <
V (0)e−2ζt . Considering (26), we obtain λmin (P)kxr (t)k2 ≤ V (t) < V (0)e−2ζt ≤ λmax (P)e−2ζt kxr (0)k2 , which
results in kxr (t)k< ce−ζt kxr (0)k, with the c defined in Theorem 1. Therefore, the condition given in (27) is
the sufficient constraint to ensure ζ-exponential stability according to Definition 2. Now V̇ (t) is expanded
according to the reduced closed-loop system (10) as follows
V̇ (t) = ẋT
(t)P xr (t) + xT
(t)P ẋr (t)
r
r
T
(t)P Ax xr (t) + Aer (t) + L̂hni Bσ1 + L̂hni Bσ2 ,
= Ax xr (t) + Aer (t) + L̂hni Bσ1 + L̂hni Bσ2 P xr (t) + xT
r
(t), σ1T , σ2T ]T ,
(t), eT
where Ax = ( AhN−1i + A ), σ1 = ∆K (t)Lxr (t), and σ2 = ∆K (t)Ler (t). Defining Ω = [ xT
r
r
one can rearrange (27) in terms of Ω to obtain the following matrix structure
T
Ax P + P Ax + 2ζP P A P L̂hni B P L̂hni B
∗
0
0
0
Ω < 0.
ΩT
∗
∗
0
0
∗
∗
∗
(28)
0
To formulate proper quadratic conditions with respect to σ1 and σ2 , we use the upper bound norm for
control uncertainties based on Assumption 1
σ1T σ1 = xT
(t)LT ∆2K (t)Lxr (t) ≤ δ 2 xT
(t)LT Lxr (t),
r
r
(29)
11
σ2T σ2 = eT
(t)LT ∆2K (t)Ler (t) ≤ δ 2 eT
(t)LT Ler (t).
r
r
(30)
The performance-related constraints derived in (21), (29), and (30) are required to be included in the
stability constraint (28). Repeatedly using Lemma 2, the aforementioned constraints along with the new
slack variables τ1 , τ2 and τ3 appear in the
π̄11
∗
Π̄ =
∗
∗
with
following integrated matrix inequality
π̄12 P L̂hni B P L̂hni B
π̄22
0
0
< 0,
∗
−τ1 I
0
∗
∗
(31)
−τ2 I
T
π̄11 = ATx P + P Ax + 2ζP + τ1 δ 2 LT L + τ3 Mhni
Φ2 Mhni ,
T
Φ2 Mhni ,
π̄12 = P A + τ3 Mhni
and
T
Φ2 Mhni .
π̄22 = τ2 δ 2 LT L − τ3 I + τ3 Mhni
Now, we apply Lemma 1 to obtain the inequality below
AT P + P Ax + 2ζP + τ1 δ 2 LT L
PA
P L̂hni B
x
2
T
∗
−τ3 I + τ2 δ L L
0
∗
∗
−τ1 I
∗
∗
∗
∗
∗
∗
P L̂hni B
0
0
−τ2 I
∗
T
Φ
τ3 Mhni
T
Φ
τ3 Mhni
< 0.
0
0
−τ3 I
(32)
Pre and post-multiplying (32) with the positive definite matrix Q = diag (I, I, I, I, Φ−1 ) results in inequality
T
ATx P + P Ax + 2ζP + τ1 δ 2 LT L
PA
P L̂hni B P L̂hni B τ3 Mhni
T
∗
−τ3 I + τ2 δ 2 LT L
0
0
τ3 Mhni
(33)
∗
∗
−τ1 I
0
0 < 0.
∗
∗
∗
−τ2 I
0
∗
∗
∗
∗
−τ3 Φ−2
The matrix inequality derived in (33) is not linear as long as the optimization variables are multiplied by
each other or their inverse form exists. Hence, the blocks containing P Ax , P A, and τ3 Φ−2 need to be handled
in such a way that the ultimate inequality turns into a linear structure. To this end, we first expand P A
according to P defined in (25). Using Hadamard and Kronecker product, it is straightforward to derived the
equation below
P A = L̂ ⊗ 1n 1Tn ◦ 1N −1 ⊗ [ PB1 K1 , . . . , PBN KN ] L.
Defining Θi = PBi Ki , (1 ≤ i ≤ N ), as alternative variables, the inequality (33) becomes linear with respect to
Θi ’s, thus the term Ξ given in (25) is obtained. The same procedure is applicable for handling τ3 Φ−2 . Defining
Γ = τ3 Φ−2 = γIn(N −1 ) , the maximization problem over Φ is converted to an equivalent convex minimization
problem over Γ. The resulting inequality will now be a linear one with respect to Γ and τ3 . Note that τ3 is
not multiplied by any other variables except Φ. Finally, to make the obtained control gains implementable,
12
we minimize the size of Ki ’s by adding appropriate constraints. As the two components which are used to
compute control gains, we restrict the norm of Θi ’s and P−1 by setting the following minimization conditions
on µ and Υ [37]
P−1 < µI,
µ > 0,
(34)
and ΘTi Θi < υi I,
υi > 0.
(35)
The Schur complement lemma is enough to derive the two corresponding LMIs in (23). Once the optimization
problem is solved, unknown variables τ3 , P , Θi and γ are obtained. The Control gains and event-triggering
threshold coefficient are consequently derived from (22).
Remark 4. We note that, design parameters can alternatively be obtained directly from (32) by solving
BMIs. However, this approach is more computationally challenging and provides no guarantee of global
optimization.
D. Zeno Behavior Exclusion
From an implementation point of view, in an event-triggering scheme, there must always be a finite number
of triggering instants within a given finite time interval. Otherwise, the triggering mechanism would exhibit
Zeno behavior [38]. It is essential to prove that the time interval between any two events are strictly positive
for all agents. The following theorem provides the lower bound on the interval between two consecutive
triggering instants.
Theorem 2. Considering system (2), control law (5), event-triggering function (12), and design parameters
(22), the inter-event interval for agent i is strictly positive and lower bounded by the following term
!
i
1
φkAkk
X̂
(t
)k
i
k
tik+1 − tik ≥
ln
+1 ,
k
kAk
F
(36)
i
k
where F i =
max
t∈[tik ,tik+1 )
kAx̂i (t) + Bi (Ki + ∆K i (t))X̂i (t)k, and kX̂i (t)k < δc i . Positive-valued scalar δc i is the
given stopping threshold for agent i, (1 ≤ i ≤ N ).
Proof. Consider an interval t ∈ [tik , tik+1 ) for agent i. Based on the event-triggering mechanism discussed in
section IV-B, ei (tik ) = 0. Then, ei (t) evolves from zero with the following dynamics until tik+1 is determined
by (12) and φ from Theorem 1.
d
kei (t)k≤ kẋi (t)k≤ A ( x̂i (t) − ei (t) ) + Bi ( Ki + ∆K i (t) ) X̂i (t) ≤ kAk kei (t)k+Fi (t),
dt
where Fi (t) = kAx̂i (t) + Bi ( Ki + ∆K i (t) ) X̂i (t)k. One can solve (37) for ei (t) as follows
Z t
Fi (η)ekAk(t−η) dη, t ∈ [tik , tik+1 ).
kei (t)k≤
tik
(37)
13
Algorithm 1 : Proposed Event-based Consensus
Input: Adjacency Weighted Matrix A = {aij }, Agents’ dynamics given in (2).
Output: Multi-objective Event-triggered Consensus.
Parameter Design: (D1 – D5)
I. Initialization
D1. Transformation Matrix: Remove N th row of Laplacian matrix L in order to determine the reduced
Laplacian matrix, L̂.
D2. System Transformation: Derive reduced system (10).
D3. Correlation Matrix: Using Lemma 6, determine correlation matrix Mhni .
II. Optimization
D4. Solving the LMIs: Using convex optimization solvers, solve the LMIs (23) for given parameters {δ, ζ}.
D5. Feasibility Verification: If a solution exists for (23), obtain φ, and Ki ’s from (22). Otherwise, change
parameters {δ, ζ}, and repeat step D4.
Event-triggered Consensus: (C1 – C3)
C1. Initialization: Initialize consensus process by allowing all agents to transmit their initial states xi (0) to
neighbours.
C2. Execution: Using Ki ’s derived in Step D4, the states of agent i in (2) are excited by local controller given
in (5). Triggering condition (12) is responsible to determine the next state transmission to neighbours
for agent i as the states evolves to reach consensus.
C3. Consensus Achievement: Agent i repeats Step C2 until convergence is achieved for the disagreement
state vector, i.e., kX̂i (t)k< δc i .
Incorporating (37) with the event-triggering mechanism kei (t)k ≤ φkX̂i (t)k, the next triggering instant for
agent i does not happen until the right-hand side of (37) evolves from zero to reach φkX̂i (tik )k. Then it
follows from (37) that
k
d
dt kei (t)k ≤ kAk kei (t)k+F i ,
or equivalently
k
kei (t)k≤
The next event is triggered at t = tik+1 when
F i kAk(t−tik )
e
−1 .
kAk
(38)
k
kei (tik+1 )k= φ kX̂i (tik )k≤
F i kAk(tik+1 −tik )
e
−1 ,
kAk
(39)
which simplifies to (36). Observe that during the consensus process X̂i (tik ) > δc i , thus the right hand side of
(36) is strictly positive and tik+1 − tik > 0.
The event-based consensus algorithm is summarized in Algorithm 1.
14
V. Simulations
A. Deterministic example
Consider a network of six second-order heterogeneous agents with the following dynamics [39]
m̄i r̈i (t) = [ 1 + ∆ui ] ui (t),
1 ≤ i ≤ 6,
(40)
where ri (t) ∈ R and m̄i > 0 defines, respectively, the position and inertia of agent i. Inequality |∆ui | < 1
represents the uncertainty in the control input ui (t) ∈ R due to dis-adjustment of actuators. Equation (40)
can be rearranged as ( m̄i /(1 + ∆ui ) )r̈i (t) = ui (t). As suggested in [39], the term (m̄i /(1 + ∆ui )), denoted as
mi , is treated as the new inertia for agent i. The state space representation for (40) is, therefore, given by
ṙi (t) = vi (t)
mi v̇i (t) = ui (t),
1≤i≤6
(41)
where vi (t)∈R denotes the velocity of agent i. Unlike previous work where the inertia mi ’s and uncertainty
∆ui ’s in the second-order MASs are not considered separately for each agent, i.e., mi = 1 are assumed, model
(41) considers a more practical scenario with heterogeneous inertias. Among various choices suggested to
simulate heterogeneous inertias in literature, we consider mi = 0.8 + 0.1i, (1 ≤ i ≤ 6), as in [39]. The state
space representation for (41) with respect to (2) is, hence, given by xi (t) = [ ri (t), vi (t) ]T , A = [ 0, 1; 0, 0 ],
B1 = [0, 0.9]T ; B2 = [ 0, 1.0 ]T ; B3 = [ 0, 1.1 ]T ; B4 = [ 0, 1.2 ]T ; B5 = [ 0, 1.3 ]T ; and B6 = [ 0, 1.4 ]T . The
directed network configuration corresponded to the (41) is described by the asymmetric Laplacian matrix
L = [3, 0, 0, −1, −1, −1; 0, 2, 0, 0, −1, −1; 0, 0, 2, −1, 0, −1; 0, −1, 0, 2, 0, −1; −1, −1, 0, −1, 3, 0; −1, −1, −1, 0, 0, 3].
To solve the consensus problem using Theorem 1, we initialize the LMI optimization with ζ = 0.4 and
δ = 0.02. Using the YALMIP parser and SDPT3 solver, we solve (23) with the aforementioned values for the
optimization and system parameters [40]. The parameters obtained from the LMI optimization (23) are
0.0608 −0.0363
, τ3 = 0.1312, γ = 5.032.
P=
(42)
−0.0363 0.1020
Using (22), the control gains are derived as follows: K1 = [ −0.1187, −0.2952 ], K2 = [ −0.1933, −0.2953 ],
K3 = [ −0.1965, −0.4292 ], K4 = [ −0.1391, −0.2523 ], K5 = [ −0.2413, −0.3472 ], and K6 = [ −0.1842, −0.1659 ].
The maximum transmission threshold is also calculated from (22) as φ = 0.1614. In order to observe the state
trajectories of the closed-loop heterogeneous multi-agent system (40) with the designed parameters, we pick
initial values for xi (0) = [ i + 5, i − 2 ]T , (1 ≤ i ≤ 6). Moreover,the time-varying uncertainty in the control
gains is assumed to be ∆K i (t) = √12 sin(t) [ 0.02, 0.02 ] for all agents. Recall that k∆K (t)k ≤ δ. Computed
with discretization intervals Ts = 10−3 sec, the state trajectories of the six agents are shown in Figure 1(a).
We further define the convergence criteria for the consensus process as δc i = 5 × 10−3 . Figure 1(b) plots the
control inputs as defined in (5) with respect to the aforementioned free or obtained parameters. Figure 2
is included to verify that the obtained parameters, i.e., Ki ’s and φ, are capable of ensuring ζ-exponential
convergence among the agents for ζ = 0.4 and δ = 0.02.
15
(b)
(a)
25
agent 1
agent 2
agent 3
agent 4
agent 5
agent 6
4
20
3
Position, xi (t)
15
Control input
Position and Velocity Trajectories
5
Velocity, vi (t)
10
5
2
1
0
-1
-2
-3
0
-4
-5
0
2
4
6
8
10
12
-5
0
Time (sec)
2
4
6
8
10
12
Time (sec)
Fig. 1: trajectories of the multi-agent system; (a): State consensus, (b): Control inputs ui (t).
The simulation results show that it takes 12412 iterations to achieve consensus in this experiment. However,
the six agents, respectively, transmit their information on 807, 781, 317, 222, 311, and 140 occasions during
the consensus process. The average number of data transmission per agent in this configuration is, therefore,
429.67 times. We also define a criteria to measure the efficiency of the event-based method in terms of the
saved amount of transmission in percentage, denoted by ST. The ST index is expressed as a function of
the average number of transmissions per agent, defined as AT , and total number of iterations to reach
consensus, denoted by TI, i.e,
ST% = 1 − AT/TI × 100.
(43)
Using the definition given in (43), 96.54% of the total possible transmissions are saved in this example.
Another crucial performance-related factor in a comparison is the amount of control force consumption
during the process. In order to consider the effect of control input expense, we measure the well-known
R∞
P
T
input cost function Ju = N
i=1 0 ui (t) ui (t) dt in our analysis [41]. In the current experiment, the control
cost is calculated as Ju = 57.0516. We will use the value of Ju to compare the control expense in consensus
processes.
It is also interesting to study how the performance indices are affected if agents are intentionally allowed
to transmit at a higher rate. To this end, we manually reduce the initially obtained φ = 0.1614 to let agents
benefit from receiving more data from their neighbors. The results are summarized in Table I. All other
parameters remain the same as denoted previously. According to Table I, when the agents are allowed to
transmit more data with lower values of φ, the save transmission ST is reduced as a result of higher average
transmission per agent AT.
However, transmitting more information to the neighbors does not necessarily result in a faster consensus
after a certain value as illustrated by TI. Moreover, the control cost Ju constantly gets reduced as a result of
more communication. In other words, more transmission helps achieving smoother trajectories and control
16
TABLE I: Consensus performance, dicreasing φ with {δ = 0.02, ζ = 0.4}
Transmission
TI
AT
ST%
Ju
0.12
11920
461.17
96.13
50.18
0.08
12651
742.00
94.13
44.05
0.04
13272
1439.00
89.15
39.06
0.00
13586
13586
0
34.72
threshold, φ
TABLE II: Consenssus performance, increasing δ with ζ = 0.4
control
φ
TI
AT
ST%
Ju
0.010
0.1612
19184
699.83
96.35
55.26
0.020
0.1614
12412
429.67
96.54
57.05
0.030
0.1616
10400
221.33
97.87
61.81
0.040
0.1617
12438
191.67
98.45
65.81
0.050
0.1617
14369
366.50
97.44
66.78
uncertainty δ
inputs, but not necessarily at a faster convergence rate. In conclusion, once the φ is derived from (22), it is
always possible to run the algorithm at a lower φ, as a trade-off between AT and Ju .
In this part, we investigate how different choices of constraint-related values {ζ, δ}, will affect the consensus
performance. Table II shows the effect of δ being increased until the LMIs (23) become infeasible while ζ
is fixed. According to Table II, a larger uncertainty δ leads to a slightly larger threshold φ in the feasibility
region of (23). On the other hand, an increment in δ always results in relatively larger control gains. These
two contrasting outcomes, result in a parabolic trend in TI. Therefore, it is difficult to predict AT and
TI beforehand. Greater uncertainty in control gains, results in a more fluctuating control input with huge
jumps, thus Ju is ascending.
In the next experiment we focus on the effect of ζ on consensus performance. The results are summarized
in Table III. As we increase the value of the ζ, larger Ki ’s are obtained from the optimization LMI’s to
accelerate the convergence process. Consequently, the actuators are forced to implement greater control
60
ceζt kxr (0)k
Exponential Stability
50
kxr (t)||
40
30
20
10
0
0
2
4
6
8
10
Time (sec)
Fig. 2: Exponential stability for xr (t)
12
17
2.5
(a)
×104
(b)
1100
TI N=8
TI N=12
TI N=16
1000
AT N=8
AT N=12
2
AT N=16
N=8
N=16
800
1.5
N=12
700
Ju
TI & AT
900
600
1
500
400
0.5
300
200
0
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.2
0.25
0.3
Damping coefficient ζ
3.5
×104
0.35
0.4
0.45
(c)
(d)
1500
TI N=8
TI N=12
TI N=16
AT N=8
AT N=12
AT N=16
3
2.5
N=8
N=12
N=16
1000
u
2
J
TI & AT
0.5
Damping coefficient ζ
1.5
500
1
0.5
0
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
0
0.004
0.006
0.008
0.01
Control Uncertainty (δ)
0.012
0.014
0.016
0.018
0.02
Control Uncertainty (δ)
Fig. 3: Monte Carlo results for N = {8, 12, 16}; (a): TI and AT, varying ζ with fixed δ = 0.01, (b): Control cost Ju ,
varying ζ with fixed δ = 0.01, (c): TI and AT, varying δ with fixed δ = 0.4, (d): Control cost Ju , varying δ with fixed
δ = 0.4.
TABLE III: Consensus performance; increasing ζ with δ = 0.02
damping
φ
TI
AT
ST%
Ju
0.10
0.1601
24998
424.83
98.30
40.59
0.20
0.1607
20916
720.17
96.55
45.69
0.30
0.1611
16466
739.83
95.50
50.71
0.40
0.1614
12412
429.67
96.53
57.05
0.50
0.1616
9249
199.00
97.84
63.20
0.60
0.1617
9064
354.50
96.08
69.29
coefficient ζ
input with more fluctuation which results in more control cost. In conclusion, as ζ is increased the TI
constantly gets reduced at the expense of more Ju , while ST remains relatively the same.
B. Monte-Carlo Simulations
All scenarios in section V-A were based on a single network of six agents. As the adjacency matrix A
affects the feasibility region of (23), a Monte-Carlo simulation is included to study heterogeneous eventbased consensus in randomly generated networks including a spanning tree. The inertias are also generated
18
randomly as mi = 1 + 0.1 Xi , where Xi is a Gaussian random variable with zero mean and unit variance. In
the first Monte-Carlo experiment the control uncertainty δ = 0.01 is fixed, and ζ is gradually increased from
0.2 to 0.5 for different network sizes with N = {8, 12, 16} ( Figures 3(a) and 3(b) ). In the second scenario, δ
is gradually increased with fixed ζ = 0.4 ( Figures 3(c) and 3(d) ). The following facts are observed through
the two scenarios. (i) According to Figure 3(a), the value of save transmission ST is reduced as the network
is expanded; (ii) As shown in Figure 3(a), TI and AT are more close to each other in larger networks as
compared to smaller configurations; (iii) As observed in Figure 3(b), increasing the value of ζ results in a
faster consensus convergence rate; (iv) As shown in Figure 3(c), the parabolic patten is detectable for TI
and AT over randomly generated networks of different size; (v) As illustrated in Figure 3(d), the control
cost is raised when Ki ’s are subjected to a larger uncertainty.
VI. Conclusion
The paper addresses the problem of event-based consensus with predefined objectives for a class of
heterogeneous (multi-agent systems) MASs configured in directed networks. The augmented closed-loop
system is transformed to an equivalent reduced system. The Lyapunov stability theorem is then used to
incorporate the control objectives (exponential convergence rate, resilient control design, and minimum
gain design) within an event-triggered function. The heterogeneous control gains and the transmission
threshold are co-designed by solving an LMI-based optimization problem. It is also proved that the triggering
mechanism does not exhibit the Zeno behavior. The effectiveness of the proposed algorithm is studied through
simulations for heterogeneous second-order MASs. In future, we are interested in applying the proposed
algorithm to even-based state estimation problems in sensor networks.
References
[1] A. Amini, A. Azarbahram, and M. Sojoodi, “H∞ consensus of nonlinear multi-agent systems using dynamic output feedback
controller: an lmi approach,” Nonlinear Dynamics, vol. 85, no. 3, pp. 1865–1886, 2016.
[2] Z. Li, W. Ren, X. Liu, and M. Fu, “Consensus of multi-agent systems with general linear and lipschitz nonlinear dynamics
using distributed adaptive protocols,” IEEE Trans. Autom. Control, vol. 58, no. 7, pp. 1786–1791, 2013.
[3] A. Mohammadi and A. Asif, “Distributed consensus innovation particle filtering for bearing/range tracking with communication constraints,” IEEE Trans. Signal Process., vol. 63, no. 3, pp. 620–635, 2015.
[4] S. Zhu, C. Chen, X. Ma, B. Yang, and X. Guan, “Consensus based estimation over relay assisted sensor networks for
situation monitoring,” IEEE J. Sel. Top. Signal Process., vol. 9, no. 2, pp. 278–291, 2015.
[5] G. Battistelli, L. Chisci, and D. Selvi, “Distributed averaging of exponential-class densities with discrete-time eventtriggered consensus,” IEEE Trans. Control Network Syst., 2016.
[6] A. Amini, A. Mohammadi, and A. Asif, “Event-based consensus for a class of heterogeneous multi-agent systems: An lmi
approach,” in ICASSP, 2017.
[7] G. Xie, H. Liu, L. Wang, and Y. Jia, “Consensus in networked multi-agent systems via sampled control: fixed topology
case,” in American Control Conference, pp. 3902–3907, IEEE, 2009.
[8] W. Ren and Y. Cao, “Convergence of sampled-data consensus algorithms for double-integrator dynamics,” in CDC 2008.
47th IEEE Conference on, pp. 3965–3970, IEEE, 2008.
[9] P. Tabuada, “Event-triggered real-time scheduling of stabilizing control tasks,” IEEE Trans. Autom. Control, vol. 52, no. 9,
pp. 1680–1685, 2007.
19
[10] D. V. Dimarogonas and K. H. Johansson, “Event-triggered control for multi-agent systems,” in CDC/CCC 2009.
Proceedings of the 48th IEEE Conference on, pp. 7131–7136, IEEE, 2009.
[11] M. Cao, F. Xiao, and L. Wang, “Event-based second-order consensus control for multi-agent systems via synchronous
periodic event detection,” IEEE Trans. Autom. Control, vol. 60, no. 9, pp. 2452–2457, 2015.
[12] W. Hu, L. Liu, and G. Feng, “Output consensus of heterogeneous linear multi-agent systems by distributed eventtriggered/self-triggered strategy,” IEEE Trans. Cybern., vol. PP, no. 99, pp. 1–11, 2016.
[13] H. Li, X. Liao, T. Huang, and W. Zhu, “Event-triggering sampling based leader-following consensus in second-order multiagent systems,” IEEE Trans. Autom. Control, vol. 60, no. 7, pp. 1998–2003, 2015.
[14] M. Cao, F. Xiao, and L. Wang, “Second-order leader-following consensus based on time and event hybrid-driven control,”
Syst. Control Lett., vol. 74, pp. 90–97, 2014.
[15] N. Mu, X. Liao, and T. Huang, “Event-based consensus control for a linear directed multiagent system with time delay,”
IEEE Trans. Circuits Syst. II Express Briefs, vol. 62, no. 3, pp. 281–285, 2015.
[16] D. Yang, W. Ren, X. Liu, and W. Chen, “Decentralized event-triggered consensus for linear multi-agent systems under
general directed graphs,” Automatica, vol. 69, pp. 242–249, 2016.
[17] R. T. Marler and J. S. Arora, “Survey of multi-objective optimization methods for engineering,” Struct. Multidiscip. Optim.,
vol. 26, no. 6, pp. 369–395, 2004.
[18] X. Zhou, P. Shi, C.-C. Lim, C. Yang, and W. Gui, “Event based guaranteed cost consensus for distributed multi-agent
systems,” J. Franklin Inst., vol. 352, no. 9, pp. 3546–3563, 2015.
[19] J. Mei, W. Ren, and J. Chen, “Consensus of second-order heterogeneous multi-agent systems under a directed graph,” in
2014 ACC, pp. 802–807, IEEE, 2014.
[20] C. Scherer, P. Gahinet, and M. Chilali, “Multiobjective output-feedback control via lmi optimization,” IEEE Trans. Autom.
Control, vol. 42, no. 7, pp. 896–911, 1997.
[21] C. Wang, Z. Zuo, Z. Lin, and Z. Ding, “A truncated prediction approach to consensus control of lipschitz nonlinear multiagent systems with input delay,” IEEE Trans. Control Netw. Syst.
[22] H. Li, X. Jing, and H. R. Karimi, “Output-feedback-based control for vehicle suspension systems with control delay,” IEEE
Trans. Ind. Electron., vol. 61, no. 1, pp. 436–446, 2014.
[23] P. Badri, A. Amini, and M. Sojoodi, “Robust fixed-order dynamic output feedback controller design for nonlinear uncertain
suspension system,” Mech. Syst. Sig. Process., vol. 80, pp. 137–151, 2016.
[24] L. Jiang, W. Yao, Q. Wu, J. Wen, and S. Cheng, “Delay-dependent stability for load frequency control with constant and
time-varying delays,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 932–941, 2012.
[25] C. Wang, Z. Zuo, Z. Lin, and Z. Ding, “A truncated prediction approach to consensus control of lipschitz nonlinear multiagent systems with input delay,” IEEE Trans. Control Network Syst., 2016.
[26] H. Zhao and J. H. Park, “Dynamic output feedback consensus of continuous-time networked multiagent systems,”
Complexity, vol. 20, no. 5, pp. 35–42, 2015.
[27] X. Zhou, P. Shi, C.-C. Lim, C. Yang, and W. Gui, “Event based guaranteed cost consensus for distributed multi-agent
systems,” J. Franklin Inst., vol. 352, no. 9, pp. 3546–3563, 2015.
[28] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix inequalities in system and control theory. SIAM,
1994.
[29] Z. Li, G. Wen, Z. Duan, and W. Ren, “Designing fully distributed consensus protocols for linear multi-agent systems with
directed graphs,” IEEE Trans. Autom. Control, vol. 60, no. 4, pp. 1152–1157, 2015.
[30] W. Hu, L. Liu, and G. Feng, “Consensus of linear multi-agent systems by distributed event-triggered strategy,” IEEE
Trans. Cybern., vol. 46, no. 1, pp. 148–157, 2016.
[31] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings
of the IEEE, vol. 95, no. 1, pp. 215–233, 2007.
[32] M. S. Mahmoud, Resilient control of uncertain dynamical systems, vol. 303. Springer Science & Business Media, 2004.
[33] A. D. Ames, K. Galloway, K. Sreenath, and J. W. Grizzle, “Rapidly exponentially stabilizing control lyapunov functions
and hybrid zero dynamics,” IEEE Trans. Autom. Control, vol. 59, no. 4, pp. 876–891, 2014.
[34] Y. Liu and Y. Jia, “h∞ consensus control of multi-agent systems with switching topology: a dynamic output feedback
protocol,” Int. J. Control, vol. 83, no. 3, pp. 527–537, 2010.
20
[35] J. Y. Ishihara and M. H. Terra, “On the lyapunov theorem for singular systems,” IEEE Trans. Autom. Control, vol. 47,
no. 11, pp. 1926–1930, 2002.
[36] V. Phat, Y. Khongtham, and K. Ratchagit, “Lmi approach to exponential stability of linear systems with interval timevarying delays,” Linear Algebra Appl., vol. 436, no. 1, pp. 243–251, 2012.
[37] D. Šiljak and D. Stipanovic, “Robust stabilization of nonlinear systems: the lmi approach,” Mathematical problems in
Engineering, vol. 6, no. 5, pp. 461–493, 2000.
[38] V. S. Dolk, P. Tesi, C. De Persis, and W. Heemels, “Event-triggered control systems under denial-of-service attacks,” IEEE
Trans. Control Network Syst., 2016.
[39] J. Mei, W. Ren, and J. Chen, “Distributed consensus of second-order multi-agent systems with heterogeneous unknown
inertias and control gains under a directed graph,” IEEE Trans. Autom. Control, vol. 61, no. 8, pp. 2019 – 2034, 2016.
[40] J. Lofberg, “Yalmip: A toolbox for modeling and optimization in matlab,” in Computer Aided Control Systems Design,
2004 IEEE International Symposium on, pp. 284–289, IEEE, 2005.
[41] L. Yu and J. Chu, “An lmi approach to guaranteed cost control of linear uncertain time-delay systems,” Automatica, vol. 35,
no. 6, pp. 1155–1159, 1999.
| 3cs.SY
|
Fast Computation of Isochrones in Road
Networks
Moritz Baum1 , Valentin Buchhold1 , Julian Dibbelt1 and Dorothea
Wagner1
1
Karlsruhe Institute of Technology, Germany, [email protected]
November 12, 2015
Abstract
We study the problem of computing isochrones in road networks, where the objective
is to identify the region that is reachable from a given source within a certain amount
of time. While there is a wide range of practical applications for this problem (e. g.,
reachability analyses, geomarketing, visualizing the cruising range of a vehicle), there
has been little research on fast computation of isochrones on large, realistic inputs.
In this work, we formalize the notion of isochrones in road networks and present a
basic approach for the resulting problem based on Dijkstra’s algorithm. Moreover, we
consider several speedup techniques that are based on previous approaches for one-tomany shortest path computation (or similar scenarios). In contrast to such related
problems, the set of targets is not part of the input when computing isochrones. We
extend known Multilevel Dijkstra techniques (such as CRP) to the isochrone scenario,
adapting a previous technique called isoGRASP to our problem setting (thereby, enabling
faster queries). Moreover, we introduce a family of algorithms based on (single-level)
graph partitions, following different strategies to exploit the efficient access patterns of
PHAST, a well-known approach towards one-to-all queries. Our experimental study
reveals that all speedup techniques allow fast isochrone computation on input graphs
at continental scale, while providing different tradeoffs between preprocessing effort,
space consumption, and query performance. Finally, we demonstrate that all techniques
scale well when run in parallel, decreasing query times to a few milliseconds (orders of
magnitude faster than the basic approach) and enabling even interactive applications.
1. Introduction
Web-based map services, autonomous navigation systems, and other location-based applications have gained wide currency in the last decades, motivating a great deal of research
1
on practical algorithms for routing in road networks [2]. Most work focused on computing distances (or shortest paths) between pairs (or sets) of vertices. On the other hand,
isochrones (more generally, isolines) are defined as curves of constant distance (from a given
source). Thereby, isochrones represent the area that is in range for a given time limit (or
some other limited resource). Hence, rather than actual distances, information about the
subgraph within range is required. Isolines are relevant in a wide range of applications,
such as reachability analyses [3, 17, 18], geomarketing [13], range visualization for (electric)
vehicles [20], and a variety of online applications [22].
Related Work. Dijkstra’s well-known algorithm [12] computes shortest paths from a given
source to all (reachable) vertices in (almost) linear time. The MINE algorithm [17] is a
Dijkstra-based search to compute isochrones in transportation networks, using a spatial
network database. An improved variant, MIMEX [18], has reduced space requirements.
Both approaches work on databases, prohibiting interactive applications (with running
times beyond the order of minutes).
Regarding shortest-path computation, on the other hand, various speedup techniques [2]
were developed for faster queries, using an offline preprocessing phase (sometimes including
an additional customization step to incorporate, e. g., user preferences or traffic updates),
and an online query phase. Customizable Route Planning (CRP) [5] uses overlays based
on a (multilevel) graph partition to speed up queries [8, 21, 23]. Metric customization
is practical, as only shortcut costs in the overlay need to be updated. In Contraction
Hierarchies (CH) [19], vertices are contracted in increasing order of importance, creating
shortcuts between neighbors to maintain distances. A customizable variant of CH (CCH)
was introduced in [11].
Both CH and CRP were extended to related scenarios, such as one-to-many or one-to-all
queries. Delling et al. introduce PHAST [4], exploiting CH for fast one-to-all and all-pairs
shortest-path computations. Queries consist of a CH search and a linear sweep over the
vertices to obtain distances. For one-to-many scenarios, RPHAST [7] uses a target selection
phase to speed up the second phase of PHAST queries. Efentakis et al. [14] propose GRASP,
a one-to-all (and one-to-many) technique that builds upon CRP, enabling customization
in a few seconds. Many-to-many queries can be handled by bucket-based approaches [24],
storing pairs of target and distance at vertices. This generic approach can be implemented
by several speedup techniques [7]. Other related query types include point-of-interest (POI)
queries, best-via queries (shortest paths that visit certain types of POI), and k-nearest
neighbors (kNN) queries. Bucket-based approaches can be extended to such scenarios [1].
Delling et al. [9] support POI and kNN queries based on CRP. Efentakis et al. [15] extend
GRASP to cover kNN queries. All techniques exploit the fact that targets (or POIs) are
known in advance. To the best of our knowledge, the only speedup technique extended to
2
isochrone queries is GRASP [14].1 However, GRASP computes distances to all vertices in
range, which can be wasteful if only the actual isoline is required.
Our Contribution. This work studies speedup techniques for isochrone computation. Since
no canonical definition exists in the literature, we formally define the isochrone problem
(specifying the output of our algorithms) and present a variant of Dijkstra’s algorithm
to solve it. For faster queries, we propose a new variant of the CRP speedup technique
that computes isochrones. As an alternative, we also extend the existing approach of
isoGRASP [14] to our definition. Next, we introduce a family of novel approaches that
exploit different strategies to combine graph partitions and (R)PHAST. We also show how
all approaches can be parallelized for further speedup. Our experimental evaluation on
realistic, large-scale input reveals that all techniques are orders of magnitude faster than
the basic approach. Providing queries in the order of milliseconds, they enable a range
of new (e. g., interactive) applications. Since all proposed techniques differ in terms of
customization effort, memory consumption, and query time, each is suitable for certain
applications.
Outline. Section 2 introduces basic terminology and building blocks. Section 3 formally
defines the problem and provides a basic approach. Section 4 presents techniques based on
CRP and GRASP, while Section 5 introduces algorithms that combine graph partitions
and PHAST. Section 6 experimentally evaluates all approaches. Section 7 concludes with
final remarks.
2. Preliminaries
We consider road networks given as directed graph G = (V, E) with length function len : E →
R≥0 (representing, e. g., travel time). An s–t-path in G is a sequence Ps,t = [s =
v1 , v2 . . . , vk = t] of vertices, such that (vi , vi+1 ) ∈ E for 1 ≤ i ≤ k − 1. The length
of Ps,t is the sum of its edge lengths. We presume that graphs are strongly connected, i. e.,
there exists an s–t-path for each pair s, t ∈ V . The distance d(s, t) from s to t is the length
of the shortest s–t-path in G.
A (vertex) partition is a family V = {V1 , . . . , Vk } of cells Vi ⊆ V , such that Vi ∩ Vj = ∅
S
for i 6= j and ki=1 Vi = V . A (nested) multilevel partition with L levels is a family Π =
{V 1 , . . . , V L } of partitions of nested cells, i. e., for each level ` ≤ L and cell Vi` ∈ V ` , there
is a cell Vj`+1 ∈ V `+1 at level ` + 1 with Vi` ⊆ Vj`+1 . For consistency, we define V 0 := {{v} |
v ∈ V } (the trivial partition where each vertex has its own cell) and V L+1 := {V } (the
trivial single-cell partition). An edge (u, v) ∈ E is a boundary edge (u and v are boundary
vertices) on level `, if u and v are in different cells of V ` . Similar to vertex partitions, we
1
An extension of the CRP approach in [9] to isochrones is outlined in a patent description (US Patent
App. 13/649,114; http://www.google.com/patents/US20140107921), however, in a simpler than our
intended scenario. Furthermore, the approach was neither implemented nor evaluated.
3
define edge partitions E = {E1 , . . . , Ek }, with Ei ∩ Ej = ∅ for i 6= j and ki=1 Ei = E. A
vertex v ∈ V is distinct (wrt. E) if all its incident edges belong to the same cell, else v is a
boundary vertex or ambiguous.
Dijkstra’s algorithm [12] computes, for a given source s, the distances d(s, v) to all v ∈ V .
It maintains distance labels d(·) for each vertex (initially, d(s) = 0 and d(v) = ∞ for v 6= s).
In each iteration, the algorithm extracts a vertex u with minimum d(u) from a priority
queue (initialized with s) and settles it. At this point, d(u) is final, i. e., d(u) = d(s, u). It
then scans all edges (u, v), i. e., if d(u) + len(u, v) < d(v), it updates d(v) accordingly and
adds (or updates) v in the queue.
S
Speedup Techniques. The three-phase workflow of CRP [5] distinguishes preprocessing
and metric customization. First, a (multilevel) partition Π of the road network is computed,
inducing for each level ` of Π an overlay graph H ` that contains all boundary vertices and
boundary edges in V ` and shortcut edges between boundary vertices of each cell Vi` ∈ V ` .
During customization, the lengths of all shortcuts are computed (using, e. g., Dijkstra’s
algorithm). For fast integration of new length functions, previously computed overlays are
used to compute shortcuts of higher levels. For s–t-queries, Dijkstra’s algorithm is run on
the union of the top-level overlay H L and the subgraphs (of H ` for 0 ≤ ` ≤ L − 1, H 0 := G)
induced by the cells containing s or t.
In CH [19], vertices are contracted in a (heuristic) order rank : V → {1, . . . , |V |}. To
contract a vertex v, shortcut edges are added between uncontracted neighbors (if necessary)
to preserve distances. Also, levels `(·) are assigned to vertices, initially set to zero. When
contracting u ∈ V , we set `(v) = max{`(v), `(u) + 1} for each uncontracted neighbor v.
Let E + denote the set of shortcuts added during preprocessing. An s–t-query is bidirectional,
with a forward search (from s) on G↑ = (V, E ↑ ), where E ↑ = {(u, v) ∈ E ∪ E + : rank(u) <
rank(v)}, and a backward search (from t) on G↓ = (V, E ↓ ), where E ↓ = {(u, v) ∈ E ∪
E + : rank(u) > rank(v)}.
Batched Shortest Paths. Both CH and CRP compute distances between pairs of vertices.
GRASP [14] extends CRP to batched query scenarios. In addition to shortcuts between
boundary vertices (within a cell), each level-` boundary vertex (for 0 ≤ ` < L) stores
(incoming) downward shortcuts from boundary vertices of its corresponding supercell at
level ` + 1. Customization works similar to CRP, storing downward edges in a separate
downward graph H ↓ . For one-to-all queries, the upward phase runs a CRP search from the
source, obtaining distances to settled vertices. Then, the scanning phase processes cells in
descending level order, sweeping over downward edges to propagate distance labels from
boundary vertices to those at the level below.
PHAST [4] exploits CH preprocessing for fast one-to-all queries. The upward phase
runs a forward CH search, while the scanning phase processes vertices in descending order
of CH level and propagates distances by scanning incoming edges in E ↓ . Vertices are
reordered during preprocessing, so the scanning phase is a linear sweep over an edge array.
4
RPHAST [7] enables one-to-many queries by introducing an additional target selection
phase. Given a target set T , it extracts a (restricted) subgraph G↓T of the original downward
graph with a BFS in G↓ (from all t ∈ T ). Queries resemble PHAST queries, running the
scanning phase on G↓T instead of G↓ .
3. Problem Statement and Basic Approach
We formalize the notion of isochrones and describe a basic algorithm. Given a graph G =
(V, E) and a length function len on its edges, the isochrone problem takes as input a
source s ∈ V and a time limit τ ∈ R≥0 . We say that a vertex v ∈ V is in range if d(s, v) ≤ τ ,
else it is out of range. There is no canonical definition of the output of the isochrone
problem. Usually, a compact representation of all vertices in range (or edges with at least
one incident vertex in range) is required [14, 18, 25]. Therefore, we define the output of
the isochrone problem as the set of all isochrone edges that separate vertices in range from
those out of range. Observe that these are the edges (u, v) ∈ E with exactly one endpoint
in range, i. e., either d(s, u) ≤ τ, d(s, v) > τ or d(s, u) > τ, d(s, v) ≤ τ . This set of edges
compactly represents the area in range (thus, generating output is unlikely to become a
performance bottleneck). All approaches presented below can be modified to serve other
definitions (requiring, e. g., the set of vertices in range).
Next, we describe a variant of Dijkstra’s algorithm, called isoDijkstra, to compute all
isochrone edges for a given source s and time limit τ . The search runs from s as described
in Section 2, but stops once the stopping criterion is fulfilled, i. e., the distance label of
the minimum element in the queue exceeds τ . To determine isochrone edges, we then
sweep over all vertices v still contained in the queue (note that these must be out of range).
For each, we check all incoming edges (u, v) and output exactly those where u is in range.
Clearly, this way we find all isochrone edges (u, v) where u is in range (Dijkstra’s algorithm
scans exactly the vertices in range, so the out-of-range endpoints of all isochrone edges were
added to the queue, but not extracted). We make the following modification to ensure that
isochrone edges (u, v) where u is out of range are found as well. When settling a vertex u,
we also scan incoming edges (v, u). If d(v) = ∞, we insert v into the queue with a key of
infinity. Thereby, we guarantee that both types of isochrone edges are contained in the
queue when the search terminates.
4. Multilevel Dijkstra Approaches
The basic idea for Multilevel Dijkstra (MLD) approaches is to skip cells that are in range,
but descend into lower levels where necessary to determine isochrone edges. Following
this approach, we have to cope with several challenges. First, it is not sufficient to check
whether the time limit is exceeded at boundary vertices (we may miss isochrone edges
that are part of no shortcut, see Appendix A for an example). Therefore, we compute
additional data during customization. Second, when descending into lower levels, distance
5
labels must be consistent within each cell. This motivates a two-phase approach. We say
that a cell is active, if its induced subgraph contains at least one isochrone edge. The first
phase determines active cells (running a CRP query), while the second phase descends into
lower-level overlays to determine isochrone edges.
Customization. Metric customization works along the lines of plain CRP. We compute
distances of clique shortcuts by running Dijkstra searches (restricted to the respective cells).
Cliques are represented as square matrices in contiguous memory for efficiency. To improve
data locality and simplify index mapping, vertices are reordered such that boundary vertices
are pushed to the front (ordered by descending level), breaking ties by cell [5].
Given a boundary vertex u of a cell Vi` , we define its (level-`) eccentricity, denoted ecc` (u),
as the maximum (finite) distance d(u, v) to some v ∈ Vi` on the subgraph induced by Vi` .
Note that this subgraph is not strongly connected in general, but restricting eccentricites to
cells allows faster customization (see Appendix B for alternative approaches). We compute
eccentricities for all boundary vertices on each level (storing them adds a single column
to each clique matrix). To this end, we make use of the Dijkstra searches that compute
shortcut lengths. Given a boundary vertex v at level 1, we run the search from v until the
queue is empty (instead of aborting it once all boundary vertices were settled) at negligible
overhead. The distance label of the last settled vertex is the eccentricity ecc1 (v) of v. At
higher levels (1 < ` ≤ L), we compute upper bounds on eccentricities as follows. When
settling a vertex v, we check whether the sum of the label d(v) and ecc`−1 (v) exceeds the
current bound and update it if necessary. While upper bounds may lead to cells falsely
being marked active, this does not violate correctness of queries.
isoCRP. Given a source s ∈ V and a limit τ , queries run in two phases. First, the upward
phase phase is similar to a plain CRP query from s, but running isoDijkstra on the search
graph (the union of the top-level overlay and all induced subgraphs of cells containing s).
Thus, the upward phase ends when all elements in the queue have a distance label greater
than τ . Additionally, to determine active cells, we maintain two flags i(·) (initially false)
and o(·) (initially true) for each cell (per level), to indicate whether the the cell contains
at least one vertex that is in or out of range, respectively. When settling a vertex u ∈ Vi` ,
we set i(Vi` ) to true if d(u) ≤ τ . Next, we check whether d(u) + ecc` (u) ≤ τ . Observe
that this condition is not sufficient to unset o(Vi` ), because ecc` (u) was computed on the
subgraph of Vi` . If this subgraph is not strongly connected, d(u) + ecc` (u) is (in general)
not an upper bound on the distance to any vertex in Vi` (there may be shorter paths to
internal vertices of Vi` via boundary vertices that are unreachable from u in the cell-induced
subgraph). Therefore, when scanning an outgoing shortcut (u, v) with length ∞ (this
shortcut exists due to the matrix representation), we also check whether d(v) + ecc` (v) ≤ τ .
If the condition holds for u and all boundary vertices v unreachable from u (wrt. Vi` ), we
can safely unset o(Vi` ). Note that all checks are cache efficient. Toggled flags are final, so
we no longer need to perform any checks for such flags. After the upward phase finished,
6
we mark all cells Vi` that have both i(Vi` ) and o(Vi` ) set as active (since isochrone edges
can only be contained in cells having vertices both in and out of range).
The downward phase then consists of L subphases. In decreasing order of levels, and
for every active cell at the current level `, each subphase runs isoDijkstra restricted to the
respective cell in H`−1 . Initially, all boundary vertices are inserted into the queue using their
distance labels (according to the previous phase) as keys. Again, we check eccentricities
on-the-fly to mark active cells for the next phase. Isochrone edges are determined at the end
of each isoDijkstra search (see Section 3). On overlays, only boundary edges are reported
(in contrast to shortcuts).
isoGRASP. We describe a modified version of isoGRASP [14] that is more suitable for
our scenario (in [14], distances to all vertices in range are computed). Customization of
our variant is similar to isoCRP, computing eccentricities of boundary vertices as described
above. Additionally, we generate downward edges required by GRASP on-the-fly. We apply
edge reduction (removing shortcuts via other boundary vertices) [14] to downward edges,
but stick to the matrix representation for overlay edges. Queries again consist of two phases.
The upward phase is identical to the one of isoCRP. The scanning phase handles levels
from top to bottom in L subphases and processes active cells. For an active cell Vi` at
level `, isoGRASP sweeps over its internal vertices (all vertices in H`−1 that lie in Vi` and
are no boundary vertex of Vi` ). For each internal vertex v, its incoming downward edges
are scanned, obtaining the distance at v. To determine active cells for the next subphase,
we maintain flags i(·) and o(·) as described for isoCRP. This requires checks at unreachable
boundary vertices (from v within Vi` ). For speedup, we precompute these vertices and store
them in a separate adjacency array.
Similar to isoCRP, the upward phase collects all isochrone edges that are boundary edges
at level L or contained in the cell-induced subgraph of the source. For remaining isochrone
edges, we sweep over internal vertices a second time after processing a cell in the scanning
phase. To avoid duplicates and to ensure that endpoints of scanned edges have correct
distances, we skip edges leading to vertices with higher index (recall that we reorder vertices
during preprocessing). Scanned boundary edges are then added to the output if exactly
one endpoint is reachable.
Parallelization. During customization, cells of each level are processed in parallel [5]. To
avoid concurrent memory access when customizing isoGRASP, we maintain thread-local
containers for downward edges, which are concatenated at the end. Regarding queries,
the (more expensive) downward phase is parallelized in a natural way, as cells at a certain
level can be handled independently. Thus, we assign cells to threads and synchronize them
between subphases. To reduce the risk of false sharing (i. e., concurrent access to the same
cache line), we assign blocks of consecutive cells (wrt. vertex ordering) to the same thread.
Moreover, to reduce synchronization overhead, we process cells on lower levels in a top-down
fashion within the same thread.
7
5. Contraction-Based Approaches
This section introduces isoPHAST, an approach based on (R)PHAST to compute isochrones.
Since the targets are not part of the input, we make use of graph partitions to restrict
the area that is examined for isochrone edges. Queries have three phases, starting with
a forward CH search from the source. Next, active cells are determined. Finally, we use
PHAST sweeps (restricted to active cells) to compute distances to vertices of active cells.
We now describe the (generic) preprocessing of isoPHAST, before we present strategies to
determine active cells.
First, we compute a (single-level) partition V = {V1 , . . . , Vk } of the graph. Then, we use
CH to contract all cell-induced subgraphs, but leave boundary vertices of the partition
uncontracted. Afterwards, we reorder vertices in the graph such that uncontracted core
vertices are pushed to the front, breaking ties by cell (providing the same benefits as in
CRP). We order contracted vertices within cells by their CH levels (as required by PHAST,
see Section 2). As a result of preprocessing, we obtain a (single) upward graph G↑ , containing
(for all cells) edges to higher-ranked vertices, added shortcuts between core vertices, and
all boundary edges. Conversely, the downward graph G↓ stores only downward edges (i. e.,
boundary vertices have no incoming edges). Further steps of preprocessing depend on the
query strategy and are described below.
isoPHAST-CD. Our first strategy (CoreDijkstra) executes isoDijkstra on the core graph
to determine active cells. This requires eccentricities for core vertices, which are efficiently
obtained from preprocessed data. To compute ecc(u) of some vertex u, we run (as a last step
of preprocessing) Dijkstra’s algorithm on the core graph (restricted to the cell Vi containing u), followed by a PHAST sweep over the internal vertices of Vi . Whenever processing a
vertex v (of Vi ), we update the eccentricity by setting ecc(u) = max{ecc(u), d(v)}.
A query starts by running (iso)Dijkstra from the source s in G↑ . Within the cell of s, this
corresponds to an upward CH search, since G↑ stores only upward edges. At core vertices,
we update flags i(·) and o(·) to determine active cells (as described in Section 4, using
an adjacency array to store vertices required for additional checks, similar to isoGRASP).
In the special case that the core is not reached, only the cell of s is set active. Then, we
process every active cell by running PHAST on its internal vertices. Thereby, we obtain
distances to all vertices that are both in range and contained in an active cell (due to
the stopping criterion, we get upper bounds for vertices out of range, which suffices for
isochrone detection).
Regarding output, boundary (isochrone) edges are found by isoDijkstra. Isochrone
edges connecting internal vertices are obtained in linear sweeps. When settling a vertex v,
higher-ranked neighbors have final distance labels. The label of v is final after scanning
incoming edges (u, v) ∈ G↓ . Thus, looping through incoming edges a second time suffices
to find isochrone edges after a few modifications to G↓ . First, as only original edges are
relevant candidates, we indicate this with an additional flag (per edge). Second, (directed)
edges (v, u) ∈ E to vertices u of higher rank are (in general) not contained in G↓ . To ensure
8
that adjacent vertices (wrt. G) are also adjacent in G↓ , we add dummy edges (with a length
of infinity) to G↓ .
isoPHAST-CP. Instead of isoDijktra, this strategy (CorePHAST) runs PHAST within
the core. We compute eccentricities after the (generic) preprocessing (as described above).
Next, CH preprocessing is run on the core, too. We reorder core vertices according to
CH levels. Finally, we update G↑ and G↓ (obtained after cell contraction), adding core
shortcuts and removing the obsolete boundary edges of the partition (that were needed to
compute eccentricites).
Queries strictly follow the three-phase pattern discussed above. First, we perform a
forward CH search in G↑ , until the priority queue is empty. Second, we determine active
cells and compute correct distance labels for all core vertices. To achieve this, we run a
PHAST sweep over the core vertices (that were sorted accordingly), relaxing edges in G↓ to
propagate distance values to vertices of lower rank. Again, we maintain flags, i(·) and o(·),
and an adjacency array for vertex checks to determine active cells. To find isochrone edges
between boundary vertices on-the-fly, we add dummy edges and edge flags inside the core
(similar to the third phase of isoPHAST-CD). The third phase (sweeps over active cells) is
identical to isoPHAST-CD.
isoPHAST-DT. Our last strategy (DistanceTable) uses a distance (bounds) table to speed
up the second phase (determining active cells). Working with such tables (instead of a
dedicated core search) benefits from edge partitions, since the unique assignment of edges
to cells simplifies isochrone edge retrieval (we avoid checking for boundary edges in the
second phase). Given a partition E = {E1 , . . . , Ek } of the edges of G, the table stores for
each pair Ei , Ej of cells a lower bound d(Ei , Ej ) and an upper bound d(Ei , Ej ) on the
distance from Ei to Ej , i. e., d(Ei , Ej ) ≤ d(u, v) ≤ d(Ei , Ej ) for all u ∈ Ei , v ∈ Ej (we
abuse notation, saying v ∈ Ei if v is an endpoint of at least one edge e ∈ Ei ). Given a
source s ∈ Ei (if s is ambiguous, pick any cell containing s) and a limit τ , cells Ej with
d(Ei , Ej ) ≤ τ < d(Ei , Ej ) are set active.
To (quickly) compute (not necessarily tight) distance bounds, we make use of the data
computed during generic preprocessing (that is now based on an edge partition). On the
resulting core graph we run, for each cell Ei , a (multi-source) forward CH search from
all boundary vertices in Ei . Next, we perform a PHAST sweep on G↓ (i. e., the complete
downward graph), keeping track of the minimum and maximum distance label per target
cell. This yields, for all cells, lower bounds d(Ei , ·), and upper bounds on the distance from
boundary vertices of Ei to each cell. To obtain the desired bounds d(Vi , ·), we increase
these values by the (backward) boundary diameter of Ei , i. e., the maximum distance from
any vertex in Ei to a boundary vertex in Ei . To compute the diameter of a cell Ei , we
run, for each boundary vertex v ∈ Ei , Dijkstra’s algorithm on the reverse core graph (i. e.,
inverting edge directions), until all boundary vertices of Ei were scanned (this can be easily
verified using a counter). Using a PHAST sweep over the downward graph of Ei , we obtain
9
the unlimited (i. e., not limited to Ei ) backward eccentricity of v. The diameter of Vi
is the maximum eccentricity of its boundary vertices. Next, we perform the actual CH
preprocessing on G, using a purely greedy order (boundary vertices are not artificially kept
on top of the contraction order). This results in a better contraction order and sparser
graphs G↑ and G↓ , reducing preprocessing and query times. Finally, we extract (and store)
the relevant search graph G↓i for each Ei ∈ E. This requires an RPHAST selection phase
per cell, using all (ambiguous and distinct) vertices of a cell as input.
Queries start with a forward CH search on G↑ (applying no stopping criterion). Next,
active cells are determined in a linear sweep over one row (corresponding to the source cell)
of the distance table. The third phase runs an RPHAST sweep for each active cell, obtaining
distances to all its vertices [7]. Although vertices can be contained in several search graphs,
we do not reinitialize distance labels between sweeps (the source remains unchanged). To
obtain isochrone edges, we proceed as before, looping through incoming downward edges
twice (again, we add dummy edges to G↓i for correctness). To avoid duplicates in the output
(due to vertices contained in several search graphs), edges in G↓i get an additional flag to
indicate whether the edge belongs to the corresponding cell Ei . Isochrone edges are added
to the output only if this flag is set.
As mentioned above, search graphs may share some vertices (e. g., the vertex with
maximum CH rank is contained in every search graph). Vertices contained in many graphs
increase memory consumption and slow down queries. Therefore, we use search graph
compression, i. e., we store the topmost vertices of the hierarchy (and their incoming edges)
in a separate graph G↓c and remove them from all search graphs G↓i . During queries, we
first perform a linear sweep over G↓c (obtaining distances for all v ∈ G↓c ), before processing
search graphs of active cells.
Parallelization. The first steps of preprocessing are executed in parallel, namely, building
cell graphs, contracting internal vertices, inserting dummy edges, and reordering internal
vertices by level. Afterwards, threads are synchronized. The upward and downward graphs
are computed sequentially. Finally, computation of eccentricities (or cell diameters) is
parallelized again. Since threads operate on distinct vertex sets, there are no concurrent
accesses to distance labels, and we can share labels among threads. However, each thread
needs its own priority queue. All remaining steps of preprocessing (depending on the
strategy) are executed sequentially.
As for queries, the first two phases are run sequentially. Both the CH search and
isoDijkstra are difficult to parallelize. Running PHAST (on the core) in parallel does not
pay off (the core is rather dense, resulting in many CH levels). Distance table operations,
on the other hand, are very fast, so parallelization is not necessary. In the third phase,
however, active cells can be assigned to different threads. Again, we share distance labels
among threads, except for isoPHAST-DT (search spaces may overlap, so each thread uses
own labels to avoid concurrent accesses). As a result, in isoPHAST-DT each thread runs a
10
Table 1. Performance of basic one-to-all and one-to-many building blocks. Execution
times are sequential. One-to-many algorithms compute distances to all targets in a ball of
size |B| = 214 .
Algorithm
[our]
Dijkstra (1-to-all)
PHAST
GRASP
[7, 15]
2 653.18
–
144.16 136.92
171.11 169.00
Dijkstra (1-to-many)
RPHAST (selection)
RPHAST (query)
7.34
1.29
0.16
7.43
1.80
0.17
forward CH search to initialize its labels. We store a copy of the graph G↓ once per NUMA
node for faster access during queries.
Running the third phase in parallel can make the second phase of isoPHAST-CP a
bottleneck. Therefore, we alter the way of computing i(·) and o(·). When settling a
vertex v ∈ Vi , we set o(Vi ) if d(v) + ecc(v) > τ , and i(Vi ) if d(v) ≤ τ . These checks are
less accurate (marking more cells), but we no longer have to check unreachable boundary
vertices (wrt. Vi ). Observe that correctness of isoPHAST-CP is maintained (no stopping
criterion is applied and maxv∈Vi (d(v) + ecc(v)) is a valid upper bound on the distance to
each vertex in Vi , so no active cells are missed).
6. Experiments
Our code is written in C++ (using OpenMP for parallelization), compiled with g++ 4.8
(-O3). The input for all experiments is the road network of Western Europe made available
for the 9th DIMACS Implementation Challenge (|V | = 18 010 173, |E| = 42 188 664), with
travel time (in seconds) as edge lengths [10]. For isoCRP and isoGRASP, we use a 4-level
partition obtained from PUNCH [6], with maximum cell sizes of 28 , 212 , 216 , 220 , respectively.
For (single-level) vertex partitions required by isoPHAST, we used BUFFOON [27]. Edge
partitions for isoPHAST-DT were retrieved following the approach described in [26, 28].
Our CH implementation follows [19], but adopting tuning parameters (e. g. for priority
terms and staged hop limits) from [4]. Our GRASP implementation includes implicit
initialization [14] and (downward) edge reduction [16]. As opposed to [14], we resort to
a 4-level partition, with only minor effects on running times in preliminary experiments
(similar observations were made in [15]). For comparability, Table 1 shows running times of
our implementations of the basic batched query algorithms on a single core of a 4-core Intel
Xeon E5-1630v3 clocked at 3.7 GHz, with 128 GiB of DDR4-2133 RAM, 10 MiB of L3 and
256 KiB of L2 cache (chosen because it most closely resembles the machines used in [7, 15]).
We report average running time over 1 000 random one-to-all queries for Dijkstra, PHAST,
11
Table 2. Performance of sequential queries. We report customization time (excluding
metric-independent partitioning) and space requirements (space per additional metric given
in brackets, if it differs). The table shows the average number of settled vertices (# Settled,
in thousands) and running times of queries, for time limits τ = 100 and τ = 500. Best
values (except Dijkstra wrt. space) are highlighted in bold.
Custom
Queries (τ = 100 min) Queries (τ = 500 min)
Algorithm
Time [s]
Space [MiB]
# Settled
RangeDijkstra
isoCRP
isoGRASP
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
–
646
1.70
900 (138)
2.50 1 856 (1 094)
26.11
785
1 221.84 781
1 079.11 2 935
460 k
101 k
120 k
440 k
626 k
597 k
Time [ms] # Settled
68.32
15.44
10.06
6.09
15.02
9.96
7 041 k
354 k
387 k
1 501 k
2 029 k
1 793 k
Time [ms]
1 184.06
60.67
37.77
31.63
31.00
24.80
and GRASP. For one-to-many techniques (Dijkstra and RPHAST), we follow a procedure
of [7], picking a center vertex c at random and running Dijkstra’s algorithm from c until |T |
vertices were scanned, making them the target set T . Distances from a random source s ∈ T
to all vertices in T are computed. For RPHAST, we report both target selection (sel) and
query (qry) time. We also report numbers (as-is) from [7, 15] for comparison. Even when
accounting for hardware differences, we see that query times of our implementations are
similar to original publications (target selection of RPHAST is even slightly faster).
Regarding isoPHAST, we performed preliminary studies to obtain reasonable parameters
for partition sizes (k) and search graph compression (|G↓c |, denoting the number of vertices
in the search graph for isoPHAST-DT). For sequential CD (CP) queries, a value of k = 212
(211 ) yields best query times. For fewer cells (i. e., coarser partitions), the third phase
scans a large portion of the graph and becomes the bottleneck. Using a more fine-grained
partition yields a larger core graph, slowing down the second phase. Consequently, a smaller
value for k (256) becomes favorable when queries are executed in parallel (as the third
phase becomes faster). For DT, similar effects occur for different values of k. Additionally,
subgraph compression has a major effect on running times (and space consumption). For
small sizes of G↓c , high-ranked vertices occur in multiple cells, but a larger graph G↓c leads to
unnecessary vertex scans during queries. Setting k = 214 (212 ) and |G↓c | = 216 (213 ) resulted
in fastest sequential (parallel) queries for isoPHAST-DT.
Table 2 shows performance of all algorithms discussed in Sections 4 and 5, reporting
figures on customization and sequential queries. Experiments were conducted on a dual
8-core Intel Xeon E5-2670 clocked at 2.6 GHz, with 64 GiB of DDR3-1600 RAM, 20 MiB
of L3 and 256 KiB of L2 cache. Customization is executed in parallel (using all 16 cores).
Timings for PHAST include all steps of preprocessing except computing the partition (since
it is independent of the metric). For queries, we report average running times of 1 000
random queries for mid-range (τ = 100) and long-range time limits (τ = 500, this was the
12
Table 3. Performance of parallel queries. We report the same figures as in Table 2, but
queries use 16 threads. Best values (per column) are highlighted in bold.
Custom
Queries (τ = 100 min) Queries (τ = 500 min)
Algorithm
Time [s]
Space [MiB]
# Settled
isoCRP
isoGRASP
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
1.70
900 (138)
2.50 1 856 (1 094)
38.07
769
1 432.39 766
865.50 1 066
100 k
120 k
918 k
944 k
914 k
Time [ms] # Settled
2.73
2.35
1.61
4.47
1.74
354 k
387 k
4 578 k
5 460 k
2 979 k
time [ms]
7.86
5.93
8.22
7.86
3.80
hardest limit for most approaches, as it maximized the number of active cells). As expected,
techniques based on MLD provide better customization times, while isoPHAST achieves
the lowest query times (CD for mid-range and DT for long-range queries, respectively).
Customization of isoCRP and isoGRASP is very practical (below three seconds). For
isoPHAST-CD, the lightweight preprocessing pays of as well, allowing customization in less
than 30 seconds. The comparatively expensive preprocessing times of isoPHAST-CP and
DT are dominated by expensive core contraction. Still, (metric-dependent) preprocessing
is far below half an hour, which is suitable for applications that do not require realtime
metric updates. Compared to isoCRP, isoGRASP requires almost an order of magnitude
of additional space for the downward graph (having about 110 million edges). Even when
executed sequentially, all approaches take well below 100 ms, which is significantly faster
than isoDijkstra. The number of settled vertices is significantly larger for isoPHAST (but
data access is more cache friendly). Query times of isoPHAST are below MLD techniques for
both limits, with the exception of isoPHAST-CP for small ranges (because the whole core
graph is scanned). Again, the performance of isoPHAST-CD is quite noteable, providing the
fastest queries for (reasonable) mid-range limits and decent query times for the long-range
limit. Finally, query times of isoPHAST-DT show best scaling behavior, with lowest times
for harder queries.
Table 3 shows parallel times for the same set of random queries, using the same machine.
Note that preprocessing time of isoPHAST changes due to different parameter choices.
Most approaches scale very well with the number of threads, providing a speedup of
(roughly) 8 using 16 threads. Note that factors (according to tables) are much lower for
isoPHAST, since we use tailored partitions for single-thread queries. In fact, isoPHAST-DT
scales best when run on the same preprocessed data (speedup of 11), since its sequential
workflow (upward CH search, table lookups) is very fast. Considering MLD techniques,
isoGRASP scales worse than isoCRP (speedup of 6.5 compared to 7.7), probably because it
is memory bandwidth bounded (while isoCRP comes with more computational overhead).
Consequently, isoGRASP benefits greatly from storing a copy of the downward graph on
each NUMA node. As one may expect, speedups are slightly lower for mid-range queries
(τ = 100). The isoPHAST approaches yield best query times, below 2 ms for mid-range
13
70
isoDijkstra
isoCRP
isoGRASP
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
60
Query Time [ms]
50
40
30
20
10
0
10
30
100
300
1000
3000
Limit [min]
Figure 1. Sequential query times for varying time limits, ranging from 10 to (roughly)
4 700 minutes (the diameter of our input graph). Each point in the plot is the average of
1 000 random queries.
queries, and below 4 ms for the long-range limit. In summary, all algorithms enable queries
that are fast enough for practical applications, achieving speedups by more than two orders
of magnitude compared to isoDijkstra.
Figure 1 shows how (sequential) query times scale for varying time limits. Timings of all
algorithms (except isoDijkstra) follow a characteristic curve. Query times first increase with
the limit τ (the isoline is extended, cutting more active cells), before dropping again once τ
exceeds 500 (the isochrone reaches the boundary of the network, decreasing the number
of active cells). For values τ > 4 710 (minutes), all vertices are in range, making queries
very fast (as there are no active cells). For small values of τ , the MLD techniques and
isoPHAST-CD are fastest. Conversely, isoPHAST-CP is slowed down by the linear sweep
over the core graph (taking about 6 ms, independent of the value τ ), while isoPHAST-DT
suffers from distance bounds in the table that are not tight. On the other hand, the
second phase of isoPHAST-CD has no stopping criterion, making it the slowest variant for
large values of τ (while the other isoPHAST variants benefit from good scaling behavior).
Regarding MLD techniques, isoGRASP is up to almost twice as fast as isoCRP, providing a
decent trade-off between customization effort and query times. In the parallel query setting
(not reported in the figure), query times follow the same characteristic curves. The linear
sweep in the second phase of isoPHAST-CP becomes slightly faster, since the core graph
14
is smaller (due to a different partition). Also ,the performance gap between isoCRP and
isoGRASP is slightly smaller when using multiple threads.
7. Final Remarks
We proposed a formal definition of the isochrone problem, asking for a compact representation
of the region in a road network that is in range (given some resource limit). We defined
output that is reasonable, e. g., for actual visualization of isochrones. The approach of
[25] (used as state-of-the-art in recent work [13]) can be modified easily to take isochrone
edges as input. Given a (planar) graph representation of the network, this can be achieved
by traversing the faces of the graph along the border of the region in range, switching
to the corresponding twin face at isochrone edges. Preliminary experiments showed that
extracting the isochrone region (according to the definition in [25]) this way takes only tens
of milliseconds. While the desired output may differ for other applications, our algorithms
can be readily extended to most reasonable formats.
We introduced a portfolio of speedup techniques for the isochrone problem. While no
single approach is best in all criteria (preprocessing effort, space, query time, simplicity), the
right choice depends on the application and its requirements. If user-dependent metrics are
needed, the fast and lightweight customization isoCRP is helpful. Fast query times subject
to frequent updates (e. g., due to live traffic) are enabled by isoGRASP. If customization
time below a minute is acceptable and ranges are low, isoPHAST-CD is a candidate that
may achieve even faster query times. The remaining variants of isoPHAST show best
scaling behavior, making them suitable for long-range isochrones on large graphs, or if
customizability is not required.
Current and future work includes algorithms for computing actual isolines that are fast
enough for interactive applications (exploiting the techniques presented here). Moreover, we
are interested in visualizing the cruising range of electric vehicles, taking both travel time
and energy consumption into account. Finally, we investigate the possibility of customizable
isoPHAST (exploiting, e. g., CCH [11]), and variants that do not (explicitly) require a
partition of the graph.
15
References
[1] Ittai Abraham, Daniel Delling, Amos Fiat, Andrew V. Goldberg, and Renato F.
Werneck. HLDB: Location-based services in databases. In Proceedings of the 20th
ACM SIGSPATIAL International Symposium on Advances in Geographic Information
Systems (GIS’12), pages 339–348. ACM Press, 2012. Best Paper Award.
[2] Hannah Bast, Daniel Delling, Andrew V. Goldberg, Matthias Müller–Hannemann,
Thomas Pajor, Peter Sanders, Dorothea Wagner, and Renato F. Werneck. Route
planning in transportation networks. Technical Report abs/1504.05140, ArXiv e-prints,
2015.
[3] Veronika Bauer, Johann Gamper, Roberto Loperfido, Sylvia Profanter, Stefan Putzer,
and Igor Timko. Computing isochrones in multi-modal, schedule-based transport
networks. In Proceedings of the 16th ACM SIGSPATIAL international conference on
Advances in geographic information systems (GIS ’08), GIS ’08, pages 78:1–78:2. ACM
Press, 2008.
[4] Daniel Delling, Andrew V. Goldberg, Andreas Nowatzyk, and Renato F. Werneck.
PHAST: Hardware-accelerated shortest path trees. Journal of Parallel and Distributed
Computing, 73(7):940–952, 2013.
[5] Daniel Delling, Andrew V. Goldberg, Thomas Pajor, and Renato F. Werneck. Customizable route planning in road networks. Transportation Science, 2015.
[6] Daniel Delling, Andrew V. Goldberg, Ilya Razenshteyn, and Renato F. Werneck. Graph
partitioning with natural cuts. In 25th International Parallel and Distributed Processing
Symposium (IPDPS’11), pages 1135–1146. IEEE Computer Society, 2011.
[7] Daniel Delling, Andrew V. Goldberg, and Renato F. Werneck. Faster batched shortest
paths in road networks. In Proceedings of the 11th Workshop on Algorithmic Approaches
for Transportation Modeling, Optimization, and Systems (ATMOS’11), volume 20 of
OpenAccess Series in Informatics (OASIcs), pages 52–63, 2011.
[8] Daniel Delling, Martin Holzer, Kirill Müller, Frank Schulz, and Dorothea Wagner.
High-performance multi-level routing. In The Shortest Path Problem: Ninth DIMACS
Implementation Challenge, volume 74 of DIMACS Book, pages 73–92. American
Mathematical Society, 2009.
[9] Daniel Delling and Renato F. Werneck. Customizable point-of-interest queries in road
networks. IEEE Transactions on Knowledge and Data Engineering, 2014. to appear.
[10] Camil Demetrescu, Andrew V. Goldberg, and David S. Johnson, editors. The Shortest
Path Problem: Ninth DIMACS Implementation Challenge, volume 74 of DIMACS
Book. American Mathematical Society, 2009.
16
[11] Julian Dibbelt, Ben Strasser, and Dorothea Wagner. Customizable contraction hierarchies. In Proceedings of the 13th International Symposium on Experimental Algorithms
(SEA’14), volume 8504 of Lecture Notes in Computer Science, pages 271–282. Springer,
2014.
[12] Edsger W. Dijkstra. A note on two problems in connexion with graphs. Numerische
Mathematik, 1:269–271, 1959.
[13] Alexandros Efentakis, Nikos Grivas, George Lamprianidis, Georg Magenschab, and
Dieter Pfoser. Isochrones, traffic and DEMOgraphics. In Proceedings of the 21st
ACM SIGSPATIAL International Conference on Advances in Geographic Information
Systems, SIGSPATIAL’13, pages 548–551. ACM Press, 2013.
[14] Alexandros Efentakis and Dieter Pfoser. GRASP. Extending graph separators for
the single-source shortest-path problem. In Proceedings of the 22nd Annual European
Symposium on Algorithms (ESA’14), volume 8737 of Lecture Notes in Computer
Science, pages 358–370. Springer, September 2014.
[15] Alexandros Efentakis, Dieter Pfoser, and Yannis Vassiliou. SALT. A unified framework
for all shortest-path query variants on road networks. In Evripidis Bampis, editor,
Experimental Algorithms, volume 9125 of Lecture Notes in Computer Science, pages
298–311. Springer, 2015.
[16] Alexandros Efentakis, Dimitris Theodorakis, and Dieter Pfoser. Crowdsourcing computing resources for shortest-path computation. In Proceedings of the 20th ACM
SIGSPATIAL International Symposium on Advances in Geographic Information Systems (GIS’12), pages 434–437. ACM Press, 2012.
[17] Johann Gamper, Michael Böhlen, Willi Cometti, and Markus Innerebner. Defining
isochrones in multimodal spatial networks. In Proceedings of the 20th ACM International
Conference on Information and Knowledge Management, CIKM ’11, pages 2381–2384.
ACM Press, 2011.
[18] Johann Gamper, Michael Böhlen, and Markus Innerebner. Scalable computation of
isochrones with network expiration. In Proceedings of the 24th International Conference
on Scientific and Statistical Database Management, SSDBM’12, pages 526–543. Springer,
2012.
[19] Robert Geisberger, Peter Sanders, Dominik Schultes, and Christian Vetter. Exact
routing in large road networks using contraction hierarchies. Transportation Science,
46(3):388–404, August 2012.
[20] S. Grubwinkler, T. Brunner, and M. Lienkamp. Range prediction for evs via crowdsourcing. In Vehicle Power and Propulsion Conference (VPPC), 2014 IEEE, pages
1–6, Oct 2014.
17
[21] Martin Holzer, Frank Schulz, and Dorothea Wagner. Engineering multilevel overlay
graphs for shortest-path queries. ACM Journal of Experimental Algorithmics, 13(2.5):1–
26, December 2008.
[22] Markus Innerebner, Michael Böhlen, and Johann Gamper. Isoga: A system for
geographical reachability analysis. In Proceedings of the 12th International Conference
on Web and Wireless Geographical Information Systems, W2GIS’13, pages 180–189.
Springer, 2013.
[23] Sungwon Jung and Sakti Pramanik. An efficient path computation model for hierarchically structured topographical road maps. IEEE Transactions on Knowledge and
Data Engineering, 14(5):1029–1046, September 2002.
[24] Sebastian Knopp, Peter Sanders, Dominik Schultes, Frank Schulz, and Dorothea Wagner. Computing many-to-many shortest paths using highway hierarchies. In Proceedings
of the 9th Workshop on Algorithm Engineering and Experiments (ALENEX’07), pages
36–45. SIAM, 2007.
[25] Sarunas Marciuska and Johann Gamper. Determining objects within isochrones in
spatial network databases. In Proceedings of the 14th East European Conference on
Advances in Databases and Information Systems, ADBIS’10, pages 392–405. Springer,
2010.
[26] Alex Pothen, Horst D. Simon, and Kang-Pu Liou. Partitioning sparse matrices with
eigenvectors of graphs. SIAM Journal on Matrix Analysis and Applications, 11:430–452,
1990.
[27] Peter Sanders and Christian Schulz. Distributed evolutionary graph partitioning. In Proceedings of the 14th Meeting on Algorithm Engineering and Experiments (ALENEX’12),
pages 16–29. SIAM, 2012.
[28] Christian Schulz. High Quality Graph Partitioning. PhD thesis, Karlsruhe Institute of
Technology, July 2013.
18
A. Omitted Figures
Figure 2 shows two examples for which a naïve MLD approach may produce wrong outputs
(see also the challenges described in Section 4). Below, Figure 3 illustrates the work of the
isoPHAST algorithms described in Section 5.
22
1
8
v
21
7
3
6
4
1
s
2
w
3
1
u
1
Figure 2. Two examples for which a naïve approach produces wrong outputs. Gray edges
denote shortcut edges. Left: Imagine a mountain in the middle of the cell. The shortcut
edge from the left to the right boundary vertex corresponds to a fast road through a tunnel
(the green edges) and can be traversed within the time limit. However, this does not imply
that we can reach the summit of the mountain via a mountain road (the red edges). Thus,
without using eccentricites, it may happen that we do not descend into the cell and hence
miss an isochrone edge on the mountain road. Right: Assume a limit τ = 4. When we scan
u, we can traverse the shortcut edge (u, v), since the distance at u plus the shortcut length
equals the limit. Thus, we do not descend into the cell at u, but we relax the outgoing
shortcut edge (u, v). Next, scan v, where the sum of distance at v and the shortcut length
is greater than the τ . Therefore, we descend into the cell and relax the edge (v, w). Since
w has a distance label greater than τ and the queue contains no other vertices, the search
stops. Now, we erroneously report (v, w) as an isochrone edge, since the distance label
at v is below τ , while the distance label of w is greater than τ . The s − w path via u that
makes w reachable within τ is not found. This motivates out two-phase approach (which
ensures that distance labels in cells are correct).
B. Omitted Experiments
This section contains experiments omitted in the main document. Below, we present
alternative variants of isoCRP and isoGRASP, using different strategies to determine active
cells. We also include detailed experiments considering parameter tuning of isoPHAST
variants.
Alternative Multilevel Strategies. Given a boundary vertex u on some cell Vi` at level `,
we define its unrestricted (level-`) eccentricity, denoted ecc0` (u), as the maximum distance
d(u, v) to a vertex v ∈ Vi` (in G) in the same cell (i. e., distances are not restricted to
19
s
dist(s,·) ≤ x dist(s,·) > x
Figure 3. Illustration of the isoPHAST algorithms. Thick black lines denote the isoline,
which is intersected by isochrone edges (thin black lines). Cells that are colored gray denote
active cells. Note that there are cells that are marked active despite containing no isochrone
edges (however, this does not violate correctness). Also note that besides the active cells
along the primary isoline, there may be active cells in the interior of the area bounded by
the primary isoline, for example due to mountains whose summits we cannot reach within
the time limit.
cell-induced subgraphs). We compute, for every boundary vertex v, and for each level `,
the unrestricted level-` eccentricity of v. The procedure described in Section 4 yields
upper bounds on (unrestricted) eccentricities if the corresponding subgraphs are strongly
connected (otherwise, we just set the eccentricity to ∞).
During queries, we may now check for boundary vertices whether the sum of their distance
label and their eccentricity exceeds the time limit. If it does, we mark it as an active cell.
We refer to this approach (which is simpler but less accurate than the approach of Section 4)
as isoCRP (none). Since we compute upper bounds on eccentricities during customization,
this approach may falsely mark cells as active. While this does not violate correctness, it
can significantly slow down the second phase of the query. This holds in particular for
cells that induce subgraphs which are not strongly connected. Recall that eccentricities
are computed on this subgraph, and hence we assign eccentricities of ∞ to some vertices.
In other words, the corresponding cell will always be marked active if at least one of its
boundary vertices is in range. Also note that a significant number of cells induces subgraphs
that are not strongly connected (think of cuts at directed edges representing lanes of a
highway). Therefore, we present an approach to improve upper bounds on eccentricities.
We run customization as described above, but add a second sweep over all boundary vertices.
For each boundary vertex, we run another Dijkstra search, but no longer limit it to the
current cell (note that outside of cells, we can make use of all preprocessed shortcuts, even
on higher levels). This search is aborted once that all vertices of the cell were reached (we
can use a counter to easily check this condition). This yields our second approach, isoCRP
20
Table 4. Performance of (sequential) MLD query variants.
queries
custom
algorithm
isoCRP
isoCRP
isoCRP
isoCRP
time [s]
(none)
(all)
(inf)
(scc)
space [MiB]
limit τ = 100 min
limit τ = 500 min
# scans time [ms]
# scans time [ms]
1.70
8.42
4.91
2.48
900
900
900
900
(138)
(138)
(138)
(138)
200 020
122 550
133 276
140 344
26.12 1 956 102
17.33
436 301
18.48
492 876
19.34
544 784
isoCRP (updown)
1.70
900
(138)
100 789
15.44
354 291
60.67
isoGRASP
isoGRASP
isoGRASP
isoGRASP
2.50
9.65
5.85
3.26
1 856
1 856
1 856
1 856
(1 094)
(1 094)
(1 094)
(1 094)
215 302
150 079
157 614
163 483
13.99 1 846 034
10.59
499 641
10.84
543 677
11.42
588 081
123.01
40.88
43.31
48.06
2.50 1 856 (1 094)
2.50 1 856 (1 094)
2.69 1 879 (1 117)
156 766
120 327
120 327
11.38
11.56
10.06
(none)
(all)
(inf)
(scc)
isoGRASP (up)
isoGRASP (updown)
isoGRASP (sep)
547 253
387 053
387 053
267.40
69.62
77.46
85.17
44.62
45.42
37.77
(all). To speed up customization (at the cost of slightly increased query times), we only
run this search for boundary vertices that have an infinite eccentricity after the first run
(isoCRP (inf)). In fact, we can even try to make use of eccentricities previously updated
during the second run: before running a Dijkstra search for a vertex u with ecc` (u) = ∞,
we check all outgoing shortcuts (u, v) (on the current level `). If len((u, v)) + ecc` (v) 6= ∞,
we can use this sum as finite upper bound and omit the Dikstra search, yielding isoCRP
(scc).
Table 4 shows (sequential) query timings of different MLD variants (for 1 000 random
queries). The first block corresponds to isoCRP queries based on unrestricted eccentricites.
Next, isoCRP (updown) corresponds to the approach presented in the main document. For
isoGRASP, similar figures are reported. We also distinguish several variants of the approach
described in the main document. In particular, (up) uses unrestricted eccentricites only in
the second phase, (updown) uses restriced eccentricites (as described in Section 4), but does
not store vertices for additional checks in a dedicated adjacency array (instead, we scan
matrices of the overlay). Finally, (sep) is the variant described in the main text. Table 5
shows figures for parallel queries.
21
Table 5. Performance of (parallel) MLD query variants.
queries
custom
algorithm
isoCRP
isoCRP
isoCRP
isoCRP
time [s]
(none)
(all)
(inf)
(scc)
space [MiB]
limit x = 100 min
limit x = 500 min
# scans time [ms]
# scans time [ms]
1.70
8.42
4.91
2.48
900
900
900
900
(138)
(138)
(138)
(138)
200 020
122 550
133 276
140 344
3.92 1 956 102
2.84
436 301
2.99
492 876
3.02
544 784
isoCRP (updown)
1.70
900
(138)
100 789
2.73
354 291
7.86
isoGRASP
isoGRASP
isoGRASP
isoGRASP
2.50
9.65
5.85
3.26
1 856
1 856
1 856
1 856
(1 094)
(1 094)
(1 094)
(1 094)
215 302
150 079
157 614
163 483
2.73 1 846 034
2.35
499 641
2.40
543 677
2.47
588 081
13.59
6.04
6.37
6.97
2.50 1 856 (1 094)
2.50 1 856 (1 094)
2.69 1 879 (1 117)
156 766
120 327
120 327
2.37
2.50
2.35
(none)
(all)
(inf)
(scc)
isoGRASP (up)
isoGRASP (updown)
isoGRASP (sep)
22
547 253
387 053
387 053
25.02
9.08
8.87
9.36
6.39
6.19
5.93
Table 6. Performance of isoPHAST-CD and isoPHAST-CP for varying number of cells.
Queries are executed sequentially (seq.) and in parallel (par.) and given in milliseconds.
The variant with the least absolute (relative) deviations is highlighted in dark (light) gray.
# cells
algorithm
exec.
isoPHAST-CD seq.
isoPHAST-CP
τ
128
256
512
1 024
2 048
4 096
8 192 16 384
10 1.92 1.17
100 12.51 10.37
500 65.45 51.81
0.74 0.51 0.38 0.32
8.88 7.75 6.83 6.09
42.24 35.14 31.87 31.24
0.33
5.63
33.70
0.41
5.56
40.03
0.91
1.62
8.19
0.59 0.34 0.40 0.43
1.34 1.33 1.50 1.83
9.02 10.51 13.18 17.27
par.
10
100
500
1.13
2.63
23.58
2.08
4.10
38.68
seq.
10 4.36 4.50
5.08 6.35 8.25 10.86 14.56
100 15.30 14.05 13.69 13.98 14.93 16.59 19.19
500 68.26 54.21 43.40 35.41 31.10 28.44 28.41
19.30
23.21
30.64
par.
10
100
500
1.87
2.26
8.30
3.85
4.33
8.53
3.68
4.44
7.90
4.68
4.80
7.74
4.69
5.56
7.97
5.87 7.94
6.77 8.60
8.83 10.14
10.41
11.10
12.10
13.77
14.16
15.27
Parameter Tuning for isoPHAST. The following Table 6 shows running times of isoPHASTCD and isoPHAST-CP for partitions with different numbers of cells. We report sequential
and parallel running times for different time limits (given in minutes). Table 7 shows
sequential query times obtained for different choices of partition and compressed subgraph
size. Table 8 shows parallel running times for the same parameter set.
23
Table 7. Sequential query times [ms] of isoPHAST-DT for varying numbers of cells
in underlying partitions and compressed subgraph sizes |G↓c |. Query times are reported
for different time limits (given in minutes) The variant with the least absolute (relative)
deviations is highlighted in dark (light) gray.
# cells
|G↓c |
τ
128
256
512
1 024
2 048
4 096
8 192
16 384
128
10 11.23 6.50 4.14 2.88
2.40
2.42
2.80
3.32
100 28.96 23.51 21.98 22.39 26.56 37.02 56.37 85.55
500 111.79 98.99 95.56 94.05 105.11 134.85 195.08 280.13
256
10 11.13 6.39 4.04 2.77
100 28.87 23.23 21.17 21.41
500 109.61 96.93 91.62 89.42
2.29
2.23
2.49
3.07
25.19 33.99 50.37 77.16
99.34 122.52 174.39 250.95
512
10 10.97 6.50 3.86 2.60
100 28.11 22.36 19.93 19.55
500 108.24 92.88 85.50 79.81
2.09
1.95
2.13
2.63
22.26 28.78 41.79 63.52
84.85 100.98 138.64 200.31
1 024
10 10.77 5.99 3.66 2.38
100 27.44 21.16 18.45 17.16
500 104.46 87.64 77.81 68.71
1.88
18.83
70.07
1.65
22.84
75.56
1.68
2.02
29.75 44.99
94.72 132.81
2 048
10 10.62 5.85 3.52 2.27
100 26.96 20.31 17.11 15.42
500 102.93 83.64 71.59 61.26
1.73
15.61
57.49
1.47
17.41
57.40
1.41
20.76
64.10
1.54
27.35
78.22
4 096
10 10.58 5.84 3.49 2.27
100 26.73 19.77 16.23 14.49
500 100.66 81.07 67.56 55.98
1.71
14.07
49.85
1.43
14.15
46.75
1.34
15.32
46.39
1.39
18.17
51.84
8 192
10
100
500
10.72 5.96 3.61 2.40
26.46 19.50 15.75 13.42
99.99 79.52 65.14 52.90
1.82
12.31
44.87
1.53
11.83
39.33
1.42
11.79
36.61
1.43
12.68
36.31
16 384
10 11.52 6.25 3.91 2.69
100 26.33 19.56 15.70 13.37
500 101.44 79.08 63.83 50.95
2.12
11.87
42.51
1.83
10.92
35.80
1.75
10.32
32.17
1.81
10.51
29.02
32 768
10
100
500
11.96 6.84 4.46 3.26
26.94 19.94 16.05 13.50
97.85 78.36 63.57 50.28
2.69
12.03
41.49
2.39
11.17
34.30
2.28
10.09
29.35
2.62
9.72
26.66
65 536
10
100
500
12.48 7.81 5.44 4.22
27.54 21.01 17.47 14.34
98.19 79.07 63.54 50.44
3.65
12.76
41.63
3.36
11.85
34.07
3.26
10.64
28.80
3.24
10.03
25.03
24
Table 8. Parallel query times [ms] of isoPHAST-DT for varying numbers of cells in
underlying partitions and compressed subgraph sizes |G↓c |. Query times are reported
for different time limits (given in minutes) The variant with the least absolute (relative)
deviations is highlighted in dark (light) gray.
# cells
|G↓c |
τ
128
256
512 1 024 2 048
4 096
8 192 16 384
128
10 4.38 2.92 2.02
100 4.65 3.54 2.91
500 10.42 8.67 8.05
1.53
2.65
7.67
1.09 1.05 1.29
2.62 3.79 4.89
8.12 10.78 14.43
2.18
7.68
21.29
256
10 4.34 2.86 1.98
100 4.67 3.49 2.85
500 10.31 8.50 7.79
1.50
2.56
7.56
1.07
2.49
7.60
1.03 1.27
3.14 4.49
9.29 13.02
2.16
7.04
18.69
512
10 4.32 2.82 1.92
100 5.19 3.40 2.74
500 10.57 8.21 7.33
1.80
2.53
6.70
0.94
2.26
6.69
1.01 1.24
2.77 3.97
7.82 11.28
2.13
6.14
15.27
1 024
10
100
500
4.27 2.73 1.84
4.51 3.30 2.62
9.90 7.82 6.78
1.39
2.24
5.90
0.88
2.01
5.65
0.97
2.34
6.07
1.24
3.08
7.56
2.10
4.92
10.70
2 048
10
100
500
4.29 2.71 1.82
4.48 3.25 2.53
9.71 7.56 6.37
1.04
1.85
5.12
0.95
1.82
4.81
0.89
1.98
4.86
1.28
2.50
5.49
2.17
3.83
7.13
4 096
10
100
500
4.26 2.75 1.87
4.49 3.26 2.52
9.64 7.48 6.14
1.07
1.82
4.84
1.01
1.73
4.34
0.89
1.80
4.17
1.32
2.20
4.38
2.25
3.34
5.51
8 192
10
100
500
4.42 3.05 1.96
4.60 3.82 2.60
9.65 7.94 6.09
1.19
1.88
4.75
1.13
1.75
4.14
1.14
1.79
3.81
1.41
2.10
3.84
2.37
3.12
4.64
16 384
10
100
500
4.54 3.01 2.17
4.78 3.54 2.80
9.81 7.58 6.23
1.45
2.07
4.86
1.49
1.93
4.20
1.46
1.94
3.80
1.65
2.21
3.73
2.61
3.18
4.37
32 768
10 4.88 3.41 2.57
100 5.16 3.94 3.19
500 10.15 7.94 6.57
2.07
2.52
5.20
2.14
2.39
4.53
2.14
2.36
4.07
2.11
2.57
3.97
2.99
3.49
4.55
65 536
10 5.55 4.10 3.44
100 5.91 4.69 4.02
500 10.89 8.67 7.32
3.15
3.43
5.94
3.18
3.32
5.24
3.15
3.30
4.78
3.31
3.42
4.65
3.73
4.18
5.20
25
Table 9. Sequential queries of isoPHAST
prepro [s]
algorithm
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
queries
contract ecc./
contract
comp.
upward
cells diam. core/graph dist. tbl. time [ms]
17.37
19.43
13.41
0.58
0.74
5.26
–
1 171.57
672.19
–
–
307.94
scanning # active
time [ms]
cells
15.76
15.87
1.20 7.83/21.97
0.38
24.37
329
221
1 439
Table 10. Parallel queries of isoPHAST
prepro [s]
algorithm
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
queries
contract ecc./
contract
comp.
upward scanning # active
cells diam. core/graph dist. tbl. time [ms] time [ms]
cells
29.51
29.49
17.76
1.49
0.95
3.00
–
1 392.66
717.72
–
–
75.08
4.47
3.76
1.04 2.13/4.68
0.40
3.38
64
77
587
Detailed Timings for isoPHAST. Tables 9 and 10 show details about sequential and
parallel query times, respectively. We report duration of different phases of preprocessing
and queries. Moreover, we show the average number of active cells.
Scaling of Parallel Queries. Finally, Figure 4 shows parallel query times (using 16 cores)
for varying time limits for the same random queries as in Section 6. As mentioned before,
curves mainly resemble the ones in the single-core setup. For the isoPHAST-CP, the linear
sweep through the core graph is slightly faster now, since the core graph is smaller. The
isoPHAST-CD variant is the best technique for mid-range queries again, however, as before,
query performance get worse for larger time limits. Note that the performance gap between
isoCRP and isoGRASP is smaller when using multiple cores (as explained in the main text,
we conjecture that isoGRASP is limited by the memory bandwidth).
26
12
isoCRP
isoGRASP
isoPHAST-CD
isoPHAST-CP
isoPHAST-DT
11
10
9
Query Time [ms]
8
7
6
5
4
3
2
1
0
10
30
100
300
1000
3000
Limit [min]
Figure 4. Parallel query times using 16 threads for varying time limits.
27
| 8cs.DS
|
Examining the Classification Accuracy of TSVMs with Feature Selection
in Comparison with the GLAD Algorithm
Hala Helmi, Jonathan M. Garibaldi and Uwe Aickelin
Abstract- Gene expression data sets are used to classify
and predict patient diagnostic categories. As we know, it
is extremely difficult and expensive to obtain gene
expression labelled examples. Moreover, conventional
supervised approaches cannot function properly when
labelled data (training examples) are insufficient using
Support Vector Machines (SVM) algorithms. Therefore,
in this paper, we suggest Transductive Support Vector
Machines (TSVMs) as semi-supervised learning
algorithms, learning with both labelled samples data and
unlabelled samples to perform the classification of
microarray data. To prune the superfluous genes and
samples we used a feature selection method called
Recursive Feature Elimination (RFE), which is supposed
to enhance the output of classification and avoid the local
optimization problem. We examined the classification
prediction accuracy of the TSVM-RFE algorithm in
comparison with the Genetic Learning Across Datasets
(GLAD) algorithm, as both are semi-supervised learning
methods. Comparing these two methods, we found that
the TSVM-RFE surpassed both a SVM using RFE and
GLAD.
I.
INTRODUCTION
Data mining techniques have traditionally been used to
extract hidden predictive information in many diverse
contexts. Usually datasets contain thousands of examples.
Recently the growth in biology, medical science, and DNA
analysis has led to the accumulation of vast amounts of
biomedical data that require in-depth analysis.
After years of research and development, many data
mining, machine learning, statistical analysis systems and
tools are available to be used in biodata analysis.
Consequently, this paper will examine a relatively new
technique in data mining. This technique is called
Transductive Supervised Support Vector Machines [2], also
named Semi-Supervised Support Vector Machines S3VMs,
located between supervised learning with fully-labelled
training data and unsupervised learning without any labelled
training data [1]. In this method, we used both labelled and
unlabelled samples for training: a small amount of labelled
data and a large amount of unlabelled data.
Hala Helmi (PhD student), Jon Garibaldi (Assoc. Prof.) and Professor Uwe
Aickelin are with the Intelligent Modelling and Analysis (IMA) Research
Group, School of Computer Science, The University of Nottingham, Jubilee
Campus, Wollaton Road, Nottingham, NG8 1BB, U.K.
{hqh, jmg, uxa} @cs.nott.ac.uk
The purpose of this paper is to observe the performance of
Transductive SVMs combined with a feature selection
method called Recursive Feature Elimination (RFE), which
is used to select molecular descriptors for Transductive
Support Vector Machines (TSVMs).
The paper is organized as follows: section 2 provides a
literature review on Support Vector Machines, Transductive
Support Vector Machines and finally Recursive Feature
Elimination. In section 3, the TSVM algorithm combined
with RFE is detailed, as well as a brief summary of the
GLAD algorithm based on a recently published paper [3]
which aims to compare the prediction accuracy of these two
algorithms. Section 4 is dedicated to comparing and
analysing the experimental results of both algorithms
(TSVM and GLAD). Finally, a summary of the results and
discussion will be presented in section 5.
II.
BACKGROUND
A. Support Vector Machines
Support Vector Machine (SVMs), as a supervised
machine learning technique, perform well in several areas of
biological research, including evaluating microarray
expression data [4], detecting remote protein homologies [5]
and recognizing translation initiation sites [6]. SVMs have
demonstrated the ability not only to separate the entities
correctly into appropriate classes, but also to identify
instances where established classification is not supported by
the data [7]. SVMs are a technique that makes use of training
that utilizes samples to determine beforehand which data
should be clustered together [4].
B. Tranductive Support Vector Machines
Transductive learning is a method strongly connected to
semi-supervised learning, where semi-supervised learning is
intermediate between supervised and unsupervised learning.
Vapnik introduced Semi-Supervised Learning for Support
Vector Machines in the 1990s. His view was that
transduction (TSVM) is preferable to induction (SVM),
since induction needs to solve a more general problem
(inferring a function) before solving a more detailed one
(computing outputs for new cases) [8] [9].
Transductive Support Vector Machines attempt to
maximize the hyperplane classifier between two classes
using labelled training data; at the same time this forces the
hyperplane to be far away from the unlabelled samples.
TSVMs seem to be a perfect semi-supervised learning
algorithm because they combine the regularization of
Support Vector Machines with a straight forward
implementation of the clustering assumption [10].
C. Recursive Feature Elimination
Most prediction model algorithms are less effective when
the size of the data set is large. There are several methods for
decreasing the amount of the feature set. From among these
methods we selected a technique called Recursive Feature
Elimination (RFE). The basis for RFE is to begin with all the
features, select the least useful, remove this feature, and then
repeat until some stopping condition is reached.
Finding the best subset features is too expensive, so RFE
decreases the difficulty of feature selection by being
‘greedy’ [11].
III.
METHODS
This section of the paper is focused on describing TSVMRFE, the problem motivated by the task of classifying
biomedical data. The goal is to examine classifier accuracy
and classification errors using the Transductive Support
Vector Machines method, in order to determine whether this
method is an effective model when combined with Recursive
Feature Elimination (RFE) compared with the Genetic
Learning Across Datasets (GLAD) algorithm.
A. Support Vector Machines
The purpose of SVMs is to locate a classifier with the
greatest margin between the samples relating to two different
classes, where the training error is minimized. Therefore, to
achieve this we used a set of ݊-dimensional training samples
ࢄ = {ݔ}ୀଵ labelled {ࣳ}ୀଵ and their mapping { (ݔ)}ୀଵ via
kernel function:
′
ܓ൫ݔ, ݔ൯= (ईࣻ)் ൫ݔ൯
SVM has the following primal form:
Minimize over ( ܟ, ܾ, ξଵ, … , ξ )
‖‖ ܟ
+ ܥ ξ
ୀଵ
Subject to:
∀ୀଵ:ࣳ ( ܟ ( ∗ܢईࣻ)′ + ܾ) ≥ 1 − ξ, ξ ≥ 0
(1)
The SVM predictor for samples x, as shown below, was
settled on by the vector inner product between the ܟand the
mapped vector (x), plus the constant ܾ.
ࣳ = ܟ(ܖܛ (࢞)′ + ܾ)
The predictor actually corresponds to a separating
hyperplane in the mapped feature space. The prediction for
each training samples x୧ is connected with a violation term
ξ୧. The C is a user-specified constant to manage the penalty
for these violation terms.
The parameter p in the above (1) points to which kind of
norm of ܟis assessed. It is usually set to 1 or 2, resulting in
the 1-norm (lଵ-SVM) and 2-norm SVM (lଶ-SVM)
respectively. The 1-norm and 2-norm TSVMs have been
discussed in [12] and [13].
B. Transductive Support Vector Machines
In this paper, we are using the extended SVM technique
of transductive SVMs and we methodically adept the 2norm for the TSVM.
The standard setting can be illustrated as:
Minimize over (ࣳଵ∗ , … . , ࣳ ୩∗ , w, b, ξଵ, … , ξ୫ , ξଵ∗ )
୫
୩
୧ୀଵ
୨ୀଵ
1
‖w‖ଶଶ + C ξ୧ + C ∗ ξ∗୨
2
∀୫୧ୀଵ:ࣳ (w
∀୩୨ୀଵ:
ࣳ ౠ∗
subject to:
(z ∗ ईࣻ)′ + b) ≥ 1 − ξ୧, ξ୧ ≥ 0
′
ቀw ൫z ∗ x୨∗ ൯ + bቁ≥ 1 − ξ∗୨, ξ∗୨ ≥ 0
(2)
Where each ࣳ ୨∗ is the unknown label for x୨∗ which is one
of the k unlabelled samples; compared with SVM (1), the
formulation (2) of the TSVMs takes the unlabelled data into
consideration by representing the violation terms ξ∗୨ caused
by forecasting each unlabelled pattern (x୨∗ ) into ࣳ ୨∗ . The
penalty for these violation terms is controlled by a new
constant C ∗ labelled with unlabelled samples, while C
consists of labelled samples only.
Precisely solving the transductive problem needs a
search of all potential assignments of ࣳଵ∗ , … . , ࣳ ୩∗ and
identifying the various terms of ξ∗ which are regularly
intractable for large data sets. It is worth mentioning the lଶTSVM implemented in the SVMLight [18] [8].
C. Recursive Feature Elimination
Recursive Feature Elimination (RFE) has the advantage
of decreasing the number of redundant and recursive
features. RFE decreases the difficulty of feature selection by
being greedy.
To extend SVM feature selection techniques to
transductive feature selection is specifically straightforward,
as we can produce TSVM-RFE by iteratively eliminating
features with weights calculated from TSVM models. We
can explain the TSVM-RFE approach as the following
standard process.
1. Pre-process data and calculate filtering scores ̅ܛ.
Moreover, optionally further normalize data. This
approach first filters some features based on scores like
Pearson correlation coefficients.
2. Adjust ܢas an all-one input vector.
3. Set ̅ܛ ∗ܢ ← ܢ. Set part of the small entries of ܢzero
according to a proportion/threshold, and probably
discrete non-zero ܢto 1.
4. Obtain a (sub-) optimal TSVM as calculated by crossvalidation accuracy (2).
5. For RFE approaches, estimate feature weights ̅ܛfrom the
model in step 4 according to:
ଵ
J୲′ = − ∑୫୧,୨α୧ α୨ࣳ ୧ࣳ ୨ (x୧/ x୧୲) (x୨, x୨୲)′ + ∑୬୧ୀଵ α୧ (3)
ଶ
Where (x୧/x୧୲) indicates the input samples ݅with feature
ݐremoved. The weight of the ݐ-th feature can be clarified as
s୲ = ඥ|∆ J୲| = ට หJ − J୲′ ห
The following estimation suggested in [11] is easier to
measure.
s୲ଶ
୫
≈ α୧ α୨ࣳ ୧ࣳ ୨ ( x୧୲) (x୨୲)′
Step 3 comprises selecting a proportion/number of
features according to a threshold cutting the vector z. For
filtering scores and the RFE method, the vector z is changed
to a binary vector. Then the z ∗ x has the effect of pruning
or deactivating some features.
The threshold is usually found to prune a (fixed)
number/proportion of features at each iteration. The value of
the remaining features is then measured by the optimality of
the TSVM model obtained in step 4. We then apply crossvalidation accuracy as the performance measure for the
TSVM algorithm. For a subset of features selected by
choosing a threshold value, we extend the model search
upon the free parameters, such as [C, C ∗ σ (RBF), d(Poly)]
and choose the preferable parameter set which results in the
highest cross-validation accuracy.
D. Genetic Learning Across Datasets (GLAD)
The GLAD algorithm is different from prior algorithms
of semi-supervised learning. The GLAD algorithm has been
applied as a wrapper method for feature selection. A Genetic
Algorithm (GA) was implemented for generating a
population of related feature subsets. The labelled data and
the unlabelled data samples were computed separately.
Linear Discriminant Analysis (LDA) and K-means (K = 2)
for these two data forms of cluster algorithms were used
[14]. A distinctive two-term scoring function resulted to
independently score the labelled and unlabelled data
samples. Generally, the score was calculated as a weighted
average of the two terms as shown below.
(4)
As the typical leave-one-out-cross-validation accuracy
for the labelled training samples, they identified the labelled
data samples score. The unlabelled data samples score
consists of two terms: a cluster separation term and a steady
ratio term.
∑ஷหܥ − ܥห
݁ݎܿݏ௨ௗ =
−
1
∑ஷหܥ − ܥห+ ∑
∑หx − ܥห
ܰ
(5)
C୧ = centroid of cluster; π୧= ratio of data in cluster i ; πୣ୶୮ =
expected ratio in cluster i ; Nେ = number of data samples in
cluster i; nୡ = number of clusters.
IV.
୧,୨ୀଵ
Specifically, the feature weights are identical to the ܟif the
SVM is built upon a linear kernel.
Return to step 3 unless there is an acceptable number of
features/iterations. Output the closing predictor and features
highlighted by large values of z.
score = w × score୪ୟୠୣ୪୪ୣୢ + (1 − w)
× score୳୬୪ୟୠୣ୪୪ୣୢ
1
ඨ ݊ (ߨ− ߨୣ୶୮)ଶ
EXPERIMENTS AND RESULTS
This section discusses the results of the experiments
which were carried out in order to assess the effectiveness of
the classification model accuracy proposed in the previous
section.
A. Datasets
Leukaemia (AML-ALL): including 7129 probes,
two variants of leukaemia are available: acute
myeloblastic leukaemia (AML), 25 samples; and
acute lymphoblastic leukaemia (ALL), 47 samples
[15].
Lymphoma (DLBCL): consisting of 7129 genes
and 58 DLBCL samples. Diffuse large B-cell
lymphoma (DLBCL) and 19 samples of follicular
lymphoma (FL) [16].
Chronic Myeloid Leukaemia (CML): contained
30 samples (18 severe emphysema, 12 mild or no
emphysema) with 22,283 human gene probe sets
[17].
B. TSVM Recursive Feature Elimination (TSVM-RFE)
Result
Leukaemia (AML-ALL): the results for the
Leukaemia ALL/AML dataset are summarized in
Figure 1 in the diagram on the left. TSVM-RFE
gave the smallest minimal error of 3.68%, and
compassionately smaller errors compared with
SVM-RFE: 3.97% for 30, 40, ..., 70 genes.
Interestingly, in our experiments both methods gave
the lowest error when 60 genes were used. This
provided a reasonable suggestion for the number of
relevant genes that should be used for the
leukaemia data.
Lymphoma (DLBCL): the results for the
Lymphoma (DLBCL) dataset are summarized in
Figure 1 in the middle diagram. TSVM-RFE gave
the smallest minimal error of 3.89%, and quite
firmly smaller errors compared to the SVM-RFE:
4.72% for 30, 40, ..., 70 genes. For TSVM, the
methods gave the lowest error for 60 genes, while
SVM methods gave the lowest error at 50 genes
with 4.72% compared to 4.97% for 60 genes. This
could give a sensible suggestion for the number of
relevant genes that should be used for the
lymphoma (DLBCL) data.
30
40
50
60
Number of Genes
70
6.5
6
5.5
5
4.5
4
3.5
3
SVM-RFE
TSVM-RFE
Error Rate
SVM-RFE
TSVM-RFE
Error Rate
Error Rate
7.5
7
6.5
6
5.5
5
4.5
4
3.5
3
CML Dataset
DLBCL Dataset
AML/ALL Dataset
30
40
50
60
70
9
8.5
8
7.5
7
6.5
6
5.5
5
SVM-RFE
30
Number of Genes
40
50
60
70
Number of Genes
Figure 1: Testing error for three data sets. The 5-fold cross-validated pair t-test shows the SVM-RFE and the TSVM-RFE
have relative differences when comparing the two methods at the confidence rate of 95% (Linear kernel, C = 1).
Leukaemia (CML): lastly, the TSVM-RFE and
SVM-RFE results for the Leukaemia (CML) dataset
are provided in Figure 1 in the diagram on the right.
TSVM-RFE gave the smallest minimal error of
6.52%, and critically smaller errors in contrast to the
7.85% SVM-RFE for 30, 40, ..., 70 genes. Both
algorithms showed the lowest error when 50 genes
were used. This presented a sensible proposal for the
number of related genes that should be used for the
Leukaemia (CML) data.
C. Comparing the TSVM Algorithm result with the GLAD
Algorithm
Implementing Genetic Learning Across Datasets involved
conducting three experiments using previous datasets, each
addressing a different cancer diagnostic problem: ALL/AML
for disparity in diagnosis; in CML a dataset predicting the
response of imatinib; and in DLBCL for forecasting outcome.
In the AML-ALL dataset, the accuracy range using only
labelled samples was 73.46%. Combining unlabelled samples
with labelled samples increased the range to 75.14%. Adding
unlabelled samples increased the accuracy from 59.34% to
65.57% in the CML experiments. The addition of the
unlabelled samples to the unlabelled samples for DLBCL
raised the accuracy from 49.67% to 55.79%. This shows that
the GLAD algorithm outperformed the SVM-RFE and TSVMRFE in some cases when we made use of the labelled data only
without gene selection. In Table 1, for example the AML-ALL
dataset, the GLAD algorithm gives 73.46% while SVM-RFE
and TSVM-RFE accuracy are 52.8% and 55.6% respectively.
However, in the second dataset DLBCL showed that
GLAD algorithm accuracy was 49.67% and SVM-RFE 55.8%.
Furthermore, the third dataset of CML, SVM-RFE gave
59.02% without gene selection, while GLAD gave 59.34%.
On the other hand, TSVM exceeded GLAD when making use
of unlabelled data along with labelled data and selecting genes.
The results are shown in Table 1.
For instance, for the CML dataset using all the samples
without gene selection TSVM gave 72.6% when selecting
genes based on REF, TSVM exceeded 93.48%, while GLAD
gave 65.57% with gene selection. In the same vein, the
accuracy for the DLBCL dataset achieved 96.11% by TSVM
with gene selection.
On the other hand, the GLAD algorithm gave 55.79% with
gene selection. As well as this, the TSVM with the AML-ALL
dataset with gene selection gave 96.32% while the GLAD
algorithm gave 75.14%. This means that the TSVM performed
better than the GLAD algorithm, and the performance with
gene selection showed a superior result.
Table 1: Accuracy obtained with SVM-RFE,
TSVM-RFE and GLAD
Dataset
ALL-AML
Without
Selection
With
Selection
DLBCL
Without
Selection
With
Selection
CML
Without
Selection
With
Selection
7219 Genes,
72 Samples
60 Genes,
72 Samples
7219 Genes,
77 Samples
60 Genes,
77 Samples
22,283
Genes,
30 Samples
50 Genes,
30 Samples
V.
SVM-RFE
Accuracy
(labelled)
TSVMRFE
Accuracy
52.8%
55.6%
73.46%
(labelled)
96.03%
96.32%
75.14%
55.8%
57.1%
95.03%
96.11%
49.67%
(labelled)
55.79%
59.02%
72.6%
59.34%
(labelled)
92.15%
93.48%
65.57%
GLAD
Accuracy
CONCLUSION
This paper has investigated topics focused on semisupervised learning. This was achieved by comparing two
different methods for semi-supervised learning using
previously classified cancer datasets.
The results on average for semi-supervised learning
surpassed those for supervised learning. However, this shows
that the GLAD algorithm outperformed SVM-RFE when we
made use of the labelled data only. On the other hand, TSVMRFE exceeded GLAD when unlabelled data along with
labelled data were used; it performed much better with gene
selection and performed well even if the labelled dataset was
small.
On the other hand, TSVM still had some drawbacks when
increasing the size of the labelled dataset, as the performance
did not significantly improve accordingly. Moreover, when the
size of the unlabelled samples was extremely small, the time
complexity was correspondingly high.
As with almost all semi-supervised learning algorithms,
TSVM showed some instability, as some results of different
runs were not the same. This occurred because unlabelled
samples may have been wrongly labelled during the learning
process. If we find a way in future to select and eliminate the
unlabelled sample first, we can then limit the number of newlylabelled samples for re-training the classifiers.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Abney, S. "Semisupervised Learning for Computational
Linguistics" (1st ed.). Chapman & Hall/CRC, 2007.
Zhu, X., "Semi-Supervised Learning Tutorial".
Department of Computer Sciences University of
Wisconsin, Madison, USA, 2007.
Harris, C. and Ghaffari, N. "Biomarker discovery across
annotated and unannotated microarray datasets using
semi-supervised learning". BMC Genomics 9 Suppl 2:S7,
2008.
Brown, M.P., Grundy, W.N., Lin, D., Cristianini, N.,
Sugnet, C.W., Furey, T.S., Ares, M. Jr., Haussler, D.
"Knowledge-based analysis of microarray gene expression
data by using support vector machines". Proc. Natl. Acad.
Sci. USA 97:262–267. 2001.
Jaakkola, T., Haussler, D. and Diekhans, M. "Using the
Fisher kernel method to detect remote protein
homologies". In Proceedings of ISMB, 1999.
Zien, A., Ratsch, G., Mika, S., Scholkopf, B., Lemmen,
C., Smola, A., Lengauer, T., and Muller, K.-R.
"Engineering support vector machine kernels that
recognize translation initiation sites", Bioinformatics,
16:799- 807, 2000.
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Valafar, F. "Pattern Recognition Techniques in Microarray
Data Analysis". Annals of the New York Academy of
Sciences 980(1): 41-64, 2002.
Gommerman, A., Vovk, V. and Vapnik, V. "Learning by
Transduction", In Uncertainty in Artificial Intelligence,
pp.148-155,1998.
Collobert, R., Sinz, F., Weston, J. and Bottou, L. "Large
Scale Transductive SVMs". J. Mach. Learn. Res. 7
(December 2006), 1687-1712, 2006.
Zhang, R., Wang, W. et al. "Least Square Transduction
Support Vector Machine." Neural Processing Letters
29(2): 133-142, 2009.
Guyon, I., Weston, J. et al. "Gene Selection for Cancer
Classification using Support Vector Machines Machine
Learning". 46(1): 389-422, 2002.
Bennett, K. and Demiriz, A. "Semi-supervised support
vector machines". In NIPS,1998.
Joachims, T. "Transductive inference for text classification
using support vector machines". In Proceedings of ICML99, pages 200-209, 1999.
Han, J. and Kamber, M. "Data Mining: Concepts and
Techniuqes". Morgan Kaufmann Publishers, San
Francisco, CA, 2001.
Golub, T.R. et al. “Molecular classification of cancer:
class discovery and class prediction by gene expression
monitoring”. Science, 286:531-537, 1999.
Shipp, M.A. et al. "Diffuse large B-cell lymphoma
outcome prediction by gene expression profiling and
supervised machine learning". Nature Medicine, 8:68-74.
2002.
Yong, A.S.M. et al. "Molecular profiling of CD34+ cells
identifies low expression of CD7, along with high
expression of proteinase 3 or elastase, as predictors of
longer survival in patients with CML". Blood 2006,
107:205-12, 2006.
Joachims, T. "Making large-scale support vector machine
learning practical". In Advances in Kernel Methods:
Support Vector Machines, 1999.
| 5cs.CE
|
On the Duality of Fractional Repetition Codes
arXiv:1710.07909v1 [cs.IT] 22 Oct 2017
Bing Zhu†, Kenneth W. Shum‡ , and Hui Li†
† School of Electronic and Computer Engineering, Peking University
‡ Institute of Network Coding, The Chinese University of Hong Kong
E-mail: [email protected], [email protected], [email protected]
Abstract—Erasure codes have emerged as an efficient technology for providing data redundancy in distributed storage systems.
However, it is a challenging task to repair the failed storage nodes
in erasure-coded storage systems, which requires large quantities
of network resources. In this paper, we study fractional repetition
(FR) codes, which enable the minimal repair complexity and also
minimum repair bandwidth during node repair. We focus on the
duality of FR codes, and investigate the relationship between the
supported file size of an FR code and its dual code. Furthermore,
we present a dual bound on the supported file size of FR codes.
Index Terms—Distributed storage systems, regenerating codes,
repair bandwidth, fractional repetition codes, dual code.
I. I NTRODUCTION
In the recent decades, erasure coding techniques are gaining
more and more popularity owing to the high storage efficiency.
For the same storage overhead, erasure codes can improve the
data reliability than the conventional replication-based scheme
[1]. Therefore, enterprise storage systems are also transitioning
to erasure coding techniques for better storage features [2], [3].
In an erasure coding scheme, an original data object is mapped
into n coded blocks and spread across n distinct storage nodes,
satisfying that any k out of the n nodes are sufficient to retrieve
the source file. Even though traditional erasure codes (such as
Reed-Solomon codes) are capable of reducing the data storage
overhead while preserving the fault tolerance guarantees, they
are less efficient in failure recovery as compared to replication.
When replicated data is lost, it can be simply reconstructed by
downloading copies from other surviving nodes. However, for
erasure codes, the conventional method for recovering the lost
data is to reconstruct the original file by connecting to k nodes
and then re-encode to obtain the data stored in the failed node.
In other words, repairing a single failed node requires the data
from k other nodes, resulting in a k times higher reconstruction
traffic than the replication scheme.
Regenerating codes are proposed in [4] aiming to reduce the
amount of data transmission during the node repair operation.
Specifically, an (n, k, d, M, α, β) regenerating code encodes a
data object of size M into nα coded packets, which are spread
across n storage nodes, with each node containing α packets.
The stored data can be reconstructed by contacting any subset
of k nodes. Upon failure of a storage node, it can be replaced
by a new node, whose content is generated by downloading β
packets from each of any set of d surviving nodes. Particularly,
Hui Li is also with the Shenzhen Key Lab of Information Theory and Future
Internet Architecture, Shenzhen Eng. Lab of Converged Networks Technology,
and the Future Network PKU Lab of National Major Research Infrastructure,
Shenzhen Graduate School, Peking University, China.
for the case that the total amount of network transfer (i.e., dβ)
equals to the size of lost content stored in the failed node, then
the corresponding code is referred to as a minimum-bandwidth
regenerating (MBR) code [4]. However, the reduction in repair
bandwidth is at the cost of imposing an encoding capability on
the helper nodes, i.e., the β transferred packets are obtained as
a function of the packets stored in the contacted node.
Rashmi et al. [5] introduced the notion of repair-by-transfer
for node repair, where each helper node transfers a portion of
the stored data to the replacement node without any arithmetic
operations. Such a repair model eliminates the computational
burden at the helper nodes. El Rouayheb and Ramchandran [6]
further generalized the code constructions in [5] and proposed
fractional repetition codes. The encoding process of FR codes
consists of the concatenation of two codes: an outer maximumdistance-separable (MDS) code followed by an inner fractional
repetition code. The data objects are first encoded by an MDS
code, and then the encoded packets are replicated and stored in
the system according to the FR code. The outer MDS code is
used for data reconstruction, and the inner FR code is designed
to facilitate the node recovery process. In the presence of node
failures, the replacement node simply downloads the required
data from a specific set of surviving nodes without the need of
computations. The transferred packets by the contacted helper
nodes are the same as the content of the failed storage node. In
such a sophisticated manner, FR codes achieve uncoded exact
repair at the MBR point.
El Rouayheb and Ramchandran presented in [6] two explicit
constructions of FR codes based on regular graphs and Steiner
systems. Several recent studies extend the design of FR codes
to a larger set of parameters, which are based on randomized
algorithm [7], bipartite cage graphs [8], resolvable designs [9],
incidence matrix [10], group divisible designs [11], transversal
designs [12], symmetric designs [13], Kronecker product [14],
and perfect difference families [15]. Moreover, the supported
file size of constructed FR codes are also investigated in [12],
[14] and [15].
Contributions: In this paper, we consider the duality of fractional repetition codes. Based on the relationship between the
supported file size of an FR code and its dual code, we present
a dual bound on the supported file size of the original FR code.
Organization: The rest of this paper is organized as follows.
Section II introduces the duality of fractional repetition codes.
We elaborate on the file size hierarchy of FR codes in Section
III. Section IV presents a dual bound on the supported file size
of FR codes. Finally, we conclude in Section V.
structure specified by C. Each coded packet is associated with
a certain point in P, and the storage nodes are associated with
the blocks in B. A packet is stored in a node if and only if the
corresponding point and block are incident. Hence, each node
contains α packets, and each packet is stored in ρ nodes. Fig. 1
shows the encoding process of C. Since a packet is replicated
ρ times, the system can tolerate any ρ − 1 node failures.
The dual of C is defined as the FR code (B, P, I t ), where
t
I is the subset of B × P defined by
I t := {(B, p) : (p, B) ∈ I}.
Let C t denote the dual of C. We note that the incidence matrix
of C and C t are the transpose of each other. In [6], the authors
refer to the dual FR code as the Transpose code.
Figure 1. The encoding process of an (n, α, v, ρ)-FR code. A data object of
size M is first encoded into v coded packets using an MDS code. Then each
coded packet is equally replicated ρ times and distributed to n storage nodes
N1 , N2 , . . . , Nn according to the FR code, where each node stores α packets.
II. F RACTIONAL R EPETITION C ODE
AND I TS
D UAL
An incidence structure is a triple (P, B, I), where P and B
are two finite sets, and I is a subset of P ×B. The elements in
P are called points, and the elements in B are called blocks.
A point p ∈ P is incident with a block B ∈ B if (p, B) ∈ I.
The incidence matrix of an incidence structure is a |B| × |P|
zero-one matrix, with rows indexed by the blocks and columns
indexed by the points, such that the entry corresponding to a
point p and a block B is equal to 1 if and only if p is incident
with B [16].
In this setting, it is permissible that two distinct blocks are
incident with the same set of points. We say that the incidence
structure contains repeated blocks. For the case that there are
no repeated blocks in an incidence structure (P, B, I), we say
that the incidence structure is simple, and any block is uniquely
determined by the points incident with it. In other words, each
block in B can be viewed as a subset of P.
A fractional repetition (FR) code is an incidence structure in
which each point is incident with ρ blocks, and each block is
incident with α points, for some constants ρ and α. Therefore,
the incidence matrix of an FR code has constant row sum α,
and constant column sum ρ. Notice that a fractional repetition
code is also known as a tactical configuration in the literature
of combinatorial designs. Let n denote the number of blocks,
and v be the number of points in an FR code. We refer to such
an FR code as an (n, α, v, ρ)-FR code. These code parameters
satisfy the following relationship
nα = vρ.
(1)
Example 1. When every block is incident with α = 2 points,
then the notion of an FR code coincides with a ρ-regular graph
(possibly with multiple edges).
An (n, α, v, ρ)-FR code C is used to distribute data packets
across a storage system. We encode a data object into v coded
packets by using a proper MDS code, which are then replicated
and spread across n storage nodes according to the incidence
Lemma 1. Let C be an (n, α, v, ρ)-FR code.
(i) The dual code of C is a (v, ρ, n, α)-FR code.
(ii) The double dual of C is C itself.
Example 2. Let C be the incidence structure obtained from
the line graph of the complete graph on five vertices. This gives
the (5, 4, 10, 2)-FR code with incidence matrix
1 1 1 1 0 0 0 0 0 0
1 0 0 0 1 1 1 0 0 0
0 1 0 0 1 0 0 1 1 0
0 0 1 0 0 1 0 1 0 1
0 0 0 1 0 0 1 0 1 1
as discussed in [5]. We note that this is a 5 × 10 matrix with
constant row sum α = 4 and constant column sum ρ = 2. The
dual of this FR code is a (10, 2, 5, 4)-FR code with incidence
matrix
1 1 0 0 0
1 0 1 0 0
1 0 0 1 0
1 0 0 0 1
0 1 1 0 0
0 1 0 1 0.
0 1 0 0 1
0 0 1 1 0
0 0 1 0 1
0 0 0 1 1
III. T HE H IERARCHY
OF
S UPPORTED F ILE S IZE
We now formally define the supported file size Mk (C) of an
(n, α, v, ρ)-FR code C = (P, B, I). For k = 1, 2, . . . , n, the
supported file size Mk (C) of C is defined as
Mk (C) :=
min
K⊂B,|K|=k
|{p ∈ P : ∃B ∈ K, (p, B) ∈ I}|, (2)
where the minimum is taken over all k-subsets K of the block
set B. The parameter k is called the reconstruction degree. By
definition, the value of Mk (C) essentially refers to the smallest
number of distinct packets one can download from k storage
nodes. For a fixed reconstruction degree k, we can encode the
data object by an MDS code with parameters [v, Mk (C)], such
that the original data can be recovered from any Mk (C) out of
the v coded symbols. Treating each symbol as a coded packet,
and
Nk (C) < Nk0 (C) for all k > k0 ,
Nℓ (C t ) < Nℓ0 (C t ) for all ℓ > ℓ0 .
Therefore, we obtain
v,
v − 1,
v − 2,
Nk (C) = .
..
1,
0,
for
for
for
..
.
for
for
k = 0,
0 = Nv (C t ) < k ≤ Nv−1 (C t ),
Nv−1 (C t ) < k ≤ Nv−2 (C t ),
(6)
N2 (C t ) < k ≤ N1 (C t ),
N1 (C t ) < k ≤ N0 (C t ) = n.
Based on the analysis, we obtain the following theorem.
Figure 2. The graphic illustration of (k, Nk (C)) and (Nℓ (C t ), ℓ).
we can distribute the coded packets to n storage nodes based
on the FR code C such that any k storage nodes are sufficiently
in decoding the data object.
If we increase the value of k, then the resulting supported
file size also increases, i.e.,
α = M1 (C) ≤ M2 (C) ≤ · · · ≤ Mn (C) = v.
(3)
We call the above the hierarchy of supported file size of C.
We shall also define M0 (C) := 0 by convention. Similarly, the
file size hierarchy of the dual code C t is
ρ = M1 (C t ) ≤ M2 (C t ) ≤ · · · ≤ Mv (C t ) = n.
(4)
Notice that there is a close relationship between Mk (C) and
Mℓ (C t ). This property can be seen from the fact that if we can
find an x × y all-zero submatrix in the incidence matrix of C,
then we have
Mx (C) ≤ v − y and My (C t ) ≤ n − x.
This motivates us to define
max
K⊂B,|K|=k
|{p ∈ P :6 ∃B ∈ K, (p, B) ∈ I}|
i=1
where I(P ) is the indicator function equal to 1 if the condition
P is true and 0 otherwise. In this case, the right-hand side term
of (8) counts the number of i ∈ {1, 2, . . . , v} such that Ni (C t )
is strictly less than k. Thus,
v
X
I(k > Ni (C t )) = v − ℓ for Nℓ+1 (C t ) < k ≤ Nℓ (C t ),
where k = 1, 2, . . . , n.
(5)
with the maximum taken over all subsets K ⊂ B of size k. By
definition, Nk (C) is the largest integer ℓ such that we can find
an k × ℓ all-zero submatrix in the incidence matrix of C. From
(3) and (4), we have
v = N0 (C) > N1 (C) ≥ N2 (C) ≥ · · · ≥ Nn (C) = 0,
n = N0 (C t ) > N1 (C t ) ≥ N2 (C t ) ≥ · · · ≥ Nv (C t ) = 0.
We shall plot the points (k, Nk (C)) for k = 0, 1, . . . , n, and
(Nℓ (C t ), ℓ) for ℓ = 0, 1, . . . , v in the same figure. The results
can be found in Fig. 2. A Pareto optimal point, say (k0 , ℓ0 ), is
a vertex of the graph that satisfies
ℓ0 = Nk0 (C) and k0 = Nℓ0 (C t ),
Remark 3. Note that the identities in (7) can be expressed in a
more compact way by
v
X
I(k > Ni (C t )),
(8)
Mk (C) =
i=1
Nk (C) := |P| − Mk (C)
=
Theorem 2. Let C be an (n, α, v, ρ)-FR code. With Nℓ (C t ) as
defined in (5), we have
v,
for N1 (C t ) < k ≤ n = N0 (C t ),
t
t
v − 1, for N2 (C ) < k ≤ N1 (C ),
v − 2, for N3 (C t ) < k ≤ N2 (C t ),
Mk (C) = .
..
..
.
2,
for Nv−1 (C t ) < k ≤ Nv−2 (C t ),
1,
for Nv (C t ) = 0 < k ≤ Nv−1 (C t ).
(7)
Example 2 (continued). For k = 1, 2, . . . , 5, we can compute
that the supported file size Mk (C) of the complete graph based
FR code C is
10, for k = 4, 5,
9, for k = 3,
Mk (C) =
7, for k = 2,
4, for k = 1,
and the values of Nℓ (C t ) for ℓ = 1, 2, . . . , 10 are
0 = N7 (C t ) = N8 (C t ) = N9 (C t ) = N10 (C t ),
1 = N4 (C t ) = N5 (C t ) = N6 (C t ),
2 = N2 (C t ) = N3 (C t ),
3 = N1 (C t ).
Theorem 4. ([12, Lemma 32]) If we store a data object of size
M by using an (n, α, v, ρ)-FR code C, then the reconstruction
degree k is lower bounded by
l n M−1 m
α
k≥
+ 1.
(11)
v
α
Proof: By applying the bound in (9) to the dual code of
C, we obtain
v−α
t
ℓ
,
(12)
Mℓ (C ) ≤ n 1 − v
ℓ
for ℓ = 1, 2, . . . , v. (We can remove the floor operator without
loss of generality.) Hence,
v−α
v−α
t
ℓ
ℓ
Nℓ (C ) ≥ n − n 1 − v
=n v .
(13)
ℓ
ℓ
Given an integer M between 1 and v, we let ℓ be the integer
that satisfies
M = v − ℓ + 1.
Figure 3. The relationship between Mk (C) and Mℓ (C t ).
By Theorem 2, we obtain
Furthermore, the supported file size hierarchy of C t is 5 −
Nℓ (C t ), i.e,
5 = M10 (C t ) = M9 (C t ) = M8 (C t ) = M7 (C t ),
4 = M6 (C t ) = M5 (C t ) = M4 (C t ),
3 = M3 (C t ) = M2 (C t ),
2 = M1 (C t ).
Fig. 3 illustrates the relationship between Mk (C) and Mℓ (C t ).
We can obtain the two supported file size functions if we view
the stair-case graph from two different perspectives, which are
distinguished with different colors.
IV. D UAL B OUND
ON
S UPPORTED F ILE S IZE
t
k ≥ Nℓ (C ) + 1 ≥ n
v−α
v−M+1
v
v−M+1
+1=n
M−1
α
v
α
+ 1.
(14)
The proof of this theorem is completed by taking the ceiling of
both sides.
The above proof illustrates that Theorem 4 is essentially the
same as the basic bound (9) on the supported file size. We can
obtain a slight improvement if we use the bound given in (10)
instead.
Theorem 5. Given an FR code C with parameters (n, α, v, ρ),
we define the function g ′ (ℓ) recursively by
l αg ′ (ℓ) − ℓρ m
,
g ′ (1) := ρ, g ′ (ℓ + 1) := g ′ (ℓ) + ρ −
v−ℓ
Two upper bounds on the supported file size of FR codes are
presented in [6]. Specifically, the supported file size Mk (C) of
an (n, α, v, ρ)-FR code C is upper bounded by
n−ρ k
j
k
,
(9)
Mk (C) ≤ v 1 − n
for ℓ = 1, 2, . . . , v − 1. Then, for all k = 1, 2, . . . , n, we have
Mk (C) ≤ g(k),
Mℓ (C t ) ≤ g ′ (ℓ).
(16)
Nℓ (C t ) ≥ n − g ′ (ℓ),
(17)
k
and
v
X
I(k > n − g ′ (i)).
(15)
i=1
Proof: We notice that the function g ′ (ℓ) is the counterpart
of the recursive bound (10) on the dual code. Thus,
(10)
Since
where g(k) is defined recursively by
g(1) := α, g(k + 1) := g(k) + α −
Mk (C) ≤
l ρg(k) − kα m
n−k
.
The bound in (10) is not explicit and is relatively more difficult
to compute as compared to (9). However, it can be shown that
(10) is tighter than (9) in general.
We note that Theorem 2 provides a link between an FR code
and its dual. Using the mechanism in the previous section, we
give an alternate proof of a bound on the reconstruction degree
k, which is first obtained in [12].
for all ℓ, in view of the remark after the proof of Theorem 2,
we have
v
v
X
X
Mk (C) =
I(k > Nℓ (C t )) ≤
I(k > n − g ′ (ℓ)), (18)
ℓ=1
ℓ=1
which completes the proof.
We refer to the inequality in (15) as the dual bound on the
supported file size.
Example 3. Consider an explicit FR code C with parameters
(n, α, v, ρ) = (15, 2, 10, 3). The bounds in (9) and (10) suggest
that the supported file size with reconstruction degree k = 6 is
upper bounded by
k
j
12
15
M6 (C) ≤ 10 1 −
/
= 8, and
6
6
M6 (C) ≤ g(6) = 7, respectively.
Moreover, the recursive bound applied to the dual code yields
Mℓ (C t ) ≤ g ′ (ℓ) with
g ′ (1) = 3, g ′ (2) = 5, g ′ (3) = 7, g ′ (4) = 9, g ′ (5) = 11,
g ′ (6) = 12, g ′ (7) = 13, g ′ (8) = 14, g ′ (9) = g ′ (10) = 15.
Then, the dual bound in (15) gives
M6 (C) ≤
10
X
i=1
I(6 > 15 − g ′ (i)) =
10
X
I(g ′ (i) > 9) = 6.
i=1
This bound can be achieved by the (15, 2, 10, 3)-FR code listed
in the database in [10] with the following incidence matrix:
1 1 0 0 0 0 0 0 0 0
1 0 1 0 0 0 0 0 0 0
1 0 0 1 0 0 0 0 0 0
0 1 0 0 1 0 0 0 0 0
0 1 0 0 0 1 0 0 0 0
0 0 1 0 0 0 1 0 0 0
0 0 1 0 0 0 0 1 0 0
0 0 0 1 0 0 0 0 1 0 .
0 0 0 1 0 0 0 0 0 1
0 0 0 0 1 1 0 0 0 0
0 0 0 0 1 0 1 0 0 0
0 0 0 0 0 1 0 1 0 0
0 0 0 0 0 0 1 0 1 0
0 0 0 0 0 0 0 1 0 1
0 0 0 0 0 0 0 0 1 1
We note that the six storage nodes associated to rows 1, 2, 3,
4, 5, and 10 contain precisely 6 distinct packets. Therefore, this
FR code can support a file size of M = 6 with k = 6, implying
that it is optimal by the dual bound.
V. C ONCLUSION
In this paper, we investigate fractional repetition codes with
an emphasize on the duality of FR codes. We start by studying
the relationship between the supported file size of an FR code
and its dual code. Furthermore, we present an improved upper
bound on the supported file size of FR codes, which we refer to
as the dual bound.
Even though we focus on FR codes whose incidence matrix
has constant row sum and also constant column sum, the main
Theorem 2 works for any incidence matrix based code. These
codes can be viewed as a generalization of FR codes, and are
called heterogeneous FR codes. Explicit constructions of such
codes can be found in [17]–[21]. The relationship between the
file size of a heterogeneous FR code and its dual code can be
determined following a similar procedure as in Theorem 2.
ACKNOWLEDGMENT
This work was partially supported by the National Natural
Science Foundation of China under Grant 61671001, National
Key R&D Program of China under Grant 2016YFB0800101
and 2017YFB0803204, Shenzhen Basic Research Project under Grant 20170306092030521 and 20150331100723974, and
a grant from the University Grants Committee of Hong Kong
Special Administrative Region (Project No. AoE/E-02/08).
R EFERENCES
[1] H. Weatherspoon and J. D. Kubiatowicz, “Erasure coding vs. replication:
A quantitative comparison,” in Proc. Int. Workshop on Peer-to-Peer Syst.
(IPTPS), Mar. 2002, pp. 328–338.
[2] C. Huang, H. Simitci, Y. Xu, A. Ogus, B. Calder, P. Gopalan, J. Li,
and S. Yekhanin, “Erasure coding in Windows Azure Storage,” in Proc.
USENIX Annu. Tech. Conf., Jun. 2012, pp. 15–26.
[3] M. Sathiamoorthy, M. Asteris, D. Papailiopoulos, A. G. Dimakis, R.
Vadali, S. Chen, and D. Borthakur, “XORing elephants: Novel erasure
codes for big data,” in Proc. 39th Int. Conf. Very Large Data Bases
(VLDB), Aug. 2013, pp. 325–336.
[4] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” IEEE Trans. Inf.
Theory, vol. 56, no. 9, pp. 4539–4551, Sep. 2010.
[5] K. V. Rashmi, N. B. Shah, P. V. Kumar, and K. Ramchandran, “Explicit
construction of optimal exact regenerating codes for distributed storage,”
in Proc. 47th Annu. Allerton Conf., Sep. 2009, pp. 1243–1249.
[6] S. El Rouayheb and K. Ramchandran, “Fractional repetition codes for
repair in distributed storage systems,” in Proc. 48th Annu. Allerton Conf.,
Oct. 2010, pp. 1510–1517.
[7] S. Pawar, N. Noorshams, S. El Rouayheb, and K. Ramchandran, “Dress
codes for the storage cloud: simple randomized constructions,” in Proc.
IEEE Int. Symp. Inf. Theory (ISIT), Jul. 2011, pp. 2338–2342.
[8] J. Koo and J. Gill, “Scalable constructions of fractional repetition codes
in distributed storage systems,” in Proc. 49th Annu. Allerton Conf., Sep.
2011, pp. 1366–1373.
[9] O. Olmez and A. Ramamoorthy, “Repairable replication-based storage
systems using resolvable designs,” in Proc. 50th Annu. Allerton Conf.,
Oct. 2012, pp. 1174–1181.
[10] S. Anil, M. K. Gupta and T. A. Gulliver, “Enumerating some fractional
repetition codes,” arXiv:1303.6801 [cs.IT], Mar. 2013.
[11] B. Zhu, K. W. Shum, H. Li, and S.-Y. R. Li, “On low repair complexity
storage codes via group divisible designs,” in Proc. 19th IEEE Symp.
Comput. Commun. (ISCC), Jun. 2014, pp. 1–5.
[12] N. Silberstein and T. Etzion, “Optimal fractional repetition codes based
on graphs and designs,” IEEE Trans. Inf. Theory, vol. 61, no. 8, pp.
4164–4180, Aug. 2015.
[13] B. Zhu and H. Li, “Adaptive fractional repetition codes for dynamic
storage systems,” IEEE Commun. Lett., vol. 19, no. 12, pp. 2078–2081,
Dec. 2015.
[14] O. Olmez and A. Ramamoorthy, “Fractional repetition codes with
flexible repair from combinatorial designs,” IEEE Trans. Inf. Theory,
vol. 62, no. 4, pp. 1565–1591, Apr. 2016.
[15] H. Park and Y.-S. Kim, “Construction of fractional repetition codes with
variable parameters for distributed storage systems,” Entropy, vol. 18,
no. 12, pp. 1–8, Dec. 2016.
[16] E. F. Assmus and J. D. Key, Designs and Their Codes. Cambridge, U.K.:
Cambridge Univ. Press, 1992.
[17] M. K. Gupta, A. Agrawal, and D. Yadav, “On weak Dress codes for
cloud storage,” arXiv:1302.3681 [cs.IT], Feb. 2013.
[18] B. Zhu, K. W. Shum, H. Li, and H. Hou, “General fractional repetition
codes for distributed storage systems,” IEEE Commun. Lett., vol. 18,
no. 4, pp. 660–663, Apr. 2014.
[19] Q. Yu, C. W. Sung, and T. H. Chan, “Irregular fractional repetition
code optimization for heterogeneous cloud storage,” IEEE J. Sel. Areas
Commun., vol. 32, no. 5, pp. 1048–1060, May 2014.
[20] B. Zhu, K. W. Shum, and H. Li, “Heterogeneity-aware codes with
uncoded repair for distributed storage systems,” IEEE Commun. Lett.,
vol. 19, no. 6, pp. 901–904, Jun. 2015.
[21] K. G. Benerjee and M. K. Gupta, “On Dress codes with flowers,” in
Proc. Int. Workshop Signal Design Appl. Commun. (IWSDA), Sep. 2015,
pp. 108–112.
| 7cs.IT
|
Robust Stochastic Configuration Networks with
Kernel Density Estimation
arXiv:1702.04459v2 [cs.NE] 29 May 2017
Dianhui Wang∗, Ming Li
Department of Computer Science and Information Technology
La Trobe University, Melbourne, VIC 3086, Australia
Email:[email protected]
Abstract
Neural networks have been widely used as predictive models to fit data
distribution, and they could be implemented through learning a collection of
samples. In many applications, however, the given dataset may contain noisy
samples or outliers which may result in a poor learner model in terms of
generalization. This paper contributes to a development of robust stochastic
configuration networks (RSCNs) for resolving uncertain data regression problems.
RSCNs are built on original stochastic configuration networks with weighted least
squares method for evaluating the output weights, and the input weights and biases
are incrementally and randomly generated by satisfying with a set of inequality
constrains. The kernel density estimation (KDE) method is employed to set
the penalty weights for each training samples, so that some negative impacts,
caused by noisy data or outliers, on the resulting learner model can be reduced.
The alternating optimization technique is applied for updating a RSCN model
with improved penalty weights computed from the kernel density estimation
function. Performance evaluation is carried out by a function approximation, four
benchmark datasets and a case study on engineering application. Comparisons to
other robust randomised neural modelling techniques, including the probabilistic
robust learning algorithm for neural networks with random weights and improved
RVFL networks, indicate that the proposed RSCNs with KDE perform favourably
and demonstrate good potential for real-world applications.
1 Introduction
For many real-world applications, sample data collected from various sensors may be contaminated
by some noises or outliers [7], which makes troubles for building neural networks with sound
generalization. Over the past years, robust data modelling techniques have received considerable
attention in the field of applied statistics [7, 8, 11] and machine learning [2, 3, 5, 12, 19]. It is
well known that the cost function plays an important role in robust data modelling. In [2], the
M-estimator and Hampels hyperbolic tangent estimates were employed in the cost function, aiming
to alleviate the negative impacts of outliers on the modelling performance. Under an assumption that
the additive noise of the output follows Cauchy distribution, the mean log squared error was used as
a cost function in [12]. In [5], a robust learning algorithm based on the M-estimator cost function
with random sample consensus was proposed to deal with outliers, and this algorithm has been
successfully applied in computer vision and image processing [15, 20, 22]. Besides these methods
mentioned above, some results on robust data regression using support vector machine (SVM) have
been reported in [3, 19], where SVM-based approaches demonstrate some limits to handle uncertain
data regression problems with higher level outliers.
∗
Corresponding author.
Back-propagation algorithms for training neural networks suffer from many shortcomings, such
as learning parameter setting, slow convergence and local minima. Thus, it is useful to develop
advanced learning techniques for resolving data regression problems, in particular, for stream data
or online data modelling tasks. With such a background, randomized methods for training neural
networks have been developed in the last decades [9, 14, 16]. Readers may refer to a recently
published survey paper for more details about some milestones on this topic [18]. Studies on
the robust data modelling techniques based on Random Vector Functional-link (RVFL) networks
have been reported in [1, 4]. Specifically, a hybrid regularization model with assumption on the
sparsity of outliers was used in training process, and a probabilistic robust learning algorithm for
neural networks with random weights (PRNNRW) was proposed in [1]. However, some learning
parameters used in PRNNRW must be set properly and this is quite difficult to be done in practice. In
[4], an improved version of RVFL networks built by using a KDE-based weighted cost function was
suggested. Unfortunately, the significance of the scope setting of the random weights and biases for
RVFL networks has not been addressed. In [13], we looked into some practical issues and common
pitfalls of RVFL networks, and clearly revealed the impact of the scope setting on the modelling
performance of RVFL networks. Our findings reported in [13] motivates us to further investigate the
robust data regression problem using an advanced randomized learner model, termed as Stochastic
Configuration Networks (SCNs), which are built incrementally by assigning the random weights
and biases with a supervisory mechanism [21].
This paper aims to develop a robust version of SCNs for uncertain data regression. Based on the
construction process of SCNs, we utilise a weighted least squares objective function for evaluating
the output weights of SCNs, and the resulting approximation errors from the present SCN model are
used to incrementally configure the hidden nodes with constrained random parameters. During the
course of building RSCNs, the penalty weights representing the degree of contribution of individual
data samples to the objective function are updated according to a newly constructed KDE function.
In this work, an alternating optimization (AO) technique is employed to implement the RSCN model.
Our proposed algorithm, termed as RSC-KDE, is evaluated by using a function approximation, four
benchmark datasets with different levels of artificial outliers, and an engineering application [4].
Experimental results indicate that the proposed RSC-KDE outperforms other existing methods in
terms of effectiveness and robustness.
The remainder of paper is organized as follows: Section 2 briefly reviews stochastic configuration
networks and the kernel density estimation method. Section 3 details our proposed RSC-KDE
algorithm. Section 4 reports experimental results with comparisons and discussion. Section 5
concludes this paper with some remarks.
2 Revisit of Stochastic Configuration Networks
This section reviews our proposed SCN framework, in which the input weights and biases are
randomly assigned in the light of a supervisory mechanism, and the output weights are evaluated by
solving a linear least squares problem. More details about SCNs can be read in [21].
Let Γ := {g1 , g2 , g3 ...} be a set of real-valued functions, span(Γ) denote a function space spanned by
Γ; L2 (D) denote the space of all Lebesgue measurable functions f = [f1 , f2 , . . . , fm ] : Rd → Rm
defined on D ⊂ Rd , with the L2 norm defined as
!1/2
m Z
X
2
kf k :=
|fq (x)| dx
< ∞.
(1)
q=1
D
The inner product of φ = [φ1 , φ2 , . . . , φm ] : Rd → Rm and f is defined as
m
m Z
X
X
fq (x)φq (x)dx.
hf, φi :=
hfq , φq i =
q=1
q=1
(2)
D
In the special case that m = 1,R for a real-valued function ψ : Rd → R defined on D ⊂ Rd ,
its L2 normRbecomes kψk := ( D |ψ(x)|2 dx)1/2 , while the inner product of ψ1 and ψ2 becomes
hψ1 , ψ2 i = D ψ1 (x)ψ2 (x)dx.
Given a target function f : Rd → Rm , suppose that we have already built a single layer
PL−1
feed-forward network (SLFN) with L − 1 hidden nodes, i.e, fL−1 (x) = j=1 βj gj (wjT x + bj )
2
(L = 1, 2, . . ., f0 = 0), βj = [βj,1 , . . . , βj,m ]T , and the current residual error, denoted as
eL−1 = f − fL−1 = [eL−1,1 , . . . , eL−1,m], does not reach an acceptable tolerance level. The
framework of SCNs provides an effective solution for how to add βL , gL (wL and bL ) leading
to fL = fL−1 + βL gL until the residual error eL = f − fL falls into an expected tolerance ǫ.
The following Theorem 1 restates the universal approximation property of SCNs (corresponding to
Theorem 7 in [21]).
Theorem 1. Suppose that span(Γ) is dense in L2 space and ∀g ∈ Γ, 0 < kgk < bg for some
bg ∈ R+ . Given 0 < r < 1 and a nonnegative real number sequence {µL } with limL→+∞ µL = 0
and µL ≤ (1 − r). For L = 1, 2, . . ., denoted by
δL =
m
X
q=1
δL,q , δL,q = (1 − r − µL )keL−1,q k2 , q = 1, 2, . . . , m.
(3)
If the random basis function gL is generated with the following constraints:
heL−1,q , gL i2 ≥ b2g δL,q , q = 1, 2, . . . , m,
(4)
and the output weights are evaluated by
[β1∗ , β2∗ , . . . , βL∗ ] = arg min kf −
β
Then, limL→+∞ kf − fL k = 0, where fL =
PL
j=1
L
X
j=1
βj gj k.
(5)
∗
∗
βj∗ gj , βj∗ = [βj,1
, . . . , βj,m
]T .
The construction process of SCNs starts with a small sized network, then incrementally adds hidden
nodes followed by computing the output weights. This procedure keeps going on until the model
meets some certain termination criterion. Some remarkable merits of our SCNs can be summarized
as follows : (i) There is no requirement on any prior knowledge about the architecture of the
constructed network for a given task; (ii) The scope of the random weights and biases is adjustable
and automatically determined by the data rather than a fixed setting from end-users; and (iii) The
input weights and biases are randomly assigned with an inequality constraint, which can guarantee
the universal approximation property.
Remark 1: In the past two decades, randomized methods for training neural networks suffer from
some misunderstandings on the constraint of random assignment (i.e., randomly assign the input
weights and biases in a fixed interval, even any intervals, or no specification at all). Unfortunately,
in many published works the authors blindly and wrongly use the randomness in constructing
randomized learner models and mindlessly made some misleading statements without any scientific
justification [13]. It should be pointed out that the universal approximation theorems for RVFL
networks established in [9] are fundamental and significant to the randomized learning theory.
However, these theoretical results cannot provide us with practical and useful guidance on structure
and learning parameter settings, and algorithm design and implementation aspects. Our proposed
SCN framework firstly touches the base of randomized learning techniques, and draws researches
on this topic into right tracks through proper uses of constrained random parameters. Indeed,
the inequalities (4) proposed originally in [21] are essential to ensure the universal approximation
property. From algorithm implementation perspectives, a scheme to prevent SCNs from over-fitting
must in place. As usually done in machine learning, the performance over a validation set can be
used to terminate the learning process.
At the end of this section, we briefly introduce a kernel density estimation method for weighting
contributions of each training data in the learning process. Basically, a kernel density estimator
computes a smooth density estimation from data samples by placing on each sample point a function
representing its contribution to the density. Readers can refer to [10] and [17] for more details on
the kernel density estimation (KDE) method.
Based on KDE method, the underlying probability density function of a random variable η can be
estimated by
N
X
ρk K(η − ηk ),
(6)
Φ(η) =
k=1
3
where K represents a kernel function (typically a Gaussian function) centered at the data points
ηk , and ρk are weighting coefficients (uniform weights are commonly used, i.e., ρk = 1/N , k =
1, 2, . . . , N ).
3 Robust Stochastic Configuration Networks
Robust data regression seeks for a capable learner model that can successfully learn a true
distribution from uncertain data samples. This is very important for industrial applications, where
the collected data samples are always contaminated by outliers caused by the failure of measuring
or transmission devices or unusual disturbances. This section details the development of robust
stochastic configuration networks (RSCNs). For a target function f : Rd → Rm , given a
training dataset with inputs X = {x1 , x2 , . . . , xN }, xi = [xi,1 , . . . , xi,d ]T ∈ Rd and outputs
T = {t1 , t2 , . . . , tN }, where ti = [ti,1 , . . . , ti,m ]T ∈ Rm , i = 1, . . . , N , a RSCN model can
be built by solving a weighted least squares (WLS) problem, that is,
min
β,θ
N
X
i=1
θi k
L
X
j=1
βj g(wj , bj , xi ) − ti k2 ,
(7)
where θi ≥ 0 (i = 1, 2, . . . , N ) is the i-th penalty weight, representing the contribution of the
P
corresponding sample to the objective function (7). GL (x) = L
j=1 βj g(wj , bj , x) is a SCN, in
which g is the activation function and L is the number of hidden nodes, wj , bj are the input weights
and biases that are stochastically configured according to Theorem 1, and βj represents the output
weights.
Generally speaking, the penalty weights θi (i = 1, 2, . . . , N ) can be determined according to the
reliability of the sample xi . It is easy to understand that a higher reliability means more trust in the
data that correctly represents the process behavior, and a lower reliability indicates less confidence
on the sample that may be an outlier or noisy one. Thus, decreasing (increasing) the penalty weights
of training samples with lower (higher) reliability can eliminate or even remove negative impacts on
the learner model building.
A logical thinking to combine the original SCN framework with the WLS-based learning is
to use a weighted version of the model’s residual error in the process of building SCNs. In
other words, a RSCN model can be incrementally built by stochastically configuring the hidden
parameters based on a redefined constraint (4), and evaluating the output weights by using the
WLS solution of (7). Given training samples X = {x1 , x2 , . . . , xN }, xi = [xi,1 , . . . , xi,d ]T ∈
Rd , denoted by eL−1 (X) = [eL−1,1 (X), eL−1,2 (X), . . . , eL−1,m(X)]T ∈ RN ×m , where
T
eL−1,q (X) = [eL−1,q (x1 ), . . . , eL−1,q (xN )] ∈ RN , q = 1, 2, . . . , m. Let hL (X) = [gL (wL
x1 +
T
T
bL ), . . . , gL (wL xN + bL )] be the output vector of the new hidden node for each input xi ,
i = 1, 2, . . . , N . Then, we can obtain the current hidden layer output matrix, HL = [h1 , h2 , . . . , hL ].
According to (7), a weighted form of eL−1 (X) can be defined if the penalty weights are
available during the constructive process of SCNs. Denoted the weighted eL−1 (X) and hL (X)
by ẽL−1 (X) = [ẽL−1,1 (X), ẽL−1,2 (X), . . . , ẽL−1,m (X)]T = ΘeL−1 (X), and h̃L (X) = ΘhL (X),
√ √
√
P
respectively, where Θ = diag{ θ1 , θ2 , . . . , θN }. Let ξ˜L = m
q=1 ξ̃L,q and ξ̃L,q be defined as
2
ẽT
L−1,q (X) · h̃L (X)
− (1 − r − µL )ẽT
(8)
ξ˜L,q =
L−1,q (X)ẽL−1,q (X).
h̃T
(X)
·
h̃
(X)
L
L
Based on Theorem 1, the hidden parameters (w and b) can be stochastically configured by choosing
a maximum ξ˜L among multiple tests, subjected to ξ̃L,q ≥ 0, q = 1, 2, . . . , m.
Now, the remaining question is how to assign penalty weights θ1 , θ2 , . . . , θN along with the process
of building SCNs. Recall that if the probability density function of the residuals can be obtained
or estimated, the reliabilities of the samples will be determined. Inspired by the work in [25], we
construct a probability density function of the residual error eL as follows (here eL is regarded as a
random variable)
N
1 X
keL − eL (xk )k
Φ(eL ) =
,
(9)
K
τN
τ
k=1
4
where eL (xk ) = [eL−1,1 (xk ), . . . , eL−1,m (xk )]T ∈ Rm , τ = 1.06σ̂N −1/5 is an estimation window
width, σ̂ is the standard deviation of the residual errors, K is a Gaussian function defined by
1
t2
K(t) = √ exp (− ).
2
2π
(10)
With these preparation, the probability of each residual error eL (xi ) (i = 1, 2, . . . , N ) can be
obtained by calculating Φ(eL (xi )). Then, the penalty weights θi (i = 1, 2, . . . , N ) can be assigned
as
N
1 X
keL (xi ) − eL (xk )k
θi = Φ(eL (xi )) =
.
(11)
K
τN
τ
k=1
With these penalty weights, the output weights β ∗ = [β1∗ , β2∗ , . . . , βL∗ ] can be evaluated by solving
the following WLS problem:
β∗
=
arg min (HL β − T )T Λ(HL β − T )
=
(HLT ΛHL )† HLT ΛT,
β
(12)
2
where β = [β1 , β2 , . . . , βL ], HL = [h1 , h2 , . . . , hL ], Λ = Θ = diag{θ1 , θ2 , . . . , θN }.
In this paper, an alternating optimization (AO) strategy is applied for implementing RSCNs, which
includes the process of building a SCN model with a set of suitable penalty weights that control the
contribution of contaminated samples. The whole procedure begins with assigning equal penalty
weights for all samples (i.e., θi = 1, and Λ is an identity matrix) and building the SCN model,
followed by updating these penalty weights according to (11), then repeating these two steps
alternatively until some certain stopping criterion is reached. It should be clarified that the penalty
weights are updated only when a round of the process of building the SCN model is completed.
Specifically, the penalty weights θi (i = 1, 2, . . . , N ) and the output weights β can be calculated
iteratively by applying the AO procedure, that is,
!
N
(ν)
(ν)
eL (xi ) − eL (xk )
1 X
(ν+1)
=
θi
(13)
K
τN
τ
k=1
and
β (ν+1) = (HLT Λ(ν+1) HL )† HLT Λ(ν+1) T,
(14)
(ν+1)
where ν denotes the ν-th iteration of the alternating optimization process, and Λ
=
(ν+1) (ν+1)
(ν+1)
(ν)
diag{θ1
, θ2
, . . . , θN
}. Here, we use eL (xi ) to represent the residual error value for
(ν)
xi with θi used as the present penalty weights in the RSCN model.
Distinguished from the original SCN framework that all training samples contribute equally to the
objective function, our newly developed RSCNs treat individual samples differently and put more
emphasis on data samples with higher reliability, which indeed corresponds to lager values of the
penalty weights. That means if a training output yj is corrupted by outliers or noises, the sample
pair (xj , yj ) will provide less contribution to the cost function due to the relatively small value of its
corresponding penalty weight.
Remark 2: An important issue in design of RSCNs is about the termination criterion. As mentioned
in Remark 1, the performance over a validation data set can be employed as a stopping condition.
Unfortunately, this method does not make sense and cannot be applied for building RSCNs due to
the presence of uncertainties in the validation data. However, the validation testing criterion can still
be used if a clean data set extracted from the true data distribution is available. In this work, we
assume that a clean validation data set is ready to be used for this purpose. Our proposed RSC-KDE
algorithm is summarized in the following pseudo code.
4 Performance Evaluation
This section reports some simulation results on a function approximation, four benchmark datasets
from KEEL2, and an industrial application [4]. The proposed RSC-KDE algorithm is compared to
2
KEEL: http://www.keel.es/
5
RSC-KDE Algorithm
Given inputs X = {x1 , x2 , . . . , xN }, xi ∈ Rd , outputs T = {t1 , t2 , . . . , tN }, ti ∈ Rm ; Set Lmax as
the maximum number of hidden nodes, ǫ as the expected error tolerance, Pmax as the maximum times
of random configuration, Imax as the maximum number of alternating optimization; Choose a set of
positive scalars Υ = {λmin : ∆λ : λmax };
1. Initialize e0 := [t1 , . . . , tN ]T , 0 < r < 1, θi = 1, Λ = diag{θ1 , θ2 , . . . , θN }, Ω, W := [ ];
2. While ν ≤ Imax AND ke0 kF > ǫ, Do
3. While L ≤ Lmax AND ke0 kF > ǫ, Do
4.
For λ ∈ Υ, Do
5.
For k = 1, 2 . . . , Pmax , Do
6.
Randomly assign ωL and bL from [−λ, λ]d and [−λ, λ], respectively;
7.
Calculate ẽL−1 , h̃L , ξ˜L,q by Eq. (8), set µL = (1 − r)/(L + 1);
8.
If min{ξ˜L,1 , ξ̃L,2 , ..., ξ˜L,m } ≥ 0
9.
Save wL and bL in W , ξ̃L in Ω, respectively;
10.
Else Go back to Step 5
11.
End If
12.
End For (corresponds to Step 5)
13.
If W is not empty
∗
14.
Find wL
, b∗L maximizing ξ̃L in Ω, and set HL = [h∗1 , h∗2 , . . . , h∗L ];
15.
Break (go to Step 19);
16.
Else Randomly take τ ∈ (0, 1 − r), renew r := r + τ , return to Step 5;
17.
End If
18.
End For (corresponds to Step 4)
∗
19.
Calculate β ∗ = [β1∗ , β2∗ , . . . , βL
] based on Eq. (14);
∗
20.
Calculate eL = HL β − T and obtain ẽL ;
21.
Renew e0 := ẽL , L := L + 1;
22. End While
23. Update θi by Eq. (13), renew Λ = diag{θ1 , θ2 , . . . , θN } and ν := ν + 1;
24. End While
∗
∗
], ω ∗ = [ω1∗ , ω2∗ , . . . , ωL
], b∗ =
25. Return Λ = diag{θ1 , . . . , θN }, β ∗ = [β1∗ , β2∗ , . . . , βL
[b∗1 , b∗2 , . . . , b∗L ].
other three randomized algorithms: RVFL [9], improved RVFL [4], and the probabilistic learning
algorithm PRNNRW [1]. All comparisons are conducted under several scenarios with different
system settings on learning parameters and noise levels. The Root Mean Squared Error (RMSE) is
used to evaluate the generalization capability of each algorithm over the outlier-free test datasets. In
addition, a robustness analysis on the setting of ν and Lmax is given for the case study.
The input and output values are normalized into [0,1] before artificially adding certain level of
outliers. The maximum times of random configuration Tmax in RSC-KDE is set as 100, and the
sigmoidal activation function g(x) = 1/(1 + exp(−x)) is used in all simulations.
4.1 Function Approximation
Consider the following function approximation problem [6]:
2
2
2
y = 0.2e−(10x−4) + 0.5e−(80x−40) + 0.3e−(80x−20) , x ∈ [−1, 1].
The training dataset contains 600 points which are randomly generated from the uniform distribution
[-1,1]. The test dataset, of size 600, is generated from a regularly spaced grid on [-1,1]. We purposely
introduce outliers into the training dataset: A variable percentage ξ of the data points is selected
randomly and their corresponding function values (y) are substituted by background noises with
values uniformly distributed over [-0.2, 0.8]. To show the advantage of RSC-KDE in uncertain data
modelling, we make comparisons on the performance among these four algorithms at each outlier
percentage, i.e., ζ = {0%, 5%, 10%, 15%, 20%, 25%, 30%}.
Figure 1 (a) depicts the training samples with the outlier percentage at ξ = 25%, Figure 1
(b) shows the learners’ performance on the test dataset from the four algorithms, in which the
proposed RSC-KDE exhibits the best performance compared with the other three methods. For
RVFL, Improved RVFL and PRNNRW, we examined different scope settings for the random
6
0.6
0.5
0.6
Training Samples
Target Function
0.5
0.4
0.4
Test Samples
Improved RVFL
PRNNRW
RVFL
RSC-KDE
0.3
Y
y
0.3
0.2
0.2
0.1
0.1
0
0
-0.1
-0.2
-1
-0.5
0
x
0.5
1
(a)
-0.1
-1
-0.5
0
x
0.5
1
(b)
Figure 1: (a) 600 training samples used for function approximation at ζ = 25%, along with target
function shown in red line; (b) Approximation performance on the test dataset by four learning
algorithms at ζ = 25%.
parameters, i.e., w, b ∈ [−λ, λ], λ = 1, 10, 30, 50, 100, 150, to demonstrate its significant impact
on the randomized learners’ performance. For each λ, different network architectures (e.g. L =
40, 60, 100, 120, 150, 200) are used to find the pair (λ, L) leading to the most favorable performance.
For λ = 1, 30, 50, 100, we demonstrate the test results of the four algorithms in Figure 2, where the
average errors and standard deviations of RMSE (based on 100 trials) are plotted for each outlier
percentage. It is clear that our proposed RSC-KDE algorithm outperforms the other methods for
each case. In particular, for λ = 1, the approximation performance of all other three algorithms
are far worse than an acceptable level. Obviously, if the scope setting is improper, the randomized
learner models can not be expected to perform at all, as reported in [13, 21]. As the outlier percentage
becomes very low, RSCN outperforms other randomized learner models, which aligns well with the
consequence reported in [21]. Given a relatively higher outlier percentage, for example ζ = 30%,
RSC-KDE produces a promising result with RMSE of 0.0138 ± 0.0027, that is much better than
the other three algorithms. More results for λ = 1, 30, 50, 100 at ζ = 10%, 15%, 20%, 25% are
reported in Table 1.
Table 1: Performance comparisons on the function approximation
Scope Setting
λ=1
λ = 30
λ = 50
λ = 150
Algorithm
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
Test Performance at Different Outlier Percentage (MEAN, STD)
10%
15%
20%
25%
0.0463,2.0e-05 0.0464,2.0e-05 0.0469,2.0e-05 0.0478,2.0e-05
0.0556,5.1e-05 0.0555,3.0e-05 0.0560,4.0e-05 0.0563,3.3e-05
0.0781,6.6e-06 0.0781,7.3e-06 0.0781,7.3e-06 0.0781,6.8e-06
0.0090,0.0016
0.0098,0.0022
0.0104,0.0035
0.0117,0.0022
0.0225,0.0037
0.0249,0.0039
0.0264,0.0033
0.0326,0.0032
0.0344,0.0030
0.0326,0.0036
0.0319,0.0034
0.0322,0.0028
0.0491,0.0029
0.0490,0.0013
0.0490,0.0017
0.0492,0.0010
0.0090,0.0016
0.0098,0.0022
0.0104,0.0035
0.0117,0.0022
0.0248,0.0042
0.0274,0.0079
0.0288,0.0062
0.0343,0.0065
0.0247,0.0056
0.0229,0.0047
0.0218,0.0041
0.0240,0.0038
0.0498,0.0014
0.0497,0.0020
0.0496,0.0024
0.0494,0.0015
0.0090,0.0016
0.0098,0.0022
0.0104,0.0035
0.0117,0.0022
0.0345,0.0166
0.0353,0.0062
0.0380,0.0099
0.0386,0.0086
0.0229,0.0082
0.0214,0.0074
0.0228,0.0084
0.0284,0.0076
0.0445,0.0031
0.0439,0.0032
0.0438,0.0030
0.0441,0.0037
0.0090,0.0016
0.0098,0.0022
0.0104,0.0035
0.0117,0.0022
4.2 Benchmark Datasets
Table 2 gives some statistics on the four datasets used in our simulations. We randomly chose
75% of the whole samples as the training dataset while take the remainders as the test dataset. A
7
0.08
0.055
Improved RVFL
PRNNRW
RVFL
RSC-KDE
0.07
0.05
0.045
0.06
0.04
0.05
0.035
0.04
0.03
0.025
0.03
0.02
0.02
0.015
0.01
0.01
0
0.005
0
5%
10%
15%
20%
25%
30%
0
5%
10%
(a) λ = 1
15%
20%
25%
30%
20%
25%
30%
(b) λ = 30
0.06
0.055
0.05
0.05
0.045
0.04
0.04
0.035
0.03
0.03
0.025
0.02
0.02
0.015
0.01
0.01
0.005
0
0
5%
10%
15%
20%
25%
30%
0
5%
10%
(c) λ = 50
15%
(d) λ = 100
Figure 2: Test RMSE comparison on the function approximation among the four algorithms with
different outlier percentage ζ. λ = 1, 10, 50, 100 and L = 100 are used in RVFL, Improved RVFL
and PRNNRW.
similar strategy as done in the function approximation problem is applied to introduce different
percentages of outliers into the training dataset. That is, for each normalized training dataset, a
variable percentage ζ of the data points are selected at random and the associated output values are
substituted with background noises that are uniformly distributed on the range [-0.5,0.5]. Finally,
the contaminated output values are distributed over [-0.5,1.5] instead of [0,1], while the test dataset
is outlier-free for the assessment purpose.
Table 2: Statistics of the benchmark datasets
No.
Name
Instances
Features
1
2
3
4
stock
laser
concrete
treasury
950
993
1030
1049
9
4
8
15
For each benchmark dataset, we evaluate the performance of RVFL, Improved RVFL and
PRNNRW with different settings of λ and L, for example, λ = {0.1, 0.5, 1, 3, 5} and L =
{30, 50, 100, 150, 200}, respectively. We conduct 50 independent trials for each case (λ and
L) and calculate their mean values and standard deviations of RMSE at different percentage of
outliers. In Figure 3, we plot the comparison results for RVFL, Improved RVFL, and PRNNRW
with λ = 1, and it shows that our proposed RSC-KDE algorithm outperforms the others at each
outlier percentage. From the results reported in Table 3, it is observed that our proposed RSC-KDE
algorithm outperform the others for all these four datasets at each outlier percentage, despite that the
results obtained by other three algorithms are the ‘best’ ones selected from all results with various
8
settings of λ and L. Specifically in Figure 3, PRNNRW (with λ = 1) exhibits the worst accuracy on
stock, laser, concrete, but performs better than RVFL and Improved RVFL on treasury. Also in Table
3, the results from PRNNRW, as obtained by the most appropriate combination of λ and L from the
set {0.1, 0.5, 1, 3, 5} and {30, 50, 100, 150, 200}, respectively, are much better than that shown in
Figure 3. Indeed, for RVFL, Improved RVFL and PRNNRW, a common practice to determine a
suitable scope setting and a reasonable network architecture is based on the trial-and-error method,
while the proposed RSC-KDE works robustly with much less human intervention on the parameter
setting.
0.09
0.08
0.08
0.07
0.07
0.06
0.06
0.05
0.05
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0
5%
10%
15%
20%
25%
30%
0
5%
10%
(a) stock
15%
20%
25%
30%
20%
25%
30%
(b) laser
0.14
0.04
0.13
0.035
0.12
Improved RVFL
PRNNRW
RVFL
RSC-KDE
0.03
0.11
0.025
0.1
0.02
0.09
0.015
0.08
0.01
0.07
0.06
0.005
0
5%
10%
15%
20%
25%
30%
(c) concrete
0
5%
10%
15%
(d) treasury
Figure 3: Test RMSE comparison on four benchmark datasets among four algorithms with different
outlier percentage ζ. λ = 1 and L = 150 are used in RVFL, Improved RVFL, and PRNNRW.
4.3 Particle Size Estimation of Mineral Grinding Process: A Case Study
In this section, we make a further investigation on the merits of our proposed RSC-KDE algorithm
by using a dataset from process industry [4]. Figure 4 depicts the grinding process where the
coarse fresh ore OF is fed into the ball mill through the conveyor. Meanwhile, a certain amount
of mill water QF is added through a pipe to maintain a proper pulp density. At this stage, the
steel balls within the mill crush the coarse ore to a finer size alone with the knocking and tumbling
actions. After grinding, the mixed ore pulp that includes both coarser and finer particles is discharged
continuously from the mill into the spiral selector for further classification with assistance of dilution
water QD mixed to the ore pulp. Next, the pulp is separated into the overflow and underflow pulp.
Finally, the underflow pulp with coarser particles is recycled back to the mill for re-grinding, whilst
the overflow pulp with finer particles (as product) is further proceeded. As can be seen that the
particle size estimation plays an important role in this circling procedure.
Particle size estimation of mineral grinding process can be formulated as a regression problem, with
three input variables including the fresh ore feed rate OF , the mill water flow rate QF , and the
dilution water flow rate QD , and with a single output of the unmodelled dynamics, namely △r.
Denoted by x = [OF , QF , QD ]T and y as the input vector and the output (refers to the estimation
of the unmodelled dynamics △r̃), respectively. Let r̃ represent an estimate of the particle size from
9
Table 3: Performance comparisons on the benchmark datasets
Dataset
stock
laser
concrete
treasury
Algorithm
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
Test Performance at Different Outlier Percentage (MEAN,STD)
10%
15%
20%
25%
0.0495,0.0024 0.0531,0.0027 0.0554,0.0032 0.0590,0.0036
0.0373,0.0023 0.0404,0.0022 0.0425,0.0014 0.0456,0.0023
0.0378,0.0014 0.0387,0.0014 0.0388,0.0011 0.0392,0.0017
0.0317,0.0014 0.0322,0.0016 0.0328,0.0012 0.0342,0.0012
0.0318,0.0033 0.0323,0.0029 0.0343,0.0030 0.0359,0.0038
0.0239,0.0023 0.0260,0.0024 0.0264,0.0021 0.0277,0.0039
0.0424,0.0022 0.0424,0.0024 0.0421,0.0026 0.0428,0.0027
0.0161,0.0013 0.0202,0.0020 0.0195,0.0027 0.0233,0.0014
0.0975,0.0039 0.1038,0.0047 0.1068,0.0047 0.1092,0.0037
0.0903,0.0070 0.0967,0.0064 0.0991,0.0043 0.1022,0.0032
0.1019,0.0029 0.1008,0.0030 0.1013,0.0025 0.1034,0.0031
0.0749,0.0016 0.0773,0.0014 0.0805,0.0026 0.0812,0.0025
0.0231,0.0012 0.0253,0.0011 0.0265,0.0013 0.0325,0.0016
0.0135,0.0005 0.0145,0.0004 0.0145,0.0005 0.0166,0.0005
0.0130,0.0004 0.0130,0.0004 0.0131,0.0002 0.0131,0.0002
0.0084,0.0002 0.0086,0.0002 0.0093,0.0003 0.0109,0.0004
a mathematical (mechanism) model. Thus, the final estimated value of the particle size r̃ can be
evaluated by r̃ := r̃ + △r̃.
Figure 4: Flow chart of mineral grinding process [4].
In this case study, 300 training samples and 300 test samples were collected from a
hardware-in-the-loop (HIL) platform [4], which is composed of the following five subsystems: an
optimal setting control subsystem, a human supervision subsystem, a DCS control subsystem, a
virtual actuator and sensors subsystem, and a virtual operation process subsystem. For detailed
descriptions on the operational functionalities of these subsystems, readers can refer to [4]. Both the
input and output values are normalized into [0,1]. Then, different levels of outliers are added into
the normalized training dataset in the similar way as done in the previous simulations, i.e, a variable
percentage ζ of data points are selected randomly and the corresponding output values are corrupted
by background noises followed the uniform distribution [-0.5,0.5]. As a result, the output values are
distributed in the range [-0.5,1.5], while the test samples are outlier-free.
The performance of these four algorithms are evaluated at several outlier percentages, i.e., ζ =
{0%, 5%, 10%, 15%, 20%, 25%, 30%}. Specifically, for RVFL, Improved RVFL, and PRNNRW,
different settings of λ and L are used in this study. Each comparison is based on 50 independent
trials, the mean value and standard deviation of RMSE are recorded at each outlier percentage. The
10
test performance of the four algorithms are depicted in Figure 5, where the presented results for
RVFL, Improved RVFL and PRNNRW with each scope setting (λ = 0.1, 0.5, 1, 5) correspond to
the ‘best’ records among the trails using different L (L = 10, 30, 50, 100, 150). It can be easily
found that our proposed RSC-KDE algorithm outperforms the other three methods in most cases.
Although the Improved RVFLN exhibits very close performance when the outlier percentage is
relatively lower, RSC-KDE algorithm has demonstrated the best at robustness even at high outlier
contamination rate. When λ = 0.5 and λ = 1, the performance of PRNNRW has been improved
a lot compared with λ = 0.1, but are still unacceptable in comparison with Improved RVFL and
RSC-KDE. For both RVFL and Improved RVFLN, there is no remarkable difference between the
results with λ = 0.5 and λ = 1. Interestingly, Figure 5 (d) shows that results from PRNNRW
(with λ = 5) at relatively higher outlier percentages (e.g. ζ = 20%, 25%, 30%) are slightly
better than RSC-KDE. However, this result needs a suitable scope setting (i.e. λ = 5), which is
time-consuming in practice. In contrast, the proposed RSC-KDE can lead to good performance than
the others without any user-oriented trials for parameter setting. Table 4 summarizes the records
of our proposed RSC-KDE algorithm as well as the ‘best’ results obtained from RVFL, Improved
RVFL and PRNNRW with the ‘most appropriate’ parameter setting on L for each λ. The impact
of the scope setting for the other three randomized algorithms can be seen through comparing their
records at each outlier percentage. Specifically for λ = 1, the test results from the four algorithms
are shown in Figure 6, where both the real and estimated values of the normalized particle size (of
the whole test samples) are plotted.
0.16
0.04
0.14
0.035
0.12
0.03
0.1
0.025
0.08
0.02
0.06
0.015
0.04
0.01
0.02
0.005
0
Improved RVFL
PRNNRW
RVFL
RSC-KDE
0
0
5%
10%
15%
20%
25%
30%
0
5%
(a) λ = 0.1
10%
15%
20%
25%
30%
20%
25%
30%
(b) λ = 0.5
0.04
0.07
0.035
0.06
0.03
0.05
0.025
0.04
0.02
0.03
0.015
0.02
0.01
0.01
0.005
0
0
0
5%
10%
15%
20%
25%
30%
(c) λ = 1
0
5%
10%
15%
(d) λ = 5
Figure 5: Test RMSE comparisons on the case study between four algorithms with different outlier
percentage ζ. λ = 0.1, 0.5, 1, 5 and L = 50 are used in RVFL, Improved RVFL, and PRNNRW.
Before ending up this work, we conduct a robustness analysis on the key parameters (L and ν) to
investigate their impacts on the performance of our proposed RSC-KDE algorithm. The test results
for different combination of L and ν are reported in Table 5 with ζ = 0%, ζ = 10% and ζ = 30%,
respectively. For ζ = 0%, there is no much difference among the results with different ν for each
setting of L, implying that the AO process is not necessary. In this case, RSC-KDE is identical
to the original SC algorithm (SC-III in [21]). When ζ = 10%, the most appropriate setting of the
11
3.5
1.2
Real
Estimation
Real
Estimation
3
2.5
Normalized Particle Size
Normalized Particle Size
1
2
1.5
1
0.5
0.8
0.6
0.4
0.2
0
-0.5
0
0
50
100
150
200
Sample Sequence Number
250
300
0
50
(a) RVFL
100
150
200
Sample Sequence Number
250
300
250
300
(b) Improved RVFL
1.2
1
Real
Estimation
Real
Estimation
1
Normalized Particle Size
Normalized Particle Size
0.8
0.8
0.6
0.4
0.2
0.6
0.4
0.2
0
-0.2
0
0
50
100
150
200
Sample Sequence Number
250
300
(c) PRNNRW
0
50
100
150
200
Sample Sequence Number
(d) RSC-KDE
Figure 6: Test results of the four algorithms at ζ = 30% on the case study. λ = 1 and L = 50 are
used in RVFL (a), Improved RVFL (b), and PRNNRW (c).
Table 4: Performance comparisons on the case study
Scope Setting
λ = 0.1
λ = 0.5
λ=1
λ=5
Algorithm
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
RVFL
Improved RVFL
PRNNRW
RSC-KDE
Test Performance at Different Outlier Percentage (MEAN,STD)
10%
15%
20%
25%
0.0213,0.0027 0.0279,0.0025 0.0274,0.0025 0.0288,0.0030
0.0018,0.0003 0.0033,0.0003 0.0066,0.0006 0.0097,0.0010
0.1437,0.0063 0.1451,0.0053 0.1469,0.0053 0.1465,0.0048
0.0006,0.0003 0.0020,0.0007 0.0026,0.0006 0.0028,0.0004
0.0217,0.0026 0.0278,0.0024 0.0274,0.0022 0.0295,0.0035
0.0023,0.0003 0.0064,0.0011 0.0078,0.0010 0.0103,0.0032
0.0238,0.0005 0.0238,0.0005 0.0234,0.0006 0.0230,0.0005
0.0006,0.0003 0.0020,0.0007 0.0026,0.0006 0.0028,0.0004
0.0213,0.0024 0.0275,0.0020 0.0269,0.0016 0.0292,0.0029
0.0025,0.0003 0.0069,0.0013 0.0085,0.0035 0.0106,0.0044
0.0145,0.0014 0.0153,0.0014 0.0150,0.0015 0.0148,0.0014
0.0006,0.0003 0.0020,0.0007 0.0026,0.0006 0.0028,0.0004
0.0305,0.0039 0.0463,0.0126 0.0429,0.0100 0.0468,0.0144
0.0050,0.0021 0.0100,0.0049 0.0123,0.0043 0.0169,0.0087
0.0017,0.0007 0.0021,0.0009 0.0023,0.0009 0.0022,0.0007
0.0006,0.0003 0.0020,0.0007 0.0026,0.0006 0.0028,0.0004
architecture is L = 20 while the iteration times in AO can be selected as ν = 3, 5, 8, 10, 12. At this
percentage of outliers, it is fair to say that the accuracy of RSC-KDE with L = 20 is preferable and
stay within a stable level (i.e. RMSE is around 0.0010) provided that ν is set equal or larger than
3. Similar to the case of ζ = 30%, the most appropriate setting of the architecture is L = 10 while
the value of ν can be selected from the set {5, 8, 10, 12}. All these records suggest that RSC-KDE
performs robustly for uncertain data modelling with smaller iteration times in AO (between 5 and 12).
These empirical results offer us some information on the setting of ν, although it is data dependent.
12
Table 5: Robustness analysis of ν and L with different outlier percentage ζ on the case study
Outlier Percentage
ζ = 0%
ζ = 10%
ζ = 30%
Number of AO
ν=2
ν=3
ν=5
ν=8
ν = 10
ν = 12
ν=2
ν=3
ν=5
ν=8
ν = 10
ν = 12
ν=2
ν=3
ν=5
ν=8
ν = 10
ν = 12
Test Performance with Different Setting of L (Mean RMSE)
L = 10 L = 20 L = 30 L = 50 L = 60 L = 80
0.0025
0.0003
0.0003
0.0002
0.0002
0.0002
0.0025
0.0003
0.0003
0.0002
0.0002
0.0002
0.0025
0.0003
0.0003
0.0002
0.0002
0.0002
0.0026
0.0003
0.0003
0.0002
0.0002
0.0002
0.0026
0.0003
0.0003
0.0002
0.0002
0.0002
0.0028
0.0003
0.0003
0.0002
0.0002
0.0002
0.0026
0.0030
0.0072
0.0216
0.0334
0.0501
0.0024
0.0010
0.0024
0.0100
0.0200
0.0444
0.0024
0.0009
0.0012
0.0076
0.0128
0.0290
0.0022
0.0010
0.0012
0.0055
0.0091
0.0228
0.0023
0.0010
0.0012
0.0060
0.0094
0.0194
0.0025
0.0009
0.0012
0.0057
0.0088
0.0244
0.0101
0.0173
0.0544
0.0950
0.1315
0.2322
0.0051
0.0081
0.0347
0.0932
0.1493
0.2057
0.0040
0.0056
0.0161
0.0966
0.1574
0.2182
0.0039
0.0051
0.0102
0.0870
0.1604
0.2326
0.0038
0.0053
0.0095
0.1017
0.1688
0.2366
0.0041
0.0050
0.0099
0.0959
0.1621
0.2308
5 Concluding Remarks
Uncertain data modelling problems appear in many real-world applications, it is significant to
develop advanced machine learning techniques to achieve better modelling performance. This paper
proposes a robust version of stochastic configuration networks for problem solving. Empirical
results reported in this work clearly indicate that the proposed RSCNs, as one of the extensions of
our recently developed SCN framework in [21], have great potential in dealing with robust data
regression problems. From the implementation perspective, our design methodology needs an
assumption on the availability of a clean validation dataset, which helps to prevent the learner from
over-fitting during the course of incrementally constructing stochastic configuration networks. In
practice, however, such a hypothesis is not always applicable. Thus, further research on this topic is
necessary. A plenty of explorations are being expected, such as the use of various cost functions for
evaluating the output weights, development of online version of RSCNs, and distributed RSCNs for
large-scale data modelling.
Acknowledgment
The authors would like to thank Dr Wei Dai from China University of Mining Technology for his
sharing the data used in the case study.
References
[1] F. Cao, H. Ye, D. Wang, A probabilistic learning algorithm for robust modeling using neural networks with
random weights, Information Sciences 313 (2015) 62–78.
[2] D. S. Chen, R. C. Jain, A robust backpropagation learning algorithm for function approximation, IEEE
Transactions on Neural Networks, 5 (3) (1994) 467–479.
[3] C. C. Chuang, S. F. Su, J. Tsong, C. C. Hsiao, Robust support vector regression networks for function
approximation with outliers, IEEE Transactions on Neural Networks, 13 (6) (2002) 1322–1330.
[4] W. Dai, Q. Liu, T. Y. Chai, Particle size estimate of grinding processes using random vector functional link
networks with improved robustness, Neurocomputing, 169 (2015) 361–372.
[5] M. T. El-Melegy, Random sampler M-estimator algorithm with sequential probability ratio test for robust
function approximation via feed-forward neural networks, IEEE Transactions on Neural Networks and
Learning Systems, 24 (7) (2013) 1074–1085.
13
[6] A. N. Gorban, I. Tyukin, D. V. Prokhorov, K. Sofeikov, Approximation with random bases: Pro et contra,
Information Sciences 364–365 (2016) 129–145.
[7] F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, W. A. Stahel, Robust Statistics: The Approach Based on
Influence Functions, John Wiley & Sons, 2011.
[8] P. Huber, Robust Statistics, Wiley Series in Probability and Mathematical Statistics, Wiley, 1981.
[9] B. Igelnik, Y. H. Pao, Stochastic choice of basis functions in adaptive function approximation and the
functional-link net, IEEE Transactions on Neural Networks, 6 (6) (1995) 1320–1329.
[10] L. J. Latecki, A. Lazarevic, D. Pokrajac, Outlier detection with kernel density functions, in: Proceedings
of the 5th International Conference on Machine Learning and Data Mining in Pattern Recognition, Berlin,
Heidelberg, (2007) 61–75.
[11] A. M. Leroy, P. J. Rousseeuw, Robust Regression and Outlier Detection, Wiley Series in Probability and
Mathematical Statistics, Wiley, 1987.
[12] K. Liano, Robust error measure for supervised neural network learning with outliers, IEEE Transactions
on Neural Networks, 7 (1) (1996) 246–250.
[13] M. Li, D. Wang, Insights into randomized algorithms for neural networks: practical issues and common
pitfalls, Information Sciences 382-383 (2017) 170-178.
[14] D. Lowe, Multi-variable functional interpolation and adaptive networks, Complex Systems 2 (1988)
321–355.
[15] P. Meer, D. Mintz, A. Rosenfeld, D. Y. Kim, Robust regression methods for computer vision: A review,
International Journal of Computer Vision 6 (1) (1991) 59–70.
[16] Y. H. Pao, Y. Takefji, Functional-link net computing, IEEE Computer Journal 25 (5) (1992) 76–79.
[17] P. Ramsay, D. Scott, Multivariate Density Estimation, Theory, Practice, and Visualization, Wiley, 1993.
[18] S. Scardapane, D. Wang, Randomness in neural networks: an overview, Wiley Interdisciplinary Reviews:
Data Mining and Knowledge Discovery 7 (2) (2017) e1200. doi:10.1002/widm.1200.
[19] J. A. Suykens, J. De Brabanter, L. Lukas, J. Vandewalle, Weighted least squares support vector machines:
robustness and sparse approximation, Neurocomputing 48 (1) (2002) 85–105.
[20] P. H. Torr, A. Zisserman, MLESAC: A new robust estimator with application to estimating image
geometry, Computer Vision and Image Understanding 78 (1) (2000) 138–156.
[21] D. Wang, M. Li, Stochastic configuration networks: Fundamentals and algorithms, arXiv:1702.03180
[cs.NE], 10 Feb, 2017.
[22] X. Zhuang, T. Wang, P. Zhang, A highly robust estimator through partially likelihood function modeling
and its application in computer vision, IEEE Transactions on Pattern Analysis and Machine Intelligence
14 (1) (1992) 19–35.
14
| 9cs.NE
|
1
Distributed Multi-Relay Selection in
Accumulate-then-Forward Energy Harvesting Relay
Networks
arXiv:1602.00339v3 [cs.IT] 8 Oct 2017
Yifan Gu, He Chen, Yonghui Li, Ying-Chang Liang and Branka Vucetic
Abstract—This paper investigates a wireless-powered cooperative network (WPCN) consisting of one source-destination pair
and multiple decode-and-forward (DF) relays. We develop an
energy threshold based multi-relay selection (ETMRS) scheme
for the considered WPCN. The proposed ETMRS scheme can be
implemented in a fully distributed manner as the relays only need
local information to switch between energy harvesting and information forwarding modes. By modeling the charging/discharging
behaviours of the finite-capacity battery at each relay as a finitestate Markov Chain (MC), we derive an analytical expression for
the system outage probability of the proposed ETMRS scheme
over mixed Nakagami-m and Rayleigh fading channels. Based on
the derived expression, the optimal energy thresholds for all the
relays corresponding to the minimum system outage probability
can be obtained via an exhaustive search. However, this approach
becomes computationally prohibitive when the number of relays
and the associated number of battery energy levels is large. To
resolve this issue, we propose a heuristic approach to optimize
the energy threshold for each relay. To gain some useful insights
for practical relay design, we also derive the upper bound for
system outage probability corresponding to the case that all
relays are equipped with infinite-capacity batteries. Numerical
results validate our theoretical analysis. It is shown that the
proposed heuristic approach can achieve a near-optimal system
performance and our ETMRS scheme outperforms the existing
single-relay selection scheme and common energy threshold
scheme.
Index Terms—Wireless energy harvesting, cooperative communications, relay selection, accumulate-then-forward, Markov
Chain, outage probability.
I. I NTRODUCTION
The performance of many wireless communication networks
in practice is largely confined by the energy constrained
devices that require replenishment periodically. Recently, a
novel radio-frequency (RF) energy transfer and harvesting
technique has been proposed as a new viable and promising
solution to prolong the lifetime of energy constrained wireless
networks [1]. RF energy transfer and harvesting enables wireless devices to harvest energy from RF signals broadcast by
ambient/dedicated energy transmitters to charge their batteries
[2], [3]. This technique has opened a new research paradigm,
termed wireless-powered communication (WPC), which has
Y. Gu, H. Chen, Y. Li, Y.-C. Liang and B. Vucetic are with
School of Electrical and Information Engineering, The University of
Sydney, Sydney, NSW 2006, Australia (email: [email protected],
[email protected],
[email protected],
[email protected],
[email protected]).
The material in this paper was presented in part at the IEEE International
Conference on Communications, Kuala Lumpur, Malaysia, in May 2016.
become a hot research topic recently (see, e.g., [4], [5] and
references therein).
The WPC technique has brought new research opportunities to cooperative communications, which have attracted
an upsurge of interest during the past decade due to its
various advantages [6]. In this paper, we refer to a cooperative
communication network with wireless-powered1 relay(s) as a
wireless-powered cooperative network (WPCN). In [7], Nasir
et. al first investigated a classical three-node WPCN consisting of one source-destination pair and one energy harvesting amplify-and-forward (AF) relay. Two practical relaying
protocols, namely time switching-based relaying and power
splitting-based relaying, were proposed and analyzed in [7].
Inspired by this seminal work, a plenty of works focusing on
the design and/or analysis of WPCNs have published in open
literature very recently (see [8]–[22] and references therein).
All aforementioned works on WPCNs assumed that the
wireless-powered node(s) exhausts the harvested energy
in the current time slot to perform information transmission/forwarding straight away. Equipping each wirelesspowered node with an energy storage (e.g., a rechargeable
battery) such that they can accumulate the harvested energy
and then perform information tranmission/forwarding in an
appropriate time slot can improve the system performance
significantly. The energy accumulation process of the classical
three-node WPCNs with a single wireless-powered relay was
modeled and the resulting network performance was analyzed
in [23], [24] for finite and infinite storage scenario, respectively. [25] studied a multi-user network where all the users
can harvest and accumulate energy from the base station
simultaneously. Based on the considered system, the authors
derived the transmission probability, the signal-to-interference
ratio coverage probability and the overall success probability.
It is also of great importance to investigate the network setup
with multiple wireless-powered relays. Specifically, the relay
selection problem was studied in [26] for a time-division and
full-duplex block structure. Inspired by the max-max relay
selection strategy, a new relay selection scheme is proposed
that the relay with the best source-to-relay link is selected to
receive information and store it in its buffer while the relay
with the best relay-to-destination link is selected to transmit information. The corresponding outage probability and
throughput were then analyzed. Inspired by the opportunistic
1 Throughout this paper, we use the terms “wireless-powered” and “energy
harvesting” interchangeably.
2
2 For more information about CSI acquisition for energy harvesting devices,
interested readers could see [29] and [30].
݄ଵ
S
݄ଷ
݄ଶ
݄ே
ࡾ
ࡾ
ࡾ
……
relaying (OR) originally proposed in [27], a battery-aware
relay selection (BARS) scheme was proposed and analyzed.
In the BARS, the relays with accumulated energy exceeding a
predetermined threshold will first form a subset, which need to
feedback their channel state information (CSI) to the source.
Then, the “best” relay among the subset with maximum endto-end signal-to-noise ratio (SNR) is selected by the source to
forward its information, while other relays harvest energy in
the first hop. [28] studied a WPCN with multiple randomly
distributed relays. A distributed beamforming (DB) scheme
was proposed that all relay nodes which are fully charged at
the beginning of the transmission block form a forwarding
subset. Among this subset of relays, the relays that are able
to decode the source’s signal create a virtual multiple antenna
array and transmit source’s signal to destination coherently.
In this paper we develop an efficient energy threshold
based multi-relay selection (ETMRS) scheme with energy
accumulation capability at each relay for WPCNs. In the
proposed ETMRS, each relay can flexibly switch between
energy harvesting and information forwarding modes in each
transmission block. It should be noted that the proposed
scheme is purely distributed and only the local battery status
and local channel state information2 (CSI) are required at each
relay to perform mode selection, thus not involving extensive
inter-relay information exchanges as in existing schemes. We
also consider a practical and more general channel fading than
[28], [31], where source-relay links and relay-destination links
are assumed to experience independent but not necessarily
identical distributed (i.n.i.d.) channel fading. Different from
[28], we adopt a multiple-level battery model to characterize
the charging/discharging behavior at relay batteries. In this
case, the amount of harvested energy available at relays could
be different. As such, to adequately exploit the relays, in
our ETMRS scheme we set an individual energy threshold
for each relay. This is in contrast to the existing schemes
that considered independent and identical distributed channel
fading and adopted a common energy threshold for all relays
[28], [31]. The performance of the proposed ETMRS scheme
can be further improved by jointly optimizing the energy
thresholds of all relays. Moreover, compared with the DB
scheme in which the weights in distributed beamforming for
the relays within the forwarding set are obtained based on full
CSI of the whole forwarding subset, the weight at each relay
in the proposed ETMRS approach is calculated based only
on the local CSI such that it can be implemented in a fully
distributed manner. This can effectively reduce the network
overhead and latency. Note that in our ERMRS scheme the
relays that are selected to forward source’s information to
destination in each transmission block can be any combination
of all relays, which depends on not only the instantaneous
CSI but also the long-term evolution of all relay batteries.
This makes the performance analysis and system design of
the considered network a non-trivial task since we need to
develop a systematic approach to characterize the probability
of different relay combinations as well as the statistics of the
ࡾࡺ
݃ଵ
݃ଶ
݃ଷ
݃ே
D
Fig. 1. The considered WPCN with one source-destination pair and N
wireless-powered relays.
corresponding end-to-end SNR as a sum of i.n.i.d random
variables. Furthermore, the energy threshold of each relay that
determines its long-term energy evolution should be jointly
designed to boost the system performance.
Notation: Throughout this paper, we use fX (x) and FX (x) to
denote the probability density function (PDF) and cumulative
distribution function (CDF) of a random variable X. Γ (·) is
the Gamma function [32, Eq. (8.310)], γ (·, ·) is the lower
incomplete gamma function [32, Eq. (8.350.1)] and ⌈·⌉ is the
ceiling function. We use (·)∗ and (·)T to represent the complex
conjugate and the transpose of a matrix or vector, respectively.
E [·] is the expectation operator and I denotes the identity
matrix. P { A |B } is the conditional probability of A under a
given condition B.
II. S YSTEM M ODEL AND S CHEME D ESIGN
A. System Model
As depicted in Fig. 1, in this paper we investigate a WPCN
consisting of one source-destination pair and N decode-andforward (DF) relays, which are deployed to assist the source’s
information transmission. We assume that there is no direct
link between source and destination due to obstacles or severe
attenuation. Also, all nodes are equipped with single antenna
and work in half-duplex mode. As in [28], we consider the
scenario that all DF relays are wireless-powered devices and
purely rely on the energy harvested from RF signals broadcast
by source to perform information forwarding. Moreover, these
relays are equipped with separate energy and information receivers [33]. As such, they can flexibly switch between energy
harvesting (EH) mode and information forwarding (IF) mode
at the beginning of each transmission block. We also assume
that each relay is equipped with a finite-capacity rechargeable
battery such that it can perform energy accumulation and
scheduling across different transmission blocks. Specifically,
they can accumulate the harvested energy to a certain amount
before assisting the source’s information transmission.
We denote by T the duration of each transmission block,
which is further divided into two time slots with equal length
T /2. During the first time slot, the source broadcasts its
signal to all relays. At each relay operating in EH mode,
the received signal is delivered to the energy receiver to
convert to direct current and charge the battery. In contrast,
the received signal at each relay in IF mode is connected to
its information receiver to decode the information sent by the
source. All relays that operate in IF mode and decode the
source’s information correclty form a decoding set. In the
second time slot, all relays in the decoding set will jointly
3
forward the source’s information to destination by consuming
part of the accumulated energy from their batteries. On the
other hand, other relays outside the decoding set keep in
silence during the second time slot3 .
We hereafter use subscript-S and subscript-D to denote
the source and destination respectively. We denote by Ru ,
u ∈ {1, 2, · · · , N }, the u-th wireless-powered relay. Among
existing fading models, Rician fading would be the most
appropriate one to characterize the channel fading of S-Ru
links. This is mainly motivated by the fact that the up-to-date
wireless energy transfer techniques could only be operated
within a relatively short communication range such that the
line-of-sight (LoS) path is very likely to exist in these links.
However, the statistical functions (e.g., cumulative density
function (CDF) and probability density function (PDF)) of
Rician fading are very complicated, which would make the
analysis extremely difficult [34]. Fortunately, the Rician distribution could be well approximated by the more tractable
Nakagami-m fading model. Thus, in this paper we adopt an
asymmetric scenario for the fading distributions of sourceto-relay links and relay-to-destination links. Specifically, the
S-Ru link is assumed to be subject to Nakagami-m fading
with fading severity parameter mu and average power gain
λSRu , while the Ru -D link suffers from Rayleigh fading with
average power gain λRu D as the distance between them may
be much further. Besides, all channels in the considered system
experience slow, independent, and frequency-flat fading such
that instantaneous channel gains remain unchanged within
each transmission block but change independently from one
block to the other. It is worth mentioning that we do not require
the channel fading parameters to be identical or non-identical
for both hops. That is, we investigate a general independent
but not necessarily identical fading model, which includes the
independent and identical one as well as the independent and
non-identical one for special cases. Without loss of generality,
we consider a normalized transmission block (i.e., T = 1)
hereafter.
Let P denote the source transmit
power and x denote the
transmitted symbol with E | x| 2 = 1. The received signal at
the u-th relay during the first time slot is thus given by
√
(1)
yu = Phu x + n,
where hu is the channel coefficient between S and Ru , and n
denotes the additive white Gaussian noise (AWGN) with zero
mean and variance N0 at the receiver side.
When the u-th relay works in EH mode, the received signal
yu will be delivered to its energy receiver and converted to
direct current to charge the battery. The amount of harvested
energy at Ru during the first time slot can thus be expressed
as
1
E˜u = ηPHu,
(2)
2
where 0 < η < 1 is the energy conversion efficiency and
Hu = |hu | 2 is the channel power gain between S and Ru . Note
3 For simplicity, we consider that the relays out of the decoding set will not
harvest energy in the second time slot since the amount of energy harvested
from the forwarded signals would be negligible compared with the energy
harvested from the source.
that in (2), we ignore the amount of energy harvested from the
noise since the noise power is normally very small and below
the sensitivity of the energy receiver. On the other hand, if Ru
opts to decode information in the first time slot, it will harvest
zero energy.
Let Φ denote the current decoding set. In the second time
slot, all relays in the decoding set Φ will jointly forward
the source’s information to the destination by implementing
the distributed beamforming technique [35]. Specifically, the
transmitted signal at Ru ∈ Φ is given by
p
xu = wu Pu x,
(3)
where wu is the weight of Ru in distributed beamforming and
Pu is the transmit power of Ru . In the considered WPCN, we
assume that each relay only knows its local CSI of the second
hop. In this case, the optimal weight for the u-th relay that
maximizes the overall end-to-end SNR can be expressed as
wu = gu∗ /|gu | [35], where gu is the complex channel coefficient
between Ru -D. We define gˆu = |gu | for notation simplicity.
The received signal at the destination can thus be expressed
as
Õ
p
yd =
gˆu Pu x + n.
(4)
u:Ru ∈Φ
As a result, the conditional end-to-end SNR for a given
decoding set Φ can be written as
2
Í √
Pu gˆu
γΦ =
u:Ru ∈Φ
N0
.
(5)
B. Energy Threshold Based Multi-Relay Selection
In this paper we develop an energy threshold based
multi-relay selection (ETMRS) framework for the considered
WPCN. In our ETMRS scheme, each relay Ru determines
its individual energy threshold, denoted by χ̃u . This energy
threshold includes two parts: the first part is the energy
consumption for circuit operation (e.g., information decoding),
denoted by α̃ for all the relays; the second part is the energy
consumption for information forwarding, denoted by β̃u for
relay Ru . Each relay decides to operate in EH mode or IF
mode based on its own battery status at the beginning of each
transmission block. Specifically, relay Ru will perform the IF
operation only when its accumulated energy is not less than
its associated energy threshold χ̃u . Otherwise, it will opt EH
mode to further accumulate energy in its battery. Moreover,
if Ru works in the IF mode and falls in the decoding set, it
will decode and forward the source signal to destination by
consuming the amount of energy β˜u from its battery in the
second time slot. The conditional end-to-end SNR given in
(5) can now be updated by substituting Pu = β˜u /(1/2) = 2 β˜u .
That is,
2
Í p ˜
2 βu gˆu
u:Ru ∈Φ
.
(6)
γΦ =
N0
Note that we are investigating a distributed scheme such
that each relay only requires local CSI and battery status to
4
determine its operation modes. In this case, the optimal transmit power for each transmission block cannot be obtained as
it requires global CSI and battery status. As a result, the fixed
transmit power strategy is still preferable in our considered
case and we can thus set fixed energy thresholds to each
relay, χ̃u, u = 1, 2, · · · , N. Moreover, the energy thresholds for
all the relays should be different in order to achieve the best
system performance when the relays are in different locations.
For the special case that all the relays are co-located in a
cluster, the channels of each hop are now independent and
identically distributed (i.i.d.). In this sense, all relays can adopt
the same energy threshold χ̃ = χ̃1 = χ̃2 = · · · = χ̃N and
consume the same power to forward information, denoted by
β̃ = β̃1 = β̃2 = · · · = β̃ N . We then have the conditional endto-end SNR of this i.i.d. case given by
2
Í
2 β̃
gˆu
will also be verified later in the simulation section. For each
relay node, we define state Si as the relay residual energy in
i, j
the battery being εi . The transition probability Tu is defined
as the probability of transition from state Si to state S j at the
u-th relay. With the adopted discrete-level battery model, the
amount of harvested energy can only be one of the discrete
energy level. Thus, the discretized amount of harvested energy
at the u-th relay during an EH operation is defined as
∆
˜
(8)
εi : εi ≤ E u .
Eu = ε j , j = arg max
III. P ERFORMANCE A NALAYSIS
The energy consumption for information forwarding β̃u and
the energy threshold for each relay χ̃u should be chosen from
one of the energy levels of the battery excluding the empty
level. The descretized energy consumption for information
forwarding βu and discretized energy threshold for each
relay χu can be defined as βu ∈ {ε1, ε2, · · · , ε L − α} and
χu = α+ βu ∈ {α + ε1, α + ε2, · · · , ε L }, respectively. Note that
the system will work properly only when α < ε L , otherwise
when α ≥ ε L , the fully charged battery even cannot support
the circuit energy consumption of the relay and βu does not
exist.
We now evaluate the state transition probabilities of the MC
for each relay Ru , u ∈ {1, 2, · · · , N }. Different from [23] that
summarizes all the transition probabilities into eight general
cases, we propose a compact mode-based approch which
summarizes all the transition probabilities into the following
two cases.
1) The relay Ru operates in EH mode (Si to S j with
0 ≤ i < χεu1 ≤ L and ∀ j): When the relay Ru operates in
the EH mode, it harvests energy from the source during the
first time slot while remains in silence during the second time
slot. Due to the fact that the relay battery is not discharged
in the transition, the transition probability is none-zero only
when the end state falls into the set S j ∈ {Si, Si+1, · · · , SL }.
Specifically, S j = Si indicates that the battery level of the
relay remains unchanged and the harvested energy during the
transition E˜u is discretized to zero (i.e., Eu = 0). S j = SL
denotes the case that the battery is fully charged during the
transition and the harvested energy should be larger than ε L−i .
From the definition of discretization given in (8), the transition
probabilities for Ru operates in EH mode can be summarized
in (10) on top of the next page.
Recall that the channels between source and relays are
assumed to suffer from Nakagami-m fading. As such, the
PDF and CDF of Hu are given by [37, eq.2.21] fHu (x) =
bu mu mu −1
u ,bu x)
exp(−bu x), FHu (x) = γ(mΓ(m
, where bu =
Γ(mu ) x
u)
mu /λSRu . With the CDF, the transition probabilities for this
case can now be expressed as
⌢
γΦ =
u:Ru ∈Φ
N0
.
(7)
To analyze the performance of the proposed ETMRS
scheme, in this section we first characterize the dynamic charging/discharging behaviors of the relay batteries. We consider
a discrete-level and finite-capacity battery model. Thus, it is
natural to use a finite-state Markov chain (MC) to model the
dynamic behaviors of relays’ batteries. From the MC model
and the derived stationary distribution of the battery, we then
derive an approximate analytical expression of the system
outage probability for our ETMRS scheme. Based on the
derived analytical expression, we subsequently discuss how
to optimize the energy threshold of the relays to reduce the
system outage probability.
A. Markov Chain of Relay Batteries
Thanks to the fact that our ETMRS scheme is a decentralized relay selection approach and the relays make decisions of
their operation modes based only on their local CSI and battery
status, we thus can evaluate the steady state distributions of
all relays’ batteries separately as they are independent to
each other. In [23], the authors investigated a single relay
WPCN and proposed an adaptive information forwarding
scheme such that the relay forwards the source information
only when its residual energy can guarantee an outage-free
transmission in the second hop. The transition probabilities of
the relay battery was summarized into eight general cases.
Recall that our ETMRS scheme implements fixed transmit
powers for multiple relays. By adequately using this feature, a
compact mode-based method is used to evaluate the transition
probabilities of the MC for each relay.
Let C denote the capacity of all relays’ batteries and L
denote the number of discrete levels excluding the empty level
in each battery. Then, the i-th energy level of the relay battery
can be expressed as εi = iC/L, i ∈ {0, 1, 2 · · · L}. It is worth
pointing out that as shown in [36], the adopted discrete battery
model can tightly approximate its continuous counterpart when
the number of energy levels (i.e., L) is sufficiently large, which
i ∈{0,1, ··· , L }
Moreover, for relays operating in the IF operation mode, the
energy consumption for decoding operation α should also be
discretized to one specific energy level of the battery with the
definition give by
ε
:
ε
≥
α
, if α̃ ≤ ε L
arg
min
i
i
∆
i ∈{0,1, ··· , L }
(9)
α = ε j, j =
∞, if α̃ > ε
L
5
Tui, j
Tui, j
2 ( j − i) C
2 ( j − i + 1) C
,
≤ Hu <
Pr
˜
ηPL
ηPL
Pr ε j−i ≤ Eu < ε j−i+1 , if i ≤ j < L
2 (L − i) C
=
= Pr E˜u ≥ ε L−i , if i ≤ j = L
, if i ≤ j = L
Pr Hu ≥
0, if i > j
ηPL
0, if i > j
2( j − i + 1)C
FHu
ηPL , if i ≤ j < L
− FH 2( j − i)C
u
.
=
ηPL
2(L
−
i)C
1 − FHu
, if i ≤ j = L
ηPL
0, if i > j
(11)
2) The relay Ru operates in IF mode (Si to S j with
≤ i ≤ L and ∀ j): In this case, the relay Ru will try
to decode the received signal and forward it to destination
if the decoding is successful. In the first time slot, the relay
consumes energy α to decode the received signal from S. If
the decoding is unsuccessful, it remains in silence during the
second time slot. On the other hand, if the relay Ru decodes the
information successfully, it forwards the decoded information
to the destination by further consuming βu from its battery
during the second time slot. We can now conclude that after
the transition, the end state is none-zero only when j = i − εα1
or j = i − χεu1 . Recall that ϕu is the decoding indicator of relay
Ru and the transition probabilities for Ru operating in IF mode
can be summarized as
α
Pr {ϕu = 0} , if j = i −
ε1
i, j
Tu = 1 − Pr {ϕu = 0} , if i ≤ j = i − χu .
(12)
ε1
0, Otherwise
χu
ε1
We now analyze the term Pr {ϕu = 0} for the u-th relay. Let
u
γu = PH
N0 denote the received SNR at relay Ru . The channel
capacity of the S-Ru link is given by Θu = 21 log2 (1 + γu ).
According to the channel capacity, the term Pr {ϕu = 0} can
be evaluated as
vN0
,
(13)
Pr {ϕu = 0} = Pr {Θu < κ} = FHu
P
where κ is the system transmission rate and v = 22κ − 1 is the
SNR threshold for system outage.
Based on the above analysis, the transition probabilities can
be re-written as
vN0
α
, if j = i −
FHu
P
ε
1
vN0
χu .
Tui, j =
(14)
1 − FHu
, if i ≤ j = i −
P
ε
1
0, Otherwise
i, j
We now define Zu = (Tu ) to denote the (L + 1) × (L +
1) state transition matrix of the MC for each relay Ru , u ∈
{1, 2, · · · , N }. By using similar method in [23], we can easily
verify that the MC transition matrix Zu derived from the above
if i ≤ j < L
.
(10)
MC model is irreducible and row stochastic. Thus for each
relay Ru , there must exist a unique stationary distribution π u
that satisfies the following equation
T
π u = πu,0, πu,1, · · · , πu, L = (Zu )T π u,
(15)
where πu,i , i ∈ {0, 1, · · · , L}, is the i-th component of π u
representing the stationary distribution of the i-th energy level
at relay Ru . The battery stationary distribution of relay Ru can
be solved from (15) and expressed as [23]
−1
π u = (Zu )T − I + B b,
(16)
where Bi, j = 1, ∀i, j and b = (1, 1, · · · , 1)T .
Moreover, when it comes to the i.i.d. channel model, all
relays are equipped with an identical energy threshold and they
have the same transition matrix. The corresponding identical
stationary distribution of all relays can be similarly obtained
as (16) and denoted by π 1 = π 2 = · · · = π N = π =
(π0, π1, · · · , π L )T .
B. System Outage Probability
With the above derived stationary distribution of the relay
batteries, we now characterize the system outage probability of
the proposed ETMRS scheme. Let O denote the outage event
of the considered system employing our ETMRS scheme.
According to the full probability theory, we can express the
system outage probability as
Õ
Pout = Pr {O} =
Pr {Φ} Pr {O |Φ } ,
(17)
Φ∈Λ
where Λ = {R1, R2 · · · , RN } denotes the set of all relays in
the considered network and incorporates the decoding set Φ
as its subset, Pr {Φ} denotes the probability that the current
decoding set is Φ, and Pr {O |Φ } denotes the probability that
system outage occurs under the decoding set Φ. In order to
further expand (17), we define Φk,n to denote the n-th k-subset
of Λ (i.e., the n-th k-subset of Λ contains exactly k elements,
k = 1, 2, · · · , N, n = 1, 2, · · · , Nk ). Then the outage probability
of the ETMRS scheme can be further expanded as
(k )
N Õ
Õ
N
Pout = Pr {∅} +
k=1 n=1
Pr Φk,n Pr O Φk,n ,
(18)
by realizing that the system outage probability equals to one
when the decoding set is an empty set ∅. The empty decoding
set can be caused by two kinds of event: One is that none
of the relays working in IF mode decodes the received signal
from source correctly. The other is that all relays operate in
EH mode and no relay performs IF. Based on the derived
6
Pr Φk,n =
Ö
u:Ru ∈Φk, n
=
Ö
u:Ru ∈Φk, n
Pr {ζu = ζI , ϕu = 1}
Ö
u:Ru <Φk, n
u:Ru ∈Λ
L
Ö ©
Õ
ª
=
πu,i +
πu,i ®.
Pr {ϕu = 0}
u:Ru ∈Λ «
i=0
i=χu /ε1
¬
Similarly, the term Pr Φk,n can be computed as (20) on top
of the next page.
To evaluate the third probability term in (18), we first
characterize the distribution of the conditional end-to-end SNR
for a given decoding set. For notation simplicity, we use
γk,n to denote the received SNR at the destination when the
decoding set is Φn,k . Recall that the conditional end-to-end
SNR for a certain decoding set is given in (6), which includes
a weighted sum of Rayleigh random variables. However, to
the best knowledge of authors, the exact distribution of a
weighted sum of Rayleigh random variables does not exist in
open literature. As a result, we cannot further characterize the
exact distribution of γk,n . Fortunately, with the aid of a tight
approximation for the CDF of a weighted sum of Rayleigh
random variables derived in [38], the CDF of γk,n can be
approximated as a gamma distribution and expressed as
!
χÕ
u /ε1
γ k,
Fγk, n (x) ≈
4
Í
N0
u:R u ∈Φk, n
Γ (k)
βu σu2
(19)
x
,
(21)
p
where σu = λRu D /2 is the scale parameter of the Rayleigh
fading channel between Ru and D.
In order to further expand the summation term in (21), we
use a similar method
we
adopted in [39]–[41]. To this end,
√
βu σu : Ru ∈ Λ, u ∈ {1, 2, · · · , N } with
define a set A =
the same cardinality as the set Λ (i.e., |A| = |Λ| ). Similarly,
let Ak,n denote the n-th k-subset of A. The j-th element of the
subset Ak,n are denoted by φk,n, j ∈ Ak,n , j = 1, 2, · · · , k. To
be more clear,
√ we list the corresponding relationship between
φk,n, j and βu σu in Table I.
√
With the corresponding relation between φk,n, j and βu σu ,
(21) can now be expressed as
i
k−1
Õ
ak,n x
γ k, ak,n x
= 1 − exp(−ak,n x)
, (22)
Fγk, n (x) ≈
Γ (k)
i!
i=0
where ak,n =
4
k
Í
N0
j=1
Pr {ζu = ζI , ϕu = 0} ∪ ζu = ζE
L
Õ
(1 − Pr {ϕu = 0})
π
u,i
i=χ
/ε
u 1
stationary discrete distribution of relay batteries given in (16),
we can calculate the first probability term in (18) as follows
Ö
Pr {∅} =
Pr {ζu = ζI , ϕu = 0} ∪ ζu = ζE
and the last equality in (22) holds
2
φk,
n, j
according to [32, Eq. (8.352.6)] with integer k. Based on the
Ö
χÕ
L
u /ε1
Õ
©
ª
πu,i +
πu,i ®.
Pr {ϕu = 0}
u:Ru <Φk, n «
i=0
i=χu /ε1
¬
Ȁ
ൌͳ
ൌʹ
ǥ
ൌǦͳ
ൌ
ൌͳ
ൌʹ
ߚଶ ɐଶ
ࣘଵǡଵǡଵ ൌ
ߚଵ ɐଵ
ࣘଵǡଶǡଵ ൌ
ࣘଶǡଵǡଵ ൌ
ࣘଶǡଵǡଶ ൌ
ߚଵ ɐଵ
ߚଶ ɐଶ
ࣘଶǡଶǡଵ ൌ ߚଵ ɐଵ
ࣘଶǡଶǡଶ ൌ ߚଷ ɐଷ
ǥ
ࣘேିଵǡଵǡଵ ൌ ߚଵ ɐଵ
ࣘேିଵǡଵǡଶ ൌ ߚଶ ɐଶ
ǥ
ࣘேିଵǡଵǡேିଵ ൌ
ࣘேǡଵǡଵ ൌ
ࣘேǡଵǡଶ ൌ
ǥ
ࣘேǡଵǡேିଵ ൌ
ߚேିଵ ɐேିଵ
ߚଵ ɐଵ
ߚଶ ɐଶ
ߚே ɐே
ǥ
ࣘேିଵǡଶǡଵ ൌ
ࣘேିଵǡଶǡଶ ൌ
ǥ
ߚଵ ɐଵ
ߚଶ ɐଶ
ࣘேିଵǡଶǡேିଶ ൌ ߚேିଶ ɐேିଶ
ࣘேିଵǡଶǡேିଵ ൌ ߚே ɐே
െ
ǥ
ǥ
ǥ
ǥ
ǥ
ǥ
(20)
ൌ
ࣘଵǡேǡଵ ൌ
ࣘ
ே ேǦଵ
ଶǡ ଶ ǡଵ
ࣘ
ൌ
ߚே ɐே
ߚேିଵ ɐேିଵ
ൌ
ߚே ɐே
ࣘேିଵǡேǡଵ ൌ
ࣘேିଵǡேǡଶ ൌ
ߚଶ ɐଶ
ߚଷ ɐଷ
ே ேିଵ
ଶǡ
ǡଶ
ଶ
ǥ
ǥ
ࣘேିଵǡேǡேିଶ ൌ ߚேିଵ ɐேିଵ
ࣘேିଵǡேǡேିଵ ൌ ߚே ɐே
െ
TABLE I√
T HE RELATION BETWEEN φ k, n, j AND βu σu , u = 1, 2, · · · , N , FOR
k = 1, · · · , N , n = 1, 2, · · · , N
k AND j = 1, 2, · · · , k.
distribution of γk,n , the third probability term in (18) can now
be further expanded as
Pr O Φk,n = Pr γk,n < v = Fγk, n (v)
i
k−1
Õ
(23)
ak,n v
,
≈ 1 − exp(−ak,n v)
i!
i=0
where v = 22κ − 1 is the SNR threshold for system outage and
κ is the system transmission rate. By substituting (19), (20)
and (23) into (18), we have derived an approximate analytical
expression of the outage probability for the proposed ETMRS
scheme.
In terms of the i.i.d channel fading case where the relays
have the same energy threshold, the number of different
decoding sets reduces to N + 1 and the expression of system
outage probability can be simplified to
N
n⌢o Õ
n⌢ o n ⌢ o
Pout = Pr {O} = Pr ∅ +
Pr Φk Pr O Φk ,
⌢
k=1
(24)
n⌢o
where Pr ∅ is the probability that the decoding set is empty
n⌢ o
for the i.i.d. case, Pr Φk is the probability that the decoding
n ⌢ o
set contains k relays, and Pr O Φk is the conditional outage
probability when k relays falls in the decoding set. Similar to
7
the analysis of the general case, the first and second probability
terms in (24) can be expressed as
χ/ε
L
n⌢ o
Õ
Õ1 ª
©
πi +
Pr ∅ = Pr {ϕ = 0}
πi ® ,
i=0
i=χ/ε
1
¬
«
k
L
n ⌢ o N
Õ
(1 − Pr {ϕ = 0})
π
Pr Φk =
i ×
k
i=χ/ε
1
N
L
Õ
χ/ε
Õ1
N−k
(25)
(26)
©
ª
πi +
πi ®
,
Pr {ϕ = 0}
i=0
i=χ/ε
1
«
¬
where χ = α + β is the identical energy threshold for all the
relays and Pr {ϕ = 0} is the outage probability of the relays
for the special i.i.d. channel fading case, which can be derived
from that for the
case given in (13) and written as
general
mv N0
P
γ m, λ
SR
, where m is the identical severity
Pr {ϕ = 0} =
Γ(m)
parameter of the Nakagami-m fading channels between source
and relays and λSR is the identical channel power gain of the
first hop.
The third probability term in (24) can be deduced from (23)
as
i
N0 v
k−1
n ⌢ o
Õ
2
4kβσ
N0 v
,
(27)
Pr O Φk ≈ 1 − exp(−
)
i!
4k βσ 2 i=0
p
where σ = λRD /2 is the identical scale parameter for the
Rayleigh fading channels between relays and destination. Note
that different from (23), the expression of the conditional
system outage probability given in (27) does not require the
parameters defined in Table I. The outage probability for
the special i.i.d. channel fading case can be obtained by
substituting (25), (26) and (27) into (24).
C. Energy Threshold Optimization
We first consider the special i.i.d. channel fading case where
the associated energy thresholds for all the relays is identical,
denoted by χ = α + β. When the energy used for information
forwarding β grows, the conditional outage probability derived
in (27) reduces and the overall system outage probability correspondingly decreases. On the other hand, increasing of β will
lead to an increase of χ. According to (26), it decreases the
probability of each relay working in IF mode and the number
of relays falling into the decoding set. This will increase the
overall system outage probability. In simple words, when the
designed energy threshold χ for the relays is small, most of
relays could fall into the decoding set but their transmit power
is low. When the energy threshold is large, only a few relays
fall into the decoding set but their associated transmit power
is high. As χ = α + β ≥ α + ε1 , we can now infer that for the
special i.i.d. channel fading case, there should exist an optimal
energy threshold χ ∈ {α + ε1, α + ε2, · · · , ε L } that minimize
the system outage probability of the proposed ETMRS scheme.
Similarly, for the general i.n.i.d. channel fading case where
the relays may be located dispersively. There should also
exist an optimal energy threshold set { χ1, χ2, · · · , χN } , χu ∈
{α + ε1, α + ε2, · · · , ε L } , ∀u = 1, 2, · · · , N, for all relays that
minimize the overall system outage probability.
Due to the complexity of the adopted MC model, it is
difficult to derive an analytical expression for the optimal
energy threshold of each relay. However, for the special i.i.d.
channel fading case, the optimal energy threshold can be easily
achieved by performing a one-dimensional exhaustive search
from all the possible energy levels with the derived analytical
outage probability given in (24). The computation complexity
of this search is given by O (L).
When it comes to the general i.n.i.d. case, the optimal
set of energy thresholds can be found by a N-dimension
exhaustive search from all the possible combinations of energy
thresholds with the analytical expression
derived in (18). The
computation complexity is O L N , which grows exponentially
with the number of relays N. Thus, finding the optimal set of
energy thresholds becomes computationally prohibitive when
the number of relays N and the level of batteries L are large.
To overcome this problem, in the following subsection, we
provide a heuristic approach to design the energy thresholds
for the general i.n.i.d. case.
D. A Heuristic Approach for i.n.i.d Scenario
Intuitively, a higher energy threshold and forwarding transmit power should be set for those relays closer to the source
node. This is because that the average amount of harvested
energy is higher for those relays compared to the ones far away
from the source node. On the other hand, relays closer to the
source are relatively further away from the destination node
such that the associated second hop channels are relatively
weaker, a larger transmit power should be used to overcome
the higher path loss. On the other hand, for the relays near to
the destination node, smaller forwarding transmit power can
be adopted due to their limited harvested energy and stronger
second hop channels. Inspired by this fact, we set the energy
consumption for information forwarding at each relay Ru as
β̃u = zλSRu /λRu D,
(28)
where z is a scalar factor to adjust the overall transmit power
of all the relays. For the considered discrete-battery model,
the designed energy consumption for information forwarding
β̃u should be discretized to one specific energy level of
the battery. Note that the definition of discretization given
below (9) no longer holds as we cannot simply choose
βu ∈ {ε1, ε2, · · · , ε L − α}. We now define the discretized value
∆
of β̃u for the u-th relay as βu = ε j ,
˜
εi : εi ≥ βu , if β˜u ≤ ε L − α
min
arg i ∈{0,1,
··· , L }
. (29)
j=
α
L − , if β˜u > ε L − α
ε1
Based on the above definitions, the discretized energy threshold for relay Ru can be further expressed as
zλSRu
,C − α .
(30)
χu = α + βu = α + min ε1
λ R u D ε1
Similar to the analysis given in the previous subsection,
we can deduce that there should exist an optimal value of
8
z that minimizes the system outage probability. To find the
optimal value of z, we first define λmax and λmin as the
maximum and minimum value of the term λSRu /λRu D , among
all relays, respectively. The optimal z should exist within
the interval [ε1 /λmax, ε L /λmin ], where z = ε1 /λmax could
make all relays choose ε1 + α as their energy thresholds
and z = ε L /λmin will force all relays to adopt energy
level L as their energy thresholds. The optimal z thus can
now be achieved by performing a one-dimension exhaustive
search from this interval over the derived outage probability
expression. In order to capture all the possible combinations,
we search z with an increment of ε1 /λmax each time and the
computation complexity for the proposed heuristic approach
can be expressed by O (Lλmax /λmin ). Note that the complexity
of the heuristic approach is obviously much
lower than that
of the exhaustive search given by O L N .
1
ηPλSRu .
(32)
2
With the fact that relay Ru falls in the decoding set only
when it performs IF operation and decodes the source’s
information correctly, we thus have
IV. P ERFORMANCE U PPER B OUND
qu = (1 − pu ) (1 − Pr {ϕu = 0}) ,
As described in Section III, we adopt a practical finitecapacity battery model in this paper. With this model, it can
be readily deduced that the system performance could be
improved when we increase the battery capacity (i.e., C). This
is because a larger battery capacity can reduce the energy loss
caused by energy overflow (i.e., the battery cannot be charged
when it is full) and thus the relays have more available energy
to support their information forwarding operation. On the other
hand, we can infer that the system performance improvement
speed actually decreases as the battery capacity increases since
the energy overflow happens more rarely when the battery
capacity keeps increasing. A natural question that comes up
here is “For a given network setup, how large of the battery
capacity C will be sufficient?” This question is particularly
important for the considered system as one of its potential
applications will be low-cost and lower-power networks (e.g.,
wireless sensor networks), in which the network deployment
cost should be kept as low as possible by carefully selecting
the battery capacity. However, it is hard to find the answer of
this question based on the derived expressions in the previous
section due to their complexity. We are thus motivated to
adopt an indirect way: We first derive performance upper
bounds of the considered system and the sufficient battery
capacity will then be obtained when a certain value of C can
make the performance expressions derived in previous sections
approach their corresponding upper bounds. In this sense, in
this section we analyze the performance upper bounds of the
considered system with infinite battery capacity (i.e., C → ∞
and implicitly L → ∞).
When the relay batteries are infinite, there will be no energy
overflow. As such, we can implement the well-known flow
conservation law to evaluate the system outage probability.
Specifically, in our ETMRS scheme, the relay battery is
charged if operated in EH mode and discharged if operated
in IF mode. In this sense, for each relay, the total amount
of harvested energy in a long term should equal to the total
amount of energy consumed for information decoding and
forwarding. Mathematically, we have the following formula
for the u-th relay
pu Ēu = (1 − pu ) [(1 − qu ) α + qu (α + βu )] ,
(31)
where pu denotes the probability that relay Ru performs EH
operation, qu denotes the probability that the u-th relay falls in
the decoding set, Ēu is the average amount of harvested energy
during each EH operation, α is the circuit energy consumption
for each decoding operation and βu is the energy consumption
for information forwarding of Ru adopted in the finite-capacity
case. Since both C and L are infinity, the term Ēu can be
obtained from (2) given by
Ēu =
(33)
where the term Pr {ϕu = 0} has been derived in (13). By
jointly considering (31)-(33), we can now obtain qu expressed
as
1
.
(34)
qu =
2α+2β
1
u (1−Pr{ϕu =0})
1−Pr{ϕu =0} + η Pλ S R (1−Pr{ϕu =0})
u
The first and second probability terms in the system outage
probability defined in (18) can now be evaluated by qu as
Ö
Pr {∅} =
(1 − qu ) ,
(35)
Ö
u:Ru ∈Λ
Pr Φk,n =
qu
u:Ru ∈Φk, n
Ö
u:Ru <Φk, n
(1 − qu ) .
(36)
The third probability term in (18) is independent of C and
L, and only depends on the energy threshold of the relays.
In this sense, the third probability term in (18) remains the
same as the finite-capacity case, which has been derived in
(23). We thus can express the upper bound of system outage
probability for the proposed ETMRS scheme corresponding to
infinite battery capacity as (37) on top of the next page.
When the special i.i.d. channel fading case is considered,
similar to the above analysis, the system outage probability is
upper bounded by (38) on top of the next page.
where q is the probability of each relay falling in the
decoding set when the channels are i.i.d. and it can be easily
derived based on (34) and expressed as
q=
1
1
1−Pr{ϕ=0}
+
2α+2β(1−Pr{ϕ=0})
η Pλ S R (1−Pr{ϕ=0})
.
(39)
Remark 1: By substituting (13) into (34), we can see that the
probability for the u-th relay falling in the decoding set (i.e.,
qu ) is proportional to the term PλSRu . This is understandable
since as the value of PλSRu increases, the u-th relay harvests
more energy on average and is more likely to decode the
source signal successfully in the first hop. Moreover, as
expected, the value of qu is inversely proportional to that
of the energy threshold βu . Furthermore, using the above
performance upper bounds of system outage probability, we
9
ub
Pout
i #
"
( Nk ) Ö
k−1
N Õ
Ö
Õ
Õ
ak,n v
qu
(1 − qu ) 1 − exp(−ak,n v)
≈
(1 − qu ) +
.
i!
u:Ru <Φk, n
u:Ru ∈Λ
i=0
k=1 n=1 u:Ru ∈Φk, n
Ö
⌢ ub
P out ≈ (1 − q) N +
N
Õ
N
k
k=1
q k (1 − q) N−k
(37)
i
N0 v
k−1
Õ
2
4kβσ
N0 v
)
,
1 − exp(−
i!
4k βσ 2 i=0
(38)
N=8
0
1
10
0.9
L = 20
Outage Probability
Outage Probability Ratio
0.8
−1
10
L = 40
−2
10
L = 200
−3
10
Finite−Capacity Battery, Analytical Eq. (18)
Infinite−Capacity Battery, Analytical Eq. (37)
Monte Carlo Simulation
0.7
0.6
0.5
0.4
P = 34dBm, dSR = 4, ∀u
0.3
P = 34dBm, dSR = 5, ∀u
0.2
P = 35dBm, dSR = 4, ∀u
u
u
u
P = 35dBm, dSR = 5, ∀u
0.1
0
u
1
−4
15
20
25
30
Source Transmit Power P (dBm)
35
Fig. 2. The outage probability of the proposed ETMRS scheme versus the
source transmit power for different battery levels L, where transmission bit
rate κ = 1, C = 2 × 10−5 , α = 10−7 , dS R = {5, 5.5, 6, 6, 6, 6, 6.5, 7} and
χ = {3, 3, 3, 3, 3, 3, 4, 4} × 10−6 .
are able to judge whether a certain value of battery capacity
C is sufficiently large for a given network setup via numerical
results as shown in next section.
AND
2.5
3
−5
x 10
(a) Combinations of P and dS R u
P = 35dBm
1
0.9
0.8
0.7
0.6
0.5
0.4
N = 6, dSR = 4, ∀u
0.3
N = 6, dSR = 5, ∀u
0.2
N = 9, dSR = 4, ∀u
0.1
N = 9, dSR = 5, ∀u
u
u
0
V. S IMULATIONS
1.5
2
Battery Capacity C
40
Outage Probability Ratio
10
D ISCUSSIONS
In this section, we present some simulation and numerical
results to validate and illustrate the above theoretical analysis.
In order to capture the effect of path-loss, we use the model
10−3
that λXY = 1+d
ω , where λ XY is the average channel power
XY
gain between node X and Y, dXY denotes the distance between
node X and Y, and ω ∈ [2, 5] is the path-loss factor. Note that
a 30 dB average signal power attenuation is assumed at a
reference distance of 1 meter (m) in the above channel model
[19]. For simplicity, we consider a linear topology that the
relays are located on a straight line
between the source and
destination and denote by dSR = dSR1 , dSR2 , · · · , dSR N the
set of distances between source and all relays. We use χ =
{ χ1, χ2, · · · , χN } to represent the energy threshold set for all
the relays. In all the following simulations, we set the distance
between the source and destination dSD = 20m, the path-loss
factor ω = 3, the severity parameter mu = 2, ∀u, the noise
power N0 = −90dBm, and the energy conversion efficiency
η = 0.5.
We first compare the analytical system outage probability
with its Monte Carlo simulation, which corresponds to the
case that the charging of the relay batteries is continuous (i.e.,
L → ∞). To this end, we plot the system outage probability
u
u
1
1.5
2
Battery Capacity C
2.5
3
−5
x 10
(b) Combinations of N and dS R u
Fig. 3. The ratio between the upper bound outage probability of the ETMRS
scheme with infinite battery capacity and the outage probability with finite
battery capacity versus battery capacity C for different source transmit power
and relay topologies, where κ = 1, α = 10−7 , χu = 4×10−6 , ∀u and L = 600.
of our ETMRS scheme versus the source transmit power for
different battery levels L in Fig. 2. We can see that the derived
analytical expression of outage probability approaches the
corresponding Monte Carlo simulation result as the discrete
battery level L increases. Specifically, when L = 200, the
analytical expression coincides well with the simulation result,
which verifies the effectiveness of the adopted MC model and
the correctness of our theoretical analysis presented in Sec.
II-IV. Moreover, as expected, the performance of the ETMRS
scheme with finite-capacity battery is bounded by the one with
infinite-capacity battery. As the analytical results agree well
with the simulation results and for the purpose of simplicity,
in the following, we will only plot the analytical results of the
ETMRS scheme.
Recall that we derived system performance upper bound in
10
⌢ ub
N=8
0
10
−3
10
Optimal χ
−1
10
Outage Probability
Section IV, which can be used to judge whether a certain
value of C is sufficiently large via numerical results. To
show this, we now plot the outage probability ratio of the
proposed ETMRS scheme with finite battery capacity over its
upper bound with infinite battery capacity versus relay battery
capacity C in Fig. 3. For simplicity, we consider the special
i.i.d. case where the relays are located in a cluster. The outage
−4
10
0.5
1
1.5
2
−6
−2
x 10
10
P = 37dBm, dSR = 4, ∀u
u
P = 37dBm, dSR = 6, ∀u
−3
u
10
⌢
P = 40dBm, dSR = 4, ∀u
probability ratio is formally defined as Pout /Pout ∈ (0, 1],
u
P = 40dBm, dSR = 6, ∀u
⌢ ub
u
10
0.2
0.4
0.6
0.8
1
1.2
1.4
Identical Energy Threshold χ
1.6
1.8
2
−5
x 10
(a) Combinations of P and dS R u
P = 40dBm
0
10
N = 6, dSR = 4
Optimal χ
u
N = 6, dSR = 6
−1
u
10
N = 9, dSR = 4
Outage Probability
u
N = 9, dSR = 6
u
−2
10
−3
10
−4
10
1
−5
10
0.2
0.4
0.6
0.8
1
1.2
1.4
Identical Energy Threshold χ
2
3
−6
1.6x 10 1.8
2
−5
x 10
(b) Combinations of N and dS R u
Fig. 4. The outage probability of the proposed ETMRS scheme versus the
identical energy threshold for different source transmit power, relay topologies
where κ = 1, C = 2 × 10−5 , α = 10−7 and L = 200.
1
0.9
0.8
dSR = {2, 4, 6, 8}
0.7
Outage Probability
where Pout derived in (38) is the system outage probability
⌢
for infinite-capacity battery case and Pout given in (24) is
the system outage probability for finite-capacity battery case.
From Fig. 3, we can first observe that the outage probability
ratio monotonically increases as the battery capacity C grows
and gradually converges to 1 when the value of C is large
enough. This indicates that the performance gap between the
finite-capacity battery and infinite-capacity battery decays to
0 as C increases. However, the convergence speed varies for
different network setups. Specifically, from Fig. 3(a), we can
see that the outage probability ratio converges to 1 slower
with either higher source transmit power or shorter distances
between source and relays. This can be understood as follows.
The amount of harvested energy at relays increases when
the source transmit power increases or the distances between
source and relays reduce. In this case, a larger battery capacity
is required to avoid the potential energy overflow (i.e., the
battery cannot be charged when it is full). From Fig. 3(b),
we find out that as the number of relays N increases, the
convergence speed is also reduced. This is understandable
since energy overflow happens in a higher probability as N
grows. Similarly, larger value of C is required to make the
outage probability ratio close to 1. We can now summarize
that larger capacity batteries should be equipped at energy
harvesting relays for those network setups with higher source
transmit power, shorter distances between source and relays
or larger number of relays.
In Fig. 4, we plot the outage probability versus the energy
threshold of the i.i.d. channel fading case for different source
transmit power and relay location. Recall that the energy
threshold χ is discretized to one of the L + 1 energy levels of
the battery excluding the empty level. Thus, the outage probability is plotted in stair curve in the figure. First of all, Fig.
4 demonstrates that there exists an optimal energy threshold
that minimizes the outage probability in all considered cases,
which validates our deduction in Remark 1. Moreover, we
can see from Fig. 4(a) that the higher the transmit power at
the source, the larger the value of optimal energy threshold.
This is because the relays can harvest more energy when
the transmit power of the source increases and thus a larger
energy threshold can be supported. We can also see from Fig.
4(a) that a smaller energy threshold should be chosen when
the relays are located away from the source. This can be
explained as two folds. Firstly, the harvested energy at the
relays is limited when they are far away from the source.
Secondly, the second hop channel condition becomes better
when the relays are away from the source (i.e., close to the
destination) and a small energy threshold is enough to avoid
system outage. The number of the relays N also affects the
−4
0.6
dSR = {4, 5, 5, 5, 6}
0.5
0.4
0.3
0.2
Exhaustive Search Approach
Proposed Heuristic Approach
Common Energy Threshold Approach [29]
0.1
0
10
12
14
16
18
20
22
24
Source Transmit Power P (dBm)
26
28
30
Fig. 5. Outage probability of the proposed ETMRS scheme and the common
energy threshold scheme for different relay locations, where κ = 1, C =
2 × 10−5 , α = 10−7 and L = 20.
optimal energy threshold. From Fig. 4(b) we can see that as the
number N increases, the optimal energy threshold decreases.
This is understandable since in cooperative relay networks, a
decoding set with more relays can achieve the same outage
probability with less transmit power. Note that similar results
can be observed for the general i.n.i.d case, which are not
provided here due to space limitation..
We then compare the system performance of the proposed
ETMRS scheme with optimized energy threshold and the common energy threshold scheme in Fig. 5. The optimal energy
threshold of the ETMRS is obtained by the exhaustive search
based approach and the proposed heuristic one. As the optimal
energy threshold for the i.i.d. case can be easily obtained
11
0
10
−1
Outage Probability
10
−2
10
−3
10
ETMRS Analytical, κ = 1
BARS Simulation, κ = 1
ETMRS Analytical, κ = 1.5
BARS Simulation, κ = 1.5
−4
10
15
20
25
30
Source Transmit Power P (dBm)
35
40
Fig. 6. The outage probability of the proposed ETMRS scheme and the
existing BARS scheme with optimal settings versus source transmit power for
different transmission bit rate, where N = 8, dS R u = 3, ∀u, C = 2 × 10−5 ,
α = 10−7 and L = 200.
via a one-dimension exhaustive search, we only consider the
general i.n.i.d. case where the optimal energy thresholds for
the relays could be different. In Fig. 5, we plot the system
outage probability for these two approaches versus the source
transmit power for two different relay topologies. The outage
probability of exhaustive search approach is obtained via a
N-dimension exhaustive search from all the possible combinations of energy thresholds. Due to the intense computation
complexity mentioned before, we consider the case that each
relay only has 20 energy levels. For the proposed heuristic
approach, the computation complexity can be dramatically
reduced and it is thus particularly suitable to those networks
with large number of relays and energy levels. From Fig. 5,
we can observe that the proposed heuristic approach with
reduced complexity can achieve the near optimal system
performance in both simulated scenarios. We also can see
that our proposed scheme outperforms the common-energy
strategy for all simulated cases. For the case that the relays
are located closely, our proposed scheme and the one in [28]
performs similarly. When the relays are located differently,
our proposed scheme improves the performance significantly
by using different energy thresholds for each relay.
Finally, we compare the proposed ETMRS scheme with the
existing BARS scheme [31] in Fig. 6. As the BARS scheme
proposed in [31] only considers a clustered topology, we thus
compared BARS scheme with the special i.i.d. case of the
proposed ETMRS scheme. In Fig. 6, we plot the optimal
outage probabilities of the ETMRS and BARS schemes versus
source transmit power for different transmission rate. It can
be observed that the proposed ETMRS scheme can achieve
a lower outage probability than the BARS scheme. And their
performance gap is enlarged as the transmission rate grows.
VI. C ONCLUSIONS
In this paper, we proposed an energy threshold based multirelay selection (ETMRS) scheme for accumulate-then-forward
energy harvesting relay networks. We modeled the finitecapacity battery of the relays by a finite-state Markov Chain
(MC) in order to evaluate their stationary distribution. We
then derived an approximate analytical expression for system
outage probability of the proposed ETMRS scheme over
independent but not necessarily identical mixed Nakagami-m
and Rayleigh fading channels. A heuristic approach was then
designed to minimize the system outage probability, which was
shown to achieve near-optimal performance with reduced computational complexity. Moreover, we derived upper bounds for
the concerned system performance corresponding to the case
that all relays are equipped with infinite-capacity batteries.
Numerical simulations validated the accuracy of the analytical
results, demonstrated the impact of various system parameters
and provided some insights of practical relay battery design.
Numerical results showed that larger capacity battery should
be equipped at energy harvesting relays for those network
setups with higher transmit power, shorter distance in the first
hop or larger number of relays. Furthermore, the proposed
ETMRS scheme can considerably outperform the existing single relay selection scheme and the common energy threshold
scheme.
R EFERENCES
[1] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
with RF energy harvesting: A contemporary survey,” IEEE Communications Surveys Tutorials, vol. 17, no. 2, pp. 757–789, 2015.
[2] T. Paing, J. Shin, R. Zane, and Z. Popovic, “Resistor emulation approach
to low-power RF energy harvesting,” IEEE Trans. on Power Electronics,,
vol. 23, no. 3, pp. 1494–1501, May 2008.
[3] C. Mikeka and H. Arai, “Design of a cellular energy-harvesting radio,”
in Wireless Technology Conference, 2009. EuWIT 2009. European, Sept
2009, pp. 73–76.
[4] S. Bi, C. Ho, and R. Zhang, “Wireless powered communication: opportunities and challenges,” IEEE Communications Magazine, vol. 53,
no. 4, pp. 117–125, 2015.
[5] S. Bi, Y. Zeng, and R. Zhang, “Wireless Powered Communication
Networks: An Overview,” ArXiv e-prints, 2015.
[6] M. Dohler and Y. Li, Cooperative communications: hardware, channel
and PHY. John Wiley & Sons, 2010.
[7] A. Nasir, X. Zhou, S. Durrani, and R. Kennedy, “Relaying protocols for
wireless energy harvesting and information processing,” IEEE Transactions on Wireless Communications, vol. 12, no. 7, pp. 3622–3636, 2013.
[8] Z. Chen, B. Xia, and H. Liu, “Wireless information and power transfer in
two-way amplify-and-forward relaying channels,” in 2014 IEEE Global
Conference on Signal and Information Processing (GlobalSIP), Dec
2014, pp. 168–172.
[9] Z. Ding, S. Perlaza, I. Esnaola, and H. Poor, “Power allocation strategies
in energy harvesting wireless cooperative networks,” IEEE Transactions
on Wireless Communications, vol. 13, no. 2, pp. 846–860, 2014.
[10] Z. Fang, X. Yuan, and X. Wang, “Distributed energy beamforming for
simultaneous wireless information and power transfer in the two-way
relay channel,” IEEE Signal Process. Lett., vol. 22, no. 6, pp. 656–660,
June 2015.
[11] C. Zhong, H. Suraweera, G. Zheng, I. Krikidis, and Z. Zhang, “Wireless
information and power transfer with full duplex relaying,” IEEE Trans.
Commun., vol. 62, no. 10, pp. 3447–3461, Oct 2014.
[12] Y. Zeng and R. Zhang, “Full-duplex wireless-powered relay with selfenergy recycling,” IEEE Wireless Commun. Lett., vol. 4, no. 2, pp. 201–
204, April 2015.
[13] I. Krikidis, S. Sasaki, S. Timotheou, and Z. Ding, “A low complexity
antenna switching for joint wireless information and energy transfer in
MIMO relay channels,” IEEE Trans. Commun., vol. 62, no. 5, pp. 1577–
1587, May 2014.
[14] Z. Zhou, M. Peng, Z. Zhao, and Y. Li, “Joint power splitting and antenna
selection in energy harvesting relay channels,” IEEE Signal Process.
Lett., vol. 22, no. 7, pp. 823–827, July 2015.
[15] G. Zhu, C. Zhong, H. Suraweera, G. Karagiannidis, Z. Zhang, and
T. Tsiftsis, “Wireless information and power transfer in relay systems
with multiple antennas and interference,” IEEE Trans. Commun., vol. 63,
no. 4, pp. 1400–1418, April 2015.
[16] K. Xiong, P. Fan, C. Zhang, and K. Letaief, “Wireless information
and energy transfer for two-hop non-regenerative MIMO-OFDM relay
networks,” Accepted to appear in IEEE J. Sel. Areas Commun., 2015.
12
[17] M. Mohammadi, H. Suraweera, G. Zheng, C. Zhong, and I. Krikidis,
“Full-duplex MIMO relaying powered by wireless energy transfer,” in
2015 IEEE 16th International Workshop on Signal Processing Advances
in Wireless Communications (SPAWC), 2015, pp. 296–300.
[18] Z. Ding, I. Krikidis, B. Sharif, and H. Poor, “Wireless information and
power transfer in cooperative networks with spatially random relays,”
IEEE Transactions on Wireless Communications, vol. 13, no. 8, pp.
4440–4453, 2014.
[19] H. Chen, Y. Li, J. Luiz Rebelatto, B. F. Uchoa-Filho, and B. Vucetic,
“Harvest-Then-Cooperate: Wireless-Powered Cooperative Communications,” IEEE Trans. on Signal Processing, vol. 63, pp. 1700–1711, Apr.
2015.
[20] H. Chen, Y. Li, Y. Jiang, Y. Ma, and B. Vucetic, “Distributed power
splitting for SWIPT in relay interference channels using game theory,”
IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 410–420, Jan 2015.
[21] D. S. Michalopoulos, H. A. Suraweera, and R. Schober, “Relay selection
for simultaneous information transmission and wireless energy transfer:
A tradeoff perspective,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 8, pp. 1578–1594, 2015.
[22] B. K. Chalise, W. K. Ma, Y. D. Zhang, H. A. Suraweera, and M. G.
Amin, “Optimum performance boundaries of ostbc based AF-MIMO
relay system with energy harvesting receiver,” IEEE Transactions on
Signal Processing, vol. 61, no. 17, pp. 4199–4213, 2013.
[23] I. Krikidis, S. Timotheou, and S. Sasaki, “RF energy transfer for
cooperative networks: Data relaying or energy harvesting?” IEEE Communications Letters, vol. 16, no. 11, pp. 1772–1775, November 2012.
[24] A. Nasir, X. Zhou, S. Durrani, and R. Kennedy, “Wireless-powered relays in cooperative communications: Time-switching relaying protocols
and throughput analysis,” IEEE Trans. Commun., vol. 63, no. 5, pp.
1607–1622, May 2015.
[25] A. H. Sakr and E. Hossain, “Analysis of k-tier uplink cellular networks
with ambient rf energy harvesting,” IEEE Journal on Selected Areas in
Communications, vol. 33, no. 10, pp. 2226–2238, 2015.
[26] K. H. Liu and T. L. Kung, “Performance improvement for RF energyharvesting relays via relay selection,” IEEE Transactions on Vehicular
Technology, vol. PP, no. 99, pp. 1–1, 2017.
[27] A. Bletsas, A. Khisti, D. Reed, and A. Lippman, “A simple cooperative
diversity method based on network path selection,” IEEE Journal on
Selected Areas in Communications, vol. 24, no. 3, pp. 659–672, 2006.
[28] I. Krikidis, “Relay selection in wireless powered cooperative networks
with energy storage,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 12, pp. 2596–2610, 2015.
[29] J. Zhang and G. Pan, “Outage analysis of wireless-powered relaying
mimo systems with non-linear energy harvesters and imperfect csi,”
IEEE Access, vol. 4, pp. 7046–7053, 2016.
[30] M. Haghifam, B. Makki, M. Nasiri-Kenari, and T. Svensson, “On
Wireless Energy and Information Transfer in Relay Networks,” ArXiv
e-prints, 2016.
[31] K. Liu, “Performance analysis of relay selection for cooperative relays
based on wireless power transfer with finite energy storage,” IEEE Trans.
on Vehicular Technology, vol. PP, no. 99, pp. 1–1, 2015.
[32] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products,
ser. Table of Integrals, Series, and Products Series. Elsevier Science,
2007.
[33] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless
information and power transfer,” IEEE Trans. on Wireless Communications, vol. 12, no. 5, pp. 1989–2001, May 2013.
[34] C. Zhong, X. Chen, Z. Zhang, and G. Karagiannidis, “Wireless powered
communications: Performance analysis and optimization,” To appear in
IEEE Trans. Commun., 2015.
[35] J.-B. Kim and D. Kim, “Cooperative system with distributed beamforming and its outage probability,” in Vehicular Technology Conference,
2007. VTC2007-Spring. IEEE 65th, 2007, pp. 1638–1641.
[36] W.-J. Huang, Y.-W. Hong, and C.-C. Kuo, “Lifetime maximization
for amplify-and-forward cooperative networks,” IEEE Trans. Wireless
Commun., vol. 7, no. 5, pp. 1800–1805, May 2008.
[37] M. Simon and M. Alouini, Digital Communication over Fading Channels, ser. Wiley Series in Telecommunications and Signal Processing.
Wiley, 2005.
[38] M. Hanif, N. Beaulieu, and D. Young, “Two useful bounds related
to weighted sums of rayleigh random variables with applications to
interference systems,” IEEE Trans. on Communications, vol. 60, no. 7,
pp. 1788–1792, 2012.
[39] D. Michalopoulos and G. Karagiannidis, “Performance analysis of
single relay selection in rayleigh fading,” IEEE Trans. on Wireless
Communications, vol. 7, no. 10, pp. 3718–3724, 2008.
[40] H. Chen, J. Liu, L. Zheng, C. Zhai, and Y. Zhou, “An improved selection
cooperation scheme for decode-and-forward relaying,” IEEE Commun.
Lett., vol. 14, no. 12, pp. 1143–1145, December 2010.
[41] H. Chen, J. Liu, Z. Dong, Y. Zhou, and W. Guo, “Exact capacity
analysis of partial relay selection under outdated CSI over rayleigh
fading channels,” IEEE Trans. on Vehicular Technology, vol. 60, no. 8,
pp. 4014–4018, Oct 2011.
| 7cs.IT
|
MD. ABDUL LATIF SARKER, CHONBUK NATIONAL UNIVERSITY, KOREA
1
Distortion-free Golden Precoded Alamouti for
MISO Systems
arXiv:1710.09142v2 [cs.IT] 15 Mar 2018
Md. Abdul Latif Sarker
Abstract—Although the discrete Fourier transform (DFT)
precoded Alamouti (DPA) scheme with limited feedback offers
numerous benefits in the Rayleigh fading channel, this scheme
inherently generates geometric mean distortion due to the error exponential of the DFT precoding (DP). Consequently, the
receiver suddenly losses the average received channel power.
Instead of the conventional DPA scheme, we propose a Golden
precoded Alamouti (GPA) scheme to construct a distortion-free
codeword generator that can be used in multiple-input and singleoutput (MISO) systems. The Golden section with continuous
geometric proportion and the Golden precoding (GP) pattern can
be used to construct some good space-time block codes (STBCs)
with orthogonal properties and simple inverse computation for
use in future wireless communications. In this paper, we overcome
the Grassmannian subspace packing problem and compute the
pairwise error probability (PEP) to verify the bit-error-rate
(BER) performance. Our Monte Carlo simulations show that
the proposed GPA scheme performs better than the conventional
DPA scheme.
Index Terms—MISO systems, Alamouti code, the DPA and
GPA scheme, subspace packing problem, BER performance.
I. I NTRODUCTION
T
HE use of orthogonal space-time block codes (OSTBCs)
is an important technique in wireless communications
to improve the network reliability and offer a high data rate
and diversity gain [1]. A well-known space-time block code
(STBC) is the Alamouti code, which is a complex orthogonal
space-time code specialized for the use with two transmission
antennas to achieve full diversity [1]. In addition, OSTBCs
provide the benefit of low-complexity symbol-wise decoding
[2].
An open loop is generally used in STBCs because they operate
without the benefit of channel state information (CSI) at the
transmitter [3] and because of the results of studies on closed
loops with STBCs to improve the performance of MIMO
systems [4]. Closed-loop schemes are based on the type of
CSI that is available at the transmitter, such as quantized CSI,
although occasionally only statistical CSI may be available on
the channel. However, we assume that the channel is known at
both the transmitter and the receiver. The CSI is estimated at
the receiver and is returned to the transmitter with limited
feedback [3], [5]. Thus, in this study, we consider limited
feedback with updated CSI, where the precoder index can be
given based on the CSI. Instead of full CSI, the corresponding
index can be given as feedback to the transmitter.
The discrete Fourier transform precoding (DP) codebook was
first proposed in [3], [5], where the authors used the optimal
codebook selection criterion to drive a criterion for finding
codebooks with low average distortion. Hai et al. [6] showed
that the performance of discrete fractional sine transform
(DFRST) is equivalent to that of DFT precoding, although they
attempted to reduce the hardware complexity. Unfortunately,
most existing works have not completely removed the mean
distortion from the codeword generator. Thus, in this work,
we design a distortion-free Golden precoding codebook (GPC)
that constructs an error-free codeword generator.
STBCs with Golden numbers were proposed in [7], [8]. In
[7], the authors showed that the Golden code satisfies the
properties of full-rank, full-rate, and non-vanishing determinants, which increase the minimum determinants. The Golden
and Silver code were compared with diagonal precoding for
non-OSTBCs to evaluate the pairwise error probability [8].
Many existing works have introduced a Golden number for
STBCs, but Golden precoding has not yet been introduced for
OSTBCs. Thus, this paper designs a Hadamard-based GPC
scheme to produce a distortion-free codeword generator that
measures the minimum chordal distance and computes the
pairwise error probability to verify the system bit-error-rate
(BER) performance. The rest of this paper is organized as
follows. First, we investigate the system model in Section II.
Second, we describe the conventional DFT precoded Alamouti
(DPA) scheme in Section III. We then propose the Golden
precoded Alamouti (GPA) scheme and compute the pairwise
error probability (PEP) in Section IV and Section V, respectively. Finally, the Monte Carlo simulations and conclusions
are presented in Sections VI and VII, respectively.
II. S YSTEM M ODEL
We consider multiple-input and single-output (MISO) systems with a flat Rayleigh fading channel hC1×NT that is
employed with multiple transmit antennas NT and a single
receive antenna NR = 1. The receive signal yC1×T at a
period T can be modeled as
y = hX + z,
(1)
where XCNT ×T is a transmit codeword structure of the
STBCs and zC1×T is a circularly symmetric complex Gaussian with zero mean and unit variance.
III. DPA S CHEME
Md. Abdul Latif Sarker is with the Department of Electronics and Information Engineering, Chonbuk National University, Jeonju 54896, Republic of
Korea. e-mail: ([email protected]).
Manuscript received xxx, 2017; revised xxx, 2017.
The DPA codeword generator X is given by [3]
X = WD S,
(2)
MD. ABDUL LATIF SARKER, CHONBUK NATIONAL UNIVERSITY, KOREA
2
where the M × T complex Alamouti code with a length of
M ≤ NT is
s −s∗
Sk = 11 ∗21 , k = 1, 2, .., T,
(3)
s21 s11
where WD CNT ×M is the NT × M DFT precoding matrix
j2πkl
1
given in [3] with √N
e NT at entry (k, l), which is chosen
T
from the codebook generator in [3], [5-6]:
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
-0.4
-0.2
-0.6
-0.4
-0.8
-0.6
i−1
FD = {θD
WD }, i = 1, 2, ..., L − 1.
(4)
The NT × M remaining DFT precoding matrices are given by
i−1
WD,i = θD
WD,1 , i = 2, ..., L,
(5)
where L is the size of the codebook; WD,1 is an NT × M
DFT-based first precoding matrix, the (k, l) of which is given
j2π(k−1)(l−1)
1
NT
as √N
e
; and θD is the diagonal matrix with the
T
NT
free variable {ui }i=1 to be determined in [3].
Problem statement: The given DPA scheme has two main
problems. The first problem is the geometric mean distortion
of the DPA scheme due to the error exponential of the DP
matrix, and the second problem is the minimum chordal
distance problem.
A. Geometric mean distortion of the DPA scheme
The DP matrix contains an error exponential that produces
a slight distortion at a low-order DPA scheme, although this
problem gradually arises at a high-order DPA scheme. Consequently, codeword matrix suffers a spectral leakage problem;
for example, the 2 × 2 DPA third codeword matrix is given
by (2)-(5) as follows:
θ2 h 1 1 i s11 −s∗21
[X3 ] = √D
jπ
s21 s∗11
2 1 e
h
i
0.0000 + 0.0000i 0.8315 − 0.5556i
=
, (6)
0.1379 + 0.6935i 0.1379 + 0.6935i
where i = 3, s11 = 0.7071 + 0.7071i and s21 = −0.7071 −
0.7071i are accounted for by the quadrature amplitude modulation (QAM) constellation, and (·)∗ is the complex conjugate
operator. Now, we increase the size of the DP matrix in (6).
For example, we set i = 3,M = 2, and NT = 4 in (6), and
we obtain the 4 × 2 DPA third codeword matrix in (7):
1 1
2
jπ
θ 1 e
s11 −s∗21
[X3 ] = √D
∗
j2π s
21 s11
4 1 e
j3π
1 e
0.0000 + 0.0000i 0.8315 − 0.5556i
0.1379 + 0.6935i 0.1379 + 0.6935i
=
. (7)
0.0000 − 1.0000i −0.0000 − 0.0000i
0.7071 − 0.0000i −0.7071 + 0.0000i
Using (6) and (7), we draw Fig. 1, which depicts the mean
distortion with the DPA generator. We observe in (6) and Fig.
1 (a) that the DPA generator contains a very small distortion,
which has gradually increased in the high-order DP matrix in
Fig. 1 (b). Hence, the codeword generator cannot be generated
as anti-diagonal blocks in the DPA matrix, and consequently,
the decoder requires more frequent CSI feedback to meet the
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
(a) Mean distortion with a 2 × 2(b) Mean distortion with a 4 × 2
DPA generator obtained using (6). DPA generator obtained using (7).
Fig. 1: Mean distortion of the DPA scheme. (a) Mean distortion with a 2 × 2 DPA generator obtained using (6). (b) Mean
distortion with a 4 × 2 DPA generator obtained using (7).
constraint. This problem arises from the error exponential of
the DP matrix, which degrades the BER performance by at
least 10−1 steps. However, we solve this problem in Section
IV.
B. Minimum chordal distance problem
The minimum chordal distance is the major issued for
packing subspaces in the Grassmann manifold. We formulate
the minimum chordal distance problem as follows:
The optimal DP matrix, Wopt , is given by
Wopt = argmaxWD,i F k HWD,i S̄ k2F ,
(8)
where the error matrix S̄ = Si − Sj and S̄S̄H = αI with a
constant α. When WD,i F is not imposed, the above optimum
solution Wopt is not unique. Thus, the average loss in the
received channel power is
E minW F (k HWopt k2F − k HW k2F )
2
1
H
H 2
≤ E λ1 (H) E minWF (k V̄V̄ − WW kF ) , (9)
2
where the singular value decomposition (SVD) technique
of channel H = UΣVH , the diagonal entry of Σ, is in
descending order; V̄ is unitary; λ21 (Wopt ) = 1; and E [·] is
called the expectation with regard to the random channel H.
Hence, the optimum codebook is designed to maximize the
minimum chordal distance [3], [5-6], [9]:
i−1
δmin,D = min2≤i≤L d(WD,1 , θD
WD,1 ).
(10)
IV. P ROPOSED GPA S CHEME
In this section, we will design a distortion-free GPA scheme.
For this purpose, we first design a Golden Hadamard precoding
matrix. Let NT = 2q and a 2q × M precoding matrix, W(2q ),
is constructed by M columns of the 2q × 2q recursive Golden
Hadamard precoding matrices:
1 θG WG (2q−1 )
θG WG (2q−1 )
, (11)
WG (2q ) = √
−1
q−1
q−1
ξ θG WG (2 ) θG (θG − θG )WG (2 )
for 2 ≤ qNT in [10], where W(1) = [1], ξ = n{(1 + n)q −
(1 − n)q }/2q , n indicates the root of the geometric number,
MD. ABDUL LATIF SARKER, CHONBUK NATIONAL UNIVERSITY, KOREA
−1
θG is the Golden number in [11] that satisfies θG θG
= 1,
−1
(θG − θG ) = −1, and its continuous geometric proportion
is θG : 1 : 1/θG , which indicates that the geometric mean
between θG and 1/θG is unity. Thus, we incorporate (11)
in(5)and construct the remaining Hadamard-based Golden
precoding (GP) matrices as follows:
WG,i (2q ) = (−1)i−1 WG,1 (2q ), i = 2, ..., L,
(13)
−1
where (θG
− θGr ) = −1 , WGr ,2 (4) = (−1)WGr ,1 (4), and
r
1 1 1 1
θG 1 −1 1 −1
(14)
WGr ,1 (4) = √ r
.
5 1 1 −1 −1
1 −1 −1 1
Case II (Complex Golden Number): Let θGc √
be the realj+ 3
2
valued root of µθGc (X) = X −jX −1; θGc = 2 is known
as the complex Golden
√ number, and its algebraic conjugate is
¯ = j− 3 . Similarly, we incorporate i = 3, q =
obtained√as θG
c
2
2, n = 3, and θGc values in (11-12) to obtain the 4 × 4
complex Golden Hadamard matrices as
WGc ,3 (4) = (−j)2 WGc ,2 (4), i = 2, ..., L,
(15)
−1
(θG
c
− θGc ) = −j , WGc ,2 (4) = (−j)WGc ,1 (4), and
1 1 1 1
θG 1 −1 1 −1
WGc ,1 (4) = √ c
(16)
.
3 1 1 −1 −1
1 −1 −1 1
Thus, we observe that (14) and (16) are the exact real and
complex Golden Hadamard matrix, respectively. Finally, our
designed GPC matrices are given by
where
(2q )},f or−case I,
{(−1)i−1 W
,i
FG = {(−j)i−1 WGG,ir(2
q )},f or−case II,
c
1.5
1.2
1
1
0.8
0.5
0.6
0
0.4
0.2
-0.5
0
(12)
where WG,1 (2q ) is an NT × M Hadamard-based first Golden
precoding matrix. Now, we state Theorem 1 as follows.
Theorem 1: The DPA codeword generator X becomes
distortion-free if the precoding matrix WG (2q ) is a Golden
Hadamard.
Proof of Theorem 1:
A. Golden Hadamard Precoding
Case I (Real Golden Number): We consider that θGr is
√ the
real-valued root of µθGr (X) = X 2 − X − 1; θGr = 1+2 5 is
known as the real Golden number [7,11], and its algebraic
√
¯ = 1 − θG = 1− 5 . We
conjugate is obtained as θG
r
r
2
√
incorporate i = 3, q = 2, n = 5, and θGr values in (1112) to obtain the 4 × 4 real Golden Hadamard matrices as
WGr ,3 (4) = (−1)2 WGr ,2 (4), i = 2, ..., L,
3
(17)
where i=1,2,...,L-1.
B. Distortion-free DPA Generator
Matrix: Let i = 3, q =
√
2, NT = 4, M = 2, and n = 5. Then, by substituting (14) in
(7), we obtain the 4 × 2 GPA third codeword generator matrix
by using case I as
0.0000 + 0.0000i 1.0233 − 1.0233i
1.0233 + 1.0233i 0.0000 + 0.0000i
[X3 ] =
(18)
,
0.0000 + 0.0000i 1.0233 − 1.0233i
1.0233 + 1.0233i 0.0000 + 0.0000i
-1
-0.2
-1.5
-0.4
0
0.2
0.4
0.6
0.8
1
1.2
(a) Case I
0
0.2
0.4
0.6
0.8
1
1.2
(b) Case II
Fig. 2: Distortion-free GPA scheme: (a) case I (18) and (b)
case II (19).
where s11 = 0.7071 + 0.7071i and s21 = −0.7071 − 0.7071i.
Similarly, by substituting (16) in (7), we obtain the 4 × 2 GPA
third codeword generator matrix by using case II as
0.0000 + 0.0000i 1.1153 − 0.2988i
0.2988 + 1.1153i 0.0000 + 0.0000i
[X3 ] =
(19)
.
0.0000 + 0.0000i 1.1153 − 0.2988i
0.2988 + 1.1153i 0.0000 + 0.0000i
Using (18) and (19), we draw Fig. 2 against Fig. 1 and observe that the proposed GPA scheme is completely distortionfree. Thus, Theorem 1 is proved.
V. PAIRWISE E RROR P ROBABILITY (PEP)
Let Si and Sj be the transmitted and decoded space-time
codewords, respectively, where j 6= i. Thus, the codeword PEP
is Pr (Si − Sj |H), where h is the given channel and WG,i (2q )
is the precoding matrix. An upper bound of the total PEP can
be given via the union bound in [12, Sec 5.2], and the average
error rate [13, Eq.4.2.2] is upper bounded by
1 X X
Pr,tot ≤
Pr (Si −→ Sj ) ,
(20)
|χ|
Si χ Sj χ,j6=i
where | χ | is the number of elements in the matrix constellation χ. The PEP depends upon the product distance S̄S̄H ,
where the error matrix S̄ = Si − Sj for the STBC scheme and
(·)H represents the Hermitian. The upper bound of the PEP is
!
r
γ0 kHWG,i (2q )S̄k2F
Pr (Si −→ Sj |H) = Q
2 · 2q
!
r
γ0 kHWG,i (2q )S̄k2F
≤ exp
, (21)
4 · 2q
where γ0 is the signal-to-noise ratio (SNR)and Q(·) is the
Gaussian Q-function. Thus, the proposed performance measures of the packing minimum chordal distance for both real
and complex cases are as follows:
δmin,Gr = min2≤i≤L d WGr ,1 (2q ), (−1)i−1 WGr ,1 (2q ) , (22)
and
δmin,Gc = min2≤i≤L d WGc ,1 (2q ), (−j)i−1 WGc ,1 (2q ) . (23)
MD. ABDUL LATIF SARKER, CHONBUK NATIONAL UNIVERSITY, KOREA
4
TABLE I: Computation of Minimum Chordal Distance
0
10
Minimum chordal distance (δmin )
when i = 2, q = 2, NT = 4, and M = 2
−1
10
Proposed GPC Scheme
Case I(22)
Case II(23)
DPC Scheme(10)
DPA(64−QAM)[3,5−6]
Proposed GPAc(64−QAM)
Proposed GPAr(64−QAM)
−2
10
1.0000
1.0000
BER
0.8392
−3
10
0
10
−4
DPA(16−QAM)[3,5−6]
Proposed GPAc(16−QAM)
Proposed GPAr(16−QAM)
10
−5
10
−1
10
−6
BER
10
−2
0
5
10
15
SNR[dB]
20
25
30
10
Fig. 4: BER performance of GPA scheme against DPA with
64-QAM, when q = 2, NT = 4,M = 2,L = 64, and γ =
30dBs.
−3
10
−4
10
0
5
10
SNR[dB]
15
20
Fig. 3: BER performance of GPA scheme against DPA with
16-QAM, when q = 2, NT = 4,M = 2,L = 64, and γ =
20dBs.
VI. S IMULATION R ESULTS
We compare our designed GPC scheme in (17) with the
conventional DPA scheme in (4) for a MISO flat block
fading Rayleigh channel. Our designed GPC matrix leads to
the minimal chordal distance δmin,G = 1 = sin(900 ); in
comparison, the minimal chordal distance of the DPC matrix
is δmin,D = 0.8392 = sin(57.050 ) in Table I. Thus, the
resulting GPC provides distortion-free performance, as shown
in Fig. 3 and Fig. 4. We apply two different codebook schemes,
16-QAM and 64-QAM, at SNR values of 20 and 30 dB,
respectively. In both the low- and high-SNR regimes, in Case I
(the real case), the GPA scheme performs significantly better
than DPA. In contrast, in Case II (the complex case), the
performance of the GPA scheme is lower because of the
reduction of the Golden ratio. However, the average array
gains of both the real (Case I) and complex (Case II)
GPA schemes are increased by almost 1.0 ∼ 3.0 dB with
log2 L = log2 26 = 6-bit feedback.
VII. C ONCLUSIONS
We investigated the problem of geometric mean distortion
of the DPA scheme due to the error exponential of the DP
matrix and proposed the GPC approach, which is achieved
with a distortion-free codeword. We addressed the problem
of the packing minimum chordal distance and demonstrated
the distortion-free performance of our proposed GPA scheme.
We may further deal with the presented results in multiuser
systems. In future work, we will investigate multi-layer GPC
scheme to further improve the sum-rate performance.
R EFERENCES
[1] S. M. Alamouti, A Simple Transmit Diversity Scheme for Wireless
Communication, IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 14511458, Oct. 1998.
[2] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, Space-time Block Codes
from Orthogonal Designs, IEEE Trans. Inf. Theory, vol. 45, no. 5, pp.
1456-1467, Jul. 1999.
[3] D. J. Love and R. W. Heath, Limited Feedback Unitary Precoding for
Orthogonal Space-time Block Codes, IEEE Trans. Signal Process., vol.
53, no. 1, pp. 64-73, Jan. 2005.
[4] M. Vu and A. Paulraj, Optimal Linear Precoders for MIMO Wireless
Correlated Channels with Non-zero Mean in Space-time Coded Systems,
IEEE Trans. Signal Process., vol. 54, no. 6, pp. 2318-2332, Jun. 2006.
[5] B. M. Hochwald, T. L. Marzetta, T. J. Richardson, W. Sweldens, and
R. Urbanke, Systematic Design of Unitary Space-time Constellations,
IEEE Trans. Inf. Theory, vol. 46, pp. 1962-1973, Sep. 2000.
[6] H. Hai, X. Q. Jiang, Y. Jeong and M. H. Lee, Low complexity Precoding
Codebook Design Based on DFRST for Limited Feedback Precoded
OSTBCs System, EURASHIP J. Wireless Commun. Netw., vol. 103,
pp. 1-6, Jun 2017.
[7] J. C. Belfiore, G. Rekaya, and E. Viterbo, The Golden Code: A 2 ×
2 Full-rate Space-time Code with Nonvanishing Determinants, IEEE
Trans. Inf. Theory, vol. 51, no. 4, pp. 1432-1436, Apr. 2005.
[8] A. Garg, V. Venu, and M. R. Bhatnagar, One Bit Feedback Based
Diagonal Precoding for Non-Orthogonal Space-Time Block Codes, IEEE
Wireless Commun. Lett., vol. 5. no. 5, Oct. 2015.
[9] D. J. Love, R. W. Heath, and T. Strohmer, Grassmannian Beamforming
for Multiple-Input Multiple-Output Wireless Systems, IEEE Trans. Inf.
Theory, vol. 49, no. 10, pp.2735-2747, Oct. 2003.
[10] H. Seleem, I.S. Ahmed, and A. Alsanie, Hybrid Precoding-Beamforming
Design with Hadamard RF Codebook for mmWave Large-Scale MIMO
Systems, IEEE Access, vol. 5, pp. 6813-6823, June 2017.
[11] S. Olsen,The Golden Section, Bloomsbury, New York, USA, 2006.
[12] J. G. Proakis, Digital Communications, 4/E, McGraw-Hill, New York,
USA, 2001.
[13] E. G. Larsson and P. Stoica, Space-Time Block Coding for Wireless
Communications, Cambridge University Press, Cambridge, UK, 2003.
| 7cs.IT
|
arXiv:1502.04365v3 [math.LO] 20 Dec 2017
Definably amenable NIP groups
Artem Chernikov∗†and Pierre Simon∗‡
December 21, 2017
Abstract
We study definably amenable NIP groups. We develop a theory of
generics showing that various definitions considered previously coincide, and study invariant measures. As applications, we characterize
ergodic measures, give a proof of the conjecture of Petrykowski connecting existence of bounded orbits with definable amenability in the
NIP case, and prove the Ellis group conjecture of Newelski and Pillay
connecting the model-theoretic connected component of an NIP group
with the ideal subgroup of its Ellis enveloping semigroup.
1
Introduction
In the same way as algebraic or Lie groups are important in algebraic or
differential geometry, the understanding of groups definable in a given firstorder structure (or in certain classes of first-order structures) is important
for model theory, as well as its applications. On the one hand, even if one
is only interested in abstract classification of first-order structures (i.e. in
understanding combinatorial complexity of definable sets), unavoidably one
is forced to study definable groups. (This realization probably started with
∗
The research leading to this paper has been partially supported by the European Research Council under the European Union’s Seventh Framework Programme (FP7/20072013)/ERC Grant Agreement No. 291111 and by ValCoMo (ANR-13-BS01-0006).
†
Partially supported by the Fondation Sciences Mathematiques de Paris (ANR-10LABX-0098), by NSF (grants DMS-1600796 and DMS-1651321) and by the Sloan foundation
‡
Partially supported by NSF (grant DMS-1665491), and by the Sloan foundation.
1
Zilber’s work on totally categorical structures [Zil], and later it was made
clear by Hrushovski’s theorem on unidimensional theories [Hru90].) On the
other hand, some of the most striking applications of model theory are based
on a detailed understanding of definable groups in certain structures. The
class of stable groups is at the core of model theory, and the corresponding theory was developed in the 1970s-1980s borrowing many ideas from the
study of algebraic groups over algebraically closed fields (with corresponding
notions of connected components, stabilizers, generics, etc., see e.g. [Poi01]).
In particular, this general theory was applied to groups definable in differentially closed and separably closed fields, and was used by Hrushovski to
prove the Mordell-Lang conjecture for function fields [Hru96] . The theory of
stable groups was generalized in the 90s to groups definable in a larger class
of simple theories, centered around the model-theoretic notion of forking
(see [Wag00]) and led to a number of results including Hrushovski’s proof
of the Manin-Mumford conjecture [Hru01] and other applications to algebraic dynamics (e.g. [MS14]). More recently, inspired by the ideas of stable
and simple group theory, Hrushovski has obtained a general stabilizer-type
theorem and found striking applications to approximate subgroups [Hru12],
which led to a complete classification by Breuillard, Green and Tao [BGT12].
On the other hand, groups definable in o-minimal structures were studied extensively, generalizing the theory of real Lie groups. This study culminated
in a recent resolution of Pillay’s conjecture for compact o-minimal groups
[HPP08], and the proof has brought to light the importance of the general
theory of groups definable in NIP structures (a common generalization of
stable and o-minimal structures, see below) and the study of invariant measures on definable subsets of the group. In parallel, methods and objects of
topological dynamics were introduced into the picture by Newelski [New09]
and gave rise to some new invariants coming from topological dynamics and
conjectures concerning their relationship to the more familiar model-theoretic
invariants. This circle of ideas has rapidly become a very active research area.
The present paper contributes to this direction and, continuing the work in
[HPP08], develops the theory of groups definable in NIP structures which
admit a translation invariant probability measure on the boolean algebra of
definable subsets.
The NIP condition (the negation of the Independence Property) is a combinatorial tameness assumption on a first-order structure M which says, in
modern terms, that if D ⊆ Mm+n is a definable set, then the family {D(a),
a ∈ Mm } of its fibers has finite VC-dimension (see Section 2.1). Roughly
2
speaking, it says that the collection of definable subsets of M is very structured. One can think of NIP as capturing the notion of a geometric structure — as opposed to say arithmetic or random-like — and of NIP groups
as groups arising in geometric settings. This condition was introduced by
Shelah [She71], and the connection to VC-dimension was discovered later in
[Las92]. In the past ten or fifteen years, the role of NIP theories has grown to
become a central notion in model theory thanks to applications to o-minimal
structures, valued fields and combinatorics (see e.g. [Sim15a] or [Sta16] for
a survey). Typical examples of NIP structures are given by stable structures
(such as algebraically closed fields), o-minimal structures (such as real closed
fields) and many Henselian valued fields. On the other hand, an ultraproduct of finite fields is an example of a structure which is not NIP, essentially
because of arithmetic phenomena that enter the picture.
Let now G be a group definable in an NIP structure M (i.e. both the
underlying set and multiplication are definable by formulas with parameters
in M). Such a group comes equipped with a collection of definable subsets of
cartesian powers of G which is closed under boolean combinations, projection
and cartesian products. For example, if M = (R, +, ×, 0, 1) is the field of
reals, then G is a real semi-algebraic group and definable sets are all semialgebraic subsets. As it is typical in model theory, we prefer to work in
a saturated model of our group (which in the case of an algebraic group
corresponds to working in the universal domain in the sense of Weil). More
precisely, let U M be a sufficiently saturated and homogeneous elementary
extension of M, a “monster model” for the first-order theory of M. We write
G(U) to denote the group obtained by evaluating in U the formulas used to
define G in M (and G(M) will refer to the set of the M-points of G). So
e.g. if we start with M the field of reals, and G(M) its additive group, then
G(U) is the additive group of a large real closed field extending R which
now contains infinitesimals, infinitesimals relatively to those infinitesimals,
etc. — i.e., it satisfies a saturation condition: every small enough finitely
consistent family of definable sets has non-empty intersection.
It was shown by Shelah [She08] that any NIP group G(U) admits a unique
maximal compact quotient denoted G/G00 , which plays an important role in
this theory and for which we will give a dynamical interpretation below.
Our goal in this paper is to adapt techniques from stable group theory to
the NIP context in order to have tools at our disposal potentially as useful as
those for stable groups. However, one main difference with the stable case is
that we cannot deal with all groups any more. As we show, to have any hope
3
of having well-behaved notions of generic types and large subsets, the group
must be definably amenable (this strengthens results of Hrushovski, Peterzil
and Pillay [HP11] who first observed this).
We say that a definable group G is definably amenable if there is a finitelyadditive probability measure on the boolean algebra of definable subsets of G
which is moreover invariant under the group action (this property holds for
G(U) if and only if it holds for G(M), see the remark after Definition 3.1).
This notion has been introduced and studied in [HPP08] and [HP11]. The
emphasis in those papers is on the special case of the so-called fsg groups,
which will not be relevant to us here. Of course, if G(M) is amenable as a
discrete group, then it is definably amenable since we have such a measure
on all subsets, not just the definable ones, but the converse need not hold
(e.g. deep work of Sela [Sel13] demonstrates that any non-commutative free
group, viewed as a first-order structure in the group language, is stable,
hence definably amenable; but of course it is not amenable). Here are some
important examples of definably amenable NIP groups:
• stable groups;
• definable compact groups in o-minimal theories or in p-adics (e.g.
SO3 (R));
• solvable NIP groups, or more generally, any NIP group G such that
G(M) is amenable as a discrete group.
Examples of definable NIP groups which are not definably amenable are
SL2(R) or SL2 (Qp ) (see [HPP08]).
It is classical in topological dynamics to consider the action of a discrete
group G on the compact space of ultrafilters on G, or in other words ultrafilters on the boolean algebra of all subsets of G. In the definable setting, given
a definable group G(M), we let SG (M) denote the space of ultrafilters on the
boolean algebra of definable subsets of G(M), hence the space SG (M) (called
the space of types of G(M)) is a “tame” analogue of the Stone-Čech compactification of the discrete group G. Then G(M) acts on SG (M) by homeomorphisms. The same construction applies to G(U) giving the space SG (U)
of ultrafilters on the definable subsets of G(U). Our main object of study
in this paper are the dynamical systems (SG (M), G(M)) and (SG (U), G(U))
and related objects. In this context, we classify regular ergodic measures
and show in particular that minimal flows are uniquely ergodic. We also
give various characterizations of definable subsets of G which have positive
measure for some (resp. for all) invariant measures, connecting topological
dynamics of the system with Shelah’s model-theoretic notion of forking.
4
A starting point of this theory is a theorem of Shelah stating that any
NIP group G(U) admits a maximal compact quotient G/G00 (the kernel G00
is characterized as the smallest subgroup of G(U) which is an intersection
of definable subsets and has small index in G(U)). We give a dynamical
interpretation of this compact quotient by establishing an isomorphism between the ideal subgroup of the Ellis semigroup of a certain extension of
(SG (M), G(M)) and G/G00 . Those results settle several questions in the
area.
Now we state the main results more precisely. In the case of stable groups,
a natural notion of a generic set (or type) was given by Poizat (generalizing
the notion of a generic point in an algebraic group), and a very satisfactory
theory of such generics was developed in [Poi87]. In a non-stable group,
however, generic types need not exist, and several substitutes were suggested
in the literature, either motivated by the theory of forking as in simple groups
([HPP08, HP11]), or by topological dynamics ([NP06]). First we show that
in a definably amenable NIP group all these notions coincide, and that in
fact nice behaviour of these notions characterizes definable amenability.
Theorem 1.1. Let G = G(U) be a definable NIP group, with U a sufficiently
saturated model. Then the following are equivalent:
1. G is definably amenable (i.e. admits a G-invariant measure on its
definable subsets).
2. The action of G on SG (U) admits a small orbit.
The proof is contained in Theorem 3.12. It confirms a conjecture of
Petrykowski in the case of NIP groups [New12, Conjecture 0.1] and solves
Conjecture 4.13 of [CP12].
Theorem 1.2. Let G = G(U) be a definably amenable NIP group. Then the
following are equivalent for a definable set φ(x):
1. φ(x) does not G-divide (i.e. there is no infinite sequence (gi )i<ω of
elements of G and natural number k such that any k sets in {gi φ(x)}i<ω
have empty intersection, see Definition 3.2);
2. φ(x) is weakly generic (i.e. there is some non-generic ψ(x) such that
φ(x) ∨ ψ(x) is generic, see Definition 3.28);
5
3. µ(φ(x)) > 0 for some G-invariant measure µ;
4. φ(x) is f-generic (meaning that for any small model M over which
φ(x) is defined, no G-translate of φ(x) forks over M, see Definition
3.2);
Moreover, for a global type p ∈ SG (U) the following are equivalent:
1. p is f-generic (i.e. every formula in p is f-generic);
2. p has a small G-orbit;
3. Stab(p) = G00 .
This is given by Theorem 3.35 and Proposition 3.8, and combined with
Theorem 1.1 solves in particular [CP12, Problem 4.13].
We continue by studying the space of G-invariant measures using VCtheory, culminating with a characterization of regular ergodic measures (Section 4) and unique ergodicity (Section 3.4). Generalizing slightly a construction from [HP11], we associate to every generic type p ∈ SG (U) a measure µp ,
which is a lifting of the Haar measure on the compact group G/G00 via p (see
Definition 3.16). It follows from Theorem 1.2 that the supports of the measures µp are exactly the minimal subflows of (SG (U), G(U)) (see Proposition
3.31).
Theorem 1.3. Let G = G(U) be a definably amenable, NIP group. Then
regular ergodic measures on SG (U) are precisely the measures of the form
µp , for p an f-generic type in SG (U). If two such measures have the same
support, then they are equal (i.e., minimal subflows of (SG (U), G(U)) are
uniquely ergodic).
The first statement is Theorem 4.5 and the second follows from Proposition 3.24. The following is Theorem 3.36.
Theorem 1.4. Let G = G(U) be a definably amenable, NIP group. Then
G has a unique invariant measure if and only if it admits a unique minimal
subflow if and only if it admits a global generic type. Moreover, in such a
group all the notions in Theorem 1.2 coincide with “φ(x) is generic”, and in
the moreover part we can add “p is almost periodic”.
6
Next we study enveloping semigroups. This notion from topological dynamics (see [Gla07a]) was introduced in model theory by Newelski [New09].
He observed that it behaved better when one replaced the dynamical system
(SG (M), G(M)) with an extension of it: The set G(M) embeds into SG (U)
as realized types and we let SG (Mext ) be its closure. Then G(M) acts on
SG (Mext ) and this flow admits SG (M) as a factor. We consider the enveloping semigroup E of the dynamical system (SG (Mext ), G(M)). In view of the
results in [CPS14], E can be identified with (SG (Mext ), ·), where · is a naturally defined operation extending multiplication on G(M) (see Section 5.3
for details).
Fix a minimal flow M in (SG (Mext ), G(M)) (i.e. a closed G(M)-invariant
set), and an idempotent u ∈ M. Then general theory of Ellis semigroups implies that uM is a subgroup of E, which we call the Ellis group. The canonical
surjective homomorphism G → G/G00 factors naturally through the space
SG (Mext ), so we have a well-defined continuous surjection π : SG (Mext ) →
G/G00 , tp(g/M) 7→ gG00 , and the restriction of π to the group uM is a
surjective homomorphism. Newelski asked if under certain model-theoretic
assumptions this map could be shown to be an isomorphism. Pillay later
formulated a precise conjecture which we are able to prove here.
Theorem 1.5 (Ellis group conjecture). Let G be definably amenable and
NIP. Then π : uM → G/G00 is an isomorphism.
In particular, this demonstrates that the Ellis group is indeed a model
theoretic object, i.e. it only depends on the first-order theory of the group
and does not depend on the choice of a small model M over which it is
computed. Some special cases of the conjecture were previously known (see
[CPS14]). For the proof, we establish a form of generic compact domination
for minimal flows in definably amenable groups with respect to the Baire
ideal — Theorem 5.3.
Remark 1.6. We remark that the study of NIP definably amenable groups
can be thought of as a model-theoretic version of tame dynamics as studied
by Glasner, Megrelishvili and others, see [Gla07b] (in fact, we discovered
the connection only after having essentially completed this work). The NIP
assumption implies that the dynamical system (SG (M), G(M)) is tame—
and even null—in the sense of [Gla07b], [KL07], but it is not equivalent to
it. Nullness of this system is equivalent to the fact that the definable family
of translates of any given definable set has finite VC-dimension (see [KL07,
7
Proposition 5.4(2)]), whereas the NIP condition implies that any uniformly
defined family of sets has finite VC-dimension.
Acknowledgements
We are grateful to the referees for pointing out a mistake in the previous
version of the article (see Question 3.18) and for other numerous suggestions
on improving the presentation.
2
Preliminaries
In this section we summarize some of the context for our results including the
theory of forking and groups in NIP, along with some general results about
families of sets of finite VC-dimension.
2.1
Combinatorics of VC-families
Let X be a set, finite or infinite, and let F be a family of subsets of X. Given
A ⊆ X, we say that it is shattered by F if for every A ′ ⊆ A there is some
S ∈ F such that A ∩ S = A ′. A family F is said to have finite VC-dimension
if there is some n < ω such that no subset of X of size n is shattered by
F. If this is the case, we let VC(F) be the largest integer n such that some
subset of X of size n is shattered by it.
If S ⊆ X is a subset and x1 , . . . , xn ∈ X, we let Av(x1 , . . . , xn ; S) =
1
|{i
≤ n : xi ∈ S}|. Similarly, if (ti )i<n is a set of truth values, we let
n
Av(ti ) = n1 |{i < n : ti = True}.
Later in the paper, we will often write a ≈ǫ b for |a − b| ≤ ǫ.
A fundamental fact about families of finite VC-dimension is the following
uniform version of the weak law of large numbers ([VC71], see also [HP11,
Section 4] for a discussion).
Fact 2.1. For any k > 0 and ε > 0 there is N < ω satisfying the following.
Let (X, µ) be a probability space, and let F be a family of subsets of X of
VC-dimension ≤ k such that:
1. every set from F is measurable;
8
2. for each n, the function fn : Xn → [0, 1] given by
(x1 , . . . , xn ) 7→ sup | Av(x1 , . . . , xn ; S) − µ(S)|
S∈F
is measurable;
3. for each n, the function gn : X2n → [0, 1]
(x1 , . . . , xn , y1, . . . , yn ) 7→ sup | Av(x1 , . . . , xn ; S) − Av(y1 , . . . , yn ; S)|
S∈F
is measurable.
Then there is some tuple (x1 , . . . , xN ) ∈ XN such that for any S ∈ F we have
|µ(S) − Av(x1 , . . . , xN ; S)| ≤ ε.
The assumptions (2) and (3) are necessary in general (but follow from (1)
if the family F is countable).
Another fundamental fact about VC-families that we will need is the
following theorem about transversal sets due to Matousek. It uses the following definition: a family G of subsets of some set X has the (p, k)- property
if among any p sets in G, some k have non-empty intersection.
Fact 2.2 ([Mat04]). Let F be a family of subsets of some set X. Assume that
F has finite VC-dimension. Then there is some k < ω such that for every
p ≥ k, there is an integer N such that: for every finite subfamily G ⊆ F, if G
has the (p, k)-property, then there is an N-point set intersecting all members
of G.
2.2
Forking in NIP theories
We will use standard notation. We work with a complete theory T in a
language L. We fix a monster model U |= T which is κ-saturated and κstrongly homogeneous for κ a sufficiently large strong limit cardinal.
Recall that a formula φ(x, y) is NIP if the family of subsets {φ(x, a) :
a ∈ U} has finite VC-dimension. The theory T is NIP if all formulas are
NIP. In this paper, we always assume that T is NIP unless explicitly stated
otherwise.
We summarize some facts about forking in NIP theories. Recall that a
set A is an extension base if every type p ∈ S(A) has a global extension
9
non-forking over A. In particular, any model of an arbitrary theory is an
extension base, and every set is an extension base in o-minimal theories,
algebraically closed valued fields or p-adics.
Definition 2.3 ([CK12]).
1. A global type q ∈ S(U) is strictly non-forking
over a small model M if q does not fork over M, and for every B ⊇ M
and a |= q|B , tp(B/aM) does not fork over M.
2. Given q ∈ S(M), we say that (bi : i < κ) is a strict Morley sequence in
q if there is some global extension q ′ ∈ S(U) of q strictly non-forking
over M satisfying bi |= q ′ |Mb<i for all i < κ.
Fact 2.4 ([CK12]). Assume that T is NIP and let A be an extension base.
1. A formula φ(x, a) ∈ L(U) forks over A if and only if it divides over A,
i.e., the set of formulas dividing over A forms an ideal.
2. Every q(y) ∈ S(M) admits a global extension strictly non-forking over
M.
3. Assume that φ(x, b) ∈ L(U) forks (equivalently, divides) over M, and
let (bi : i < κ) in U be an infinite strict Morley sequence in tp(b/M).
Then {φ(x, bi ) : i < κ} is inconsistent.
From now on, we will freely use the equivalence of forking and dividing
over models in NIP theories.
Fact 2.5. (See e.g. [HP11, Proposition].) Assume that T is NIP and M |= T .
A global type p(x) does not fork (equivalently, does not divide) over M if and
only if it is M-invariant. This is, for every φ(x, a) and a ′ ≡M a, we have
p ⊢ φ(x, a) ⇔ p ⊢ φ(x, a ′ ).
Remark 2.6. In particular, in view of Fact 2.4, if π(x) is a partial type that
does not divide over M (e.g. if π(x) is M-invariant), then it extends to a
global M-invariant type.
Let now p(x), q(y) be global types invariant over M. For any set D ⊇ M,
let b |= q|D , a |= p|Db . Then by invariance of p and q, the type tp(ab/D)
does not S
depend on the choice of a, b. Call this type (p ⊗ q)D , and let
p ⊗ q = {(p ⊗ q)D : M ⊆ D ⊆ U small}. Then (p ⊗ q)(x, y) is a welldefined, global invariant type over M.
10
Let p(x) be a global type invariant over M. Then one defines
p(n) (x0 , . . . , xn−1 ) = p(xn−1 ) ⊗ · · · ⊗ p(x0 ),
[
p(ω) (x0 , x1 , . . .) =
p(n) (x0 , . . . , xn−1 ).
n<ω
For any small set D ⊇ M and (ai )i<ω |= p(ω) |D , the sequence (ai )i<ω is
indiscernible over D.
We now discuss Borel-definability. Let p(x) be a global M-invariant type,
pick a formula φ(x, y) ∈ L, and consider the set Sp,φ = {a ∈ U : φ(x, a) ∈ p}.
By invariance, this set is a union of types over M. In fact, it can be written as
a finite boolean combination
V of M-type-definable sets ([HP11]). Specifically,
let Altn (x0 , . . . , xn−1 ) = i<n−1 ¬ (φ(xi , y) ↔ φ(xi+1 , y)) and let An (y) and
Bn (y) be the type-definable subsets of U defined by
∃x0 . . . xn−1 p(n) |M (x0 , . . . , xn−1 ) ∧ Altn (x0 , . . . , xn−1 ) ∧ φ (xn−1 , y)
and
∃x0 . . . xn−1 p(n) |M (x0 , . . . , xn−1 ) ∧ Altn (x0 , . . . , xn−1 ) ∧ ¬φ (xn−1 , y)
respectively.
S
Then for some N < ω, Sp,φ = n<N (An ∧ ¬Bn+1 ).
Note that the set of all global M-invariant types is a closed subset of S(U).
We now consider the local situation. Let φ(x, y) ∈ L be a fixed formula
and let Sφ (U) be the space of all global φ-types (i.e., maximal consistent
collections of formulas of the form φ(x, b), ¬φ(x, b), b ∈ U). Let Invφ (M)
be the set of all global M-invariant φ-types—a closed subset of Sφ (U), which
we equip with the induced topology.
Fact 2.7 ([Sim15b]). Let M be a countable model and let φ(x, y) be NIP.
For any set Z ⊆ Invφ (M) and p ∈ Invφ (M), if p ∈ Z (i.e., in the topological
closure of Z), then p is the limit of a countable sequence of elements of Z.
2.3
Keisler measures
Now we introduce some terminology and basic results around the study of
measures in model theory. A Keisler measure µ(x) (or µx ) over a set of
parameters A is a finitely additive probability measure on the boolean algebra
11
Def x (A) of A-definable subsets of U in the variable x. Alternatively, a Keisler
measure µ(x) may be viewed as assigning a measure to the clopen basis of the
space of types Sx (U). A standard argument shows that it can be extended
in a unique way to a countably-additive regular probability measure on all
Borel subsets of Sx (U) (see e.g. [Sim15a, Chapter 7] for details). From now
on we will just say “measure” unless it could create some confusion.
For a measure µ over A we denote by S(µ) its support: the set of types
weakly random for µ, i.e., the closed set of all p ∈ S(A) such that for any
φ(x), φ(x) ∈ p implies µ(φ(x)) > 0.
Remark 2.8. Let Mx (A) denote the set of measures over A in variable x, it is
naturally equipped with a compact topology as a closed subset of [0, 1]Lx(A)
with the product topology. Every type over A can be identified with the
{0, 1}-measure concentrating on it, thus Sx (A) is identified with a closed subset of Mx (A).
The following implication of Fact 2.1 was observed in [HP11, Section 4].
Fact 2.9. Let T be NIP. Let µ(x) a measure over A, ∆ = {φi (x, yi )}i<m a
finite set of L-formulas, and ε > 0 be arbitrary. Then there are some types
p0 , . . . , pn−1 ∈ Sx (A) such that for every a ∈ A and φ (x, y) ∈ ∆, we have
|µ (φ (x, a)) − Av (p0 , . . . , pn−1 ; φ (x, a))| ≤ ε.
Furthermore, we may assume that pi ∈ S (µ), the support of µ, for all i < n.
Corollary 2.10. Let T be an NIP theory in a countable language L, and
let µ be a measure. Then the support S(µ) is separable (with respect to the
topology induced from S(U)).
Proof. By Fact 2.9, for any finite ∆ ⊆ L and k < ω, we can find some
p∆0 , . . . , p∆n∆ −1 ∈ S(µ) such that for any φ(x, y) ∈ ∆ and any a ∈ U we have
k
S
1
µ(φ(x, a)) ≈ k Av(p∆0 , . . . , p∆n∆−1 ; φ(x, a)). Let S0 = k<ω,∆⊆L finite {p∆i : i <
k
n∆k }. Then S0 is a countable subset of S(µ), and we claim that it is dense.
Let U be a non-empty open subset of S(µ). Then there is some formula
φ(x) ∈ L(U) such that ∅ =
6 φ(x) ∩ S(µ) ⊆ U. In particular µ(φ(x)) > 0,
hence for some k and ∆ large enough we have by the construction of S0 that
necessarily φ(x) ∈ p∆i for at least one i < n∆k .
A measure µ ∈ Mx (U) is non-forking over a small model M if for every
formula φ(x) ∈ L(U) with µ(φ(x)) > 0, φ(x) does not fork over M. A
12
theory of forking for measures in NIP generalizing the previous section from
types to measures is developed in [HP11, HPS13]. In particular, a global
measure non-forking over a small model M is in fact Aut(U/M)-invariant.
Moreover, using Fact 2.9 along with results in Section 2.2 one shows that a
global measure µ invariant over M is Borel definable over M, i.e., for any
φ(x, y) ∈ L the map fφ : Sy (M) → [0, 1], q 7→ µ(φ(x, b)), b |= q is Borel
(and it is well defined by M-invariance of µ). This allows to define a tensor
product of M-invariant measures: given µ ∈ Mx (U), ν ∈ My (U) M-invariant
and φ(x, y) ∈ L(U), let N ⊇ M be some
small model over which φ is defined.
R
We define µ ⊗ ν(φ(x, y)) by taking q∈Sy (N) fφ (q)dν ′ , where ν ′ = ν|N viewed
as a Borel measure on Sy (N). Then µ ⊗ ν is a global M-invariant measure.
We will need the following basic combinatorial fact about measures (see
[HPP08] or [Sim15a, Lemma 7.5]).
Fact 2.11. Let µ be a Keisler measure, φ(x, y) a formula and (bi )i<ω an
indiscernible sequence. Assume that for some ǫ > 0 we have µ(φ(x, bi )) ≥ ǫ
for every i < ω. Then the partial type {φ(x, bi ) : i < ω} is consistent.
2.4
Model-theoretic connected components
Now let G = G(U) be a definable group. Let A be a small subset of U. We
say that H ≤ G has bounded index if |G : H| is smaller than the saturation
of U, and define:
T
• G0A = {H ≤ G : H is A-definable, of finite index}.
T
• G00
{H ≤ G : H is type-definable over A, of bounded index}.
A =
T
• G∞
{H ≤ G : H is Aut (U/A)-invariant, of bounded index}.
A =
∞
Of course G0A ⊇ G00
A ⊇ GA for any A and these are all normal A-invariant
subgroups of G.
Fact 2.12 (see e.g. [Sim15a, Chapter 8] and references therein). Let T be
∞
∞
00
NIP. Then for every small set A we have G0A = G0∅ , G00
A = G∅ , GA = G∅ .
∞
|T |
Moreover, |G/G | ≤ 2 .
We will be omitting ∅ in the subscript and write for instance G00 for G00
∅ .
13
Remark 2.13. It follows that G∞ is equal to the subgroup of G generated by
the set {g−1 h : g ≡M h}, for any small model M.
Let π : G → G/G00 be the canonical projection map.
The quotient G/G00 can be equipped with a natural “logic” topology: a
set S ⊆ G/G00 is closed iff π−1 (S) is type-definable over some (equivalently,
any) small model M.
Fact 2.14 (see [Pil04]). The group G/G00 equipped with the logic topology is
a compact topological group.
Remark 2.15. If L is countable then G/G00 is a Polish space with respect to
the logic topology. Indeed, there is a countable model M such that every
closed set is a projection of a partial type over M, and {π(φ(U))c : φ(x) ∈
L(M)} is a countable basis of the topology.
In particular, G/G00 admits an invariant normalized Haar probability
measure h0 . Furthermore h0 is the unique left-G/G00 -invariant Borel probability measure on G/G00 (see e.g. [Hal50, Section 60]), as well as simultaneously the unique right-G/G00 -invariant Borel probability measure on G/G00 .
The usual completion procedure for a measure preserves G-invariance, so
we may take h0 to be complete.
3
3.1
Generic sets and measures
G-dividing, bounded orbits and definable amenability
Context: We work in an NIP theory T , and let G = G(U) be an ∅-definable
group.
We will consider G as acting on itself on the left. For any model M, this
action extends to an action of G(M) on the space SG (M) of types concentrating on G. Hence if p ∈ SG (M) and g ∈ G(M) we have g·p = tp(g·a/M)
where a |= p. The group G(M) also acts on M-definable subsets of G by
(g · φ)(x) = φ(g−1 · x) and on measures by (g · µ)(φ(x)) = µ(φ(g · x)).
One could also consider the right action of G on itself and obtain corresponding notions. Contrary to the theory of stable groups, this would not
yield equivalent definitions (see Section 6.1 for a discussion).
14
Definition 3.1. The group G is definably amenable if it admits a global
Keisler measure µ on definable subsets of G(U) which is invariant under
(left-) translation by elements of G(U).
As explained for example in [Sim15a, 8.2], if for some model M, there
is a G(M)-invariant Keisler measure on M-definable subsets of G, then G
is definably amenable (it can be seen by taking an elementary extension M
expanded by predicates for the invariant measure).
Definition 3.2.
1. Let φ(x) be a subset of G defined over some model
M. We say that φ(x) (left-)G-divides if there is an M-indiscernible
sequence (gi : i < ω) such that {gi · φ(x) : i < ω} is inconsistent.
2. The formula φ(x) is (left-)f-generic over M if no translate of φ(x)
forks over M. We say that φ(x) is f-generic if it is f-generic over some
small M. A (partial) type is f-generic if every formula implied by it is
f-generic.
3. A global type p is called (left-)strongly f-generic over M if no G(U)translate of p forks over M. A global type p is strongly f-generic if it
is strongly f-generic over some small model M.
Note that we change the usual terminology: our notion of strongly fgeneric corresponds to what was previously called f-generic in the literature
(see e.g. [HP11]). We feel that this change is justified by the development of
the theory presented here.
Note that if µ is a global G-invariant and M-invariant measure and p ∈
S(µ), then p is strongly f-generic over M since all its translates are weaklyrandom for µ. It is shown in [HP11] how to conversely obtain a measure µp
from a strongly f-generic type p. We summarize some of the results from
[HP11] in the following fact.
Recall that the stabilizer of p is StabG (p) = {g ∈ G : g · p = p}.
Fact 3.3.
1. If G admits a strongly f-generic type over some small model
M, then it admits a strongly f-generic type over any model M0 .
2. If p is strongly f-generic then StabG (p) = G00 = G∞ (= h{g−1h : g ≡M
h}i for any small model M).
3. The group G admits a G-invariant measure if and only if there is a
global strongly f-generic type in SG (U).
15
Our first task is to understand basic properties of f-generic formulas and
types.
Proposition 3.4. Let G be a definably amenable group, and let φ(x) ∈
LG (M). Let also p(x) ∈ SG (U) be strongly f-generic, M-invariant and take
g |= p|M . Then the following are equivalent:
1. φ(x) is f-generic over M;
2. φ(x) does not G-divide;
3. g−1 · φ(x) does not fork over M.
Proof. (2) ⇒ (1): Assume that some translate h · φ(x) forks over M. Then
it divides over M, and as φ(x) is over M, we obtain an M-indiscernible
sequence (hi : i < ω) such that {hi · φ(x) : i < ω} is inconsistent. This shows
that φ(x) G-divides.
(1) ⇒ (3): Clear.
(3) ⇒ (2): Assume that φ(x) does G-divide and let (gi : i < ω) be an Mindiscernible sequence witnessing it, i.e., {gi · φ(x) : i < ω} is k-inconsistent
for some k < ω. By indiscernibility, all of gi ’s are in the same G00 -coset, and
00
replacing gi by g−1
for all i.
0 gi+1 , we may assume that gi ∈ G
−1
Let h realize p over (gi )i<ω M. Then gi · h |= p|M by G00 -invariance of
p. As the set {gi · φ(x) : i < ω} is inconsistent, so is {h−1 gi · φ(x) : i < ω}.
Then the sequence (g−1
i · h : i < ω) is an M-indiscernible sequence in p|M =
tp(g/M) (as tp(h/(gi )i<ω M) is M-invariant). Therefore g−1 · φ(x) divides
over M.
Note that we do not say “G-divides over M”, because the model M does
not matter in the definition: for any M ≺ N, an M-definable φ(x) G-divides
over M if and only if it G-divides over N. Therefore the same is true for
f-genericity (i.e. if φ(x) is both M-definable and N-definable, then it is fgeneric over M if and only if it is f-generic over N) and from now on we will
just say f-generic, without specifying the base.
Corollary 3.5. Let G be definably amenable. The family of non-f-generic
formulas (equivalently, G-dividing formulas) forms an ideal. In particular,
every partial f-generic type extends to a global one.
Proof. Assume that φ(x), ψ(x) are not f-generic, and let M be some small
model over which both formulas are defined. Let also p be a global type
strongly f-generic over M (exists by Fact 3.3) and take g |= p|M . Then by
Fact 3.4(3) we have that both g−1 · φ(x), g−1 · ψ(x) fork over M, in which
16
case g−1 · (φ(x) ∨ ψ(x)) = g−1 · φ(x) ∨ g−1 · ψ(x) also forks over M. Applying
Fact 3.4(3) again it follows that φ(x) ∨ ψ(x) is not f-generic.
The “in particular” statement follows by compactness.
Lemma 3.6. Let G be definably amenable, let φ(x) ∈ LG (U) be a formula
and g ∈ G00 . Then φ(x)△g · φ(x) is not f-generic (and hence it G-divides
by Proposition 3.4).
Proof. Let M be a model over which φ(x) and g are defined. Let p ∈
SG (U) be a global strongly f-generic type which is M-invariant (exists by
Fact 3.3(1)) and let h realize p over Mg. Then h−1 · (φ(x)△g · φ(x)) =
(h−1 · φ(x))△(h−1 g · φ(x)). Since h ≡M g−1 h (as g−1 ∈ StabG (p) by Fact
3.3(2)), the latter formula cannot belong to any global M-invariant type,
and so it must fork over M by Remark 2.6. Hence φ(x)△g · φ(x) is not
f-generic.
Definition 3.7. A global type p(x) ∈ SG (U) has a bounded orbit if |G·p| < κ
for some strong limit cardinal κ such that U is κ-saturated.
Proposition 3.8. Let G be definably amenable. For p ∈ SG (U), the following
are equivalent:
1. p is f-generic,
2. p is G00 -invariant (and StabG (p) = G00 ),
3. p has a bounded orbit.
Proof. (1) ⇒ (2): If p is not G00 -invariant then φ(x)△gφ(x) ∈ p for some
g ∈ G00 , φ(x) ∈ LG (U), and so p is not f-generic by Lemma 3.6. Hence
G00 ⊆ StabG (p). Given an arbitrary a ∈ StabG (p), let M be a small model
containing a and let b |= p|M . Then a · b |= p|M , hence a = (a · b) · b−1 and
a · b ≡M b. By Fact 3.3(2) it follows that a ∈ G00 , hence StabG (p) = G00 .
(2) ⇒ (3): If p is G00 -invariant, then the size of its orbit is bounded by
the index of G00 (which is ≤ 2|T | ).
(3) ⇒ (1): If p is not f-generic, then some φ(x) ∈ p must G-divide (by
Proposition 3.4). Then, as in the proof of Proposition 3.4, we can find an
arbitrarily long indiscernible sequence (gi )i<λ in G00 such that {gi φ(x) : i < λ}
is k-inconsistent for some k < ω, which implies that the G-orbit of p is
unbounded.
17
Next we clarify the relationship between f-generic and strongly f-generic
types in definably amenable groups.
Proposition 3.9. Let G be definably amenable. A type p ∈ SG (U) is strongly
f-generic if and only if it is f-generic and M-invariant over some small model
M.
Proof. Strongly f-generic implies f-generic is clear.
Conversely, assume that p is M-invariant, but not strongly f-generic over
M. Then g · p divides over M for some g ∈ G. It follows that there is some
φ (x, a) ∈ p such that for any κ there is some M-indiscernible sequence
(gi^ai )i<κ with g0^a0 = g^a and such that {gi · φ (x, ai )}i<κ is k-inconsistent
for some k < ω. By M-invariance of p we have that φ (x, ai ) ∈ p, so
{gi · p (x)}i<κ is k-inconsistent. This implies that the orbit of p is unbounded,
and that p is not f-generic in view of Proposition 3.8.
Example 3.10. There are f-generic types which are not strongly f-generic.
Let R be a saturated model of RCF. We give an example of a G-invariant (and
so f-generic by Proposition 3.8) type in G = (R2 ; +) which is not invariant
over any small model (and so not strongly f-generic by Proposition 3.9).
Let p(x) ∈ S1 (R) denote the definable 1-type at +∞ and q(y) ∈ S1 (R) a
global 1-type which is not invariant over any small model (hence corresponds
to a cut of maximal cofinality from both sides). Then p and q are weakly
orthogonal types. Let (a, b) |= p × q (in some bigger model) and consider
r := tp(a, a + b/R). Then r ∈ SG (R) is a G-invariant type which is not
invariant over any small model.
The following lemma is standard.
Lemma 3.11. Let N ≻ M be |M|+ -saturated, and let p ∈ SG (N) be such
that g · p does not fork over M for every g ∈ G(N). Then p extends to a
global type strongly f-generic over M.
Proof. It is enough to show that
p(x) ∪ {¬(g · φ(x, a)) : g ∈ G(U), φ(x, a) ∈ L(U) forks over M}
W
is consistent. Assume not, then p(x) ⊢ i<n gi · φi (x, ai ) for some gi ∈ G(U),
φi (x, y) ∈ L and ai ∈ U such that φi (x, ai ) forks over M. By |M|+ -saturation
of N and compactness we can find some (gi′ , ai′)i<n ≡M (gi , ai )i<n in N such
18
W
that p(x) ⊢ i<n gi′ · φi (x, ai′), which implies that gi′ · φi (x, ai′) ∈ p for some
i < n, i.e., (gi′)−1 · p forks over M. But this contradicts the assumption on
p.
Finally for this subsection, we prove Theorem 1.1: for NIP groups, definable amenability is characterized by the existence of a type with a bounded
orbit, proving Petrykowski’s conjecture for NIP theories (see [New12, Conjecture 0.1]). In fact, existence of a measure with a bounded orbit is sufficient.
Theorem 3.12. Let T be NIP, U |= T and G = G(U) a definable group.
Then the following are equivalent:
1. G is definably amenable;
2. |G · p| ≤ 2|T | for some p ∈ SG (U);
3. some measure µ ∈ MG (U) has a bounded G-orbit.
Proof. (1) ⇒ (2): If G is definably amenable, then it has a strongly f-generic
type p ∈ SG (U) by Fact 3.3 and such a type is G00 -invariant. In particular
its orbit has size at most |G/G00 | ≤ 2|T | .
(2) ⇒ (3) is obvious.
(3) ⇒ (1): Assume that |Gµ| < κ, with κ strong limit and U is κsaturated. Let M be a model with |M| = |T |, let N0 ≻ M be an |M|+ saturated submodel of U of size 2|M| < κ (exists as κ is a strong limit cardinal),
and let (Ni )i<κ be a strict Morley sequence in tp (N0 /M) contained in U
(exists by κ-saturation of U and Fact 2.4(2)). In particular Ni is an |M|+ saturated extension of M for all i < κ.
Let µi = µ|Ni . It is enough to show that for some i < κ, the measure gµi
does not fork over M for any g ∈ G(Ni ), as then any type in the support of
µi extends to a global type strongly f-generic over M by Lemma 3.11, and
we can conclude by Fact 3.3.
Assume not, then for each i < κ we have some gi ∈ G (Ni ) and some
φi (x, ci ) ∈ L(Ni ) such that gi µi (φ(x, ci )) > 0 but φi (x, ci ) forks over M.
As the orbit of µ is bounded, by throwing away some i’s we may assume
that there is some g ∈ G such that gi µ = gµ for all i < κ, in particular
(gµ)|Ni = gi µi . By pigeonhole and the assumption on κ we may assume also
that there are some φ(x, y) ∈ L and ε > 0 such that φi (x, yi ) = φ(x, y)
and gµ(φ(x, ci)) > ε for all i < κ, and that the sequence (ci : i < κ) is
indiscernible (i.e. the ci ’s occupy the same place in the enumeration of Ni ,
19
for all i, and the sequence (Ni )i<κ is indiscernible by construction). Applying
Fact 2.11 to the measure gµ we conclude that {φ(x, ci ) : i < κ} is consistent.
But as (ci ) is a strict Morley sequence, this contradicts the assumption that
φ(x, ci ) divides over M for all i, in view of Fact 2.4(3).
Remark 3.13.
1. In the special case of types in o-minimal expansions of
real closed fields this was proved in [CP12, Corollary 4.12].
2. Theorem 3.12 also shows that the issues with absoluteness of the existence of a bounded orbit considered in [New12] do not arise when one
restricts to NIP groups.
3.2
3.2.1
Measures in definably amenable groups
Construction
Again, we are assuming throughout this section that G = G(U) is an NIP
group. We generalize the connection between G-invariant measures and
strongly f-generic types from Fact 3.3 to f-generic types in definably amenable
groups.
First we generalize Proposition 3.8 to measures.
Proposition 3.14. Let G be definably amenable, and let µ be a Keisler
measure on G. The following are equivalent:
1. The measure µ is f-generic, that is µ(φ(x)) > 0 implies φ(x) is fgeneric for all φ(x) ∈ LG (U).
2. All types in the support S(µ) are f-generic.
3. The measure µ is G00 -invariant.
4. The orbit of µ is bounded.
Proof. The equivalence of (1) and (2) is clear by compactness, (1) implies
(3) is immediate by Lemma 3.6, and (3) implies (4) as the size of the orbit
of a G00 -invariant measure is bounded by |G/G00 |.
(4) ⇒ (1): Assume that we have some G-dividing φ(x) with µ(φ(x)) >
ε > 0. As in the proof of Proposition 3.4, (3) ⇒ (2) we can find an arbitrarily
long indiscernible sequence (gi )i∈λ with gi ∈ G00 such that {gi φ(x)} is kinconsistent, for some fixed k < ω.
20
In view of Fact 2.11 for any fixed i < λ there can be only finitely many
j < λ such that gi µ(gj φ(x)) > ε. But gi µ(gjφ(x)) = g−1
j gi µ(φ(x)). This
implies that gi µ 6= gj µ for all but finitely many j < λ, which then implies
that the orbit of µ is unbounded.
In [HP11, Proposition 5.6] it is shown that one can lift the Haar measure
on G/G00 to a global G-invariant measure on all definable subsets of an NIP
group G using a strongly f-generic type. We point out that in a definably
amenable NIP group, an f-generic type works just as well. For this we need
a local version of the argument used there.
Fix a small model M, and let FM be the set of formulas of the form
g · φ (x) or ¬g · φ(x), for g ∈ G (U) , φ (x) ∈ LG (M).
Proposition 3.15. Let G be definably amenable, and let p be a maximal
finitely consistent set of formulas in FM . Then p is f-generic if and only if
g · p is M-invariant for every g ∈ G.
Proof. Notice that g·p(x) is also a set of formulas in FM . Assume that g·p (x)
is not M-invariant. Then gp ⊢ g0 φ (x) △g1 φ (x) for some φ (x) ∈ L (M)
−1
−1
00
(by Fact
and g0 ≡M g1 . Hence g−1
1 gp ⊢ g1 g0 φ(x)∆φ(x) and g1 g0 ∈ G
−1
3.3(2)). Then (g1 g0 )φ(x)△φ(x) is not f-generic by Lemma 3.6, and so p is
not f-generic — a contradiction.
Conversely, assume that some formula ψ(x) implied by p(x) is not fgeneric. Let N ⊇ M contain the parameters of ψ. Then there is some
(hi )i<ω indiscernible over N such that {hi ψ(x)}i<ω is k-inconsistent. Then
h0 ψ(x) ∈ h0 p, but hi ψ(x) ∈
/ h0 (p) for some i < ω. So h0 p is not Ninvariant, and thus also not M-invariant.
Definition 3.16. Let G = G(U) be definably amenable, and let p ∈ SG (U)
be f-generic. Keeping in mind that p (as well as all its translates) is G00 invariant (by Proposition 3.8), we define a measure µp on G by:
µp (φ(x)) = h0 ({ḡ ∈ G/G00 : φ(x) ∈ g · p}),
where h0 is the normalized Haar measure on the compact group G/G00 and
ḡ = g/G00 .
We have to check that this definition makes sense, that is that the set
we take the measure of is indeed measurable. Let M be a small model over
which φ(x) is defined. Let pM be the restriction of p to formulas from FM (as
21
defined above). By Proposition 3.15, pM is M-invariant. It follows that pM
extends to some complete M-invariant type (by Remark 2.6). Then we can
use Borel-definability of invariant types (applied to the family of all translates
of φ(x)) exactly as in [HP11, Proposition 5.6] to conclude.
Remark 3.17.
1. The measure µp that we just constructed is clearly Ginvariant and G00 -strongly invariant (that is, µp(φ(x)△g · φ(x)) = 0
for g ∈ G00 ). Besides µp = µgp for any g, p.
2. We have S(µp) ⊆ G · p. Indeed, if q ∈ S(µp) and φ(x) ∈ q arbitrary,
then µp (φ(x)) > 0, which by the definition of µp implies that g · p ⊢
φ(x) for some g ∈ G.
Question 3.18. 1 Let G = G(U) be an NIP group. Are the following two
properties equivalent?
1. G is definably amenable.
2. G admits a global f-generic type (equivalently, the family of all non-fgeneric subsets of G is an ideal).
3.2.2
Approximation lemmas
Throughout this section, G = G(U) is a definably amenable NIP group.
Given a G00 -invariant type p(x) ∈ SG (U) and a formula φ(x) ∈ LG (U), let
Aφ,p := {ḡ ∈ G/G00 : φ(x) ∈ ḡ · p}.
Note that Ag·φ,p = ḡ · Aφ,p and Aφ,g·p = Aφ,p · ḡ−1 , where ḡ is the image
of g in G/G00 .
Lemma 3.19. For a fixed formula φ(x, y), let Aφ ⊆ P(G/G00 ) be the family
of all Aφ(x,b),p where b varies over U and p varies over all f-generic types on
G. Then Aφ has finite VC-dimension.
Proof. Let ḡ0 , . . . , ḡn−1 be shattered by Aφ . Then for any I ⊆ n there is some
Aφ(x,bI ),pI which cuts out that subset. Take representatives g0 , . . . , gn−1 ∈ G
of the ḡi ’s. Let aI |= pI |g0 ,...,gn−1 bI , then we have φ(gi aI , bI) if and only if i ∈ I.
Hence the VC-dimension of Aφ is at most that of ψ(u; x, y) = φ(ux, y), so
finite by NIP.
1
We have claimed an affirmative answer in an earlier version of this article, however a
mistake in our argument was pointed out by the referees.
22
Replacing the formula φ(x; y) by φ ′ (x; y, u) := φ(u−1 · x; y), we may
assume that any translate of an instance of φ is again an instance of φ. Note
also that then for any parameters a, b we have
ḡ1 Aφ ′ (x;a,b),p ḡ2 = Ag1 φ ′ (x;a,b),g−1 p = Aφ ′ (x;a ′ ,b ′ ),g−1 p
2
2
for some a ′ , b ′. Using this and applying Lemma 3.19 to φ ′ (x; y, u), we get
the following corollary.
Corollary 3.20. For any φ(x, y) ∈ LG (U), the family
Fφ = {ḡ1 · Aφ(x,b),p · ḡ2 : ḡ1 , ḡ2 ∈ G/G00 , b ∈ U, p ∈ SG (U) f-generic}
has finite VC-dimension.
We would now like to apply the VC-theorem to Fφ . This requires verifying an additional technical hypothesis (assumptions (2) and (3) in Fact
2.1), which we are only able to show for certain (sufficiently representative)
subfamilies of Fφ .
Fix φ(x) ∈ LG (U) and let S be a set of global f-generic types. Let
Fφ,S := ḡ1 · Aφ(x),p · ḡ2 : ḡ1 , ḡ2 ∈ G/G00 , p ∈ S .
Lemma 3.21. If S is countable and L is countable, then Fφ,S satisfies all of
the assumptions of Fact 2.1 with respect to the measure h0 .
Proof. First of all, the family of sets Fφ,S has finite VC-dimension by Corollary 3.20 and the obvious inclusion Fφ,S ⊆ Fφ .
Next, (1) is satisfied by the assumption that S consists of f-generic types
and an argument as in the discussion after Definition 3.16 (using countability
of the language).
For a set S ′ of global f-generic types, let
fS ′ ,n (x0 , . . . , xn−1 ) := sup {| Av(x0 , . . . , xn−1 ; Y) − h0 (Y)|},
Y∈Fφ,S ′
gS ′ ,n (x0 , . . . , xn−1 , y0, . . . , yn−1 ) := sup {| Av(x0 , . . . , xn−1 ; Y)−
Y∈Fφ,S ′
− Av(y0 , . . . , yn−1 ; Y)|}.
For (2) and (3) we need to show that fS,n and gS,n are measurable for
all n < ω. Note that fS,n = supp∈S f{p},n and gS,n = supp∈S g{p},n . Since S is
23
countable, it is enough to show that for a fixed f-generic type p the functions
fn := f{p},n and gn := g{p},n are measurable.
Let A = Aφ,p . By G/G00 -invariance of h0 both on the left and on the
right, we have:
fn (x0 , . . . , xn−1 ) =
max
ḡ1 ,ḡ2 ∈G/G00
| Av(x0 , . . . , xn−1 ; ḡ1 · A · ḡ2 ) − h0 (A)|
and
gn (x0 , . . . , xn−1 , y0 , . . . , yn−1 ) =
max
ḡ1 ,ḡ2 ∈G/G00
| Av(x0 , . . . , xn−1 ; ḡ1 · A · ḡ2 )−
− Av(y0 , . . . , yn−1 ; ḡ1 · A · ḡ2 )|.
Then it is enough to show that for a fixed I ⊆ n, the set
AI = {(x0 , . . . , xn−1 ) ∈ (G/G00 )n : for some ḡ1 , ḡ2 ∈ G/G00 ,
xi ∈ ḡ1 · A · ḡ2 ⇐⇒ i ∈ I}
is measurable. But we can write AI as the projection of AI′ ⊆ (G/G00 )n+2
−1
where AI′ is the intersection of {(ḡ1 , ḡ2, x0 , . . . , xn−1 ) : ḡ−1
1 xi ḡ2 ∈ A} for i ∈ I
−1
−1
/ A} for i ∈
/ I. As group multiplication
and {(ḡ1 , ḡ2, x0 , . . . , xn−1 ) : ḡ1 xi ḡ2 ∈
is continuous and A is Borel, those sets are Borel as well. Hence AI is
analytic. Now G/G00 is a Polish space (as L is countable, by Remark 2.15)
and analytic subsets of Polish spaces are universally measurable (see e.g.
[Kec95, Theorem 29(7)]). In particular they are measurable with respect to
the complete Haar measure h0 .
The next lemma will allow us to reduce to a countable sublanguage.
Lemma 3.22. Let L0 be a sublanguage of L, T0 the L0 -reduct of T , G an L0 definable group definably amenable (in the sense of T ) and φ(x) a formula
from L0 (U). Let p ∈ SG (U) be a global L-type which is f-generic, and let
p0 = p|L0 .
1. In the sense of T0 , the group G is definably amenable NIP and p0 is an
f-generic type.
2. Let G00
L0 be the connected component computed in T0 , and let µp0 (µp )
be the G-invariant measure on L0 -definable (resp., L-definable) subsets
of G given by Definition 3.16 in T0 (resp., in T ). Then µp (φ(x)) =
µp0 (φ(x)).
24
Proof. (1) The first assertion is clear. Similarly, it is easy to see that if
ψ(x) ∈ L0 is G-dividing in T0 , then it is G-dividing in T (by extracting
an L-indiscernible sequence from an L0 indiscernible sequence). Then p0 is
f-generic by Fact 3.4 applied in T0 .
(2) Let A = {ḡ ∈ G/G00 : g · p ⊢ φ(x)} and A0 = {ḡ ∈ G/G00
L0 : g ·
p0 ⊢ φ(x)}, then by definition µp (φ(x)) = h0 (A) and µp0 (φ(x)) = h0′ (A0 ),
where h0 is the Haar measure on G/G00 and h0′ is the Haar measure on
00
00
G/G00
→ G/G00
7→ g/G00
L0 . The map f : G/G
L0 , g/G
L0 is a surjective group
homomorphism, and it is continuous with respect to the logic topology. Note
that for any g ∈ G we have g · p0 ⊢ φ(x) ⇐⇒ g · p ⊢ φ(x), so A = f−1 (A0 ).
Let h∗0 = f∗ (h0 ) be the push-forward measure, it is an invariant measure on
∗
′
G/G00
L0 . But by the uniqueness of the Haar measure, it follows that h0 = h0 ,
∗
′
and so h0 (A) = h0 (A0 ) = h0 (A0 ), i.e., µp (φ(x)) = µp0 (φ(x)) as wanted.
Proposition 3.23. For any φ(x) ∈ LG (U), ε > 0 and a countable set of
f-generic types S ⊆ SG (U) there are some g0 , . . . , gn−1 ∈ G such that: for
any g, g ′ ∈ G and p ∈ S we have µgp (g ′φ(x)) ≈ε Av(gj g ′ φ(x) ∈ gp).
Proof. First assume that the language L is countable. Using Lemma 3.21, we
can apply the VC-theorem (Fact 2.1) to the family F ′ = Fφ,S and find some
g0 , . . . , gn−1 ∈ G/G00 such that for any Y ∈ F ′ we have Av(g0 , . . . , gn−1 ; Y) ≈ε
h0 (Y). Let gi ∈ G be some representative of gi , for i < n. Let g, g ′ ∈ G and
p ∈ S be arbitrary. Recall that µgp(g ′ φ(x)) = h0 (Ag ′ φ,gp ) and that Ag ′ φ,gp =
g ′ Aφ,p g−1 , where g = g/G00 , g ′ = g ′ /G00 . Then Ag ′ φ,gp ∈ F ′ and we have
−1
′
′
µgp(g ′ φ(x)) ≈ε Av(g0 , . . . , gn−1 ; Ag ′φ,gp ) = Av(g−1
0 g φ(x), . . . , gn−1 g φ(x); gp).
Now let L be an arbitrary language, and let L0 be an arbitrary countable
sublanguage such that φ(x) ∈ L0 and G is L0 -definable, let T0 be the corresponding reduct. Let S0 = {p|L0 : p ∈ S}, by Lemma 3.22 it is a countable
set of f-generic types in the sense of T0 . Applying the countable case with
respect to S0 inside T0 , we find some g0 , . . . , gn−1 ∈ G such that for any
g, g ′ ∈ G and p0 ∈ S0 we have µgp0 (g ′ φ(x)) ≈ε Av(gjg ′ φ(x) ∈ gp0 ). Let
p ∈ S be arbitrary, and take p0 = p|L0 . On the one hand, the right hand
side is equal to Av(gj g ′ φ(x) ∈ gp). On the other hand, as g ′ φ(x) ∈ L0 (U)
and gp0 = gp|L0 is f-generic, by Lemma 3.22 the left hand side is equal to
µgp(g ′ φ(x)), as wanted.
Proposition 3.24. Let p be an f-generic type, and assume that q ∈ G · p.
Then q is f-generic and µp = µq .
25
Proof. First of all, q is f-generic because the orbit of p consists of f-generic
types and the set of f-generic types is closed.
Take a formula φ(x) ∈ LG (U) and ε > 0, and let g0 , . . . , gn−1 be as given
by Proposition 3.23 for S = {p, q}. Then we have µq (φ(x)) ≈ε Av(gi φ(x); q).
As q ∈ G · p, there is some g ∈ G such that for each i < n we have gi φ(x) ∈
q ⇐⇒ gi φ(x) ∈ gp. But we also have µgp (φ(x)) ≈ε Av(gi φ(x); gp), which
together with µgp = µp implies µp (φ(x)) ≈2ε µq (φ(x)). As φ(x) and ε were
arbitrary, we conclude.
Proposition 3.25. Let p be an f-generic type. Then for any definable set
φ(x), if µp (φ(x)) > 0, then there is a finite union of translates of φ(x) which
covers the support S(µp) (so in particular has µp-measure 1).
Proof. As S(µp) ⊆ G · p (Remark 3.17), any type q weakly random for µp is
f-generic and satisfies µq = µp by Proposition 3.24. Hence µq (φ(x)) > 0, so
some translate of φ(x) must be in q. It follows that the closed compact set
S(µp) can be covered by translates of φ, so by finitely many of them.
Lemma 3.26. Let µ be G-invariant. Then for any ε > 0 and φ(x, y), there
are some f-generic p0 , . . . , pn−1 ∈ S(µ) such that
µ(φ(x, b)) ≈ε
1X
µp (φ(x, b))
n i<n i
for any b ∈ U.
Proof. As before, we may assume that every translate of an instance of
φ(x, y) is an instance of φ(x, y). Fix ε > 0.
By Fact 2.9 there are some p0 , . . . , pn−1 ∈ S(µ) such that µ(φ(x, b)) ≈ε
Av(φ(x, b) ∈ pi ) for all b ∈ U. It follows by G-invariance of µ and the
assumption on φ that for any g ∈ G and b ∈ U, Av(gφ(x, b) ∈ pi ) ≈ε
µ(φ(x, b)).
By Proposition 3.14, all of the pi ’s are f-generic. By Proposition 3.23
with S = {p0 , . . . , pn−1 }, for every b ∈ U there are some g0 , . . . , gm−1 ∈ G
such that for any i < n, µpi (φ(x, b)) ≈ε Av(gj φ(x, b) ∈ pi ).
So let b ∈ U be arbitrary, and choose the corresponding g0 , . . . , gm−1 for
it. By the previous remarks we have
1X
1X
µpi (φ(x, b)) ≈ε
Av(gjφ(x, b) ∈ pi ) =
n i<n
n i<n
26
1X
=
n i<n
!
1 X
1 X
"gj φ(x, b) ∈ pi " =
m j<m
m j<m
!
1X
"gj φ(x, b) ∈ pi " =
n i<n
1 X
1 X
Av(gj φ(x, b) ∈ pi ) ≈ε
µ(φ(x, b)) = µ(φ(x, b)).
m j<m
m j<m
P
Thus µ(φ(x, b)) ≈2ε n1 i<n µpi (φ(x, b)).
=
Corollary 3.27. Let µ be a G-invariant measure and assume that S(µ) ⊆
G · p for some f-generic p. Then µ = µp .
Proof. Let φ(x) ∈ LG (U) and ε > 0 be arbitrary. By Lemma 3.26 we can find
some f-generic p0 , . . . , pn−1 ∈ S(µ) such that µ(φ(x)) ≈ε Av(µpi (φ(x)) : i <
n). But as pi ∈ S(µ) ⊆ G · p, it follows by Proposition 3.24 that µpi = µp
for all i < n, so µ(φ(x)) ≈ε µp (φ(x)).
3.3
Weak genericity and almost periodic types
Now we return to the notions of genericity for definable subsets of definable groups and add to the picture another one motivated by topological
dynamics, due to Newelski.
We will be using the standard terminology from topological dynamics:
Given a group G, a G-flow is a compact space X equipped with an action
of G such that every x 7→ g · x, g ∈ G is a homeomorphism of X. We will
usually write a G-flow X as a pair (G, X). A set Y ⊆ X is said to be a subflow
if Y is closed and G-invariant. The flows relevant to us are (SG (U), G(U))
and (SG (M), G(M)) for a small model M.
Definition 3.28 ([New09, Poi87]).
1. A formula φ(x) ∈ LG (U) is (left)generic
S if there are some finitely many g0 , . . . , gn−1 ∈ G such that
G = i<n gi φ(x).
2. A formula φ(x) ∈ LG (U) is (left-)weakly generic if there is formula ψ(x)
which is not generic, but such that φ(x) ∨ ψ(x) is generic.
3. A (partial) type is (weakly) generic if it only contains (weakly) generic
formulas.
4. A type p ∈ SG (U) is called almost periodic if it belongs to a minimal
flow in (SG (U), G(U)) (i.e., a minimal G-invariant closed set), equivalently if for any q ∈ G · p we have G · p = G · q.
27
Fact 3.29 ([New09], Section 1). The following hold, in an arbitrary theory:
1. The formula φ(x) is weakly generic if and only if for some finite A ⊆ G,
X \ (A · φ(x)) is not generic.
2. The set of non weakly generic formulas forms a G-invariant ideal. In
particular, there are always global weakly generic types by compactness.
3. The set of all weakly generic types is exactly the closure of the set of
all almost periodic types in SG (U).
4. Every generic type is weakly generic. Moreover, if there is a global
generic type then every weakly generic type is generic, and the set of
generic types is the unique minimal flow in (SG (U), G(U)).
5. A type p(x) is almost periodic if and only if for every φ(x) ∈ p, the set
G · p is covered by finitely many left translates of φ(x).
We connect these definitions to the notions of genericity from the previous
sections. As before, we always assume that G = G(U) is NIP.
Proposition 3.30. Let G be definably amenable and let φ(x) ∈ LG (M) be a
weakly generic formula. Then it is f-generic.
Proof. We adapt the argument from [NP06, Lemma 1.8]. As φ(x) is weakly
generic, let ψ(x) be non-generic and A ⊂ G a finite set such that A · (φ(x) ∨
ψ(x)) = X. We may assume that A ⊂ M and that ψ(x) is defined over M.
Assume that φ(x) is not f-generic over M. The set of formulas which are not
f-generic is G-invariant, and moreover it is an ideal by Corollary 3.5. Thus
A · φ(x) is not f-generic, which implies that there is some g ∈ G such that
g · A · φ(x) divides
T over M. That is, there is an M-indiscernible sequence
(gi )i<k such that i<k gi · A · φ(x) = ∅.
As A · φ(x) ∪ A · ψ(x) S
= G, we also have gTi · A · φ(x) ∪ gi · A · ψ(x) = G
for every i < k. Thus G \ i<k gi · A · ψ(x) ⊆ i<k gi · A · φ(x) = ∅. But this
means that ψ(x) is generic, a contradiction.
Proposition 3.31. Assume that G is definably amenable.
1. If p is almost periodic then it is f-generic and G · p = S(µp ).
2. Minimal flows in SG (U) are exactly the sets of the form S(µp) for some
f-generic p.
28
3. If p, q are almost periodic and µp = µq then G · p = G · q.
Proof. (1) An almost periodic type p contains only weakly generic formulas
and hence is f-generic by Proposition 3.30. As S(µp) ⊆ G · p (see Remark
3.17), it follows by minimality that S(µp) = G · p.
(2) For an f-generic p, the set S(µp) is a subflow by G-invariance of µp .
If q ∈ S(µp ) and φ(x) ∈ q, then µp (φ(x)) > 0 and by Proposition 3.25 there
are finitely many translates of φ(x) which cover S(µp), so in particular they
cover G · q ⊆ S(µp). Thus q is almost periodic (by the usual characterization
of almost periodic types from Fact 3.29(5)).
(3) is clear.
In particular, for any f-generic type p there is some almost periodic type
q with µp = µq . However, the following question remains open 2 .
Question 3.32. Is every f-generic type almost periodic? Equivalently, does
p ∈ S(µp) always hold?
Now towards the converse.
Proposition 3.33. Let G be definably amenable. Assume that φ (x) does
not G-divide. Then there are some global almost periodic types p0 , . . . , pn−1 ∈
SG (U) such that for any g ∈ G there is some i < n such that gφ(x) ∈ pi
holds.
Proof. Let k ∈ ω be as given by Fact 2.2 for the VC-family F = {gφ (x) :
g ∈ G}. We claim that F satisfies the (p, k)-property for some p < ω. If not,
then by compactness we can find an infinite indiscernible sequence (gi )i<ω in
G such that {gi φ(x) : i < ω} is k-inconsistent, and so G-divides.
By Fact 2.2 and compactness it follows that there are some p0 , . . . , pN−1 ∈
SG (U) which satisfy:
(∗) for every g ∈ G, for some i < N, we have gφ (x) ∈ pi .
Now consider the action of G on (SG (U))N with the product topology, and
let
F = {g · (p0 , . . . , pN−1 ) : g ∈ G}.
It is a subflow, and besides every (q0 , . . . , qN−1 ) ∈ F satisfies
V (∗) (it is clear for
translates of (p0 , . . . , pN−1 ); if for some g ∈ G we have i<N ¬g · φ (xi ) ∈ qi ,
2
While this paper was under review, a negative answer was obtained in [PY16].
29
V
then since i<N ¬g · φ (xi ) is an open subset of (SG (U))N with respect to the
product topology containing (q0 , . . . , qN−1 ), it follows that h · (p0 , . . . , pN−1 )
belongs to it for some h ∈ G, which is impossible). Let F ′ be a minimal
subflow of F, and notice that the projection of F ′ on any coordinate is a
minimal subflow of (G, SG (U)). Thus, taking (q0 , . . . , qN−1 ) ∈ F ′ , it follows
that qi is almost periodic for every i < N, and every translate of φ (x)
belongs to one of the qi , i < N.
Corollary 3.34. Let G be definably amenable. If φ(x) is f-generic, then
µq (φ(x)) > 0 for some global f-generic type q.
Proof. Let p0 , . . . , pn−1 be some global almost periodic types given by Propo00
sition 3.33, they are
S also f-generic00by Proposition 3.31. Let Yi = {ḡ ∈ G/G :
gφ(x) ∈ pi }. As i<n Yi = G/G and each of Yi ’s is measurable, it follows
that h0 (Yi ) ≥ n1 for some i < n. But then µpi (φ(x)) ≥ n1 .
Summarizing, we have demonstrated that all notions of genericity that
we have considered coincide in definable amenable NIP groups.
Theorem 3.35. Let G be definably amenable, NIP. Let φ(x) be a definable
subset of G. Then the following are equivalent:
1. φ(x) is f-generic;
2. φ(x) is not G-dividing;
3. φ(x) is weakly-generic;
4. µ(φ(x)) > 0 for some G-invariant measure µ;
5. µp (φ(x)) > 0 for some global f-generic type p.
Proof. (1) and (2) are equivalent by Proposition 3.4, (1) implies (3) by Proposition 3.33 and (3) implies (1) by Proposition 3.30. Finally, (1) implies (5)
by Corollary 3.34, (5) implies (4) is obvious and (4) implies (1) by Lemma
3.14.
30
3.4
Unique ergodicity
We now characterize the case when G admits a unique G-invariant measure. Following standard terminology in topological dynamics, we call such
a G uniquely ergodic (indeed, it will follow from the next section that this
condition is equivalent to SG (U) having a unique regular ergodic measure).
Recall that a G-invariant measure µ is called generic if for any definable
set φ(x), µ(φ(x)) > 0 implies that φ(x) is generic. It follows that any
p ∈ S(µ) is generic.
Theorem 3.36. A definably amenable NIP group G is uniquely ergodic if
and only if it admits a generic type (in which case it has a unique minimal
flow — the support of the unique measure).
Proof. If G admits a generic type p, then
S for any type q, p belongs to the
closure G · q (if φ(x) ∈ p then X = i<n gi · φ(x) for some gi ∈ G, so
φ(x) ∈ g−1
i q for some i < n). In particular, for an arbitrary f-generic type
q we have µq = µp (by Proposition 3.24). By Lemma 3.26, this implies that
any invariant measure µ is equal to µp , hence there is a unique invariant
measure.
Conversely, assume that G admits a unique G-invariant measure µ. We
claim that µ is generic. Assume not, and let φ(x) be a definable set of positive
µ-measure and
W assume that φ(x) is not generic. Then for any g1 , . . . , gn ∈ G,
the union i<n gi · φ(x) is not generic. Hence its complement is weakly
generic. By Theorem 3.35 we conclude that the partial type {¬g · φ(x) :
g ∈ G(U)} is f-generic and hence extends to a complete f-generic type p.
The measure µp associated to p gives φ(x) measure 0, so µp 6= µ, which
contradicts unique ergodicity.
Remark 3.37. In particular, in a uniquely ergodic group every f-generic type
is almost periodic and generic.
Recall from [HP11] that an NIP group G is fsg if it admits a global type
p such that for some small model M, all translates of p are finitely satisfiable
over M. It is proved that an fsg group admits a unique invariant measure
and that this measure is generic. So the previous proposition was known in
this special case. We now give an example (pointed out to us by Hrushovski)
of a uniquely ergodic group which is not fsg.
Remark 3.38. Let Kv be a model of ACVF and consider G = (Kv , +) the
additive group. By C-minimality, the partial type p concentrating on the
31
complement of all balls is a complete type and is G-invariant. There can be
no other G-invariant measure since any non-trivial ball in (Kv , +) G-divides,
hence cannot have positive measure for any G-invariant measure. Finally,
the group G is not fsg since p is not finitely satisfiable.
4
Regular ergodic measures
In this section, we are going to characterize regular ergodic measures on
SG (U) for a definably amenable NIP group G = G(U), but first we recall
some general notions and facts from functional analysis and ergodic theory
(see e.g. [Wal00]). As we are going to deal with more delicate measuretheoretic issues here, we will be specific about our measures being regular
or not. The reader should keep in mind that all the results in the previous
sections only apply to regular measures on SG (U).
The set of all regular (Borel, probability) measures on SG (U) can be
naturally viewed as a subset of C∗ (SG (U)), the dual space of the topological
∗
vector space of continuous functions on
R SG (U), Rwith the weak topology of
pointwise convergence (i.e., µi → µ if fdµi → fdµ for all f ∈ C(SG (U))).
It is easy to check that this topology coincides with the logic topology on the
space of measures (Remark 2.8). This space carries a natural structure of a
real topological vector space containing a compact convex set of G-invariant
measures.
We will need the following version of a “converse” to the Krein-Milman
Theorem (see e.g. [Jer54, Theorem 1]. We refer to e.g. [Sim11, Chapter 8]
for a discussion of convexity in topological spaces).
Fact 4.1. Let E be a real, locally convex, Hausdorff topological vector space.
Let C be a compact convex subset of E, and let S be a subset of C. Then the
following are equivalent:
1. C = convS, the closed convex hull of S.
2. The closure of S includes all extreme points of C.
Now we recall the definition of an ergodic measure.
Fact 4.2 ([Phe01, Proposition 12.4]). Let G be a group acting on a topological
space X with x 7→ gx a Borel map for each g ∈ G, and let µ be a G-invariant
Borel probability measure on X. Then the following are equivalent:
32
1. The measure µ is an extreme point of the convex set of G-invariant
measures on X.
2. For every Borel set Y such that µ(gY△Y) = 0 for all g ∈ G, we have
that either µ(Y) = 0 or µ(Y) = 1.
A G-invariant measure is ergodic if it satisfies any of the equivalent conditions above. Under many natural conditions on G and X the two notions
above are equivalent to the following property of µ: for every G-invariant
Borel set Y, either µ(Y) = 0 or µ(Y) = 1. However this is not the case in
general.
Proposition 4.3. The map p 7→ µp from the (closed) set of global f-generic
types to the (closed) set of global G-invariant measures on SG (U) is continuous.
Proof. Fix φ (x) ∈ LG (U) and r ∈ [0, 1], and let Y be the set of all global
f-generic p ∈ SG (U) with µp (φ (x)) ≥ r. It is enough to show that Y is
closed. Let q belong to the closure of Y, in particular q is f-generic. Let L0
be some countable language such that G is L0 -definable and φ(x) ∈ L0 (U),
and let T0 = T |L0 .
Now let M be some countable model of T0 over which φ(x) is defined,
and let ψ(x, y) = φ(y−1 x). Let q ′ (x) = q|ψ , i.e., the restriction of q to all
formulas of the form g · φ(x), ¬g · φ(x), g ∈ G, and let Y ′ = {p|ψ : p ∈ Y}.
By Lemma 3.22, q ′ and all elements of Y ′ are f-generic in the sense of T0 .
By Lemma 3.15 applied in T0 we know that q ′ and all elements of Y ′ are
M-invariant. Working in T0 , let Invψ (M) be the space of all global ψ-types
invariant over M. It follows from the assumption that q ′ ∈ Y ′ (i.e., the
closure of Y ′ in the sense of the topology on Invψ (M)).
By Fact 2.7 we know that q ′ is a limit of a countable sequence (pi′ : i < ω)
of types from Y ′ . Each of pi′ is f-generic in T0 , so in T as well (easy to verify
using equivalence to G-dividing both in T and T0 ), and extends to some global
f-generic L-type pi ∈ Y by Corollary 3.5.
Now work in T , and let ε > 0 be arbitrary. By Proposition 3.23, with
S = {q} ∪ {pi : i < ω}, there are some g0 , . . . , gm ∈ G such that µpi (φ(x)) ≈ε
Av(gjφ(x) ∈ pi ) for all i < ω, as well as µq (φ(x)) ≈ε Av(gjφ(x) ∈ q). As
for any g ∈ G, gφ(x) ∈ pi ⇐⇒ gφ(x) V
∈ pi′ , and the same for q, q ′ , it follows
that for all i < ω large enough we have j<m (gjφ(x) ∈ q ⇐⇒ gj φ(x) ∈ pi ).
But this implies that for any ε > 0, µq (φ (x)) ≥ r − ε, and so µq (φ (x)) ≥ r
and q ∈ Y.
33
Corollary 4.4.
1. The set {µp : p is f-generic} is closed in the set of all
G-invariant measures.
2. Given a G-invariant measure µ, the set of f-generic types p for which
µp = µ is a subflow.
Proof. This follows from Proposition 4.3.
Theorem 4.5. Let G be definably amenable. Then regular ergodic measures
on SG (U) are exactly the measures of the form µp for some f-generic p ∈
SG (U).
Proof. Fix a global f-generic type p, and assume that µp is not an extreme
point. Then there is some 0 < t < 1 and some G-invariant measures µ1 6= µ2
such that µp = tµ1 + (1 − t)µ2. First, it is easy to verify using regularity of
µp that both µ1 and µ2 are regular. Second, it follows that S(µ1), S(µ2) ⊆
S(µp) ⊆ Gp. By Corollary 3.27 which we may apply as µ1 , µ2 are regular, it
follows that µ1 = µp = µ2, a contradiction.
Now for the converse, let µ be an arbitrary regular G-invariant measure
which is an extreme point, and let S = {µp : p ∈ SG (U) is f-generic}. Let
convS be the closed convex hull of S. By Lemma 3.26, µ is a limit of the
averages of measures from S, so µ ∈ convS and it is still an extreme point
of convS. Then we actually have µ ∈ S (by Fact 4.1, as (1) is automatically
satisfied for C = convS, then (2) holds as well). But S = S by Corollary
4.4(1).
Corollary 4.6. The set of all regular ergodic measures in SG (U) is closed.
Let FGen denote the closed G-invariant set of all f-generic types in SG (U).
By Proposition 3.8 we have a well-defined action of G/G00 on FGen (not
necessarily continuous, or even measurable). If ν is an arbitrary regular
G-invariant measure, then S(ν) ⊆ FGen by Proposition 3.14, and we can
naturally view ν as a G/G00 -invariant measure on Borel subsets of FGen.
Question 4.7. Consider the action f : G/G00 × FGen → FGen, (g, p) 7→
g · p. Is it measurable? It is easy to see that f is continuous for a fixed g
and measurable for a fixed p. In many situations this is sufficient for joint
measurability of the map, but our case does not seem to be covered by any
result in the literature.
34
5
Generic compact domination and the Ellis
group conjecture
5.1
Baire-generic compact domination
Let G = G(U) be a definably amenable NIP group, and let M be a small
model of T . Let p ∈ SG (U) be a global type strongly f-generic over M. Let
π : G → G/G00 be the canonical projection. It naturally lifts to a continuous
map π : SG (U) → G/G00 . Fix a formula φ (x) ∈ LG (U), and we define
Uφ(x) = g/G00 : g · p ⊢ φ (x) ⊆ G/G00 .
Proposition 5.1. The set U = Uφ(x) is a constructible subset of G/G00
(namely, a Boolean combination of closed sets).
Proof. Note that U = π (S) with S = {g ∈ G : φ
S(gx) ∈ p}.
As explained in Section 2.2 we have S = n<N (An ∧ ¬Bn+1 ) for some
N < ω, where:
^
Altn (x0 , . . . , xn−1 ) =
¬ (φ(gxi ) ↔ φ(gxi+1 )) ,
i<n−1
An = {g ∈ G : ∃x0 . . . xn−1 (p(n) |M (x0 , . . . , xn−1 ) ∧ Altn (x0 , . . . , xn−1 )∧
∧φ (gxn−1 ))},
Bn = {g ∈ G : ∃x0 . . . xn−1 (p(n) |M (x0 , . . . , xn−1 ) ∧ Altn (x0 , . . . , xn−1 )∧
∧¬φ (gxn−1 ))}.
Note that An , Bn are type definable (over M and the parameters of φ(x)).
Define
An′ := g ∈ G : ∃h ∈ G g−1 h ∈ G00 ∧ h ∈ An ,
Bn′ := g ∈ G : ∃h ∈ G g−1 h ∈ G00 ∧ h ∈ Bn .
S
′
These are also type-definable sets. Let S ′ = n<N An′ ∧ ¬Bn+1
. We
′
check that S = S. Note:
1. S is G00 -invariant (because p is),
2. all of An′ , Bn′ , S ′ are G00 -invariant (by definition),
35
3. An ⊆ An′ , Bn ⊆ Bn′ .
′
First, if g ∈ S ′, say g ∈ An′ ∧ ¬Bn+1
, then there is h ∈ G such that
−1
00
′
′
hg ∈ G and h ∈ An . As g ∈ ¬Bn+1 , also h ∈ ¬Bn+1
, and so h ∈ ¬Bn+1
(by (2) and (3)). Hence h ∈ S, and by (1) also g ∈ S. So S ′ ⊆ S.
Assume that g ∈ S \ S ′, and let n < N be maximal for which there is
h ∈ gG00 such that h ∈ An ∧¬Bn+1 . Then for a corresponding h, we still have
′
h ∈ S \ S ′ by (1) and (2). In particular, h ∈
/ An′ ∧ ¬Bn+1
. As h ∈ An ⊆ An′ ,
′
necessarily h ∈ Bn+1
. This means that there is some h ′ ∈ hG00 = gG00 such
′
that h ∈ Bn+1 . As h ′ is still in S by (1), it follows that h ′ ∈ Am ∧ ¬Bm+1 for
some m, but by the definition of the Bn ’s this is only possible if m+1 > n+1,
contradicting the choice of n. Thus
S = S ′.
S
′
Now, we have π(S ′) = π(S) = n<N π(An′ ) ∧ ¬π(Bn+1
) since An′ and Bn′
are all G00 -invariant. As π(An′ ), π(Bn′ ) are closed, we conclude that π(S) is
constructible.
Let C := G · p ⊆ SG (U), and we define
Eφ(x) = h̄ ∈ G/G00 : π−1 h̄ ∩ φ (x) ∩ C 6= ∅ and π−1 h̄ ∩ ¬φ (x) ∩ C 6= ∅ .
Remark 5.2. Let X be an arbitrary topological space, and let Y ⊆ X be a
constructible set. Then the boundary ∂Y has empty interior.
Proof. This is easily verified as Y is a Boolean combination of closed sets,
∂(Y1 ∪ Y2 ) ⊆ ∂Y1 ∪ ∂Y2 for any sets Y1 , Y2 ⊆ X, and ∂Y has empty interior if
Y is either closed or open.
Theorem 5.3. (Baire-generic compact domination) The set Eφ(x) is closed
and has empty interior. In particular it is meagre.
Proof. We have Eφ(x) = π(φ(x) ∩ C) ∩ π(¬φ(x) ∩ C) and φ(x) ∩ C, ¬φ(x) ∩ C
are closed subsets of SG (U), hence Eφ(x) is closed.
We may assume that p concentrates on G00 , as replacing p by g · p for
some g ∈ G(U) does not change C, and thus does not change Eφ(x) .
Let ḡ ∈ Eφ(x) be given, and let V be an arbitrary open subset of G/G00
containing ḡ. As the map π is continuous, the set S = π−1 (V) is an open
subset of SG (U). By the definition of Eφ(x) , there must exist q, q ′ ∈ C such
that π(q) = π(q ′ ) = ḡ and q ∈ S ∩ φ(x), q ′ ∈ S ∩ ¬φ(x). As C = G · p,
it follows that there are some h, h ′ ∈ G(U) such that h · p ∈ S ∩ φ(x) and
h ′ ·p ∈ S∩¬φ(x). But then, as p concentrates on G00 , π(h) = π(h·p) ∈ V ∩U
36
and π(h ′) = π(h ′ · p) ∈ V ∩ Uc (where U = Uφ(x) is as defined before
Proposition 5.1). As V was an arbitrary neigbourhood of ḡ, it follows that
ḡ ∈ ∂U, hence Eφ(x) ⊆ ∂U. By Proposition 5.1, U is constructible. Hence
∂U has empty interior by Remark 5.2, and so Eφ(x) has empty interior as
well.
5.2
Connected components in an expansion by externally definable sets
Given a small model M of T , an externally definable subset of M is an intersection of an L(U)-definable subset of U with M. One defines an expansion
Mext in a language L ′ by adding a new predicate symbol for every externally
definable subset of Mn , for all n. Recall that a global type p ∈ S(U) is
finitely satisfiable in M if p lies in the topological closure of M, where M
is identified with its image in S(U) under the map sending a ∈ M to the
type x = a. There is a canonical bijection (even homeomorphism) between
S(Mext ) and the subspace of types in S(U) finitely satisfiable in M. Recall
also that a coheir of a type p ∈ S(M) is a type over a larger model N which
extends p and is finitely satisfiable in M.
Let T ′ = ThL ′ (Mext ). Note that automatically any quantifier-free L ′ -type
over Mext is definable (using L ′-formulas). The following is a fundamental
theorem of Shelah [She09] (see also [CS13] for a refined version).
Fact 5.4. Let T be NIP, and let M be a model of T . Then T ′ eliminates
quantifiers.
It follows that T ′ is NIP and that all (L ′-)types over Mext are definable.
Assume now that G is an L-definable group, and let U ′ be a monster
model for T ′ such that U ↾ L is a monster for T . In general there will be
many new L ′ -definable subsets and subgroups of G(U ′ ) which are not Ldefinable. In [CPS14] it is demonstrated however that many properties of
definable groups are preserved when passing to T ′ .
Fact 5.5. Let T be NIP and let M be a small model of T . Let G be an
L-definable group.
1. If G is definably amenable in the sense of T , then it is definably amenable
in the sense of T ′ as well.
37
2. The group G00 (U) computed in T coincides with G00 (U ′ ) computed in
T ′.
In particular this implies that G/G00 is the same group when computed
in T or in T ′ . Note also that the logic topology on G/G00 computed in T
coincides with the logic topology computed in T ′ : any open set in the sense of
T is also open in the sense of T ′ and both are compact Hausdorff topologies,
therefore they must coincide.
Remark 5.6. In view of Remark 2.15, if L is countable then G/G00 is still a
Polish space with respect to the L ′-induced logic topology.
5.3
Ellis group conjecture
We recall the setting of definable topological dynamics and enveloping semigroups (originally from [New09, Section 4], but we are following the notation
from [CPS14]).
Let M0 be a small model of a theory T , and assume that all types over
M0 are definable. Then G(M0 ) acts on SG (M0 ) by homeomorphisms, and
the identity element 1 has a dense orbit. The set SG (M0 ) admits a natural
semigroup structure · extending the group operation on G(M0 ) and continuous in the first coordinate: for p, q ∈ SG (M0 ), p · q is tp(a · b/M0 ) where
b realizes q and a realizes the unique coheir of p over M0 b. This semigroup is precisely the enveloping Ellis semigroup of (SG (M0 ), G(M0 )) (see
e.g. [Gla07a]). In particular left ideals of (SG (M0 ), ·) are precisely the closed
G(M0 )-invariant subflows of G(M0 ) y SG (M0 ), there is a minimal subflow
M and there is an idempotent u ∈ M. Moreover, u · M is a subgroup of
the semigroup (SG (M0 ), ·) whose isomorphism type does not depend on the
choice of M and u ∈ M. It is called the Ellis group (attached to the data).
The quotient map from G = G(U) to G/G00
M0 factors through the tautological
map g 7→ tp(g/M0) from G to SG (M0 ), and we let π denote the resulting
map from SG (M0 ) → G/G00
M0 . It is a surjective semigroup homomorphism,
and for any minimal subflow M of SG (M0 ) and u ∈ M, the restriction of π
to u · M is a surjective group homomorphism.
Now, let T be NIP, and let M be an arbitrary model. Then we consider
M0 := Mext , an expansion of M by naming all externally definable subsets
of Mn for all n ∈ N, in a new language L ′ extending L. Then T ′ = ThL ′ (M0 )
is still NIP, and all L ′ -types over M0 are definable (by Fact 5.4), so the
construction from the previous paragraph applies to (SG (M0 ), G(M0 )). Let
38
U ′ be a monster model for T ′ , so that U = U ′ ↾ L is a monster model for T .
By Fact 5.5, if G(U ′ ) is definably amenable in the sense of T , then it remains
definably amenable in the sense of T ′ , and G00 (U) = G00 (U ′ ) (the first one
is computed in T with respect to L-definable subgroups, while the second
one is computed in T ′ with respect to L ′-definable subgroups). Newelski
asked in [New09] if the Ellis group was equal to G/G00 for some nice classes
of groups. Gismatullin, Penazzi and Pillay [GPP15] show that this is not
always the case for NIP groups (SL2 (R) is a counterexample). The following
modified conjecture was then suggested by Pillay (see [CPS14]):
Ellis group conjecture: Suppose G is a definably amenable NIP group.
Then the restriction of π : SG (M0 ) → G/G00 to u · M is an isomorphism, for
some/any minimal subflow M of SG (M0 ) and idempotent u ∈ M (i.e., π is
injective).
Theorem 5.7. The Ellis group conjecture is true, i.e., π : u · M → G/G00
is an isomorphism.
Proof. Fix notations as above. Throughout this proof, we work in T ′ . Let
p ∈ SG (U ′ ) be strongly f-generic over M0 . Let C := G · p, and let V :=
{p|M0 : p ∈ C}. Note that V is a subflow of G (M0 ) y SG (M0 ): it is closed
as a continuous image of a compact set C into a Hausdorff space, and it is
G (M0 )-invariant as C is G (U ′ )-invariant. Let M be a minimal subflow of
V. It has to be of the form G (M0 ) · (p ′ |M0 ) for some p ′ ∈ C. So replacing
p by p ′ (which is still strongly f-generic over M0 ) we may assume that
M = G (M0 ) · (p|M0 ) is minimal.
Let u ∈ M be an idempotent. We will show that if p1 , p2 ∈ u · M and
π (p1 ) = π (p2 ), i.e., they determine the same coset of G00 , then there is some
r ∈ M such that r · p1 = r · p2 . By the general theory of Ellis semigroups (see
e.g. [Gla07a, Proposition 2.5(5)]) this will imply that p1 = p2 , as wanted.
Let F be the filter of comeagre subsets of G/G00 , and let F ′ be some
ultrafilter extending it. Let q1 , q2 ∈ C be some global types extending p1 , p2
respectively. For each ḡ ∈ G/G00 , let rḡ ∈ SG (M0 ) be a type in M with
π (rḡ ) = ḡ. Let r = limF ′ rḡ . Note that r ∈ M.
Let U∗ ≻ U ′ be a larger monster of T ′ . Let ai ∈ U∗ be such that ai |= qi
for i = 1, 2. For each ḡ ∈ G/G00 let rḡ′ be the unique coheir of rḡ over U∗ ,
and let bḡ |= rḡ′ |U ′ a1 a2 . Finally, let r ′ = limF ′ rḡ′ , the unique coheir of r over
U∗ , and let b ∈ U∗ realize r ′ |U ′ a1 a2 .
Claim 1. limF ′ tp (bḡ · ai /U ′ ) = tp (b · ai /U ′ ) for i = 1, 2.
39
This follows by left continuity of the semigroup operation, but we give
the details. Let φ (x) ∈ L ′ (U ′ ) be arbitrary, and let ai′ ∈ U ′ be such that
ai′ |= qi |N , where N M0 is some small model over which φ (x) is defined.
Then we have:
′
φ(x) ∈ lim
(tp(b
·
a
/U
))
⇔
ḡ ∈ G/G00 :|= φ(bḡ · ai ) ∈ F ′ ⇔
ḡ
i
′
F
⇔ ḡ ∈ G/G00 :|= φ(bḡ · ai′ ) ∈ F ′ ⇔ φ(x · ai′ ) ∈ lim
(tp(bḡ /U ′ )) ⊆ r ′ ⇔
′
F
⇔ φ(x · ai ) ∈ r ′ ⇔|= φ(b · ai ).
The second equivalence is by M0 -invariance of rḡ′ , and the fourth one is
by M0 -invariance of r ′ .
Claim 2. r · p1 = r · p2 .
Assume not, say there exists some φ (x) ∈ L ′ (U ′ ) such that φ (x) ∈ r · p1 ,
¬φ (x) ∈ r · p2 , so |= φ (b · a1 ) ∧ ¬φ (b · a2 ) (according to the choice of
a1 , a2, b and the definition of the semigroup operation on SG (M0 )). We
may
assume that both q1 and q2 concentrate on G00 . By Claim 1 we have
ḡ ∈ G/G00 :|= φ (bḡ · a1 ) ∧ ¬φ (bḡ · a2 ) ∈ F ′ . As Eφ(x) ⊆ G/G00 is meagre
c
by Theorem 5.3, we have Eφ(x) ∈ F ′ , and so there is some ḡ ∈
/ Eφ(x) such
that |= φ (bḡ · a1 ) ∧ ¬φ (bḡ · a2 ).
For an arbitrary open set V ⊆ G/G00 containing ḡ, we can choose h ∈
G(U ′ ) such that π(h) ∈ V and φ (h · a1 ) ∧ ¬φ (h · a2 ) holds. Indeed, let
S = π−1 (V) ⊆ SG (U ′ ), which is open by continuity of π. Then there is an
L ′ (U ′ )-definable set ψ(x) ⊆ S such that π(ψ(x)) ⊆ V and |= ψ(bḡ ). By finite
satisfiability of rḡ′ , take h ∈ G(U ′ ) satisfying φ (x · a1 ) ∧ ¬φ (x · a2 ) ∧ ψ(x).
As ḡ ∈
/ Eφ(x) and Eφ(x) is closed by Theorem 5.3, we find such an h with
π(h) ∈
/ Eφ(x) .
Note that π (h · a1 ) = π (h) = π (h · a2 ) as q1 , q2 concentrate on G00 ,
and that tp(h · a1 /U ′ ) = h · q1 ∈ C, tp(h · a2 /U ′ ) = h · q2 ∈ C. It follows
that π(h) ∈ Eφ(x) — a contradiction.
Corollary 5.8. In a definably amenable NIP group, the Ellis group of the
dynamical system (SG (Mext ), G(M)) is independent of the model M.
40
6
Further remarks
6.1
Left vs. right actions
Until now, we have only considered the action of the group G on itself by
left-translations. One could also let G act on the right and define analogous
notions of right-f-generic, right-invariant measure etc. In a stable group, a
type is left-generic if and only if it is right-generic so we obtain nothing new.
However, in general, left and right notions may differ.
We start with an example of a left-invariant measure which is not rightinvariant.
Example 6.1. Let G = (R, +) ⋊ {±1}, where the two-element group {±1}
acts on R by multiplication. Consider G as a group defined in a model R
of RCF with universe R × {−1, 1} and multiplication defined by (x0 , ǫ0 ) ·
(x1 , ǫ1 ) = (x0 + ǫ0 x1 , ǫ0 ǫ1 ). Let p+
+∞ (x, y) be the type whose restriction to
x is the type at +∞ and which implies y = 1. Define similarly p−
−∞ . Then
1
+
−
µ = 2 (p+∞ + p−∞ ) is left-invariant, but not right-invariant.
However, some things can be said.
Lemma 6.2. Let G = G(U) be definably amenable, then there is always a
measure on G which is both left and right invariant.
Proof. Let µ be a left invariant measure on G which is also invariant over
some small model M (always exists in a definably amenable NIP group, e.g.
by [HP11, Lemma 5.8]).
Let µ−1 be defined by µ−1 (X) := µ(X−1 ) for every definable set X ⊆ G,
where X−1 := {a−1 : a ∈ X}. Then µ−1 is also a measure, M-invariant (as
µ−1 (σ(X)) = µ(σ(X)−1 ) = µ(σ(X−1 )) = µ(X−1 ) = µ−1 (X) for any automorphism σ ∈ Aut(U/M)) and right invariant (as µ−1 (X · g) = µ(g−1 · X−1 ) =
µ(X−1 ) = µ−1 (X) for any g ∈ G).
For any φ(x) ∈ LG (U), we define ν(φ(x)) := µ ⊗ µ−1 (φ(x · u)). That is,
for any definable set X ⊆ G and Ra model N containing M and such that X
is N-definable, we have ν(X) = SG (N) fX dµ−1 , where for every q ∈ SG (N),
fX (q) = µ(X · h−1 ) for some/any h |= q (well-defined by M-invariance of µ,
see Section 2.3). Then ν is an M-invariant measure, and given any g ∈ G
and N such that g and X are N-definable, for any q ∈ SG (N) and h |= q we
have:
41
1. fg·X(q) = µ((g · X) · h−1) = µ(g · (X · h−1)) = µ(X · h−1) = fX (q), by left
invariance of µ.
R
R
2. fX·g (q) = µ((X·g)·h−1) = fX (q·g−1 ), and SG (N) fX (q)dµ−1 = SG (N) fX (q·
R
g−1 )d(µ−1 · g) = SG (N) fX (q · g−1 )d(µ−1) as µ−1 = µ−1 · g by right invariance.
Hence ν is both left and right invariant.
Proposition 6.3. Let G be definably amenable and let φ(x) ∈ LG (U). If
φ(x) is left-generic, then it is right-f-generic.
Proof. By the previous lemma, let µ(x) be a left and right invariant measure
on G. Then as φ(x) is left-generic, we must have µ(φ(x)) > 0. But as µ
is also right-invariant, this implies that φ(x) is right-f-generic (by the “right
hand side” counterpart of Proposition 3.14).
As the following example shows, no other implication holds.
Example 6.4. Let R be a saturated real closed field and let G = (R2 , +) ⋊
SO(2) with the canonical action, seen as a definable group in R. For 0 < a <
1 let Ca ⊂ R2 be the angular region defined by {(x, y) : x ≥ 0 & |y| ≤ a · x}.
Finally, let Xa = Ca × SO(2) ⊆ G.
Note that any two translates of Ca intersect. Hence any two
S right translates of Xa intersect: let g = (xg , σg ) ∈ G, then Xa · g = τ∈SO(2) (Ca +
τ(xg )) × {τ · σg }; hence Xa · g ∩ Xa is non-empty and in fact has surjective
projection on SO(2). This shows that Xa is right-f-generic.
On the other hand, multiplying Xa on the left has the effect of turning it:
g · Xa = (xg + σg (Ca )) × SO(2). If a is infinitesimal, then there are infinitely
many pairwise disjoint left-translates of Xa , hence Xa is not left-f-generic. If
however a is not infinitesimal, then we can cover R2 by finitely many SO(2)conjugates of Ca , and hence cover G by finitely many left-translates of Xa .
We conclude that if a is infinitesimal, then Xa is right-f-generic but not
left-f-generic, and if a is not infinitesimal, then Xa is left-generic but not
right-generic.
6.2
Actions on definable homogeneous spaces
While the theory above was developed for the action of a definably amenable
group G on SG (U), we remark that (with obvious rephrasements) it works
42
just as well for a definably amenable group G = G(U) acting on SX (U) for X
a definable homogeneous G-space (i.e. X is a definable set, the graph of the
action map G × X → X is definable and the action is transitive). We show
that given a definable homogeneous space X for a definably amenable group
G, every G-invariant measure on G pushes forward to a G-invariant measure
on X, and conversely any G-invariant measure on X lifts to a G-invariant
measure on G, possibly non-uniquely.
Lemma 6.5. Let B0 ⊆ Def(U) be a Boolean algebra and let I ⊆ Def(U) be
an ideal such that I ∩ B0 is contained in the zero-ideal of ν0 , a measure on
B0 .
Let B be the collection of all sets U ∈ Def(U) for which there is some
V ∈ B0 such that U△V ∈ I. Then B is a Boolean algebra with B0 , I ⊆ B.
Moreover, ν0 extends to a global measure ν on Def(U) such that all sets from
I have ν-measure 0.
Proof. It can be checked straightforwardly that B is a Boolean algebra containing B0 and I. Now for U ∈ B, let ν ′ (U) = ν0 (V) where V is some set in
B0 with U△V ∈ I.
1. ν ′ is well-defined. If we have some V ′ ∈ B0 with U△V ′ ∈ I, then
V△V ′ ⊆ (U△V) ∪ (U△V ′ ) ∈ I, so V△V ′ ∈ I; but by assumption this
implies that ν0 (V△V ′ ) = 0, so ν0 (V) = ν0 (V ′ ).
2. ν ′ is a measure on B extending ν0 . Given Ui ∈ B, i ≤ 2, let Vi ∈ B0
be such that Ui △Vi ∈ I, i ≤ 2. Then ν ′ (U1 ∪ U2 ) = ν(V1 ∪ V2 ) =
ν(V1 ) + ν(V2 ) − ν(V1 ∩ V2 ) = ν ′ (U1 ) + ν ′ (U2 ) − ν ′ (U1 ∩ U2 ), as wanted.
Now ν ′ extends to a global measure ν by compactness, see e.g. [Sim15a,
Lemma 7.3].
Proposition 6.6. Let X be a definable homogeneous G-space, and let x0 be
an arbitrary point in X.
1. Let µ̃ be a measure on G. For every definable subset φ(x) of X, let
µ(φ(x)) = µ̃(φ(u · x0 )). Then µ is a measure on X. Moreover, if µ̃ is
G-invariant, then µ is G-invariant as well. If µ̃ is also right-invariant,
then µ does not depend on the choice of x0 .
43
2. Assume moreover that G is definably amenable, NIP. Let µ be a Ginvariant measure on X. Then there is some (possibly non-unique)
G-invariant measure µ̃ on G such that the procedure from (1) induces
µ.
Proof. (1) It is clearly a measure as µ(∅) = µ̃(∅), µ(X) = µ̃(G) and if
φi (x), i < n are disjoint subsets of X, then φi (u · x0 ), i < n are disjoint
subsets of G. If µ̃ is G-invariant then for any g ∈ G we have µ(φ(g−1 · x)) =
µ̃(φ(g−1 · u · x0 )) = µ̃(φ(u · x0 )) = µ(φ(x)).
Finally, assume that µ̃ is also right invariant. Let x1 ∈ X and φ(x) be
arbitrary, then by transitivity of the action there is some g ∈ G such that
x1 = g · x0 . We have µ̃(φ(u · x1 )) = µ̃(φ(u · (g · x0 ))) = µ̃(φ((u · g) · x0 )) =
µ̃(φ(u · x0 ) · g−1 ) = µ̃(φ(u · x0 )), as wanted.
(2) Now let µ be a G-invariant measure on X, and fix x0 ∈ X. Let
B0 ⊆ Def G (U) be the family of subsets of G of the form {g ∈ G : g · x0 ∈ Y},
where Y is a definable subset of X. For U ∈ B0 , define ν0 (U) = µ(Y). The
following can be easily verified using that µ is a G-invariant measure:
Claim. The family B0 is a Boolean algebra closed under G-translates
and ν0 is a G-invariant measure on B0 .
Next, let I ⊆ Def G (U) be the collection of all non-f-generic definable
subsets of G. We know by Corollary 3.5 that it is an ideal. As in Proposition
3.14, B0 ∩ I is contained in the zero-ideal of ν0 . Then, applying Lemma
6.5, we obtain a global measure ν on Def G (U) extending ν0 and such that
all types in its support are f-generic. Note that ν is G00 -invariant: for any
φ(x) ∈ L(U) and ε > 0 there are some p0 , . . . , pn−1 ∈ S(ν) such that for any
g ∈ G, ν(gφ(x)) ≈ε Av(p0 , . . . , pn−1 ; gφ(x)) (by Fact 2.9), and each pi is
G00 -invariant (by Proposition 3.8). Consider the map fφ : G/G00 → R, g 7→
ν(gφ(x)). It is well-defined and h0 -measurable (using an
R argument as in the
proof of Lemma 3.21). Finally, we define µ̃(φ(x)) = g∈G/G00 fφ (g)dh0 . It
is easy to check that µ̃ is a G-invariant measure on Def G (U) (and that the
procedure from (1) applied to µ̃ returns µ).
References
[BGT12] Emmanuel Breuillard, Ben Green, and Terence Tao. The structure of approximate groups. Publications mathématiques de l’IHÉS,
116(1):115–221, Nov 2012.
44
[CK12]
Artem Chernikov and Itay Kaplan. Forking and dividing in NTP2
theories. J. Symbolic Logic, 77(1):1–20, 2012.
[CP12]
Annalisa Conversano and Anand Pillay. Connected components of
definable groups and o-minimality I. Adv. Math., 231(2):605–623,
2012.
[CPS14] Artem Chernikov, Anand Pillay, and Pierre Simon. External definability and groups in NIP theories. Journal of the London Mathematical Society, 90(1):213–240, 2014.
[CS13]
Artem Chernikov and Pierre Simon. Externally definable sets and
dependent pairs. Israel Journal of Mathematics, 194(1):409–425,
2013.
[Gla07a] Eli Glasner. Enveloping semigroups in topological dynamics. Topology and its Applications, 154(11):2344 – 2363, 2007.
[Gla07b] Eli Glasner. The structure of tame minimal dynamical systems.
Ergodic Theory and Dynamical Systems, 27(06):1819–1837, 2007.
[GPP15] Jakub Gismatullin, Davide Penazzi, and Anand Pillay. Some model
theory of SL(2, R). Fundamenta Mathematicae, 229(2):117–128,
2015.
[Hal50]
Paul Richard Halmos. Measure theory, volume 2. van Nostrand
New York, 1950.
[HP11]
Ehud Hrushovski and Anand Pillay. On NIP and invariant measures. J. Eur. Math. Soc. (JEMS), 13(4):1005–1061, 2011.
[HPP08] Ehud Hrushovski, Ya’acov Peterzil, and Anand Pillay. Groups,
measures, and the NIP. J. Amer. Math. Soc., 21(2):563–596, 2008.
[HPS13] Ehud Hrushovski, Anand Pillay, and Pierre Simon. Generically
stable and smooth measures in NIP theories. Trans. Amer. Math.
Soc., 365(5):2341–2366, 2013.
[Hru90]
Ehud Hrushovski. Unidimensional theories are superstable. Annals
of Pure and Applied Logic, 50(2):117 – 137, 1990.
45
[Hru96]
Ehud Hrushovski. The Mordell-Lang conjecture for function fields.
J. Amer. Math. Soc, 9:667–690, 1996.
[Hru01]
Ehud Hrushovski. The manin–mumford conjecture and the model
theory of difference fields. Annals of Pure and Applied Logic,
112(1):43–115, 2001.
[Hru12]
Ehud Hrushovski. Stable group theory and approximate subgroups.
J. Amer. Math. Soc., 25(1):189–243, 2012.
[Jer54]
Meyer Jerison. A property of extreme points of compact convex
sets. Proceedings of the American Mathematical Society, 5(5):782–
783, 1954.
[Kec95]
Alexander S Kechris. Classical descriptive set theory, volume 156.
Springer-Verlag New York, 1995.
[KL07]
David Kerr and Hanfeng Li. Independence in topological and c*dynamics. Mathematische Annalen, 338(4):869–926, 2007.
[Las92]
Michael C. Laskowski. Vapnik-Chervonenkis classes of definable
sets. J. London Math. Soc. (2), 45(2):377–384, 1992.
[Mat04] Jiřı́ Matoušek. Bounded VC-dimension implies a fractional Helly
theorem. Discrete Comput. Geom., 31(2):251–255, 2004.
[MS14]
Alice Medvedev and Thomas Scanlon. Invariant varieties for polynomial dynamical systems. Annals of Mathematics, 179(1):81–177,
2014.
[New09] Ludomir Newelski. Topological dynamics of definable group actions. J. Symbolic Logic, 74(1):50–72, 2009.
[New12] Ludomir Newelski. Bounded orbits and measures on a group. Israel
J. Math., 187:209–229, 2012.
[NP06]
Ludomir Newelski and Marcin Petrykowski. Weak generic types
and coverings of groups. I. Fund. Math., 191(3):201–225, 2006.
[Phe01]
Robert R Phelps. Lectures on Choquet’s theorem, volume 1757.
Springer, 2001.
46
[Pil04]
Anand Pillay. Type-definability, compact Lie groups, and ominimality. Journal of Mathematical Logic, 4(02):147–162, 2004.
[Poi87]
Bruno Poizat. Groupes stables: une tentative de conciliation entre la géométrie algébrique et la logique mathématique, volume 2.
Launay, 1987.
[Poi01]
Bruno Poizat. Stable groups, volume 87 of Mathematical Surveys
and Monographs. American Mathematical Society, Providence, RI,
2001. Translated from the 1987 French original by Moses Gabriel
Klein.
[PY16]
Anand Pillay and Ningyuan Yao. On minimal flows, definably
amenable groups, and o-minimality. Advances in Mathematics,
290:483–502, 2016.
[Sel13]
Zlil Sela. Diophantine geometry over groups viii: Stability. Annals
of Mathematics, 177(3):787–868, 2013.
[She71]
Saharon Shelah. Stability, the f.c.p., and superstability; model
theoretic properties of formulas in first order theory. Ann. Math.
Logic, 3(3):271–362, 1971.
[She08]
Saharon Shelah. Minimal bounded index subgroup for dependent theories. Proceedings of the American Mathematical Society,
136(3):1087–1091, 2008.
[She09]
Saharon Shelah. Dependent first order theories, continued. Israel
Journal of Mathematics, 173(1):1–60, 2009.
[Sim11]
Barry Simon. Convexity: An analytic viewpoint, volume 187. Cambridge University Press, 2011.
[Sim15a] Pierre Simon. A guide to NIP theories. Cambridge University
Press, 2015.
[Sim15b] Pierre Simon. Rosenthal compacta and NIP formulas. Fundamenta
Mathematicae, 231(1):81–92, 2015.
[Sta16]
Sergei Starchenko. NIP, Keisler measures and combinatorics.
Séminaire BOURBAKI, page 68, 2016.
47
[VC71]
Vladimir N Vapnik and A Ya Chervonenkis. On the uniform convergence of relative frequencies of events to their probabilities. Theory
of Probability & Its Applications, 16(2):264–280, 1971.
[Wag00] Frank O. Wagner. Simple theories, volume 503 of Mathematics and
its Applications. Kluwer Academic Publishers, Dordrecht, 2000.
[Wal00] Peter Walters.
Springer, 2000.
[Zil]
An introduction to ergodic theory, volume 79.
B. Zilber. Uncountably Categorical Theories. Translations of mathematical monographs. American Mathematical Society.
48
| 4math.GR
|
JMLR: Workshop and Conference Proceedings vol 49:1–30, 2016
Optimal Best Arm Identification with Fixed Confidence
Aurélien Garivier
AURELIEN . GARIVIER @ MATH . UNIV- TOULOUSE . FR
arXiv:1602.04589v2 [math.ST] 1 Jun 2016
Institut de Mathématiques de Toulouse; UMR5219
Université de Toulouse; CNRS
UPS IMT, F-31062 Toulouse Cedex 9, France
Emilie Kaufmann
EMILIE . KAUFMANN @ INRIA . FR
Univ. Lille, CNRS, UMR 9189
CRIStAL - Centre de Recherche en Informatique Signal et Automatique de Lille
F-59000 Lille, France
Abstract
We give a complete characterization of the complexity of best-arm identification in one-parameter
bandit problems. We prove a new, tight lower bound on the sample complexity. We propose
the ‘Track-and-Stop’ strategy, which we prove to be asymptotically optimal. It consists in a new
sampling rule (which tracks the optimal proportions of arm draws highlighted by the lower bound)
and in a stopping rule named after Chernoff, for which we give a new analysis.
Keywords: multi-armed bandits, best arm identification, MDL.
1. Introduction
A multi-armed bandit model is a paradigmatic framework of sequential statistics made of K probability distributions ν1 , . . . , νK with respective means µ1 , . . . , µK : at every time step t = 1, 2, . . .
one arm At ∈ A = {1, . . . , K} is chosen and a new, independent reward Xt is drawn from νAt .
Introduced in the 1930s with motivations originally from clinical trials, bandit models have raised a
large interest recently as relevant models for interactive learning schemes or recommender systems.
A large part of these works consisted in defining efficient strategies for maximizing the expected cumulated reward E[X1 + · · · + Xt ]; see for instance Bubeck and Cesa-Bianchi (2012) for a survey. A
good understanding of this simple model has allowed for efficient strategies in much more elaborate
settings, for example including side information (Abbasi-Yadkori et al., 2011; Agrawal and Goyal,
2013), infinitely many arms (Srinivas et al., 2010; Bubeck et al., 2011), or for the search of optimal
strategies in games (Munos, 2014), to name just a few.
In some of these applications, the real objective is not to maximize the cumulated reward, but
rather to identify the arm that yields the largest mean reward µ∗ = max1≤a≤K µa , as fast and
accurately as possible, regardless of the number of bad arm draws. Let Ft = σ(X1 , . . . , Xt ) be the
sigma-field generated by the observations up to time t. A strategy is then defined by:
• a sampling rule (At )t , where At is Ft−1 -measurable;
• a stopping rule τ , which is a stopping time with respect to Ft ;
• and a Fτ -measurable decision rule âτ .
The goal is to guarantee that âτ ∈ argmax µa with the highest possible probability while minimizing
the number τ of draws. Two settings have been considered in the literature. In the fixed-budget
setting, the number of draws is fixed in advance, and one aims at minimizing the probability of
c 2016 A. Garivier & E. Kaufmann.
G ARIVIER K AUFMANN
error P(âτ ∈
/ argmax µa ). In the fixed-confidence setting a maximal risk δ is fixed, and one looks
for a strategy guaranteeing that P(âτ ∈
/ argmax µa ) ≤ δ (such a strategy is called δ-PAC) while
minimizing the sample complexity E[τ ].
The aim of this paper is to propose an analysis of the best arm identification problem in the fixedconfidence setting. For the sake of clarity and simplicity, we suppose that there is a single arm with
highest expectation, and without loss of generality that µ1 > µ2 ≥ · · · ≥ µK . We also focus on the
simple case where the distributions are parameterized by their means, as in one-parameter exponential families, and we index probabilities and expectations by the parameter µ = (µ1 , . . . , µK ). The
Kullback-Leibler divergence of two distributions of means µ1 and µ2 is a function d : (µ1 , µ2 ) →
R+ . Two cases are of particular interest: the Gaussian case, with d(x, y) = (x − y)2 /(2σ 2 ), and
binary rewards with d(x, y) = kl(x, y) := x log(x/y) + (1 − x) log((1 − x)/(1 − y)).
Several strategies have been proposed to minimize Eµ [τ ]. While racing strategies are based on
successive eliminations of apparently sub-optimal arms (Even-Dar et al. (2006); Kaufmann and Kalyanakrishnan
(2013)), another family of strategies exploits the use of upper and lower confidence bounds on the
means of the arms (Kalyanakrishnan et al. (2012); Gabillon et al. (2012); Kaufmann and Kalyanakrishnan
(2013); Jamieson et al. (2014)). They reflect some aspects of the difficulty of the problem, but are
not proved to satisfy any optimality property1 . In particular, there was still a gap between the lower
bounds, involving complexity terms reflecting only partially the structure of the problem, and the
upper bounds on Eµ [τ ] for these particular algorithms, even from an asymptotic point of view. For
the particular case K = 2, this gap was closed in Kaufmann et al. (2014). The tools used there,
however, where specific to the two-arm case and cannot be extended easily.
The first result of this paper is a tight, non-asymptotic lower bound on Eµ [τ ]. This bound
involves a ‘characteristic time’ for the problem, depending on the parameters of the arms, which
does not take a simple form like for example a sum of squared inverse gaps. Instead, it appears as
the solution of an optimization problem, in the spirit of the bounds given by Graves and Lai (1997)
in the context of regret minimization. We give a brief analysis of this optimization problem, and we
provide an efficient numerical solution as well as elements of interpretation.
The second contribution is a new δ-PAC algorithm that asymptotically achieves this lower
bound, that we call the Track-and-Stop strategy. In a nutshell, the idea is to sample so as to equalize
the probability of all possible wrong decisions, and to stop as soon as possible. The stopping rule,
which we name after Chernoff, can be interpreted in three equivalent ways: in statistical terms, as
a generalized likelihood ratio test; in information-theoretic terms, as an application of the Minimal
Description Length principle; and in terms of optimal transportation, in light of the lower bound.
The sampling rule is a by-product of the lower bound analysis, which reveals the existence of optimal proportions of draws for each arm. By estimating and tracking these proportions, our algorithm
asymptotically reaches the optimal sample complexity as δ goes to 0.
The paper is organized as follows. In Section 2, the lower bound is given with a short proof.
Section 3 contains a commented description of our stopping and sampling rules. The analysis of the
resulting algorithm is sketched in Sections 4 (validity of the stopping rule) and 5 (sample complexity
analysis), establishing its asymptotic optimality. Section 6 contains practical comments and results
of numerical experiments, in order to show the efficiency of our strategy even for moderate values
1. Optimality is mentioned in several articles, with different and sometimes weak meanings (minimax, rate-optimal,...).
In our view, BAI algorithms for which there exists a model with a sample complexity bounded, up to a multiplicative
constant, by some quantity related to some lower bound, may not be called optimal.
2
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
of δ. We also briefly comment on the gain over racing strategies, which can be explained in light of
Theorem 1. Most proofs and technical details are postponed to the Appendix.
2. Lower Bounds on the sample complexity
The pioneering work of Lai and Robbins (1985) has popularized the use of changes of distributions
to show problem-dependent lower bounds in bandit problems: the idea is to move the parameters
of the arms until a completely different behavior of the algorithm is expected on this alternative
bandit model. The cost of such a transportation is induced by the deviations of the arm distributions: by choosing the most economical move, one can prove that the alternative behavior is not
too rare in the original model. Recently, Kaufmann et al. (2014) and Combes and Proutière (2014)
have independently introduced a new way of writing such a change of measure, which relies on
a transportation lemma that encapsulates the change of measure and permits to use it at a higher
level. Here we go one step further by combining several changes of measures at the same time, in
the spirit of Graves and Lai (1997). This allows us to prove a non-asymptotic lower bound on the
sample complexity valid for any δ-PAC algorithm on any bandit model with a unique optimal arm.
We present this result in the particular case of arms that belong to a canonical exponential family,
dνθ
= exp θx − b(θ) ,
P = (νθ )θ∈Θ :
dξ
where Θ ⊂ R, ξ is some reference measure on R and b : Θ → R is a convex, twice differentiable
function. A distribution νθ ∈ P can be parameterized by its mean ḃ(θ), and for every µ ∈ ḃ(Θ) we
denote by ν µ be the unique distribution in P with expectation µ. Unfamiliar readers may simply
think of Bernoulli laws, or Gaussian distributions with known variance. As explained in Cappé et al.
(2013, see also references therein), the Kullback-Leibler divergence from νθ to νθ′ induces a divergence function d on ḃ(Θ) defined, if ḃ(θ) = µ and ḃ(θ ′ ) = µ′ , by
′
d(µ, µ′ ) = KL(ν µ , ν µ ) = KL(νθ , νθ′ ) = b(θ ′ ) − b(θ) − ḃ(θ)(θ ′ − θ) .
With some abuse of notation, an exponential family bandit model ν = (ν µ1 , . . . , ν µK ) is identified
with the means of its arms µ = (µ1 , . . . , µK ).
2.1. General Lower Bound
Denote by S a set of exponential bandit models such that each bandit model µ = (µ1 , . . . , µK ) in S
has a unique optimal arm: for each µ ∈ S, there exists an arm a∗ (µ) such that µa∗ (µ) > max{µa :
a 6= a∗ (µ)}. Fixed-confidence strategies depend on the risk level and we subscript stopping rules
by δ. A strategy is called δ-PAC if for every µ ∈ S, Pµ (τδ < ∞) = 1 and Pµ (âτδ 6= a∗ ) ≤ δ. We
introduce
Alt(µ) := {λ ∈ S : a∗ (λ) 6= a∗ (µ)} ,
the set of problems where the optimal arm is not the same as in µ, and ΣK = {ω ∈ Rk+ : ω1 + · · · +
ωK = 1} the set of probability distributions on A.
Theorem 1 Let δ ∈ (0, 1). For any δ-PAC strategy and any bandit model µ ∈ S,
Eµ [τδ ] ≥ T ∗ (µ) kl(δ, 1 − δ),
3
G ARIVIER K AUFMANN
where
∗
−1
T (µ)
:= sup
inf
w∈ΣK λ∈Alt(µ)
K
X
!
wa d(µa , λa )
a=1
.
(1)
Remark 2 As kl(δ, 1 − δ) ∼ log(1/δ) when δ goes to zero, Theorem 1 yields the asymptotic lower
bound
Eµ [τδ ]
≥ T ∗ (µ).
lim inf
δ→0 log(1/δ)
A non-asymptotic version can be obtained for example from the inequality kl(δ, 1−δ) ≥ log(1/(2.4δ))
that holds for all δ ∈ (0, 1), given in Kaufmann et al. (2015).
We will see that the supremum in Equation (1) is indeed a maximum, and we call
!
K
X
∗
w (µ) := argmax inf
wa d(µa , λa )
w∈ΣK
λ∈Alt(µ)
a=1
the corresponding distribution on the arms. The proof of Theorem 1 shows that w∗ is the proportion
of arm draws of any strategy matching this lower bound.
The particular case where S is the class of bandit models with Poisson rewards in which all
suboptimal arms are equal is considered in the very insightful paper by Vaidhyan and Sundaresan
(2015), where a closed-form formula is given for both T ∗ (µ) and w∗ (µ). In this paper, we consider
best arm identification in all possible bandit models with a single best arm, and in the rest of the
paper we fix
S :=
µ : ∃a ∈ A : µa > max µi .
i6=a
Proof of Theorem 1. Let δ ∈ (0, 1), µ ∈ S, and consider a δ-PAC strategy. For every t ≥ 1,
denote by Na (t) the (random) number of draws of arm a up to time t. The ‘transportation’ Lemma
1 of Kaufmann et al. (2015) relates the expected number of draws of each arm and the KullbackLeibler divergence of two bandit models with different optimal arms to the probability of error δ:
∗
∗
∀λ ∈ S : a (λ) 6= a (µ),
K
X
a=1
d(µa , λa )Eµ [Na (τδ )] ≥ kl(δ, 1 − δ).
(2)
Instead of choosing for each arm a a specific instance of λ that yields a lower bound on Eµ [Na (τδ )],
we combine here the inequalities given by all alternatives λ:
!
!
K
K
X
X
Eµ [Na ]
Eµ [Na ]
kl(δ, 1 − δ) ≤ inf Eµ [τδ ]
d(µa , λa ) = Eµ [τδ ] inf
d(µa , λa )
Eµ [τδ ]
Eµ [τδ ]
λ∈Alt(µ)
λ∈Alt(µ)
a=1
a=1
!
K
X
wa d(µa , λa ) ,
≤ Eµ [τδ ] sup inf
w∈ΣK λ∈Alt(µ)
as Eµ [τδ ] =
a=1
PK
a=1 Eµ [Na (τδ )].
In the last inequality, the strategy-dependent proportions of arm draws are replaced by their supremum so as to obtain a bound valid for any δ-PAC algorithm; one can see that a strategy may reach
this bound only if it meets the w∗ (µ). To make this bound useful, it remains to study T ∗ and w∗ .
4
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
2.2. About the Characteristic Time and the Optimal Proportions
We study here the optimization problem (1), so as to better understand the function T ∗ and w∗
(Proposition 6), and in order to provide an efficient algorithm for computing first w∗ (µ) (Theorem 5), then T ∗ (µ) (Lemma 3). The main ideas are outlined here, while all technical details are
postponed to Appendix A. Simplifying T ∗ (µ) requires the introduction of the following parameterized version of the Jensen-Shannon divergence (which corresponds to α = 1/2): for every
α ∈ [0, 1], let
Iα (µ1 , µ2 ) := αd µ1 , αµ1 + (1 − α)µ2 + (1 − α)d µ2 , αµ1 + (1 − α)µ2 .
(3)
The first step is, for any w, to identify the minimizer of the transportation cost:
Lemma 3 For every w ∈ ΣK ,
inf
λ∈Alt(µ)
K
X
!
wa d(µa , λa )
a=1
= min (w1 + wa )I
a6=1
w1
w1 +wa
(µ1 , µa ) .
It follows that
T ∗ (µ)−1 =
sup min (w1 + wa )I
w∈ΣK a6=1
w1
w1 +wa
w∗ (µ) = argmax min (w1 + wa )I
w∈ΣK
a6=1
(µ1 , µa ) , and
w1
w1 +wa
(µ1 , µa ) .
It is easy to see that, at the optimum, the quantities (w1 + wa )Iw1 /(w1 +wa ) (µ1 , µa ) are all equal.
Lemma 4 For all a, b ∈ {2, . . . , K},
(w1∗ + wa∗ )I
∗
w1
∗ +w∗
w1
a
(µ1 , µa ) = (w1∗ + wb∗ )I
∗
w1
∗ +w∗
w1
b
(µ1 , µb ) .
This permits to obtain a more explicit formula for w∗ (µ) involving only a single real parameter.
Indeed, for every a ∈ {2, . . . K} let
ga (x) = (1 + x)I
1
1+x
(µ1 , µa ) .
(4)
The function ga is a strictly increasing one-to-one mapping from [0, +∞[ onto [0, d(µ1 , µa )[. We
define xa : [0, d(µ1 , µa )[→ [0, +∞[ as its inverse function: xa (y) = ga−1 (y). Denoting x1 the
function constantly equal to 1, one obtains the following characterization of w∗ (µ):
Theorem 5 For every a ∈ A,
xa (y ∗ )
wa∗ (µ) = PK
,
∗
a=1 xa (y )
where y ∗ is the unique solution of the equation Fµ (y) = 1, and where
K d µ1 , µ1 +xa (y)µa
X
1+xa (y)
Fµ : y 7→
µ1 +xa (y)µa
a=2 d µa , 1+xa (y)
(5)
(6)
is a continuous, increasing function on [0, d(µ1 , µ2 )[ such that Fµ (0) = 0 and Fµ (y) → ∞ when
y → d(µ1 , µ2 )).
5
G ARIVIER K AUFMANN
Thus, w∗ can be simply computed by applying (for example) the bisection method to a function
whose evaluations requires the resolution of K smooth scalar equations. By using efficient numerical solvers, we obtain a fast algorithm of complexity, roughly speaking, proportional to the number
of arms. This characterization of w∗ (µ) also permits to obtain a few sanity-check properties, like
for example:
Proposition 6
1. For all µ ∈ S, for all a, wa∗ (µ) 6= 0.
2. w∗ is continuous in every µ ∈ S.
∗ (µ).
3. If µ1 > µ2 ≥ · · · ≥ µK , one has w2∗ (µ) ≥ · · · ≥ wK
Observe that one may have2 w2 > w1 . In general, it is not possible to give closed-form formulas
for T ∗ (µ) and w∗ (µ). In particular, T ∗ (µ) cannot be written as a sum over the arms of individual
complexity terms as in previous works (Mannor and Tsitsiklis (2004); Kaufmann et al. (2015)). But
the following particular cases can be mentioned.
Two-armed bandits. For a two-armed bandit model µ = (µ1 , µ2 ), T ∗ (µ) and w∗ (µ) can be
computed algebraically. Lemma 3 and the fact that w2 = 1 − w1 imply that
T ∗ (µ)−1 =
sup Iα (µ1 , µa ),
α∈(0,1)
Some algebra shows that the maximum is reached at α = α∗ (µ1 , µ2 ) defined by the equation
d(µ1 , µ∗ ) = d(µ2 , µ∗ ), where µ∗ = α∗ (µ1 , µ2 )µ1 +(1−α∗ (µ1 , µ2 ))µ2 . The value of the maximum
is then the ‘reversed’ Chernoff information d∗ (µ1 , µ2 ) := d(µ1 , µ∗ ) = d(µ2 , µ∗ ). This permits to
recover the bound already given in Kaufmann et al. (2014):
Eµ [τδ ] ≥
kl(δ, 1 − δ)
.
d∗ (µ1 , µ2 )
The Gaussian case. When d(x, y) = (x − y)2 /(2σ 2 ), T ∗ (µ) and w∗ (µ) can be computed by
solving a rational equation. Indeed, Equation (6) and Lemma 4 imply that
K
X
a=2
x2a = 1 and
K
X
λ2
xa
λ
=
,
thus
=1
2−λ 2
1 + xa
(µ1 − µa )2
(µ
−
µ
)
1
a
a=2
for some λ ∈ (0, (µ1 − µ2 )2 ). For K = 3, λ is the solution of a polynomial equation of degree
4 and has therefore an (obscure) algebraic expression. The following inequalities, established in
Appendix A.4 give a better insight of the order of magnitude of T ∗ (µ): if ∆1 = ∆2 and ∆a =
µ1 − µa for a ≥ 2, then
K
K
X
X
2σ 2
2σ 2
∗
≤
T
(µ)
≤
2
.
∆2a
∆2a
a=1
a=1
2. This can happen when the right-deviations of ν µ2 are smaller than the left-deviations of ν µ1 ; for example, with
Bernoulli arms of parameters µ = (0.5, 0.1, 0.02), ν ∗ (µ) ≈ (39%, 42%, 19%).
6
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
3. The Track-and-Stop Strategy
We now describe a new strategy which is the first (as far as we know) to asymptotically match the
lower bound of Theorem 1. Denote by µ̂(t) =
P(µ̂1 (t), . . . , µ̂K (t)) the current maximum likelihood
estimate of µ at time t: µ̂a (t) = Na (t)−1 s≤t Xs 1{As = a}. As seen in Section 2, a good
sampling rule should respect the optimal proportions of arm draws given by w∗ (µ). There are
several ways to ensure this, and we present two of them in Section 3.1. A good stopping rule should
determine the earliest moment when sufficient statistical evidence has been gathered for identifying
the best arm: we propose one (with several interpretations) in Section 3.2, showing in Section 4
how to tune it in order to ensure the δ-PAC property. As for the decision rule, we simply choose
âτδ = argmax µ̂a (τδ ). The optimality of the Track-and-Stop strategy is shown in Section 5.
a∈A
3.1. Sampling Rule: Tracking the Optimal Proportions
The first idea for matching the proportions w∗ (µ) is to track the plug-in estimates w∗ (µ̂(t)). In
bandit settings, using plug-in estimates is always hazardous, because bad initial estimate may lead
to abandon an arm and to prevent further observations that would correct the initial error. Indeed,
one may see (both theoretically and in numerical experiments) that a naive plug-in sampling rule
sometimes fails. But there is a very simple fix, which consists in forcing sufficient exploration of
each arm to ensure a (sufficiently fast) convergence of µ̂(t).
The shortest way to do this (which we term C-Tracking) is to slightly alter the optimization
solu
tion: for every ǫ ∈ (0, 1/K], let wǫ (µ) be a L∞ projection of w∗ (µ) onto ΣǫK = (w1 , . . . , wK ) ∈
[ǫ, 1] : w1 + · · · + wK = 1 . Choosing ǫt = (K 2 + t)−1/2 /2 and
t
X
waǫs (µ̂(s)) − Na (t) ,
At+1 ∈ argmax
1≤a≤K s=0
we prove in Appendix B that:
Lemma 7 For all t ≥ 1 and a ∈ A, the C-Tracking rule ensures that Na (t) ≥
and that
t−1
X
√
wa∗ (µ̂(s)) ≤ K(1 + t) .
max Na (t) −
1≤a≤K
√
t + K 2 − 2K
s=0
It is slightly more efficient in practice to target directly w∗ (µ̂(t)),
√ and to force exploration steps
whenever an arm is in deficit. Introducing Ut = {a : Na (t) < t − K/2}, the D-Tracking rule
(At ) is sequentially defined as
At+1 ∈
argmin Na (t) if Ut 6= ∅
(forced exploration)
a∈Ut
argmax t wa∗ (µ̂(t)) − Na (t) (direct tracking)
1≤a≤K
√
Lemma 8 The D-Tracking rule ensures that Na (t) ≥ ( t − K/2)+ − 1 and that for all ǫ > 0, for
all t0 , there exists tǫ ≥ t0 such that
sup max |wa∗ (µ̂(t)) − wa∗ (µ)| ≤ ǫ
t≥t0
a
⇒
sup max
t≥tǫ
7
a
Na (t)
− wa∗ (µ) ≤ 3(K − 1)ǫ .
t
G ARIVIER K AUFMANN
It is guaranteed under all these sampling rules that the empirical proportion of draws of each arm
converges to the optimal proportion, as proved in Appendix B.3.
Proposition 9 The C-Tracking and D-Tracking sampling rules both satisfy
Na (t)
∗
= wa (µ) = 1 .
Pw lim
t→∞
t
We actually have a little more: a minimal convergence speed of µ̂ to µ, which proves useful in
the analysis of the expected sample complexity. Of course, other tracking strategies are possible,
like for example the one introduced in Antos et al. (2008) for the uniform estimation of all arms’
expectations.
3.2. Chernoff’s Stopping Rule
From a statistical point of view, the question of stopping at time t is a more or less classical statistical
test: do the past observations allow to assess, with a risk at most δ, that one arm is larger than the
others? For all arms a, b ∈ A, we consider the Generalized Likelihood Ratio statistic
maxµ′a ≥µ′b pµ′a X aNa (t) pµ′b X bNb (t)
,
Za,b (t) := log
maxµ′a ≤µ′b pµ′a X aNa (t) pµ′b X bNb (t)
where X aNa (t) = (Xs : As = a, s ≤ t) is a vector that contains the observations of arm a available
at time t, and where pµ (Z1 , . . . , Zn ) is the likelihood of n i.i.d. observations from wµ :
pµ (Z1 , . . . , Zn ) =
n
Y
k=1
exp(ḃ−1 (µ)Zk − b(ḃ−1 (µ))) .
This statistic has a convenient closed-form expression for exponential family bandit models. Introducing for all arms a, b a weighted average of their empirical mean:
µ̂a,b (t) :=
Na (t)
Nb (t)
µ̂a (t) +
µ̂b (t) ,
Na (t) + Nb (t)
Na (t) + Nb (t)
it is well known and easy to see that if µ̂a (t) ≥ µ̂b (t),
Za,b (t) = Na (t) d µ̂a (t), µ̂a,b (t) + Nb (t) d µ̂b (t), µ̂a,b (t) ,
(7)
τδ = inf {t ∈ N : ∃a ∈ A, ∀b ∈ A \ {a}, Za,b (t) > β(t, δ)}
= inf t ∈ N : Z(t) := max min Za,b (t) > β(t, δ) ,
(8)
and that Za,b (t) = −Zb,a (t). The testing intuition thus suggests the following stopping rule:
a∈A b∈A\{a}
where β(t, δ) is an exploration rate to be tuned appropriately. The form of this stopping rule can
be traced back to Chernoff (1959)3 . As minb∈A\a Za,b (t) is non-negative if and only if µ̂a (t) ≥
3. The stopping rule τδ was proposed under equivalent form (with a different threshold) in the context of adaptive sequential hypothesis testing. Best arm identification in a bandit model can be viewed as a particular instance in which
we test K hypotheses, Ha : (µa = maxi∈A µi ), based on adaptively sampling the marginal of ν = (ν µ1 , . . . , ν µK ).
However, Chernoff (1959) considers a different performance criterion, and its analysis holds when each of the hypotheses consists in a finite set of parameters, unlike the bandit setting.
8
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
µ̂b (t) for all b 6= a, Z(t) = minb∈A\ât Zât ,b (t) whenever
there is a unique best empirical arm
ât = argmaxa∈A µ̂a (t) at time t. Obviously, µ̂a (τδ ) a has a unique maximizer, which is the final
decision.
In addition to the testing interpretation given above, the stopping rule can be explained in light
of the lower bound Eµ [τδ /T ∗ (µ)] ≥ kl(δ, 1 − δ). Indeed, one may write
Z(t) = t
min
λ∈Alt(µ̂(t))
K
X
Na (t)
a=1
t
d(µ̂a (t), λa ) ≤
T∗
t
.
µ̂(t)
Theorem 1 suggests that a δ-PAC strategy cannot stop (at least in expectation) before t/T ∗ µ̂(t) is
larger than kl(δ, 1 − δ) ∼ log(1/δ), which suggests to stop when Z(t) ≥ log(1/δ). In Section 4,
we prove that a slightly more prudent choice of the threshold β(t, δ) does lead to a PAC algorithm
(whatever the sampling rule, even if the proportions of draws are sub-optimal). And it is shown in
Section 5 that, using our sampling rule, E[τδ ] is indeed of order T ∗ (µ) log(1/δ).
It is also possible to give a Minimum Description Length (MDL) interpretation of the stopping
rule. It is well known that choosing the model that gives the shortest description of the data is
a provably efficient heuristic (see Rissanen (1978), and Grünwald (2007) for a survey). In some
sense, the stopping rule presented above follows the same principle. In fact, elementary algebra
shows that
Za,b (t) = (Na (t) + Nb (t))h (µ̂a,b (t)) − [Na (t)h(µ̂a (t)) + Nb (t)h(µ̂b (t))] ,
where h(µ) = EX∼ν µ [− log pµ (X)] = b(ḃ−1 (µ)) − ḃ−1 (µ)µ. In the Bernoulli case, the Shannon
entropy h(µ) = −µ log(µ) − (1 − µ) log(1 − µ) is well-known to represent an ideal code length per
character for binary compression. Thus, Za,b (t) appears as the difference between the ideal code
length for the rewards of arms a and b coded together, and the sum of the ideal code lengths for the
rewards coded separately. If this difference is sufficiently large4 , the shortest description of the data
is to separate arms a and b. The stopping rule (8) thus consists in waiting for the moment when it
becomes cheaper to code the rewards of the best empirical arm separately from each of the others.
It is no surprise that the proof of Proposition 10 below is based on a classical information-theoretic
argument.
4. Choosing the Threshold in the Stopping Rule
We now explain how to choose the threshold β(t, δ) so as to ensure the δ-PAC property: with
probability larger than 1 − δ, any algorithm based on the stopping rule (8) outputs the optimal arm,
provided that it stops. The interpretations of the stopping rule presented in the last section suggest
the presence of two ingredients: log(1/δ) for the risk, and log(t) for the fluctuations of the counts.
We present here two results: one is based on an information-theoretic argument used for consistency
proofs of MDL estimators, and the second is based on the probabilistic control of self-normalized
averages taken from Magureanu et al. (2014). To keep things simple, the first argument is detailed
only for the case of Bernoulli rewards (the standard framework of coding theory). The second
argument is more general, but a little less tight.
4. Universal coding theory would suggest a threshold of order log(t)/2, a term that will appear in Section 4.
9
G ARIVIER K AUFMANN
The Informational Threshold.
Theorem 10 Let δ ∈ (0, 1). Whatever the sampling strategy, using Chernoff’s stopping rule (8)
on Bernoulli bandits with threshold
2t(K − 1)
β(t, δ) = log
δ
ensures that for all µ ∈ S, Pµ (τδ < ∞, âτδ 6= a∗ ) ≤ δ.
Proof sketch. A more detailed proof is given in Appendix C.1. We proceed here similarly to Vaidhyan and Sundaresan
(2015), employing an argument used for MDL consistency proofs (see Garivier (2006) and references therein). Introducing, for a, b ∈ A, Ta,b := inf{t ∈ N : Za,b (t) > β(t, δ)}, one has
Pµ (τδ < ∞, âτδ 6= a∗ ) ≤ Pµ (∃a ∈ A \ a∗ , ∃t ∈ N : Za,a∗ (t) > β(t, δ))
X
≤
Pµ (Ta,a∗ < ∞) .
a∈A\{a∗ }
It is thus sufficient to show that if β(t, δ) = log(2t/δ),
and if µa < µb , then Pµ (Ta,b < ∞) ≤ δ.
For such a pair of arms, observe that on the event Ta,b = t time t is the first moment when Za,b
exceeds the threshold β(t, δ), which implies by definition that
1 ≤ e−β(t,δ)
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
maxµ′a ≤µ′b pµ′a (X at )pµ′b (X bt )
.
It thus holds that
Pµ (Ta,b < ∞) =
≤
≤
=
∞
X
t=1
∞
X
t=1
∞
X
t=1
∞
X
Pµ (Ta,b = t) =
t=1
t=1
"
e−β(t,δ) Eµ 1(Ta,b =t)
−β(t,δ)
"
e
Eµ 1(Ta,b =t)
e−β(t,δ)
Z
{0,1}t
∞
X
i
h
Eµ 1(Ta,b =t)
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
#
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
#
maxµ′a ≤µ′b pµ′a (X at )pµ′b (X bt )
pµa (X at )pµb (X bt )
1(Ta,b =t) (x1 , . . . , xt ) max′ pµ′a (xat )pµ′b (xbt )
µ′a ≥µb
|
Y
pµi (xit ) dx1 . . . dxt .
i∈A\{a,b}
{z
(∗)
}
Of course the maximum likelihood (∗) is not a probability density. A possible workaround (sometimes referred to as Barron’s lemma, see Barron et al. (1998) and references therein) is to use a
universal distribution like Krichevsky and Trofimov (1981), which is known to provide a tight uniform approximation:
Lemma 11 [Willems et al. (1995)] Let pu (x) be the likelihood of successive observations x ∈
{0, 1}n of a Bernoulli random variable with mean u. Then the Krichevsky-Trofimov distribution
Z 1
1
p
pu (x)du
kt(x) =
0 π u(1 − u)
10
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
is a probability law on {0, 1}n that satisfies
√
pu (x)
≤2 n.
x∈{0,1}n u∈[0,1] kt(x)
sup
Together with the inequality
is upper-bounded by
∞
X
ab ≤ (a+b)/2, this property permits to conclude that Pµ (Ta,b < ∞)
:=I(x1 ,...,xt )
−β(t,δ)
2te
t=1
=
√
sup
∞
X
t=1
Z
{0,1}t
}| Y
z
a
b
1(Ta,b =t) (x1 , . . . , xt ) kt(xt )kt(xt )
h
i∈A\{a,b}
{
pµi (xit ) dx1 . . . dxt
i
2te−β(t,δ) Ẽ 1(Ta,b =t) ≤ δ P̃(Ta,b < ∞) ≤ δ,
using that the partially integrated likelihood I(x1 , . . . , xt ) is the density of a probability measure P̃
(we denote the corresponding expectation by Ẽ).
The Deviational Threshold. The universal coding argument above can be adapted to other distributions, as shown for example in Chambaz et al. (2009). It is also possible to make use of a
deviation result like Magureanu et al. (2014) in order to prove PAC guarantees in any exponential
family bandit model. The exploration rate involved are slightly larger and less explicit.
Proposition 12 Let µ be an exponential family bandit model. Let δ ∈ (0, 1) and α > 1. There
exists a constant C = C(α, K) such that whatever the sampling strategy, using Chernoff’s stopping
rule (8) with the threshold
α
Ct
β(t, δ) = log
δ
ensures that for all µ ∈ S, Pµ τδ < ∞, âτδ 6= a∗ ≤ δ.
The proof of Proposition 12 is given in Appendix C.2.
5. Sample Complexity Analysis
Combining Chernoff’s stopping rule and an optimal-tracking sampling rule permits to approach the
lower bound of Theorem 1 for sufficiently small values of the risk δ. We first state a simple almostsure convergence result and give its short proof. Then, we sketch the (somewhat more technical)
analysis controlling the expectation of τδ .
5.1. Almost-sure Upper Bound
Proposition 13 Let α ∈ [1, e/2] and r(t) = O(tα ). Using Chernoff’s stopping rule with β(t, δ) =
log(r(t)/δ), and any sampling rule ensuring that for every arm a ∈ A, Na (t)/t converges almostsurely to wa∗ guarantees that for all δ ∈ (0, 1), Pµ (τδ < +∞) = 1 and
τδ
∗
≤ αT (µ) = 1.
Pµ lim sup
δ→0 log(1/δ)
11
G ARIVIER K AUFMANN
Proof.
Let E be the event
Na (t)
∗
→ w and µ̂(t) → µ .
E = ∀a ∈ A,
t→∞
t t→∞ a
From the assumption on the sampling strategy and the law of large number, E is of probability 1.
On E, there exists t0 such that for all t ≥ t0 , µ̂1 (t) > maxa6=1 µ̂a (t) and
Z(t) = min Z1,a (t) = min N1 (t)d(µ̂1 (t), µ̂1,a (t)) + Na (t)d(µ̂a (t), µ̂1,a (t))
a6=1
a6=1
N1 (t) Na (t)
= t min
+
I
(µ̂
(t),
µ̂
(t))
.
N1 (t)/t
1
a
a6=1
t
t
N1 (t)/t+Na (t)/t
For all a ≥ 2, the mapping (w, λ) → (w1 + wa )Iw1 /(w1 +wa ) (λ1 , λa ) is continuous at (w∗ (µ), µ).
Therefore, for all ǫ > 0 there exists t1 ≥ t0 such that for all t ≥ t1 and all a ∈ {2, . . . , K},
w∗ + wa∗
N1 (t) Na (t)
(µ̂1 (t), µ̂a (t)) ≥ 1
+
I w1∗ (µ1 , µa ) .
I
N1 (t)/t
t
t
1 + ǫ w∗ +w
∗
N1 (t)/t+Na (t)/t
a
1
Hence, for t ≥ t1 ,
Z(t) ≥
t
t
min (w1∗ + wa∗ ) I w1∗ (µ1 , µa ) =
.
1 + ǫ a6=1
(1 + ǫ)T ∗ (µ)
∗ +w∗
w1
a
Consequently,
τδ = inf{t ∈ N : Z(t) ≥ β(t, δ)}
≤ t1 ∨ inf{t ∈ N : t(1 + ǫ)−1 T ∗ (µ)−1 ≥ log(r(t)/δ)}
≤ t1 ∨ inf{t ∈ N : t(1 + ǫ)−1 T ∗ (µ)−1 ≥ log(Ctα /δ)},
for some positive constant C. Using the technical Lemma 18 in the Appendix, it follows that on E,
as α ∈ [1, e/2],
Ce((1 + ǫ)T ∗ (µ))α
C((1 + ǫ)T ∗ (µ))α
∗
τδ ≤ t1 ∨ α(1 + ǫ)T (µ) log
+ log log
.
δ
δ
Thus τδ is finite on E for every δ ∈ (0, 1), and
lim sup
δ→0
τδ
≤ (1 + ǫ) α T ∗ (µ) .
log(1/δ)
Letting ǫ go to zero concludes the proof.
5.2. Asymptotic Optimality in Expectation
In order to prove that the lower bound of Theorem 1 is matched, we now give an upper bound on
the expectation of the stopping time τδ . The proof of this result is to be found in Appendix D.
12
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
Theorem 14 Let µ be an exponential family bandit model. Let α ∈ [1, e/2] and r(t) = O(tα ).
Using Chernoff’s stopping rule with β(t, δ) = log(r(t)/δ), and the sampling rule C-Tracking or
D-Tracking,
Eµ [τδ ]
≤ αT ∗ (µ) .
lim sup
log(1/δ)
δ→0
To summarize, for every Bernoulli bandit µ, the choice β(t, δ) = log(2(K − 1)t/δ) in Chernoff’s
stopping rule is δ-PAC (by Theorem 10); with one of the sampling rules given above, the stopping
time τδ is almost surely finite (by Proposition 13) and when δ is small enough its expectation is
close to T ∗ (µ) log(1/δ) by Theorem 14, an optimal sample complexity after Theorem 1.
More generally, for exponential family bandit models, combining Proposition 12 and Theorem 14, one obtains for every α > 1 the existence of an exploration rate for which Chernoff’s
stopping rule combined with a tracking sampling rule is δ-PAC and satisfies
lim sup
δ→0
Eµ [τδ ]
≤ αT ∗ (µ) .
log(1/δ)
6. Discussion and Numerical Experiments
We give here a few comments on the practical behaviour of the Track-and-Stop (T-a-S) strategy.
Let us first emphasize that the forced exploration step are rarely useful, but in some cases really
necessary and not only for the theorems: when µ2 and µ3 are equal, they prevent the probability
that the strategy never ends from being strictly positive. Second, our simulation study suggests that
the exploration rate β(t, δ) = log((log(t) + 1)/δ), though not (yet) allowed by theory, is still overconservative in practice. Further, even though any sampling strategy ensuring that Na (t)/t → wa∗
satisfies the optimality theorems above, we propose (without formal justification) an experimentally
more efficient sampling rule: after t observations, let
ĉt = argmin Zât ,c (t)
c∈A\{ât }
be the ’best challenger’ of the empirical champion ât . We choose At+1 = ât if
wâ∗t (µ̂(t))
Nât (t)
< ∗
∗ (µ̂(t))
Nât (t) + Nĉt (t)
wât (µ̂(t)) + wĉ(t)
and At+1 = ĉt otherwise (with forced explorations steps as in the D-Tracking rule).
We consider two sample scenarios µ1 = [0.5 0.45 0.43 0.4] and µ2 = [0.3 0.21 0.2 0.19 0.18],
and we choose δ = 0.1. This choice is meant to illustrate that our algorithm performs well even for
relatively high risk values (so far, optimality is proved only for small risks). We compare the Trackand-Stop algorithm based on D-Tracking and BestChallenger to algorithms from the literature designed for Bernoulli bandit models, namely KL-LUCB and KL-Racing (Kaufmann and Kalyanakrishnan,
2013). Racing algorithms proceed in rounds: at start, all arms are active; at each round, all active
arms are drawn once; at the end of a round, a rule determines if the empirically worst arm should
be eliminated. Call âr the empirically best arm after r rounds. In KL-Racing, arm b is eliminated if
its upper confidence bound max{q ∈ [0, 1] : rd(µ̂b,r , q) ≤ β(r, δ)} is smaller than the best arm’s
13
G ARIVIER K AUFMANN
lower bound min{q ∈ [0, 1] : rd(µ̂âr ,r , q) ≤ β(r, δ)}. We also introduced in the competition the
‘hybrid’ Chernoff-Racing strategy, which eliminates b if
µ̂â ,r + µ̂b,r
µ̂â ,r + µ̂b,r
+ rd µ̂b,r , r
> β(r, δ).
Zâr ,b = rd µ̂âr ,r , r
2
2
Table 1 presents the estimated average number of draws of the five algorithms in the two scenarios.
Our (Julia) code will be available online. We see that the use of the MDL stopping rule leads to
a clear improvement. Moreover, Chernoff-Racing significantly improves over KL-Racing, and its
performance is even close to that of our optimal algorithms.
µ1
µ2
T-a-S (BC)
3968
1370
T-a-S (D-Tracking)
4052
1406
Chernoff-Racing
4516
3078
KL-LUCB
8437
2716
KL-Racing
9590
3334
Table 1: Expected number of draws Eµ [τδ ] for δ = 0.1, averaged over N = 3000 experiments: µ1 = [0.5 0.45 0.43 0.4], w∗ (µ1 ) = [0.417 0.390 0.136 0.057]; µ2 =
[0.3 0.21 0.2 0.19 0.18], w∗ (µ2 ) = [0.336 0.251 0.177 0.132 0.104].
It should be emphasized that a Racing-type algorithm cannot reach the lower bound in general:
by construction, it forces the last two arms in the race (hopefully µ1 and µ2 ) to be drawn equally
often, which is sub-optimal unless w1∗ (µ) = w∗ (µ2 ) (a condition approximately matched only if
there is a large gap between the second and the third best arms). This is illustrated in the second
scenario µ2 of Table 1, where the ratio w1∗ /w2∗ is larger.
7. Conclusion
We gave a characterization of the complexity of best arm identification in the fixed confidencesetting, for a large class of bandit models with arms parameterized by their means. Our new lower
bound reveals the existence of optimal proportions of draws of the arms that can be computed efficiently. Our Track-and-Stop strategy, by combining a track of these optimal proportions with
Chernoff’s stopping rule, asymptotically matches the lower bound. In future work, instead of using
forced exploration steps within a plugin procedure, we will investigate optimistic (or robust-tonoise) sampling strategies in order to optimize the exploration and to obtain non-asymptotic sample
complexity bounds. Furthermore, we will
investigate the fixed-budget setting, for which we conjecture that P (ât 6= 1) & exp − t/T∗ (µ) with
T∗ (µ)−1 = sup
min
inf
w∈ΣK a∈{2,...,K} µa <ma <µ1
w1 d(ma , µ1 ) + wa d(ma , µa ) .
Acknowledgments
This work was partially supported by the CIMI (Centre International de Mathématiques et d’Informatique) Excellence program while Emilie Kaufmann visited Toulouse in November 2015. The
authors acknowledge the support of the French Agence Nationale de la Recherche (ANR), under
grants ANR-13-BS01-0005 (project SPADRO) and ANR-13-CORD-0020 (project ALICIA). They
thank Sébastien Gerchinovitz for stimulating discussions and comments.
14
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
References
Y. Abbasi-Yadkori, D.Pál, and C.Szepesvári. Improved Algorithms for Linear Stochastic Bandits.
In Advances in Neural Information Processing Systems, 2011.
S. Agrawal and N. Goyal. Thompson Sampling for Contextual Bandits with Linear Payoffs. In
International Conference on Machine Learning (ICML), 2013.
A. Antos, V. Grover, and C. Szepesvári. Active learning in multi-armed bandits. In Algorithmic
Learning Theory, 2008.
A. Barron, J. Rissanen, and Bin Yu. The minimum description length principle in coding and
modeling. Information Theory, IEEE Transactions on, 44(6):2743–2760, Oct 1998. ISSN 00189448. doi: 10.1109/18.720554.
S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed bandit
problems. Fondations and Trends in Machine Learning, 5(1):1–122, 2012.
S. Bubeck, R. Munos, G. Stoltz, and C. Szepesvári. X-armed bandits. Journal of Machine Learning
Research, 12:1587–1627, 2011.
O. Cappé, A. Garivier, O-A. Maillard, R. Munos, and G. Stoltz. Kullback-Leibler upper confidence
bounds for optimal sequential allocation. Annals of Statistics, 41(3):1516–1541, 2013.
Antoine Chambaz, Aurélien Garivier, and Elisabeth Gassiat. A MDL approach to HMM with poisson and gaussian emissions. application to order identification. Journal of Statistical Planning
and Inference, 139(3):962–977, 2009.
H. Chernoff. Sequential design of Experiments. The Annals of Mathematical Statistics, 30(3):
755–770, 1959.
R. Combes and A. Proutière. Unimodal Bandits without Smoothness. Technical report, 2014.
E. Even-Dar, S. Mannor, and Y. Mansour. Action Elimination and Stopping Conditions for the
Multi-Armed Bandit and Reinforcement Learning Problems. Journal of Machine Learning Research, 7:1079–1105, 2006.
V. Gabillon, M. Ghavamzadeh, and A. Lazaric. Best Arm Identification: A Unified Approach to
Fixed Budget and Fixed Confidence. In Advances in Neural Information Processing Systems,
2012.
Aurlien Garivier. Consistency of the unlimited BIC context tree estimator. IEEE Transactions on
Information Theory, 52(10):4630–4635, 2006.
T.L. Graves and T.L. Lai. Asymptotically Efficient adaptive choice of control laws in controlled
markov chains. SIAM Journal on Control and Optimization, 35(3):715–743, 1997.
Peter D. Grünwald. The Minimum Description Length Principle (Adaptive Computation and Machine Learning). The MIT Press, 2007. ISBN 0262072815.
15
G ARIVIER K AUFMANN
K. Jamieson, M. Malloy, R. Nowak, and S. Bubeck. lil’UCB: an Optimal Exploration Algorithm
for Multi-Armed Bandits. In Proceedings of the 27th Conference on Learning Theory, 2014.
S. Kalyanakrishnan, A. Tewari, P. Auer, and P. Stone. PAC subset selection in stochastic multiarmed bandits. In International Conference on Machine Learning (ICML), 2012.
E. Kaufmann and S. Kalyanakrishnan. Information complexity in bandit subset selection. In Proceeding of the 26th Conference On Learning Theory., 2013.
E. Kaufmann, O. Cappé, and A. Garivier. On the Complexity of A/B Testing. In Proceedings of the
27th Conference On Learning Theory, 2014.
E. Kaufmann, O. Cappé, and A. Garivier. On the Complexity of Best Arm Identification in MultiArmed Bandit Models. Journal of Machine Learning Research (to appear), 2015.
Raphail E. Krichevsky and Victor K. Trofimov. The performance of universal encoding. IEEE
Transactions on Information Theory, 27(2):199–206, 1981. doi: 10.1109/TIT.1981.1056331.
URL http://dx.doi.org/10.1109/TIT.1981.1056331.
T.L. Lai and H. Robbins. Asymptotically efficient adaptive allocation rules. Advances in Applied
Mathematics, 6(1):4–22, 1985.
S. Magureanu, R. Combes, and A. Proutière. Lipschitz Bandits: Regret lower bounds and optimal
algorithms. In Proceedings on the 27th Conference On Learning Theory, 2014.
S. Mannor and J. Tsitsiklis. The Sample Complexity of Exploration in the Multi-Armed Bandit
Problem. Journal of Machine Learning Research, pages 623–648, 2004.
R. Munos. From bandits to Monte-Carlo Tree Search: The optimistic principle applied to optimization and planning., volume 7. Foundations and Trends in Machine Learning, 2014.
J. Rissanen.
Modeling by shortest data description.
Automatica, 14(5):465 – 471,
1978.
ISSN 0005-1098.
doi: http://dx.doi.org/10.1016/0005-1098(78)90005-5.
URL
http://www.sciencedirect.com/science/article/pii/0005109878900055.
N. Srinivas, A. Krause, S. Kakade, and M. Seeger. Gaussian Process Optimization in the Bandit
Setting : No Regret and Experimental Design. In Proceedings of the International Conference
on Machine Learning, 2010.
N.K. Vaidhyan and R. Sundaresan. Learning to detect an oddball target. arXiv:1508.05572, 2015.
Frans M. J. Willems, Yuri M. Shtarkov, and Tjalling J. Tjalkens. The context tree weighting method:
Basic properties. IEEE Transactions on Information Theory, 41:653–664, 1995.
16
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
Appendix A. Characterization of the optimal proportion of draws
A.1. Proof of Lemma 3
Let µ such that µ1 > µ2 ≥ · · · ≥ µK . Using the fact that
[
Alt(µ) =
{λ ∈ S : λa > λ1 } ,
a6=1
one has
T ∗ (µ)−1 =
=
sup min
inf
K
X
w∈ΣK a6=1 λ∈S:λa >λ1 a=1
sup min
inf
w∈ΣK a6=1 λ:λa ≥λ1
wa d(µa , λa )
[w1 d(µ1 , λ1 ) + wa d(µa , λa )] .
Minimizing
f (λ1 , λa ) = w1 d(µ1 , λ1 ) + wa d(µa , λa )
under the constraint λa ≥ λ1 is a convex optimization problem that can be solved analytically. The
minimum is obtained for
wa
w1
µ1 +
µa
λ1 = λa =
w1 + wa
w1 + wa
and its value can be rewritten (w1 + wa )I
w1
w1 +wa
(µ1 , µa ), using the function Iα defined in (3).
A.2. Proof of Theorem 5
The function ga introduced in (4) rewrites
ga (x) = d (µ1 , ma (x)) + xd(µa , ma (x)), with ma (x) =
µ1 + xµa
.
1+x
d
Using that m′a (x) = (µa − µ1 )/(1 + x)2 and dy
d(x, y) = (y − x)/b̈(b−1 (y)) one can show that ga
is strictly increasing, since ga′ (x) = d(µa , ma (x)) > 0. As ga (x) tends to d(µ1 , µa ) when x goes
to infinity, the inverse function xa (y) = ga−1 (y) is defined on [0, d(µ1 , µa )[ and satisfies
x′a (y) =
1
>0.
d(µa , ma (xa (y)))
Let w ∗ be an element in
argmax min (w1 + wa )I
w∈ΣK
a6=1
w1
w1 +wa
(µ1 , µa ) = argmax w1 min ga
a6=1
w∈ΣK
wa
w1
.
The equality uses that w1∗ 6= 0 (since if w1 = 0, the value of the objective is zero). Introducing
wa∗
x∗a = w
∗ for all a 6= 1, one has
1
w1∗ =
1+
1
PK
∗
a=2 xa
and, for a ≥ 2, wa∗ =
17
1+
x∗a
PK
∗
a=2 xa
G ARIVIER K AUFMANN
and (x∗2 , . . . , x∗K ) ∈ RK−1 belongs to
argmax
(x2 ,...,xK )∈RK−1
mina6=1 ga (xa )
.
1 + x2 + · · · + xK
(9)
We now show that all the ga (x∗a ) have to be equal (Lemma 4). Let
∗
∗
B = b ∈ {2, . . . , K} : gb (xb ) = min ga (xa )
a6=1
and A = {2, . . . , K}\B. Assume that A =
6 ∅. For all a ∈ A and b ∈ B, one has ga (x∗a ) > gb (x∗b ).
Using the continuity of the g functions and the fact that they are strictly increasing, there exists
ǫ > 0 such that
∀a ∈ A, b ∈ B, ga (x∗a − ǫ/|A|)) > gb (x∗b + ǫ/|B|) > gb (x∗b ) .
Introducing xa = x∗a − ǫ/|A| for all a ∈ A and xb = x∗b + ǫ/|B| for all b ∈ B, there exists b ∈ B:
gb (x∗b + ǫ/|B|)
gb (x∗b )
mina6=1 ga (x∗a )
mina6=1 ga (xa )
=
>
=
,
1 + x2 + . . . xK
1 + x∗2 + · · · + x∗K
1 + x∗2 + · · · + x∗K
1 + x∗2 + · · · + x∗K
which contradicts the fact that x∗ belongs to (9). Hence A = ∅ and there exists y ∗ ∈ [0, d(µ1 , µ2 )[
such that
∀a ∈ {2, . . . , K}, ga (x∗a ) = y ∗ ⇔ x∗a = xa (y ∗ ),
with the function xa introduced above. From (9), y ∗ belongs to
G(y) with G(y) =
argmax
y∈[0,d(µ1 ,µ2 )[
y
.
1 + x2 (y) + · · · + xK (y)
G is differentiable and, using the derivative of the xa given above, G′ (y) = 0 is equivalent to
K
X
y
d(µa , ma (xa (y))
= 1 + x2 (y) + · · · + xK (y)
K
X
d(µ1 , ma (xa (y))) + xa (y)d(µa , ma (xa (y))
= 1 + x2 (y) + · · · + xK (y)
a=2
a=2
d(µa , ma (xa (y))
K
X
d(µ1 , ma (xa (y)))
a=2
d(µa , ma (xa (y))
= 1.
(10)
For the the second equality, we use that ∀a, d(µ1 , ma (xa (y))) + xa (y)d(µa , ma (xa (y)) = y. Thus
y ∗ is solution of the equation (10). This equation has a unique solution since
Fµ (y) =
K
X
d(µ1 , ma (xa (y)))
a=2
d(µa , ma (xa (y))
(11)
is strictly increasing and satisfies Fµ (0) = 0 and limy→d(µ1 ,µ2 ) Fµ (y) = +∞. As G is positive and
satisfies G(0) = 0, limy→d(µ1 ,µ2 ) G(y) = 0, the unique local extrema obtained in y ∗ is a maximum.
18
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
A.3. Proof of Proposition 6
Let µ ∈ S, and re-label its arms in decreasing order. From Theorem 5, w1 (µ) 6= 0 and for a ≥ 2,
wa∗ (µ) = 0 ⇔ xa (y ∗ ) = 0 ⇒ y ∗ = 0 where y ∗ is the solution of (10). But 0 is not solution of (10),
since the value of the left-hand side is 0. This proves that for all a, wa∗ (µ) 6= 0. For a given µ, y ∗ is
defined by
Fµ (y ∗ ) − 1 = 0 ,
where Fµ is defined in (11). For all µ ∈ S and every y ∈ [0, d(µ1 , µ2 )[, it can be shown that
d
d
∗
∗
dy Fµ (y) 6= 0, in particular dy Fµ (y ) 6= 0. Thus y is a function of µ that is continuous in every
∗
µ ∈ S, denoted by y (µ). By composition, the function µ 7→ xa (y ∗ (µ)) are continuous in µ ∈ S,
and so does w∗ .
The proof of Statement 3 relies on the fact that if a and b are such that µ1 > µa ≥ µb , ga (x) ≤
gb (x) for all x. Thus, for all y ∈ [0, d(µ1 , µ2 )[, xa (y) ≥ xb (y) and particularizing for y ∗ (µ) yields
the result.
A.4. Bounds on the characteristic time in the Gaussian case
In the Gaussian case, with d(x, y) = (x − y)2 /(2σ 2 ), the expression in Lemma 3 can be made more
explicit and yields
w1 wa ∆2a
T ∗ (µ)−1 = sup inf
.
2
w∈ΣK a6=1 w1 + wa 2σ
Introducing w̃ ∈ ΣK defined by
1/∆2
,
∀a = 1 . . . K, w̃a = PK a
2
i=1 1/∆i
it holds that
w̃1 w̃a ∆2a
1
= PK
a6=1 w̃1 + w̃a 2σ 2
i=1
T ∗ (µ)−1 ≥ inf
inf
2σ2 a6=1
∆2i
1
1+
∆21
∆2a
.
The infimum is obtained for a = 2, and using that ∆2 = ∆1 leads to the upper bound
!
K
2
X
2σ
T ∗ (µ) ≤ 2
.
∆2i
i=1
The following lower bound was obtained by Kaufmann et al. (2015) for every PAC strategy:
K
lim inf
δ→0
X 2σ 2
E[τδ ]
≥
.
log(1/δ)
∆2i
i=1
Combining this inequality with the upper bound on lim inf δ→0 E[τδ ]/log(1/δ) obtained in Theorem 14 for the δ-PAC Track-and-Stop algorithm shows that
T ∗ (µ) ≥
K
X
2σ 2
i=1
which concludes the proof.
19
∆2i
,
G ARIVIER K AUFMANN
Appendix B. Tracking results
Lemma 7 and 8 both follow from deterministic results that we can give for procedures tracking any
cumulated sums of proportions (Lemma 15) or any changing sequence of proportions that concentrates (Lemma 17). We state and prove theses two results in this section, and also explain how they
lead to Lemma 7 and 8. We then provide a proof of Proposition 9.
B.1. Tracking a cumulated sum of proportions
Lemma 15 Let K be a positive integer, let ΣK be the simplex of dimension K − 1, and for every
i ∈ {1, . . . , K}, let δi be the vertex of ΣK with a 1 on coordinate i. For a positive integer n, let
p(1), p(2), . . . , p(n) ∈ ΣK and for every k ≤ n let P (k) = p(1) + · · · + p(k). Define N (0) = 0,
for every k ∈ {0, . . . , n − 1}
Ik+1 ∈ argmax [Pi (k + 1) − Ni (k)]
1≤i≤K
and N (k + 1) = N (k) + δIk+1 . Then
max Ni (n) − Pi (n) ≤ K − 1 .
1≤i≤K
ǫk−1 (µ̂(k − 1)), so that
To obtain Lemma
Pk 7,ǫswe start by applying Lemma 15 with p(k) = w
P (k + 1) = s=0 w (µ̂(s)). One obtains
max Na (t) −
1≤a≤K
t−1
X
s=0
waǫs (µ̂(s)) ≤ K − 1 .
(12)
Moreover, by definition of wǫ (s),
max
1≤a≤K
t−1
X
waǫs (µ̂(s))
s=0
−
t−1
X
wa∗ (µ̂(s))
s=0
≤
t−1
X
Kǫs .
s=0
Now, with the choice ǫt = (K 2 + t)−1/2 /2, one has
p
t+
K2
−K =
Z
t−1
t
X
ds
√
ǫs ≤
≤
2 K 2 + s s=0
0
Z
t−1
−1
p
p
ds
√
= t + K2 − 1 − K2 − 1 ,
2 K2 + s
which yields
max Na (t) −
1≤a≤K
t−1
X
s=0
wa∗ (µ̂(s)) ≤ K − 1 + K
p
t + K2 − 1 −
p
√
K 2 − 1 ≤ K(1 + t) .
From (12), it also follows that
Na (t) ≥
t−1
X
s=0
ǫs − (K − 1) ≥
p
t + K 2 − K − (K − 1) ≥
which concludes the proof of Lemma 7.
20
p
t + K 2 − 2K,
O PTIMAL B EST A RM I DENTIFICATION
Proof of Lemma 15.
WITH
F IXED C ONFIDENCE
First, we prove by induction on k that
max Ni (k) − Pi (k) ≤ 1.
1≤i≤K
The statement is obviously true for k = 0. Assume that it holds for some k ≥ 0. For i 6= Ik+1 one
has Ni (k + 1) − Pi (k + 1) = Ni (k) − Pi (k) − pi (k) ≤ 1 − pP
i (k) ≤ 1, whereas NIk+1 (k + 1) −
PIk+1 (k + 1) = 1 + (NIk+1 (k) − PIk+1 (k + 1) ≤ 1, using that i (Pi (k + 1) − Ni (k)) = 0, hence
the largest term in this sum (which is for i = Ik+1 by definition) is positive.
It follows that, for all k,
max Ni (k) − Pi (k) = max max Pi (k) − Ni (k), max Ni (k) − Pi (k)
1≤i≤K
1≤i≤K
1≤i≤K
X
Pi (k) − Ni (k) + , 1} .
≤ max{
1≤i≤K
Introducing, for every k ∈ {1, . . . , n}
rk =
K
X
i=1
Pi (k) − Ni (k) + ,
Lemma 15 follow from the following bound on rk , that we prove by induction on k:
rk ≤ K − 1.
Observe that r0 = 0. For every k ≥ 0, one can write
rk+1 = rk +
K
X
i=1
pi (k + 1)1(Pi (k+1)≥Ni (k+1)) − 1(PI
k+1
(k+1)≥NIk+1 (k+1)) .
We distinguish two cases. If (PIk+1 (k + 1) ≥ NIk+1 (k + 1)) one has
rk+1 ≤ rk +
K
X
i=1
pi (k + 1) − 1 = rk ≤ K − 1.
If (PIk+1 (k + 1) < NIk+1 (k + 1)), which implies PIk+1 (k + 1) − NIk+1 (k) ≤ 1, one has
rk+1 =
K
X
i=1
i6=Ik+1
=
(Pi (k + 1) − Ni (k + 1))+ ≤
K
X
i=1
i6=Ik+1
K
X
i=1
i6=Ik+1
max (Pj (k + 1) − Nj (k))
j
PIk+1 (k + 1) − NIk+1 (k) ≤ (K − 1) ,
which concludes the proof.
Remark 16 This result is probably overly pessimistic, but one cannot hope for an upper bound
independent of K: for pi (k) = 1{i ≥ k}/(K −k+1), by choosing at each step K the smallest index
in the argmax defining Ik , one gets NK (K−1) = 0 and PK (K−1) = 1/K+1/(K−1)+...+1/2 ∼
log(K).
21
G ARIVIER K AUFMANN
B.2. Tracking a changing sequence that concentrates
Lemma 17 Let K be a positive integer and let ΣK be the simplex of dimension K − 1. Let
g : N → R be a non-decreasing function such that g(0) = 0, g(n)/n → 0 when n tends to infinity
and for all k ≥ 1 and ∀m ≥ 1,
inf{k ∈ N : g(k) ≥ m} > inf{k ∈ N : g(k) ≥ m − 1} + K.
Let λ̂(k) be a sequence of elements in ΣK such that there exists λ∗ ∈ ΣK , there exists ǫ > 0 and an
integer n0 (ǫ) such that
∀n ≥ n0 , sup |λ̂i (k) − λ∗i | ≤ ǫ .
1≤i≤K
Define N (0) = 0, and for every k ∈ {0, . . . , n − 1}, Uk = {i : Ni (k) < g(k)} and
Ni (k) if Uk 6= ∅,
argmin
i∈Uk
h
i
Ik+1 ∈
argmax
k
λ̂
(k)
−
N
(k)
else,
i
i
i∈{1,...,K}
and for all i Ni (k + 1) = Ni (k) + 1(Ik+1 =i) . Then for all i ∈ {1, . . . , K}, Ni (n) > g(n) − 1 and
there exists n1 ≥ n0 (that depends on ǫ) such that for all n ≥ n1 ,
Nn (i)
− λ∗i ≤ 3(K − 1)ǫ .
n
√
First it is easy to check that g(k) = ( k − K/2)+ satisfies the assumptions of Lemma 17.
Lemma 8 then follows by choosing λ̂(k) = w∗ (µ̂(k)) and λ∗ = w∗ (µ). The constant n1 in
Lemma 17 depends on ǫ, hence the notation tǫ .
The proof of this Lemma 17 is inspired by the proof of Lemma 3 in Antos et al. (2008), although
a different tracking procedure is analyzed.
max
1≤i≤K
Proof of Lemma 17. First, we justify that Ni (n) > g(n) − 1. For this purpose, we introduce for
all m ∈ N the integer nm = inf{k ∈ N : g(k) ≥ m}. We also let Im = {nm , . . . , nm+1 −1}. From
our assumption on g, it follows that |Im | > K and by definition, for all k ∈ Im , m ≤ g(k) < m+1.
We prove by induction that the following statement holds for all m:
∀k ∈ Im , ∀i, Ni (k) ≥ m. Moreover for k ≥ nm + K, Uk = ∅ and Ni (k) ≥ m + 1 .
(13)
First, for all k ∈ I0 , one has Uk = {i : Ni (k) = 0}. Therefore (I1 , . . . , IK ) is a permutation of
(1, . . . , K), thus for k ≥ K = n0 + K, Ni (k) ≥ 1, and Uk = ∅, and the statement holds for m = 0.
Now let m ≥ 0 such that the statement is true. From the inductive hypothesis, one has
∀k ∈ Im+1 , ∀i, Ni (k) ≥ Ni (nm+1 − 1) ≥ m + 1 .
Besides, as g(k) < m + 2 for k ∈ Im+1 , one has Uk = {i : Ni (k) = m + 1} and Ik is chosen
among this set while it is non empty. For k ≥ nm+1 + K, it holds that Uk = ∅ and Ni (k) ≥ m + 2
for all i. Thus the statement holds for m + 1.
From the fact that (13) holds for all m, using that for k ∈ Im , m > g(k) − 1, it follows that for
all k, for all i, Ni (k) > g(k) − 1.
22
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
Now for all i ∈ {1, . . . , K}, we introduce Ei,n = Ni (n) − nλ∗i . Using that
K
X
Ei,n = 0,
(14)
i=1
leads to
sup |Ei,n | ≤ (K − 1) sup Ei,n .
i
i
Indeed, for every i, one has Ei,n ≤ supk Ek,n and
X
X
Ei,n = −
Ej,n ≥ −
sup Ek,n = −(K − 1) sup Ek,n .
j6=i
j6=i
k
k
To conclude the proof, we give an upper bound on supi Ei,n , for n large enough. Let n′0 ≥ n0
such that
∀n ≥ n′0 , g(n) ≤ 2nǫ and 1/n ≤ ǫ .
We first show that for n ≥ n′0 ,
(In+1 = i) ⊆ (Ei,n ≤ 2nǫ)
(15)
To prove this, we write
(In+1 = i) ⊆
Ni (n) ≤ g(n) or i = argmin Nj (n) − nλ̂j (n)
1≤j≤K
!
.
Now if Ni (n) ≤ g(n), one has Ei,n ≤ g(n) − nλ∗i ≤ g(n) ≤ 2nǫ, by definition of n′0 . In the
second case, one has
Ni (n) − nλ̂i (n) = min Nj (n) − nλ̂j (n)
j
Ei,n + n(λ∗i − λ̂i (n)) = min Ej,n + n(λ∗j − λ̂j (n)
j
Using the closeness of each λ̂j (n) to the corresponding λ∗j , as n ≥ n0 , yields
Ei,n + n(λ∗i − λ̂n (i)) ≤ min (Ej,n + nǫ) ≤ nǫ,
j
where we use that minj Ej,n ≤ 0 by (14). Using that |λ∗i − λ̂i (n)| ≤ ǫ as well, one obtains
Ei,n ≤ 2nǫ .
This proves (15).
Ei,n satisfies Ei,n+1 = Ei,n + 1(In+1 =i) − λ∗i , therefore, if n ≥ n′0 ,
Ei,n+1 ≤ Ei,n + 1(Ei,n ≤2nǫ) − λ∗i .
We now prove by induction that for every n ≥ n′0 , one has
Ei,n ≤ max(Ei,n′0 , 2nǫ + 1).
23
G ARIVIER K AUFMANN
For n = n′0 , this statement clearly holds. Let n ≥ n′0 such that the statement holds. If Ei,n ≤ 2nǫ,
one has
Ei,n+1 ≤ 2nǫ + 1 − λ∗i ≤ 2nǫ + 1 ≤ max(Ei,n′0 , 2(n)ǫ + 1)
≤ max(Ei,n′0 , 2(n + 1)ǫ + 1).
If Ei,n > 2nǫ, the indicator is zero and
Ei,n+1 ≤ max(Ei,n′0 , 2nǫ + 1) − λ∗i ≤ max(Ei,n′0 , 2(n + 1)ǫ + 1) ,
which concludes the induction.
For all n ≥ n′0 , using that Ei,n′0 ≤ n′0 and 1/n ≤ ǫ, it follows that
1 n′0
n′0
Ei,n
≤ (K − 1) max 2ǫ + ,
≤ (K − 1) max 3ǫ,
.
sup
n
n n
n
i
Hence there exists n1 ≥ n′0 such that, for all n ≥ n1 ,
sup
i
Ei,n
≤ 3(K − 1)ǫ ,
n
which concludes the proof.
B.3. Proof of Proposition 9
Because of the forced-exploration step, both tracking strategies satisfy ∀a, Na (t) → ∞. Thus, from
the law of large number, the event
E = µ̂(t) → µ
t→∞
is of probability one. For C-Tracking, it follows from Lemma 7 that
√
t−1
Na (t) 1 X ∗
K( t + 1)
−
.
wa (µ(s)) ≤
t
t
t
s=0
By continuity of w∗ in µ, for all a, wa∗ (µ̂(t)) → wa∗ (µ). Using moreover the Cesaro lemma, one
obtains that, on E, Na (t)/t → wa∗ (µ). For D-Tracking, we first use that for ω ∈ E, there exists
t0 (ǫ) such that
ǫ
,
sup max |wa∗ (µ̂(t)) − wa∗ (µ)| ≤
3(K − 1)
t≥t0 (ǫ) a
by continuity of the function λ 7→ w∗ (λ) in µ. Hence, using Lemma 8, there exists tǫ ≥ t0 such
that for all t ≥ tǫ ,
Na (t)
− wa∗ (µ) ≤ ǫ .
max
a
t
Hence, for this ω ∈ E, Na (t)/t(ω) → wa∗ (µ) for all a.
24
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
Appendix C. PAC guarantees
C.1. Proof of Proposition 10.
Recall that
Pµ (τδ < ∞, âτδ 6= a∗ ) ≤ Pµ (∃a ∈ A \ a∗ , ∃t ∈ N : Za,a∗ (t) > β(t, δ))
X
≤
Pµ (Ta,a∗ < ∞) ,
a∈A\A∗
where Ta,b := inf{t ∈ N : Za,b (t) > β(t, δ)}. To conclude the proof, we now show that for β(t, δ)
as in Proposition 10, for any a, b such that µa < µb , one has
δ
.
K −1
Pµ (Ta,b < ∞) ≤
Let a,b be such that µa < µb . One introduces fµ (xt , at ) the likelihood of observing the sequence
of rewards xt = (x1 , . . . , xt ) and sequence of actions at = (a1 , . . . , at ). One has
"
#
t
K
Y
Y
fµ (xt , at ) =
p(as |xs−1 , as−1 ) ,
pµi xit × p(a1 )
s=2
i=1
where xit is a vector that gathers the sequence of successive rewards from arm i up to time t (that
is a function of both xt and at ). fµ (xt , at ) is a probability density on X t × At . For any density h
supported on I, the (partially) integrated likelihood
#
"
Z
Z
t
Y
Y
i
a
b
p(as |xs−1 , as−1 )
pu xt h(u)du × p(a1 )
pu (xt ) h(u)du
Ih (xt , at ) =
pµi xt
R
R
i∈A\{a,b}
s=2
is also a probability distribution.
On the event (Ta,b = t), Za,b (t) exceeds for the first time the threshold β(t, δ), which implies
in particular (from the definition of Za,b (t)) that
−β(t,δ)
1≤e
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
maxµ′a ≤µ′b pµ′a (X at )pµ′b (X bt )
.
We use this fact in the first inequality below, whereas the second inequality is based on the fact that
µ satisfies µa < µb :
Pµ (Ta,b < ∞) =
≤
≤
∞
X
Pµ (Ta,b = t) =
t=1
t=1
∞
X
t=1
∞
X
t=1
∞
X
"
i
h
Eµ 1(Ta,b =t)
e−β(t,δ) Eµ 1(Ta,b =t)
−β(t,δ)
e
"
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
maxµ′a ≤µ′b pµ′a (X at )pµ′b (X bt )
Eµ 1(Ta,b =t)
|
maxµ′a ≥µ′b pµ′a (X at )pµ′b (X bt )
pµa (X at )pµb (X bt )
{z
(∗)
25
#
#
}
.
G ARIVIER K AUFMANN
We now explicit the expectation (∗) for Bernoulli distributions.
(∗) =
X
X
1(Ta,b =t) (xt , at ) max′
µ′a ≥µb
at ∈At xt ∈{0,1}t
pµ′a (xat )pµ′b (xbt )
Y
pµi (xit )
i∈A\{a,b}
"
p(a1 )
t
Y
s=2
p(as |xs−1 , as−1 )
Introducing, for a vector x,
kt(x) =
Z
0
we have that
1
1
p
πu(1 − u)
pu (x)du ,
max pµ′a (xat )pµ′b (xbt ) ≤ 2t × kt(xat )kt(xbt ),
µ′a ≥µ′b
which follows from Lemma 11 stated in the main text and the fact that for na , nb are such that
√ √
na + nb ≤ t, 4 na nb ≤ 2(na + nb ) ≤ 2t. Using this inequality to upper bound (∗), one
1[0,1] (u):
recognize the integrated likelihood associated to the density h(u) = √ 1
πu(1−u)
(∗) ≤ 2t
X
at
∈At
xt
X
1(Ta,b =t) (xt , at )Ih (xt , at ) = 2tP̃(Ta,b = t) ,
∈{0,1}t
where P̃ is an alternative probabilistic model, under which µa and µb are drawn from a Beta(1/2, 1/2)
(prior) distribution at the beginning of the bandit game. Finally, using the explicit expression of
β(t, δ),
Pµ (Ta,b < ∞) ≤
=
∞
X
t=1
2te−β(t,δ) P̃(Ta,b = t) ≤
∞
δ X
P̃(Ta,b = t)
K − 1 t=1
δ
δ
,
P̃(Ta,b < ∞) ≤
K −1
K −1
which concludes the proof.
C.2. Proof of Proposition 12.
The proof relies on the fact that Za,b (t) can be expressed using function Iα introduced in Definition 3. An interesting property of this function, that we use below, is the following. It can be
checked that if x > y,
Iα (x, y) = inf
αd(x, x′ ) + (1 − α)d(y, y ′ ) .
′
′
x <y
For every a, b that are such that µa < µb and µ̂a (t) > µ̂b (t), one has the following inequality:
Za,b (t) = (Na (t) + Nb (t))I
=
Na (t)
Na (t)+Nb (t)
(µ̂a (t), µ̂b (t))
inf Na (t)d(µ̂a (t), µ′a ) + Nb (t)d(µ̂b (t), µ′b )
µ′a <µ′b
≤ Na (t)d (µ̂a (t), µa ) + Nb (t)d(µ̂b (t), µb ) .
26
#
O PTIMAL B EST A RM I DENTIFICATION
WITH
F IXED C ONFIDENCE
One has
Pµ (τδ < ∞, âτδ 6= a∗ ) ≤ Pµ (∃a ∈ A \ a∗ , ∃t ∈ N : µ̂a (t) > µ̂a∗ (t), Za,a∗ (t) > β(t, δ))
≤ Pµ ∃t ∈ N : ∃a ∈ A \ a∗ : Na (t)d (µ̂a (t), µa ) + Na∗ (t)d(µ̂a∗ (t), µa∗ ) ≥ β(t, δ)
!
K
X
Na (t)d(µ̂a (t), µa ) ≥ β(t, δ)
≤ Pµ ∃t ∈ N :
a=1
≤
∞
X
t=1
eK+1
β(t, δ)2 log(t)
K
K
e−β(t,δ) .
The last inequality follows from a union bound and Theorem 2 of Magureanu et al. (2014), originally stated for Bernoulli distributions but whose generalization to one-parameter exponential families is straightforward. Hence, with an exploration rate of the form β(t, δ) = log(Ctα /δ), for
α > 1, choosing C satisfying
∞ K+1
X
e
(log2 (Ctα ) log t)K
t=1
KK
tα
≤C
yields a probability of error upper bounded by δ.
Appendix D. Expected sample complexity analysis
The proof of Theorem 14 relies on two ingredients: a concentration result for the empirical mean
µ̂(t), that follows from the forced exploration (Lemma 19) and the tracking lemma associated to
the sampling strategy used (Lemma 7, Lemma 8). We start with a small technical lemma that can
be checked directly, or that can be seen as a by-product of well-known bounds on the Lambert W
function.
Lemma 18 For every α ∈ [1, e/2], for any two constants c1 , c2 > 0,
c2 e
c2
α
log
+ log log α
x=
α
c1
c1
c1
is such that c1 x ≥ log(c2 xα ).
D.1. Proof of Theorem 14
To ease the notation, we assume that the bandit model µ is such that µ1 > µ2 ≥ · · · ≥ µK . Let
ǫ > 0. From the continuity of w∗ in µ, there exists ξ = ξ(ǫ) ≤ (µ1 − µ2 )/4 such that
Iǫ := [µ1 − ξ, µ1 + ξ] × · · · × [µK − ξ, µK + ξ]
is such that for all µ′ ∈ Iǫ ,
max |wa∗ (µ′ ) − wa∗ (µ)| ≤ ǫ.
a
In particular, whenever µ̂(t) ∈ Iǫ , the empirical best arm is ât = 1.
27
G ARIVIER K AUFMANN
Let T ∈ N and define h(T ) := T 1/4 and the event
ET (ǫ) =
T
\
t=h(T )
(µ̂(t) ∈ Iǫ ) .
The following lemma is a consequence of the “forced
exploration” performed by the algorithm, that
√
ensures that each arm is drawn at least of order t times at round t.
Lemma 19 There exist two constants B, C (that depend on µ and ǫ) such that
Pµ (ETc ) ≤ BT exp(−CT 1/8 ).
Then, exploiting the corresponding tracking Lemma, one can prove the following
Lemma 20 There exists a constant Tǫ such that for T ≥ Tǫ , it holds that on ET , for either CTracking or D-Tracking,
∀t ≥
√
T , max
a
Na (t)
− wa∗ (µ) ≤ 3(K − 1)ǫ
t
Proof. This statement is
√ obvious for D-Tracking, just by definition of Iǫ and by Lemma 8. For
C-Tracking, for any t ≥ T = h(T )2 , using Lemma 7, one can write, for all a,
Na (t)
− wa∗ (µ)
t
≤
≤
≤
t−1
t−1
1X ∗
Na (t) 1 X ∗
−
wa (µ̂(s)) +
w (µ̂(s)) − wa∗ (µ)
t
t s=0
t s=0 a
√
t−1
K( t + 1) h(T ) 1 X
+
+
|wa∗ (µ̂(s)) − wa∗ (µ)|
t
t
t
t=h(T )
2K
1
2K + 1
+
+ ǫ ≤ 3ǫ ,
+ǫ=
1/4
h(T )
T
T 1/4
whenever T ≥ ((2K + 1)/2ǫ)4 .
On the event ET , it holds for t ≥ h(T ) that ât = 1 and the Chernoff stopping statistic rewrites
max min Za,b (t) = min Z1,a (t) = min N1 (t)d(µ̂1 (t), µ̂1,a (t)) + Na (t)d(µ̂a (t), µ1,a (t))
a b6=a
a6=1
a6=1
N1 (t) Na (t)
(µ̂1 (t), µ̂a (t))
= t min
+
I
N1 (t)/t
a6=1
t
t
N1 (t)/t+Na (t)/t
!
Na (t) K
= tg µ̂(t),
,
t
a=1
where we introduce the function
g(µ′ , w ′ ) = min(w1′ + wa′ )I
a6=1
28
′
w1
′
w′ +wa
1
(µ′1 , µ′a ).
O PTIMAL B EST A RM I DENTIFICATION
F IXED C ONFIDENCE
WITH
Using Lemma 20, for T ≥ Tǫ , introducing
Cǫ∗ (µ) =
inf
µ′ :||µ′ −µ||≤α(ǫ)
w ′ :||w ′ −w ∗ (µ)||≤2(K−1)ǫ
g(µ′ , w ′ ) ,
√
on the event ET it holds that for every t ≥ T ,
∗
max min Za,b (t) ≥ tCǫ (µ) .
a
b6=a
Let T ≥ Tǫ . On ET ,
min(τδ , T ) ≤
≤
Introducing
√
T+
T
X
√
t= T
√
T+
T
X
√
t= T
1(τδ >t) ≤
√
T+
T
X
√
1(maxa minb6=a
Za,b (t)≤β(t,δ))
t= T
1(tCǫ∗ (µ)≤β(T,δ)) ≤
√
T+
β(T, δ)
.
Cǫ∗ (µ)
√
β(T, δ)
T0 (δ) = inf T ∈ N : T + ∗
≤T ,
Cǫ (µ)
for every T ≥ max(T0 (δ), Tǫ ), one has ET ⊆ (τδ ≤ T ), therefore
Pµ (τδ > T ) ≤ P(ETc ) ≤ BT exp(−CT 1/8 )
and
Eµ [τδ ] ≤ T0 (δ) + Tǫ +
∞
X
BT exp(−CT 1/8 ) .
T =1
We now provide an upper bound on T0 (δ). Letting η > 0 and introducing the constant
√
C(η) = inf{T ∈ N : T − T ≥ T /(1 + η)}
one has
r(T )
1
T
log
T0 (δ) ≤ C(η) + inf T ∈ N : ∗
≤
Cǫ (µ)
δ
1+η
1+α
∗
Dt
Cǫ (µ)
T ≥ log
,
≤ C(η) + inf T ∈ N :
1+η
δ
where the constant D is such that r(T ) ≤ DT α . Using again the technical Lemma 18, one obtains,
for α ∈ [1, e/2],
α(1 + η)
D(1 + η)α
De(1 + η)α
T0 (δ) ≤ C(η) +
+ log log
.
log
Cǫ (µ)
δ(Cǫ∗ (µ))α
δ(Cǫ∗ (µ))α
This last upper bound yields, for every η > 0 and ǫ > 0,
lim inf
δ→0
α(1 + η)
Eµ [τδ ]
≤
.
log(1/δ)
Cǫ∗ (µ)
29
G ARIVIER K AUFMANN
Letting η and ǫ go to zero and using that, by continuity of g and by definition of w∗ ,
lim Cǫ∗ (µ) = T ∗ (µ)−1
ǫ→0
yields
lim inf
δ→0
Eµ [τδ ]
≤ αT ∗ (µ) .
log(1/δ)
D.2. Proof of Lemma 19.
P (ETc )
≤
T
X
t=h(T )
P (µ̂(t) ∈
/ Iǫ ) =
K
T
X
X
t=h(T ) a=1
[P (µ̂a (t) ≤ µa − ξ) + P (µ̂a (t) ≥ µa + ξ)] .
√
√
Let T be such that h(T ) ≥ K 2 . Then for t ≥ h(T ) one has Na (t) ≥ ( t − K/2)+ − 1 ≥ t − K
for every arm a. Let µ̂a,s be the empirical mean of the first s rewards from arm a (such that
µ̂a (t) = µ̂a,Na (t) ). Using a union bound and Chernoff inequality, one can write
√
P (µ̂a (t) ≤ µa − ξ) = P µ̂a (t) ≤ µa − ξ, Na (t) ≥ t ≤
≤
t
X
√
s= t−K
exp(−sd(µa − ξ, µa )) ≤
t
X
√
P (µ̂a,s ≤ µa − ξ)
s= t−K
√
t−K)d(µa −ξ,µa )
−(
e
1 − e−d(µa −ξ,µa )
.
Similarly, one can prove that
√
e−( t−K)d(µa +ξ,µa )
.
P (µ̂a (t) ≥ µa + ξ) ≤
1 − e−d(µa +ξ,µa )
Finally, letting
C = min (d(µa − ξ, µa ) ∧ d(µa + ξ, µa )) and B =
a
K
X
a=1
eKd(µa −ξ,µa )
eKd(µa +ξ,µa )
+
1 − e−d(µa −ξ,µa ) 1 − e−d(µa +ξ,µa )
one obtains
P (ETc )
≤
T
X
t=h(T )
p
√
B exp(− tC) ≤ BT exp(− h(T )C) = BT exp(−CT 1/8 ) .
30
!
,
| 10math.ST
|
1
Online Thevenin Equivalent Parameter Estimation using
Nonlinear and Linear Recursive Least Square Algorithm
Md. Umar Hashmi
Department of Energy Science & Engineering
IIT Bombay
Bombay, Maharashtra
[email protected]
Rahul Choudhary
Department of Systems and Control
IIT Bombay
Bombay, Maharashtra
[email protected]
Abstract—This paper proposes method for detection, estimation of
Thevenin equivalent parameters to describe power system behavior.
Thevenin equivalent estimation is a challenge due to variation in
system states caused by power flow in the network. Thevenin
equivalent calculation based on changes in system with multiple
sources integrated with grid, isolated distributed generator system is
analysed and nonlinear least square fit estimation technique for
algorithm is adopted. Linear least square fit is used with a linearized
model. Performance evaluation of proposed method is carried out
through mathematical model, nonlinear and linear least square fit
based
algorithm
technique
and
simulation
through
MATLAB/SIMULINK package. Accurate grid and source side
impedance estimation technique is applicable for distributed
generation sources interfaced with grid to improve dynamic response,
stability, reliability when subjected to faults or any other
disturbances in network. Algorithm can accurately estimate Thevenin
equivalent of multiple sources connected in parallel simultaneously
with voltage and current phasor measurements at point of common
coupling. Mathematical analysis and simulation results validate the
effectiveness of proposed method.
Index Terms—Adaptive controller, Impedance, Least square fit
estimation, Point of common coupling, Thevenin equivalent.
I. INTRODUCTION
W
ith integration of large number of distributed generation
sources (DG) into distribution network has raised concerns
related to stability, power quality and reliability of the
supply [1-3]. Distributed power generation sources depending
upon their mode of operation affects line impedance of distributed
electricity network. Performance of grid connected converter
depends upon line impedance [4]. Islanding operation due to
faults or grid disturbances due to voltage or frequency sag if not
detected properly leads to poor power quality at point of common
coupling (PCC).Online measurement due its numerous advantages
is preferred to fulfill anti islanding requirement. The distortion in
network caused due to electrical loads and distributed generation
sources strongly depends upon grid impedance and amount of
apparent power connected [5].To optimize the operation and
performance of distributed generation sources in microgrid
network, time and frequency dependent grid parameter estimation/
calculation is key [5].
In literature several methods are proposed for impedance
estimation. Estimation based on control loop variations which
include new values of inductance and resistance in control loop to
improve performance of system is proposed in [6]. Grid
impedance estimation based on use of extra devices is proposed in
[7]. Injection of harmonics signal into the grid and use of
mathematical models to obtain new impedance parameters are
Jayesh G. Priolkar
Department of Electrical Engineering
Goa College of Engineering
Farmagudi, Ponda Goa
[email protected]
addressed in [8-10].Grid impedance using controlled excitation
based on frequency characteristics of inductance -capacitance–
inductance (LCL) filter resonance is reported in [3]. The various
methods used to determine grid impedance characteristics offline
by means of frequency response analysis is reported in [11-13].
The methods of equivalent impedance estimation can be
broadly classified as active and passive type [14]. Passive
methods utilize the existing disturbance present in power
networks, for example detection of low order harmonic frequency
impedance. In active methods in addition to regular operation,
forced disturbance is injected into the grid or distributed
generation network for parameter estimation. Regression analysis
is used to extract parameters from measured data to define
physical characteristics of system. Thevenin equivalent estimation
based on concept of regression is also reported in literature.
Recursive least square estimation technique based on varying
system states to determine Thevenin equivalent is proposed in
[15].
Detection and estimation of Thevenin equivalent impedance for
different application is reported in literature. For example, power
system fault detection, load matching for maximum power
transfer, state of charge estimation for battery bank [16],
simultaneous estimation of Thevenin equivalent of multiple
sources, for load management by load shedding in power system
network based on detection of undervoltage [17], voltage stability
margin adjustment and analysis for prevention of voltage collapse
by a real time voltage instability identification algorithm based on
local phasor measurements [18].
The main aim of our work is to develop the algorithm for
online Thevenin equivalent parameter estimation. It is active
method which is available to estimate Thevenin equivalent
voltage and current. To evaluate the performance of proposed
method, mathematical model, nonlinear least square fit based
algorithm
technique
and
simulation
through
MATLAB/SIMULINK package are done.
Paper is organized as follows, Introduction of paper is given in
section I, section II discusses in brief applications of online
Thevenin equivalent estimation algorithm, section III presents
parameter estimation for isolated power system, and section IV
discusses the parameter estimation for grid connected system.
Simulation results are discussed in section V. Section VI
concludes the paper.
II. APPLICATIONS OF THEVENIN EQUIVALENT ESTIMATION
Online Thevenin equivalent estimation algorithm finds
application in adaptive control of electrical processes for
performance improvement and preemptive corrective action can be
taken for safe system operation. Some of the applications have
been discussed in this section.
2
i) Power system fault detection: - The Thevennin equivalent of
power system will indicate the change of estimatedd parameters due
to fault in power system. The relationships betweeen the real fault
distance and the varying Thevenin equivalennt impedance is
presented in paper [19]. The proposed Thevvenin equivalent
estimation algorithm in our work uses balancced symmetrical
components for correct estimation. Balanced ground faults,
balanced short circuits and balanced impedancee change can be
identified using the presented algorithm.
ii) Load matching for maximizing power transsfer: - Maximum
power is transferred to a load resistor (RL) when the value of the
load resistor is selected to match the value oof the Thevenin
resistance (Rth) of the power source. Figure 1 dem
monstrates this for
a value of Rth of 10 ohms.
Load Power Response to Changes in Load Resistance forr Rth = 10 ohms
1000
Power (Watts)
800
600
iii) State of charge estimation of battery or battery bank: A
6] between the open circuit
linear relationship is presented in [16
voltage (VOC) and SOC
)
%) =
The values of a and b are found experrimentally [16]. Open circuit
voltage is the Thevenin equivalent voltage of a battery under no
load condition.
ustment and analysis for
iv) Voltage stability margin adju
prevention of voltage collapse: Voltage instability is a major
concern for power systems operatiion. A Real-Time voltage
instability identification algorithm Based on Local Phasor
Measurements is proposed in [18] which recognize instabilities in
voltage. The power transferred to the bus reaches its voltage
stability limit when that Thevenin impedance has the same
magnitude as the load impedance at the bus [18].
The apparent power supplied is S, Y = 1/ZL, than
400
=
200
0
0
10
20
30
Load Resistance(Ohms)
40
50
Figure 1: Maximum Power Transfer at Zload=
=Zsource
Maximum Power Transfer with Complex Impedance
)
))^
(2)
The condition for maximum load ap
pparent power is dS/dY =0,
hence the critical point of voltage insstability is Zth = ZL. Hence
maximum loading point can be accurrately monitored online by
calculating dS/dY. The value of dS/d
dY close to zero indicates
proximity to voltage collapse point [20]].
The proposed algorithm assumes no simplification in system for
estimating Thevenin equivalent parameters.
Power Transfer (Watts)
1000
III. PARAMETER ESTIMATION FOR ISOLATED
M
SYSTEM
800
600
400
200
0
5
20
15
0
Description of the Proposed Algorith
hm:
Consider a voltage source feeding a lo
oad which has an impedance
of itself, line and filter impedance befo
ore it’s actually connected to
the load. The Thevenin equivalent of the
t circuit is shown adjacent
to it in the figure below.
10
Impedance (Ohms)
5
-5
Resiistance (Ohms)
0
Figure 2: Maximum Power Transfer with complexx impedance
2
P = I R, PL = [Vth/ (Rth+ RL)]2(RL)
(1)
One example is load matching is for wirelesss power transfer
circuitry design using inductive coils. Four coill wireless power
transfer simulation diagram along with the matcching network is
shown in Figure 3. Parameter values of the matchhing network are
dependent on the Thevenin equivalent impedancee of the four coil
system.
Matching network shown in figure 3 is one of thhe many possible
architectures of a matching network.
Figure 3: Matching network design for 4 coil wireless pow
wer transfer circuit
Figure 4: Isolated system with itss Thevenin equivalent
Using Kirchhoff’s Voltage Law we can write the voltage loop
equation as:
VT
cos θ j sin θ) = VPCC cos
j sin ) IPCC cos φ
j sin φ) × R T
jXT
)
(3)
We can separate the real and imaaginary component of the
equation as,
VT
cos θ = VPCC cos
IPCC cos φ × R T
sin φ ×
XT
)
(4)
VT
sin θ = VPCC sin
IPCC co
os φ × XT
RT
× sin φ)
(5)
3
The algorithm will be able to estimate the Theevenin equivalent
voltage and current. The point of common coupliing is the point at
which the load is connected. The known and unkn
known parameters
are shown below in table 1:
Table 1: State of Parameters
Known Parameters
Voltage Magnitude at PCC (VPCC)
Start
Unknown P
Parameters
Thevenin Voltage (VThevenin) ( 1)
Voltage Angle at PCC ( )
Thevenin Voltagee Angle (θ) (
Current Magnitude at PCC (IPCC)
Thevenin Resistancee (RThevenin) ( 3)
Current Angles at PCC (φ)
Thevenin Inductancee (XThevenin)(
VPCC cos = VT
XT
)
cos θ
IPCC cos φ × R T
VPCC sin = VT
RThevenin×sinφ)
sin θ
IPCC cos φ × XT
Equation (6) and (7) are non-linear in nature.
cos )
=
sin )
Where, y = VPCC cos , = VPCC sin , a = IPCC cos φ,
b = IPCC sin φ are known parameters and = VT
,
= RT
, = XT
are unknown param
meters.
=
Where
, , )
=
Estimation of Thevenin equivalent uses local measurements of at
least two different voltage and curreent vectors (magnitude and
phase) pairs measured at different time with different values
nin equivalent parameters.
associated with same reference Theven
Initialize Simulation
Variables
2)
For i = 1: number of iterations
s
4)
1. Define Load
2. Run Simulation
p
3. Get Voltage and Current (both phase
& magnitude ) at PCC
sin φ ×
(6)
Yes
(7)
No
Initial Guess for unknown
parameters X
Use
U
previous value of
unknown
u
parameter X
No
(8)
= ,
1. Minimize the error function E 2 for X
[E2 = f(Vpcc,Ipcc, load, Vsourse
e)]
2. Get X using least square meth
hod
Store
unknown
parameter X
Display X
(9)
and
If i = 1
If i = iteration limit
=
Yes
end
= min ∑
)
(10)
It is intended to solve these equations to obtaain the values X
which satisfy the system of equations describedd in equation 10.
Initial guess values are random numbers. Equattion 10 has to be
minimized for optimal values of unknown param
meters. Ideally β
has to be zero but since the algorithm recursiveely optimizes the
unknown parameters, there is a residual error w
which has to be
maintained as low as possible for accurate ressults. Residual is
defined by equation (11).
=
(11)
Nonlinear Recursive algorithm needs two Phasoor measurements
of voltage and current at the point of commoon coupling. The
figure below shows the circuit diagram for two different loading
scenarios. Phasor measurements are done at steaddy state.
Figure 7: Flow chart for the proposed algorithm
Parameters Estimation using linear regression
Non-linearity of the Equation (8) can be
b taken care by eliminating
the trigonometric variables.
=
=
(12)
1
0
0
1
×
(13)
Where, y = VPCC cos , = VPCC sin , a = IPCC cos φ, b =
IPCCsinφ are known parameters and 1=VThevenincosθ,
= VT
sin θ,
= RT
, = XT
are unknown
parameters.
a
Additional variable can be calculated as
(14)
VT
=
)
= tan
(15)
For n different loads, Equation (13) caan be written in following
structures (augmentation of matrix),
Figure 5: Circuit description of two data set coollection
=
(16)
Where dimension of the matrices are as,
a
For given initial estimate of
Figure 6: Phasor diagram of at least two different measurem
ments is required for
operation of estimation algorithm.
=
Estimation error
=
=
=
=
×
=
×
×
= ,
(17)
(18)
(19)
4
Error between actual and estimate will be minimuum for
)
=0
=0
)
=
Vector
(20)
(21)
(22)
provides good estimate of unknown paraameters.
IV. PARAMETER ESTIMATION FO
OR GRID
CONNECTED SYSTEM
Consider DG source connected to grid as shown in figure 8.
The total impedance of network includes line,, filter and load
impedance.
Thevenin Estimation Results:
Positive sequence voltage and current angle annd magnitude are
measured. It is assumed that the input voltagee magnitude and
angle and the Thevenin impedance as seen from
m the load side is
constant while the algorithm is in operation. A
As the estimation
gives an output which is within ±1% error wiith just one load
change; and hence it does not take more than a few seconds for
accurate prediction.
The positive sequence extraction is done too eliminate the
unbalanced components in voltage and currentt. The impact of
negative sequence is in terms of losses or heeating caused by
unbalances, for this case it’s not that relevant.
The results of the estimation with multiple numbbers of iterations
are tabulated below table.
Table 2: Estimation of Parameters for Isolatedd system
No. of Phasor
Generator
Voltage
Resistaance
Inductance
Measurements
Voltage
angle
Actual
70.7107
1
0.377
2 Phasor
70.642
1.98E-05
0.984442
0.3747
Error %
0.09706
1.557333
0.5993
3 Phasor
70.642
2.98E-05
0.9843399
0.374741
Error %
0.09715
1.56001
0.5992
5 Phasor
70.6254
1.11E-05
0.978555
0.373026
Error %
0.12057
2.144227
1.05397
10 Phasor
70.576
2.76E-05
0.96009
0.3691
Error %
0.1905
3.911
2.0955
100 Phasor
70.5762
1.34E-06
0.961
0.3691
Error %
0.19015
3.899449
2.09503
The estimated values tabulated in Table 2 shows accurate
prediction with just two Phasor measurem
ments of known
parameters at point of common coupling.
•
Current from individual so
ources (Igen and Igrid in the
figure above) is needed
Equivalent Thevenin voltage (maagnitude and angle) and
impedance can be estimated. The resullts of estimation are listed in
Table 4 and 6. Generator Side Estim
mation known and unknown
parameters are listed in table 3. A hypothetical scenario is
simulated in which voltage levels off two sources connected in
parallel are drastically different. The
T
generator side voltage
magnitude is set at 70.7 V and the grid is assumed at 49.5 V. The
algorithm accurately estimates the volttage level difference.
Table 3: State of Parameters on Generator sid
de
Known Parameters
Unknown Parameters
Voltage Magnitude at PCC
Generator Voltage
Voltage Angle at PCC
Generator Voltage Angle
Current Magnitude of generator
Generator resistance
Current Angles
Generator Inductance
Parameter Estimated with their errors percentages
p
for generator
side parameter estimation is tabulated in Table 4.
Table 4: Estimation of Parameters with error on Generator side
No. of Phasor
Generator
Voltage
Resistance
Measurements
Voltage
angle
Actual
70.7107
1
2 Phasor
70.70932
-6.285
0.999933
Error %
0.00195
0.0067
3 Phasor
70.70948
-6.279
0.999942
Error %
0.00172
0.00577
5 Phasor
70.71005
-6.290
0.999971
Error %
0.00092
0.00289
0.377
0.376977
0.0061
0.376974
0.007004
0.376978
0.0059
Grid Side Estimation known and unkn
nown parameters are listed in
table 5.
Table 5: State of Parameters on grid side
Known Parameters
Voltage Magnitude at PCC
Voltage Angle at PCC
Current Magnitude from grid
Current Angles
Unknown Parameters
Grid Voltage
Grid Voltage Angle
Grid Resistance
Grid Inductance
Parameter Estimated with their errors:
Table 6: Estimation of parameters with errors on grid side
No. of Phasor Grid Voltage
Voltagee
Grid
Measurements
angle
Resistance
Actual
49.4975
0.5
2 Phasor
49.4973
-0.0019
9
0.5
Error %
0.0004
0
3 Phasor
49.49726
-6.29
0.500017
Error %
0.000483
-0.00334
5 Phasor
49.49754
-6.29
0.499
Error %
-8.26E-05
0.00112
Grid
Inductance
0.0377
0.0377
0
0.037699
0.00285
0.03769
0.00157
'n' number of Power Sources co
onnected in Parallel
Superposition theorem implementation of proposeed algorithm:
140
Known Parameters for Parame
eter Estimation
Unknown Parameters Estimat ed
120
Number of Parameters
The algorithm can be used for multiple soources parameter
estimation simultaneously. The information nneeded from the
system is that we should know the currents cominng out of each of
the sources and the common voltage vector at ppoint of common
coupling.
Inductance
100
80
60
40
20
0
Figure 8: Generator and grid connected in parallel tto feed a load
For parameter estimation for a system with multipple sources:
• Voltage magnitude and angle at PCC is needed.
0
5
10
15
20
0
25
Number of Sources in
n Parallel
30
35
Figure 9: Known & Unknown parameters forr ‘n’ sources connected in parallel
5
Generator
Voltage
70.711
15.395
78.228
47.984
32.140
70.602
0.154
70.651
0.084
70.599
0.158
70.583
0.181
70.636
0.105
Voltage
angle
39.479
-30.161
-0.151
-0.112
-0.143
-0.151
-0.110
Resistance
1.000
-8.829
982.878
-1.342
234.184
1.005
-0.483
1.006
-0.557
0.999
0.091
0.996
0.374
0.998
0.219
Inductance
0.377
-3.351
988.831
-6.005
1692.739
0.390
-3.496
0.392
-4.019
0.374
0.665
0.367
2.730
0.371
1.603
Linear recursive least square method derived for thevenin
equivalent estimation requires at least three phasor measurements
for accurate prediction of unknown parameters. The result for
linear recursive least square estimation is listed in table 7. The
results with 1st and 2ndphasor measurements are not reliable,
however the results converge with 3 phasor measurements and
own wards.
Compared to linear recursive fit, nonlinear recursive fit is better as
the error percentages are much lower as per results tabulated in
table 4, 6 and 7.
Nonlinear recursive method uses exact equation, therefore
computationally exhaustive, for estimation and hence the overall
accuracy of estimation is higher. Furthermore nonlinear recursion
estimation converges with 2 phasor measurement data but linear
method takes at least 3 phasor measurements for convergence.
The results generated using proposed algorithms has an upper
bound of 5000 as maximum function evaluations and maximum
number of iterations as 8000.
Power System Grid Impedance Change Detection
Power system undergoes various kinds of faults which lead to
overall impedance change of the system. Simulations are
conducted in MATLAB Simulink for isolated system to verify the
capability of proposed algorithm to detect impedance change in
Volts
70.711
70.7105
70.71
70.7095
0
2
4
6
8
10
Time(s)
Figure10: Voltage magnitude estimation
The voltage magnitude from source side is unchanged and only
line impedance is changed. Therefore the Thevenin voltage
magnitude and phase estimated value remains fairly stable.
-4
-9
Voltage Angle
x 10
-9.2
-9.4
-9.6
-9.8
-10
0
2
4
6
8
10
Figure 11: Voltage Phase estimation
Resitance
2.2
2
1.8
Ohms
No. of Phasor
Measurements
Actual
1 Phasor
Error %
2 Phasor
Error %
3 Phasor
Error %
4Phasor
Error %
5 Phasor
Error %
6 Phasor
Error %
10 Phasor
Error %
RMS Voltage Magnitude
70.7115
Angle
Table 7: Parameter Estimation with linear regression
power system. Single source simulations with impedance change
from 1+0.377j to 2+0.755jat 5 second in a 9.6 second simulation
is conducted. The algorithm is capable to detect the impedance
change very accurately. As the source voltage and phase remains
unchanged, the estimation remains fairly stable at its initial
values.
1.6
1.4
1.2
1
0.8
0
2
4
6
8
10
Time(s)
Figure12: Power source Resistance estimation
Reactance
0.8
0.7
Ohms
The results of estimation indicate accurate prediction of both
generator and grid side parameters simultaneously. This algorithm
can be extrapolated to ‘n’ number of parallel sources feeding a
load. The algorithm uses the shared current from the source and
voltage at point of common coupling for estimating Thevenin
equivalent parameters. Figure 9 shows the number of estimated
parameters and number of known parameters used for estimating
the unknown values using the proposed algorithm. For very large
number of sources connected in parallel, the algorithm proposed
estimates twice the number of unknown parameters. This aspect
of the proposed algorithm can help in controller design and tuning
for multiple sources i.e. inverters and/or UPS to be connected in
parallel. Virtual impedance design requires equivalent Thevenin
impedance seen at from load side for proper compensation of
parameter mismatches (Line and/or filter parameters).
Simultaneous estimation of multiple sources Thevenin’s
equivalent will help in proper selection of virtual impedance for
minimizing circulating current flowing among power sources.
0.6
0.5
0.4
0
1
2
3
4
5
Time(s)
6
7
Figure 13: Power System reactance
8
9
10
6
As the resistance and reactance values are doubled at 5sec, the
estimated values changes level around 5 sec. The error % for the
estimated quantities is well below 0.5% under steady state. The
positive sequence voltage and current angle and magnitude are
measured. It is assumed that the input voltage magnitude and
angle and the Thevenin impedance as seen from the load side is
constant while the algorithm is in operation. The positive
sequence extraction is done to eliminate the unbalanced
components in voltage and current. Proposed algorithm can be
used for detection of balanced power system faults using the
proposed parameter estimation algorithm using linearized and
nonlinear recursive least square algorithms evaluated through
simulation results shown from Figure 10 to 13.
VI. CONCLUSION
Online Thevenin equivalent parameter estimation using
nonlinear and linear recursive least square algorithm are proposed
and evaluated in this paper. Simultaneous parameter estimation of
‘n’ parallel source with known parameters at point of common
coupling i.e. current and voltage phasor are evaluated and
analysed. The proposed algorithms require at least two phasor
measurements for accurate prediction. The robustness of the
algorithms has also been tested and verified. Fault in power
system leads to impedance change. The Thevenin equivalent
estimation proposed in the paper can help to determine power
system behavior and adopt the corrective actions to improve
response of system network and stability. Simulation results
validating the performance of the proposed algorithms have also
been verified for power system impedance change detection.
From the analysis carried out it is found that nonlinear recursion is
comparatively better than linear recursion in terms of error
percentage of the estimation. However linear method is
computationally less exhaustive, more viable for online controller
implementation.
REFERENCES
[1]A. Azmy and I. Erlich, “Impact of distributed generation on the stability of
electrical power systems,” in proc. power engineering society general meeting,
IEEE 2005, pp. 1056-1063, 2005.
[2]S. Jian, “Small signal methods for AC distributed power system – A
Review,” IEEE Trans. on Power Electronics, vol.24, no.11, pp.2545-2554,
Nov.2009.
[3] M. Liserre, R. Teodorescu and F. Blaabjerg, “Stability of photovoltaic and
wind turbine grid connected inverters for large set of grid impedance values,”
IEEE Trans.on Power Electronics, vol.21, no.1, pp.263-272, Jan.2006.
[4] M.Liserre, F. Blaabjerg, and R.Teodorescu, “Grid Impedance estimation
via excitation of LCL filter resonance,” IEEE Trans. on Industry applications,
vol.43, no.5, Oct.2007.
[5] N. Hoffmann, and F. Fuchs , “Online grid estimation for control of grid
connected converters in inductive –resistive distributed power networks using
extended Kalmann filter,” IEEE conference ,pp.922-929, 2012.
[6] J. Vasquez, J. Gurerreo, A. Luna, P. Rodriguez, and R. Teoderscu,
“Adaptive control applied to voltage source inverters operating in grid
connected and islanded modes,” IEEE Transactions on Industrial Electronics,
Vol. 56 , No. 10, pp. 4088-4096, Oct. 2009.
[7] J. Huang, K. Corzine, M. Belkhyat, “ Small signal Impedance measurement
of power electronics based AC power systems using line to line current
injection,” IEEE Transactions on Power Electronics,vol.56, pp.4088-4096,
2009.
[8] M. Ciobotaru, R. Teodorescu and F. Blaabjberg, “On line grid
estimation based on harmonic injection for grid connected PV inverter,” In
proc. IEEE International Symposium on Industrial electronics, ISIE 2007.
[9]A. Timbus , P. Roriguez, R. Teodorescu, M. Ciobotaru, “Line Impedance
estimation using active and reactive power variations,” In proc. IEEE , PESC
,pp.1273-1279, 2007.
[10]L. Siminoaei, R. Teodorescu, F. Blaabjerg, U. Borup, “ A digital
controlled PV inverter with grid impedance estimation for ENS detection ,”
IEEE Trans. on Power Electronics, vol.20, no.6, pp. 1480-1490, Nov.2005.
[11]Z. Straozcyzk, “A method for real time wide identification of the source
impedance in power system,” IEEE Trans. on Instrumentation measurement,
vol.54,no.1,pp.377-385, Feb. 2005.
[12] M. Sumner, B. Palethorpe, D. Thomas,P. Zanchetta, and M. Piazza, “ A
Technique for power supply harmonic impedance estimation using controlled
voltage disturbance,” IEEE Trans. Power Electronics,vo.17,no.2, pp.207-212,
Mar.2002.
[13] M. Sumner, B. Palethorpe, and D. Thomas, “Impedance measurement
for improved power quality- part 1: the measurement technique,” IEEE Trans.
on Power Electronics, vol.19,no.3, pp.1442-1448, July 2004.
[14] S. Cobaceus, E. Bueno, D. Pizzrro,F. Rodriguez, and F.Huerta, “ Grid
Impedance monitoring system for distributed power generation electronic
interfaces,” IEEE Trans. on Instrumentation and Measurement, vol.58, no.9,
pp.3112-3121, Sept.2009.
[15] K. Vu, M. Begovic, D. Novosel, and M. Saha, “Use of local
measurements to estimate voltage stability margin,” IEEE Trans. on Power
systems, vol.14, pp.1029-1034, 1999.
[16] Vairamohan Bhaskar, “State of charge estimation for batteries,” Masters
Thesis, University of Tenesse, 2002.
[17]S. Tsai, K. Wong , “Adaptive undervoltage load shedding relay design
using Thevenin equivalent estimation”, in Proc. Power and energy society
general meeting, IEEE, 2008.
[18] S. Corsi, and G. Taranto, “A real time voltage instability identification
algorithm on local Phasor measurement,” IEEE Trans. on Power systems,
vol.23, no.3, Aug. 2008.
[19]C. Tsai, C. Chiachu, “Fault locating estimation using Thevenin equivalent
in power systems,” in proc. IPEC,IEEE conference, 2010.
[20] A. R. Phadke, M. Fozdar, K. R. Niazi, “A New Technique for on-line
monitoring of voltage Stability margin using local signals”, Fifteenth National
Power Systems Conference (NPSC), IIT Bombay, December 2008.
| 3cs.SY
|
1
Primary frequency regulation with load-side
participation: stability and optimality
arXiv:1602.02800v1 [math.OC] 8 Feb 2016
Andreas Kasis, Eoin Devane, and Ioannis Lestas
Abstract—We present a method to design distributed generation and demand control schemes for primary frequency
regulation in power networks that guarantee asymptotic stability
and ensure fairness of allocation. We impose a passivity condition
on net power supply variables and provide explicit steady state
conditions on a general class of generation and demand control
dynamics that ensure convergence of solutions to equilibria
that solve an appropriately constructed network optimization
problem. We also show that the inclusion of controllable demand
results in a drop in steady state frequency deviations. We discuss
how various classes of dynamics used in recent studies fit within
our framework and show that this allows for less conservative
stability and optimality conditions. We illustrate our results with
simulations on the IEEE 68 bus system and observe that both
static and dynamic demand response schemes that fit within our
framework offer improved transient and steady state behavior
compared with control of generation alone. The dynamic scheme
is also seen to enhance the robustness of the system to time-delays.
I. I NTRODUCTION
Large scale integration of renewable sources of energy
within the power grid is expected to cause fast changes in
generation, making power imbalances increasingly frequent
due to the inability of conventional means of generation
to counter-balance them. Load participation is considered
to be one potential solution to this problem, providing fast
response to power changes. Household appliances like air
conditioning units, heaters, and refrigerators can be controlled
to adjust frequency and regulate power imbalances. Although
the idea dates back to the 1970s [1], research attention has
recently increasingly focused on the concept of controllable
demand [2], with particular consideration given to its use for
primary control [3], [4]. Primary or droop control operates on
a timescale of tens of seconds and is responsible for ensuring
that generation and demand are always balanced [5]. It functions by adjusting governor set-points based on local frequency
deviations and is completely decentralized. Simulation studies
and field tests (e.g. [6], [7]) have shown the feasibility of
including controllable demand in the droop control process.
Furthermore, they have shown that local frequency measurements are sufficient as a control signal for load participation
in primary control.
An issue of fairness is raised if appliances are to be used as a
means of controllable demand. Recent studies have attempted
to address this issue by devising control schemes which solve
Andreas Kasis and Ioannis Lestas are with the Department of Engineering,
University of Cambridge, Trumpington Street, Cambridge, CB2 1PZ, United
Kingdom; e-mails: [email protected], [email protected]
Eoin Devane is with the Cambridge Centre for Analysis, Centre for Mathematical Sciences, University of Cambridge, Wilberforce Road, Cambridge,
CB3 0WA, United Kingdom; e-mail: [email protected]
an optimization problem guaranteeing fair allocation between
them. This approach has been studied for primary and also for
secondary control. We consider here primary rather than secondary control in order to avoid the additional communication
that would be necessary to get a fair allocation if controllable
demand were used in the latter case. This is because it is
evident that a synchronizing variable is necessary to achieve
optimality, allowing all nodes to adapt their generation and
controllable demand so as to attain equal marginal costs. In
primary control, frequency deviation from the nominal value
can be used for this purpose, allowing decentralized control to
be achieved [8], [9]. In [8] only primary control in loads and
linearised power flows are considered while in [9] these assumptions are waived, considering also generation control and
non-linear power flows. In secondary control, where frequency
deviations return to zero, a different variable needs to be
used for synchronization. A common approach is to introduce
additional information exchange between neighbouring buses
in order to achieve this [10], [11], [12], [13] or have special
network structures like star topology, as in [14].
In this paper, we consider a network model described by
nonlinear swing equations. We consider a general class of
dynamics for power generation and controllable demand, on
which we impose appropriate conditions so as to achieve
stability of the equilibrium points and an optimization interpretation of those. This allows us to guarantee, for a wide variety
of possible generation and demand dynamics, convergence to a
power allocation that solves an appropriately constructed optimization problem, thus ensuring fairness in this allocation. The
class of dynamics considered incorporates control schemes
using only local frequency measurements as input signals, and
we demonstrate that this is sufficient to enable them to take the
right decisions so as to converge to a global optimum, thus
allowing distributed control. Further, we also prove that the
inclusion of controllable demand can aid in frequency control
by reducing the steady state frequency error. We illustrate the
applicability of our approach by demonstrating that various
dynamics that have been used in recent interesting studies,
such as [8] and [9], can be incorporated within our framework,
and we show that the analysis presented in the paper allows
for less conservative stability and optimality conditions.
It should be noted that one of the distinctive features of our
analysis is that optimality of the power allocation is provided
via appropriate conditions on the input/output properties of the
systems considered. In our companion paper [16] we show
how additional local information can be exploited to relax a
passivity condition used in this paper to deduce convergence.
The paper is organized as follows. In Sections II and III
we start with some basic notation and definitions that are
2
used in the rest of the paper. In Section IV, we present the
power network model and introduce the class of generation and
load dynamics that will be considered. Section IV introduces
the optimization problem that we consider and Section V
presents our main results, which are proved in Appendix
A. In Section VI, we discuss how our analysis relates to
other important studies. Section VII illustrates our results
through simulations on the IEEE 68 bus system. Finally,
conclusions are drawn in Section VIII. In Appendix B we
present a generalization of the optimality results in the paper
whereby the cost functions considered are allowed to have
discontinuous derivatives, which, as discussed in the paper,
can be important in practical implementations.
0. The region of attraction1 of x̄ is denoted by X0 . We also
define the static input-state characteristic map kx : R → Rm ,
kx (ū) := x̄.
(2)
Based on this, we can also define the static input-output
characteristic map ky : R → R,
ky (ū) := g(kx (ū), ū).
(3)
The requirement that for any constant input to (1) there exists
a unique equilibrium point, could be relaxed to require only
isolated equilibrium points, however, we assume it here to
simplify the presentation.
IV. P ROBLEM FORMULATION
II. N OTATION
A. Network model
Real numbers are denoted by R, and the set of ndimensional vectors with real entries is denoted by Rn . For a
d
f (q).
function f (q), we denote its first derivative by f 0 (q) = dq
−1
The expression f (w) represents the preimage of the point w
under the function f , i.e. f −1 (w) = {q : f (q) = w}. When the
function f is invertible, f −1 then defines the inverse function
of f . A function f : Rn → R is said to be positive definite
on a neighbourhood D around the origin if f (0) = 0 and
f (x) > 0 for every non-zero x ∈ D. It is positive semidefinite
if the inequality > 0 is replaced by ≥ 0. Furthermore, for
a, b ∈ R, a ≤ b, the expression [q]ba will be used to denote
max{min{q, b}, a}. The indicator function 1S : Rn → {0, 1}
of a set S ⊆ Rn takes the value 1 if its argument belongs to
the set S and 0 otherwise.
For a function of time q(t), we denote its derivative with
respect
R ∞ −stto time by q̇ and its Laplace transform by q̂ =
e q(t) dt. We use L2 to denote the Hilbert
0
qR space of
∞ 2
functions f : R → R with finite norm kf k2 =
f (t)dt.
0
For a system as in (1) where x = x̄, u = y = 0 is an
equilibrium point, its L2 -gain is defined as the supremum of
the ratio of the L2 -norms of the output y and the input u,
kyk
i.e. supkuk2 6=0 kuk2 , with x(0) = x̄. For a stable linear system
2
with transfer function T (s) it is known that its L2 -gain is
given by its ∞-norm kT k∞ := sup |T (jω)|.
ω
III. P RELIMINARIES
Throughout the paper we will consider dynamical systems
with input u(t) ∈ R, state x(t) ∈ Rm , and output y(t) ∈ R
with a state space realization of the form
ẋ = f (x, u),
y = g(x, u),
The power network model is described by a graph (N, E)
where N = {1, 2, . . . , |N |} is the set of buses and E ⊆ N ×N
the set of transmission lines connecting the buses. There are
two types of buses in the network, generation and load buses.
Their main difference is that generation buses have non-zero
generation inertia in contrast with load buses. Correspondingly,
only generation buses have nontrivial generation dynamics.
Let G = {1, 2, . . . , |G|} and L = {|G| + 1, . . . , |N |} be the
sets of generation and load buses such that |G| + |L| = |N |.
Furthermore, we use (i, j) to denote the link connecting buses
i and j and assume that the graph (N, E) is directed with
arbitrary direction, so that if (i, j) ∈ E then (j, i) ∈
/ E.
For each j ∈ N , we use i : i → j and k : j → k to
denote the sets of buses that are predecessors and successors of
bus j respectively. It is important to note that the form of the
dynamics in (4)–(5) below is unaltered by any change in the
graph ordering, and all of our results are independent of the
choice of direction. We also assume that (N, E) is connected.
The following assumptions are made for the network:
1) Bus voltage magnitudes are |Vj | = 1 p.u. for all j ∈ N .
2) Lines (i, j) ∈ E are lossless and characterized by their
susceptances Bij = Bji > 0.
3) Reactive power flows do not affect bus voltage phase angles
and frequencies.
The rate of change of frequency at the generation buses can
then be represented using swing equations, while power must
be conserved at each of the load buses. This motivates the
following system dynamics (e.g. [5])
η̇ij = ωi − ωj , (i, j) ∈ E,
(4a)
X
X
M
c
u
Mj ω̇j = −pL
pjk +
pij , j ∈ G,
j +pj −(dj +dj )−
0=
(1)
−pL
j
−
(dcj
+
duj )
−
m
pjk +
k:j→k
pij = Bij sin ηij −
m
i:i→j
k:j→k
X
pnom
ij ,
X
(4b)
pij , j ∈ L, (4c)
i:i→j
(i, j) ∈ E.
(4d)
m
where f : R × R → R is locally Lipschitz and g : R ×
R → R is continuous. We assume in system (1) that given
any constant input u(t) ≡ ū ∈ R, there exists a unique locally
asymptotically stable equilibrium point x̄ ∈ Rm , i.e. f (x̄, ū) =
1 That is, for the constant input u = ū, any solution x(t) of (1) with initial
condition x(0) ∈ X0 must satisfy x(t) → x̄ as t → ∞. The definition
of local asymptotic stability also implies that this region X0 has nonempty
interior.
3
u
c
In system (4) the quantities pM
j , ωj , dj and dj are timedependent variables representing, respectively, deviations from
a nominal value2 for the mechanical power injection to the
generator bus j, and the frequency, controllable load and
uncontrollable frequency-dependent load present at any bus j.
The variables ηij and pij represent, respectively, the power
angle difference3 and the deviation from nominal value pnom
ij
for the power transmitted from bus i to bus j. The constant
Mj > 0 denotes the generator inertia. We shall study the
response of system (4) to a step change pL
j in the uncontrollable demand at each bus j. To investigate decentralized
control schemes for generation and controllable load based
upon local measurements of the frequency alone, we close
c
u
the loop in (4) by determining each of pM
j , dj , and dj as
outputs from independent systems of the form in Section III
with inputs given by the negative of the local frequency,
ẋM,j = f M,j (xM,j , −ωj ),
M,j
pM
(xM,j , −ωj ),
j =g
ẋ
c,j
dcj
ẋ
=g
u,j
duj
=f
c,j
=f
=g
c,j
(x
u,j
u,j
(x
c,j
c,j
(x
(x
, −ωj ),
, −ωj ),
u,j
u,j
, −ωj ),
, −ωj ),
j ∈ G,
j ∈ N,
j ∈ N.
X
p∗jk +
p∗ij , j ∈ G,
i:i→j
k:j→k
(6b)
0=
c,∗
u,∗
−pL
j − (dj + dj ) −
X
k:j→k
p∗jk
+
X
p∗ij ,
j ∈ L, (6c)
i:i→j
xM,j,∗ = kxM,j (−ωj∗ ), j ∈ G,
(6d)
xc,j,∗ = kxc,j (−ωj∗ ), j ∈ N,
(6e)
xu,j,∗ = kxu,j (−ωj∗ ), j ∈ N,
(6f)
where the quantities in (6b) and (6c) are given by
∗
p∗ij = Bij sin ηij
− pnom
ij , (i, j) ∈ E,
(6g)
pM,∗
= k pM
(−ωj∗ ), j ∈ G,
j
j
(6h)
∗
c
dc,∗
j = kdj (−ωj ), j ∈ N,
(6i)
du,∗
= kduj (−ωj∗ ), j ∈ N.
j
(6j)
(5a)
(5b)
(5c)
For convenience in the notation, we collect4 the variables
in (5) into the vectors xM = [xM,j ]j∈G , xc = [xc,j ]j∈N , and
xu = [xu,j ]j∈N . These quantities represent the internal states
of the dynamical systems used to update the desired outputs
c
u
M
c
pM
j , dj , and dj . The variables pj and dj are controllable, so
we have freedom in our analysis to design certain properties
of the dynamics in (5a) and (5b). By contrast, duj represents
uncontrollable load and the dynamics in (5c) are thus fixed. It
should be noted that the systems in (5) can be heterogeneous
and of arbitrary dimension.
Throughout the paper we aim to characterize broad classes
of dynamics associated with generation and demand, so that
stability and optimality can be guaranteed for the equilibrium
points of the overall interconnected system (4)–(5). Various
examples of those will be discussed in Section VI, with
simulations also provided in section VII.
B. Equilibrium analysis
We now quantify what is meant by an equilibrium of the
interconnected system (4)–(5).
Definition 1: The constants5 (η ∗ , ω ∗ , xM,∗ , xc,∗ , xu,∗ ) define an equilibrium of the system (4)–(5) if the following hold
0 = ωi∗ − ωj∗ , (i, j) ∈ E,
X
M,∗
u,∗
0 = −pL
−(dc,∗
j +pj
j +dj )−
(6a)
2 A nominal value of a variable is defined as its value at an equilibrium of
(4) with frequency equal to the nominal value of 50Hz (or 60Hz).
3 The quantities η
ij represent the phase differences between buses i and j,
given by θi − θj . The angles themselves must also satisfy θ̇j = ωj at all
j ∈ N , however, we omit this equation in (4) since the power transfers p are
functions only of the phase differences.
4 Note that each local variable (e.g. xM,j ) is a vector with multiple
components.
5 By constant we mean a variable independent of time.
We call (6) the equilibrium conditions for the system (4)–(5).
Remark 1: For any equilibrium with a given frequency
value ω ∗ , the uniqueness in the definition of the static inputstate characteristic in Section III immediately shows that the
values of pM,∗ , dc,∗ , and dc,∗ are all guaranteed to be unique.
By contrast, there can in general be multiple choices of η ∗
and p∗ such that the equilibrium equations (6) remain valid,
and thus it follows that the equilibrium power transfers p∗ij
between regions need not be unique. It can be shown that they
become unique under prescribed conditions on the network
structure, such as when (N, E) has tree topology.
Note that the static input-output characteristic maps kpM
,
j
kdcj , and kduj , relating power generation/demand with frequency, as defined in (3), completely characterize the effect
of the dynamics (5) on the behavior of the power system (4)
at equilibrium. In our analysis, we will consider a class of
dynamics within (5) for which the set of equilibrium points in
Definition 1 that satisfy Assumption 1 below is asymptotically
stable. Within this class, we then consider appropriate conditions on these characteristic maps such that the values of the
variables defined in (6h)–(6j) are optimal for an appropriately
constructed network optimization problem.
Throughout the remainder of the paper we suppose that
there exists some equilibrium of (4)–(5) as defined in Definition 1. We let (η ∗ , ω ∗ , xM,∗ , xc,∗ , xu,∗ ) denote any such
equilibrium and use (p∗ , pM,∗ , dc,∗ , du,∗ ) to represent the
corresponding quantities defined in (6g)–(6j). We now impose
an assumption on the equilibrium, which can be interpreted as
a security constraint for the power flows generated.
∗
Assumption 1: |ηij
| < π2 for all (i, j) ∈ E.
The stability and optimality properties of such an equilibrium
point will now be studied in the sections that follow.
4
C. Combined passive dynamics from generation and load
dynamics
In terms of the outputs from (5), define the net supply
variables
M
c
u
sG
(7a)
j = pj − (dj + dj ), j ∈ G,
c
u
sL
j = −(dj + dj ), j ∈ L.
(7b)
Correspondingly, their values at equilibrium can be written as
u,∗
L,∗
u,∗
sG,∗
= pM,∗
− (dc,∗
= −(dc,∗
j
j
j + dj ) and sj
j + dj ).
The variables defined in (7) evolve according to the dyL
namics described in (5). Consequently, sG
j and sj can be
viewed as outputs from these combined dynamical systems
with inputs −ωj .
We now introduce a notion of passivity for systems of the
form (1) which we will use for the dynamics of the supply
variables defined in (7) to prove our main stability results.
Definition 2: The system (1) is said to be locally input
strictly passive about the constant input values ū and the
constant state values x̄ if there exist open neighbourhoods U
of ū and X of x̄ and a continuously differentiable, positive
semidefinite function V (x) (called the storage function) such
that, for all u ∈ U and all x ∈ X,
T
V̇ (x) ≤ (u − ū) (y − ȳ) − φ(u − ū)
(8)
where φ is a positive definite function and ȳ = ky (ū).
We now suppose that the maps from negative frequency
to supply satisfy this condition of input strict passivity about
equilibrium. Note that this is a decentralized condition.
Assumption 2: Each of the systems defined in (5) with inputs
−ωj and outputs given by (7a) and (7b) respectively are locally
input strictly passive about their equilibrium values −ωj∗ and
(xM,j,∗ , xc,j,∗ , xu,j,∗ ), in the sense described in Definition 2.
Remark 2: Note that, in Assumption 2, we assume only a
passivity property for the systems (5) without specifying the
precise form of the systems, which permits the inclusion of a
broad class of generation and load dynamics. Moreover, the
fact that passivity is assumed only for the net supply dynamics,
rather than for the generation and load dynamics individually,
can permit the analysis of systems incorporating dynamics that
are not individually passive. We will see in Section VI that
various classes of generation and load dynamics that have been
investigated in the literature fit within the present framework.
It should be noted that for linear systems input strict passivity can easily be checked from strict positive realness of the
transfer function or numerically using the KYP Lemma [17]. It
holds for sign preserving static nonlinearities, first-order linear
dynamics, and also higher-order generation/load dynamics
with sufficiently large damping, as it will be discussed in
Section VI.
D. Optimal supply and load control
We aim to explore how the generated power and controllable loads may be adjusted to meet the step change pL
in frequency-independent load in a way that minimizes the
total cost that comes from the extra power generated and the
disutility of loads. We now introduce an optimization problem,
which we call the optimal supply and load control problem
(OSLC), that can be used to achieve this goal.
c
Suppose that costs Cj (pM
j ) and Cdj (dj ) are incurred for
M
c
deviations pj and dj in generation and controllable load
respectively. Furthermore, some additional cost is incurred due
to any change in frequency which alters the demand from
uncontrollable frequency-dependent loads. We represent this
by an integral cost in terms of a function hj which is explicitly
determined by the dynamics in (5c) as
hj (z) = kduj (−z) for all z̄ ∈ R.
(9)
The total cost within OSLC then sums all the above costs,
and the problem is to choose the vectors pM , dc , and du
that minimize this total cost and simultaneously achieve power
balance, while satisfying physical saturation constraints.
OSLC:
Z
X
X
c
min
Cj (pM
)
+
C
(d
)
+
dj j
j
pM ,dc ,du
j∈G
subject to
X
=
X
(dcj
j∈G
j∈N
M,min
M,max
M
pj
≤ pj ≤ pj
,
c,min
c,max
c
dj
≤ dj ≤ dj
, ∀j
h−1
j (z) dz
0
j∈N
pM
j
du
j
+
duj
+
pL
j ),
(10)
∀j ∈ G,
∈ N,
where pM,min
, pM,max
, dc,min
, and dc,max
are the bounds for
j
j
j
j
generation and controllable demand respectively at bus j. The
equality constraint in (10) represents conservation of power
by specifying that all the extra frequency-independent load
is matched by the total additional generation plus all the
deviations in frequency-dependent loads.
Remark 3: The variables pM and dc within (10) represent
the variables that can be directly controlled (generation and
controllable demand), while the variable du can be controlled
only indirectly by effecting changes in the system frequencies
(uncontrollable frequency-dependent demand). Therefore, we
aim to specify properties of the control dynamics in (5a)–
(5b) that ensure that the quantities pM and dc , along with the
system frequencies, converge to values at which optimality
in (10) can be guaranteed.
E. Additional conditions
To guarantee convergence and optimality, we will require
additional conditions on the behavior of the systems (4)–
(5) and the structure of the optimization problem (10). The
assumptions introduced are all of practical relevance, and
we will see in Section VI that the framework considered
encompasses a number of important examples that have been
investigated within the literature.
The first two of these conditions are needed for our proof
of the convergence result in Theorem 1. Within the second of
these we will denote ω G = [ωj ]j∈G and ω L = [ωj ]j∈L .
Assumption 3: The storage functions in Assumption 2 have
strict local minima at the points (xM,j,∗ , xc,j,∗ , xu,j,∗ ) and
(xc,j,∗ , xu,j,∗ ) respectively.
Remark 4: In practice, Assumption 3 is often trivially
satisfied. For instance, whenever the vector fields in (5) are
5
continuously differentiable, then by considering a linearization
about equilibrium, the KYP Lemma can be seen to generate a
storage function satisfying Assumption 3, when the linearized
system is controllable and observable.
Assumption 4: There exists an open neighbourhood T
of (η ∗ , ω G,∗ , xM,∗ , xc,∗ , xu,∗ ) such that at any time instant t, ω L (t) is uniquely determined by the system states
(η(t), ω G (t), xM (t), xc (t), xu (t)) ∈ T and equations (4)–(5).
Remark 5: Assumption 4 is a technical assumption that is
required in order for the system (4)–(5) to have a locally
well-defined state space realization. This is needed in order
to apply Lasalle’s Theorem to analyze stability in the proof
of Theorem 1 below. Without Assumption 4, stability could
be analyzed through more technical approaches such as the
singular perturbation analysis discussed in [18, Section 6.4].
Remark 6: Assumption 4 can often be verified by using the
Implicit Function Theorem to generate decentralized algebraic
conditions under which it is guaranteed to hold. For instance,
Assumption 4 always holds if in (5) we have for all j ∈ L
∂g u,j
∂g c,j
+
6= 0,
∂ωj
∂ωj
(11)
at the equilibrium point. If the functions g c,j and g u,j have no
explicit dependence on ωj satisfying the following condition
at the equilibrium point is also sufficient
Theorem 1 (Stability): Suppose that Assumptions 1–4 are
all satisfied. Then there exists an open neighbourhood S of
the equilibrium (η ∗ , ω G,∗ , xM,∗ , xc,∗ , xu,∗ ) such that whenever
the initial conditions (η(0), ω G (0), xM (0), xc (0), xu (0)) ∈ S,
then the solutions of the system (4)–(5) converge to an
equilibrium as defined in Definition 1.
Remark 8: It will be seen within the proof of Theorem 1 that
ω, xM , xc , xu converge to ω ∗ , xM,∗ , xc,∗ , xu,∗ respectively.
The phase differences η also converge to a fixed point,
however, this can be different from η ∗ .
Theorem 2 (Optimality): Suppose that Assumption 5 is satisfied. If the control dynamics in (5a) and (5b) are chosen such
pM,max
(−ωj∗ ) = [(Cj0 )−1 (−ωj∗ )]pjM,min and kdcj (−ωj∗ ) =
that kpM
j
dc,max
0 −1
[(Cdj
) (ωj∗ )]djc,min ,
j
then the values pM,∗ , dc,∗ , and du,∗ are
optimal for the OSLC problem (10).
From Theorems 1 and 2, we immediately deduce the
following result guaranteeing convergence to optimality.
Theorem 3 (Convergence to optimality): Consider equilibria
of (4)–(5) with respect to which Assumptions 1–5 are all
satisfied. If the control dynamics in (5a) and (5b) are chosen
such that
pM,max
j
0 −1
kpM
(ū)
=
[(C
)
(ū)]
j
j
pM,min
j
(13)
dc,max
0
−1
kdcj (ū) = [(Cdj ) (−ū)]djc,min
j
j
X ∂g c,j ∂f c,j X ∂g u,j ∂f u,j
i
i
+
> 0.
c,j ∂ω
u,j ∂ω
∂x
∂x
j
j
i
i
i
i
(12)
These conditions are, for example, satisfied by the demand
dynamics considered in Section VII.
In addition, we impose an assumption concerning the form
of the cost functions in the OSLC problem (10).
Assumption 5: The cost functions Cj and Cdj are continuously differentiable and strictly convex. Moreover, the first
derivative of h−1
j (z) is nonnegative for all z ∈ R.
Remark 7: The convexity and continuous differentiability
of the cost functions given by Assumption 5 are sufficient to
allow the use of the KKT conditions to prove the optimality
result in Theorem 2. We demonstrate in Appendix B how the
condition of continuity in the derivatives may be relaxed by
by means of subgradient methods.
We will explicitly state within each of our results which of
these conditions are required.
V. M AIN RESULTS
In this section we state our main results, with their proofs
provided in Appendix A. Our first result shows that the set
of equilibria of the system (4)–(5) for which the assumptions
stated are satisfied is asymptotically attracting, while our second result demonstrates sufficient conditions for equilibrium
points to be optimal for the OSLC problem (10). Based on
these results, we can guarantee convergence to optimality of
all solutions starting in the vicinity of an equilibrium. Finally,
we show that the inclusion of controllable demand in our
model reduces steady state frequency deviation, thereby aiding
in frequency control.
hold for all ū ∈ R, then there exists an open neighbourhood
of initial conditions about any such equilibrium such that the
solutions of (4)–(5) are guaranteed to converge to a global
minimum of the OSLC problem (10).
Remark 9: Theorem 3 states that if the system (4)–(5)
starts sufficiently close to any of its equilibria with respect
to which Assumptions 1 - 5 are satisfied, then the system
is guaranteed to converge to an equilibrium point which will
be optimal for the desired OSLC problem (10). The fact that
pM and dc represent controllable quantities means that we
are free to design the dynamics in (5a) and (5b) in order
that the conditions (13) are satisfied. Thus, knowledge of
the cost functions in the optimization problem we want to
solve explicitly determines classes of dynamics which are
guaranteed to yield convergence to optimal solutions.
Theorem 4 (Reduction in steady state error): Suppose that
Assumptions 1–5 are all satisfied with respect to all equilibria
of (4)–(5). If the control dynamics in (5a) and (5b) are chosen
such that (13) hold for all ū ∈ R, then the addition of
controllable demand in primary control results in a drop in
steady state frequency deviation from its nominal value.
VI. D ISCUSSION
We now discuss various examples of generation and load
dynamics that can fit within our framework.
As a first example, consider the model in [8], which
investigates a linearized version of the system (4) coupled with
0 −1
the static nonlinearities dcj = (Cdj
) (ωj ) for the controllable
demand, and with uncontrollable loads of the form duj =
Dj ωj . The damping constants Dj were assumed positive,
the cost functions Cdj were taken to be strictly convex, and
6
As shown in Corollary 1 below, under (16), the overall
M
u
system relating −ωj with6 sG
j = pj − dj becomes input
strictly passive about any equilibrium point. This follows
from a more general result which we now state describing
the connection between the L2 -gain of general generation
dynamics and the passivity of the supply dynamics. The proofs
of Proposition 1 and Corollary 1 can be found in Appendix A.
ˆ
Fig. 1. Nyquist plot for the transfer function relating p̃ˆM
j with −ω̃j when a
linearization of (14) about equilibrium is considered. A transfer function with
a Nyquist plot within the circle satisfies the gain condition (16). Our approach
allows the Nyquist plot to extend within the entire striped region.
the mechanical power injection pM was also assumed to be
constant after a step change. It is easy to see that for such a
system Assumptions 1–5 are all satisfied. Hence, this model
can be analyzed in the framework introduced above, thus
implying optimality and stability of the equilibrium points.
The present framework can also include systems in which
the generated powers satisfy any first-order dynamics, since
such schemes are passive about their equilibria for arbitrary gains. For higher-order schemes, however, the dynamics
for pM are not necessarily passive, so some additional conditions are needed to ensure stability. Second-order generation
dynamics are often considered in the literature to model
turbine-governor dynamics (e.g. [5]). These can be described
by the system
α̇j = −
ṗM
j
1
αj +
1
pcj ,
τg,j
τg,j
1
1 M
p +
αj ,
=−
τb,j j
τb,j
j∈G
(14)
where αj is the valve position of the turbine, the constants τg,j
and τb,j represent lags in the dynamics of the governor and
turbine respectively, and pcj is a static function of frequency,
corresponding to droop control. Consider the case where
generator damping and uncontrollable frequency-dependent
loads are modeled by
duj = Dj ωj , ∀j ∈ N,
(15)
and there is no controllable demand. In [9], the gain condition
|pcj (ωj ) − pcj (ωj∗ )| ≤ Kj |ωj − ωj∗ |, j ∈ G
(16)
with Kj < Dj was imposed to ensure stability of the network
when the generations dynamics are given by (14)–(15).
Proposition 1: Let equation (15) hold and consider any
generation dynamics from −ωj to pM
j of the form (5a). Given
M,∗
)
any equilibrium, if the L2 -gain from (ωj −ωj∗ ) to (pM
j −pj
is strictly less than Dj , then the system with input −ωj and
M
u
output sG
j = pj − dj is input strictly passive about the
equilibrium considered.
It easy to show that, for the dynamics (14)–(15), condition (16) is sufficient to satisfy the L2 -gain condition in
Proposition 1. Applying Proposition 1 on the system (14)–
(16) results in the following corollary.
Corollary 1: Let equation (15) hold and consider the generation dynamics in (14). Then, for any equilibrium where (16)
M
u
holds, the system with input −ωj and output sG
j = p j − dj
is input strictly passive about this equilibrium.
Our framework can allow us to include also more general
dynamics and to deduce asymptotic stability under weaker
conditions. To see this, we consider a linearization of the
system (14) about equilibrium and let q̃ denote the deviation
of any quantity q from its equilibrium value q ∗ . Expressing
1
ˆc
ˆM
p̃M
j in the Laplace domain gives p̃j = (τg,j s+1)(τb,j s+1) p̃j .
Therefore,
1
ˆj )
p̃ˆc + Dj (−ω̃
(τg,j s + 1)(τb,j s + 1) j
ˆ j ], j ∈ G,
=: Hj (s)[−ω̃
(17)
ˆ˜u
ˆM
s̃ˆG
j = p̃j − dj =
ˆ j and
where Hj (s) denotes the transfer function relating −ω̃
G
ˆ
s̃j . Since the maximum gain of the transfer function from
p̃cj to p̃M
is 1 at s = 0, the condition in (16) constrains
j
the Nyquist diagram of Hj to lie inside a ball with centre
(Dj , 0) and radius Kj < Dj . This is contained strictly within
the right half-plane, implying the required passivity condition
in Assumption 2. For instance, the Nyquist plot from input
ˆ j to output p̃ˆM
−ω̃
can be as shown by the solid line in
j
Fig. 1. However, according to our analysis any dynamics for
the command signal can be permitted provided that the supply
dynamics in (17) remain input-strictly passive. This can permit
any frequency response within the striped region in Fig. 1,
for example allowing the larger Nyquist locus shown with a
dashed line. In fact, under the reasonable assumption that pcj
has the same sign as −ωj (i.e. negative feedback is used), it
can easily be verified that the transfer function from p̃cj to p̃M
j
6 Note that this example could also include passive controllable demand
dcj (−ωj ), since showing input strict passivity about equilibrium of the system
M
u
with input −ωj and output sG
j = pj − dj is sufficient to ensure also that
M
u
c
the system with the same input and output sG
j = pj − dj − dj is input
strictly passive.
7
Maximum allowable value of ratio K/D vs. value of α
Generation system dynamics with a deadband
60
pmax
c
Deviation in power command
Maximum value of ratio K/D
50
40
30
20
0
10
max
−pc
0
0
0.5
1
1.5
Ratio of time constants, α
2
Fig. 2. Maximum value of the ratio Kj /Dj under which the passivity
property is maintained for the linearized dynamics of system (14)–(16) against
the ratio of time constants a = τb,j /τg,j . The figure demonstrates that
the maximum value for Kj that ensures passivity is allowed to be much
higher than Dj .
given by Tj (s) =
1
(τg,j s+1)(τb,j s+1)
has a minimum real value
−τg,j τb,j
√
(τg,j + τb,j )2 + 2(τg,j + τb,j ) τg,j τb,j
(18)
q
√
(τg,j +τb,j )+ τb,j τg,j
.
Thus,
the
at frequency ωj,M AX =
(τb,j τg,j )3/2
required passivity property will be maintained provided Kj
multiplied by the quantity in (18) is strictly greater than −Dj .
The maximum values of Kj for which this is satisfied are
τb,j
. We
shown in Fig. 2, where a is defined as the ratio τg,j
see that the maximum allowable value for Kj is always at
least 8Dj (obtained at a = 1) and tends to infinity as a → 0
(which corresponds to a first order system). This shows that the
stability guarantees can be preserved under significantly larger
gains Kj than those indicated in Corollary 1. Therefore, our
approach allows for a less conservative stability condition for
equilibrium points where a linearization is feasible, while also
allowing to consider a wider class of generation dynamics. It
is worth noting, however, that the use of the more conservative
L2 -gain condition described in Proposition 1 would generally
be expected to yield better robustness properties. Such tradeoffs between feedback gain and stability margin need to be
taken into account in the design of control systems.
Remark 10: The assumption of small power angles differences η is widely used in the literature. Under such an assumption the stability results stated in the paper hold globally
for the other system states if assumptions 2-4 hold globally7 .
It should be noted though that in general the system (4)–(5)
is not globally asymptotically stable due to the fact that the
sinusoids in (4d) are only locally passive. Finding the region
of attraction is an interesting problem (associated with what is
know as transient stability) which is not explicitly addressed
within this paper.
< (T j (jωj,M AX )) =
7 In particular, system (4)–(5) with the sinusoids in (4d) linearized about
about η ∗ (as in e.g. [9]), would be globally asymptotically stable if in
assumption 2 the systems are globally passive about the equilibrium point
considered, and in assumption 3 the storage functions are radially unbounded
with a minimum at the equilibrium point. For example, cases where the input
strict passivity property is satisfied globally include the class of dynamics
considered in Proposition 1 (and Corollary 1), or the dynamics in (19)–(20).
−ω1
0
−ω0
ω0
Frequency Deviation
ω1
Fig. 3. Power command signal against frequency deviations for a system with
deadband and saturation bounds.
Another class of droop control schemes used in practice
incorporates a deadband which prevents unneeded adjustments
for small variations in frequency about its nominal value. An
example of this is shown in Fig. 3. The response profile of a
static controllable load with deadband and saturation bounds is
analogous. For these systems, a minimum frequency deviation
ωj0 is required to trigger a frequency-dependent deviation. The
system then reaches its physical limits at a higher frequency
deviation ωj1 .
While the stability of these systems can be shown by
Theorem 1, provided8 Assumptions 1–4 hold, Theorem 2
cannot be applied, since the deadband makes the cost functions fail to satisfy continuously differentiability, therefore
not meeting the conditions of Assumption 5. Theorem 5 in
Appendix B shows how Theorem 2 can be generalized to
include non-continuously differentiable cost functions, thereby
permitting the application of our results to analyze systems
with such dynamics.
VII. S IMULATION ON THE IEEE 68 BUS SYSTEM
In this section we illustrate our stability and optimality
results through applications on the IEEE New York / New
England 68 bus interconnection system [20], simulated using
the Power System Toolbox [21]. This model is more detailed
and realistic than our analytical model, including line resistances, a DC12 exciter model, power system stabilizer (PSS),
and a subtransient reactance generator model. A similar model
with no PSS was used for comparison9 .
The test system contains 52 load buses serving different
types of loads including constant active and reactive loads.
The overall system has a total real power of 16.41GW. For
our simulation, we have added three loads on units 2, 9, and
17, each having a step increase of magnitude 1 p.u. (base
100MVA) at t = 1 second. We allow controllable demand
on buses 1-30. The disutility function for the aggregate load
at each bus is dcj is Cdj (dcj ) = 21 αj dcj 2 , where αj are cost
8 Although the nonlinearity depicted in Fig. 3 is itself not input strictly
passive about its equilibria, the presence of any frequency-dependent damping
of the form (15) will be sufficient to ensure that Assumption 2 holds.
9 The details of the simulation models with or without PSS can be found
in the Power System Toolbox data files data16m and data16em respectively.
8
Frequency at bus 63
Allocation of demand in controllable loads
0
60
−0.01
59.9
−0.02
59.8
59.7
No OSLC − No PSS
Static OSLC − No PSS
Dyn. OSLC − No PSS
No OSLC − PSS
Static OSLC − PSS
Dyn. OSLC − PSS
59.6
59.5
59.4
59.3
0
5
10
15
20
25
30
Allocation (p.u.)
Frequency (Hz)
60.1
Static OSLC − α = 10
Dynamic OSLC − α = 10
Static OSLC − α = 5
Dynamic OSLC − α = 5
−0.03
−0.04
−0.05
−0.06
−0.07
−0.08
0
35
5
10
Time (s)
15
20
25
30
35
Time (s)
coefficients. The cost coefficients were selected such that for
a step change in demand, the power allocated between total
generation and controllable demand would be roughly equal,
as suggested in [19]. The selected values were αj = 5 for
load buses 1-10 and αj = 10 for the rest. The loads were
controlled every 10ms.
Consider the static and dynamic control schemes given by10
0 −1
dcj = (Cdj
) (ωj ), j ∈ N,
(19)
d˙cj = −(Cj0 (dcj ) − ωj ), j ∈ N.
(20)
Fig. 5. Power allocation among controllable loads with non-equal cost
coefficients for 2 cases: (i) Static OSLC (ii) Dynamic OSLC.
Marginal Cost of OSLC
0.4
0.35
0.3
Marginal Cost
Fig. 4. Frequency at bus 63 in six cases: (i) no OSLC, no PSS, (ii) Static
OSLC, no PSS, (iii) Dynamic OSLC, no PSS, (iv) no OSLC, with PSS, (v)
Static OSLC, with PSS, (vi) Dynamic OSLC, with PSS.
Static OSLC
Dynamic OSLC
0.25
0.2
0.15
0.1
We refer to the resulting dynamics as Static OSLC and Dynamic OSLC respectively. We investigate the behavior in the
following six cases: (i) no OSLC, no PSS, (ii) Static OSLC,
no PSS, (iii) Dynamic OSLC, no PSS, (iv) no OSLC, with
PSS, (v) Static OSLC, with PSS, (vi) Dynamic OSLC, with
PSS. The frequency dynamics for bus 63 are shown in Fig. 4.
From Fig. 4, we observe that whether or not PSS is used, the
presence of OSLC results in a drop in steady state frequency
deviations, as expected due to Theorem 4. Furthermore, we
can see that the overshoot is significantly less when OSLC is
used. The responses for Static and Dynamic OSLC have no
significant differences and converge to the same exact value
at steady state. However, Dynamic OSLC appears to give a
larger overshoot than Static OSLC. In all cases, the voltage
deviation was less than 0.01 p.u., showing that the constant
voltage assumption is reasonable.
In Fig. 5 we also observe a higher power allocation at the
load buses whose cost coefficients take the lower value αj = 5
than at those with αj = 10. This demonstrates that the power
allocation among controllable loads depends upon the loads’
respective cost coefficients of demand response. This behavior
could be beneficial if a prescribed allocation were desirable,
as then the load dynamics could be designed such that the cost
coefficients chosen yield the desired allocation. Furthermore,
as shown in Fig. 6, the marginal costs at each controlled load
converge to the same value. This illustrates the optimality in
the power allocation among the loads, as equality of marginal
10 Note that both (19) and (20) are input strictly passive about the equilibria,
and both satisfy Assumption 4 (using the conditions in equations (11) and (12)
respectively).
0.05
0
0
5
10
15
20
25
30
35
Time (s)
0
Fig. 6. Marginal costs Cdj
of controllable loads with non-equal cost
coefficients for 2 cases: (i) Static OSLC (ii) Dynamic OSLC.
costs is the optimality condition for (10) when the allocations
do not saturate.
Finally, to investigate the system’s robustness, we then
introduced delays to account for the time between the arrival
of the frequency signal and the response of the controllable
demand. The simulation was repeated with 0.1 pu loads and a
delay of 0.05 seconds. Furthermore, all cost coefficients were
set to αj = 1. Dynamic OSLC was seen to offer improved
robustness to the time-delay relative to Static OSLC, since the
first converged both with and without PSS whereas the latter
became unstable in both cases. This illustrates how appropriate higher order dynamics can have improved robustness
properties. The simulation results for Dynamic OSLC are
depicted in Fig. 7. This enhanced robustness to delays can
be explained with the help of Fig. 8. The figure shows the
Nyquist plots of the transfer functions relating the increments
from equilibrium of dcj and −ωj for the systems (19) and (20)
respectively linearized about equilibrium, with an input delay
included and also both multiplied by a positive gain K > Dj ,
where Dj is the damping coefficient at bus j, as in (15). The
delayed Dynamic OSLC (dashed line) maintains the passivity
property of the bus dynamics (since the Nyquist plot remains
to right side of −Dj , as was previously discussed in section
VI). On the other hand, the delayed Static OSLC (solid line)
does not, explaining why the latter might be expected to
9
potential extensions to our analysis include incorporating
voltage dynamics as well as more advanced models for the
loads where their switching behavior is taken into account.
Frequency at bus 63
60
Frequency (Hz)
Dynamic OSLC − No PSS
Dynamic OSLC − PSS
59.95
59.9
59.85
0
5
10
15
20
Time(s)
25
30
35
Fig. 7. Frequency at bus 63 for Dynamic OSLC with time-delays in two
cases: (i) no PSS (ii) with PSS.
A PPENDIX A
P ROOFS OF OUR MAIN RESULTS
In this appendix we prove our main results, Theorems 1–4,
and we also provide proofs of Proposition 1 and Corollary 1.
Proof of Theorem 1: We will use the dynamics in (4)
together with the passivity conditions in Assumption 2 to
define a Lyapunov function for the system (4)–(5), similarly
to [9], [13].
P
Firstly, we introduce VF (ω G ) = 12 j∈G Mj (ωj − ωj∗ )2 .
The time-derivative along the trajectories of (4)–(5) is then
!
X
X
X
∗
L
G
V̇F =
(ωj − ωj ) − pj + sj −
pjk +
pij
j∈G
i:i→j
k:j→k
!
+
X
(ωj − ωj∗ )
X
L
− pL
j + sj −
j∈L
X
pjk +
k:j→k
pij ,
i:i→j
by substituting (4b) for ω̇j for j ∈ G and adding the final
term, which is equal to zero by (4c). Subtracting the product
of (ωj − ωj∗ ) with each term in (6b) and (6c), this becomes
X
X
G,∗
L,∗
V̇F =
(ωj − ωj∗ )(sG
(ωj − ωj∗ )(sL
j − sj ) +
j − sj )
j∈G
j∈L
+
X
(pij − p∗ij )(ωj − ωi ),
(21)
(i,j)∈E
Fig. 8. Nyquist plot of the transfer functions relating the deviations from
equilibrium of dc and −ω, for the systems (19) and (20) linearized about
equilibrium, multiplied by a constant K > Dj and with an input delay also
included. The figure shows that only the Dynamic delayed system (dashed
line) remains on the right of −Dj , maintaining the passivity property of the
bus dynamics when the damping coefficient is Dj as in (15). The Static
system (solid line) extends to the left of −Dj , hence the passivity property
is lost.
become unstable.
VIII. C ONCLUSION
Within the paper, we have considered the problem of
designing distributed generation and demand control schemes
for primary frequency regulation in power networks such that
asymptotic stability is guaranteed while ensuring optimality
of power allocation. We have presented a network passivity
framework which provides a systematic method to show
stability over a broad class of generation and load dynamics.
Furthermore, we have derived steady state conditions for the
generation and controllable demand control schemes that
ensure that the power generated/consumed is the solution to
an appropriately constructed network optimization problem,
thus allowing fairness in power allocation to be guaranteed.
In addition, under some minor assumptions, we have shown
that the inclusion within the model of controllable demand
has a positive effect also on secondary control, decreasing
the steady state deviation in frequency. A simulation on
the IEEE 68 bus system verifies our results. Interesting
using in the final term the equilibrium
P condition R(6a).
η
Additionally, consider VP (η) = (i,j)∈E Bij η∗ij (sin φ −
ij
∗
sin ηij
) dφ. Using (4a) and (4d), the time-derivative equals
X
∗
V̇P =
Bij (sin ηij − sin ηij
)(ωi − ωj )
(i,j)∈E
=
X
(pij − p∗ij )(ωi − ωj ).
(22)
(i,j)∈E
Furthermore, from the passivity conditions in Assumption 2
and the definition of input strict passivity from Section III, it
follows that the following hold:
1) For each j ∈ G, there exist open neighbourhoods UjG
of ωj∗ and XjG of (xM,j,∗ , xc,j,∗ , xu,j,∗ ) and a continuously differentiable, positive semidefinite function11
VjG (xM,j , xc,j , xu,j ) such that
G,∗
V̇jG ≤ ((−ωj ) − (−ωj∗ ))(sG
j − sj )
∗
− φG
j ((−ωj ) − (−ωj ))
(23)
for all ωj ∈ UjG and all (xM,j , xc,j , xu,j ) ∈ XjG ,
2) For each j ∈ L, there exist open neighbourhoods UjL
of ωj∗ and XjL of (xc,j,∗ , xu,j,∗ ) and a continuously differentiable, positive semidefinite function VjL (xc,j , xu,j )
such that
L,∗
V̇jL ≤ ((−ωj ) − (−ωj∗ ))(sL
j − sj )
∗
− φL
j ((−ωj ) − (−ωj )) (24)
11 The
time-derivates of VjG and VjL are evaluated using the dynamics (5).
10
for all ωj ∈ UjL and all (xc,j , xu,j ) ∈ XjL ,
L
where φG
j and φj are positive definite functions.
Based on the above, we define the function
V (η, ω G , xM , xc , xu ) = VF (ω G ) + VP (η)
X
X
+
VjG (xM,j , xc,j , xu,j ) +
VjL (xc,j , xu,j ),
j∈G
j∈L
which we aim to use in Lasalle’s theorem. By (21) and (22),
i
Xh
G,∗
G
(ωj − ωj∗ )(sG
V̇ =
j − sj ) + V̇j
j∈G
+
i
Xh
L,∗
L
.
(ωj − ωj∗ )(sL
j − sj ) + V̇j
j∈L
Using (23) and (24) it therefore holds that
V̇ ≤ −
X
j∈G
≤0
∗
φG
j ((−ωj ) − (−ωj )) −
X
∗
φL
j ((−ωj ) − (−ωj ))
j∈L
(25)
whenever ωj ∈ UjG and (xM,j , xc,j , xu,j ) ∈ XjG for all j ∈ G
and ωj ∈ UjL and (xc,j , xu,j ) ∈ XjL for all j ∈ L.
Clearly VF has a strict global minimum at ω G,∗ , and
G
Vj , VjL have strict local minima at (xM,j,∗ , xc,j,∗ , xu,j,∗ ),
(xc,j,∗ , xu,j,∗ ) by Assumption 3. Additionally, note that Assumption 1 guarantees the existence of some neighbourhood
∗
of each ηij
on which the respective integrand in the definition
of VP is increasing. Since the integrand is zero at the lower
∗
limit of the integration, ηij
, this immediately implies that
VP has a strict local minimum at η ∗ . Thus, V has a strict
local minimum at the point Q∗ := (η ∗ , ω G,∗ , xM,∗ , xc,∗ , xu,∗ ).
We now recall Assumption 4. This tells us that, provided
(η, ω G , xM , xc , xu ) ∈ T , ω L can be uniquely determined
from these quantities. Therefore, the states of the differential
equation system (4)–(5) within the region T can be expressed
as (η, ω G , xM , xc , xu ). Let us now choose a neighbourhood
in the coordinates (η, ω G , xM , xc , xu ) about Q∗ on which the
following all hold:
1) Q∗ is a strict minimum of V ,
2) (η, ω G , xM , xc , xu ) ∈ T ,
3) ωj ∈ UjG and (xM,j , xc,j , xu,j ) ∈ XjG for all j ∈ G and
ωj ∈ UjL and (xc,j , xu,j ) ∈ XjL for all j ∈ L 12 ,
4) xM,j , xc,j , and xu,j all lie within their respective
neighbourhoods X0 as defined in Section III.
Recalling now (25), we thus see that within this neighbourhood, V is a nonincreasing function of all the system states and
has a strict local minimum at Q∗ . Consequently, the connected
component of the level set {(η, ω G , xM , xc , xu ) : V ≤ }
containing Q∗ is guaranteed to be both compact and positively
invariant with respect to the system (4)–(5) for all sufficiently
small > 0. Therefore, there exists a compact positively
invariant set Ξ for (4)–(5) containing Q∗ .
Lasalle’s Invariance Principle can now be applied with the
function V on the compact positively invariant set Ξ. This
is possible because the requirement ωj ∈ UjL for all j ∈ L
corresponds, by Assumption 4 and the continuity of the equations in (4)–(5), to
requiring the states (η, ω G , xM , xc , xu ) to lie in some open neighbourhood
about Q∗ .
12 This
guarantees that all solutions of (4)–(5) with initial conditions (η(0), ω G (0), xM (0), xc (0), xu (0)) ∈ Ξ converge to the
largest invariant set within Ξ ∩ {(η, ω G , xM , xc , xu ) : V̇ = 0}.
We now consider this invariant set. If V̇ = 0 holds at a point
within Ξ, then (25) holds with equality, whence by Assumption 2 we must have ω = ω ∗ . Moreover, on any invariant set
on which ω = ω ∗ , the system equations (4) apply and give
precisely the equilibrium conditions (6a), (6b), (6c), and (6g).
Furthermore, if V̇ = 0, it follows from (21), (22), (23),
and (24) that all V̇jG = 0 and V̇jL = 0. But ω = ω ∗ implies
by the definitions in Section III that (xM , xc , xu ) converge
to the point (xM,∗ , xc,∗ , xu,∗ ), at which Assumption 3 states
that VjG and VjL take strict local minima. Therefore the
values of VjG and VjL must decrease along all nontrivial
trajectories within the invariant set, contradicting V̇jG = 0
and V̇jL = 0. Consequently, at all points of any invariant
set within Ξ ∩ {(η, ω G , xM , xc , xu ) : V̇ = 0}, we must also
have (xM , xc , xu ) = (xM,∗ , xc,∗ , xu,∗ ). Thus, the remaining
equilibrium conditions (6d), (6e), (6f), (6h), (6i), and (6j) are
also satisfied. Therefore, we conclude by Lasalle’s Invariance
Principle that all solutions of (4)–(5) with initial conditions
(η(0), ω G (0), xM (0), xc (0), xu (0)) ∈ Ξ converge to the set of
equilibrium points as defined in Definition 1. Finally, choosing
for S any open neighbourhood of Q∗ within Ξ completes the
proof.
0
Proof of Theorem 2: Due to Assumption 5, Cj0 and Cdj
are
strictly increasing and hence invertible. Therefore
pM,max
kpM
(−ωj∗ ) = [(Cj0 )−1 (−ωj∗ )]pjM,min
j
j
kdcj (−ωj∗ )
=
dc,max
0 −1
[(Cdj
) (ωj∗ )]djc,min
(26)
j
are well-defined. Furthermore, Assumption 5 also ensures that
the OSLC problem (10) is a convex optimization problem
with a continuously differentiable cost function. Thus, a point
(p̄M , d¯c , d¯u ) is a global minimum for (10) if and only if it
satisfies the KKT conditions [22]
+
−
Cj0 (p̄M
j ) = −ν − λj + λj , j ∈ G,
(27a)
−
0 ¯c
Cdj
(dj ) = ν − µ+
j + µj , j ∈ N,
(27b)
¯u
h−1
j (dj ) = ν, j ∈ N,
X
X
p̄M
(d¯cj + d¯uj + pL
j =
j ),
(27c)
j∈G
(27d)
j∈N
M,max
pM,min
≤ p̄M
, j ∈ G,
j ≤ pj
j
(27e)
dc,min
≤ d¯cj ≤ dc,max
, j ∈ N,
j
j
(27f)
M,max
M,min
M
M
λ+
) = 0, λ−
) = 0, j ∈ G,
j (pj − pj
j (pj − pj
(27g)
c,min
− c
µ+ (dcj − dc,max
)
=
0,
µ
(d
−
d
)
=
0,
j
∈
G,
(27h)
j
j
j
−
+
−
for some constants ν ∈ R and λ+
j , λj , µj , µj ≥ 0. We
will now show that these conditions are satisfied by the
equilibrium values (p̄M , d¯c , d¯u ) = (pM,∗ , dc,∗ , du,∗ ) defined
by equations (6h), (6i), and (6j).
11
0
are both strictly increasBy Assumption 5, Cj0 and Cdj
ing, so we can uniquely define13 frequencies ωjM,max :=
−Cj0 (pM,max
), ωjM,min := −Cj0 (pM,min
), ωjc,max :=
j
j
0
0
). Letting ω0∗ denote
(dc,min
), and ωjc,min := Cdj
Cdj
(dc,max
j
j
∗
the common value of all ωj due to (6a), we can, in terms of
these quantities, define the nonnegative constants
M,max
− ω0∗ ) 1{q : q≤ωM,max } (ω0∗ ),
λ+
j := (ωj
j
M,min
∗
λ−
) 1{q : q≥ωM,min } (ω0∗ ),
j := (ω0 − ωj
j
(28)
c,max
∗
µ+
) 1{q : q≥ωjc,max } (ω0∗ ),
j := (ω0 − ωj
c,min
µ−
− ω0∗ ) 1{q : q≤ωc,min } (ω0∗ ).
j := (ωj
j
Then, since (Cj0 )−1 (−ω0∗ ) ≥ pM,max
⇔ ω0∗ ≤ ωjM,max ,
j
M,min
M,min
0 −1
(Cj0 )−1 (−ω0∗ ) ≤ pj
⇔ ω0∗ ≥ ωj
, (Cdj
) (ω0∗ ) ≥
c,max
c,max
c,min
∗
0 −1
∗
dj
⇔ ω0 ≥ ωj
, and (Cdj ) (ω0 ) ≤ dj
⇔ ω0∗ ≤
c,min
, it follows by (6h), (6i), and (26) that the complemenωj
tary slackness conditions (27g) and (27h) are satisfied with the
choices (28).
Let us also define ν = ω0∗ . Then
M,max
−
0 −1
∗ −ωj
(Cj0 )−1 (−ν − λ+
+
λ
)
=
(C
)
[−ω
]
j
0 −ω M,min
j
j
j
=
pM,max
[(Cj0 )−1 (−ω0∗ )]pjM,min
j
= pM,∗
,
j
by the above definitions and equations (6h) and (26). Therefore, the optimality condition (27a) holds. Analogously,
ω c,max
0 −1
0 −1
(Cdj
) (ν − µ+ + µ− ) = (Cdj
)
[ω0∗ ]ωjc,min
j
=
dc,max
0 −1
[(Cdj
) (ω0∗ )]djc,min
j
= dc,∗
j ,
by (6i) and (26), satisfying (27b). Additionally, (27c) holds
because hj (ν) = duj follows from (6j) and (9).
Furthermore, summing the equilibrium conditions (6b) over
all j ∈ G and (6c) over all j ∈ L shows that (27d) is satisfied.
Finally, the saturation constrains in (26) ensure that (27e)
and (27f) are also satisfied.
Thus, we see that the values (p̄M , d¯c , d¯u )
=
M,∗ c,∗ u,∗
(p , d , d ) satisfy the KKT conditions (27) with
ν = ω0∗ . Therefore, the equilibrium values pM,∗ , dc,∗ ,
and du,∗ define a global minimum for the problem (10).
Proof of Theorem 3: If Assumptions 1–5 all hold and (13)
is true, then all of the assumptions in both Theorems 1 and 2
are satisfied, and thus the result follows.
Proof of Theorem 4: Recalling the proof of Theorem 2, we
know from (27d) and the equalities (26) that at any equilibrium
of (4)–(5) the power balance equation
X
X
X
0 −1
(Cj0 )−1 (ω0∗ ) +
((Cdj
) (ω0∗ ) + hj (ω0∗ )) = −
pL
j
j∈G
j∈N
j∈N
(29)
13 If any of these quantities do not exist, then we can just take
ωjM,max , ωjc,min := −∞ or ωjM,min , ωjc,max := ∞ as appropriate, and
the following arguments still apply.
is satisfied, where ω0∗ denotes the common steady state value
of frequency due to (6a). Now note that, because the nominal
frequency defines an equilibrium frequency prior to the step
change in load and all quantities in (4) denote deviations
from their respective values at this nominal equilibrium, the
equalities (9) and (13) imply that each term on the lefthand side in (29) must take the value zero at ω0∗ = 0.
Furthermore, Assumption 5 implies that the terms (Cj0 )−1 (ω0∗ )
0 −1
and (Cdj
) (ω0∗ ) are all strictly increasing in ω0∗ , while each
term hj (ω0∗ ) is nondecreasing in ω0∗ . Thus both the added term
due to load control and the entire left-hand side in (29) have
the same sign as ω0∗ and are strictly increasing in ω0∗ . It follows
that the presence of this load control term results in a decrease
in the value of ω0∗ , the steady state frequency deviation from
its nominal value.
Proof of Proposition 1: The L2 -gain condition implies14
s
s
Z t1
Z t1
2 dt ≤ K
)
(30)
(p̃M
ω̃j2 (t) dt.
j
j
0
0
where Kj < Dj and t1 is any positive constant. Then, input
strict passivity can be shown as follows.
s
Z t1
2
Z t1
M
M
p̃j (t)ω̃j (t) dt ≤
|p̃j (t)||ω̃j (t)| dt
(31a)
0
0
s
Z t1
Z t1
M
2
≤
|p̃j (t)| dt
|ω̃j (t)|2 dt (31b)
Z
≤
0
0
t1
Kj ω̃j (t)2 dt <
0
Z
t1
Dj ω̃j (t)2 dt,
0
(31c)
where inequality (31b) follows from the Cauchy-Schwarz
inequality and (31c) from inequality (30) and Kj < Dj ,
j ∈ G. Using (31), it is straightforward to show that
Z t1
Z t1
Dj ω̃j (t)2 dt −
p̃M
j (t)ω̃j (t)dt
0
0
Z t1
Z t1
G
=
s̃j (t)(−ω̃j (t))dt ≥ (Dj − Kj )
ω̃j (t)2 dt> 0
0
0
(32)
holds for all j ∈ G. Inequality (32) implies input strict
M
˜u
passivity of the system with output s̃G
j = p̃j − dj and
input −ω̃j , about the equilibrium point considered, since
(32) implies from [24, Lemma 1] the existence of a positive
definite storage function V satisfying the input strict passivity
condition in Definition 2.
Proof of Corollary 1: Let T j (s) be the transfer function
relating p̃ˆcj (s) and p̃ˆM
j (s) in (14), given by
T j (s) =
1
, j ∈ G.
(τg,j s + 1)(τb,j s + 1)
It is easy to show that kT j (s)k∞ = 1, hence the system from
p̃cj to p̃M
j has L2 -gain less than or equal to 1 (e.g. [25, p.18]).
Using also equation (16) we thus have
c
kp̃M
j k2 ≤ kp̃j k2 < Dj kω̃j k2 .
(33)
14 As in the main text, for a variable x that depends on time we use the
notation x̃ to denote deviations from equilibrium, i.e. x̃ := x − x∗ , where
x∗ is the value of x at the equilibrium point mentioned in the proposition.
12
It follows from (33) that the L2 -gain condition in Proposition 1 holds, therefore Proposition 1 can be used to deduce
input strict passivity of the system.
Remark 11: It can be verified that a possible choice of storage function for the input strictly passive system considered
in Corollary 1 is
X 1
1
M,∗ 2
∗ 2
V (pM , α) =
βj (pM
)
γ
(α
−
α
−
p
+
)
j
j
j
j
j
2
2
j∈G
0 −1
are therefore equal to (Cj0 )−1 and (Cdj
)
respectively in
0
0
the regime where Cj and Cdj are continuous, and otherwise
remain constant. They are thus continuous non decreasing
0 −1
functions analogous to (Cj0 )−1 and (Cdj
)
and are well
defined at all points.
Theorem 5: Suppose that Assumption 6 is satisfied. If the
control dynamics in (5a) and (5b) are chosen such that
pM,max
(−ωj∗ ) = [Dj (−ωj∗ )]pjM,min ,
k pM
j
j
τ
2(D −K )
j
j
b,j
and γj <
τg,j .
with coefficients 0 < βj < γj τg,j
Kj2
Note that this has a strict minimum at the equilibrium point
and hence also satisfies Assumption 3.
A PPENDIX B
E XTENSION TO NON -C 1 COST FUNCTIONS
In this appendix we show how our main results can be
extended to apply also to systems in which the cost functions
within the OSLC problem (10) may not be continuously differentiable. We demonstrate this by formulating a relaxed version
of Assumption 5, which we use to to prove a generalization
of our optimality result Theorem 2. Since Theorem 1 is
independent of Assumption 5 and so remains true also in this
extended case, a corresponding generalization of Theorem 3
will then follow immediately.
We weaken Assumption 5 to the following condition.
Assumption 6: The cost functions Cj and Cdj are continuous and strictly convex. Furthermore, they are continuously
differentiable except on respective sets Λj and Λdj of isolated
points. Moreover, the first derivative of h−1
j (z) is nonnegative
for all z ∈ R.
Assumption 6 therefore relaxes the conditions imposed in
the main paper by allowing the cost functions to be nondifferentiable at a discrete set of points. This permits classes of
hybrid cost functions and more involved control dynamics. To
overcome the issue of nondifferentiability within the proof of
the optimality result that follows, we will consider the notions
of subgradient and the subdifferential (e.g. [26]), defined as
follows.
Definition 3: Given a convex function f : I → R, a
subgradient of f at a point x ∈ I ⊆ R is any v ∈ R such
that f (y) − f (x) ≥ v(y − x) for all y ∈ I. The set of all
subgradients of f at x is called the subdifferential of f at x
and is denoted by ∂f (x).
As in the main text, the analysis needs to make use of the
0
inverse of Cj0 and Cdj
. Since, however, these derivatives are
allowed to be discontinuous, their inverses are not well defined
at points of discontinuity, and we hence need to define the
following functions which can be seen as generalized inverses
0
of Cj0 and Cdj
respectively. In particular, we define
(
(Cj0 )−1 (x), x ∈ Cj0 (R \ Λj ),
Dj (x) :=
γ,
x ∈ [Cj0 (γ−), Cj0 (γ+)], γ ∈ Λj ,
(
(34)
0 −1
0
(Cdj
) (x), x ∈ Cdj
(R \ Λdj ),
Ddj (x) :=
0
0
γ,
x ∈ [Cdj
(γ−), Cdj
(γ+)], γ ∈ Λdj ,
where the quantities γ± respectively denote the limits
lim↓0 (γ ± ). The functions Dj and Ddj defined in (34)
kdcj (−ωj∗ )
M,∗
=
c,∗
dc,max
[Ddj (ωj∗ )]djc,min ,
(35)
j
u,∗
then the values p , d , and d are optimal for the OSLC
problem (10).
Proof of Lemma 5: The optimality proof for non-zero values
of generation or controllable demand is similar to that of
Theorem 2. We will make use of subgradient techniques [26,
Section 23] to show that the discontinuity in the cost function’s
derivative does not affect the optimality result.
Firstly we note that strict convexity implies both that Cj0 and
0
Cdj are strictly increasing on R \ Λj and R \ Λdj respectively,
and that their jumps on Λj and Λdj are nonnegative. Therefore,
0
Cj0 and Cdj
are invertible on R \ Λj and R \ Λdj , and
0
0
0
Cj (γ−) < Cj0 (γ+) for all γ ∈ Λj and Cdj
(γ−) < Cdj
(γ+)
for all γ ∈ Λdj . Moreover, continuous differentiability on the
sets R \ Λj and R \ Λdj ensures that the relevant limits γ−
and γ+ here all exist. Consequently, the functions in (34) and
hence the controls in (35) are well-defined.
Now, since OSLC (10) is a convex optimization problem, a
point (p̄M , d¯c , d¯u ) is globally optimal if and only if it satisfies
the subgradient KKT conditions. As only Cj and Cdj are
non-continuously differentiable, these are identical to those
used in the proof of Theorem 2, with the exception that (27a)
and (27b) are replaced by the subdifferential optimality conditions
−
M
− ν − λ+
j + λj ∈ ∂Cj (p̄j ), j ∈ G,
(36a)
−
¯c
ν − µ+
j + µj ∈ ∂Cdj (dj ), j ∈ N.
(36b)
We will thus seek to show that the arguments for demonstrating (27c)–(27h) easily carry over to the present setting,
and then prove that (36a) and (36b) are also satisfied by the
equilibrium quantities, thereby guaranteeing optimality.
Due to the discontinuities in the cost function derivatives,
we slightly redefine the quantities introduced in the proof of
Theorem 2 as
ωjM,max := sup{ωj : Dj (−ωj ) ≥ pM,max
},
j
ωjM,min := inf{ωj : Dj (−ωj ) ≤ pM,min
},
j
ωjc,max := inf{ωj : Ddj (ωj ) ≥ dc,max
},
j
(37)
ωjc,min := sup{ωj : Ddj (ωj ) ≤ dc,max
}.
j
It is easy to see that these quantities reduce to those considered
in the proof of Theorem 2 in the case of continuously
differentiable cost functions.
The definitions in (37) imply that Dj (−ω0∗ ) ≥ pM,max
⇔
j
M,min
∗
ω0∗ ≤ ωjM,max , Dj (−ω0∗ ) ≤ pM,min
⇔
ω
≥
ω
,
0
j
j
c,max
∗
∗
Ddj (ω0∗ ) ≥ dc,max
⇔
ω
≥
ω
,
and
D
(ω
)
≤
dj
0
0
j
j
13
⇔ ω0∗ ≤ ωjc,min . Therefore, the exact arguments in the
dc,min
j
proof of Theorem 2 also apply here and guarantee that (27c)–
(27h) all hold with the choices (28).
Finally, let ν = ω0∗ . Then
M,max
−
∗ −ωj
Dj (−ν − λ+
+
λ
)
=
D
[−ω
]
j
0 −ω M,min
j
j
j
=
pM,max
[Dj (−ω0∗ )]pjM,min
j
= pM,∗
,
j
(38a)
by the above equivalences and equations (6h) and (35). Therefore, the optimality condition (27a) holds. Analogously,
ω c,max
Ddj (ν − µ+ + µ− ) = Ddj [ω0∗ ]ωjc,min
j
=
dc,max
[Ddj (ω0∗ )]djc,min
j
= dc,∗
j .
(38b)
We now note that the preimages of any point x ∈ R under Dj
and Ddj can be determined from the definitions (34) as
(
Cj0 (x),
x ∈ R \ Λj ,
−1
Dj (x) =
0
0
[Cj (γ−), Cj (γ+)], x = γ ∈ Λj ,
(
(39)
0
Cdj
(x),
x ∈ R \ Λdj ,
−1
Ddj (x) =
0
0
[Cdj
(γ−), Cdj
(γ+)], x = γ ∈ Λdj ,
0
It then follows from (39), the fact that Cj0 and Cdj
are strictly
increasing, and Definition 3 that, for all x ∈ R,
−1
Dj−1 (x) = ∂Cj (x) and Ddj
(x) = ∂Cdj (x).
(40)
Then, taking the preimages of (38a) and (38b) under Dj and
−
−1 M,∗
Ddj respectively yields −ν − λ+
) and ν −
j + λj ∈ Dj (pj
−
−1 c,∗
µ+
+
µ
∈
D
(d
),
whence
we
deduce
by
(40)
that
(36a)
j
j
j
dj
M,∗
¯c = dc,∗ .
and (36b) are both satisfied by p̄M
=
p
and
d
j
j
j
j
Therefore, all of the subgradient KKT conditions (36a), (36b), and (27c)–(27h) are satisfied, and so
(pM,∗ , dc,∗ , du,∗ ) defines a global optimum for the OSLC
problem (10).
Remark 12: Theorem 5 shows that the optimality result
given in Theorem 2 extends to situations in which the cost
functions are not required to be differentiable at all points.
Such non-differentiable cost functions can describe important
classes of droop control schemes, such ones involving deadbands, as discussed in Section VI.
R EFERENCES
[1] F. C. Schweppe, R. D. Tabors, J. L. Kirtley, H. R. Outhred, F. H.
Pickel, and A. J. Cox, “Homeostatic utility control,” IEEE Trans. Power
App. Syst., vol. PAS-99, no. 3, pp. 1151–1163, May 1980.
[2] J. A. Short, D. G. Infield, and L. L. Freris, “Stabilization of grid
frequency through dynamic demand control,” IEEE Trans. Power Syst.,
vol. 22, no. 3, pp. 1284–1293, Aug. 2007.
[3] D. Trudnowski, M. Donnelly, and E. Lightner, “Power-system frequency
and stability control using decentralized intelligent loads,” IEEE PES
T&D Conf. Exhib., pp. 1453–1459, May 2006.
[4] A. Molina-Garciá, F. Bouffard, and D. S. Kirschen, “Decentralized
demand-side contribution to primary frequency control,” IEEE Trans.
Power Syst., vol. 26, no. 1, pp. 411–419, May 2010.
[5] A. R. Bergen and V. Vittal, Power Systems Analysis. Prentice Hall, 1999.
[6] D. J. Hammerstrom et al., “Pacific Northwest GridWise testbed demonstration projects, part II: grid friendly appliance project,” Pacific Northwest Nat. Lab., Tech. Rep. PNNL-17079, Oct. 2007.
[7] B. J. Kirby, Spinning reserve from responsive loads. United States
Department of Energy, 2003.
[8] C. Zhao, U. Topcu, N. Li, and S. Low, “Design and stability of loadside primary frequency control in power systems,” IEEE Trans. Autom.
Control, vol. 59, no. 5, pp.1177–1189, 2014.
[9] C. Zhao, and S. Low, “Optimal decentralized primary frequency control
in power networks,” Proc. 53rd IEEE Conf. Decision Control, pp 2467–
2473, Dec. 2014.
[10] J. W. Simpson-Porco, F. Dörfler, and F. Bullo, “Synchronization and
power sharing for droop-controlled inverters in islanded microgrids,”
Automatica, vol. 49, no. 9, pp. 2603–2611, Sep. 2013.
[11] Q. Shafiee, J. M. Guerrero, and J. C. Vasquez, “Distributed secondary
control for islanded microgrids – a novel approach,” IEEE Trans. Power
Electron., vol. 29, no. 2, pp. 1018–1031, Apr. 2013.
[12] M. Andreasson, H. Sandberg, D. V. Dimarogonas, and K. H. Johansson,
“Distributed integral action: stability analysis and frequency control of
power systems,” Proc. 51st IEEE Conf. Decision Control, pp. 2077–
2083, Dec. 2012.
[13] S. Trip, M. Bürger, and C. De Persis. “An internal model approach to
frequency regulation in inverter-based microgrids with time-varying voltages,” Proc. 53rd IEEE Conf. Decision Control, pp 223–228, Dec. 2014.
[14] X. Zhang and A. Papachristodoulou, “A real-time control framework
for smart power networks with star topology,” Proc. American Control
Conf., 2013.
[15] H. Miyagi and A. R. Bergen, “Stability studies of multimachine power
systems with the effects of automatic voltage regulators,” IEEE Trans.
Autom. Control, vol. 31, no. 3, pp. 210–215, 1986.
[16] E. Devane, A. Kasis and I. Lestas, “Primary frequency regulation with
load-side participation Part II: beyond passivity approaches,” (under
review for publication) Feb. 2016.
[17] H. K. Khalil and J. W. Grizzle. Nonlinear systems, vol. 3, Prentice Hall
New Jersey, 1996.
[18] S. Sastry, Nonlinear systems: analysis, stability and control. SpringerVerlag New York, 1999.
[19] A. Molina-Garcia, F Bouffard, and D. S. Kirschen, “Decentralized
demand-side contribution to primary frequency control,” IEEE Trans.
Power Syst., vol. 26, no. 1, pp. 411–419, 2011.
[20] G. Rogers, Power System Oscillations. Kluwer Academic Publishers, 2000.
[21] K. W. Cheung, J. Chow, and G. Rogers, Power System Toolbox,
v. 3.0. Rensselaer Polytechnic Institute and Cherry Tree Scientific
Software, 2009.
[22] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge University Press, 2004.
[23] M. J. Steele. The Cauchy-Schwarz master class: an introduction to the
art of mathematical inequalities, Cambridge University Press, 2004.
[24] P. Moylan and D. Hill, “Stability criteria for large-scale systems,” IEEE
Trans. Autom. Control, vol. 23, no. 2, pp. 143–149, 1978.
[25] J. C. Doyle, B. A. Francis, and A. Tannenbaum. Feedback control theory,
vol. 1, Macmillan Publishing Company New York, 1992.
[26] R. T. Rockafellar. Convex analysis. Princeton University Press, 1970.
| 3cs.SY
|
1
Product Matrix MSR Codes with Bandwidth
Adaptive Exact Repair
arXiv:1708.03402v2 [cs.IT] 28 Dec 2017
Kaveh Mahdaviani∗, Soheil Mohajer† , and Ashish Khisti∗
∗ ECE Dept., University of Toronto, Toronto, ON M5S3G4, Canada
† ECE Dept., University of Minnesota, Minneapolis, MN 55455, USA
Email: {kaveh, akhisti}@comm.utoronto.ca, [email protected]
Abstract
In a distributed storage systems (DSS) with k systematic nodes, robustness against node failure is commonly provided by
storing redundancy in a number of other nodes and performing repair mechanism to reproduce the content of the failed nodes.
Efficiency is then achieved by minimizing the storage overhead and the amount of data transmission required for data reconstruction
and repair, provided by coding solutions such as regenerating codes [1]. Common explicit regenerating code constructions enable
efficient repair through accessing a predefined number, d, of arbitrary chosen available nodes, namely helpers. In practice, however,
the state of the system dynamically changes based on the request load, the link traffic, etc., and the parameters which optimize
system’s performance vary accordingly. It is then desirable to have coding schemes which are able to operate optimally under
a range of different parameters simultaneously. Specifically, adaptivity in the number of helper nodes for repair is of interest.
While robustness requires capability of performing repair with small number of helpers, it is desirable to use as many helpers as
available to reduce the transmission delay and total repair traffic.
In this work we focus on the minimum storage regenerating (MSR) codes, where each of the n nodes in the network is
supposed to store α information units, and the source data of size kα could be recovered from any arbitrary set of k nodes. We
introduce a class of MSR codes that realize optimal repair bandwidth simultaneously with a set of different choices for the number
of helpers, namely D = {d1 , · · · , dδ }. Our coding scheme follows the Product Matrix (PM) framework introduced in [2], and
could be considered as a generalization of the PM MSR code presented in [2], such that any di = (i + 1)(k − 1) helpers can
perform an optimal repair. As a result, the coding rate in our construction is limited by nk ≤ 12 . However, similar to the original
design of PM MSR codes, our solution can realize practical values of the parameter α. Recently [3] has presented another explicit
MSR coding scheme which is capable of performing optimal repair for various number of helpers. The solution presented in [3]
works for any arbitrary set of parameters k, D and can achieve high coding rates, but the required α for this code is exponentially
large. We show that the required value for α in the coding scheme presented in this work is exponentially smaller when compared
to [3] for the same set of other parameters. Particularly, for a DSS with n nodes, and k systematic nodes, the required value for
α is reduced from sn to sk, where s = lcm (d1 − k + 1, · · · , dδ − k + 1). We also show the required field size in the presented
coding scheme is equal to n1 .
I. I NTRODUCTION
Distributed storage systems (DSS) are compelling solutions to the fast growth of the demand in storage and accessibility of
digital content. The main feature of such systems is to provide accessibility and durability for the stored data by introducing
redundancy. In parallel, the number of storage components as well as the number of users connecting to these servers are
dramatically increasing. These increase the chance of failures in the system, due to storage device failures or inaccessible
nodes with overloaded traffic. Replication is the simplest approach to make the distributed storage system robust against such
failures, which is implemented in systems such as [5], [6], [6]. This approach however, provides simplicity in the cost of huge
storage overhead.
In the literature of erasure codes, there exists solutions, such as the Reed-Solomon (RS) code [7], which offer similar faulttolerance level as the replication does, with significantly less storage overhead. For instance, a 3-replication based DSS needs
to accommodate two mirrors for every single storage node in order to achieve reliability guarantee against any simultaneous
two node failure, which arises 200% storage overhead, while maximum distance separable (MDS) codes, such as RS code,
achieve the same level of fault-tolerance guarantee by adding only two redundant storage nodes for the whole system. In other
words, an erasure code with MDS property can guarantee to recover the whole source data from any subset of stored encoded
segments of collective size equal to the source data size.
On the other hand, when a node fails in a DSS, it needs to be replaced by a new node in order to maintain the system’s
performance. Such procedure is referred to as a repair. To perform a repair, the system downloads some data from a subset of
the surviving nodes, called helper nodes. The amount of data downloaded for a repair is referred to as the repair bandwidth.
In conventional MDS erasure codes, such as the RS codes, one has to reconstruct the entire source data and re-encode it to
recover a single lost segment. Hence, the repair bandwidth of these codes is at least as large as the size of the whole source
data.
Considering both storage overhead and repair bandwidth simultaneously, the family of regenerating codes [1], [8] offers
a very efficient coding mechanism for distributed storage systems. More precisely, a regenerating code on a Galois field Fq
1 This
work was presented in part at the 2017 IEEE Information Theory Workshop (ITW), Kaohsiung, Taiwan [4].
2
for a DSS with n storage nodes, maps the source data of size F symbols into n pieces of size α symbols each, and stores
them in n separate nodes, such that any k out of n nodes suffice to recover the data. Such a system is capable of tolerating
up to (n − k) node failures. Moreover, upon failure of one node, it can be replaced by a new node whose content of size α
symbols is determined by connecting to an arbitrary set of d (where d ≥ k) helper nodes, and downloading β symbols form
each (where β ≤ α). Hence, the total repair bandwidth in regenerating codes is dβ, which is denoted by γ. The parameter α
is also referred to as the subpacketization level.
Ideally, one would like to minimize the storage overhead, and repair bandwidth simultaneously. It turns out that for a given
file size F , there is a trade-off between the subpacketization level α and the repair bandwidth γ, and one can be minimized
only at the cost of a penalty for the other [1]. In particular, at one extreme point of this trade-off, one could first minimize
the subpacketization level, α, and then minimize the per-node repair bandwidth, β, to obtain a minimum storage regenerating
(MSR) code. As a result, MSR codes have the MDS property, and also minimize the repair bandwidth for the given α [1],
which means for an MSR code we have F = kα, and
F
.
(1)
β=
k(d − k + 1)
The total repair bandwidth (dβ) of an MSR code is then upper-bounded by that of RS code, and only coincides with that when
d = k. In other word, the total repair bandwidth is decreasing super-linearly as d grows in MSR codes.
Reversing the order of minimization between α, and β results in another extreme point of the trade-off. Such codes are
not MDS and hence have more storage overhead but provide the smallest possible repair bandwidth and are referred to as
minimum repair bandwidth (MBR) codes. Our focus in this work is on MSR codes as they minimize the storage cost.
In general, two distinct types of repair can be identified: in an exact repair scenario the replacement generated node will
contain the same data as stored in the failed node. In the so called functional repair, however, the replacement node may store
a different content, provided that the entire new system maintain the same properties as of the original one. In practice, exact
repair regenerating codes are much more appealing, mainly due to the fact that they could provide systematic encoding, which
is a significant advantage in practice.
It is shown that for functional repair MSR code design could be translated into a linear network coding problem [1]. Design
of such codes with exact repair property is more challenging, due to a large number of constraints that should be simultaneously
satisfied. Various constructions for exact repair MSR codes have been proposed for various sets of parameters [2], [3], [9]–[18].
Among the numerous available regenerating code designs, the common adopted model considers a rigid predetermined
number d (where k ≤ d ≤ n − 1) of helpers required for any repair procedure. Each of these helpers is also assumed to provide
β = γ/d repair symbols. This sets a single threshold for the system’s capability to perform the repair. On the other hand, in
practice the state of the system dynamically changes as a function of various factors including availability of nodes, traffic load,
available bandwidth, etc. It has been shown that when such characteristics change in the system the optimal number of helper
nodes for minimizing the cost of repair changes [19]. Therefore, runtime adaptation would be of great value towards optimizing
the performance. For instance, when the system is heavily loaded by many read requests, many nodes in the network might
not be capable to provide the required repair bandwidth and hence would be considered unavailable for the repair. This may
result in system’s failure to perform the repair, while there might be a few nodes, less than d, which are capable of providing
even more than β repair symbols. A natural question is whether one can download more data from the available helpers, and
accomplish the repair without the struggler helpers. Besides, from (1) it is clear that in optimal repair of MSR codes, increasing
the number of helpers reduces both β, and γ = dβ. One could consider a situation in which there are many nodes, more than
d, which are capable of contributing to the repair. It would be of great practical value then if the system could increase the
number of helpers and reduce both the per-node and total repair traffic. Note that this will also reduce the transmission delay,
which is one of the main bottlenecks in the DSS’s performance. We refer to such property as bandwidth adaptivity. Note that
the dynamic capability of service for storage nodes is a well-known characteristic for many practical distributed systems such
as peer-to-peer systems or heterogeneous network topologies [20]–[25].
The design of such codes has been of interest and the significance of bandwidth adaptivity in the performance of the system
has been emphasised in [9], [26]–[30]. However, these works mainly focus on the study of fundamental limits and appealing
properties of regenerating codes with bandwidth adaptivity, such as the significant improvement in the mean-time-to-data-loss
(MTTDL), and the coding schemes presented in these works either only provide functional repair [26]–[29], or relax some of
the constraints and consequently yield to a sub-optimum trade-of [9], [30].
When considering the optimal exact repair, it is a challenging problem to design such coding scheme with a large flexibility
degree since it needs to satisfy many optimality conditions simultaneously. As a result, this problem has only been considered
for the MSR [3], [9], and MBR [31] extreme points of the trade-off. For the MBR case, [31] provides an exact repair bandwidth
adaptive solution for a wide range of practical parameters based on the Product Matrix framework introduced in [2]. In [9]
a bandwidth adaptive code is provided based on interference alignment, which only achieves the MSR characteristics when
both α and β tend to infinity. The first explicit exact repair MSR code constructions which satisfy the bandwidth adaptivity
for finite values of α and β are introduced in [3]. These constructions work for any parameters k, n, and all values of d
such that k < d < n. However, the required values for the parameters α, and β in these constructions are still very huge
3
(i.e. exponentially large in n), and hence they only achieve optimality for extremely large contents. Recently, [17] introduced
a modified version of the code constructions in [3] which achieves MSR optimality for much lower values of α, at the cost
of losing bandwidth adaptivity. Indeed the MSR code in [17] works only for d = n − 1. We will review the related works in
more details in Section III.
In this work we will address the design problem of MSR codes with bandwidth adaptive exact repair for small α, and β,
following the Product Matrix framework [2]. For any positive integer design parameter δ, the presented coding scheme allows
us to choose the number of helper nodes for each repair scenario from the set {2(k − 1), · · · , (δ + 1)(k − 1)}, based on
availability of the nodes, network traffic and load state of the nodes. For any choice of the number of helpers, the per-node
repair bandwidth, β, takes its optimum value based on (1). To this end we change the design of the “message matrix” for the
PM MSR codes introduced in [2] to enable its partitioning into smaller symmetric submatices in many different ways. We then
also design appropriate repair and reconstruction schemes using a successive cancellation mechanism, to enable performing
repair based on different choices for the number of helpers, while keeping the required contribution of each helper at the
optimal value.
Product Matrix codes are practical regenerating codes due to their small subpacketization and repair bandwidth requirements.
Our focus in this work is to generalize the PM MSR codes to enable bandwidth adaptive repair, while keeping small practical
values for the parameters α, and β. For the construction presented in this work the required values for α and β are linear in
terms of the number of nodes in the DSS, which results in an exponential reduction compared to the constructions proposed
in [3]. However, unlike the constructions in [3] which achieve any coding rate k/n, it should be mentioned that the coding
rate in our construction is limited to k/n ≤ 1/2. The main contributions of this work are explained in the next section, after
formally defining the problem setup.
The rest of this paper is organized as follows: The following section formally introduced the problem setup and summarizes
the main contributions. Section III briefly reviews the most relevant works in the literature. Coding scheme and examples
are presented in Section IV, which is followed by a discussion on the properties of the code in Section V. Conclusion and
appendix are provided at the end.
II. M ODEL AND M AIN R ESULTS
A. Model
In this section we will briefly introduce a setup for the distributed storage system and the coding scheme of our interest.
This model is a modified version of the original setup considered in [1], and is very similar to the model considered in [31].
The first element we consider for the model of our bandwidth adaptive distributed storage system is a predefined Galois
filed alphabet, Fq of size q. Hereafter we assume all the symbols stored or transmitted through the network are elements of
Fq . Besides, we consider a homogeneous group of n storage nodes, each capable of storing α symbols.
Definition 1 (Bandwidth Adaptive Regenerating Code). Consider parameters α, n, k, δ, a set D = {d1 , · · · , dδ }, with
d1 < · · · < dδ , and a total repair bandwidth function γ : D → [α, ∞). A bandwidth adaptive regenerating code C(n, k, D, α,
γ) is a regenerating code with subpacketization level α, such that following mechanisms are guaranteed.
• Repair: In each repair process the number of helpers, d, can be chosen arbitrarily from the set D. The choice of helper
nodes is also arbitrary, and each of the chosen helpers then provides β(d) = γ(d)/d repair symbols.
• Data Reconstruction: The data collector recovers the whole source data by accessing any arbitrary set of k nodes.
Note that the flexibility of the repair procedure depends on the parameter δ, such that for a larger δ, there are more options
to select the number of helpers. In general, it is appealing to have small choices such as d1 , to guarantee the capability of code
to perform repair when the number of available helpers is small, and also large choices such as dδ , to provide the capability
of reducing the total as well as per-node repair bandwidth, and hence the transmission delay, whenever a larger number of
helpers are available. The coding scheme we present in this work allows to design such a range for the elements in D.
Definition 2 (Total Storage Capacity). For a set of parameters α, n, k, δ, a set D = {d1 , · · · , dδ }, and a given function
γ : D → [α, ∞), the total storage capacity of a bandwidth adaptive regenerating code, C(n, k, D, α, γ), is the maximum
size of a file that could be stored in a network of n storage nodes with subpacketization level α, using a bandwidth adaptive
regenerating code C(n, k, D, α, γ). We will denote the storage capacity of such a system by F (n, k, D, α, γ), or simply F ,
when the parameters could be inferred from the context.
Definition 3 (Bandwidth Adaptive MSR Codes, and Flexibility Degree). For any choice of parameters α, n, k, δ, and set
D = {d1 , · · · , dδ }, the bandwidth adaptive regenerating codes that realize both the MDS property defined by
F (n, k, D, α, γ) = kα,
(2)
as well as the the MSR characteristic equation simultaneously for all d ∈ D, given as
α = (d − k + 1)β(d), ∀d ∈ D,
(3)
4
are referred to as bandwidth adaptive MSR codes. Moreover, the number of elements in the set D is referred to as flexibility
degree of the code, and is denoted by δ.
Remark 1. It is worth mentioning that in a bandwidth adaptive MSR code, the total repair bandwidth,
dα
,
d−k+1
is a decreasing function of d. As a result, naive time-sharing between multiple MSR code components fails to satisfy optimality
in a bandwidth adaptive setting.
dβ(d) =
B. Main Results
The main contribution of this work is to provide a bandwidth adaptive MSR coding scheme with small subpacketization
level, and field size requirement. This coding scheme also guarantees exact repair of any failed node with many different
choices of the number of helpers. This result is formally stated in the following theorem. In this paper lcm() denotes the least
common multiple.
Theorem 1. For arbitrary positive integers n, k, and δ, there exists an adaptive bandwidth MSR code over a Galois field Fq
of size q ≥ n, with subpacketization level α and total storage capacity F , satisfying
α = (k − 1)lcm (1, 2, · · · , δ) ,
F = kα,
(4)
which is capable of performing exact repair using any di helpers, for
di = (i + 1)(k − 1), i ∈ {1, · · · , δ},
and simultaneously satisfies the MSR characteristic equation (3) for all di . i.e.,
α
, i ∈ {1, · · · , δ}.
β(di ) =
(di − k + 1)
We provide a constructive proof for this theorem based on introducing an explicit construction for a bandwidth adaptive
MSR coding scheme in Section IV.
III. R ELATED W ORKS
The property of bandwidth adaptivity has been of interest in the literature of regenerating codes for the last few years.
There has been a number of researchers who have addressed this problem under different settings. Some, such as [26]–[29],
have considered functional repair, in which code design problem reduces to linear network coding, and mainly focused on the
theoretical limitations and properties of a bandwidth adaptive regenerating codes. There are, however, a few other works which
have considered exact repair as well [3], [9], [31], [32]. In this section we briefly review these works and their relevance to
the problem setup introduced in the previous section.
In the MBR case, [31], [32] addressed a similar setup. For given integers dmin , dmax , authors presented a bandwidth adaptive
exact repair MBR regenerating code with D = {dmin , dmin + 1, · · · , dmax }, which is simultaneously error resilient against up
to a certain number of erroneous nodes in the network.
For the case of MSR, it was first in [26] that Shah et. al. extended the original problem of regenerating code design,
introduced in [1], [8], to provide flexibility. In [26] the authors consider the number of participating helpers to be selected
independently in each repair or data reconstruction procedure. Moreover, they also allow asymmetric participation of helpers,
as long as the per-node repair bandwidth of each helper, namely βi , i ∈ {1, · · · , d}, is less than a fixed upper bound, βmax .
While the setting considered by Shah et. al. provides much more flexibility in repair and reconstruction, they show that the
total repair bandwidth, γ, in their setting should satisfy
F
βmax ,
γ ≥ max{α − βmax , F mod α} +
α
which is always larger than that of the original regenerating codes formulations, except for the MSR case, i.e., F = kα, where
both settings achieve the same total repair bandwidth. However, the coding scheme presented in [26] only guarantees functional
repair.
in [27] Wang et. al. considered a functional repair coding scheme with symmetric per-node repair bandwidth, which supports
bandwidth adaptivity. Later [28] also considered a similar setup, while in both these works the main focus is on derivation
of the optimal trade-off between the storage overhead and repair bandwidth for the functional repair in a coordinated setting,
where more than one node failure is considered to be repaired together. Note, however, that none of these works address the
exact repair with bandwidth adaptivity in regenerating codes.
Aggrawal et. al. [29] also considered a functional repair setup to achieve bandwidth adaptivity in regenerating codes. They
analysed the mean-time-to-Data-Loss (MTTDL) in the regenerating codes with and without bandwidth adaptivity. Their analysis
5
is based on a birth-death process model in which the population of available storage node randomly changes with appropriately
chosen rates. They showed that bandwidth adaptivity provides a significant gain in terms of MTTDL.
When considering exact repair regenerating codes, Cadambe et. al. were the first to address the bandwidth adaptivity [9].
They present an interference alignment based regenerating coding scheme which is capable of performing exact repair with
bandwidth adaptivity in the repair procedure. The code presented by Cadambe et. al. is the first exact repair regenerating code
with bandwidth adaptivity, however, their coding scheme only asymptotically achieves the MSR optimality, when α and β tend
to infinity with proper ratio. The importance of this result, however, is to show that bandwidth adaptivity could be implemented
without extra cost in the optimal trade-off between the storage overhear and repair bandwidth, at least for the MSR codes,
even when the exact repair is required.
In [30], a similar setup, referred to as progressive engagement, is considered for regenerating codes with bandwidth adaptive
exact repair, and two coding schemes are provided. While the two coding schemes both preserve the MDS property, none of
them satisfy the MSR characteristic equation (3) simultaneously for different choices of d. Another main difference between
the progressive engagement setup and the one considered in this work is that the authors in [30] relax the property that any
subset of surviving nodes could be considered as helpers by assuming that all the remaining systematic nodes are always
participating as helpers. Moreover, they require the set of available choices for the number of helpers, namely D, to be
D = {k, k + 1, · · · , n − 1}, while in our formulation, D does not include all integers between k and n − 1.
The first explicit constructions for exact repair MSR codes with bandwidth adaptivity and finite subpacketization level were
introduced in [3]. These constructions work for any arbitrary values of k and n, and the set of choices for the number of helpers,
D, could be designed to contain any value di such that k < di < n. However, the required value for the subpacketization level,
α, in both constructions is still considerably huge. In particular, for a DSS with n storage nodes, and the set D = {d1 , · · · , dδ },
constructions in [3] require
n
α = (lcm (d1 − k + 1, · · · , dδ − k + 1)) .
(5)
Therefore, these constructions only achieve optimality for storage of contents which are exponentially large in terms of the
number of storage nodes in the system.
In coding schemes suggested for practical networks, the number of nodes, n, is usually larger than 10 [33], [34], and the
required per-node storage should not exceed a few hundreds of megabytes. Indeed at the scale of multiple megabytes, the
performance of the system is already limited by bandwidth of the network link and disk I/O [6], [33]. Unfortunately, the
constructions presented in [3] still struggle to satisfy such choices. For instance the smallest realization of these codes for the
setting n = 14, and k = 10 suggested for Facebook [33] requires storing more than 1015 encoded symbols in each storage node,
which translates to several peta-bytes per node. Hence the problem of designing exact repair bandwidth adaptive regenerating
code remains yet open for practical range of parameters.
In this work, as presented in Theorem 1, we address the exact repair bandwidth adaptive MSR code design problem with
small subpacketization level. While both [30] and [3] follow the approach introduced in [10], which is based on the design of
parity check equations, in this work we follow the product matrix (PM) framework introduced in [2]. Comparing (5) with (4),
one could see that the presented scheme exponentially reduces the required values of α (and β). However, this scheme works
only for 2k − 2 ≤ di , ∀di ∈ D. As a result, the construction presented in this work only solves the problem in low coding
rates. It is worth mentioning that in applications such as general-purpose storage arrays, providing high reliability and fast
degraded reads are more important than maximizing the coding rate [35]. However, the design of high-rate bandwidth adaptive
MSR codes with small α and β still remains a challenging important problem for big data storage systems such as Hadoop.
IV. C ODING S CHEME
In this section we introduce a bandwidth adaptive exact repair MSR coding scheme, which could be designed to provide
any required flexibility degree with elements of the set D evenly located between the smallest and the largest element, namely
d1 and dδ . We describe the encoding and decoding schemes for storage, repair, and data reconstruction procedures for the
adaptive bandwidth MSR code in the following subsections. This coding scheme is closely related to the product matrix MSR
code introduced in [2], and could be considered as an extension of the product matrix code that achieves bandwidth adaptivity.
As mentioned in the previous section, we will assume all the source and encoded symbols, are elements of a Galois field of an
appropriately large field size q ≥ n, denoted by Fq . Moreover, all the operations hereafter are considered to be filed operations
of Fq . We will refer to Fq as the code alphabet.
In the design of the proposed coding scheme, we chose the design parameters, namely k, and the required flexibility degree
δ. All the other parameters of the code, including α, F , D = {d1 , · · · , dδ }, and β(di ) will be then determined based on k,
and δ as follows. The subpacketization level is
α = (k − 1) · lcm (1, 2, · · · , δ) .
(6)
Moreover, we have F = kα, which satisfies the MDS property. Finally, for D we have
D = {d1 , · · · , dδ }, di = (i + 1)(k − 1), i ∈ {1, · · · , δ}.
(7)
6
and for any di ∈ D, the associated per-node and total repair bandwidths denoted by β(di ), and γ(di ) respectively are
β(di ) =
α
α
=
,
di − k + 1
i(k − 1)
γ(di )
= di β(di ) =
(i + 1)α
.
i
(8)
A. Coding for Storage
We begin the introduction of the coding scheme by describing the process of encoding the source symbols and deriving the
encoded symbols to be stored in the storage nodes. Similar to the product matrix codes, the first step in encoding for storage
in this scheme is to arrange the information symbols in a matrix, denoted by M , which we refer to hereafter as the message
matrix. Let
zδ = lcm (1, 2, · · · , δ) .
The message matrix in our coding scheme is structured as
S1 S2 O O O
S2 S3 S4 O O
O S4 S5 S6 O
O O S6 S7 S8
M = .
..
O
···
O
···
O
···
(9)
follows,
O
O
O
O
···
···
···
···
O
O
O
O
..
.
,
..
.
O S2zδ −4 S2zδ −3 S2zδ −2
O O S2zδ −2 S2zδ −1
O O
O
S2zδ
(10)
where, each Si , i ∈ {1, · · · , 2zδ } is a symmetric (k − 1) × (k − 1) matrix filled with k(k − 1)/2 source symbols, and O is a
(k − 1) × (k − 1) zero matrix. Therefore, M has (zδ + 1)(k − 1) rows and zδ (k − 1) columns. Note that the total number of
distinct source symbols in the message matrix M is
F = k(k − 1)zδ = kα.
(11)
Example 1. Consider the design parameters k = 3, and δ = 2. Therefore, from (9) we have zδ = 2, and using (11), the
maximum number of source symbols that we can arrange in the message matrix M is 12. Denoting the source symbols by
s1 , · · · , s12 , then we have
s1 s2
s4 s5
s5 s6
s2 s3
S1 S2
s4 s5
s7 s8
.
M = S2 S3 =
(12)
s5 s6
s8 s9
O S4
0
0
s10 s11
0
0
s11 s12
△
Once the message matrix is ready, the source encoder creates the vector of coded symbols for each of the n storage nodes,
by calculating the product of a node-specific coefficient vector and the message matrix. To describe this process, we first need
the following definition.
Definition 4. [Generalized Vandermonde Matrix] For distinct and non-zero elements e1 , · · · , em of Fq , and some integer
c ≥ 0, a matrix Am×ℓ with entries
Ai,j = eic+j−1 , for i ∈ {1, · · · , m}, j ∈ {1, · · · , ℓ},
is referred to as a generalized Vandermonde matrix.
The following lemma about these matrices will be used in the proofs of the following theorems,
Lemma 1. Consider distinct and non-zero elements e1 , · · · , eℓ in Fq , and an integer c ≥ 0. Then a square ℓ × ℓ generalized
Vandermonde matrix, A, as defined in Definition 4, is invertible in Fq .
The proof of the above lemma, simply follows from the fact that,
c
1 e1 e21 · · · eℓ−1
e1 0 · · · 0
1
0 ec2 · · · 0 1 e2 e22 · · · eℓ−1
2
A= .
,
..
. . ..
..
. .
.
c
ℓ−1
2
0 0 · · · eℓ
1 eℓ eℓ · · · eℓ
7
and the two matrices on the right hand side are both full-rank, as one of them is diagonal with non-zero diagonal elements
and the other one is a square Vandermonde matrix.
Back to the description of our coding scheme, for distinct and non-zero elements ei ’s in Fq , with i ∈ {1, · · · , n} we set
c = 0, and define a generalized Vandermonde matrix of size n × (zδ + 1)(k − 1) as
(z +1)(k−1)−1
1 e1 e21 · · · e1 δ
(z +1)(k−1)−1
1 e2 e22 · · · e2 δ
.
Ψ=
..
.
(z +1)(k−1)−1
1 en e2n · · · en δ
Note that all of the submatrices of Ψ are also generalized Vandermonde matrices. We refer to Ψ as the coefficient matrix and
denote the j th row of Ψ by ψ j . The vector of encoded symbols to be stored on node j, j ∈ {1, · · · , n}, denoted by xj , is
calculated as
xj = ψ j M.
(13)
The vector ψ j is the node-specific coefficient vector for storage node j. Note that the per-node storage capacity requirement
for this coding scheme is then zδ (k − 1) as given by (6).
Example 2. Let’s consider the setting in Example 1 again, and assume we have n = 7 nodes in the network. Assume that
the code alphabet is the Galois field F11 . The coefficient matrix could be formed based on (e1 , e2 , · · · , e7 ) = (1, 2, · · · , 7) as
follows,
1 1 1 1 1 1
1 2 4 8 5 10
1 3 9 5 4 1
Ψ=
(14)
1 4 5 9 3 1 .
1 5 3 4 9 1
1 6 3 7 9 10
1 7 5 2 3 10
The encoded content of nodes 1 to 7 can be calculated using (12) and (13). For instance, for node 1 we have
x1 = [x1,1 , x1,2 , x1,3 , x1,4 ] = [1, 1, 1, 1, 1, 1]M,
which gives
x1,1 = s1 + s2 + s4 + s5 ,
x1,2 = s2 + s3 + s5 + s6 ,
x1,3 = s4 + s5 + s7 + s8 + s10 + s11 ,
x1,4 = s5 + s6 + s8 + s9 + s11 + s12 ,
and similarly for node 7 we have
x7 = [x7,1 , x7,2 , x7,3 , x7,4 ] = [1, 7, 5, 2, 3, 10]M,
which gives
x7,1 = s1 + 7s2 + 5s4 + 2s5 ,
x7,2 = s2 + 7s3 + 5s5 + 2s6 ,
x7,3 = s4 + 7s5 + 5s7 + 2s8 + 3s10 + 10s11 ,
x7,4 = s5 + 7s6 + 5s8 + 2s9 + 3s11 + 10s12 .
△
B. Data Reconstruction
In order to reconstruct all the information stored in the system, the data collector accesses k arbitrary nodes in the network
and downloads all their contents. To describe the details of the decoding we use the following lemma.
Lemma 2. Let X and Φ be two known matrices of size k × (k − 1), such that Φ is a generalized Vandermonde matrix, and
assume ∆ is a known diagonal matrix of size k × k, with distinct and non-zero diagonal elements. Then the equation
X = ΦA + ∆ΦB,
(15)
8
is uniquely solvable for unknown (k − 1) × (k − 1) symmetric matrices A and B.
The proof of this lemma is based on the data reconstruction scheme of the product matrix MSR codes, introduced in [2], and
is provided in Appendix A to help keeping this paper self-contained. The following theorem explains the data reconstruction
procedure in this coding scheme.
Theorem 2. For the coding scheme presented in subsection IV-A, there exists a decoding scheme to reconstruct all the source
symbols arranged in the message matrix M from the encoded content of any arbitrary group of k storage nodes.
Proof. Let’s assume the set of accessed nodes is {ℓ1 , · · · , ℓk }. Moreover, let’s denote the k × (zδ + 1)(k − 1) submatrix
of Ψ associated with the nodes ℓ1 , · · · , ℓk , by ΨDC . We will further denote the submatrix of ΨDC consisting of columns
(i − 1)(k − 1) + 1 through i(k − 1), by ΨDC (i). In other words, we have a partitioning of ΨDC ’s columns as
ψℓ
1
(16)
ΨDC = ... = [ΨDC (1), · · · , ΨDC (zδ + 1)] .
ψℓ
k
As a result, defining the diagonal matrix
ΛDC
(k−1)
eℓ
1
0
=
..
.
0
0
(k−1)
e ℓ2
0
0 ···
0 ···
..
.
0
0
..
.
(k−1)
0 · · · e ℓk
,
for each k × (k − 1) submatrix ΨDC (i) we have
ΨDC (i + 1) = ΛDC ΨDC (i).
(17)
Similarly, let’s denote the matrix consisting of the collected encoded vectors by XDC , and its partitioning to k × (k − 1)
submatrices XDC (i), i ∈ {1, · · · , zδ } as follows
xℓ1
XDC = ... = [XDC (1), · · · , XDC (zδ )] .
(18)
xℓk
The decoding procedure for data reconstruction consists of zδ consecutive steps. The first step uses only the submatrix
XDC (1). Using (17) we have,
S
XDC (1) = [ΨDC (1), ΨDC (2)] 1
S2
= ΨDC (1)S1 + ΛDC ΨDC (1)S2 .
Using Lemma 2, the decoder recovers both S1 , and S2 , using XDC (1), in step one. Then, for i ∈ {2, · · · , zδ }, the decoder
performs step i by using submatrix XDC (i), and decodes submatrices S2i−1 , and S2i , as follows.
In step i, i ∈ {2, · · · , zδ }, of the data reconstruction decoding, the decoder uses the submatrix XDC (i). Note that
S2(i−1)
XDC (i) = [ΨDC (i − 1), ΨDC (i), ΨDC (i + 1)] S2i−1
S2i
S
= ΨDC (i − 1)S2(i−1) + [ΨDC (i), ΨDC (i + 1)] 2i−1 .
S2i
Having the submatrix S2(i−1) already recovered from step i − 1, the decoder first calculates
X̂DC (i) = XDC (i) − ΨDC (i − 1)S2(i−1) .
S
= [ΨDC (i), ΨDC (i + 1)] 2i−1 .
S2i
(19)
Then from (17), and (19), we have
X̂DC (i) = ΨDC (i)S2i−1 + ΛDC ΨDC (i)S2i .
Again using Lemma 2, decoder recovers S2i−1 , and S2i at the end of the step i of the decoding. Therefore, by finishing step zδ ,
the decoder reconstructs all the submatrices in the message matrix M , and recovers the whole data stored in the network.
9
Example 3. Following the setting described in Example 1 and Example 2, we have k = 3, and the coefficient matrix Ψ is
given in (14). Let’s assume the data collector accesses the storage nodes 1, 2, and 4. Then, with zδ = 2, we have,
1 1 1 1 1 1
ΨDC = [ΨDC (1), ΨDC (2), ΨDC (3)] = 1 2 4 8 5 10 ,
1 4 5 9 3 1
and,
1 1
1 1
1 1
ΨDC (1) = 1 2 , ΨDC (2) = 4 8 , ΨDC (3) = 5 10 .
1 4
5 9
3 1
Moreover, we have
ΛDC
12 0 0
1 0 0
= 0 22 0 = 0 4 0 ,
0 0 42
0 0 5
(20)
and
XDC
x1
x1,1 x1,2 x1,3 x1,4
= [XDC (1), XDC (2)] = x2 = x2,1 x2,2 x2,3 x2,4 ,
x4
x4,1 x4,2 x4,3 x4,4
and finally,
x1,1 x1,2
x1,3 x1,4
XDC (1) = x2,1 x2,2 , XDC (2) = x2,3 x2,4 .
x4,1 x4,2
x4,3 x4,4
(21)
The first step of decoding in the data reconstruction process based on the encoded data stored in nodes 1, 2 and 4, starts
by using only XDC (1), as given in (21). Using Lemma 2, The decoder then recovers S1 , and S2 submatrices of the message
matrix M .
In the second step then the decoder first calculates
X̂DC (2) = XDC (2) − ΨDC (1)S2 ,
which is equal to
X̂DC (2) = ΨDC (2)S3 + ΛDC ΨDC (2)S4 ,
and hence is of the desired form of (15). Therefore, again using lemma 2, the decoder recovers S3 , and S4 , which completes
the data reconstruction.
△
Algorithm IV-B summarizes the data reconstruction mechanism in this coding scheme.
Algorithm 1 Data Reconstruction
1: Input: k, zδ , xℓi , ψ , and eℓi for i ∈ {1, · · · , k}.
ℓi
2: Output: Submatrices S1 , · · · , S2zδ .
3: Form matrices ΨDC (i), for i ∈ {1, · · · , zδ + 1}, using (16).
4: Form matrices XDC (i), for i ∈ {1, · · · , zδ }, using (18).
5: Form matrix ΛDC , using (20).
6: Recover submatrices S1 , S2 from XDC (1), ΨDC (1), and ΛDC , using Lemma 2.
7: for i = 2 to zδ do
8:
Calculate X̂DC (i) = XDC (i) − ΨDC (i − 1)S2(i−1) .
9:
Recover submatrices S2i−1 , S2i from X̂DC (i), ΨDC (i), and ΛDC , using Lemma 2.
10: end for
10
C. Bandwidth Adaptive Exact Repair
We now describe the bandwidth adaptive repair procedure, by assuming that node f is failed and the set of helpers selected
for the repair are H = {h1 , · · · , hd }, for some d ∈ D. The following theorem describes the repair procedure in this bandwidth
adaptive MSR code.
Theorem 3. Consider the coding scheme presented in subsection IV-A, with design parameters k, and δ, and D as defined
in (7). For any arbitrary failed node f , and any arbitrary set of helpers H = {h1 , · · · , hd }, for some d ∈ D, there exists a
repair scheme for recovering the content of node f with per-node repair bandwidth,
α
.
(22)
β(d) =
d−k+1
Remark 2. Note that (7) and (22) are consistent with (8), which satisfies the MSR characteristic equation (3) for any d ∈ D.
Proof. Without loss of generality let d = (m+1)(k −1), for some m ∈ {1, · · · , δ}. Note that (6), and (7) guarantee that for any
d ∈ D, α is an integer multiple of d − k + 1, hence β(d) is an integer. Each helper node h ∈ H, creates β(d) = α/(d − k + 1)
repair symbols to repair node f as follows. First helper node h partitions its encoded content into β(d) equal segments, such that
for i ∈ {1, · · · , β(d)}, the segment xh (i) is of size α/β(d) = d − k + 1 = m(k − 1), and contains elements xh,(i−1)m(k−1)+1
through xh,im(k−1) . Then we have
xh = [xh (1), · · · , xh (β(d))] .
(23)
Similarly, for any node ℓ, we split the first α entries of a coefficient vector assigned to node ℓ, namely ψ ℓ , into β(d) equal
segments as
h
i
(24)
ψ ℓ (1 : α) = ψ ℓ (1), · · · , ψ ℓ (β(d)) ,
where each segment ψ ℓ (i) is of size d − k + 1 = m(k − 1).
Now each helper node h ∈ H, creates its β(d) repair symbols as
r(h, f ) = r1 (h, f ), · · · , rβ(d) (h, f )
h
⊺
⊺ i
= xh (1) ψ f (1) , · · · , xh (β(d)) ψ f (β(d))
.
The repair decoder then stacks d repair vectors r(h, f ), for h ∈ H, into a d × β(d) matrix
r(ℓ1 , f )
..
ΥH =
.
.
(25)
(26)
r(ℓd , f )
We then introduce the following partitioning of the matrix ΥH , into β(d) submatrices, as follows
ΥH = [ΥH (1), · · · , ΥH (β(d))] ,
th
(27)
where ΥH (i), i ∈ {1, · · · , β} is the i column in ΥH , of size d × 1.
Before starting to describe the repair decoding procedure, we need to introduce some notations associated to a given
repair scenario. Consider a repair procedure with d = (m + 1)(k − 1). For the corresponding β(d) = α/(d − k + 1), we
will partition matrix M as depicted in Fig. 1, and equation (28). Note that this results in β(d) non-overlapping diagonal
submatices Mi , i ∈ {1, · · · , β(d)}, each of size m(k − 1) × m(k − 1), along with (k − 1) × (k − 1) symmetric submatrices
S2m , S4m , · · · , S2β(d)m = S2zδ as shown in the figure. It is worth mentioning that the general pattern of the partitioning
shown in figure 1 is preserved the same for all d ∈ D, and only the size, and number of the Mi diagonal blocks changes
for different choices of the parameter d. From the construction of the message matrix, introduced in (10), each Mi submatrix
will be symmetric. As a result, the message matrix M could be interpreted in terms of the submatrices Mi , and S2i for
i ∈ {1, · · · , β(d)}, associated to a repair procedure with d = (m + 1)(k − 1), d ∈ D.
Finally the last notations we use to describe the adaptive repair decoding scheme, using a given set of helpers H =
{h1 , · · · , hd }, is,
(im+1)(k−1)−1
(i−1)m(k−1)+1
(i−1)m(k−1)
· · · e h1
e h1
e h1
..
, i ∈ {1, · · · , β(d)}.
(29)
ΩH (i) =
.
(i−1)m(k−1)
(i−1)m(k−1)+1
(im+1)(k−1)−1
e hd
e hd
· · · e hd
11
m(k − 1) m(k − 1)
m(k − 1)
S2m
M1
S2m
m(k − 1)
M2
S2(i−1)m
S4m
M =
0
S4m
S2im
Mi
S2(i−1)m
S2(β−1)m
S2im
0
S2(β−1)m
Mβ
S2zδ
Fig. 1. In the above figure β represents β(d), for some chosen d ∈ D, such that, d = (m + 1)(k − 1). Moreover, the white area is filled by zeros, and each
coloured square represent a non-zero symmetric submatrix of M .
S2(i−1)m+1 S2(i−1)m+2
O
O
O
···
O
S2(i−1)m+2 S2(i−1)m+3 S2(i−1)m+4
O
O
···
O
O
S2(i−1)m+4 S2(i−1)m+5 S2(i−1)m+6 O
···
O
..
..
..
Mi =
.
.
.
.
O
···
O S2im−4 S2im−3 S2im−2
O
···
O
O
(28)
S2im−2 S2im−1
Note that, ΩH (i), i ∈ {1, · · · , β(d)}, is a d × d generalized Vandermonde matrix and hence is invertible as shown in Lemma
−1
1. We denote the upper (d − k + 1) × d submatrix of (ΩH (i)) by ΘH (i), and the lower (k − 1) × d submatrix by ΞH (i).
Then we have,
ΘH (i)
−1
(ΩH (i)) =
.
(30)
ΞH (i)
The decoding procedure for the repair of node f is performed in β(d) sequential steps as will be described in the following.
Let’s begin with the first step.
For the failed node f , let φf denote the 1 × (k − 1) vector ,
h
i
(31)
φf = 1 , ef , · · · , efk−2 .
Using a partitioning similar to (23) for xf , then we have,
xf (1) = ψ f
O
M1
· · · O S2m
i
h
m(k−1)
= ψ f (1)M1 + O, · · · , O, ef
φf S2m
(k−1)×m(k−1)
.
(32)
In the first step, the decoder recovers the two terms on the right in the above equation to reconstruct xf (1) using only the first
repair symbol received from each of the helpers, namely r1 (hi , f ), for i ∈ {1, · · · , d}, as follows.
12
Using (23) to (27), and the partitioning denoted in Fig. 1, the submatrix ΥH (1), introduced in (27) can be written as,
⊺
xh1 (1) ψ f (1)
xh1 (1)
⊺
..
..
ΥH (1) =
(1)
ψ
=
.
f
.
⊺
x
(1)
hd
xhd (1) ψ f (1)
(m+1)(k−1)−1
2
1 e h1 e h1 · · · e h1
⊺
M1
..
=
(33)
.
ψ f (1) .
(m+1)(k−1)−1
1 ehd e2hd · · · ehd
O · · · O S2m
Then using (29), we have,
ΥH (1) = ΩH (1)
O
M1
···
O S2m
⊺
ψ
(1)
.
f
(34)
Multiplying the inverse of ΩH (1) from right to the both sides of (34), and using (30) the decoder derives
That gives,
O
M1
···
O S2m
⊺ Θ (i)
H
ψ
(1)
=
ΥH (1)
f
Ξ
(i)
H
⊺
M1 ψ f (1) = ΘH (1)ΥH (1),
(35)
⊺
(m−1)(k−1)
φf
S2m ef
= ΞH (1)ΥH (1).
(36)
and similarly, using (31),
Since both M1 , and S2m are symmetric, from (35) we have,
⊺
ψ f (1)M1 = (ΘH (1)ΥH (1)) ,
(k−1)
and from (36), by multiplying the scalar ef
(37)
, we get
m(k−1)
ef
φf S2m
(k−1)
= ef
⊺
(ΞH (1)ΥH (1)) .
From (37), and (38), and using (32) the decoder then recovers xf (1) as,
h
i
m(k−1)
xf (1) = ψ f (1)M1 + O, · · · , O, ef
φf S2m
(k−1)×m(k−1)
(38)
,
(39)
where, the rightmost term in the above expression is derived by padding m − 1, (k − 1) × (k − 1) zero matrices, O, to the left
of the matrix calculated in (38).
In step i for i = 2 through β(d) of the repair decoding, the decoder then recovers xf (i), using ΥH (i) received from the
(i−1)m(k−1)
helpers, along with ef
φf S2(i−1)m , recovered from the step i − 1 of decoding. To this end, first note that similar to
(33) the repair symbols in ΥH (i) can be written as,
⊺
xh1 (i) ψ f (i)
xh1 (i)
⊺
..
..
(i)
ψ
=
ΥH (i) =
.
f
.
⊺
(i)
x
h
d
xhd (i) ψ f (i)
S2(i−1)m O · · · O
(im+1)(k−1)−1
(i−1)m(k−1)+1
(i−1)m(k−1)
· · · e h1
e h1
e h1
⊺
.
ψ (i) .
.
=
M
i
.
f
(i−1)m(k−1)
(i−1)m(k−1)+1
(im+1)(k−1)−1
e hd
e hd
· · · e hd
O
· · · O S2m
13
Using (29) we can rewrite the above equation as,
(i−1)m(k−1)−µ
φℓ
e ℓ1
1
⊺
(i−1)m(k−1)
.
.
φf + ΩH (i)
ΥH (i) =
.
S2(i−1)m ef
(i−1)m(k−1)−µ
e ℓd
φℓ
d
O
⊺
ψ f (i) .
Mi
···
O S2im
The decoder first removes the contribution of the S2(i−1)m submatrix in the repair symbols in ΥH (i) by calculating
(i−1)m(k−1)−(k−1)
φℓ
e ℓ1
1
⊺
..
S2(i−1)m e(i−1)m(k−1) φ
Υ̂H (i)=ΥH (i)−
.
(40)
f
.
f
(i−1)m(k−1)−(k−1)
e ℓd
φℓ
d
⊺
(i−1)m(k−1)
(i−1)m(k−1)
In the above expression, S2(i−1)m ef
φf
is itself derived by transposing ef
φf S2(i−1)m . As a result we
have,
Υ̂H (i) = ΩH (i)
O
⊺
ψ
(i)
.
f
Mi
···
O S2im
Therefore, similar to (35) through (38) the decoder derives,
ψ f (i)Mi = (ΘH (i)ΥH (i))⊺ ,
(41)
and
im(k−1)
ef
(k−1)
φf S2im = ef
⊺
(ΞH (i)ΥH (i)) .
(42)
Finally, using (41) and (42), we have
i
h
im(k−1)
φf S2im
xf (i) = ψ f (i)Mi + O, · · · , O, ef
(k−1)×m(k−1)
.
(43)
The following algorithm summarizes the bandwidth adaptive repair procedure.
Algorithm 2 Bandwidth Adaptive Repair
1: Input: f , ef , β, m, k and r(h, f ), ψ , eh , for h ∈ H.
h
2: Form matrices ΥH (i), for i ∈ {1, · · · , β}, using (26), and (27).
3: Form vectors φ using (31).
f
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
−1
Calculate matrices ΩH (i), and (ΩH (i)) , for i ∈ {1, · · · , β}, using (29).
Calculate matrices ψ f (1)M1 using (37).
Calculate matrices ema
f φf S2m using (38).
Recover xf (1) using (39).
for i = 2 to β do
Calculate Υ̂H (i) using (40).
Calculate ψ f (i)Mi using (35).
im(k−1)
Calculate ef
φf S2im using (42).
Recover xf (i) using (43).
end for
Form xf = [xf (1), · · · , xf (β)].
−1
Remark 3. Note that calculating the inverse matrices (ΩH (i)) , for i ∈ {1, · · · , β}, can be carried out recursively since
from (29) we have,
(i−1)m(k−1)
0
···
0
e h1
(i−1)m(k−1)
···
0
0
e h2
ΩH (i) =
..
..
..
ΩH (1)
.
.
.
(i−1)m(k−1)
0
0
· · · e hd
14
Remark 4. In a DSS with n nodes, for D = {d1 , · · · , dδ }, the bandwidth adaptive MSR codes presented in [3] require
n
α = (lcm (d1 − k + 1, · · · , dδ − k + 1)) .
(44)
Comparing (44) with (6), one could see that the presented scheme reduces the required α (and β) values exponentially. However,
this scheme works only for 2k − 2 ≤ di , ∀di ∈ D. Hence, the design of high-rate bandwidth adaptive MSR codes with small
α and β still remains an open problem.
The following example provides a detailed illustration of the MSR bandwidth adaptive exact repair procedure in the same
setup as described in Examples 1 to 3.
Example 4. As in the previous examples we will consider k = 3, and δ = 2, and using (6), α = 4. The code alphabet is
F11 , and the message matrix M as given in (12). As a result, using (7), we have D = {d1 , d2 } = {4, 6}, and from (22) their
associated per-node repair bandwidths are β1 = 2, and β2 = 1. Without loss of generality, let’s assume node 7 is failed, i.e.
f = 7. In this setup, the following two repair scenarios are then possible.
A: One option is to use d = d2 = 6 helpers, and download only β2 = 1 repair symbol from each of them, which means
d = (m + 1)(k − 1) for m = 2. Consider H = {1, 2, · · · , 6}. In this case, using the coefficient matrix Ψ for this setup, which
is given in (14), from (24), with α = 4, β = 1, for any node ℓ we have
ψ ℓ (1) = ψ ℓ (1 : 4),
which is a row vector consisting of the first α = 4 elements of the coefficient vector assigned to node ℓ. As a result, using
(23), each helper node h ∈ H will use all of its encoded content xh to create a single repair symbol as
h
⊺ i
r(h, f ) = [r1 (h, f )] = xh (1) · ψ f (1)
= [xh · [1, 7, 5, 2]⊺] .
The repair decoder will then receive
ΥH = [ΥH (1)]
x1,1 + 7x1,2 + 5x1,3 + 2x1,4
r(1, f )
x2,1 + 7x2,2 + 5x2,3 + 2x2,4
x3,1 + 7x3,2 + 5x3,3 + 2x3,4
= ... =
x4,1 + 7x4,2 + 5x4,3 + 2x4,4 .
r(6, f )
x5,1 + 7x5,2 + 5x5,3 + 2x5,4
x6,1 + 7x6,2 + 5x6,3 + 2x6,4
Moreover, using (31), and with ef = 7 from (14) as also used in all previous examples, for f = 7 we get
φf = [1, 7] .
Finally notice that in this case the repair decoding will only
im(k−1)
and k = 3 we get eℓ
= e4ℓ , and hence,
1 1
1 2
1 3
ΩH (1) =
1 4
1 5
1 6
have one single step. Therefore, for step i = 1, with m = 2,
1
4
9
5
3
3
1
8
5
9
4
7
1 1
5 10
4 1
.
3 1
9 1
9 10
Then, based on the partition represented in Fing. (1), for
M1 =
S1 S2
,
S2 S3
(45)
15
the decoder has access to
⊺
M1
ψ (1)
ΥH = ΩH (1)
f
O S4
s1 s2 s4 s5
s2 s3 s5 s6
⊺
s4 s5 s7 s8
ψ (1) .
= ΩH (1)
f
s5 s6 s8 s9
0 0 s10 s11
0 0 s11 s12
Note that the ΩH (1), given in (45) is an invertible matrix in the code
6 7
10
10
3 6
ΘH (1)
−1
(ΩH (1)) =
=
1 8
ΞH (1)
2 7
1 6
(k−1)
The decoder then calculates the lost data xf , using (39), with ef
alphabet F11 , and we have
9 7 6 10
9 0 3 1
9 1 9 5
0 8 2 3 .
7 5 8 4
10 1 5 10
= 72 = 5 in F11 , as follows
⊺
⊺
xf = (ΘH (1)ΥH ) + [0, 0, 5 (ΞH (1)ΥH ) ] .
B: The second option is to use d = d1 = 4 helpers. With k = 3 we have d = (m + 1)(k − 1) for m = 1. In this case,
using (22) we will have β1 = 4/(4 − 2) = 2 repair symbols per helper node. Let’s without loss of generality assume node 7 is
failed, i.e. f = 7, and H = {1, 2, 3, 4} is the set of helper nodes chosen to perform the repair. As a result, using the coefficient
matrix in the code alphabet F11 , given in (14), from equation (24) for the helper nodes we have
ψ 1 (1) = [1, 1], ψ 1 (2) = [1, 1],
ψ 2 (1) = [1, 2], ψ 2 (2) = [4, 8],
ψ 3 (1) = [1, 3], ψ 3 (2) = [9, 5],
ψ 4 (1) = [1, 4], ψ 4 (2) = [5, 9],
and for the failed node f = 7,
ψ f (1) = [1, 7], ψ f (2) = [5, 2].
Similarly, for the coded content of each of these nodes we consider the following partition
xℓ = [xℓ (1), xℓ (2)]
= [[xℓ,1 , xℓ,2 ], [xℓ,3 , xℓ,4 ]], for ℓ ∈ {1, 2, 3, 4, 7}.
Each helper node h ∈ H then creates two repair symbols
r(h, f ) = [r1 (h, f ), r2 (h, f )]
h
⊺
⊺ i
= xℓ (1) ψ f (1) , xℓ (2) ψ f (2)
= [xh,1 + 7xh,2 , 5xh,3 + 2xh,4 ],
and the repair decoder receives
ΥH = [ΥH (1), ΥH (2)]
r1 (1, f )
r2 (1, f )
..
..
=
.
.
r1 (4, f )
r2 (4, f )
5x1,3 + 2x1,4
x1,1 + 7x1,2
x2,1 + 7x2,2 5x2,3 + 2x2,4
=
x3,1 + 7x3,2 5x3,3 + 2x3,4 .
5x4,3 + 2x4,4
x4,1 + 7x4,2
16
Moreover, using (31), and with ef = 7 from (14) as also used in all previous examples, for f = 7 we get
φf = [1, 7] .
Finally notice that in this case the repair decoding will have two steps. Therefore, for step i = 1, with m = 1, and k = 3
im(k−1)
we get eℓ
= e2ℓ , and hence, using equation (29), we have
1 1 1 1
1 2 4 8
ΩH (1) =
(46)
1 3 9 5 .
1 4 5 9
Then, according to the partition represented in Fig. (1), for
M 1 = S1 , M 2 = S3 ,
the repair decoder has access to
⊺
M1
ψ f (1)
ΥH (1) = ΩH (1)
S2
s1 s2
s2 s3 1
= ΩH (1)
s4 s5 7 .
s5 s6
As expected ΩH (1) is an invertible matrix in the code alphabet F11 , and we have
4 5 4 10
ΘH (1)
3 4 4 0
−1
(ΩH (1)) =
= 7 7 9 10 .
ΞH (1)
9 6 5 2
m(k−1)
The decoder then calculates the lost data xf (1), using (39), with ef
= 72 = 5 in F11 , as follows
⊺
⊺
xf (1) = ΘH (1)ΥH(1) + 5 (ΞH (1)ΥH (1)) .
(i−1)m(k−1)
To start the second step, i = 2, of the repair decoding, then the repair decoder first uses ef
⊺
(i−1)m(k−1)
φf
S2 e f
= 5 (ΞH (1)ΥH (1)) ,
= 72 = 5, to calculate
where 5 (ΞH (1)ΥH (1)) is already derived in the step 1. Then the decoder calculates Υ̂H (2) using equation (40) as,
((2−1)m−1)(k−1)
e1
φ1
(2−1)m(k−1) ⊺
.
S2 e
..
Υ̂H (2) = ΥH (2) −
φf
f
((2−1)m−1)(k−1)
e4
φ4
1 1
r2 (1, f )
r2 (2, f ) 1 2
=
r2 (3, f ) − 1 3 S2 (5 (ΞH (1)ΥH (1))) .
1 4
r2 (4, f )
From (41), with
the repair decoder then has access to
1
4
ΩH (2) =
9
5
1
8
5
9
1 1
5 10
,
4 1
3 1
⊺
M2
ψ f (2)
Υ̂H (2) = ΩH (2)
S4
s7 s8
s8 s9 5
.
= ΩH (2)
s10 s11 2
s11 s12
17
As expected ΩH (2) is also an invertible matrix in the code alphabet F11 ,
4 4
ΘH (2)
3 1
−1
(ΩH (2)) =
= 7 10
ΞH (2)
9 7
and we have
9 2
9 0
.
1 2
3 7
2m(k−1)
Finally, the decoder calculates the lost data xf (2), using (39), with ef
= 74 = 3 in F11 , as follows
⊺
xf (2) = ΘH (2)ΥH(2) + 3 (ΞH (2)ΥH (2))⊺ .
V. D ISCUSSION
△
AND PROPERTIES
In this section we will briefly review some of the technical requirements and properties of the coding scheme presented in
this work. Particularly, we will show that the field size and subpacketization level requirements of the presented coding scheme
are not significantly limiting factors in the practical implementations.
A. Field Size Requirement
The only factor that influences the choice of the code alphabet Fq in the presented coding scheme is the existence coefficient
matrix Ψ, and all the inverses of its square submatrices. To satisfy this requirement for a network with n storage nodes, it
is enough to have q ≥ n [36], which is the same as the field size requirement of many other coding schemes such as the
Product Matrix codes [2] or the commonly used Reed-Solomon codes [7]. Hence, the presented coding scheme in this work
is providing the bandwidth adaptivity property at no extra cost in the field size requirements, and any field of size larger
than n could be used as the code alphabet. It worth mentioning that the field size requirement of the only other bandwidth
adaptive exact repair MSR constructions, introduced in [3] is lower bounded by n2 − kn which is significantly larger for large
distributed storage networks n. Moreover, techniques such as those presented in [37] could easily be applied in the presented
coding scheme to reduce the field size to any arbitrary (e.g. binary) small field. However, this discussion is out of the scope
of this work.
B. Subpacketization Level
The requirement for the subpacketization level α for the presented coding scheme is given in by (6). A natural question that
arises is how fast does α grow as a function of the code design parameters, k, and δ. From (6) it is clear that α is proportional
to the design parameter k. Regarding the dependency of α on δ, from [38] we have,
lcm (1, 2, · · · , δ) ≤ 4δ ,
and, recently [39] showed that
lcm (1, 2, · · · , δ) ≥ 2δ , for δ ≥ 7.
As a result, for the presented coding scheme, we have
(k − 1)2δ ≤ α ≤ (k − 1)4δ , for δ ≥ 7.
VI. C ONCLUSION
We presented an alternative solution for exact repair MSR codes in which optimal exact repair is guaranteed simultaneously
with a range of choices, D = {d1 , · · · , dδ }, for the number of helpers. The introduced coding scheme is based on the Product
Matrix framework, introduced first in [2]. The repair mechanism in this framework is based on specific symmetries in the
structure of the message matrix. We proposed a novel structure for the message matrix, which preserves the required symmetries
in many different submatrices. Corresponding repair mechanisms are also introduced to utilise these symmetries to perform
optimal repair, whit different choices for the number of helpers, namely di = (i + 1)(k − 1), i ∈ {1, · · · , δ}. In addition, the
data reconstruction procedure is enhanced based on a novel successive interference cancellation scheme to perform optimally
under the new design of message matrix. Comparing to the only other explicit constructions with exact optimal bandwidth
adaptive repair, presented in [3], we showed that when di ≥ 2k − 2, ∀di ∈ D, the required values for α, is reduced from zδn
to kzδ for a DSS with n nodes, and k systematic nodes, which also reduces β exponentially.
18
A PPENDIX A
P ROOF OF L EMMA 2
Multiplying both sides of (15) by Φ⊺ from right we get
XΦ⊺ = ΦAΦ⊺ + ∆ΦBΦ⊺ .
(47)
Following the notation in [2], we introduce
P = ΦAΦ⊺ , Q = ΦBΦ⊺ .
Then using (47) and the above equations we have
XΦ⊺ = P + ∆Q.
(48)
Note that both P and Q are symmetric k × k matrices. Recall that ∆ is a diagonal matrix, with non-zero and distinct diagonal
elements. Hence, for any 1 ≤ i < j ≤ k, we now have both
(XΦ⊺ )i,j = Pi,j + ∆i,i Qi,j ,
(49)
and
(XΦ⊺ )j,i = Pj,i + ∆j,j Qj,i
= Pi,j + ∆j,j Qi,j .
(50)
Then using (49) and (50), we recover all the non-diagonal elements of P and Q. Now, let p̂i , and q̂ i denote the ith row of
matrices P and Q excluding their diagonal elements. Moreover, let Φ̂i denote the submatrix of Φ derived by removing the ith
row, and finally let’s denote the ith row of Φ by φi . Then for all i ∈ {1, · · · , k} we have
⊺
p̂i = φi A Φ̂i ,
⊺
q̂ i = φi B Φ̂i ,
⊺
and using Lemma 1, Φ̂i is an invertible matrix, as it is a transposed generalized Vandermonde matrix. Then we calculate
the following two matrices using the above equations for i ∈ {1, · · · , k − 1},
φ1
φ1
..
.
. A = Φ̂k A, .. B = Φ̂k B.
φ(k−1)
φ(k−1)
Now since Φ̂k is invertible we have both A, and B.
R EFERENCES
[1] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” IEEE Transactions
on Information Theory, vol. 56, no. 9, pp. 4539–4551, Sep. 2010.
[2] K. V. Rashmi, N. B. Shah, and P. V. Kumar, “Optimal exact regenerating codes for distributed storage at the MSR and MBR points via a product-matrix
construction,” IEEE Transactions on Information Theory, vol. 57, no. 8, pp. 5227–5239, August 2011.
[3] M. Ye and A. Barg, “Explicit constructions of high-rate MDS array codes with optimal repair bandwidth,” IEEE Transactions on Information Theory,
vol. 63, no. 4, pp. 2001–2014, April 2017.
[4] K. Mahdaviani, S. Mohajer, and A. Khisti, “Product matrix minimum storage regenerating codes with flexible number of helpers,” in Proc. IEEE
Information Theory Workshop (ITW), Kaohsiung, Taiwan, Nov. 2017.
[5] K. Shvachko, H. Kuang, S. Radia, and R. Chansler, “The Hadoop distributed file system,” in Proc. of IEEE Symposium on Mass Storage Systems and
Technologies (MSST), Lake Tahoe Incline Villiage, NV, USA, May 2010.
[6] “HDFS architecture guide,” Available online: https://hadoop.apache.org/docs/r1.2.1/hdfs design.html#Data+Blocks, 2013.
[7] I. S. Reed and G. Solomon, “Polynomial codes over certain finite fields,” Journal of the Society for Industrial and Applied Mathematics (SIAM), vol. 8,
no. 2, pp. 300—-304, Feb. 1960.
[8] A. G. Dimakis, P. B. Godfrey, M. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” in Proc. IEEE International
Conference on Computer Communications (INFOCOM), Anchorage, Alaska, USA, May 2007, pp. 2000–2008.
[9] V. Cadambe, S. Jafar, H. Maleki, K. Ramchandran, and C. Suh, “Asymptotic interference alignment for optimal repair of MDS codes in distributed
storage,” IEEE Transactions on Information Theory, vol. 59, no. 5, pp. 2974–2987, May 2013.
[10] V. Cadambe, C. Huang, and J. Li, “Permutation code: optimal exact-repair of a single failed node in MDS code based distributed storage systems,” in
Proc. IEEE International Symposium on Information Theory (ISIT), Saint Petersburg, Russia, Aug. 2011, pp. 1225–1229.
[11] I. Tamo, Z. Wang, and J. Bruck, “Zigzag codes: MDS array codes with optimal rebuilding,” IEEE Transactions on Information Theory, vol. 59, no. 3,
pp. 1597–1616, March 2013.
[12] B. Sasidharan, G. K. Agarwal, and P. V. Kumar, “A high-rate MSR code with polynomial sub-packetization level,” in Proc. IEEE International Symposium
on Information Theory (ISIT), Hong Kong, June 2015, pp. 2051–2055.
[13] V. Cadambe, C. Huang, and J. Li, “An alternate construction of an access-optimal regenerating code with optimal sub-packetization level,” in Proc. 21st
Nat. Conf. Commun. (NCC), Bombay, India, Feb. 2015, pp. 1–6.
19
[14] K. Kralevska, D. Gligoroski, and H. Øverby, “General sub-packetized access-optimal regenerating codes,” IEEE Communication Letters, vol. 20, no. 7,
pp. 1281–1284, July 2016.
[15] N. Raviv, N. Silberstein, and T. Etzion, “Constructions of high-rate minimum storage regenerating codes over small fields,” IEEE Transactions on
Information Theory, vol. 63, no. 4, pp. 2015–2038, April 2017.
[16] Z. Wang, I. Tamo, and J. Bruck, “Explicit minimum storage regenerating codes,” IEEE Transactions on Information Theory, vol. 62, no. 8, pp. 4466–4480,
Aug. 2016.
[17] M. Ye and A. Barg, “Explicit constructions of optimal-access MDS codes with nearly optimal sub-packetization,” IEEE Transactions on Information
Theory, vol. 63, no. 10, pp. 6307–6317, Oct. 2017.
[18] B. Sasidharan, M. Vajha, and P. V. Kumar, “An explicit, coupled-layer construction of a high-rate MSR code with low sub-packetization level, small
field size and d < (n − 1),” in Proc. IEEE International Symposium on Information Theory (ISIT), Aachen, Germany, June 2017, pp. 2048–2052.
[19] S. Jiekak, A.-M. Kermarrec, N. L. Scouarnec, G. Straub, and A. V. Kempen, “Regenerating codes: A system perspective,” Proc. ACM SIGOPS Oper.
Syst. Rev., vol. 47, no. 2, pp. 23–32, July 2013.
[20] “Spacemonkey project,” http://www.spacemonkey.com.
[21] “Tahoe: The least-authority file system,” http://www.tahoe-lafs.org/trac/tahoe-lafs.
[22] F. Dabek, F. Kaashoek, D. Karger, R. Morris, and I. Stoica, “Wide-area cooperative storage with CFS,” in Proc. ACM Symposium on Operating Systems
Principles (SOSP), Chateau Lake Louise, Banff, Canada, October 2001, pp. 202–215.
[23] S. Ghemawat, H. Gobioff, and S.-T. Leung, “The Google file system,” in Proc. ACM Symposium on Operating Systems Principles (SOSP), NewYork,
USA, October 2003.
[24] S. Rhea, P. Eaton, D. Geels, H. Weatherspoon, B. Zhao, and J. Kubiatowicz, “Pond: The OceanStore prototype,” in Proc. USENIX Conference on File
and Storage Technologies (FAST), San Francisco, USA, March-April 2003.
[25] R. Bhagwan, K. Tati, Y.-C. Cheng, S. Savage, and G. M. Voelker, “Total recall: System support for automated availability management,” in Proc. USENIX
Conference on Networked System Design and Implementation (NSDI), San Francisco, USA, March 2004.
[26] N. B. Shah, K. V. Rashmi, and P. V. Kumar, “A flexible class of regenerating codes for distributed storage,” in Proc. IEEE International Symposium on
Information Theory (ISIT), Austin, Texas, USA, June 2010, pp. 1943–1947.
[27] X. Wang, Y. Xu, Y. Hu, and K. Ou, “MFR: Multi-loss flexible recovery in distributed storage systems,” in Proc. IEEE International Conference on
Communications (ICC), Cape Town, South Africa, May 2010, pp. 1–5.
[28] A.-M. Kermarrec, N. L. Scouarnec, and G. Straub, “Repairing multiple failures with coordinated and adaptive regenerating codes,” in Proc. IEEE
International Symposium on Network Coding (NetCod), Beijing, China, July 2011, pp. 1–6.
[29] V. Aggarwal, C. Tian, V. A. Vaishampayan, and Y.-F. R. Chen, “Distributed data storage systems with opportunistic repair,” in Proc. IEEE International
Conference on Computer Communications (INFOCOM), Toronto, Canada, April–May 2014, pp. 1833–1841.
[30] M. Hajiaghayi and H. Jafarkhani, “MDS codes with progressive engagement property for cloud storage systems,” Available online: https://arxiv.org/abs
/1605.06927, 2016.
[31] K. Mahdaviani, A. Khisti, and S. Mohajer, “Bandwidth adaptive & error resilient regenerating codes with minimum repair bandwidth,” in Proc. IEEE
International Symposium on Information Theory (ISIT), Barcelona, Spain, July 2016, pp. 235–239.
[32] ——, “Bandwidth adaptive & error resilient MBR exact repair regenerating codes,” Available online: https://arxiv.org/pdf/1711.02770.pdf, 2017.
[33] K. V. Rashmi, N. B. Shah, D. Gu, H. Kuang, D. Borthakur, and K. Ramchandran, “A solution to the network challenges of data recovery in erasure-coded
distributed storage systems: A study on the facebook warehouse cluster,” in Proc. 5th USENIX Workshop Hot Topics Storage File Syst. (HotStorage),
San Jose, USA, June 2013.
[34] C. Huang, H. Simitci, Y. Xu, A. Ogus, B. Calder, P. Gopalan, J. Li, and S. Yekhanin, “Erasure coding in windows azure storage,” in Proc. of the USENIX
Annual Technical Conference (ATC), Boston, USA, June 2012.
[35] S. Narayanamurthy, “Modern erasure codes for distributed storage systems,” in Storage Developer Conference (SDC), SNIA, Santa Clara, 2016.
[36] H. Althaus and R. Leake, “Inverse of a finite-field vandermonde matrix,” IEEE Transactions on Information Theory, vol. 15, no. 1, pp. 173–173, Jan.
1969.
[37] N. Raviv, “Asymptotically optimal regenerating codes over any field,” in Proc. IEEE International Symposium on Information Theory (ISIT), Aachen,
Germany, June 2017, pp. 1416–1420.
[38] M. Nair, “On chebyshev-type inequalities for primes,” The American Mathematical Monthly, vol. 88, no. 2, pp. 126–129, Feb. 1982.
[39] V. Diekert, M. Kufleitner, G. Rosenberger, and U. Hertrampf, Discrete Algebraic Methods: Arithmetic, Cryptography, Automata and Groups. Walter
de Gruyter GmbH & Co KG., 2016.
| 7cs.IT
|
ADA: A Game-Theoretic Perspective on Data Augmentation for Object Detection
Sima Behpour
University of Illinois at Chicago
Kris M. Kitani
Carnegie Mellon University
arXiv:1710.07735v2 [cs.CV] 12 Dec 2017
[email protected]
[email protected]
Brian D. Ziebart
University of Illinois at Chicago
[email protected]
Abstract
The use of random perturbations of ground truth data,
such as random translation or scaling of bounding boxes,
is a common heuristic used for data augmentation that
has been shown to prevent overfitting and improve generalization. Since the design of data augmentation is largely
guided by reported best practices, it is difficult to understand if those design choices are optimal. To provide a more
principled perspective, we develop a game-theoretic interpretation of data augmentation in the context of object detection. We aim to find an optimal adversarial perturbations
of the ground truth data (i.e., the worst case perturbations)
that forces the object bounding box predictor to learn from
the hardest distribution of perturbed examples for better
test-time performance. We establish that the game theoretic
solution, the Nash equilibrium, provides both an optimal
predictor and optimal data augmentation distribution. We
show that our adversarial method of training a predictor
can significantly improve test time performance for the task
of object detection. On the ImageNet object detection task,
our adversarial approach improves performance by over
16% compared to the best performing data augmentation
method.
Figure 1. Tiger localization example with three different bounding
box annotations illustrates ambiguity in ‘ground truth’ labels.
tics for augmenting the ground truth through consensus
[30, 23, 35, 22]. Despite these efforts, it is not clear if there
is a principled approach for identifying the optimal ground
truth distribution in the context of supervised learning.
To address in part the uncertainty of the ‘ground truth’
annotations, dataset augmentation methods can be used to
synthesize new annotations of images by perturbing annotations. In fact, heuristic data augmentation preprocessing
such as random translation, flipping or scaling, has been
shown to be essential for many modern visual learning tasks
using deep networks. However, manually choosing perturbations to improve performance can be an error-prone process. While increasing the modes of data perturbations may
effectively increase the amount of training data, it can also
cause the learned predictor to drift. In this way, current techniques for data augmentation are more of an art than a science.
Towards a more principled approach to data augmentation, we propose to integrate annotation perturbations directly into the learning process. We do this by introducing
an adversarial function that generates maximally perturbed
version of the ground truth, which makes it as hard as possible for the predictor to learn. The adversary, however, is not
completely free to perturb the data. It must retain certain
1. Introduction
There is no guarantee that human-labeled ‘ground-truth’
annotations of an image dataset are error free. Consider the
bounding box annotations of three annotators of the image
in Figure 1. Do all boxes contain the object? Are all three
bounding boxes equally correct? Is there one bounding
box which is most helpful for learning a detection model?
These questions highlight the ambiguity in the annotation
task. In response, many helpful heuristics have been utilized in the literature to obtain more consistent annotations.
To deal with inter-annotator disagreement [33, 19, 27],
previous work has relied primarily on reasonable heuris1
feature statistics (e.g., the features of the new bounding box
distribution should still be close to the features of the original bounding box). Formally, we pose the data augmentation problem as a zero-sum game between a player (the
predictor model) seeking to maximize performance and a
constrained adversary (augmented data distribution) seeking to minimize expected performance [8, 1, 32].
To the best of our knowledge, this is the first work to provide a theoretic basis for data augmentation in terms of an
adversarial two player zero-sum game. As a consequence
of our game-theoretic formulation, we develop a novel adversarial loss function that identifies the optimal data augmentation strategy which leads to the most robust predictor possible (i.e., trained for the worst case perturbation
of data). In our experiments, we focus on the task of object detection and show that our proposed adversarial data
augmentation technique consistently improves performance
over various competing loss functions, data augmentation
levels, and deep network architectures.
2. Related Work
It is common to assume that the ground truth is singular and error-free. However, disagreement between annotators is a widely-known problem for many computer vision
tasks [33], as well as a major concern [19] when constructing an annotated computer vision corpora. In large part, the
difficulty arises because the set of possible “ground truth”
annotations is typically extremely large for vision tasks. It
is a powerset of possible descriptions (e.g., words, noun
phrases) in annotation tasks, multi-partitions of the pixels
(exponential in the number of pixels) in segmentation tasks,
and the possible bounding boxes (quadratic in the number
of pixels) for localization tasks.
Methods to form a “consensus” annotation and to improve the annotation process through crowd-sourcing have
been developed by averaging or combining together different independent annotations [30], verifying annotations
with other independent annotators [23], and other strategies [35, 22]. For example, the ILSVRC2012 image dataset
employs boundary box drawing, quality verification, and
coverage verification as three separate subtasks [25] in a
crowd-sourcing pipeline. In the construction of that dataset,
proposed bounding boxes are rejected 37.8% of the time
[25], illustrating the inherent disagreement between annotators and the uncertainty of the task. Despite the added safeguards of the verification process, recent evaluations have
also been performed by removing a substantial fraction of
the training examples that are considered to have poor quality bounding boxes [27, 6, 34, 26, 21].
Many state-of-the-art methods for object detection [14]
are based on CNN, and incorporate other improvements
such as the use of very large scale datasets, more efficient
GPU computation, and data augmentation [3]. Recently,
most of the literature on data augmentation studies effective data augmentation methods for CNN features that increases the performance of different tasks (e.g., classification, object recognition) [28, 20, 16, 11]. Chatfield [3] applies the data augmentation techniques commonly applied
to CNN-based methods to shallow methods and shows an
analogous performance boost [3]. Paulin et al. [20] claim
that given a large set of possible transformations, all transformations are not equally informative and adding uninformative transformations increases training time with no gain
in accuracy. They propose Image Transformation Pursuit
(ITP) algorithm for the automatic selection of a compact
set of transformations.
Complementary to our work, data augmentation can also
be used to guard against adversarial attacks [7, 18, 9, 13,
29]. Total variance minimization and image quilting are presented as very effective defenses adversarial-example attacks on image-classification systems [9]. The strength of
these data augmentation lies in their non-differentiable nature and their inherent randomness resulting difficult defenses for an adversary. Our work is different in that we
seek to optimize the data augmentation process as part of a
supervised learning problem.
3. Problem Formulation
In order to understand the underlying theory of adversarial data augmentation proposed in this work, we must first
understand the role of the annotation distribution, p(y|x),
which describes the distribution over labels y (e.g., a bounding box annotation) given a feature vector x (e.g., an RGB
image). Note that a training dataset D = {yn , xn }N
n=1 ,
induces an annotation distribution p(y|x). In other words,
each label yn in the training set can be interpreted to be
a sample from the annotation distribution which is conditioned on a feature vector x ∈ RD . When there is absolute
certainty in the ground truth annotation, the annotation distribution p(y|x) is an indicator function where it is one for
the true label y ∗ and zero otherwise.
3.1. Data Augmentation
The process of data augmentation is a method of altering the annotation distribution. A typical method for data
augmentation generates new examples D̃ by perturbing the
training data D. For example, if the label is a structured output like a bounding box (i.e., a vector of four values), we
can generate a new structured label ỹ for the same image by
slightly perturbing the original ‘ground truth’ bounding box
y ∗ . This data augmentation process creates a new underlying annotation distribution p̃(y|x). Since data augmentation
can be used to generate multiple new labels for the same
feature vector x, the annotation probability p(y|x) becomes
a soft distribution over labels.
Now if we are given a loss function `(ŷ, y) describing the
distance between an estimated label ŷ and annotation label
y, we can compute the expected loss of the estimated label
under the annotation distribution as:
X
P (y|x)`(ŷ, y).
(1)
y∈Y
Notice that the expected loss is the smallest when the
estimated label matches the annotation distribution. Conversely, the expected loss grows larger when the estimated
label is far from the annotation distribution. It is important
to note here that this marginalization over the annotation
distribution is rarely made explicit in the loss function in
most modern supervised learning objective functions because the distribution is assumed to be an indicator function
at the ‘ground truth’ label.
Now consider the probabilistic predictor f (y|x) which
maps a feature vector x to a distribution over labels y. The
expected loss over the entire dataset D under the predictor
distribution and annotation distribution is defined as:
expected loss for input x
}|
{
X
X zX
0
0
f (y |x)
P (y|x)`(y , y) .
(2)
min
f ∈Γ
x∈D y 0
y
The goal of supervised learning is to find the optimal predictor f (from some set of predictors Γ), that minimizes the
above expected loss over the labeled training data. Understanding this verbose form of the supervised learning objective function is critical for the formulation that follows.
3.2. Adversarial Data Augmentation
If we adopt a pessimistic view of the annotated data
and assume uncertainty in the ‘ground truth’ annotations,
we can use data augmentation to perturb the ‘ground truth’
annotations to reflect this uncertainty. We can go further
and assume the worst case: that the quality of the annotation distribution is maximally perturbed. In other words, we
make a strong pessimistic assumption that the annotation
distribution was generated by an adversary. By making this
worst case assumption, we hypothesize that we can train a
more robust predictor that is resilient to large perturbations
it might encounter at test time. Figure 2 illustrates the three
possible choices of annotation distributions for a single image.
More formally, instead of the common Empirical Risk
Minimization (ERM) objective of Eq. (2), we aim to learn a
predictor f that optimizes the following adversarial objective function:
XX
X
min
f (y 0 |x) max
P (y|x)`(y 0 , y). (3)
f ∈Γ
x∈D y 0
P (y|x)
y
Notice that the maximization sub-problem has been inserted
into the objective function which reflects our assumption
Figure 2. Types of annotation distributions. Adversarial augmentation selects bounding boxes that are maximally different from the
ground truth but still contain important object features.
that the annotation distribution is adversarial (i.e., the worst
case distribution). One might quickly notice that this is an
unreasonable objective function without some additional
constraints because the adversarial annotation distribution
can be arbitrarily bad. In the next section we will incorporate constraints that limit the adversary from deviating very
far from the original ground truth annotations.
3.3. Game Formulation
Our claim is that data augmentation should be included
in the learning problem instead of being an independent
data pre-processing step. By incorporating data augmentation into the predictor learning problem, we obtain a saddle
point optimization problem where we pit a predictor trying
to minimize the loss, against an adversarial annotation distribution that is trying to maximize the loss. In this form,
the optimization can be seen as a minimax problem over a
zero-sum two-player game.
In the language of game theory, the player (predictor)
selects a label from a mixed strategy y 0 ∼ f (y 0 |x) to
minimize the loss, while the opponent (annotation distribution) selects an annotation from the adversarial distribution
y ∼ P (y|x) to maximize the loss. The equilibrium point of
the game yields both the optimal predictor and an optimal
data annotation distribution. The game is zero-sum because
the negative loss of the player (predictor) is exactly the gain
of the adversary (annotation distribution).
The value or payoff of the game for a particular feature
vector x is the expected loss of the predictor distribution
against the adversary’s annotation distribution:
X
E y0 |x ∼f [`(y 0 , y)] =
f (y 0 |x)`(y 0 , y)P (y|x) (4)
y|x ∼P
y 0 ,y
= f > Gp.
(5)
The expected loss of the game can also be written in matrix
form, where f is the vector of probabilities obtained from
the predictor over all labels, G is the game matrix where
each element contains the loss between two labels, and p is
the annotation distribution vector.
The adversarial objective function, Eq. (3) in its current
form is problematic because the adversarial annotation distribution is free to perturb the ground truth annotations in
arbitrary ways that have no similarity to the original annotations. This can be prevented by constraining the adversarial annotation distribution to choose label distributions in a
way that retains feature statistics of the original ground truth
annotation. For example, we may want the mean of a set of
augmented bounding box annotations to be the same as the
mean of the original bounding box annotation. Formally, we
can define the first-order statistic of the ground truth data as:
Ey,x∼D [φ(y, x)] =
N
1 X
φ(yn , xn ),
N n=1
(6)
where (yn , xn ) is the nth training example in D. We are
now ready to define the constrained adversarial optimization problem.
Definition 1. The Primal Adversarial Data Augmentation (ADA-P) game is:
min max Ex ∼ D, [`(y 0 , y)] such that:
f
P
y 0 |x ∼ f,
y|x ∼ P
(7)
Ex ∼ D, [φ(y, x)] = Ey,x∼D [φ(y, x)]
y|x ∼ P
where f (y 0 |x) and P (y|x) are distributions over all potential predicted labels for each feature vector x.
Due to strong Lagrangian duality [2], a dual problem
with an equivalent solution can be formulated by including the constraint in the objective function using a vector of
Lagrangian multipliers, θ. This resulting Lagrangian potential θ> φ(·, ·) links together a set of otherwise independent
zero-sum games.
Definition 2. The Dual Adversarial Data Augmentation
(ADA-D) game is:
h
h
min Ex,y∗ ∼D min max Ey0 ∼ f, `(y 0 , y)
(8)
P
θ
f
y ∼ P
i
+ θ> {φ(y, x) − φ(y ∗ , x)} .
We make two important observations based on this dual
optimization perspective. First, since P is adversarially chosen, there is no need to restrict or parameterize f to avoid
overfitting to P as is typically done in supervised learning.
Instead, the feature potential based on θ is learned to provide constraints on the adversary that make prediction easier. Further, since P is chosen after f in Eq. (8), the predictor is incentivized to randomize.
4. Adversarial Object Localization
Up to this point, we have described our proposed adversarial data augmentation learning approach in general
terms, as it can apply to many structured output tasks. Now,
we shift our focus to the concrete problem of object detection. This explicit focus will help us to describe our approach in concrete terms.
Label Space. Each structured output label y is represented
by the four coordinates of a bounding box. The domain of
a label is denoted Y. The set of all possible bounding boxes
Y is very large for an image of modest size and therefore it
is rarely practical to evaluate all possible bounding boxes.
This means that the sums over labels used in the formulation above (Section 3) are not tractable and that some form
of distribution approximation is needed. To discretize the
label space Y, we use a bounding box proposal algorithm,
Edgebox [15], to generate a set of k bounding boxes to define the label space Y.
Feature Statistics. To represent the feature statistics
φ(y, x) of a bounding box y over an image x, we use
the FC7 features of the VGG16 [24] network. Concretely,
it is a 4096 dimensional vector over a sub-image defined
by the bounding box y. The feature statistic constraint
{φ(y 0 , x) − φ(y ∗ , x)} described in the ADA-D definition
represents the difference between the FC7 features of an arbitrary bounding box y 0 and the FC7 features of the ground
truth bounding box y ∗ . Also known as the perceptual loss,
this quantity ensures that the adversarial bounding box label
remains perceptually similar to the ground truth bounding
box label.
Loss Function. The loss function used for object localization is based on the classical intersection over union
(IoU) score, IoU (y, y 0 ) = area(y ∩ y 0 )/area(y ∪ y 0 ). Here,
y and y 0 are two bounding boxes. In this work, we focus
on losses defined in terms of the amount of non-overlap,
`(y, y 0 ) = 1 − IoU(y, y 0 ), which equals to one when y and
y 0 are disjoint, zero when they are identical, and smoothly
transitions in between those extremes. Another loss function we use is the overlap loss with a threshold:
(
1 IoU(y, y 0 ) < α
`tα (y, y 0 ) =
(9)
0 IoU(y, y 0 ) ≥ α,
which assigns binary loss to bounding boxes depending on
the overlap threshold.
4.1. Game Matrix
As noted in Eq. (5), the expected loss of the adversarial
game can be written in matrix form, f > Gp. The game (or
payoff) matrix G for ADA-D can be constructed from Eq.
(8) as an |Y| × |Y| matrix, where each element is defined
as:
G(y 0 , y) = `(y 0 , y) + θ> {φ(y, x) − φ(y ∗ , x)},
(10)
where is the first term `(·, ·) is the IoU based loss and the
second term is the weighted difference between FC7 features of the annotation distribution label y and the ground
Algorithm 1 ADA Equilibrium Computation
Figure 3. Example Game Matrix for a duck image with three
bounding boxes. Each black bounding box is a potential label for
the same duck image.
truth y ∗ label. To better understand the structure of the game
GΦ . The elements
matrix, we can decompose it as G = G` +G
of each matrix are illustrated for a toy example in Figure 3.
The first matrix G ` contains the pairwise loss between the
label of the predictor and the label of the adversary, `(y 0 , y).
The second matrix G Φ contains the difference in feature
statistics between the adversarial label and the ground truth
label, θ> {φ(y, x) − φ(y ∗ , x)}. Since the constraint matrix
G Φ does not depend on the predictor label y 0 , each row is
identical. The elements of the last column are all −1 in this
example because the feature statistic of the bounding box y3
are very different from the ground truth bounding box y ∗ ,
whereas the first two columns are zero because the content
of their bounding boxes are similar.
4.2. Nash Equilibria
The solution of the game, the Nash equilibrium pair
(f , p), is defined as the optimal strategy for each player such
that:
>
f 0 G p.
max
f >G p0 ≤ f >G p ≤ min
0
0
p
f
(11)
For the example in Figure 3, the lower potential for the
third bounding box in G Φ offsets the larger loss that might
be produced in G ` by having p(y3 ) > 0. Due to the
symmetries in G , the equilibrium solution is then simply:
P (y1 |x) = P (y2 |x) = f (y10 |x) = f (y20 |x) = 0.5 and
P (y3 |x) = f (y3 |x) = 0.
In general, this equilibrium solution pair can be obtained
efficiently using a pair of linear programs:
min v such that: f >G ≤ v1 and f > 1 = 1; and
v,f ≥0
(12)
max v such that: G p ≥ v1> and p> 1 = 1,
v,p≥0
where v is the value of the game (i.e., the expected loss).
This first linear program finds f that produces the maximum
value against the worst choice of p0 using the left-hand side
of Eq. (11) via constraints for each deterministic choice of
p0 (i.e., the 1 vector). The second linear program is constructed in a likewise manner to obtain p.
Input: Image x; Parameters θ; Ground Truth y ∗
Output: Nash equilibrium, (f , p)
1: Y ← EdgeBox(x)
2: Φ = CNN(Y, x)
3: ψ ← θ > (Φ − Φ(y ∗ ))
4: Sp ← Sf ← argmaxy ψ(y)
5: repeat
6:
(f , p, vp ) ← solveGame(ψ(Sp ), loss(Sf , Sp ))
7:
(ynew , vmax ) ← maxy Ey0 ∼f [loss(y, y 0 )+ ψ(y)]
8:
if (vp 6= vmax ) then
9:
Sp ← Sp ∪ ynew
10:
end if
11:
(f , p, vf ) ← solveGame(ψ(Sp ), loss(Sf , Sp ))
12:
(y 0 new , vmin ) ← minŷ Ey∼p [loss(y, y 0 )]
13:
if (vf 6= vmin ) then
14:
Sf ← Sf ∪ y 0 new
15:
end if
16: until vp = vmax = vf = vmin
17: return (f , p)
4.3. Constraint Generation for Large Games
In practice, forming and solving a zero-sum adversarial
game over a very large set of labels (e.g., the set of all possible bounding boxes in an image) for each image is computationally expensive. To obtain the same result more efficiently, we employ a constraint-generation method [17, 32]
to solve ADA-D without explicitly constructing the entire
payoff matrix G . It is based on the key insight that the
equilibrium distributions, f and p, both assign zero probabilities to many bounding boxes and those bounding boxes
can then be effectively removed from the game matrix without changing the solution. The basic strategy of constraintgeneration is to use a set of the most violated constraints to
grow a game matrix that supports the equilibrium distribution, but is much smaller than the full game matrix.
The approach works by iteratively obtaining a Nash
equilibrium for a game defined over a subset of the possible labels (not all of them), finding a player’s best response
strategy (either the predictor or the annotation distribution)
to that equilibrium distribution. Then the best response to
the set of opponent strategies defining the game is added as
a new strategy. When additional best responses no longer
improve either player’s game value, the subgame equilibrium is guaranteed to be an equilibrium to the larger game
[17].
4.4. Algorithm Details
Algorithm 1 details the ADA equilibrium computation
structure. The pre-processing step, extracting the image box
proposals (EdgeBox) and their CNN features, is addressed
in Lines 1-4. The CNN features are extracted from the last
convolutional layer of the deep networks in respect with the
related experiments. Sp and Sf stand for strategies(chosen
box proposals) of p and f game players. The algorithm
presents the main structure in lines 5-16. solveGame is
the process of obtaining a Nash equilibrium using linear
programming. Gurobi LP solver [10] is used to solve the
linear programming. The constraint generation is presented
with max and min operations in lines 7,12. After reaching
the loop termination condition (line 16), the f and p distributions are returned.
Model parameters θ are obtained using convex optimization (gradient-based methods [4]) that reach convergence
exactly when the data augmentations match the features
of the training data: Ex ∼ D, [φ(y, x)] = Ey,x∼D [φ(y, x)].
y|x ∼ P
The gradient is simply the difference between these two expectations.
At testing time, we employ a similar inference procedure. It likewise uses constraint generation—in the form
of best responses to equilibria using sets of strategies—to
find a set of strategies supporting the equilibrium of the
full game. A final prediction is produced by taking the most
probable predictor strategy under the equilibrium distribution, ŷ = argmaxy P (y|x), as the predicted bounding box.
5. Experiments
In the following experiments, our aim is to show that
our proposed adversarial optimization for data augmentation will provide meaningful improvements in test time prediction performance. We first compare our adversarial data
augmentation (ADA) objective function against baselines
models with no data augmentation on the task of localization in Section 5.1. Second, we compare our approach to
baseline models with varying levels of data augmentation
in Section 5.2. Third, we evaluate our approach on the task
of detection (joint recognition and localization) in Section
5.3. Finally, we use our adversarial optimization over various deep features to show consistent improvements across
networks in Section 5.4.
Baselines. To analyze the performance of our adversarial
objective function (ADA) for object detection, we benchmark it against two classical objective functions.
(1) SSVM: The structured output support vector machine
(SSVM)[31] is a large margin classifier with a variable margin depending on a structured loss function `. The objective
function is defined as:
X
θ̂ = argmin λ||θ||2 +
ξn
(13)
θ
s.t.
n
>
θ (φ(yn∗ , xn ) − φ(y, xn )) ≥ `(yn∗ , y) − ξn ∀ y
where θ is the weight vector, φ is the feature function (image feature statistic), ` is the loss function and ξ is the slack
variable. To solve the SSVM objective function, we employ an iterative constraint generation strategy to accelerate
Table 1. No augmentation baseline comparison (IoU>0.5)
Model
Plane
ADA+VGG (Ours) 92.0
Softmax+VGG
84.0
SSVM+VGG
90.0
Bird
93.5
86.5
82.5
ImageNet Object Categories
Bus Car Cat Cow Dog Hors
92.0 100.0 89.1 100.0 93.0 96.4
84.0 87.0 70.9 77.5 62.0 72.7
82.0 82.0 40.0 87.5 72.0 72.7
Moni
96.0
72.0
90.0
Sofa mAP
90.0 94.2
80.0 77.7
78.0 77.7
Table 2. No augmentation baseline comparison (IoU>0.7)
Model
Plane
ADA+VGG (Ours) 58.0
Softmax+VGG
47.6
SSVM+VGG
51.8
Bird
61.5
45.7
55.5
ImageNet Object Categories
Bus Car Cat Cow Dog Hors
64.0 91.0 30.9 77.4 58.0 58.2
40.0 62.8 20.0 42.5 25.1 25.4
44.0 61.7 21.8 54.7 31.6 43.6
Moni
61.8
31.4
56.0
Sofa mAP
61.9 62.3
44.2 38.5
57.3 47.8
the learning process by adding a few constraints per iteration (instead of the entire constraint set defined by each
label y ∈ Y). At test time, we generate a set of bounding box proposals (EdgeBox) and take the bounding box
with the highest potential using the learned weight vector,
ŷ = argmaxy θ> φ(y, x) to identify the predicted structured output.
(2) Softmax: The soft maximum (logistic regression) objective function is a probabilistic predictor. For the softmax objective function, we estimate a distribution over all
proposed bounding boxes y that maximizes the conditional
likelihood of proposed bounding boxes with an IoU above
a given threshold.
Y
θ̂ = argmax
P (yn |xn ; θ),
(14)
θ
n
Y eθ> φ(yn ,xn )
P θ> φ(y,x )
= argmax
n
θ
ye
n
(15)
where θ is the weight vector, φ is the potential function (FC7
feature) and ` is the loss function. At test time, we compute
the Bayesian optimal decision to identify the most likely
bounding box from a set of proposed bounding boxes according to:
X
ŷ = argmin
P (y 0 |x; θ)`(y, y 0 ),
(16)
y
y 0 ∈Y
where P (y|x; θ) is the learned conditional distribution parameterized by θ and ` is the loss function.
5.1. Baseline Comparisons with No Augmentation
We begin with the simplest evaluation, where we compare our proposed adversarial data augmentation approach
with two baseline method that use only the ground truth
annotation, without augmenting the training data, to learn
a predictor. We compare our method ADA+VGG against
SSVM+VGG and Softmax+VGG. The suffix +VGG for
each objective function specifies the deep network from
which the features are used, in this case VGG16 [24]. We
compute the mean Average Precision (mAP) score for several classes in ImageNet dataset for each of the competing
methods. We train a bounding box predictor for each object
category, and consider an object to be correctly detected
when the IoU is greater than a threshold. We emphasize
here that we are decoupling the recognition task from the
localization task by learning class specific bounding box regressor and testing only on images that contain the target
class. Later experiments will evaluate on both recognition
and localization. For this experiment, we give results for
two thresholds, 50% IoU and 70% IoU for each object category. Since our method explicitly augments the dataset as
part of the optimization process whereas the two baselines
have no data augmentation, we expect our approach will
outperform the two baselines.
The test time localization accuracy at 50% IoU on object 10 classes from ImageNet are given in Table 1. As expected, we observed significant test time improvement in
bounding box regression accuracy for every object category.
On average, our proposed adversarial data augmentation approach improved mean average precision by 17% percentage points.
We repeated the same experiment for a more strict loss,
a 0.7 thresholded IOU loss function. The mean average precision of the predicted bounding boxes are given in Table 2.
Since we are evaluating performance with a more strict loss
function, the absolute mAP values decrease as expected.
However, notice that our proposed approach still obtains a
significant improvement over the baseline algorithms improving mAP by 15% percentage points over the strongest
baseline SSVM+VGG.
5.2. Baseline Comparisons with Augmentation
We now compare the performance of our approach to
the strongest baseline model, SSVM+VGG, trained with
different levels of data augmentation. As mention earlier,
data augmentation such as random translations of bounding
boxes, is a common heuristic used to help supervised learning methods avoid overfitting. We prepare five levels of data
augmentation to train the SSVM+VGG baseline. Instead of
using random translations within a range of the ground truth
bounding box annotation, which can contain many similar
bounding boxes, we use the EdgeBox proposal network to
generate a diverse set of bounding boxes. We keep the top
250 EdgeBox proposals with the highest scores and filter
them according to 5 thresholds with respect to the original
ground truth bounding box: (1) IoU>50%; (2) IoU>60%;
(3) IoU>70%; (4) IoU>75%; and (5) IoU>80%. We denote the experiment using the subscript t50 to represent
a model trained on a collection of bounding boxes with
IoU>50%. We consider the bounding boxes that pass the
threshold test, as new ‘ground truth’ and use them as the
training set. We note here again that our proposed method
automatically selects (gives weights to) the bounding box
proposals during the learning process and does not require
a separate augmentation step.
Table 3. Effect of Data Augmentation (IoU > 70%)
Augmentation
Plane
SSVMt50 +VGG
53.8
SSVMt60 +VGG
54.7
SSVMt70 +VGG
56.4
SSVMt75 +VGG
52.6
SSVMt80 +VGG
49.8
ADA+VGG (Ours) 58.0
Bird
57.9
58.9
61.6
61.0
52.0
61.5
AlexNet Object Category
Bus Car Cat Cow Dog Hors
49.7 64.0 22.6 59.9 37.5 45.5
52.7 67.7 23.7 64.9 42.0 48.6
56.8 70.8 25.4 67.3 49.1 51.9
51.7 64.4 20.2 61.2 42.6 44.0
44.9 60.3 20.2 55.8 33.1 41.4
64.0 91.0 30.9 77.4 58.0 58.2
Moni
56.7
57.3
58.6
57.3
55.8
61.8
Sofa
57.8
58.4
58.8
56.0
52.7
61.9
Avg
50.5
52.9
55.7
51.1
46.6
62.3
Table 4. Effect of Number of Augmented Data Annotations. ADA
outperforms best configuration SSVM+VGG baseline by 12%.
SSVM+VGG k=1 k=2 k=4 k=6 k=8 k=10 k=12 ADA+VGG
mAP
77.6 79.7 81.4 83.8 83.7 79.8 75.3
94.2
Table 5. Detection Performance Comparison (IoU > 70%).
Model
Plane
ADA+VGG (Ours) 46.0
SSVM+VGG
42.0
Softmax+VGG
40.0
Bird
55.5
46.0
42.5
Image Net Object Category
Bus Car Cat Cow Dog Hors
60.0 86.0 25.4 70.0 47.0 52.7
38.0 53.0 16.4 52.5 25.0 36.4
42.0 55.0 16.4 32.5 16.0 29.1
Moni
60.0
42.0
22.0
Sofa
48.0
42.0
34.0
Avg
55.1
39.3
33.0
The results of this experiment are shown in Table 3. We
note that augmenting SSVM in this manner improves test
performance compared to the model without data augmentation (Table 2). We also observe that the performance gain
is maximized around the 70% overlap threshold. However,
we find that with the exception of the Bird class, the performance gains of data augmentation do not reach the performance of our ADA approach. Our proposed approach still
outperforms the model with the best level data augmentation by 6.4% percentage points.
We also performed a second data augmentation experiments by only varying the number of bounding boxes with
top the EdgeBox scores (instead of using IoU) for every image label to understand how the amount of augmented data
affects test time performance. At test time, we use the 50%
overlap criteria for successful localization. Table 4 shows
the mAP performance over same 10 object categories in
ImageNet as a function of the number of augmented data
annotations per image. The augmented data annotations are
selected from a rank list of EdgeBox proposals from each
image. The performance of the SSVM+VGG tops out at 8
augmented data annotations and is still 12% points below
our propose approach (94.2% mAP).
5.3. Detection Performance Comparison
We now address the object detection task of jointly locating and recognizing the category an unknown object, to
evaluate the performance of our approach on a harder task.
In order for a model to obtain a correct result, the predictor
must output the correct category label and also generate a
bounding box that overlaps with the ground truth by at least
70% IoU. For our baseline models, SSVM+VGG and Softmax+VGG, we use the best performing data augmentation
scheme from Table 3 that includes EdgeBox proposals that
have 70% IoU threshold with the original ground truth annotation.
Adversarial Boxes (proposed)
Top-K EdgeBoxes by score
Top-K EdgeBoxes by IoU
Figure 4. The top five augmented bounding boxes (left to right) under the different augmentation strategies. Our adversarial augmentation
approach generates much more diverse bounding boxes to augment the training data than choosing solely based on IoU yet the boxes are
at the same time still relevant to the target object unlike general EdgeBoxes.
Table 6. ADA Generalization Across Deep Architectures. VOC2007 mAP for IoU>0.5.
Model
Aero
ADA+VGG16
68.5
SSVM+VGG16
73.6
SVM+VGG16 [5]
73.4
ADA+AlexNet fc7
62.4
SSVM+AlexNet fc7
68.2
SVM+AlexNet fc7 [5] 68.1
ADA+ResNet101
76.4
SSVM+ResNet101
68.0
Bike
71.5
76.4
77.0
70.0
72.9
72.8
74.8
70.2
Bird
67.8
63.7
63.4
63.6
57.3
56.8
72.4
69.3
Boat
63.3
46.1
45.4
63.0
44.2
43.0
64.0
54.3
Bott
48.6
44.0
44.6
44.8
41.8
36.8
52.5
46.5
Bus
76.5
76.0
75.1
72.2
66.0
66.3
84.0
76.2
Car
78.8
78.4
78.1
75.5
74.3
74.2
81.9
78.8
Cat
80.9
80.0
79.8
79.5
69.2
67.6
86.0
85.0
VOC 2007 Object Category
Chair Cow DinT Dog Horse mbike person Plant Sheep Sofa Train TV mAP
50.9 78.5 64.5 79.6 71.8 73.2 66.4 30.2 70.6 72.6 80.8 62.8 67.9
41.6 74.2 62.8 79.8 78.0 72.5 64.3 35.0 67.2 67.2 70.8 71.4 66.1
40.5 73.7 62.2 79.4 78.1 73.1 64.2 35.6 66.8 67.2 70.4 71.1 66.0
44.6 81.6 64.0 81.5 70.2 68.5 71.4 69.5 65.0 71.2 81.4 59.8 68.0
34.6 54.7 54.3 61.3 69.8 68.7 58.5 34.6 63.6 52.5 62.6 63.5 58.6
34.4 63.5 54.5 61.2 69.1 68.6 58.7 33.4 62.9 51.1 62.5 64.8 58.5
48.5 83.5 64.8 82.0 73.5 77.0 72.4 36.6 74.4 74.8 81.4 65.6 71.3
46.8 80.2 63.2 78.1 69.5 71.4 61.8 36.8 68.1 69.1 73.6 64.5 66.5
Table 5 shows the object detection performance when
evaluated at the 70% IoU threshold for correctness. We
again find strong support for our adversarial approach to
dealing with uncertainty. Specifically, we find that ADA70
provides the best performance for all object classes. Though
the relative performance advantage differs by object type,
for classes like Dog, the improvement over the other approaches is nearly double. On average, ADA provides a
significant performance improvement of 15.8% percentage points on this task over the strongest performing
SSVM+VGG baseline.
proccessing step of theses experiments, the box proposals
of every image are extracted using EdgeBox. These box
proposals are employed as input to the respective deep networks. The CNN features are extracted from the last convolutional layer of the respective deep network and these
box proposals and their CNN features are employed in
training the respective classifier. We report the generalized
version of ADA and SSVM with Alexnet[14], VGG, and
ResNet101[12]. For the VOC 2007 dataset, we use 5000
training images and 4952 testing images.
5.4. Generalization Across Deep Features
6. Conclusions
To understand how our proposed adversarial loss function, ADA, generalizes across various deep architectures,
we compare mAP over 20 categories from the VOC2007
dataset for several combinations of deep architecture using our proposed adversarial loss function. In the pre-
In this paper, we have developed a game-theoretic formulation for data augmentation that perturbs image annotations adversarially. This provides robustness in the learned
predictor that is achieved by training from a number of
augmentations that are adaptively selected to be difficult,
while still approximating the ground truth annotation. We
demonstrated the benefits for object localization and detection using experiments over ten different object classes
for the ILSVRC2012 dataset and twenty different object
classes for the VOC2007 dataset, showing significant improvements for our approach under 50% and 70% thresholded IoU evaluation measures.
References
[1] K. Asif, W. Xing, S. Behpour, and B. D. Ziebart. Adversarial
cost-sensitive classification. In Proceedings of the Conference on UAI, 2015. 2
[2] S. Boyd and L. Vandenberghe. Convex optimization. Cambridge university press, 2004. 4
[3] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman.
Return of the devil in the details: Delving deep into convolutional nets. arXiv preprint arXiv:1405.3531, 2014. 2
[4] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic optimization.
Journal of Machine Learning Research, 12(Jul):2121–2159,
2011. 6
[5] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation tr. In arXiv:1311.2524, 2014. 8
[6] R. Gokberk Cinbis, J. Verbeek, and C. Schmid. Multi-fold
mil training for weakly supervised object localization. In
Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 2409–2416, 2014. 2
[7] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems, pages 2672–2680, 2014. 2
[8] P. D. Grünwald and A. P. Dawid. Game theory, maximum entropy, minimum discrepancy, and robust Bayesian decision
theory. Annals of Statistics, 32:1367–1433, 2004. 2
[9] C. Guo, M. Rana, M. Cisse, and L. van der Maaten. Countering adversarial images using input transformations. arXiv
preprint arXiv:1711.00117, 2017. 2
[10] I. Gurobi Optimization. Gurobi optimizer reference manual,
2016. 5
[11] K. He, X. Zhang, S. Ren, and J. Sun. Spatial pyramid pooling
in deep convolutional networks for visual recognition. IEEE
transactions on pattern analysis and machine intelligence,
37(9):1904–1916, 2015. 2
[12] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
770–778, 2016. 8
[13] S. Joon Oh, M. Fritz, and B. Schiele. Adversarial image perturbation for privacy protection–a game theory perspective.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 1482–1491, 2017. 2
[14] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in neural information processing systems, pages
1097–1105, 2012. 2, 8
[15] P. D. Larry Zitnick. Edge boxes: Locating object proposals
from edges. In ECCV. European Conference on Computer
Vision, September 2014. 4
[16] I. Masi, A. T. Trn, T. Hassner, J. T. Leksut, and G. Medioni.
Do we really need to collect millions of faces for effective
face recognition? In European Conference on Computer Vision, pages 579–596. Springer, 2016. 2
[17] H. B. McMahan, G. J. Gordon, and A. Blum. Planning in
the presence of cost functions controlled by an adversary.
In Proceedings of the International Conference on Machine
Learning, pages 536–543, 2003. 5
[18] S.-M. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and
P. Frossard. Universal adversarial perturbations. arXiv
preprint arXiv:1610.08401, 2016. 2
[19] S. Nowak and S. Rüger. How reliable are annotations via
crowdsourcing: a study about inter-annotator agreement for
multi-label image annotation. In Proceedings of the international conference on Multimedia information retrieval, pages
557–566. ACM, 2010. 1, 2
[20] M. Paulin, J. Revaud, Z. Harchaoui, F. Perronnin, and
C. Schmid. Transformation pursuit for image classification.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 3646–3653, 2014. 2
[21] S. Reed, H. Lee, D. Anguelov, C. Szegedy, D. Erhan, and
A. Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596,
2014. 2
[22] C. Sagonas, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic.
A semi-automatic methodology for facial landmark annotation. In The IEEE Conference on CVPR Workshops, June
2013. 1, 2
[23] J. M. Saragih, S. Lucey, and J. F. Cohn. Deformable model
fitting by regularized landmark mean-shift. International
Journal of Computer Vision, 91(2):200–215, 2011. 1, 2
[24] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014. 4, 6
[25] H. Su, J. Deng, and L. Fei-Fei. Crowdsourcing annotations
for visual object detection. In Workshops at the Twenty-Sixth
AAAI Conference on Artificial Intelligence, 2012. 2
[26] S. Sukhbaatar and R. Fergus. Learning from noisy labels
with deep neural networks. arXiv preprint arXiv:1406.2080,
2(3):4, 2014. 2
[27] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4,
inception-resnet and the impact of residual connections on
learning. arXiv preprint arXiv:1602.07261, 2016. 1, 2
[28] L. Taylor and G. Nitschke.
Improving deep learning using generic data augmentation.
arXiv preprint
arXiv:1708.06020, 2017. 2
[29] F. Tramèr, A. Kurakin, N. Papernot, D. Boneh, and P. McDaniel. Ensemble adversarial training: Attacks and defenses.
arXiv preprint arXiv:1705.07204, 2017. 2
[30] P. Upchurch, D. Sedra, A. Mullen, H. Hirsh, and K. Bala.
Interactive consensus agreement games for labeling images.
In Fourth AAAI Conference on Human Computation and
Crowdsourcing, 2016. 1, 2
[31] A. Vedaldi.
A MATLAB wrapper of SVMstruct .
http://www.vlfeat.org/˜vedaldi/code/
svm-struct-matlab, 2011. 6
[32] H. Wang, W. Xing, K. Asif, and B. Ziebart. Adversarial prediction games for multivariate losses. In Advances in Neural
Information Processing Systems, pages 2710–2718, 2015. 2,
5
[33] P. Welinder, S. Branson, P. Perona, and S. J. Belongie. The
multidimensional wisdom of crowds. In Advances in neural
information processing systems, pages 2424–2432, 2010. 1,
2
[34] T. Xiao, T. Xia, Y. Yang, C. Huang, and X. Wang. Learning from massive noisy labeled data for image classification.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 2691–2699, 2015. 2
[35] X. Zhu and D. Ramanan. Face detection, pose estimation,
and landmark localization in the wild. In CVPR, 2012 IEEE
Conference on, pages 2879–2886. IEEE, 2012. 1, 2
| 2cs.AI
|
arXiv:1703.07458v1 [math.GR] 21 Mar 2017
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
G. CHRISTOPHER HRUSKA AND HOANG THANH NGUYEN
Abstract. Let S # N be an immersed horizontal surface in a 3–
dimensional graph manifold. We show that the fundamental group of
the surface S is quadratically distorted whenever the surface is virtually embedded (i.e., separable) and is exponentially distorted when the
surface is not virtually embedded.
1. Introduction
In the study of 3–manifolds, much attention has focused on the distinction
between surfaces that lift to an embedding in a finite cover and those that
do not. A π1 –injective immersion S # N of a surface S into a 3–manifold
N is virtually embedded if (after applying a homotopy) the immersion lifts
to an embedding of S into a finite cover of N .
A major part of the solution of the virtual Haken conjecture is Agol’s theorem that every immersed surface in a closed hyperbolic 3–manifold is virtually embedded [Ago13]. In contrast, Rubinstein–Wang [RW98] constructed
non–virtually embedded surfaces g : S # N in many 3–dimensional graph
manifolds N . The examples of Rubinstein–Wang are horizontal in the sense
that in each Seifert component M of N , the intersection g(S) ∩ M is transverse to the Seifert fibration. Furthermore, they introduced a combinatorial
invariant of horizontal surfaces called the dilation and proved that dilation
is the obstruction to a surface being virtually embedded.
If G and H are groups with finite generating sets S and T , the distortion
of H in G is given by
∆G
h ∈ H and |h|S ≤ n .
H (n) = max |h|T
Distortion does not depend on the choice of finite generating sets S and
T (up to a natural equivalence relation). The purpose of this article is to
address the following question of Dani Wise:
Question 1.1. Given a 3–dimensional graph manifold N , which surfaces in
N have nontrivial distortion?
Question 1.1 arises naturally in the study of cubulations of 3–manifold
groups. The typical strategy for constructing an action of the fundamental
group on a CAT(0) cube complex is to find a suitable collection of immersed
Date: March 23, 2017.
2010 Mathematics Subject Classification. 57M50, 20F65, 20F67.
Key words and phrases. graph manifold, separable subgroup, distortion.
1
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
2
surfaces and then to consider the CAT(0) cube complex dual to that collection of surfaces (see for instance [Wis12]).
Whenever a group G acts properly and cocompactly on a CAT(0) cube
complex X, the stabilizer of each hyperplane must be an undistorted subgroup of G, since hyperplanes are convex. Hagen–Przytycki [HP15] show
that chargeless graph manifolds do act cocompactly on CAT(0) cube complexes. It is clear from their construction that many surfaces in graph manifolds are undistorted.
However the situation for horizontal surfaces turns out to be quite different. Our main result states that horizontal surfaces in graph manifolds
always have a nontrivial distortion, and this distortion is directly related to
virtual embedding in the following sense:
Theorem 1.2. Let S # N be a horizontal surface immersed in a graph
manifold N . The distortion of π1 (S) in π1 (N ) is quadratic if S is virtually
embedded, and exponential if S is not virtually embedded.
The main tool used in the proof of this theorem is a simple geometric
interpretation of Rubinstein–Wang’s dilation in terms of slopes of lines in
the JSJ planes of the universal cover.
We note that, throughout this paper, the term “graph manifold” specifically excludes two trivial cases: Seifert manifolds and Sol manifolds. By
Proposition 6.7 horizontal surfaces in Seifert manifolds are always undistorted. Distortion of surfaces in Sol manifolds is not addressed in this article.
Although Theorem 1.2 deals only with horizontal surfaces, understanding these is the key to a general understanding of all π1 –injective surfaces
in graph manifolds in the following sense. It was shown by Rubinstein–
Wang that in a graph manifold, every π1 –injective, properly immersed, orientable surface can be homotoped to a properly immersed surface g′ : S # N
such that for each Seifert component M each component of the intersection
g′ (S) ∩ M is either vertical (i.e., a union of fibers) or horizontal [RW98].
While vertical surfaces in Seifert components are easy to understand, most
complications appear to arise from the horizontal part of an immersed surface.
We note the following corollary about surfaces in 3–manifolds of mixed
type, i.e., those whose geometric decomposition contains at least one JSJ
torus and at least one hyperbolic component. (For clarity, we remind the
reader of our convention that Seifert manifolds are not considered to be
“graph manifolds.”)
Corollary 1.3. Let g : S # M be a π1 –injective immersed surface in a
mixed 3–manifold M . If π1 (S) is undistorted in π1 (M ), then the pieces of
S have the following simple form.
(1) For each hyperbolic block N of M , the components of g(S) ∩ N are
all geometrically finite.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
3
(2) For each maximal graph manifold block (or isolated Seifert component) N of M , the components of g(S)∩N correspond to undistorted
subgroups of π1 (N ).
In particular, the intersection of g(S) with a graph manifold component N
cannot be a horizontal surface.
Indeed, by [KL95, DS05] the group π1 (M ) is relatively hyperbolic with
respect to the fundmental groups of maximal graph manifold components,
isolated Seifert components, and isolated JSJ tori (see [BW13] for an elementary proof of this fact). Corollary 1.3 can now be deduced from this
relative hyperbolicity using results of [Hru10] as follows. The undistorted
subgroup π1 (S) is relatively quasiconvex. Since the hyperbolic blocks of M
are also relatively quasiconvex in M and the intersection of two relatively
quasiconvex subgroups is relatively quasiconvex [Hru10], we obtain assertion (1). On the other hand, assertion (2) follows immediately from [Hru10,
Theorem 1.4].
Hagen–Przytycki [HP15] and Tidmore [Tid] have shown that chargeless
graph manifolds and chargeless mixed manifolds act cocompactly on CAT(0)
cube complexes. The cubulations they construct are each dual to a family
of immersed surfaces, none of which is entirely horizontal. The following
corollary of Theorem 1.2 shows that in order to obtain a proper, cocompact
cubulation one is required to avoid horizontal surfaces. Thus, in a sense,
the cocompact cubulations of Hagen–Przytycki are canonical, and could not
have been constructed using different surfaces.
Corollary 1.4. Let G be the fundamental group of a graph manifold. Let
{H1 , . . . , Hk } be a collection of codimension–1 subgroups of G. Let X be
the corresponding dual CAT(0) cube complex. If at least one Hi is the fundamental group of a horizontal surface, then the action of G on X is not
proper and cocompact.
An analogous result could also be obtained in the setting of mixed manifolds using Corollary 1.3.
After seeing an early version of this paper, Hung Cong Tran discovered
an alternate proof of the quadratic distortion of certain horizontal surfaces
whose fundamental groups can also be regarded as Bestvina–Brady kernels.
This observation is part of a broader study by Tran of the distortion of
Bestvina–Brady kernels in right-angled Artin groups [Tra].
1.1. Connections to previous work. In [Woo16], Woodhouse exploited
strong parallels between graph manifolds and tubular spaces to study actions of tubular groups on CAT(0) cube complexes. In the tubular setting,
immersed hyperplanes play the role of immersed surfaces in graph manifolds
Woodhouse extended the Rubinstein–Wang theory of dilation and found a
connection between dilation and distortion. In particular, he proves that
if an immersed hyperplane in a tubular group has nontrivial dilation then
its distortion is at least quadratic. Inspired by Woodhouse’s work, we use
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
4
analogous techniques in the cleaner geometric setting of graph manifolds to
obtain Theorem 1.2.
We remark that the main proof of Theorem 1.2 in Sections 5 and 6 includes both virtually embedded and non–virtually embedded surfaces in a
unified treatment that prominently uses a simple geometric interpretation
of dilation in terms of “slopes of lines” in the Euclidean geometry of JSJ
planes. This interpretation was inspired by Woodhouse’s earlier work on
tubular groups. It seems likely that the main proofs here could be translated back to the tubular setting, where they may lead to new advances in
that setting as well.
However there is an alternate (shorter) proof of the quadratic distortion
of virtually embedded horizontal surfaces using the following result, which
combines work of Gersten and Kapovich–Leeb. This alternate approach,
while more direct, does not give any information about the structure of
non–virtually embedded surfaces.
Theorem 1.5 (Gersten, Kapovich–Leeb). Let N be a graph manifold that
fibers over the circle with fiber surface S. Then π1 (S) is quadratically distorted in π1 (N ).
Kapovich–Leeb implicitly use the quadratic upper bound (without stating
it explicitly) in [KL95], where they attribute it to Gersten [Ger94]. The quadratic lower bound also follows easily from results of Gersten and Kapovich–
Leeb [Ger94, KL95], but was not specifically mentioned by them. The discussion of Theorem 1.5 in [KL95] and its precise derivation from [Ger94] is
brief and was not the main purpose of either article. For the benefit of the
reader we have included a more detailed exposition of Kapovich–Leeb’s elegant proof of Theorem 1.5 in Section 7, which relies on Thurston’s geometric
classification of 3–manifolds that fiber over the circle.
The virtually embedded case of Theorem 1.2 can be derived as a corollary
of Theorem 1.5 as follows. If a horizontal surface g : S # N is virtually
embedded then there exist finite covers Ŝ → S and N̂ → N such that N̂
is an Ŝ–bundle over S 1 (see for instance [WY97]). By Theorem 1.5, the
distortion of π1 (Ŝ) in π1 (N̂ ) is quadratic. Distortion is unchanged when
passing to subgroups of finite index, so the distortion of π1 (S) in π1 (N ) is
also quadratic.
1.2. Overview. In Section 2 we review some concepts in geometric group
theory. In Section 3 we give several lemmas about curves on hyperbolic
surfaces that will be used in Section 5. Section 4 is a review background
about graph manifolds and horizontal surfaces. A convenient metric on a
graph manifold that will use in this paper will be discussed.
In Section 5 we prove the distortion of a horizontal surface in a graph
manifold is at least quadratic. We also show that if the horizontal surface
is not virtually embedded then the distortion is at least exponential. The
strategy in this proof was inspired by the work of Woodhouse (see Section 6
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
5
in [Woo16]). In Section 6 we prove that distortion of a horizontal surface
in a graph manifold is at most exponential. Furthermore, if the horizontal
surface is virtually embedded then the distortion is at most quadratic.
Section 7 contains a detailed exposition of the proof of Theorem 1.5.
1.3. Acknowledgments. The authors benefited from conversations with
Daniel Woodhouse about the connection between distortion and dilation
in the setting of tubular groups. The authors are grateful to Hung Cong
Tran for suggesting the possibility of a connection between the virtually
embedded case of Theorem 1.2 and the earlier work of Gersten on quadratic
divergence.
The first author wishes to thank Noel Brady for many conversations about
distortion of subgroups and divergence of 3–manifold groups. In particular
the present article would not exist without the influence of Noel Brady’s
courses on geometric group theory when the first author was a graduate
student. The first author also thanks Dani Wise for many discussions about
distortion and cubulating graph manifolds, which motivated the work in this
paper, and Becca Winarski for discussions about curves and surfaces. The
second author would like to thank Joseph Tidmore for useful discussions
about graph manifolds, and Boris Okun for useful comments.
This work was partially supported by a grant from the Simons Foundation
(#318815 to G. Christopher Hruska).
2. Preliminaries
In this section, we review some concepts in geometric group theory: quasi–
isometry, distortion of a subgroup and the notions of domination and equivalence.
Definition 2.1. Let (X, d) be a metric space. A path γ : [a, b] → X is a
geodesic if d(γ(s), γ(t)) = |s − t| for all s, t ∈ [a, b]. A simple loop f : S 1 → X
is a geodesic loop if f is an isometric embedding with respect to some length
metric on S 1 .
Definition 2.2 (quasi-isometry). Let (X1 , d1 ) and (X2 , d2 ) be metric spaces.
A (not necessarily continuous) map f : X1 → X2 is an (L, C)–quasi-isometric
embedding if there exist constants L ≥ 1 and C ≥ 0 such that for all
x, y ∈ X1 we have
1
d1 (x, y) − C ≤ d2 f (x), f (y) ≤ L d1 (x, y) + C.
L
If, in addition, there exits a constant D ≥ 0 such that every point of X2 lies
in the D–neighborhood of the image of f , then f is an (L, C)–quasi-isometry.
When such a map exists, X1 and X2 are quasi-isometric.
Let (X, d) be a metric space, and γ a path in X. We denote the length
of γ by |γ|.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
6
Definition 2.3 (quasigeodesic). A path γ in a metric space (X, d) is an
(L, C)–quasigeodesic with respect to constants L ≥ 1 and C ≥ 0 if |γ[x,y] | ≤
L d(x, y) + C for all x, y ∈ γ. A quasigeodesic is a path that is (L, C)–
quasigeodesic for some L and C.
Definition 2.4. A geodesic space (X, d) is δ–hyperbolic if every geodesic
triangle with vertices in X is δ–thin in the sense that each side lies in the
δ–neighborhood of the union of the other two sides.
Definition 2.5 (deviation). Let (X, d) be a geodesic space and c ≥ 0 be a
fixed number. Consider a pair of geodesic segments [x, y] and [z, t] such that
d(y, z) ≤ c. The c–deviation of [x, y] and [z, t], denoted devc [x, y], [z, t] , is
the quantity
sup max{d(u, y), d(v, z)} u ∈ [x, y],v ∈ [z, t],d(u, v) ≤ c .
The following theorem gives a criterion for determining that a piecewise
geodesic in a δ–hyperbolic space is a quasigeodesic. The proof, which is very
similar to the proof of Lemma 19 of [GH90, Chapter 5], is left as an exercise
to the reader.
Theorem 2.6. Let (X, d) be a δ–hyperbolic space. For any κ ≥ 0 and
D ≥ 0, there exist constants L = L(δ, κ, D), and C = C(δ, κ, D) such that
the following holds: Suppose a piecewise geodesic
c = [x1 , x2 ] ∪ [x2 , x3 ] ∪ · · · ∪ [x2m−1 , x2m ]
satisfies
(1) d(x2i , x2i+1 ) ≤ κ
(2) d(x2i−1 , x2i ) ≥ 11κ + 25δ + 2D,
and
(3) devQ [x2i−1 , x2i ], [x2i+1 , x2i+2 ] ≤ D, where Q = 4δ + κ.
Then c is an (L, C)–quasigeodesic.
Definition 2.7. Let f, g be functions from positive reals to positive reals.
The function f is dominated by g, denoted f g, if there are positive
constants A, B, C, D and E such that
f (x) ≤ A g(Bx + C) + Dx + E
for all x.
The functions f and g are equivalent, denoted f ∼ g, if f g and g f .
The relation is an equivalence relation. Polynomial functions with
degree at least one are equivalent if and only if they have the same degree.
Furthermore, all exponential functions are equivalent.
Definition 2.8 (Subgroup distortion). Let H ≤ G be a pair of groups with
finite generating sets T and S respectively. The distortion of (H, T ) in
(G, S) is the function
∆G
h ∈ H and |h|S ≤ n
H (n) = max |h|T
Up to equivalence, the function ∆G
H does not depend on the choice of finite
generating sets S and A.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
7
Let H ≤ G be a pair of finitely generated groups. We say that the distortion ∆G
H is at least quadratic [exponential] if it dominates a quadratic
polynomial [exponential function]. The distortion ∆G
H is at most quadratic
[exponential] if it is dominated by a quadratic polynomial [exponential function].
The following proposition is routine, and we leave the proof as an exercise
for the reader.
Proposition 2.9. Let G, H, K be finitely generated groups with K ≤ H ≤
G.
G
(1) If H is a finite index subgroup of G then ∆H
K ∼ ∆K .
G
(2) If K is a finite index subgroup of H then ∆G
K ∼ ∆H .
It is well known that a group acting properly, cocompactly, and isometrically on a geodesic space is quasi-isometric to the space. The following
corollary of this fact allows us to compute distortion using the geometries
of spaces in place of word metrics.
Corollary 2.10. Let X and Y be compact geodesic spaces, and let g : (Y, y0 ) →
(X, x0 ) be π1 –injective. We lift the metrics on X and Y to geodesic metrics on the universal
covers X̃ and Ỹ respectively. Let G = π1 (X, x0 ) and
H = g∗ π1 (Y, y0 ) . Then the distortion ∆G
H is equivalent to the function
f (n) = max dỸ (ỹ0 , h(ỹ0 )) h ∈ H and dX̃ (x̃0 , h(x̃0 )) ≤ n .
3. Curves in hyperbolic surfaces
In this section, we will give some results about curves on surfaces that
will play an essential role in the proof of Theorem 5.1.
Definition 3.1 (Multicurves). A closed curve in a surface S is essential if it
is not freely homotopic to a point or a boundary component. A multicurve
in S is a finite collection of disjoint, essential simple closed curves such that
no two are freely homotopic. If S has a metric, a multicurve is geodesic if
each member of the family is a geodesic loop.
Lemma 3.2. Let S be a compact surface with negative Euler characteristic.
Let C be a multicurve in S. Let L be the family of lines that are lifts of
loops of C or boundary loops of S. Equip S with any length metric dS . Let
d be the induced metric on the universal cover S̃. For any r > 0 there exists
D = D(r) < ∞ such that for any two disjoint lines ℓ1 and ℓ2 of L we have
diam Nr (ℓ1 ) ∩ Nr (ℓ2 ) ≤ D
Proof. Since H = π1 (S) acts cocompactly on the universal cover S̃, there
exists a closed ball B(x0 , R) whose H–translates cover S̃. Note that L
is locally finite in the sense that only finitely many lines of L intersect the
closed ball B = B(x0 , R+2r). Since distinct lines of L are not parallel, there
exists a finite upper bound D = D(r) on the diameter of the intersection
Nr (ℓ) ∩ Nr (ℓ′ ) for all lines ℓ 6= ℓ′ ∈ L that intersect B.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
8
Consider ℓ1 6= ℓ2 ∈ L. If d(ℓ1 , ℓ2 ) ≥ 2r then their r–neighborhoods have
empty intersection, and the result is vacuously true. Thus it suffices to
assume that d(ℓ1 , ℓ2 ) < 2r. By cocompactness, there exists h ∈ H so that
h(ℓ1 ) intersects B(x0 , R). But then h(ℓ1 ) and h(ℓ2 ) both intersect B, so that
diam Nr (ℓ1 ) ∩ Nr (ℓ2 ) ≤ D,
as desired.
Lemma 3.3. Let S be a compact hyperbolic surface with totally geodesic
(possibly empty) boundary, and let C be a nonempty geodesic multicurve.
Then there exists a geodesic loop γ in S such that γ and C have nonempty
intersection.
Proof. Choose any loop c ∈ C. Cutting S along c produces a surface each of
whose components has negative Euler characteristic since each of its components is a union of blocks of S joined along circles. For any such simple
closed curve c in a compact surface S, we claim that there exists another
closed curve γ whose geometric intersection number with c is nonzero. We
leave the proof of this claim as an easy exercise for the reader (using, for
example, the techniques in Sections 1.2.4 and 1.3 of [FM12]). After applying
a homotopy, we can assume that γ is a geodesic loop.
Lemma 3.4. Let S be a compact hyperbolic surface with totally geodesic
(possibly empty) boundary, and let C be a nonempty geodesic multicurve.
Let L be the family of lines that are lifts of loops of C or boundary loops of
S. For any κ > 0, there exist numbers µ, L, and C such that the following
holds.
Consider a piecewise geodesic c = α1 β1 · · · αn βn in the universal cover S̃.
Suppose each segment αj is contained in a line of L, and each segment βj
meets the lines of L only at its endpoints. If |βj | ≤ κ and |αj | ≥ µ for all j,
then c is an (L, C)–quasigeodesic.
Proof. In order to show that the piecewise geodesic c in the δ–hyperbolic
space S̃ is uniformly quasigeodesic, we will show that there exists a constant
D such that whenever µ is sufficiently large, c satisfies the hypotheses of
Theorem 2.6. Thus c is (L, C)–quasigeodesic for constants L and C not
depending on c.
Let Q = 4δ + κ. Let D = D(Q) be the constant given by applying
Lemma 3.2 to the hyperbolic surface S and the multicurve C. Let L =
L(δ, κ, D) and C = C(δ, κ, D) be the constants given by Theorem 2.6. We
define µ = 11κ + 25δ + 2D. It follows that c satisfies Conditions (1) and (2)
of Theorem 2.6.
In order to to
verify Condition (3) of Theorem 2.6, we need to show that
devQ αj , αj+1 ≤ D. Let u ∈ αj and v ∈ αj+1 such that d(u, v) ≤ Q. Let
u′ andv ′ be the initial and terminal points of βj . We need to show that
d u, u′ ≤ D and d v, v ′ ≤ D. It is easy to see the elements u, v, u′ , and
v ′ belong to NQ αj ∩ NQ αj+1 because |βj | ≤ κ ≤ Q. We have αj+1 and
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
9
αj do not belong to the same line of L because any two distinct geodesics
in S̃ intersect at most one point. Hence
diam NQ (αj ) ∩ NQ (αj+1 ) ≤ D
by Lemma 3.2. It follows that d u, u′ ≤ D and d v, v ′ ≤ D. It follows
immediately from the definition of the deviation that devQ αj , αj+1 ≤
D.
4. Graph manifolds and horizontal surfaces
In this section, we review background about graph manifolds and horizontal surfaces. In addition, we discuss a convenient metric for a graph
manifold that will be used in next sections. We refer the reader to [RW98],
[BS04] and [KL98] for more details.
Definition 4.1. A graph manifold is a compact, irreducible, connected,
orientable 3-manifold N that can be decomposed along embedded incompressible tori T into finitely many Seifert manifolds. We specifically exclude
Sol and Seifert manifolds from the class of graph manifolds. Up to isotopy,
each graph manifold has a unique minimal collection of tori T as above
[JS79, Joh79]. This minimal collection is the JSJ decomposition of N , and
each torus of T is a JSJ torus.
Throughout this paper, a graph consists of a set V of vertices and a set of
E of edges, each edge being associated to an unordered pair of vertices by a
function ends : ends(e) = {v, v ′ } where v, v ′ ∈ V. In this case we call v and
v ′ the endpoints of the edge e and we also say v and v ′ are adjacent.
Definition 4.2. A simple graph manifold N is a graph manifold with the
following properties:
(1) Each Seifert component is a trivial circle bundle over an orientable
surface of genus at least 2.
(2) The intersection numbers of fibers of adjacent Seifert components
have absolute value 1.
Theorem 4.3 ([KL98]). Any graph manifold N has a finite cover N̂ that
is a simple graph manifold.
Definition 4.4. Let M be a Seifert manifold with boundary. A horizontal
surface in M is an immersion g : B # M where B is a compact surface with
boundary such that the image g(B) is transverse to the Seifert fibration.
We also require that g is properly immersed, ie, g(B) ∩ ∂M = g(∂B).
A horizontal surface in a graph manifold N is a properly immersed surface
g : S # N such that for each Seifert component M , the intersection g(S)∩M
is a horizontal surface in M . A horizontal surface g in a graph manifold N
is always π1 –injective and lifts to an embedding
of S in the cover of N
corresponding to the subgroup g∗ π1 (S) by [RW98]. Consequently, g also
lifts to an embedding S̃ → Ñ of universal covers.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
10
Definition 4.5. A horizontal surface g : S # N in a graph manifold N is
virtually embedded if g lifts to an embedding of S in some finite cover of N .
By a theorem
of Scott [Sco78], a horizontal surface is virtually embedded if
g∗ π1 (S) is a separable subgroup of π1 (N ), i.e., it is equal to an intersection
of finite index subgroups. Przytycki–Wise have shown that the converse
holds as well [PW14].
The following result about separability allows one to pass to finite covers
in the study of horizontal surfaces, as explained in Corollary 4.7.
Proposition 4.6 ([Sco78], Lemma 1.1). Let G0 be a finite index subgroup
of G. A subgroup H ≤ G is separable in G if and only if H ∩ G0 is separable
in G0 .
Corollary 4.7. Let q : (N̂ , x̂0 ) → (N, x0 ) be a finite covering of graph manifolds. Let g : (S, s0 ) # (N, x0 ) be a horizontal surface. Let p : (Ŝ, ŝ0 ) →
(S, s0 ) be the finite cover corresponding to the subgroup g∗−1 q∗ π1 (N̂ , x̂0 ).
Then g lifts to a horizontal surface ĝ : (Ŝ, ŝ0 ) → (N̂ , x̂0 ). Furthermore g
is a virtual embedding if and only if ĝ is a virtual embedding.
Definition 4.8. Suppose g : S # N is a horizontal surface in a graph
manifold N with JSJ decomposition T . Let Tg denote the collection of
components of g−1 (T ) in S. After applying a homotopy to g, we may
assume that the image g(c) of each curve c ∈ Tg is a multiple of a simple
closed curve on the corresponding JSJ torus. The connected components of
the splitting S|g −1 (T ) are the blocks B of S.
Remark 4.9. If g : S # N is a horizontal surface then Tg is always nonempty.
Indeed, g(S) has nonempty intersection with each JSJ torus of N because
a properly immersed horizontal surface in a connected graph manifold must
intersect every fiber of every Seifert component (see Lemma 6.3 for details).
In [RW98], Rubinstein–Wang introduced the dilation of a horizontal surface, and proved that dilation is the obstruction to a surface being virtually
embedded (see Theorem 4.11).
Definition 4.10 (Dilation). Let g : (S, s0 ) # (N, x0 ) be a horizontal surface
in a simple graph manifold N . Choose an orientation for the graph manifold
N , an orientation for the fiber of each Seifert component, and an orientation
for each curve c ∈ Tg .
The dilation of a horizontal surface S in N is a homomorphism w : π1 (S, s0 ) →
∗
Q+ defined as follows. Choose [γ] ∈ π1 (S, s0 ) such that γ is transverse to Tg .
In the trivial case that γ is disjoint from the curves of the collection Tg , we
set w(γ) = 1. Let us assume now that this intersection is nonempty. Then
Tg subdivides γ into a concatenation γ1 · · · γm with the following properties.
Each path γi starts on a circle ci ∈ Tg and ends on the circle ci+1 . The
image g(γi ) of this path in N lies in a Seifert component Mi . The image of
the circle g(ci ) in N lies in a JSJ torus Ti obtained by gluing a boundary
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
11
←
−
→
−
←
−
→
−
torus Ti of Mi−1 to a boundary torus Ti of Mi . Let fi and fi be fibers of
←
−
→
−
Mi−1 and Mi in the torus Ti . By Definition 4.2, the 1–cycles [ fi ] and [ fi ]
generate the integral homology group H1 (Ti ) ∼
= Z2 , so there exist integers
ai and bi such that
←
−
→
−
g(ci ) = ai [ fi ] + bi [ fi ] in H1 (Ti ).
Since the immersion is horizontal, these coefficients
ai and bi must be nonzero.
Q
b
/a
The dilation w(γ) is the rational number m
i
i . Note that w(γ) dei=1
pends only on the homotopy class of γ, since crossings of γ with a curve
c ∈ Tg in opposite directions contribute terms to the dilation that cancel
each other. For the rest of this paper we write wγ instead of w(γ).
The following result is a special case of [RW98, Theorem 2.3].
Theorem 4.11. A horizontal surface g : S # N in a simple graph manifold
is virtually embedded if and only if the dilation w is the trivial homomorphism.
Remark 4.12. Let g : S # N be a horizontal surface in a simple graph manifold M , then each block B is a connected surface with non-empty boundary
and negative Euler characteristic. Indeed, the immersion g : S # N maps
B to the corresponding Seifert component M with base surface F . The
composition of g|B with the projection of M to F yields a finite covering
map from B to F . Since χ(F ) < 0, it follows that χ(B) < 0 as well.
We note that the collection Tg is always a non-empty multicurve. Indeed,
Tg is nonempty by Remark 4.9. Since the blocks of S have negative Euler
characteristic, it follows that Tg is a multicurve.
Remark 4.13. We now are going to describe a convenient metric on a
simple graph manifold N introduced by Kapovich–Leeb [KL98]. For each
Seifert component Mi = Fi × S 1 of N , we choose a hyperbolic metric on
the base surface Fi so that all boundary components are totally geodesic
of unit length, and then equip each Seifert component Mi = Fi × S 1 with
the product metric di such that the fibers have length one. Metrics di on
Mi induce the product metrics on M̃i which by abuse of notations is also
denoted by di .
Let Mi and Mj be adjacent Seifert components in the simple graph manifold N , and let T ⊂ Mi ∩ Mj be a JSJ torus. Each metric space (T̃ , di )
and (T̃ , dj ) is a Euclidean plane. After applying a homotopy to the gluing
map, we may assume that at each JSJ torus T , the gluing map φ from the
←
−
→
−
boundary torus T ⊂ Mi to the boundary torus T ⊂ Mj is affine in the
sense that the identity map (T̃ , di ) → (T̃ , dj ) is affine. We now have a product metric on each Seifert component Mi = Fi × S 1 . These metrics may
not agree on the JSJ tori but the gluing maps are bilipschitz (since they
are affine). The product metrics on the Seifert components induce a length
metric on the graph manifold N denoted by d (see Section 3.1 of [BBI01]
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
12
for details). Moreover, there exists a positive constant K such that on each
Seifert component Mi = Fi × S 1 we have
1
di (x, y) ≤ d(x, y) ≤ K di (x, y)
K
for all x and y in Mi . (See Lemma 1.8 of [Pau05] for a detailed proof of the
last claim.) Metric d on N induces metric on Ñ , which is also denoted by
d (by abuse of notations). Then for all x and y in M̃i we have
1
di (x, y) ≤ d(x, y) ≤ K di (x, y)
K
The following remark introduces certain invariants of a horizontal surface
g that will be used in the proof of Theorem 6.1.
Remark 4.14.
(1) Since N is compact, there exists a positive lower bound ρ for the
distance between any two distinct JSJ planes in Ñ .
(2) We recall that for each curve ci in Tg , there exist non-zero integers
ai and bi such that
←
−
→
−
g(ci ) = ai [ fi ] + bi [ fi ] in H1 (Ti ).
The governor of a horizontal surface
g : S # N in a graph manifold
is the quantity ǫ = ǫ(g) = max ai /bi , bi /ai ci ∈ Tg .
Proposition 4.15. For each [γ] ∈ π1 (S, s0 ) as in Definition 4.10, we define
Λγ = max
k
Y
bi /ai
1≤j≤k≤m
i=j
If the horizontal surface g is a virtual embedding, then there exists a positive
constant Λ such that Λγ ≤ Λ for all [γ] ∈ π1 (S, s0 ).
Proof. Let Γg be the graph dual to Tg . Since the graph Γg is finite, it follows
that the set of non backtracking paths in Γg is finite. We denote this set
by {β1 , · · · , βm }. Let ni be the number of edges in βi with i = 1, . . . , m,
and let n = max ni i = 1, . . . , m . Let ǫ = ǫ(g) be the governor of the
horizontal surface g and let Λ = ǫn . Since the horizontal surface g is virtually
embedded, it follows that the dilation function is trivial by Theorem 4.11.
In other words, every loop α in S based at s0 has the dilation wα = 1.
Therefore, it follows that Λγ ≤ ǫn = Λ.
Remark 4.16. Let g : S # N be a horizontal surface in a simple graph
manifold. Equip N with the metric d described in Remark 4.13. By [Neu01,
Lemma 3.1] the surface g can be homotoped to another horizontal surface
g′ : S # N such that the following holds: For each curve c in g ′−1 (T ), let T
be the JSJ torus in N such that g′ (c) ⊂ T . Then g′ (c) is straight in T in the
sense that lifts of g′ (c) to Ñ are straight lines in the JSJ planes containing
it.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
13
Remark 4.17. Let f : S 1 # T = S 1 × S 1 be a horizontal immersion. Then
every fiber {x} × S 1 in the torus T has non-empty intersection with f (S 1 ).
Indeed, this follows from the fact the composition of f with the natural
projection of T to the first factor S 1 is a finite covering map.
5. A lower bound for distortion
The main theorem in this section is the following.
Theorem 5.1. Let g : (S, s0 ) # (N, x0 ) be a horizontal surface in a graph
manifold M . Let G = π1 (N, x0 ) and H = g∗ (π1 (S, s0 )). Then the distortion
∆G
H is at least quadratic. Furthermore, if the horizontal surface g is not
virtually embedded, then the distortion ∆G
H is at least exponential.
To see the proof of Theorem 5.1, we need several lemmas. For the rest of
this section we fix a horizontal surface g : S # N in a simple graph manifold
N . We equip N with the metric d described in Remark 4.13 and equip S
with a hyperbolic metric dS such that the boundary (if nonempty) is totally
geodesic and the simple closed curves of Tg are geodesics.
Lemma 5.2. Let γ be any geodesic loop in S such that γ and Tg have
nonempty intersection and such that wγ ≥ 1. There exists a positive number
A = A(γ) such that for all µ > 0 the following holds: Let {c1 , . . . , cm } be the
sequence of curves of Tg crossed by γ. The image of the circle g(ci ) in M
lies in a JSJ torus Ti . For each i = 1, 2, . . . , m, let ai and bi be the integers
such that
←
−
→
−
g(ci ) = ai [ fi ] + bi [ fi ] in H1 (Ti ; Z).
Extend the sequence a1 , . . . , am to a periodic sequence {aj }∞
j=1 with aj+m =
aj for all j > 0, and similarly extend b1 , . . . , bm to an m–periodic sequence
∞
{bj }∞
j=1 . Then there exists a (nonperiodic) sequence of integers t(j) j=1 ,
depending on our choice of the constant µ and the loop γ, with the following
properties:
(1) t(j) ≥ µ for all j.
(2) t(j) aj + t(j − 1) bj−1 ≤ A for all j > 1.
P
2
n
(3) The partial sum f (n) = nm
j=1 t(j) satisfies f (n) n + wγ .
Qm
Proof. By the definition
of the dilation function, we have wγ =
i=1 bi /ai .
Let A = max 1 + |ai | i = 1, 2, . . . , m . Set ξ = min |1/aj | , and
choose λ ∈ (0, 1] so that λ ≤ bj−1 /aj for all j > 1. Starting from an initial
value t(1) ≥ µ/λm−1 , we recursively construct an infinite sequence t(j)
satisfying (2). Suppose that t(j − 1) has been defined for some j > 1, and
we wish to define t(j). As A ≥ 1 + |aj |, we have
1≤
A + t(j − 1) bj−1
1 + t(j − 1) bj−1
A−1
=
−
.
|aj |
|aj |
|aj |
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
14
It follows that there is an integer t(j) such that
A + t(j − 1) bj−1
1 + t(j − 1) bj−1
≤ |t(j)| ≤
,
|aj |
|aj |
(♣)
which is equivalent to
1 ≤ t(j) aj − t(j − 1) bj−1 ≤ A.
Furthermore, we are free to choose the sign of t(j) so that t(j) aj and
t(j − 1) bj−1 have
opposite signs, which immediately gives (2). By induction,
the sequence t(j) satisfies both (♣)
and (2) for all j > 1.
We next show that any sequence t(j) satisfying (♣) also satisfies (3).
Indeed, the first inequality of (♣) implies
(♦)
t(j) ≥
bj−1
1
+
|aj |
aj
t(j − 1) ≥
bj−1
aj
t(j − 1)
for all j > 1.
For any j >
m, we apply (♦) iteratively m times (and use that the sequences
aj and bj are m–periodic) to get
bj−1
1
t(j − 1)
+
|aj |
aj
bj−1 bj−2
1
≥
t(j − 2) ≥ · · ·
+
|aj |
aj aj−1
b1 · · · bm
1
t(j − m)
+
≥
|aj |
a1 · · · am
t(j) ≥
(♥)
≥ ξ + wγ t(j − m) .
Further applying (♥) iteratively k times (and using that wγ ≥ 1) gives
t(j) ≥ ξ + wγ t(j − m)
≥ ξ + wγ ξ + wγ t(j − 2m)
(♠)
≥ 2ξ + wγ2 t(j − 2m) ≥ · · ·
≥ ξk + wγk t(j − km)
for all j > km
The inequality (♠) can be rewritten in the form
t(km + 1) ≥ ξk + t(1) wγk
for all k > 0.
Finally for each positive n we observe that
nm
X
j=1
t(j) ≥
n−1
X
k=1
t(km + 1) ≥ ξ
n−1
X
k=1
k + t(1)
n−1
X
wγk
k=1
which implies (3) as desired.
In order to establish (1), recall that γ satisfies wγ ≥ 1. Therefore (♥)
implies that t(j) ≥ |t(j −m)| for any j > m. In particular, it follows that
the terms of the sequence t(j) have absolute value bounded below by the
absolute values of the first m terms: t(1), . . . , t(m).
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
15
Using our choice of t(1) ≥ µ/λm−1 and the fact λ ≤ bj−1 /aj , the inequality (♦) shows that for all i = 1, . . . , m we have
µ
t(i) ≥ λi−1 t(1) ≥ m−i .
λ
As λ ≤ 1, we conclude that t(i) ≥ µ, completing the proof of (1).
Definition 5.3 (spirals). Let γ be a closed curve in S satisfying the conclusion of Lemma 3.3. Let {c1 , . . . , cm } be the sequence of curves of C crossed
by γ. Let {γ1 , . . . , γm } be the sequence of subpaths of γ introduced in Definition 4.10. Extend the finite sequences of curves c1 , . . . , cm and γ1 , . . . , γm to
∞
m–periodic infinite sequences {cj }∞
j=1 and {γj }j=1 . Let t(j) be a sequence
t(j)
of integers. We denote αj = cj .
Choose the basepoint s0 ∈ S to be a point of intersection between γ and
one of the curves of the family C. For each n ∈ N, we define a spiral loop
σn in S based at s0 as a concatenation:
σn = α1 γ1 · · · αnm γnm
and a double spiral loop ρn of σn in S based at s0 as ρn = σn α′nm+1 σn−1
t(1)
where α′nm+1 = cnm+1 .
∞
Lemma 5.4. Let t(j) j=1 be the sequence of integers given by Lemma 5.2.
Let ρn be the double spiral loop of σn coresponding to the curve γ and the
sequence t(j) . Let ρ̃n be the lift of ρn in S̃. Then the distance in Ñ
between the endpoints of g̃(ρ˜n ) is bounded above by a linear function of n.
Proof. First we describe informally the idea of the proof, which is illustrated
in Figure 1. The lift σ̃n of σn is the spiral-shaped curve running around the
outside of the left-hand diagram. The path σ̃n alternates between long
segments α̃j belonging to JSJ planes and short segments γ̃j belonging to
Seifert components. Each long segment α̃j is one side of a large triangle
in the JSJ plane whose other two sides are fibers of the adjacent blocks,
which meet in a corner yj opposite to α̃j . Connecting each pair of adjacent
corners yj produces a thin trapezoid that interpolates betweentwo adjacent
JSJ triangles. We will see that the sequence of exponents t(j)} in the
construction of σ̃n was chosen carefully to ensure that the distances between
adjacent corners yj are all short, i.e., bounded above. Thus the path running
around the inside of the spiral has at most a linear length—except for its
last segment, which is a long side of a large JSJ triangle. The double spiral
gives rise to a diagram similar to the spiral diagram, expect that it has
been doubled along this long triangular side, so that the long side no longer
appears on the boundary of the diagram but rather appears in its interior
(see Figure 2).
To be more precise, let Mj be the Seifert component of N containing γj ,
with its given product metric as a hyperbolic surface crossed with a circle
of length one. Let (M̃j , dj ) be the Seifert component of Ñ that contains
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
16
xj
α̃j
zj
γ̃j
xj+1
yj
α̃j+1
yj+1
zj+1
Figure 1. On the left, the piecewise geodesic σ̃n is the
spiral-shaped path running around the outside of the diagram. On the right is a magnified portion of the left-hand
picture showing more detail.
γ̃j , with the product metric dj induced by lifting the given metric on Mj .
Let T̃j be the JSJ plane containing α̃j . Recall that T̃j is a topological plane
that is a subspace of both M̃j−1 and M̃j . However the metrics dj−1 and
dj typically do not agree on the common subspace T̃j . Each metric space
←
−
−
→
Ej = (T̃j , dj−1 ) and Ej = (T̃j , dj ) is a Euclidean plane and the identity map
←
−
−
→
Ej → Ej is affine.
The plane T̃j universally covers a JSJ torus Tj obtained by identifying
←
−
→
−
boundary tori Tj and Tj of Seifert components Mj−1 and Mj . The initial
and terminal points xj and zj of α̃j are contained in Euclidean geodesics
←
− ←
→ −
−
←
− ←
→ −
−
−
→
−
→
ℓj ⊂ Ej and ℓj ⊂ Ej that project to fibers fj ⊂ Tj and fj ⊂ Tj respectively.
←
−
→
−
The lines ℓj and ℓj intersect in a unique point yj . Similarly, we consider
the subpath α̃−1
j in the double spiral ρ̃n . Let y j be the intersection point of
the two fibers that contain its endpoints.
Our goal is to find a linear upper bound for the distance in (N, d) between
the endpoints of ρ̃n . By the triangle inequality it suffices to produce an
upper bound for the distance between successive points of the linear sequence
y1 , . . . , ynm , y nm , . . . , y 1 . Recall that the inclusions (M̃j , dj ) → (Ñ , d) are
K–bilipschitz for some universal constant K, as explained in Remark 4.13.
Thus it is enough to bound the distance between these points with respect
to the given product metrics on each Seifert component.
Let η be the maximum of lengths of γ̃j with respect to metric dj . Let A
be the constant given by Lemma 5.2. We claim that
(⋆)
dj (yj , yj+1 ) ≤ η + A.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
17
Figure 2. The outer path represents the piecewise geodesic ρ̃n .
We prove this claim by examining the quadrilateral with vertices yj , zj , xj+1 ,
and yj+1 , illustrated on the right-hand side of Figure 1. This quadrilateral
is a trapezoid in the sense that the opposite sides [yj , zj ] and [yj+1 , xj+1 ] lie
in fibers of M̃j that are parallel lines in the product metric dj .
To find the lengths of these parallel sides, we must first examine the JSJ
triangle ∆(xj , yj , zj ), also shown on the right-hand side of Figure 1. We
consider this triangle to be a pair of homotopic paths α̃j and τj from xj to
←
−
yj in the plane T̃j , where τj = [xj , yj ]∪[yj , zj ] is a segment of ℓj concatenated
→
−
with a segment of ℓj . In particular, α̃j and τj project to homotopic loops
←
−
→
−
in Tj of the form g(αj ) and ( fj )rj ( fj )sj for some rj , sj ∈ Z. Recall that the
←
−
→
−
fiber fj has length one in the product metric on Mj−1 . Similarly the fiber fj
has length one in Mj . It follows that dj−1 (xj , yj ) = |rj | and dj (yj , zj ) = |sj |.
Since g(αj ) = g(cj )t(j) , the homology relation (written additively)
←
−
→
−
t(j) g(cj ) = t(j) aj [ fj ] + t(j) bj [ fj ] in H1 (Tj ; Z) ≈ π1 (Tj ),
implies that rj = t(j) aj and sj = t(j) bj .
Let βj be the lift of γj based at yj . The fiber which contains yj+1 and
xj+1 will intersect βj exactly at one point. We denote this point by uj . It
follows that dj (uj , yj ) ≤ η and dj (uj , yj+1 ) = t(j + 1) aj+1 + t(j) bj . Using
the triangle inequality for ∆(yj , uj , yj+1 ), we have
dj (yj , yj+1 ) ≤ dj (yj , uj ) + dj (uj , yj+1 )
≤ η + t(j + 1) aj+1 + t(j) bj
≤η+A
by Lemma 5.2(2), completing the proof of (⋆).
For a similar reason, the distance in the corresponding Seifert component
between y j and y j−1 is at most η + A. Thus it suffices to find an upper
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
18
bound for dnm (ynm , y nm ). Let R be the length of c̃1 in the metric dnm . Let
t(1)
′
z nm be the initial point of α̃−1
nm . Since αnm = cnm+1 , the dnm –distance between the endpoints of α̃′nm+1 is at most R t(1) . By the triangle inequality,
dnm (znm , z nm ) ≤ 2η + R t(1) .
We note that [znm , ynm ] and [z nm , y nm ] lie in parallel fibers of M̃nm . Furthermore they are oriented in the same direction with respect to the fiber
and have the same length. Therefore they form opposite sides of a Euclidean parallelogram in M̃nm . In particular the distances dnm (ynm , y nm )
and dnm (znm , z nm ) are equal, so that dnm (ynm , y nm ) ≤ 2η + R t(1) .
Proof of Theorem 5.1. Any graph manifold has a simple finite cover by Theorem 4.3. By Corollary 4.7 and Proposition 2.9, it suffices to prove the theorem for all horizontal surfaces in this cover. Thus we assume, without loss
of generality that N is a simple graph manifold.
Let γ be any geodesic loop in S such that γ and Tg have nonempty
intersection. The existence of such a loop is guaranteed by Lemma 3.3.
Replacing γ with γ −1 if necessary, we may assume that wγ ≥ 1. Let κ be
the maximum of lengths of γi with respect to the metric dS .
Let µ, L and C be the constants given by Lemma 3.4. Let t(j) be the
sequence of integers given by Lemma 5.2. For each n,
let σn be the spiral
loop coresponding to the curve γ and the sequence t(j) . Let ρn be the
double spiral of σn . Let L be the family of lines that are lifts of loops of Tg
or boundary loops of S. Since ρ̃n satisfies the hypotheses of Lemma 3.4, it
is an (L, C)–quasigeodesic in S̃.
Let hn be the homotopy
class of loop ρn at the basepoint s0 . We claim
2
that dS̃ s̃0 , hn (s̃0 ) n + wγn . Indeed, let r be the minimum of lengths of
ci with respect to P
the metric dS . By the construction of the spiral loop σn
2
n
we have σ̃n ≥ r nm
j=1 t(j) . By Lemma 5.2(3), we have σ̃n n + wγ .
It is obvious that |ρ̃n | n2 + wγn because |ρ̃n | ≥ 2|σ̃n |. Since ρ̃n is an
(L, C)–quasigeodesic, it follows that dS̃ s̃0 , hn (s̃0 ) n2 + wγn .
Furthermore, we have d x̃0 , hn (x̃0 ) n by Lemma 5.4. Therefore, n2 +
G
wγn ∆G
H by Corollary 2.10. It follows that ∆H is at least quadratic. If
the horizontal surface is not virtually embedded, then we may choose the
geodesic loop γ in S such that wγ > 1 by Theorem 4.11. In this case wγ is
an exponential function, and wγn ∆G
H.
6. Upper Bound of Distortion
In this section, we will find the upper bound of the distortion of horizontal
surface. The main theorem in this section is the following.
Theorem 6.1. Let g : (S, s0 ) # (N, x0 ) be a horizontal surface in a graph
manifold N . Let G = π1 (N, x0 ) and H = g∗ π1 (S, s0 ) . Then the distortion ∆G
H is at most exponential. Furthermore, if the horizontal surface g is
virtually embedded then ∆G
H is at most quadratic.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
19
Definition 6.2. Lift the JSJ decomposition of the simple graph manifold N
to the universal cover Ñ , and let TN be the tree dual to this decomposition
of Ñ . Lift the collection Tg to the universal cover S̃. The tree dual to
this decomposition of S̃ will be denoted by TS . The map g̃ induces a map
T : TS → TN .
The following lemma plays an important role in the proof of Theorem 6.1.
Lemma 6.3. Let F be a connected compact surface with non-empty boundary and χ(F ) < 0. Let M = F × S 1 . Let g : (B, b) # (M, x) be a horizontal
surface. Then each fiber in M̃ intersects with g̃(B̃) exactly at one point.
Proof. According to Lemma 2.1 in [RW98], there exists a finite covering map
p : B × S 1 → M and an embedding i : B → B×S 1 given by i(x) = (x, 1) such
that g = p ◦ i. Let ĩ : B̃ → B̃ × R be the lift of i such that ĩ(b̃) = (b̃, 0). Let
φ : B̃ → B and Ψ : F̃ → F be the universal covering maps. Let π : R → S 1
be the usual covering space. Since B̃ × R and F̃ × R both uninversal cover
M , there exists a homeomorphism ω : B̃ × R, (b̃, 0) → F̃ × R, g̃(b̃) such
that (Ψ × π) ◦ ω = p ◦ (φ × π). By the unique
lifting property, we have
ω ◦ ĩ = g̃. It follows that g̃(B̃) = ω B̃ × {0} . Since ω maps each fiber in
B̃ × R to a fiber in F̃ × R. It follows that each fiber in F̃ × R intersects
ω B̃ × {0} exactly at one point.
Proposition 6.4. The map T is bijective.
Proof. A simplicial map between trees is bijective if is locally bijective. Thus
it suffices to show that the map T is locally injective and locally surjective
(see [Sta83] for details).
Suppose by way of contradiction that T is not locally injective. Then
there exist three distinct blocks B̃1 , B̃2 and B̃3 in S̃ such that B̃1 ∩ B̃2 6= ∅
and B̃2 ∩ B̃3 6= ∅ such that the images g̃(B̃1 ) and g̃(B̃3 ) lie in the same
block M̃1 of M̃ . Let M̃2 denote the block containing the image g̃(B̃2 ). We
denote ℓ1 = B̃1 ∩ B̃2 and ℓ3 = B̃2 ∩ B̃3 . We have g̃(ℓ1 ) and g̃(ℓ3 ) are subsets
of the JSJ plane T̃ = M̃1 ∩ M̃2 . Since the lines ℓ1 and ℓ3 are disjoint and
the map g̃ is an embedding, it follows that g̃(ℓ1 ) and g̃(ℓ3 ) are disjoint lines
in the plane T̃ . By Remark 4.17, any fiber in the plane T̃ intersects g̃(ℓ1 )
and g̃(ℓ3 ) at distinct points. This contradicts with Lemma 6.3 because g̃(ℓ1 )
and g̃(ℓ3 ) are subsets of g̃(B̃2 ). Therefore, T is locally injective.
By Lemma 6.3, the block g̃(B̃) must intersect every fiber of the Seifert
component M̃ containing it. In particular, g̃(B̃) intersects every JSJ plane
adjacent to M̃ . Therefore the map T is locally surjective.
The following corollary is a combination of Lemma 6.3 and Proposition 6.4.
Corollary 6.5. Each fiber of Ñ intersects with g̃(S̃) in one point.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
20
Remark 6.6. Let G and H be finitely generated groups with generating sets
A and B respectively. Let φ : G → H be a homomorphism. Then there exists
a positive number L such that φ(g) B ≤ L|g|A for all g in G. Indeed, suppose
that A = {g1 , g2 , . . . , gn } we define L = max |φ(gi )|B i = 1, 2, . . . , n .
Since φ is a homomorphism, it is not hard to see that φ(g) B ≤ L|g|A for
all g ∈ G.
The following proposition shows that the distortion of a horizontal surface
in a trivial Seifert manifold is linear.
Proposition 6.7. Let F be a connected compact surface with non-empty
boundary and χ(F ) < 0. Let g : (B, b) # (M, x) be a horizontal surface. Let
H = g∗ π1 (B, b) and G = π1 (M, x). Then H ֒→ G is a quasi-isometric
embedding.
Proof. We first choose generating sets for π1 (B), π1 (F ) and π1 (S 1 ). The
generating sets of π1 (F ) and π1 (S 1 ) induce a generating set on π1 (M ). Let
g1 : B → F and g2 : B → S 1 be the maps such that g = (g1 , g2 ). We have
g1 : B → F is a finite covering
map because g is a horizontal surface in M .
It follows that g1∗ π1 (B) is a finite index subgroup of π1 (F ). As a result,
g1∗ is an (L′ , 0)–quasi-isometry for some constant L′ . Since g∗ = (g1∗ , g2∗ )
we have g∗ (h) ≥ g1∗ (h) ≥ |h| L′ for all h ∈ π1 (B). Applying Remark 6.6
to the homomorphism g∗ , the constant L′ can be enlarged so that we can
show that g∗ is an (L′ , 0)–quasi–isometric embedding.
For the rest of this section, we fix g : (S, s0 ) # (N, x0 ) a horizontal surface
in a simple graph manifold N , the metric d which given by Remark 4.13, and
the hyperbolic metric dS on S which described in Section 5. By Remark 4.16,
we also assume that for each curve c in Tg then g(c) is a straight in the
JSJ torus T where T is the JSJ torus such that g(c) ⊂ T . We define a
metric dg̃(S̃) on g̃(S̃) as the following: for any u = g̃(x) and v = g̃(y), we
define dg̃(S̃) (u, v) = dS̃ (x, y). The following corollary follows by combining
Proposition 6.7 with several earlier results, using the fact that S has only
finitely many blocks, and N has only finitely many Seifert components.
Corollary 6.8. There exist numbers L and C such that the following holds:
For each block B̃ in S̃, let M̃ = F̃ × R be the Seifert component of Ñ such
that g̃(B̃) ⊂ M̃ . The map g̃|B̃ : B̃ → M̃ = F̃ × R can be expressed as a pair
of maps g̃1 : B̃ → F̃ and g̃2 : B̃ → R. Then g̃1 and g̃|B̃ are (L, C)–quasiisometric embeddings, and
g̃2 (u) − g̃2 (v) ≤ L d g̃1 (u), g̃1 (v) + C
for all u, v ∈ B̃.
Proof. The map g|B : B → F × S 1 can be written as (g1 , g2 ). Since the map
g1 : B → F is a finite covering map, the lift g̃1 is a quasi–isometry. It follows
from Proposition 6.7 that g̃|B̃ is a quasi–isometric embedding. The facts g̃1
and g̃|B̃ are quasi–isometrically embedded imply the final claim.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
21
Now, we describe informally the strategy of the proof of Theorem 6.1 in
the case N is a simple graph manifold. For each n ∈ N, let u ∈ g̃(S̃) such
that d(x̃0 , u) ≤ n. We would like to find an upper bound (either quadratic or
exponential) of dS̃ (x̃0 , u) in terms of n. We first show the existence of a path
β in Ñ connecting x̃0 to u such that β passes through a sequence of Seifert
components M̃1 , . . . , M̃k and β intersects the plane T̃i = M̃i ∩ M̃i+1 exactly
at one point that is denoted by yi with i = 1, . . . , k − 1 if k ≥ 2. We will
show that k is bounded above by a linear function of n. Therefore it suffices
to find an upper bound for dS̃ (x̃0 , u) in terms of k. Let di be the given
product metric on M̃i . By Corollary 6.5 the fiber of M̃i passing through yi
intersects g̃(S̃) in a unique point which is denoted by xi . Similarly, the fiber
of M̃i+1 passing through yi intersects g̃(S̃) in one point which is denoted by
zi .
On the one hand, proving the distance in g̃(S̃) of endpoints of β is domi
Pk−1
di (yi , xi ) + di+1 (yi , zi ) is easy. On the other hand,
nated by the sum i=1
finding a quadratic or exponential upper bound for this sum as a function
of k requires more work. Our goal is to analyze the growth of the sequence
d1 (y1 , x1 ), d2 (y1 , z1 ), d2 (y2 , x2 ), . . . ,
di−1 (yi−1 , xi−1 ), di (yi−1 , zi−1 ), di (yi , xi ), . . . ,
dk−1 (yk−1 , xk−1 ), dk (yk−1 , zk−1 )
A relation between di (yi−1 , zi−1 ) and di (yi , xi ) in the Seifert component M̃i
will be described in Lemma 6.9. The ratio of di (yi−1 , zi−1 ) to di−1 (yi−1 , xi−1 )
in the JSJ plane T̃i−1 will be described in Lemma 6.10.
Lemma 6.9 (Crossing a Seifert component). There exists a positive constant L′ such that the following holds: For each block B̃ in S̃, let M̃ = F̃ ×S 1
be the Seifert component such that g̃(B̃) ⊂ M̃ . Let dM be the given product
metric on M̃ , and let T̃ and T˜′ be two disjoint JSJ planes in the Seifert
component M̃ . For any two points y ∈ T̃ and y ′ ∈ T˜′ , let ℓ ⊂ T̃ and ℓ′ ⊂ T˜′
be the lines that project to the fiber S 1 in M such that y ∈ ℓ and y ′ ∈ ℓ′ . Let
x (resp. x′ ) be the unique intersection point of ℓ (resp. ℓ′ ) with g̃(B̃) ∩ M̃
given by Lemma 6.3. Then
dM (y ′ , x′ ) ≤ dM (y, x) + L′ dM (y, y ′ )
Proof. Let ρ and K be the constants given by Remark 4.13 and Remark 4.14
respectively, and let D = ρ/K. Let L and C be the constants given by
Corollary 6.8. Let L′ = L + 2 + C/D.
Let a and b be the projection points of y and y ′ to F̃ respectively. We
write g̃|B̃ = (g̃1 , g̃2 ) where g̃1 : B̃ → F̃ and g̃2 : B̃ → R. Since g̃|B̃ is an
embedding map and x, x′ ∈ g̃(B̃), there exist a′ , b′ ∈ B̃ such that g̃(a′ ) = x
and g̃(b′ ) = x′ . It follows that g̃1 (a′ ) = a and g̃1 (b′ ) = b. By Corollary 6.8
we have
g̃2 (a′ ) − g̃2 (b′ ) ≤ L dM g̃1 (a′ ), g̃1 (b′ ) + C = L dM (a, b) + C.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
22
Since ρ ≤ d(a, b) ≤ K dM (a, b), it follows that D ≤ dM (a, b). Therefore
g̃2 (a′ ) − g̃2 (b′ ) ≤ (L + C/D)dM (a, b) ≤ (L + C/D)dM (y, y ′ ).
With respect to the orientation of the factor R of M̃ , let ∆(y, a) and
∆(y ′ , b) be the displacements of pairs of points (y, a) and (y ′ , b) respectively.
We would like to show that ∆(y, a) − ∆(y ′ , b) ≤ 2 dM (y, y ′ ). Indeed, let
s and t be the real numbers such that y = (a, s) and y ′ = (b, t). We
note that ∆(y, a) = −s if s ≥ 0 and ∆(y, a) = s if s ≤ 0 as well as
∆(y ′ , b) = −t if t ≥ 0 and ∆(y ′ , b) = t if t ≤ 0. Since dM (a, b) ≤ dM (y, y ′ ),
it follows that ∆(y, a) − ∆(y ′ , b) ≤ 2dM (y, y ′ ). Moreover, we have that
dM (y, x) = g̃2 (a′ ) + ∆(y, a) and dM (y ′ , x′ ) = g̃2 (b′ ) + ∆(y ′ , b) . Therefore
the previous inequalities imply
dM (y ′ , x′ ) − dM (y, x) ≤ g̃2 (a′ ) − g̃2 (b′ ) + ∆(y, a) − ∆(y ′ , b)
≤ (L + 2 + C/D)dM (y, y ′ ) = L′ dM (y, y ′ ).
Lemma 6.10 (Crossing a JSJ plane). Let M̃ and M̃ ′ be the two adjacent
Seifert components. Let dM and dM ′ be the given product metrics on M̃
and M̃ ′ respectively. Let T̃ = M̃ ∩ M̃ ′ , and let α ⊂ g̃(S̃) ∩ T̃ be the line
such that α universally covers a curve g(c) for some c in Tg . We moreover
assume that α is a straight line in (T̃ , dM ). For any two points x and z in
←
−
→
−
the line α, let ℓ ⊂ (T̃ , dM ) and ℓ ⊂ (T̃ , dM ′ ) be the Euclidean geodesics
←
−
−
→
←
− ←
→ −
−
−
→
such that x ∈ ℓ and z ∈ ℓ and they project to fibers f ⊂ T and f ⊂ T
←
−
→
−
respectively. Let y be the unique intersection point of ℓ and ℓ . Let a and
b be the integers such that
←
−
→
−
g(c) = a[ f ] + b[ f ] in H1 (T ; Z).
←
−
→
−
where T is the JSJ torus obtained from gluing T to T . Then dM (y, x) =
a/b dM ′ (y, z)
→
−
Proof. Let κ be the positive constant such that length of the fiber f with
respect to the metric dM equals to 1/κ. Let c̃ be a path lift of c such
that g̃(c̃) ⊂ α. Let x′ and z ′ be the initial point and the terminal point
←
−
→
−
of g̃(c̃) respectively. Let ℓ′ and ℓ′ be the Euclidean geodesics in (T̃ , dM )
←
−
→
−
←
−
and (T̃ , dM ′ ) respectively such that x′ ∈ ℓ′ and z ′ ∈ ℓ′ and both lines ℓ′
→
−
←
−
→
−
and ℓ′ project to fibers in T and T respectively. Let y ′ be the unique
←
−′
→′
−
intersection point of ℓ and ℓ . It was shown in the proof of Lemma 5.4
that dM (y ′ , x′ ) = |a| and dM ′ (y ′ , z ′ ) = |b|. By the definiton of κ, it follows
that dM ′ (y ′ , z ′ ) = κ dM (y ′ , z ′ ).
In the Euclidean plane (T̃ , dM
z)
), consider the similar
triangles ∆(x, y,
and ∆(x′ , y ′ , z ′ ). Since dM (y,x) dM (y, z) = dM(y ′ , x′ ) dM (y ′ , z ′ ) = κ a b ,
it follows that dM (y, x) = a b κ dM (y, z) = a b dM ′ (y, z).
Proof of Theorem 6.1. We may assume that N is a simple graph manifold
for the same reason as in the first paragraph of the proof of Theorem 5.1.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
23
Let K be the constant given by Lemma 4.13. Let L and C be the constants
given by Corollary 6.8. Moreover, the constant L can be enlarged so that
L ≥ K and L is greater than the constant L′ given by Lemma 6.9. Moreover,
we assume that the base point s0 belongs to a curve
in the collection Tg .
For any h ∈ π1 (S, s0 ) such that d x̃0 , h(x̃0 ) ≤ n. We will show that
dS̃ s̃0 , h(s̃0 ) is bounded above by either a quadratic or exponential function in term of n. The theorem is confirmed by an application of By Corollary 2.10. We consider the following cases:
Case 1: s̃0 and h(s̃0 ) belong to the
same a block B̃. In this degenerate
case, we will show that dS̃ s̃0 , h(s̃0 ) n. Indeed, let M̃ be the Seifert
component such that g̃(B̃) ⊂ M̃ . Let dM̃ be the given product metric on
M̃ . Since g̃|B̃ is an (L, C)–quasi–isometric embedding by Corollary 6.8 , it
follows that
dg̃(B̃) x̃0 , h(x̃0 ) ≤ LdM̃ x̃0 , h(x̃0 ) + C ≤ L2 d x̃0 , h(x̃0 ) + C ≤ L2 n + C.
Therefore,
dS̃ (s̃0 , h(s̃0 )) = dg̃(S̃) x̃0 , h(x̃0 ) ≤ dg̃(B̃) x̃0 , h(x̃0 ) ≤ L2 n + C.
Case 2: s̃0 and h(s̃0 ) belong to distinct blocks of S̃. Let L be the family
of lines in S̃ that are lifts of curves of Tg . Since we assume that s0 belongs
to a curve in the collection Tg , thus there are distinct lines α and α′ in L
such that s̃0 ∈ α and h(s̃0 ) ∈ α′ . Let e and e′ be the non-oriented edges in
the tree TS corresponding to the lines α and α′ respectively. Let β be the
non backtracking path joining e to e′ in the tree TS , with ordered vertices
v0 , v1 , . . . , vk−1 where v1 is not a vertex on the edge e and vk−2 is not a
vertex on the edge e′ . We denote the Seifert components corresponding to
the vertices T(vi ) by M̃i with i = 0, 1, · · · , k − 1. We note that the Seifert
components M̃i are distinct because T is injective by Proposition 6.4. Let
di be the given product metric on the Seifert component M̃i .
For convenience, relabel x̃0 by y0 , and h(x̃0 ) by yk . Let γ be a geodesic
in (Ñ , d) joining y0 to yk . Replace the path γ by a new path β, described
as
follows. For i = 1, 2, . . . , k − 1, let T̃i = M̃i−1 ∩ M̃i , and let Ai = t ∈ [0, 1]
γ(t) ∈ T̃i . Note that the set Ai is non-empty. Let ti = sup {t ∈ Ai }, and
let yi = γ(ti ). Let βi be the geodesic segment in (M̃i , di ) joining yi−1 to yi .
Let β be the concatenation β1 β2 · · · βk . Since di (yi−1 , yi ) ≤ L d(yi−1 , yi ), it
follows that |β| ≤ L |γ| ≤ Ln.
←
−
→
−
For each i = 1, 2, . . . , k − 1, let ℓi and ℓi be the Euclidean geodesics in
(T̃i , di−1 ) and (T̃i , di ) passing through yi respectively. Let αi = g̃(S̃) ∩ T̃i .
←
−
→
−
By Corollary 6.5, each line ℓi and ℓi intersects αi at one point, we denote
←
−
→
−
{xi } = αi ∩ ℓi and {zi } = αi ∩ ℓi .
Pk−1
di+1 (yi , zi ).
Claim 1: dg̃(S̃) x̃0 , h(x̃0 ) i=1
We recall that Tg is the collection of components of g −1 (T ) in S where T
is the JSJ decomposition of N . We note that αi universally covers a closed
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
24
←
−
→
−
curve g(ci ) ∈ Ti for some ci ∈ Tg and g(ci ) = ai [ fi ] + bi [ fi ] in H1 (Ti ; Z)
for some ai , bi ∈ Z. By Lemmas 6.10 and 6.9, we have
(∗)
di (yi , xi ) = ai /bi di+1 (yi , zi )
and
(†)
di yi , xi − di yi−1 , zi−1 ≤ Ldi yi , yi−1 ≤ L2 βi .
For each i = 2, . . . , k − 1 we have
di (zi−1 , zi ) ≤ di (zi−1 , yi−1 ) + di (yi , zi ) + di (yi−1 , yi ).
The triangle inequality implies
d1 (y0 , z1 ) ≤ d1 (y0 , y1 ) + d1 (y1 , z1 )
and
dk−1 (zk−1 , yk ) ≤ dk−1 (zk−1 , yk−1 ) + dk−1 (yk−1 , yk ).
By Corollary 6.8, we have
dg̃(S̃) (x̃0 , u) ≤ 2L
k−1
X
di+1 (yi , zi ) + L|β| + kC.
i=1
Let ρ be the minimum distance between two distinct JSJ planes in Ñ (see
Remark 4.14). Since d x̃0 , h(x̃0 ) ≤ n, it follows that kρ ≤ n, and therefore
k ≤ n/ρ.
Pk−1
Claim 2: i=1
di+1 (yi , zi ) en .
Indeed, by using (∗) and (†) we have
(‡)
bi−1
di−1 (yi−1 , xi−1 )
di yi−1 , zi−1 =
ai−1
bi−1
≤
di−1 (yi−2 , zi−2 ) + L2 βi−1
ai−1
bi−1 2
bi−1 bi−2
di−2 (yi−3 , xi−3 ) +
L βi−1 ≤ · · ·
≤
ai−1 ai−2
ai−1
Let ǫ = ǫ(g) be the governor of g, as defined in Remark 4.14. We note that
ǫ > 1 if g is not virtually embedded. By (‡) we have
di (yi−1 , zi−1 ) ≤ ǫ2 di−2 (yi−3 , zi−3 ) + ǫ2 L2 βi−2 + ǫL βi−1
≤ ǫi−2 d2 (y1 , z1 ) + L2 (Ln)
i−2
X
ǫj
j=1
We have d2 (y1 , z1 ) ≤ ǫd1 (y1 , x1 ) and d1 (y1 , x1 ) ≤ Ld1 (y1 , y0 ) ≤ L2 n by using
(∗) and (†). Hence d2 (y1 , z1 ) ≤ L2 nǫ ≤ L3 nǫ. Therefore
di (yi−1 , zi−1 ) ≤ L3 n
i−1
X
j=1
ǫj .
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
25
Pk−1
When ǫ > 1 we obtain i=1
di+1 (yi , zi ) en . If ǫ = 1 then since k ≤ n/ρ,
Pk−1
it is not hard to see i=1 di+1 (yi , zi ) n2 en .
Pk−1
Claim 3: If g is virtually embedded then i=1
di+1 (yi , zi ) n2 .
Indeed, let Λ be the constant given by Remark 4.14. By (‡) we have
di (yi−1 , zi−1 ) ≤ Λ
k−1
X
βm ≤ ΛLn.
m=1
It follows that
k−1
X
di+1 (yi , zi ) ≤ 2kΛLn ≤ 2(n/ρ)ΛLn,
i=1
establishing the claim. The theorem now follows from Corollary 2.10.
7. Fiber surfaces have quadratic distortion
In this section we show in detail how results of Gersten and Kapovich–
Leeb [Ger94, KL95] can be combined with Thurston’s geometric description
of 3–manifolds that fiber over the circle to prove Theorem 1.5. This section
is an elaboration of ideas that are implicitly used by Kapovich–Leeb in
[KL95] but not stated explicitly there. As explained in the introduction,
Theorem 1.5 is the main step in an alternate proof of the virtually embedded
case of Theorem 1.2.
The following theorem relates distortion of normal subgroups to the notion of divergence of groups. Roughly speaking, divergence is a geometric
invariant that measures the circumference of a ball of radius n as a function
of n. (See [Ger94] for a precise definition.)
Theorem 7.1 ([Ger94], Thm. 4.1). If G = H ⋊φ Z, where G and H are
finitely generated, then the divergence of G is dominated by the distortion
∆G
H.
Let H be generated by a finite set T . An automorphism φ ∈ Aut(H) has
polynomial growth of degree at most d if there exist constants α, β such that
φi (t)
T
≤ αnd + β
for all t ∈ T and all i with |i| ≤ n.
Gersten claims the following result in the case that H is a free group.
However his proof uses only that H is finitely generated, so we get the
following result using the same proof.
Theorem 7.2 ([Ger94], Prop. 4.2). If G = H ⋊φ Z, where G and H are
finitely generated and φ ∈ Aut(H) has polynomial growth of degree at most
d+1 .
d, then ∆G
H n
Proof of Theorem 1.5. Let N be a graph manifold that fibers over S 1 with
surface fiber S. Then N is the mapping torus of a homeomorphism f ∈
Aut(S). In particular, if we let G = π1 (N ) and H = π1 (S), then G = H ⋊φ Z,
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
26
where φ ∈ Aut(H) is an automorphism induced by f . Passing to finite
covers, we may assume without loss of generality that N and S are orientable
and that the map f is orientation preserving.
Kapovich–Leeb show that the divergence of the fundamental group of any
graph manifold is at least quadratic [KL98, §3]. Therefore by Theorem 7.1
the distortion of H in G is also at least quadratic.
By Theorem 7.2, in order to establish a quadratic upper bound for ∆G
H,
it suffices to show that φ has linear growth. We will first apply the Nielsen–
Thurston classification of surface homeomorphisms to the map f (see, for example, Corollary 13.2 of [FM12]). By the Nielsen–Thurston theorem, there
exists a multicurve {c1 , . . . , ck } with the following properties. The curves ci
have disjoint closed annular neighborhoods S1 , . . . , Sk . Let Sk+1 , . . . , Sk+ℓ
S
be the closures of the connected components of S − ki=1 Si . Then there is
a map g isotopic to f and a positive number m such that gm leaves each
subsurface Si invariant. Furthermore gm is a product of homeomorphisms
g1 · · · gk+ℓ such that each gi is supported on Si . For i = 1, . . . , k, the map
gi (supported on the annulus Si ) is a power of a Dehn twist about ci . For
i = k + 1, . . . , k + ℓ, each gi is either the identity or a map that restricts to
a pseudo-Anosov map of Si .
Consider the mapping torus N̂ for the homeomorphism gm of S, which
finitely covers N . Apply Thurston’s geometric classification of mapping tori
to N̂ to conclude that the family of tori ci × S 1 in the mapping torus N̂ is
equal to the family of JSJ tori of N̂. It follows that for each i = k+1, . . . , k+ℓ
the map gi is equal to the identity on S. Indeed if any gi were pseudoAnosov, then the corresponding JSJ component of N̂ would be atoroidal
and hyperbolic, which is impossible in a graph manifold.
Therefore gm is a product of powers of Dehn twists about disjoint curves.
In particular the automorphism φm has linear growth. Clearly φ itself must
also have linear growth. Therefore the distortion of H in G is at most
quadratic as desired.
References
[Ago13] I. Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013. With an
appendix by Agol, Daniel Groves, and Jason Manning.
[BBI01] D. Burago, Y. Burago, and S. Ivanov. A course in metric geometry, volume 33 of
Graduate Studies in Mathematics. American Mathematical Society, Providence,
RI, 2001.
[BS04] S.V. Buyalo and P.V. Svetlov. Topological and geometric properties of graph
manifolds. Algebra i Analiz, 16(2):3–68, 2004.
[BW13] H. Bigdely and D.T. Wise. Quasiconvexity and relatively hyperbolic groups that
split. Michigan Math. J., 62(2):387–406, 2013.
[DS05] C. Druţu and M. Sapir. Tree-graded spaces and asymptotic cones of groups.
Topology, 44(5):959–1058, 2005. With an appendix by D. Osin and M. Sapir.
[FM12] B. Farb and D. Margalit. A primer on mapping class groups, volume 49 of Princeton Mathematical Series. Princeton University Press, Princeton, NJ, 2012.
[Ger94] S.M. Gersten. Quadratic divergence of geodesics in CAT(0) spaces. Geom. Funct.
Anal., 4(1):37–51, 1994.
DISTORTION OF SURFACES IN GRAPH MANIFOLDS
27
[GH90] É. Ghys and P. de la Harpe, editors. Sur les groupes hyperboliques d’après Mikhael
Gromov, volume 83 of Progress in Mathematics. Birkhäuser Boston, Inc., Boston,
MA, 1990.
[HP15] M.F. Hagen and P. Przytycki. Cocompactly cubulated graph manifolds. Israel J.
Math., 207(1):377–394, 2015.
[Hru10] G.C. Hruska. Relative hyperbolicity and relative quasiconvexity for countable
groups. Algebr. Geom. Topol., 10(3):1807–1856, 2010.
[Joh79] K. Johannson. Homotopy equivalences of 3–manifolds with boundaries, volume
761 of Lecture Notes in Mathematics. Springer, Berlin, 1979.
[JS79] W.H. Jaco and P.B. Shalen. Seifert fibered spaces in 3–manifolds. Mem. Amer.
Math. Soc., 21(220):1–192, 1979.
[KL95] M. Kapovich and B. Leeb. On asymptotic cones and quasi-isometry classes of
fundamental groups of 3–manifolds. Geom. Funct. Anal., 5(3):582–603, 1995.
[KL98] M. Kapovich and B. Leeb. 3–manifold groups and nonpositive curvature. Geom.
Funct. Anal., 8(5):841–852, 1998.
[Neu01] W.D. Neumann. Immersed and virtually embedded π1 –injective surfaces in graph
manifolds. Algebr. Geom. Topol., 1:411–426, 2001.
[Pau05] G. Paulik. Gluing spaces and analysis. Bonner Mathematische Schriften [Bonn
Mathematical Publications], 372. Universität Bonn, Mathematisches Institut, Bonn, 2005. Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn,
Bonn, 2005.
[PW14] P. Przytycki and D.T. Wise. Graph manifolds with boundary are virtually special. J. Topol., 7(2):419–435, 2014.
[RW98] J.H. Rubinstein and S. Wang. π1 –injective surfaces in graph manifolds. Comment. Math. Helv., 73(4):499–515, 1998.
[Sco78] P. Scott. Subgroups of surface groups are almost geometric. J. London Math.
Soc. (2 ), 17(3):555–565, 1978.
[Sta83] J.R. Stallings. Topology of finite graphs. Invent. Math., 71(3):551–565, 1983.
[Tid]
J. Tidmore. Cocompact cubulations of mixed 3–manifolds. Preprint.
arXiv:1612.06272 [math.GR].
[Tra]
H.C. Tran. Geometric embedding properties of Bestvina–Brady subgroups. To
appear in Algebr. Geom. Topol. arXiv:1601.00539.
[Wis12] D.T. Wise. From riches to RAAGs: 3–manifolds, right-angled Artin groups, and
cubical geometry, volume 117 of CBMS Regional Conference Series in Mathematics. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2012.
[Woo16] D.J. Woodhouse. Classifying finite dimensional cubulations of tubular groups.
Michigan Math. J., 65(3):511–532, 2016.
[WY97] S. Wang and F. Yu. Graph manifolds with non-empty boundary are covered by
surface bundles. Math. Proc. Cambridge Philos. Soc., 122(3):447–455, 1997.
Department of Mathematical Sciences, University of Wisconsin–Milwaukee,
P.O. Box 413, Milwaukee, WI 53201, USA
E-mail address: [email protected], [email protected]
| 4math.GR
|
arXiv:1412.4524v2 [math.GR] 1 Aug 2016
THE NIELSEN AND REIDEMEISTER THEORIES OF
ITERATIONS ON INFRA-SOLVMANIFOLDS OF TYPE (R) AND
POLY-BIEBERBACH GROUPS
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Abstract. We study the asymptotic behavior of the sequence of the Nielsen
numbers {N (f k )}, the essential periodic orbits of f and the homotopy minimal
periods of f by using the Nielsen theory of maps f on infra-solvmanifolds
of type (R). We develop the Reidemeister theory for the iterations of any
endomorphism ϕ on an arbitrary group and study the asymptotic behavior
of the sequence of the Reidemeister numbers {R(ϕk )}, the essential periodic
[ϕ]-orbits and the heights of ϕ on poly-Bieberbach groups.
1. Introduction
Let f : X → X be a map on a connected compact polyhedron X. A point
x ∈ X is a fixed point of f if f (x) = x and is a periodic point of f with period n if
f n (x) = x. The smallest period of a periodic x is called the minimal period. We
will use the following notations:
Fix(f ) = {x ∈ X | f (x) = x},
Per(f ) = the set of all minimal periods of f ,
Pn (f ) = the set of all periodic points of f with minimal period n,
\
HPer(f ) =
{n ∈ N | Pn (g) 6= ∅}
g≃f
= the set of all homotopy minimal periods of f .
Let p : X̃ → X be the universal covering projection onto X and f˜ : X̃ → X̃ a
fixed lift of f . Let Π be the group of covering transformations of the projection
p : X̃ → X. Then f induces an endomorphism ϕ = ϕf : Π → Π by the following
identity
ϕ(α)f˜ = f˜α, ∀α ∈ Π.
The subsets p(Fix(αf˜)) ⊂ Fix(f ), α ∈ Π, are called fixed point classes of f . A
fixed point class is called essential if its index is nonzero. The number of essential
fixed point classes is called the Nielsen number of f , denoted by N (f ) [27].
Date: XXX 1, 2014 and, in revised form, YYY 22, 2015.
2000 Mathematics Subject Classification. Primary 37C25; Secondary 55M20.
Key words and phrases. Infra-solvmanifold, Nielsen number, Nielsen zeta function, periodic
[ϕ]-orbit, poly-Bieberbach group, Reidemeister number, Reidemeister zeta function.
The second author is supported in part by Basic Science Researcher Program through
the National Research Foundation of Korea(NRF) funded by the Ministry of Education
(No. 2013R1A1A2058693) and by the Sogang University Research Grant of 2010 (10022).
1
2
ALEXANDER FEL’SHTYN AND JONG BUM LEE
The Nielsen number is always finite and is a homotopy invariant lower bound for
the number of fixed points of f . In the category of compact, connected polyhedra
the Nielsen number of a map is, apart from in certain exceptional cases, equal to
the least number of fixed points of maps with the same homotopy type as f .
Let ϕ : Π → Π be an endomorphism on an arbitrary group Π. Consider the
Reidemeister action of Π on Π determined by the endomorphism ϕ and defined
as follows:
Π × Π −→ Π, (γ, α) 7→ γαϕ(γ)−1 .
The Reidemeister class containing α will be denoted by [α], and the set of Reidemeister classes of Π determined by ϕ will be denoted by R[ϕ]. Write R(ϕ) = #R[ϕ],
called the Reidemeister number of ϕ. When the endomorphism ϕ : Π → Π is
induced from a self-map f : X → X, i.e., when ϕ = ϕf , we also refer to R[ϕ] as the
set R[f ] of Reidemeister classes of f , and R(ϕ) as the Reidemeister number R(f )
of f .
It is easy to observe that if ψ is an automorphism on Π, then ψ sends the
Reidemeister class [α] of ϕ to the Reidemeister class [ψ(α)] of ψϕψ −1 . Hence
the Reidemeister number is an automorphism invariant. For any β ∈ Π, let τβ
denote the inner automorphism determined by β. We will compare R[ϕ] with
R[τβ ϕ]. Observe that the right multiplication rβ −1 by β −1 on Π induces a bijection
R[ϕ] → R[τβ ϕ], [α] 7→ [αβ −1 ]. Indeed,
rβ −1 : γ · α · ϕ(γ)−1 7−→ (γ · α · ϕ(γ)−1 )β −1
= γ · (αβ −1 ) · βϕ(γ)−1 β −1 = γ · (αβ −1 ) · (τβ ϕ)(γ)−1 .
Similarly, we can show that r(βϕ(β)···ϕn−1 (β))−1 induces a bijection R[ϕn ] → R[(τβ ϕ)n ],
[α]n 7→ [α(βϕ(β) · · · ϕn−1 (β))−1 ]n . Hence the Reidemeister number is a conjugacy
invariant. This is not surprising because if f and g are homotopic, then their
induced endomorphisms differ by an inner automorphism τβ .
The set Fix(f n ) of periodic points of f splits into a disjoint union of periodic
point classes p(Fix(αf˜n )) of f , and these sets are indexed by the Reidemeister
classes [α]n ∈ R[ϕn ] of the endomorphism ϕn where ϕ = ϕf . Namely,
a
(D)
p Fix(αf˜n ) .
Fix(f n ) =
[α]n ∈R[ϕn ]
From the dynamical point of view, it is natural to consider the Nielsen numbers
N (f k ) and the Reidemeister numbers R(f k ) of all iterations of f simultaneously.
For example, N. Ivanov [21] introduced the notion of the asymptotic Nielsen number, measuring the growth of the sequence N (f k ), and found the basic relation
between the topological entropy of f and the asymptotic Nielsen number. Later
on, it was suggested in [11, 42, 12, 13, 14] to arrange the Nielsen numbers N (f k ),
the Reidemeister numbers R(f k ) and R(ϕk ) of all iterations of f and ϕ into the
Nielsen and the Reidemeister zeta functions
!
∞
X
N (f k ) k
z ,
Nf (z) = exp
k
k=1
Rf (z) = exp
∞
X
R(f k )
k=1
k
zk
!
,
Rϕ (z) = exp
∞
X
R(ϕk )
k=1
k
zk
!
.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
3
The Nielsen and Reidemeister zeta functions are nonabelian analogues of the Lefschetz zeta function
!
∞
X
L(f k ) k
Lf (z) = exp
z ,
k
k=1
where
L(f n ) :=
dim
XX
k=0
h
i
n
(−1)k tr f∗k
: Hk (X; Q) → Hk (X; Q)
is the Lefschetz number of the iterate f n of f .
Nice analytic properties of Nf (z) [14] indicate that the numbers N (f k ), k ≥ 1,
are closely interconnected. Other manifestations of this are Gauss congruences
X k
N (f d ) ≡ 0 mod k,
µ
d
d|k
for any k > 0, where f is a map on an infra-solvmanifold of type (R) [15]. Whenever
all R(f k ) are finite, we also have
X k
X k
µ
µ
R(f d ) =
N (f d ) ≡ 0 mod k.
d
d
d|k
d|k
It is known that the Reidemeister numbers of the iterates of an automorphism ϕ
of an almost polycyclic group also satisfy Gauss congruences [17, 18].
The fundamental invariants of f used in the study of periodic points are the
Lefschetz numbers L(f k ), and their algebraic combinations, the Nielsen numbers
N (f k ) and the Nielsen–Jiang periodic numbers N Pn (f ) and N Φn (f ), and the Reidemeister numbers R(f k ) and R(ϕk ).
The study of periodic points by using the Lefschetz theory has been done extensively by many authors in the literature such as [27], [9], [2], [24], [41]. A natural
question is to ask how much information we can get about the set of essential periodic points of f or about the set of (homotopy) minimal periods of f from the
study of the sequence {N (f k )} of the Nielsen numbers of iterations of f . Utilizing the arguments employed mainly in [2] and [24, Chap. III] for the Lefschetz
numbers of iterations, we study the asymptotic behavior of the sequence {N (f k )},
the essential periodic orbits of f and the homotopy minimal periods of f by using
the Nielsen theory of maps f on infra-solvmanifolds of type (R). We will give a
brief description of the main results in Section 4 whose details and proofs can be
found in [16]. From the identity (D), the Reidemeister theory for the iterations of
f is almost parallel to the Nielsen theory of the iterates of f . Motivated from this
parallelism, we will develop in Section 2 the Reidemeister theory for the iterations
of any endomorphism ϕ on an arbitrary group Π. In this paper, we will study the
asymptotic behavior of the sequence {R(ϕk )}, the essential periodic [ϕ]-orbits and
the heights of ϕ on poly-Bieberbach groups. We refer to [15, 16] for background to
our present work.
Acknowledgments. The first author is indebted to the Max-Planck-Institute for
Mathematics(Bonn) and Sogang University(Seoul) for the support and hospitality
and the possibility of the present research during his visits there. The authors also
would like to thank Thomas Ward for making careful corrections and suggestions
to a few expressions.
4
ALEXANDER FEL’SHTYN AND JONG BUM LEE
2. Preliminaries
Recall that the periodic point set Fix(f n ) splits into a disjoint union of periodic
point classes
G
p Fix(αf˜n ) .
Fix(f n ) =
[α]n ∈R[ϕn ]
Consequently, there is a 1-1 correspondence η from the set of periodic point classes
p(Fix(αf˜n )) to the set of Reidemeister classes [α]n of ϕn . When m | n, Fix(f m ) ⊂
Fix(f n ). Let x ∈ Fix(f m ) and x̃ ∈ p−1 (x). Then there exist unique α, β ∈ π such
that αf˜m (x̃) = x̃ and β f˜n (x̃) = x̃. It can be easily derived that
β = αϕm (α)ϕ2m (α) · · · ϕn−m (α).
This defines two natural functions, called boosting functions,
γm,n : p Fix(αf˜m ) 7→ p Fix(αϕm (α)ϕ2m (α) · · · ϕn−m (α)f˜n ) ,
ιm,n = ιm,n (ϕ) : [α]m 7→ [αϕm (α)ϕ2m (α) · · · ϕn−m (α)]n
so that the following diagram is commutative
γm,n
✤
/ p Fix(αϕm (α)ϕ2m (α) · · · ϕn−m (α)f˜n )
p Fix(αf˜m )
❴
❴
η
η
✤
[α]m
ιm,n
/ [αϕm (α)ϕ2m (α) · · · ϕn−m (α)]n
Moreover, it is straightforward to check the commutativity of the diagram
[α]m
❴
r(βϕ(β)···ϕm−1 (β))−1
/ [α(βϕ(β) · · · ϕm−1 (β))−1 ]m
❴
✤
ιm,n (ϕ)
[αϕm (α) · · · ϕn−m (α)]n ✤
ιm,n (τβ ϕ)
r(βϕ(β)···ϕn−1 (β))−1
/ [αϕm (α) · · · ϕn−m (α)(βϕ(β) · · · ϕm−1 (β))−1 ]n
On the other hand, for x ∈ p(Fix(αf˜n )) we choose x̃ ∈ p−1 (x) so that αf˜n (x̃) =
x̃. Then
ϕ(α)f˜n f˜(x̃) = ϕ(α)f˜f˜n (x̃) = f˜αf˜n (x̃) = f˜(x̃)
and so f (x) ∈ p(Fix(ϕ(α)f˜n )). Namely, p(Fix(ϕ(α)f˜n )) is the periodic point class
determined by f (x). Therefore, f induces a function on the periodic point classes
of f n , which we denote by [f ], defined as follows:
[f ] : p Fix(αf˜n ) 7−→ p Fix(ϕ(α)f˜n ) .
Similarly, ϕ induces a well-defined function on the Reidemeister classes of ϕn , which
we will denote by [ϕ], given by [ϕ] : [α]n 7→ [ϕ(α)]n . Then the following diagram
commutes:
[f ]
✤
/ p Fix(ϕ(α)f˜n )
p Fix(αf˜n )
O
O
[α]n ✤
[ϕ]
/ [ϕ(α)]n
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
5
By [27, Theorem III.1.12], [f ] is an index-preserving bijection on the periodic point
classes of f n . We say that [α]n is essential if the corresponding class p(Fix(αf˜n ))
is essential. Evidently,
Fix(αf˜n )
f˜
˜n−1
/ Fix(ϕ(α)f˜n ) αf
/ Fix(αf˜n ).
5
identity
This implies that for each α ∈ Π, the restrictions of f
f | : p Fix(αf˜n ) −→ p Fix(ϕ(α)f˜n )
are homeomorphisms such that [f ]n is the identity. In particular,
p Fix(αf˜n ) = ∅ ⇐⇒ p Fix(ϕ(α)f˜n ) = ∅.
Moreover, [ϕ]n is the identity, ιm,n ◦ [ϕ] = [ϕ] ◦ ιm,n and γm,n ◦ [f ] = [f ] ◦ γm,n .
The length of the element [α]n ∈ R[ϕn ], denoted by ℓ([α]n ), is the smallest
positive integer ℓ such that [ϕ]ℓ ([α]n ) = [α]n . The [ϕ]-orbit of [α]n is the set
h[α]n i = {[α]n , [ϕ]([α]n ), · · · , [ϕ]ℓ−1 ([α]n )},
where ℓ = ℓ([α]n ). We must have that ℓ | n. The element [α]n ∈ R[ϕn ] is reducible
to m if there exists [β]m ∈ R[ϕm ] such that ιm,n ([β]m ) = [α]n . Note that if [α]n is
reducible to m, then m | n. If [α]n is not reducible to any m < n, we say that [α]n
is irreducible. The depth of [α]n , denoted by d([α]n ), is the smallest integer m
to which [α]n is reducible. Since clearly d([α]n ) = d([ϕ]([α]n )), we can define the
depth of the orbit h[α]n i: d(h[α]n i) = d([α]n ). If n = d([α]n ), the element [α]n or
the orbit h[α]n i is called irreducible.
Clearly, as a set p(Fix(αf˜m )) ⊂ p(Fix(αϕm (α)ϕ2m (α) · · · ϕn−m (α)f˜n )). This
implies that if p(Fix(αf˜m )) is the periodic point class of f m determined by x, then
p Fix(αϕm (α)ϕ2m (α) · · · ϕn−m (α)f˜n )
is the periodic point class of f n determined by x.
Note that if [α]n is irreducible, then every element of the fixed point class
p(Fix(αf˜n )) is a periodic point of f with minimal period n. Let [α]n be an essential class with depth m and let ιm,n ([β]m ) = [α]n . Then there is a periodic point
x of f with minimal period m. Consequently, the irreducibility of a periodic Reidemeister class of ϕ is an algebraic counterpart of the minimal period of a periodic
point of f . We say that a periodic Reidemeister class [α]n of ϕ has height n if it
is irreducible. The set IR(ϕn ) of all classes in R[ϕn ] with height n is an algebraic
analogue of the set Pn (f ) of periodic points of f with minimal period n. Let I(ϕ)
be the set of all irreducible classes of ϕ. That is,
I(ϕ) = {[α]k ∈ R[ϕk ] | α ∈ Π, k > 0, [α]k is irreducible}.
We define the set H(ϕ) of all heights of ϕ to be
H(ϕ) = {k ∈ N | some [α]k has height k}.
Then H(ϕ) is an algebraic analogue of the set Per(f ) of all minimal periods of
f . Motivated from homotopy minimal periods of f , we may define the set of all
homotopy heights of ϕ as follows:
\
HI(ϕ) =
{n ∈ N | IR((τβ ϕ)n ) 6= ∅} .
β∈Π
6
ALEXANDER FEL’SHTYN AND JONG BUM LEE
However, as we have observed before, since the boosting functions ιm,n commute
with “right multiplications”, i.e.,
ιm,n (τβ ϕ) ◦ r(βϕ(β)···ϕm−1 (β))−1 = r(βϕ(β)···ϕn−1 (β))−1 ◦ ιm,n (ϕ),
it follows that the height is a conjugacy invariant. Consequently, we have HI(ϕ) =
H(ϕ).
3. Poly-Bieberbach groups
The fundamental group of an infra-solvmanifold is called a poly-Bieberbach
group, which is a torsion free poly-crystallographic group. It is known (see for
example [10, Theorem 2.12]) that every poly-Bieberbach group is a torsion-free
virtually poly-Z group. We refer to [46, Theorem 3] for a characterization of
poly-crystallographic groups. Recall also from [46, Corollary 4] that for any polyBieberbach group Π there exist a connected simply connected supersolvable Lie
group S, a compact subgroup K of Aut(S) and an isomorphism ι of Π onto a discrete cocompact subgroup of S ⋊ K such that ι(Π) · S is dense in S ⋊ K. By [16,
Lemma 2.1], the supersolvable Lie groups are the Lie groups of type (R), that is,
Lie groups for which if ad X : S → S has only real eigenvalues for all X in the Lie
algebra S of S. Assuming ι to be an inclusion or identifying Π with ι(Π), we have
the following commutative diagram
1 −−−−→
S
x
p
−−−−→ S ⋊ K −−−−→
x
1 −−−−→ Π ∩ S −−−−→
Π
K
x
−−−−→ 1
−−−−→ p(Π) −−−−→ 1
Here, we cannot assume that the subgroup p(Π) of K ⊂ Aut(S) is a finite group
and that the translations Π ∩ S form a lattice in the solvable Lie group S of type
(R).
In this paper, we will assume the following: Let Π be a poly-Bieberbach group
which is the fundamental group of an infra-solvmanifold of type (R), i.e., Π is a
discrete cocompact subgroup of Aff(S) := S ⋊ Aut(S), where S is a connected,
simply connected solvable Lie group of type (R) and Π ∩ S is of finite index in Π
and a lattice of S. The finite group Φ := Π/Π∩S is called the holonomy group of
the poly-Bieberbach group Π or the infra-solvmanifold Π\S of type (R). Naturally
Φ sits in Aut(S). Let ϕ : Π → Π be an endomorphism. Then by [33, Theorem 2.2],
ϕ is semi-conjugate by an “affine map”. Namely, there exist d ∈ S and a Lie group
endomorphism D : S → S such that ϕ(α)(d, D) = (d, D)α for all α ∈ Π ⊂ Aff(S).
From this identity condition, the affine map f˜ := (d, D) : S → S restricts to a map
f : Π\S → Π\S for which it induces the endomorphism ϕ. Conversely, if f is a
self-map on an infra-solvmanifold Π\S of type (R), f induces an endomorphism
ϕ = ϕf , see Section 1. As remarked above, f is homotopic to a map induced by an
affine map on S. Since the Lefschetz, Nielsen and Reidemeister numbers of f are
homotopy invariants, we may assume that our f has an affine lift (d, D) on S.
Theorem 3.1 ([15, Corollary 7.6]). Let ϕ : Π → Π be an endomorphism on a
poly-Bieberbach group Π of S with holonomy group Φ. If ϕ is the semi-conjugate
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
7
by an affine map (d, D) on S, then we have
1 X
R(ϕk ) =
σ det(I − A∗ D∗k )
#Φ
A∈Φ
where σ : R → R ∪ {∞} is given by σ(0) = ∞ and σ(x) = |x| for all x 6= 0.
Furthermore, if R(ϕk ) < ∞ then R(ϕk ) = N (f k ) where f is a map on Π\S which
induces ϕ.
When all R(ϕk ) are finite, Theorem 3.1 says that the Reidemeister theory for
poly-Bieberbach groups follows directly from the Nielsen theory for infra-solvmanifolds
of type (R). In this paper, whenever possible, we will state our results in the language of Reidemeister theory.
Proposition 3.2 ([15, Proposition 9.3]). Let f be a map on an infra-solvmanifold
Π\S of type (R) induced by an affine map. Then every essential fixed point class
of f consists of a single element.
Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group Π. We assume
as before that (d, D) be an affine map on S and f : Π\S → Π\S be the map
induced by (d, D) and inducing ϕ. We assume further that all R(ϕn ) < ∞. Hence
by Theorem 3.1, R(ϕn ) = N (f n ) for all n > 0. This implies that for every n > 0
all fixed point classes of f n are essential and hence consist of a single element by
Proposition 3.2. Consequently, we can refer to essential fixed point classes of f n
as essential periodic points of f with period n. Moreover, for every n > 0 all
Reidemeister classes of ϕn are essential.
For m | n and for β ∈ Π, let α = βϕm (β) · · · ϕn−m (β) and consider the commuting diagram
η
/ p Fix(α(α)f˜n ) = {x}
[α]n ✤
O
O
ιm,n
❴
✤
[β]m
γm,n
η
❴
/ p Fix(β f˜m ) = {x}
This shows that the observation in Section 2 can be refined as follows: [α]n is
irreducible if and only [α]n has height n if and only if the corresponding essential
periodic point x of f has minimal period n. Moreover, [α]n has depth d if and only
if the corresponding essential periodic point x of f has minimal period d. Let ℓ
be the length of [α]n . That is, [ϕℓ (α)]n = [α]n . Equivalently, we have f ℓ (x) = x.
This implies that [α]n is reducible to ℓ. Further, d = ℓ. In particular, if [α]n is
irreducible, then its length is the height, ℓ = n, and so #h[α]n i = n.
We denote by O([ϕ], k) the set of all (essential) periodic orbits of [ϕ] with length
≤ k. Then we have
O([ϕ], k) = {h[α]m i | α ∈ Π, m ≤ k}
= {hxi | x is an essential periodic point of f with length ≤ k}
= O(f, k).
Recall that the set of essential periodic points of f with minimal period k is
[
Fixe (f d ).
EPk (f ) = Fixe (f k ) −
d|k,d<k
8
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Then we have the algebraic counterpart. Namely,
EPk (ϕ) = {[α]k ∈ R[ϕk ] | [α]k is irreducible (and essential)}
= {[α]k ∈ R[ϕk ] | [α]k has height k}
= IR(ϕk ),
and #IR(ϕk ) = #EPk (f ). Hence the set of (essential) Reidemeister classes of
ϕk can be identified with a disjoint union of irreducible classes, that is, R[ϕk ] is
decomposed by heights:
G
IR(ϕd ),
R[ϕk ] =
d|k
and hence its cardinality satisfies
R(ϕk ) = #R[ϕk ] =
X
#IR(ϕd ).
d|k
Recall that if we denote by Ok (ϕ) the number of essential and irreducible periodic
orbits of R[ϕk ], i.e., if Ok (ϕ) = #{h[α]k i | [α]k ∈ EPk (ϕ)} then by definition, the
prime Nielsen–Jiang periodic number of period k is
NPk (ϕ) = k × Ok (ϕ).
As observed earlier, each such orbit h[α]k i has length k. Therefore, NPk (ϕ) =
#IR(ϕk ) and topologically NPk (f ) = #EPk (f ), the number of essential periodic
points of f with minimal period k.
Theorem 3.3. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Then
NPk (ϕ) = #IR(ϕk ).
Let f : Π\S → Π\S be a map on an infra-solvmanifold Π\S of type (R). Then
NPk (f ) = #EPk (f ).
Consider all periodic orbits h[α]m i of ϕ with m | k. A set S of periodic orbits
h[α]m i of ϕ with m | k is said to be a set of k-representatives if every essential orbit
h[β]m i with m | k is reducible to some element of S. Then the full Nielsen–Jiang
periodic number of period k is defined to be
(
)
X
NFk (ϕ) = min
d(A) | S is a set of k-representatives .
A∈S
Recall that all periodic orbits are essential, and every periodic orbit h[β]m i is
boosted to a periodic orbit h[α]k i. Hence to compute NFk (ϕ), we need first to
consider only R[ϕk ] and a set of representatives [αi ]k of the orbits in R[ϕk ]. Then
NFk (ϕ) is the sum of depths of all [αi ]n . Topologically, NFk (f ) is the sum of
minimal periods of all essential periodic point classes hxi in Fixe (f k ).
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
9
4. The Nielsen theory of iterations on an infra-solvmanifold of type
(R)
In this section, we shall assume that f : M → M is a continuous map on an
infra-solvmanifold M = Π\S of type (R) with holonomy group Φ. Then f admits
an affine homotopy lift (d, D) : S → S. Concerning the Nielsen numbers N (f k ) of
all iterates of f , we begin with the following facts:
Averaging Formula: ([33, Theorem 4.2])
N (f k ) =
1 X
| det(I − A∗ D∗k )|.
#Φ
A∈Φ
Gauss Congruences: ([15, Theorem 11.4])
X k
N (f d ) ≡ 0
µ
d
mod k.
d|k
Indeed, we have shown in [15, Theorem 11.4] that the left-hand side is non-negative
because it is equal to the number of isolated periodic points of f with least period
k. By [43, Lemma 2.1], the sequence {N (f k )} is exactly realizable.
Rationality of Nielsen zeta function: ([7, Theorem 4.5],[15])
!
∞
X
N (f k ) k
z
Nf (z) = exp
k
k=1
is a rational function with coefficients in Q.
Using these facts as our main tools, we study the asymptotic behavior of the
sequence {N (f k )}, the essential periodic orbits of f and the homotopy minimal
periods of f by using the Nielsen theory of maps f on infra-solvmanifolds of type
(R). We will give a brief description of the main results in this section whose details
and proofs can be found in [16].
Consider the sequences of algebraic multiplicities {Ak (f )} and Dold multiplicities
{Ik (f )} associated to the sequence {N (f k )}:
X k
1X
k
d
Ak (f ) =
µ
µ
N (f ), Ik (f ) =
N (f d ).
k
d
d
d|k
d|k
Then Ik (f ) = kAk (f ) and all Ak (f ) are integers by (DN). From the Möbius
inversion formula, we immediately have
X
(M)
d Ad (f ).
N (f k ) =
d|k
On the other hand, since Nf (0) = 1 by definition, z = 0 is not a zero nor a pole
of the rational function Nf (z). Thus we can write
Q
r
Y
(1 − βi z)
u(z)
= Qi
=
(1 − λi z)−ρi
Nf (z) =
v(z)
j (1 − γj z)
i=1
10
ALEXANDER FEL’SHTYN AND JONG BUM LEE
with all λi distinct nonzero algebraic integers (see for example [5] or [2, Theorem 2.1]) and ρi nonzero integers. This implies that
r(f )
N (f k ) =
X
ρi λki .
i=1
Note that r(f ) is the number of zeros and poles of Nf (z). Since Nf (z) is a homotopy
invariant, so is r(f ).
We define
λ(f ) = max{|λi | | i = 1, · · · , r(f )}.
The number λ(f ) will play a similar role as the “essential spectral radius” in [24]
or the “reduced spectral radius” in [2].
The following theorem shows that 1/λ(f ) is the “radius” of the Nielsen zeta
function Nf (z). Note also that λ(f ) is a homotopy invariant.
Theorem 4.1 ([16, Theorem 3.2]). Let f be a map on an infra-solvmanifold of type
(R) with an affine homotopy lift (d, D). Let R denote the radius of convergence of
the Nielsen zeta function Nf (z) of f . Then λ(f ) = 0 or λ(f ) ≥ 1, and
1
= λ(f ).
R
In particular, R > 0. If D∗ has no eigenvalue 1, then
^
1
= sp
D∗ = λ(f ).
R
Next, we recall the asymptotic behavior of the Nielsen numbers of iterates of
maps.
Theorem 4.2 ([16, Theorem 4.1]). For a map f of an infra-solvmanifold of type
(R), one of the following three possibilities holds:
(1) λ(f ) = 0, which occurs if and only if Nf (z) ≡ 1.
(2) The sequence
(f k )/λ(f )k } has the same limit points as a periodic seP {N
k
quence { j αj ǫj } where αj ∈ Z, ǫj ∈ C and ǫqj = 1 for some q > 0.
(3) The set of limit points of the sequence {N (f k )/λ(f )k } contains an interval.
In order to give an estimate from below for the number of essential periodic
orbits of maps on infra-solvmanifolds of type (R), we recall the following:
Theorem 4.3 ([45]). If f : M → M is a C 1 -map on a smooth compact
manifold
S
M and {L(f k )} is unbounded, then the set of periodic points of f , k Fix(f k ), is
infinite.
This theorem is not true for continuous maps. Consider the one-point compactification of the map of the complex plane f (z) = 2z 2 /||z||. This is a continuous
degree two map of S 2 with only two periodic points but with L(f k ) = 2k+1 .
However, when M is an infra-solvmanifold of type (R), the theorem is true for
all continuous maps f on M . In fact, using the averaging formula, we obtain
1 X
|L(f k )| ≤
| det(I − A∗ D∗k )| = N (f k ).
|Φ|
A∈Φ
k
If L(f ) is unbounded, then so is N (f k ) and hence the number of essential fixed
point classes of all f k is infinite.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
11
Recall that any map f on an infra-solvmanifold of type (R) is homotopic to a
map f¯ induced by an affine map (d, D). By [15, Proposition 9.3], every essential
fixed point class of f¯ consists of a single element x with index sign det(I − dfx ).
Hence N (f ) = N (f¯) is the number of essential fixed point classes of f¯. It is a
classical fact that a homotopy between f and f¯ induces a one-one correspondence
between the fixed point classes of f and those of f¯, which is index preserving.
Consequently, we obtain
|L(f k )| ≤ N (f k ) ≤ #Fix(f k ).
This suggests the following conjectural inequality (see [44, 45]) for infra-solvmanifolds
of type (R):
1
1
lim sup log |L(f k )| ≤ lim sup log #Fix(f k ).
k→∞ k
k→∞ k
We denote by O(f, k) the set of all essential periodic orbits of f with length
≤ k. Thus O(f, k) = {hFi | F is a essential fixed point class of f m with m ≤ k}.
We can strengthen Theorem 4.3 as follows:
Theorem 4.4. Let f be a map on an infra-solvmanifold of type (R). Suppose that
the sequence N (f k ) is unbounded. Then there exists a natural number N0 such that
k ≥ N0 =⇒ #O(f, k) ≥
k − N0
.
r(f )
Proof. As mentioned earlier, we may assume that every essential fixed point class
F of any f k consists of a single element F = {x}. Denote by Fixe (f k ) the set of
essential fixed point (class) of f k . Thus N (f k ) = #Fixe (f k ). Recalling also that f
acts on the set Fixe (f k ) from the proof of [15, Theorem 11.4], we have
O(f, k) = {hxi | x is a essential periodic point of f with length ≤ k}.
Observe further that if x is an essential periodic point of f with least period p,
then x ∈ Fixe (f q ) if and only if p | q. The length of the orbit hxi of x is p, and
[
Fixe (f d ),
Fixe (f k ) =
d|k
Fixe (f d )
Recalling that
Am (f ) =
\
′
′
Fixe (f d ) = Fixe (f gcd(d,d ) ).
1 X m
1 X m
µ
µ
N (f k ) =
#Fixe (f k ),
m
k
m
k
k|m
k|m
Fixe (f i ) to be
1 X m
µ
# hxi ∩ Fixe (f k ) .
Am (f, hxi) =
m
k
we define Am (f, hxi) for any x ∈
S
i
k|m
Then we have
Am (f ) =
X
hxi
x∈Fixe (f m )
Am (f, hxi).
12
ALEXANDER FEL’SHTYN AND JONG BUM LEE
We begin with new notation. For a given integer k > 0 and x ∈
let
S
m
Fixe (f m ),
A(f, k) = {m ≤ k | Am (f ) 6= 0} ,
A(f, hxi) = {m | Am (f, hxi) 6= 0} .
Notice that if Am (f ) 6= 0 then there exists an essential periodic point x of f with
period m such that Am (f, hxi) 6= 0. Consequently, we have
[
A(f, hxi)
A(f, k) ⊂
hxi∈O(f,k)
Since N (f k ) is unbounded, we have that λ(f ) > 1 by the definition of λ(f ). By
[16, Corollary 4.6], there is N0 such that if n ≥ N0 then there is i with n ≤ i ≤
n + n(f ) − 1 such that Ai (f ) 6= 0. This leads to the estimate
#A(f, k) ≥
k − N0
n(f )
∀k ≥ N0 .
Assume that x has least period p. Then we have
1 X m
p X m
Am (f, hxi) =
µ
#hxi =
µ
.
m
n
m
n
p|n|m
p|n|m
Thus if m is not a multiple of p then by definition Am (f, hxi) = 0. It is clear that
′
′
Ap (f, hxi) = µ(1) = 1, i.e.,Pp ∈ A(f, hxi). Because
Pp | n | rp ′⇔ n = r p with r | r,
we have Arp (f, hxi) = 1/r p|n|rp µ(rp/n) = 1/r r′ |r µ(r/r ) which is 0 when and
only when r > 1. Consequently, A(f, hxi) = {p}.
In conclusion, we obtain the required inequality
k − N0
≤ #A(f, k) ≤ #O(f, k).
r(f )
Finally, we study (homotopy) minimal periods of maps f on infra-solvmanifolds
of type (R). We seek to determine HPer(f ) only from the knowledge of the sequence
{N (f k )}. This approach was used in [1, 19, 28] for maps on tori, in [25, 26, 23,
24, 35, 36] for maps on nilmanifolds and some solvmanifolds, and in [32, 34] for
expanding maps on infra-nilmanifolds.
Utilizing new results obtained from the Gauss congruences and the rationality
of the Nielsen zeta function, together with Dirichlet’s prime number theorem, we
obtain:
Theorem 4.5. Let f be a map on an infra-solvmanifold of type (R). Suppose
that the sequence {N (f k )/λ(f )k } is asymptotically periodic (i.e., Case (2) of Theorem 4.2). Then there exist m and an infinite sequence {pi } of primes such that
{mpi } ⊂ Per(f ). Furthermore, {mpi } ⊂ HPer(f ).
Next we recall that:
Theorem 4.6 ([22, Theorem 6.1]). Let f : M → M be a self-map on a compact
PL-manifold of dimension ≥ 3. Then f is homotopic to a map g with Pn (g) = ∅ if
and only if NPn (f ) = 0.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
13
The infra-solvmanifolds of dimension 1 or 2 are the circle, the torus and the
Klein bottle. Theorem 4.6 for dimensions 1 and 2 is verified respectively in [3], [1]
and [38, 23, 31]. Hence we have
n ∈ HPer(f ) ⇐⇒ ∃ g ≃ f such that Pn (f ) 6= ∅
⇐⇒ NPn (f ) 6= 0
(Definition)
(Theorem 4.6)
⇐⇒ EPn (f ) 6= ∅ (Theorem 3.3)
⇐⇒ In (f ) 6= 0
([16, Proposition 5.4])
⇐⇒ An (f ) 6= 0
With the identity (M), we have the following result.
Theorem 4.7. Let f be a map on an infra-solvmanifold of type (R). Then
HPer(f ) = {k | Ak (f ) 6= 0} ⊂ {k | N (f k ) 6= 0}.
Moreover, if N (f k ) 6= 0, then there exists a divisor d of k such that d ∈ HPer(f ).
Corollary 4.8. Let f be a map on an infra-solvmanifold of type (R). Suppose that
the sequence {N (f k )} is strictly monotone increasing. Then HPer(f ) is cofinite.
Proof. By the assumption, we have λ(f ) > 1. Thus by [16, Theorem 4.4] (cf.
Theorem 7.3), there exist γ > 0 and N such that if k > N then there exists ℓ =
ℓ(k) < r(f ) such that N (f k−ℓ )/λ(f )k−ℓ > γ. Then for all k > N , the monotonicity
implies that
N (f k−ℓ )
N (f k−ℓ )
γ
γ
N (f k )
≥
=
≥
≥
.
k
k
λ(f )
λ(f )
λ(f )k−ℓ λ(f )ℓ
λ(f )ℓ
λ(f )r(f )
Applying [16, Proposition 4.5] (cf. Proposition 7.4) with ǫ = γ/λ(f )r(f ) , we see
that Ik (f ) 6= 0 and so Ak (f ) 6= 0 for all k sufficiently large. Now our assertion
follows from Theorem 4.7.
Remark 4.9. Note that in the above Corollary we may use the weaker assumption
that the sequence {N (f k )} is eventually strictly monotone increasing, i.e., there
exists k0 > 0 such that N (f k+1 ) > N (f k ) for all k ≥ k0 .
Thus the main result of [34] follows from Corollary 4.8.
Corollary 4.10 ([32, Theorem 4.6], [34, Theorem 3.2]). Let f be an expanding
map on an infra-nilmanifold. Then HPer(f ) is cofinite.
In [8], the authors also discussed homotopy minimal periods for hyperbolic maps
on infra-nilmanifolds. A map f on an infra-nilmanifold with affine homotopy lift
(d, D) is hyperbolic if D∗ has no eigenvalues of modulus 1. We now give another
proof of each of the main results, Theorems 3.9 and 3.16, in [8]. In our proof, we
use some useful results such as Lemma 3.7 and Proposition 3.14 in [8].
Theorem 4.11 ([8, Theorem 3.9]). If f is a hyperbolic map on an infra-nilmanifold
with affine homotopy lift (d, D) such that D∗ is not nilpotent, then HPer(f ) is
cofinite.
Proof. [8, Lemma 3.7] says that if f is such a map, then the sequence {N (f k )} is
eventually strictly monotone increasing; by our Corollary 4.8, HPer(f ) is cofinite.
14
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Theorem 4.12 ([8, Theorem 3.16]). If f is a hyperbolic map on an infra-nilmanifold
with affine homotopy lift (d, D) such that D∗ is nilpotent, then HPer(f ) = {1}.
Proof. If f is such a map, then by [8, Proposition 3.14] all N (f k ) = 1. Now because
of the identity (M), all Ad (f ) = 0 except A1 (f ) = N (f k ) = 1. Hence Theorem 4.7
implies that HPer(f ) = {1}.
5. Reidemeister numbers R(ϕk )
Concerning the Reidemeister numbers R(ϕk ) of all iterates of ϕ, we shall assume
that all R(ϕk ) are finite. Whenever all R(ϕn ) are finite, we can consider the
Reidemeister zeta function of ϕ
!
∞
X
R(ϕk ) k
z .
Rϕ (z) = exp
k
k=1
Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group Π with Π ⊂
Aff(S) = S ⋊ Aut(S), where S is a connected, simply connected solvable Lie group
of type (R). By Section 3, ϕ is a homomorphism induced by a self-map f on the
infra-solvmanifold Π\S of type (R). First we recall the following result.
Theorem 5.1 ([15, Theorem 11.4]). Let ϕ : Π → Π be an endomorphism on a polyBieberbach group Π such that all R(ϕk ) are finite. Then the sequences {R(ϕk )} and
{N (f k )} are exactly realizable and
X k
X k
µ
µ
(DN)
R(ϕd ) ≡
N (f d ) ≡ 0 mod k
d
d
d|k
d|k
for all k > 0.
Consider the sequences of algebraic multiplicities {Ak (f )} and Dold multiplicities
{Ik (f )} associated to the sequence {N (f k )}:
X k
k
1X
µ
µ
N (f d ), Ik (f ) =
N (f d ).
Ak (f ) =
k
d
d
d|k
d|k
Then Ik (f ) = kAk (f ) and all Ak (f ) are integers by (DN). From the Möbius
inversion formula, we immediately have
X
d Ad (f ).
N (f k ) =
d|k
Because we are assuming that all R(ϕk ) are finite, by Theorem 3.1, R(ϕk ) =
N (f k ). Consequently, we obtain the sequences of algebraic multiplicities {Ak (ϕ)}
and Dold multiplicities {Ik (ϕ)} associated to the sequence {R(ϕk )}. Thus Ik (ϕ) =
kAk (ϕ) and all Ak (ϕ) are integers. Furthermore, we immediately have R(ϕk ) =
P
d|k d Ad (ϕ).
Theorem 5.2 ([15, Theorem 7.8]). Let ϕ : Π → Π be an endomorphism on a
poly-Bieberbach group Π such that all R(ϕk ) are finite. Then the Reidemeister zeta
function of ϕ
!
∞
X
R(ϕk ) k
Rϕ (z) = exp
z
k
k=1
is a rational function.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
15
Since Rϕ (0) = 1 by definition, z = 0 is not a zero nor a pole of the rational
function Rϕ (z). Thus we can write
Q
r
Y
u(z)
i (1 − βi z)
Q
(1 − λi z)−ρi
=
=
Rϕ (z) =
v(z)
j (1 − γj z)
i=1
with all λi distinct nonzero algebraic integers (see for example [5] or [2, Theorem 2.1]) and ρi nonzero integers. This implies that
r(ϕ)
k
(R1)
R(ϕ ) =
X
ρi λki .
i=1
Note that r(ϕ) is the number of zeros and poles of Rϕ (z). Since Rϕ (z) is a homotopy
invariant, so is r(ϕ).
Consider another generating function associated to the sequence {R(ϕk )}:
Sϕ (z) =
∞
X
R(ϕk )z k−1 .
k=1
Then it is easy to see that
Sϕ (z) =
d
log Rϕ (z).
dz
Moreover,
Sϕ (z) =
∞ r(ϕ)
X
X
r(ϕ)
ρi λki z k−1
=
X
i=1
k=1 i=1
ρi λi
1 − λi z
is a rational function with simple poles and integral residues, and 0 at infinity. The
rational function Sϕ (z) can be written as Sϕ (z) = u(z)/v(z) where the polynomials
u(z) and v(z) are of the form
u(z) = R(ϕ) +
s
X
ai z i ,
v(z) = 1 +
i=1
t
X
bj z j
j=1
with ai and bj integers, see (3) ⇒ (5), Theorem 2.1 in [2] or [24, Lemma 3.1.31]. Let
ṽ(z) be the conjugate polynomial of v(z), i.e., ṽ(z) = z t v(1/z). Then the numbers
{λi } are the roots of ṽ(z), and r(ϕ) = t.
The following can be found in the proof of (3) ⇒ (5), Theorem 2.1 in [2], see
also [16, Lemma 2.4].
Lemma 5.3. If λi and λj are roots of the rational polynomial ṽ(z) which are
algebraically conjugate (i.e., λi and λj are roots of the same irreducible polynomial),
then ρi = ρj .
Q
Let ṽ(z) = sα=1 ṽα (z) be the decomposition of the monic integral polynomial
P
ṽ(z) into irreducible polynomials ṽα (z) of degree rα . Of course, r = r(ϕ) = sα=1 rα
and
ṽ(z) = z r + b1 z r−1 + b2 z r−2 + · · · + br−1 z + br
=
s
Y
rα −1
rα −2
α
+ bα
+ · · · + bα
(z rα + bα
1z
2z
rα −1 z + brα ) =
α=1
s
Y
α=1
ṽα (z).
16
ALEXANDER FEL’SHTYN AND JONG BUM LEE
(α)
If {λi } are the roots of ṽα (z), then the associated ρ’s are the same ρα . Consequently, we can rewrite (R1) as
!
rα
s
X
X
(α)
(λi )k
R(ϕk ) =
ρα
α=1
=
X
ρα >0
ρ+
α
i=1
rα
X
(α)
(λi )k
i=1
Consider the rα × rα -integral square
0 0
1 0
Mα = ... ...
0 0
0 0
!
−
matrices
···
···
0
0
..
.
···
···
0
1
X
ρ−
α
ρα <0
rα
X
(α)
(λi )k
i=1
!
.
−bα
rα
−bα
rα −1
..
.
.
−bα
2
−bα
1
(α)
The characteristic polynomial is det(zI − Mα ) = ṽα (z)
P and therefore {λi } is the
set of eigenvalues of Mα . This implies that R(ϕk ) = sα=1 ρα tr Mαk . Set
M
M
M+ =
ρ+
M− =
ρ−
α Mα ,
α Mα .
ρα >0
ρα <0
Then
(R2)
k
k
R(ϕk ) = tr M+
− tr M−
= tr (M+
M
−M− )k .
We will show in Proposition 8.2 that if Ak (f ) 6= 0 then N (f k ) 6= 0 and hence
f has an essential periodic point of period k. In the following we investigate some
other necessary conditions under which N (f k ) 6= 0. Recall that
N (f k ) = the number of essential fixed point classes of f k .
If F is a fixed point class of f k , then f k (F) = F and the length of F is the smallest
number p for which f p (F) = F, written p(F). We denote by hFi the f -orbit of F,
i.e., hFi = {F, f (F), · · · , f p−1 (F)} where p = p(F). If F is essential, so is every f i (F)
and hFi is an essential periodic orbit of f with length p(F) and p(F) | k. These
are variations of Corollaries 2.3, 2.4 and 2.5 of [2].
Assuming that all R(ϕk ) are finite, we have
Corollary 5.4. If r(ϕ) 6= 0, then R(ϕi ) 6= 0 for some 1 ≤ i ≤ r(ϕ). In particular,
ϕ has an essential periodic orbit with the length p | i, i ≤ r(ϕ).
Pr(ϕ)
Recalling the identity R(ϕk ) = i=1 ρi λki , we define
r(ϕ)
X
X
X
ρi , M (ϕ) = max
ρi , −
ρj .
ρ(ϕ) =
i=1
ρi ≥0
ρj <0
Corollary 5.5. If ρ(ϕ) = 0 and r(ϕ) ≥ 1, then r(ϕ) ≥ 2 and R(ϕi ) 6= 0 for
some 1 ≤ i < r(ϕ). In particular, ϕ has an essential periodic orbit with the length
p | i, i ≤ r(ϕ) − 1.
Corollary 5.6. If r(ϕ) > 0, then R(ϕi ) 6= 0 for some 1 ≤ i ≤ M (ϕ). In particular,
ϕ has an essential periodic orbit with the length p | i, i ≤ M (ϕ).
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
17
6. Radius of convergence of Rϕ (z)
From the Cauchy–Hadamard formula, we can see that the radii R of convergence
of the infinite series Rϕ (z) and Sϕ (z) are the same and given by
1/k
R(ϕk )
1
= lim sup R(ϕk )1/k .
= lim sup
R
k
k→∞
k→∞
We will understand the radius R of convergence from the identity R(ϕk ) =
Pr(ϕ)
−1
k
are the poles or the zeros of the rational function
i=1 ρi λi . Recall that the λi
Rϕ (z). We define
λ(ϕ) = max{|λi | | i = 1, · · · , r(ϕ)}.
If r(ϕ) = 0, i.e., if R(ϕk ) = 0 for all k > 0, then Rϕ (z) ≡ 1 and 1/R = 0. In this
case, we define customarily λ(ϕ) = 0. We shall assume now that r(ϕ) 6= 0. In what
follows, when λ(ϕ) > 0, we consider
n(ϕ) = #{i | |λi | = λ(ϕ)}.
Pr(ϕ)
Remark that if λ(ϕ) < 1 then R(ϕk ) = i=1 ρi λki → 0 and so the sequence of
integers are eventually zero, i.e., R(ϕk ) = 0 for all k sufficiently large. This shows
that 1/R = 0 and furthermore, Rϕ (z) is the exponential of a polynomial. Hence
the rational function Rϕ (z) has no poles and zeros. This forces Rϕ (z) ≡ 1; hence
λ(ϕ) = 0 = 1/R.
Assume |λj | 6= λ(ϕ) for some j; then we have
k
X λi k
λi
R(ϕk ) X
=
ρ
+
ρ
,
lim
= ∞.
ρi
i
j
λj
λj
λkj
i6=j
i6=j
P
It follows from the above observations that 1/R = lim sup( i6=j ρi λki )1/k . ConseP
quently, we may assume that R(ϕk ) = j ρj λkj with all |λj | = λ(ϕ) and then we
have
1/k
1
= lim sup
R
X
|λj |=λ(ϕ)
ρj λkj
.
If λ(ϕ) > 1, then R(ϕk ) → ∞ and by L’Hopital’s rule we obtain
P
k
k
log
ρ
λ
j j j
1
log R(ϕ )
lim sup
= lim sup
= log λ(ϕ) ⇒
= λ(ϕ).
k
k
R
k→∞
k→∞
P
If λ(ϕ) = 1, then R(ϕk ) ≤ j |ρj | < ∞ is a bounded sequence and so it has a
convergent subsequence. If lim sup R(ϕk ) = 0, then R(ϕk ) = 0 for all k sufficiently
large and so by the same reason as above, λ(ϕ) = 0, a contradiction. Hence
lim sup R(ϕk ) is a finite nonzero integer and so 1/R = 1 = λ(ϕ).
Summing up, we have obtained that
Theorem 6.1. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
Π such that all R(ϕk ) are finite. Let R denote the radius of convergence of the
Reidemeister zeta function Rϕ (z) of ϕ. Then λ(ϕ) = 0 or λ(ϕ) ≥ 1, and
1
= λ(ϕ).
R
In particular, R > 0.
18
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Recall that
r(ϕ)
Sϕ (z) =
X
ρi λi
,
1 − λi z
i=1
r(ϕ)
Rϕ (z) =
Y
i=1
−ρi
(1 − λi z)
Q
ρ <0 (1
= Qj
− λj z)−ρj
ρi >0 (1
− λi z)ρi
.
These show that all of the 1/λi are the poles of Sϕ (z), whereas the 1/λi with
corresponding ρi > 0 are the poles of Rϕ (z). The radius of convergence of a power
series centered at a point a is equal to the distance from a to the nearest point where
the power series cannot be defined in a way that makes it holomorphic. Hence the
radius of convergence of Sϕ (z) is 1/λ(ϕ) and the radius of convergence of Rϕ (z) is
1/ max{|λi | | ρi > 0}. In particular, we have shown that
λ(ϕ) = max{|λi | | i = 1, · · · , r(ϕ)} = max{|λi | | ρi > 0}.
Theorem 6.2. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group Π
of S such that all R(ϕk ) are finite. Let R denote the radius of convergence of the
Reidemeister zeta function of ϕ. If ϕ is the semi-conjugate by an affine map (d, D)
on S and if D∗ has no eigenvalue 1, then
^
1
= sp
D∗ = λ(ϕ).
R
Proof. Recall
V that Rϕ (z) = Rf (z) and the radius R of convergence of Rf (z) satisfies
1/R = sp ( D∗ ) by [16, Theorem 3.4]. With Theorem 6.1, we obtain the required
assertion.
We recall that the asymptotic Reidemeister number of ϕ is defined to be
R∞ (ϕ) := max 1, lim sup R(ϕk )1/k .
k→∞
We also recall that the most widely used measure for the complexity of a dynamical
system is the topological entropy h(f ). A basic relation between these two numbers
is h(f ) ≥ log N ∞ (f ), which was found by Ivanov in [21]. There is a conjectural
inequality h(f ) ≥ log(sp(f )) raised by Shub [44]. This conjecture was proven for all
maps on infra-solvmanifolds of type (R), see [39, 40] and [15]. Consider a continuous
map f on a compact connected manifold M , and consider a homomorphism ϕ
induced by f of the group Π of covering transformations on the universal cover of
M . Since M is compact, Π is finitely generated. Let T = {τ1 , · · · , τn } be a set of
generators for Π. For any γ ∈ Π, let L(γ, T ) be the length of the shortest word in
the letters T ∪ T −1 which represents γ. For each k > 0, we put
Lk (ϕ, T ) = max L(ϕk (τi ), T ) | i = 1, · · · , n .
Then the algebraic entropy halg (f ) = halg (ϕ) of f or ϕ is defined as follows:
1
halg (f ) = lim log Lk (ϕ, T ).
k→∞ k
The algebraic entropy of f is well-defined, i.e., independent of the choices of a set
T of generators for Π and a homomorphism ϕ induced by f ([30, p. 114]). We refer
to [30] for the background. We recall that R. Bowen in [4] and A. Katok in [29],
among others, have proved that the topological entropy h(f ) of f is at least as large
as the algebraic entropy halg (ϕ) of ϕ. Furthermore, for any inner automorphism τγ0
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
19
by γ0 , we have halg (τγ0 ϕ) = halg (ϕ) ([30, Proposition 3.1.10]). Now we can make a
statement about the relations between R∞ (ϕ), λ(ϕ), h(f ) and halg (ϕ).
Corollary 6.3. Let ϕ : Π → Π be a homomorphism on a poly-Bieberbach group
Π of S and all R(ϕk ) are finite. Let (d, D) be an affine map on S such that
ϕ(α) ◦ (d, D) = (d, D) ◦ α for all α ∈ Π. Let f¯ be the map on Π\S induced by (d, D)
and let f be any map on Π\S which is homotopic to f¯. Then
^
R∞ (ϕ) = sp
D∗ = λ(ϕ),
halg (ϕ) = halg (f¯) = halg (f ) ≤ h(f¯) = log R∞ (ϕ) ≤ h(f ),
provided that 1 is not an eigenvalue of D∗ .
Proof. From [15,
V Theorem 4.3] and Theorem 6.2, we obtain the first assertion,
R∞ (ϕ) = sp ( D∗ ) = λ(ϕ). By [15, Theorem 5.2], h(f ) ≥ h(f¯) = log λ(ϕ) and
by the remark mentioned just above, we have that h(f¯) ≥ halg (f¯) = halg (f ) =
halg (ϕ).
Remark 6.4. The inequality
log R∞ (ϕ) ≥ halg (ϕ)
in Corollary 6.3 can be regarded as an algebraic analogue of the Ivanov inequality
h(f ) ≥ log N ∞ (f ).
7. Asymptotic behavior of the sequence {R(ϕk )}
In this section, we study the asymptotic behavior of the Reidemeister numbers
of iterates of maps on poly-Bieberbach groups.
Theorem 7.1. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Then one of the following two possibilities holds:
(1) λ(ϕ) = 0, which occurs if and only if Rϕ (z) ≡ 1.
k
(2) The sequence
)/λ(ϕ)k } has the same limit points as a periodic seP {R(ϕ
k
quence { j αj ǫj } where αj ∈ Z, ǫj ∈ C and ǫqj = 1 for some q > 0.
(3) The set of limit points of the sequence {R(ϕk )/λ(ϕ)k } contains an interval.
V
In Theorem 6.2, we showed that if D∗ has no eigenvalue 1 then λ(ϕ) = sp( D∗ ).
In fact, we have the following:
Lemma 7.2. Let ϕ be a homomorphism on a poly-Bieberbach group Π of S and
let ϕ be the
V semi-conjugate by an affine map (d, D) on S. If λ(ϕ) ≥ 1, then
λ(ϕ) = sp( D∗ ).
It is important to know not only the rate of growth of the sequence {R(ϕk )} but
also the frequency with which the largest Reidemeister number is encountered. The
following theorem shows that this sequence grows relatively densely. The following
are variations of Theorem 2.7, Proposition 2.8 and Corollary 2.9 of [2].
Theorem 7.3. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group Π
such that all R(ϕk ) are finite. If λ(ϕ) ≥ 1, then there exist γ > 0 and a natural
number N such that for any m > N there is an ℓ ∈ {0, 1, · · · , n(ϕ) − 1} such that
R(ϕm+ℓ )/λ(ϕ)m+ℓ > γ.
20
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Proposition 7.4. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite and such that λ(ϕ) > 1. Then for any ǫ > 0, there
exists N such that if R(ϕm )/λ(ϕ)m ≥ ǫ for m > N , then the Dold multiplicity
Im (ϕ) satisfies
ǫ
|Im (ϕ)| ≥ λ(ϕ)m .
2
Theorem 7.3 and Proposition 7.4 immediately imply the following:
Corollary 7.5. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite and such that λ(ϕ) > 1. Then there exist γ > 0 and a
natural number N such that if m ≥ N then there exists ℓ with 0 ≤ ℓ ≤ n(ϕ)−1 such
that |Im+ℓ (ϕ)|/λ(ϕ)m+ℓ ≥ γ/2. In particular Im+ℓ (ϕ) 6= 0 and so Am+ℓ (ϕ) 6= 0.
Remark 7.6. We can state a little bit more about the density of the set of algebraic
periods A(ϕ) = {m ∈ N | Am (ϕ) 6= 0}. We consider the notion of the lower
density DA(ϕ) of the set A(ϕ) ⊂ N:
DA(ϕ) = lim inf
k→∞
#(A(ϕ) ∩ [1, k])
.
k
By Corollary 7.5, when λ(ϕ) > 1, we have DA(ϕ) ≥ 1/n(ϕ). On the other hand,
Theorem 7.1 implies the following: If λ(ϕ) = 0 then R(ϕk ) = 0 for all k > 0; by
Theorem 4.7 Ak (ϕ) = 0 for all k > 0, hence DA(ϕ) = ∅. Consider Case (2) of
Theorem 7.1, that is, the sequence {R(ϕk )/λ(ϕ)k } has the same limit points as
Pn(ϕ)
the periodic sequence { j=1 ρj e2iπ(kθj ) } of period q = LCM(q1 , · · · , qn(ϕ) ). By
Theorem 7.3, we have DA(ϕ) ≥ 1/q. Finally considerP
Case (3). Then the sequence
k
k
{R(ϕ )/λ(ϕ) } asymptotically has a subsequence { j∈S ρj e2iπ(kθj ) } where S =
{j1 , · · · , js } and {θj1 , · · · , θjs , 1} is linearly independent over the integers. Therefore
by [6, Theorem 6, p. 91], the sequence (kθj1 , · · · , kθjs ) is uniformly distributed. It
follows that DA(ϕ) = 1.
8. Periodic [ϕ]-orbits
In this section, we shall give an estimate from below the number of periodic
[ϕ]-orbits of an endomorphism ϕ on a poly-Bieberbach group based on facts discussed in Section 3. We keep in mind that all periodic classes are essential, see
Proposition 3.2.
We denote by O([ϕ], k) the set of all (essential) periodic orbits of [ϕ] with length
≤ k. Thus
O([ϕ], k) = {h[α]m i | α ∈ Π, m ≤ k}.
Recalling from Section 3 that O([ϕ], k) = O(f, k), we can restate Theorem 4.4 as
follows:
Theorem 8.1. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Suppose that the sequence R(ϕk ) is unbounded. Then
there exists a natural number N0 such that
k ≥ N0 =⇒ #O([ϕ], k) ≥
k − N0
.
r(ϕ)
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
21
Proposition 8.2. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. For every k > 0, we have
X k
k
R(ϕd ) = Ik (ϕ).
µ
#IR(ϕ ) =
d
d|k
Proof. We apply the Möbius inversion formula to the identity
X
#IR(ϕd )
R(ϕk ) =
d|k
k
in Section 3 to obtain #IR(ϕ ) =
multiplicity Ik (ϕ).
P
d|k
µ
k
d
R(ϕd ), which is exactly the Dold
Definition 8.3. When all R(ϕk ) are finite, we consider the mod 2 reduction of
the Reidemeister number R(ϕk ) of f k , written R(2) (ϕk ). A positive integer k is a
R(2) -period of ϕ if R(2) (ϕk+i ) = R(2) (ϕi ) for all i ≥ 1. We denote the minimal
R(2) -period of ϕ by α(2) (ϕ).
Proposition 8.4 ([41, Proposition 1]). Let p be a prime number and let A be a
square matrix with entries in the field Fp . Then there exists k with (p, k) = 1 such
that
tr Ak+i = tr Ai
for all i ≥ 1.
k
k
Recalling (R2): R(ϕk ) = N (f k ) = tr M+
− tr M−
= tr (M+ ⊕ −M− )k , we can
(2)
(2)
see easily that the minimal R -period α (ϕ) always exists and must be an odd
number.
Now we obtain a result which resembles [41, Theorem 2].
Theorem 8.5. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Let k > 0 be an odd number. Suppose that α(2) (ϕ)2 | k
or p | k where p is a prime such that p ≡ 2i mod α(2) (ϕ) for some i ≥ 0. Then
#IR(ϕk )
NPk (ϕ)
=
k
k
is even.
9. Heights of ϕ
In this section, we study (homotopy) heights HI(ϕ) = H(ϕ) of Reidemeister
classes of endomorphisms ϕ on poly-Bieberbach groups. We wish to determine the
set H(ϕ) of all heights only from the knowledge of the sequence {R(ϕk )}. Recalling
Pr(ϕ)
that when all R(ϕk ) are finite, R(ϕk ) = i=1
ρi λki and λ(ϕ) = max{|λi | | i =
1, · · · , r(ϕ)}, we define
X
1 |λ| k
ρi λki , R̃|λ| (ϕk ) =
R|λ| (ϕk ) =
R (ϕ ).
|λ|k
|λi |=|λ|
Lemma 9.1. When all R(ϕk ) are finite, if λ(ϕ) ≥ 1, then we have
lim sup
k→∞
R(ϕk )
= lim sup |R̃λ(ϕ) (ϕk )|.
λ(ϕ)k
k→∞
22
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Proof. We have
1
R(ϕk )
= R̃λ(ϕ) (f k ) +
k
λ(ϕ)
λ(ϕ)k
X
ρi λki .
|λi |<λ(ϕ)
Since for |λi | < λ(ϕ), lim λki /λ(ϕ)k = 0, it follows that the proof is completed.
Theorem 9.2. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Suppose that the sequence R(ϕk )/λ(ϕ)k is asymptotically periodic. Then there exist an integer m > 0 and an infinite sequence {pi }
of primes such that {mpi } ⊂ H(ϕ).
Proof. Since the sequence R(ϕk ) is unbounded, by Theorem 7.1, there exists q
such that all λi /|λi | with |λi | = λ(ϕ) are roots of unity of degree q, and the sequence {R̃λ(ϕ) (ϕk )} is periodic and nonzero, because lim supk→∞ |R̃λ(ϕ) (ϕk )| >
0 by Lemma 9.1. Consequently, there exists m with 1 ≤ m ≤ q such that
R̃λ(ϕ) (ϕm ) 6= 0.
Let ψ = ϕm . Then λ(ψ) = λ(ϕm ) = λ(ϕ)m ≥ 1. The periodicity R̃λ(ϕ) (ϕm+ℓq ) =
λ(ϕ)
R̃
(ϕm ) implies that R̃λ(ψ) (ψ 1+ℓq ) = R̃λ(ψ) (ψ) for all ℓ > 0. By Lemma 9.1 or
Theorem 7.1, we can see that there exists γ > 0 such that R(ψ 1+ℓq ) ≥ γλ(ψ)1+ℓq >
0 for all ℓ sufficiently large. From Proposition 7.4 it follows that the Dold multiplicity I1+ℓq (ψ) satisfies |I1+ℓq (ψ)| ≥ (γ/2)λ(ψ)1+ℓq when ℓ is sufficiently large.
According to Dirichlet prime number theorem, since (1, q) = 1, there are infinitely many primes p of the form 1 + ℓq. Consider all primes pi satisfying
|Ipi (ψ)| ≥ (γ/2)λ(ψ)pi .
By Proposition 8.2, #IR(ψ pi ) = Ipi (ψ) > 0, each pi is the height of some (essential) Reidemeister class [α]pi ∈ R[ψ pi ]. That is, [α]pi is an irreducible Reidemeister
class of ψ pi . Consider the Reidemeister class [α]mpi determined by α of ϕmpi . Let
di be the depth of the Reidemeister class [α]mpi ∈ R[ϕmpi ]. Then di = mi pi for
some mi | m and so there is an irreducible Reidemeister class [β]di ∈ R[ϕdi ] which is
boosted to [α]mpi . This means that di is the height of [β]di . Choose a subsequence
{mik } of the sequence {mi } bounded by m which is constant, say m0 . Consequently,
the infinite sequence {m0 pik } consists of heights of ϕ, or {m0 pi } ⊂ H(ϕ).
In the proof of Theorem 9.2, we have shown the following, which proves that the
algebraic period is a (homotopy) height when it is a prime number.
Corollary 9.3. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. For all primes p, if Ap (ϕ) 6= 0 then p ∈ H(ϕ).
Corollary 9.4. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. If the sequence {R(ϕk )} is strictly monotone increasing, then there exists N such that the set H(ϕ) contains all primes larger than
N.
Proof. By the assumption, we have λ(ϕ) > 1. Thus by Theorem 7.3, there exist
γ > 0 and N such that if k > N then there exists ℓ = ℓ(k) < r(ϕ) such that
R(ϕk−ℓ )/λ(ϕ)k−ℓ > γ. Then for all k > N , the monotonicity gives
R(ϕk )
R(ϕk−ℓ )
R(ϕk−ℓ )
γ
γ
.
≥
=
≥
≥
λ(ϕ)k
λ(ϕ)k
λ(ϕ)k−ℓ λ(ϕ)ℓ
λ(ϕ)ℓ
λ(ϕ)r(ϕ)
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
23
Applying Proposition 7.4 with ǫ = γ/λ(ϕ)r(ϕ) , we see that Ik (ϕ) 6= 0 and so
Ak (ϕ) 6= 0 for all k sufficiently large. Now our assertion follows from Corollary 9.3.
Example 9.5. There are examples of groups and endomorphisms satisfying the
conditions of the above Corollary. The simplest one is the endomorphism ϕ : Z → Z
such that ϕ(1) = d. Then ϕk (1) = dk and so R(ϕk ) = |1 − dk | for all k > 0. When
d ≥ 2, it is easy to see that all R(ϕk ) are finite and the sequence {R(ϕk )} is strictly
increasing.
By a direct computation, we can show that
{1}
if d = 0 or −1
∅
if d = 1
H(ϕ) =
N − {2} if d = −2
N
if d ≥ 2 or d ≤ −3.
In fact, when d = 1, all the Reidemeister classes are inessential and hence by
definition H(ϕ) = ∅. For another instance, consider the case d = −2. For any
k ≥ 1, ϕk (1) = (−2)k and so the Reidemeister class [n]k ∈ R[ϕk ] is
[n]k = {m + n − (−2)k m | m ∈ Z} = n + (1 − (−2)k )Z.
Since ι1,2 ([n]1 ) = [n + ϕ(n)]2 = [n − 2n]2 = [−n]2 , it follows that 2 ∈
/ H(ϕ). Next,
we remark that
ιk,ℓ ([n]k ) = [n + ϕk (n) + · · · + ϕℓ−k (n)]ℓ
= [(1 + (−2)k + · · · + (−2)ℓ−k )n]ℓ
ℓ
1 − (−2)ℓ
n
.
=
1 − (−2)k
For 0 ≤ n < |1 − (−2)k |, when ℓ 6= 2 we see that
1 − (−2)ℓ
6= 1,
1 − (−2)k
0≤
1 − (−2)ℓ
n < |1 − (−2)ℓ |.
1 − (−2)k
This implies that if ℓ 6= 2 then ℓ ∈ H(ϕ). Consequently, H(ϕ) = N − {2}. Now the
remaining cases can be treated in a similar way and we omit a detailed computation.
An endomorphism ϕ : Π → Π is essentially reducible if any Reidemeister class
of ϕk being boosted to an essential Reidemeister of ϕkn is essential, for any positive
integers k and n. The group Π is essentially reducible if every endomorphism
on Π is essentially reducible.
Lemma 9.6 ([16, Lemma 6.7]). Every poly-Bieberbach group is essentially reducible.
This means that for any n, if [α]n is essential and if ιm,n ([β]m ) = [α]n then [β]m
is essential.
Lemma 9.7 ([1, Proposition 2.2]). Let ϕ : Π → Π be an endomorphism such that
all R(ϕk ) are finite. If
X
R(ϕk ) < R(ϕm ),
m
k :
prime
then ϕ has a periodic Reidemeister class with height m, i.e., m ∈ H(ϕ).
24
ALEXANDER FEL’SHTYN AND JONG BUM LEE
Proof. Let r = R(ϕm ) and let [α1 ]m , · · · , [αr ]m be the Reidemeister classes of ϕm .
If some [αj ]m is irreducible, then we are done. So assume no [αj ]n is irreducible.
Then, for each j, there is a kj so that m/k
[αj ]m is reducible to [βj ]kj ∈
Pj is prime and
k
m
kj
R[ϕ ]. But this shows that R(ϕ ) ≤ m :prime R(ϕ ), a contradiction.
k
We can not only extend but also strengthen Corollary 9.4 as follows:
Proposition 9.8. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach
group such that all R(ϕk ) are finite. Suppose that the sequence {R(ϕk )} is strictly
monotone increasing. Then:
(1) All primes belong to H(ϕ).
(2) There exists N such that if p is a prime > N then {pn | n ∈ N} ⊂ H(ϕ).
Proof. Observe that for any prime p
X
R(ϕp ) −
R(ϕk ) = R(ϕp ) − R(ϕ) = Ip (ϕ).
p
k:
prime
The strict monotonicity implies Ap (ϕ) = pIp (ϕ) > 0, and hence p ∈ H(ϕ), which
proves (1).
Under the same assumption, we have shown in the proof of Corollary 9.4 that
there exists N such that k > N ⇒ Ik (ϕ) > 0. Let p be a prime > N and n ∈ N.
Then
n
X
X
n−1
n
Ipi (ϕ) − R(ϕp ) = Ipn (ϕ) > 0.
R(ϕp ) −
R(ϕk ) =
pn
k
: prime
i=0
By Lemma 9.7, we have pn ∈ H(ϕ), which proves (2).
In Remark 7.6, we made a statement about the lower density DA(ϕ) of the set
of algebraic periods A(ϕ) = {m ∈ N | Am (ϕ) 6= 0}. We can consider as well the
lower density of the set H(ϕ) of heights, see also [37], [20] and [16]:
DH(ϕ) = lim inf
k→∞
#(H(ϕ) ∩ [1, k])
.
k
Since Ik (ϕ) = #IR(ϕk ) by Proposition 8.2, it follows that A(ϕ) ⊂ HI(ϕ) = H(ϕ).
Hence we have DA(ϕ) ≤ DH(ϕ).
Corollary 9.9. Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group
such that all R(ϕk ) are finite. Suppose that the sequence {R(ϕk )} is strictly monotone increasing. Then H(ϕ) is cofinite and DA(ϕ) = DH(ϕ) = 1.
Proof. Under the same assumption, we have shown in the proof of Corollary 9.4
that there exists N such that if k > N then Ik (ϕ) > 0. This means IR(ϕk ) is
nonempty by Proposition 8.2 and hence k ∈ H(ϕ).
Let ϕ : Π → Π be an endomorphism on a poly-Bieberbach group Π of S such
that all R(ϕk ) are finite. When ϕ is the semi-conjugate by an affine map (d, D) on
S, we say that ϕ is expanding if all the eigenvalues of D∗ have modulus > 1.
Now we can prove the main result of [34].
Corollary 9.10 ([32, Theorem 4.6], [34, Theorem 3.2]). Let ϕ be an expanding
endomorphism on an almost Bieberbach group. Then HPer(ϕ) is cofinite.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
25
V
Proof. Since ϕ is expanding, we have that λ(ϕ) = sp( D∗ ) > 1. For any k > 0,
we can write R(ϕk ) = Γk + Ωk , where
r(ϕ)
n(ϕ)
X
X
ρi λki with |λi | < λ(ϕ).
ρj e2iπ(kθj ) , Ωk =
Γk = λ(ϕ)k
j=1
i=n(ϕ)+1
Here Ωk → 0 and Γk → ∞ as k → ∞. This implies that R(ϕk ) is eventually strictly
monotone increasing. We can use Corollary 9.4 and then Corollary 9.9 to conclude
the assertion.
References
[1] L. Alsedà, S. Baldwin, J. Llibre, R. Swanson and W. Szlenk, Minimal sets of periods for torus
maps via Nielsen numbers, Pacific J. Math., 169 (1995), 1–32.
[2] I. K. Babenko and S. A. Bogatyı̌, The behavior of the index of periodic points under iterations
of a mapping, Izv. Akad. Nauk SSSR Ser. Mat., 55 (1991), 3–31 (Russian); translation in
Math. USSR-Izv., 38 (1992), 1–26.
[3] L. Block, J. Guckenheimer, M. Misiurewicz and L. S. Young, Periodic points and topological
entropy of one-dimensional maps, Global theory of dynamical systems (Proc. Internat. Conf.,
Northwestern Univ., Evanston, Ill., 1979), pp. 18–34, Lecture Notes in Math., 819, Springer,
Berlin, 1980.
[4] R. Bowen, Entropy and the fundamental group, The structure of attractors in dynamical
systems (Proc. Conf., North Dakota State Univ., Fargo, N.D., 1977), pp. 21–29, Lecture
Notes in Math., Vol. 668, Springer, Berlin, 1978.
[5] R. Bowen and O. E. Lanford, III, Zeta functions of restrictions of the shift transformation,
1970 Global Analysis (Proc. Sympos. Pure Math., Vol. XIV, Berkeley, Calif., 1968) pp. 43–49
Amer. Math. Soc., Providence, R.I.
[6] K. Chandrasekharan, Introduction to analytic number theory, Die Grundlehren der mathematischen Wissenschaften, Band 148, Springer-Verlag New York Inc., New York, 1968.
[7] K. Dekimpe and G.-J. Dugardein, Nielsen zeta functions for maps on infra-nilmanifolds are
rational, to appear in J. Fixed Point Theory Appl. (arXiv:1302.5512).
[8] K. Dekimpe and G.-J. Dugardein, A note on homotopy minimal periods for hyperbolic maps
on infra-nilmanifolds, (arXiv:1408.5579).
[9] A. Dold, Fixed point indices of iterated maps, Invent. Math., 74 (1983), 419–435.
[10] F. Farrell and L. Jones, Classical aspherical manifolds, CBMS Regional Conference Series in
Mathematics 75, Amer. Math. Soc., Providence, RI, 1990.
[11] A. L. Fel’shtyn, New zeta function in dynamic, in Tenth Internat. Conf. on Nonlinear Oscillations, Varna, Abstracts of Papers, B, 1984.
[12] A. L. Fel’shtyn, New zeta functions for dynamical systems and Nielsen fixed point theory,
Lecture Notes in Math., 1346, Springer, 1988, 33–55.
[13] A. L. Fel’shtyn, The Reidemeister zeta function and the computation of the Nielsen zeta
function, Colloq. Math., 62, (1991), 153–166.
[14] A. Fel’shtyn, Dynamical zeta functions, Nielsen theory and Reidemeister torsion, Mem. Amer.
Math. Soc., 699, Amer. Math. Soc., Providence, R.I., 2000.
[15] A. Fel’shtyn and J. B. Lee, The Nielsen and Reidemeister numbers of maps on infrasolvmanifolds of type (R), Topology Appl., 181 (2015), 62–103.
[16] A. Fel’shtyn and J. B. Lee, The Nielsen numbers of iterations of maps on infra-solvmanifolds
of type (R) and periodic orbits, (arXiv:1403.7631).
[17] A. Fel’shtyn and E. Troitsky, Twisted Burnside-Frobenius theory for discrete groups, J. reine
Angew. Math., 613 (2007), 193–210.
[18] A. Fel’shtyn and E. Troitsky, Geometry of Reidemeister classes and twisted Burnside theorem,
J. K-Theory, 2 (2008), 463–506.
[19] B. Halpern, Periodic points on tori, Pacific J. Math., 83 (1979), 117–133.
[20] J. W. Hoffman, Z. Liang, Y. Sakai and X. Zhao, Homotopy minimal period self-maps on flat
manifolds, Adv. Math., 248 (2013), 324–334.
[21] N. V. Ivanov, Entropy and the Nielsen numbers. Dokl. Akad. Nauk SSSR 265 (2) (1982),
284-287 (in Russian); English transl.: Soviet Math. Dokl. 26 (1982), 63–66.
26
ALEXANDER FEL’SHTYN AND JONG BUM LEE
[22] J. Jezierski, Wecken theorem for fixed and periodic points, Handbook of Topological Fixed
Point Theory, 555–615, Springer, Dordrecht, 2005.
[23] J. Jezierski, J. Kȩdra and W. Marzantowicz, Homotopy minimal periods for solvmanifolds
maps, Topology Appl., 144 (2004), 29–49.
[24] J. Jezierski and W. Marzantowicz, Homotopy methods in topological fixed and periodic points
theory, Topological Fixed Point Theory and Its Applications, 3, Springer, Dordrecht, 2006.
[25] J. Jezierski and W. Marzantowicz, Homotopy minimal periods for nilmanifolds maps,
Math. Z., 239 (2002), 381–414.
[26] J. Jezierski and W. Marzantowicz, Homotopy minimal periods for maps of three dimensional
nilmanifolds, Pacific J. Math., 209 (2003), 85–101.
[27] B. Jiang, Lectures on Nielsen fixed point theory, Contemp. Math., 14, Amer. Math. Soc.,
Providence, R.I., 1983.
[28] B. Jiang and J. Llibre, Minimal sets of periods for torus maps, Discrete Contin. Dynam.
Systems, 4 (1998), 301–320.
[29] A. B. Katok, The entropy conjecture(Russian), Smooth dynamical systems(Russian), pp.181–
203. Izdat. “Mir”, Moscow, 1977.
[30] A. Katok and B. Hasselblatt, Introduction to the modern theory of dynamical systems, Encyclopedia of Mathematics and its Applications, 54, Cambridge University Press, Cambridge,
1995.
[31] J. Y. Kim, S. S. Kim and X. Zhao, Minimal sets of periods for maps on the Klein bottle, J.
Korean Math. Soc., 45 (2008), 883–902.
[32] J. B. Lee and K. B. Lee, Lefschetz numbers for continuous maps, and periods for expanding
maps on infra-nilmanifolds, J. Geom. Phys., 56 (2006), 2011–2023.
[33] J. B. Lee and K. B. Lee, Averaging formula for Nielsen numbers of maps on infra-solvmanifolds
of type (R), Nagoya Math. J., 196 (2009), 117–134.
[34] J. B. Lee and X. Zhao, Homotopy minimal periods for expanding maps on infra-nilmanifolds,
J. Math. Soc. Japan, 59 (2007), 179–184.
[35] J. B. Lee and X. Zhao, Nielsen type numbers and homotopy minimal periods for maps on
the 3-nilmanifolds, Sci. China Ser. A, 51 (2008), 351–360.
[36] J. B. Lee and X. Zhao, Nielsen type numbers and homotopy minimal periods for maps on
the 3-solvmanifolds, Algebr. Geom. Topol., 8 (2008), 563–580.
[37] J. B. Lee and X. Zhao, Density of the homotopy minimal periods of maps on infrasolvmanifolds, arXiv:1404.5114.
[38] J. Llibre, A note on the set of periods for Klein bottle maps, Pacific J. Math., 157 (1993),
87–93.
[39] W. Marzantowicz and F. Przytycki, Entropy conjecture for continuous maps of nilmanifolds,
Israel J. Math., 165 (2008), 349–379.
[40] W. Marzantowicz and F. Przytycki, Estimates of the topological entropy from below for
continuous self-maps on some compact manifolds, Discrete Contin. Dyn. Syst., 21 (2008),
501–512.
[41] T. Matsuoka, The number of periodic points of smooth maps, Ergod. Th. & Dynam. Sys., 9
(1989), 153–163.
[42] V. B. Pilyugina and A. L. Fel’shtyn, The Nielsen zeta function, Funktsional. Anal. i Prilozhen.,
19 (1985) 61–67 (in Russian); English transl.: Functional Anal. Appl., 19, (1985) 300–305.
[43] Y. Puri and T. Ward, Arithmetic and growth of periodic orbits, J. Integer Seq., 4 (2001),
Article 01.2.1, 18 pp.
[44] M. Shub, Dynamical systems, filtrations and entropy, Bull. Amer. Math. Soc., 80 (1974),
27–41.
[45] M. Shub and D. Sullivan, A remark on the Lefschetz fixed point formula for differentiable
maps, Topology, 13 (1974), 189–191.
[46] B. Wilking, Rigidity of group actions on solvable Lie groups, Math. Ann., 317 (2000), 195–
237.
THE NIELSEN AND REIDEMEISTER THEORIES OF ITERATIONS
27
Instytut Matematyki, Uniwersytet Szczecinski, ul. Wielkopolska 15, 70-451 Szczecin,
Poland
E-mail address: [email protected]
Department of mathematics, Sogang University, Seoul 121-742, KOREA
E-mail address: [email protected]
| 4math.GR
|
KOSZUL ALMOST COMPLETE INTERSECTIONS
arXiv:1801.00763v1 [math.AC] 2 Jan 2018
MATTHEW MASTROENI
Abstract. Let R = S/I be a quotient of a standard graded polynomial ring
S by an ideal I generated by quadrics. If R is Koszul, a question of Avramov,
Conca, and Iyengar asks whether the Betti numbers of R over S can be bounded
above by binomial coefficients on the minimal number of generators of I. Motivated by previous results for Koszul algebras defined by three quadrics, we give
a complete classification of the structure of Koszul almost complete intersections
and, in the process, give an affirmative answer to the above question for all such
rings.
1. Introduction
Let k be a field, S be a standard graded polynomial ring over k, I ⊆ S be a
graded ideal, and R = S/I. We say that R is a Koszul algebra if k ∼
= R/R+ has
a linear free resolution over R. Many rings arising from algebraic geometry are
Koszul, including the coordinate rings of Grassmannians [Kem90], sets of r ≤ 2n
points in general position in Pn [Kem92], and canonical embeddings of smooth
curves under mild restrictions [Pol95], as well as all suitably high Veronese subrings of any standard graded algebra [Bac86]. However, the simplest examples
of Koszul algebras, due to Fröberg [Fro75], are quotients by quadratic monomial
ideals, and a guiding heuristic in the study of Koszul algebras has been that any
reasonable property of algebras defined by quadratic monomial ideals should also
hold for Koszul algebras; for example, see [ACI10], [Con14], [ACI15]. Among such
properties, considering the Taylor resolution for an algebra defined by a quadratic
monomial ideal leads to the following question about the Betti numbers of a Koszul
algebra.
Question 1.1 ([ACI10, 6.5]). If R is Koszul and I is minimally generated by g
elements, does the following inequality hold for all i?
g
S
βi (R) ≤
i
In particular, is pdS R ≤ g?
Key words and phrases. Betti numbers, Koszul algebras, almost complete intersections.
1
2
MATTHEW MASTROENI
The above questions are known to have affirmative answers when R is LGquadratic (see next section) and for arbitrary Koszul algebras when g ≤ 3 by
[BHI16, 4.5]. Recall that R or I is called an almost complete intersection if I
is minimally generated by ht I + 1 elements. The motivation for studying Koszul
almost complete intersections comes from the fact that the above question is easily
seen to have an affirmative answer when I is a complete intersection or has height
one so that the interesting case for Koszul algebras defined by three quadrics is
precisely when I is an almost complete intersection. Our main results (3.1, 3.3, 3.4,
4.3) show that Question 1.1 has an affirmative answer for Koszul almost complete
intersections generated by any number of quadrics; they are summarized in the
theorem below.
Main Theorem. Let R = S/I be a Koszul almost complete intersection with I
S
minimally generated by g + 1 quadrics for some g ≥ 1. Then β2,3
(R) ≤ 2, and:
S
(a) If β2,3
(R) = 1, then I = (xz, zw, q3 , . . . , qg+1 ) for some linear forms x, z,
and w and some regular sequence of quadrics q3 , . . . , qg+1 on S/(xz, zw).
S
(b) If β2,3
(R) = 2, then I = I2 (M) + (q4 , . . . , qg+1 ) for some 3 × 2 matrix of
linear forms M with ht I2 (M) = 2 and some regular sequence of quadrics
q4 , . . . , qg+1 on S/I2 (M).
Furthermore, R is LG-quadratic and, therefore, satisfies βiS (R) ≤
g+1
i
for all i.
The division of the rest of the paper is as follows. We recount various properties
and examples of Koszul algebras and their Betti tables in §2 which will be important in the sequel. In §3, we determine the structure of Koszul almost complete
intersections with either one or two linear second syzygies. We then complete the
classification of Koszul almost complete intersections in §4 by showing that every
quadratic almost complete intersection has at most two linear second syzygies.
Notation. Throughout the remainder of the paper, the following notation will
be in force unless specifically stated otherwise. Let k be a fixed ground field of
arbitrary characteristic, S be a standard graded polynomial ring over k, I ⊆ S be
a proper graded ideal, and R = S/I. Recall that the ideal I is called nondegenerate
if it does not contain any linear forms. We can always reduce to a presentation for
R with I nondegenerate by killing a basis for the linear forms contained in I, and
we will assume that this is the case throughout. We denote the irrelevant ideal of
L
R by R+ = n≥1 Rn .
KOSZUL ALMOST COMPLETE INTERSECTIONS
3
2. Koszul Algebras and Their Betti Tables
If R is a Koszul algebra, it is well-known that its defining ideal I must be
generated by quadrics, but not every ideal generated by quadrics defines a Koszul
algebra. We have already noted in the introduction that every quadratic monomial
ideal defines a Koszul algebra. More generally, we say that R or I is G-quadratic
if, after a suitable linear change of coordinates ϕ : S → S, the ideal ϕ(I) has a
Gröbner basis consisting of quadrics. We also say that R or I is LG-quadratic
if R is a quotient of a G-quadratic algebra A by an A-sequence of linear forms.
Every G-quadratic algebra is Koszul by upper semicontinuity of the Betti numbers; see [BC03, 3.13]. It then follows from Proposition 2.4 below that every
LG-quadratic algebra is also Koszul. In particular, every complete intersection
generated by quadrics is LG-quadratic by an argument due to Caviglia. Indeed, if
R = S/(q1 , . . . , qg ) where q1 , . . . , qg is a regular sequence of quadrics, we can take
A = S[y1, . . . , yg ]/(y12 + q1 , . . . , yg2 + qg ) so that A/(y1 , . . . , yg ) ∼
= R. By choosing a
monomial order in which the yi are greater than every monomial in the variables
of S, it follows from [Eis95, 15.15] that A is G-quadratic and that the yi2 + qi form
a regular sequence so that htA (y1 , . . . , yg ) = dim A − dim R = dim S − dim R = g
and y1 , . . . , yg is an A-sequence. In summary, we have the following implications.
G-quadratic
LG-quadratic
Koszul
Quadratic CI’s
Each of the above implications is strict. Clearly, any quadratic monomial ideal
which is not a complete intersection, such as (xy, xz, xw) ⊆ k[x, y, z, w], is LGquadratic. In [Con14, 1.14], it is observed that (x2 +yz, y 2 +xz, z 2 +xy) ⊆ Q[x, y, z]
is an Artinian quadratic complete intersection which cannot be G-quadratic since
it does not contain the square of a linear form. We will see an example of a Koszul
algebra which is not LG-quadratic below.
Remark 2.1. If R = S/I is G-quadratic and J is a quadratic initial ideal of I,
then β1S (R) = β1S (S/J) since I is generated by quadrics and R and S/J have the
same Hilbert function. Consequently, βi (R) ≤ βi (S/J) ≤ gi for all i by upper
semicontinuity of the Betti numbers and the Taylor resolution for S/J. Since
killing a regular sequence of linear forms does not affect the Betti numbers of R, it
follows that Question 1.1 has an affirmative answer for every LG-quadratic algebra.
We will be specifically interested in the graded Betti numbers of a Koszul algebra
S
R, which are defined by βi,j
(R) = dimk TorSi (k, R)j and related to the usual Betti
4
MATTHEW MASTROENI
S
numbers by βiS (R) =
j βi,j (R). This information is usually organized into a
table, called the Betti table of R; see below for an example. As we have already
pointed out in the the introduction, quadratic monomial ideals serve as a useful
benchmark in the study of Koszul algebras. In particular, we note that the squarefree quadratic monomial ideals are precisely edge ideals. Recall that, if G is a
graph with vertex set [n] = {1, . . . , n}, the edge ideal of G is the ideal of the
polynomial ring S = k[x1 , . . . , xn ] defined by IG = (xi xj | ij ∈ E(G)). We note
that every quadratic monomial ideal can be obtained as the image of an edge ideal
modulo a regular sequence of linear forms via polarization [BH93, 4.2.16], and
hence, studying the Betti tables of all quadratic monomial ideals with g generators
is equivalent to studying the Betti tables of edge ideals of graphs with g edges,
which are reasonably simple to enumerate in practice for small values of g.
In fact, a byproduct of the proof in [BHI16] that every Koszul algebra defined
by g ≤ 3 quadrics satisfies Question 1.1 is that every such algebra has the Betti
table of some edge ideal. This suggests taking the Betti tables of edge ideals as our
guide for exploring possible patterns in Betti tables of Koszul algebras with more
generators. One can then easily compute that, for various values of g, there are
only two possible Betti tables for almost complete intersection edge ideals with g
generators, one with a single linear syzygy and another with two. The purpose of
this paper is to show that this pattern holds more generally for all Koszul almost
complete intersections. However, the following example shows that the mantra
that Koszul algebras are similar to quotients by quadratic monomial ideals must
be taken with a grain of salt.
P
Example 2.2 ([Con14, 3.8]). The ring R = k[x, y, z, w]/(xy, xw, (x − y)z, z 2 , x2 +
zw) is Koszul by a filtration argument. The minimal free resolution of R over
S = k[x, y, z, w] can be computed via iterated mapping cones using the fact that
((xy, xw, z 2 ) : (x − y)z) = (xy, xw, z) and ((xy, xw, z 2 , (x − y)z) : x2 + zw) =
(xw, y, z). This yields the following Betti table for R, where the entry in column
S
i and row j is βi,i+j
(R) and zero entries are represented by “−” for readability.
0
0 1
1 –
2 –
1
–
5
–
2
–
4
4
3
–
–
6
4
–
–
2
From the above Betti table, we see that the Hilbert series of R is
HR (t) =
1 + 2t − 2t2 − 2t3 + 2t4
.
(1 − t)2
KOSZUL ALMOST COMPLETE INTERSECTIONS
5
If R were LG-quadratic, then the numerator of the Hilbert series must be the hpolynomial of a 5-generated edge ideal, since killing a regular sequence of linear
forms and passing to a quadratic initial ideal do not change either the h-polynomial
or the minimal number of generators of the defining ideal. As noted above, we can
easily compute that the h-polynomial of R does not belong to any 5-generated
edge ideal. Hence, R is not LG-quadratic, and in particular, the Betti table of R
is not the Betti table of any edge ideal. This points to unexpected complications
in trying to answer Question 1.1 for Koszul algebras defined by g ≥ 5 quadrics.
Remark 2.3. We can compute the Betti tables of 5-generated edge ideals in the
above example over a field of any characteristic by a result of Katzman, [Kat06,
4.1]. However, Katzman also shows that the Betti tables of edge ideals do depend
on the characteristic of the ground field in general.
In the remainder of this section, we collect a few results about Koszul algebras
that will be useful in the sequel. The first of these results states how the Koszul
property can be passed to and from quotient rings.
Proposition 2.4 ([CDR13, §3.1, 2]). Let S be a standard graded k-algebra and R
be a quotient ring of S.
(a) If S is Koszul and regS (R) ≤ 1, then R is Koszul.
(b) If R is Koszul and regS (R) is finite, then S is Koszul.
Compared with general quadratic algebras, the Betti tables of Koszul algebras
are much more restricted. The following result, discovered in [Bac88] and [Kem90,
4], says that the Betti tables of Koszul algebras have nonzero entries only on or
above the diagonal; see [Con14, 2.10] for an easier argument using regularity.
S
Lemma 2.5. If R = S/I is a Koszul algebra, then βi,j
(R) = 0 for all i and j > 2i.
In addition, the extremal portions of the Betti table of a Koszul algebra R,
namely the diagonal entries and the linear strand of I, satisfy bounds similar to
those in Question 1.1.
Proposition 2.6 ([BHI16, 3.4, 4.2]). Suppose that R = S/I is Koszul and that I
is minimally generated by g elements. Then:
S
(a) βi,i+1
(R) ≤ gi for 2 ≤ i ≤ g, and if equality holds for i = 2, then I has
height one and a linear resolution of length g.
S
(b) βi,2i
(R) ≤ gi for 2 ≤ i ≤ g, and if equality holds for some i, then I is a
complete intersection.
Corollary 2.7. If R = S/I is a Koszul algebra which is not a complete intersection, then I has a linear syzygy.
6
MATTHEW MASTROENI
Proof. Suppose that I is minimally generated by q1 , . . . , qg . If the Koszul syzygies
S
on the qi are all minimal generators of SyzS1 (I), then β2,4
(R) ≥ g2 contradicting the
preceding proposition. Hence, some k-linear combination of the Koszul syzygies is
not minimal, and therefore, it is an S-linear combination of linear syzygies.
Lastly, we will need a fact about the syzygies of a Koszul algebra which will be
clear to experts, but for completeness, we give a quick proof. The proof relies on
the product structure on TorS∗ (R, k). We briefly recall how this product is defined
and refer the reader to [Avr98] for further details. To simplify notation, all tensor
products below are over S.
If F• denotes the minimal free resolution of R over S, we have the Künneth map
κ
TorSi (R, k) ⊗ TorSj (R, k) = Hi (F• ⊗ k) ⊗ Hj (F• ⊗ k) −→ Hi+j ((F• ⊗ k) ⊗ (F• ⊗ k))
sending cls(v)⊗cls(w) 7→ cls(v ⊗w). Denoting by µk : k ⊗k → k and µR : R⊗R →
R the respective product maps, we have a chain map µF : F• ⊗ F• → F• lifting
µR . The product structure on TorS∗ (R, k) is the composition of the Künneth map
with the map induced on homology by the chain map
F
k
µ ⊗µ
F• ⊗ k ⊗ F• ⊗ k ∼
= F• ⊗ F• ⊗ k ⊗ k −→ F• ⊗ k .
Proposition 2.8. If R = S/I is a Koszul algebra, then SyzS1 (I) is minimally
generated by linear syzygies and Koszul syzygies.
Proof. By Lemma 2.5, we know that SyzS1 (I) is minimally generated by linear
S
and quadratic syzygies. We may assume that β2,4
(R) 6= 0 or else the conclusion holds trivially. In that case, it follows from [ACI10, 3.1] that TorS2 (R, k)4 =
(TorS1 (R, k)2 )2 , so it suffices to note that the products of the generators of TorS1 (R, k)2
correspond to the Koszul syzygies on a minimal set of generators q1 , . . . , qg for I.
If e1 , . . . , eg denotes the standard basis of S(−2)g = F1 such that ∂(ei ) = qi for
each i, we can choose µF so that µF (ei ⊗ 1) = µF (1 ⊗ ei ) = ei for all i. Since the
ei ⊗ 1 span TorS1 (R, k)2 , it follows that the µF (ei ⊗ ej ) ⊗ 1 span TorS2 (R, k)4 . As
µF is a chain map, we see that
∂(µF (ei ⊗ ej )) = µF (∂(ei ⊗ ej )) = µF (qi ⊗ ej − ei ⊗ qj ) = qi ej − qj ei
so that µF (ei ⊗ej ) corresponds to a Koszul syzygy in SyzS1 (I) as wanted. Hence, the
Koszul syzygies together with multiples of the linear syzygies must span SyzS1 (I)4 ,
and the proposition easily follows.
KOSZUL ALMOST COMPLETE INTERSECTIONS
7
3. Koszul Almost Complete Intersections
Recall that a standard graded k-algebra R = S/I with ht I = g is called an
almost complete intersection (or ACI for short) if I is minimally generated by
g + 1 elements.
S
Theorem 3.1. Let R = S/I be a Koszul algebra with β2,3
(R) = 1. Then there are
independent linear forms x and w and a linear form z such that I = (xz, zw, q3 , . . . , qg+1 )
for some regular sequence of quadrics q3 , . . . , qg+1 on S/(xz, zw), and conversely,
S
every ideal of this form defines a Koszul algebra with β2,3
(R) = 1. Hence, R is an
g−1
almost complete intersection with e(R) = 2
and Betti table
0
1
2
..
.
0
1
1
–
– g+1
–
–
g−1 –
g
–
–
–
2
–
1
3
–
–
g+2 g−1
2
1
···
g−1
1
..
.
..
g
–
–
–
–
g−1
g+i
i
i−1
g+1
–
–
–
.
2g−1 g−1
g−1 g−2
–
–
S
Specifically, we have βi,2i
(R) =
g+1
S
βi (R) = i for all i.
g−1
–
–
–
S
and βi,2i−1
(R) =
g−1
g−2
2
g−1
i−2
–
1
for i ≥ 2 so that
Proof. Since I has a linear syzygy, it is not a complete intersection. In particular,
we can write I = (q1 , . . . , qg+1 ) for some linear independent quadrics qi with g ≥ 1.
Let U = SyzS1 (I), W ⊆ U4 denote the k-span of the Koszul syzygies on the qi , and
ℓ ∈ U denote the unique linear syzygy up to scalar multiple. If W ∩ S+ U = 0, then
S
β2,4
(R) ≥ g+1
so that Proposition 2.6 implies I is a complete intersection, which
2
is a contradiction. Hence, there is a linear form z such that zℓ ∈ W is nonzero.
P
Write zℓ = 1≤i<j≤g+1 ai,j (qj ei −qi ej ) for some ai,j ∈ k, where e1 , . . . , eg+1 denotes
the standard basis of S(−2)g+1 . After suitably relabeling the qi and rescaling the
equality, we may assume that a1,2 = 1. Reading off the first two coordinates of the
Pg+1
P
preceding equality then gives zℓ1 = q2 + g+1
j=3 a2,j qj .
j=3 a1,j qj and zℓ2 = −q1 +
Using these equalities, we can replace q1 and q2 with zℓ2 and zℓ1 as generators of
I and assume that q1 = xz and q2 = zw for some linear forms x, z, and w. Note
that x and w must be independent since the qi are.
After making this change, we have ℓ = (w, −x, 0, . . . , 0) is the unique linear
syzygy on the qi , and (q2 , −q1 , 0, . . . , 0) = zℓ. Let W ′ ⊆ U4 denote the k-span of
the Koszul syzygies other than (q2 , −q1 , 0, . . . , 0). If W ′ ∩ S+ U 6= 0, then there is
8
MATTHEW MASTROENI
a linear form v such that vℓ ∈ W ′ is nonzero. Write
X
bi,j (qj ei − qi ej )
vℓ =
1≤i<j≤g+1
j≥3
for some bi,j ∈ k. Since bi,j 6= 0 for some j ≥ 3, reading off the j-th coordinate
of the above equality yields a linear dependence relation on the qi , which is a
contradiction. Hence, we must have W ′ ∩ S+ U = 0 so that all of the Koszul
syzygies except (q2 , −q1 , 0, . . . , 0) are part of a minimal set of generators for U.
S
R
Because β2,4
(R) ≤ g+1
− 1 and β2,j
(S) = 0 for j > 4 by Lemma 2.5, it follows
2
S
that U = Syz1 (I) is minimally generated by all the Koszul syzygies on the qi ,
except (q2 , −q1 , 0, . . . , 0), together with the linear syzygy ℓ.
P
If fg+1 ∈ ((xz, zw, q3 , . . . , qg ) : qg+1 ), then we can write fg+1 qg+1 = − gi=1 fi qi
for some fi ∈ S so that (f1 , . . . , fg+1 ) ∈ U. It follows from the preceding paragraph
that fg+1 ∈ (xz, zw, q3 , . . . , qg ) so that qg+1 is regular on R′ = S/(xz, zw, q3 , . . . , qg ).
As regR′ R = 1, it follows from Corollary 2.4 that R′ is also Koszul. Moreover, because we can obtain the resolution of R over S by taking the mapping cone of multiS
(R′ ) = 1 and that
plication by qg+1 on the resolution of R′ over S, it follows that β2,3
(w, −x, 0, . . . , 0) is the unique linear syzygy on xz, zw, q3 , . . . , qg . Hence, induction
on g implies that q3 , . . . , qg+1 is a regular sequence on S/(xz, zw). Conversely,
if I = (xz, zw, q3 , . . . , qg+1 ) for some regular sequence of quadrics q3 , . . . , qg+1
on S/(xz, zw), it also follows from Corollary 2.4 that R = S/I is Koszul since
S/(xz, zw) is Koszul.
From the preceding paragraph, we see that ht I = ht(xz, zw) + g − 1 = g so that
I is an almost complete intersection. If F• denotes the minimal free resolution of
S/(xz, zw) over S, we obtain the minimal resolution of R by repeatedly taking the
mapping cone of multiplication by qi+1 on the resolution of S/(xz, zw, q3 , . . . , qi ).
Since taking the mapping cone of multiplication by qi+1 is the same as tensoring
with the Koszul complex on qi+1 , we see that F• ⊗S K• (q3 , . . . , qg+1 ) is the minimal
free resolution of R over S, from which the Betti table is easily deduced. In
particular, we have
g+1
g
g
g−1
g+i g−1
S
.
=
+
=
+
βi (R) =
i
i−1
i
i−2
i−1
i
Similarly, we note that the multiplicity of S/(xz, zw, q3 , . . . , qi+1 ) is twice the multiplicity of S/(xz, zw, q3 , . . . , qi ), and so, since e(S/(xz, zw)) = 1, we see that
e(R) = 2g−1 .
Remark 3.2. In the statement of the above theorem, we can choose x and w
so that zw, q3 , . . . , qg+1 is a maximal S-regular sequence contained in I. Indeed,
KOSZUL ALMOST COMPLETE INTERSECTIONS
9
since q3 , . . . , qg+1 is a regular sequence on S/(xz, zw), we know that q3 , . . . , qg+1 is a
regular sequence on S by Auslander’s Zerodivisor Theorem, which is a consequence
of the Peskine-Szpiro Intersection Theorem for arbitrary Noetherian local rings and
follows from results of Serre in the regular case; see [PS73, II.0]. Each associated
prime of (q3 , . . . , qg+1 ) cannot contain both xz and zw, otherwise we would have
ht I ≤ g −1 since the former ideal is unmixed. If either of xz or zw is not contained
in every associated prime of (q3 , . . . , qg+1 ), we are done after possibly switching the
roles of x and w. Otherwise, if xz and zw are both contained in different associated
primes of (q3 , . . . , qg+1 ), then we can replace w with w + x.
S
Theorem 3.3. Let R = S/I be a Koszul almost complete intersection with β2,3
(R) =
2. Then there is a 3 × 2 matrix of linear forms M with ht I2 (M) = 2 such that
I = I2 (M) + (q4 , . . . , qg+1 ) for some regular sequence of quadrics q4 , . . . , qg+1 on
S/I2 (M), and conversely, every ideal of this form defines a Koszul almost complete
S
intersection with β2,3
(R) = 2. Hence, R has multiplicity e(R) = 3 · 2g−2 and Betti
table
0
1
2
..
.
0
1
1
–
– g+1
–
–
3
g−2 –
g−1 –
–
–
2
–
2
g−2
+
1
3
–
–
g−2
2
2
···
g−2
1
..
.
–
–
S
Specifically, we have βi,2i
(R) = 3
g+1
S
so that βi (R) ≤ i for all i.
..
g−2
–
–
–
g−1
–
–
–
.
3
g−2
g−3
+1 2
–
g−2
i−1
+
g−2
i
g
–
–
–
S
and βi,2i−1
(R) = 2
g−2
g−3
3
g−2
i−2
–
2
for i ≥ 2
Proof. Since I has a linear syzygy, it is not a complete intersection. In particular,
we can write I = (q1 , . . . , qg+1 ) for some linear independent quadrics qi with g ≥ 1.
In fact, we must have g ≥ 2 since it is easily seen that a 2-generated graded ideal
cannot have two independent linear syzygies. Let U = SyzS1 (I), W ⊆ U4 denote
the k-span of the Koszul syzygies on the qi , and ℓ, h ∈ U denote independent
linear syzygies. Arguing as in the proof of the previous theorem, we see there
are linear forms z and v such that zℓ + vh ∈ W is nonzero. Write zℓ + vh =
P
1≤i<j≤g+1 ai,j (qj ei − qi ej ) for some ai,j ∈ k, , where e1 , . . . , eg+1 denotes the
standard basis of S(−2)g+1 . After suitably relabeling the qi and rescaling the
equality, we may assume that a1,2 = 1. Reading off the coordinates of the preceding
10
MATTHEW MASTROENI
equality then gives
q˜2 = zℓ1 + vh1 = q2 +
g+1
X
a1,j qj
j=3
−q˜1 = zℓ2 + vh2 = −q1 +
g+1
X
a2,j qj
j=3
zℓp + vhp = −
X
ai,p qi +
X
ap,i qi
(p ≥ 3) .
i>p
i<p
Using the above equalities, we can replace q1 and q2 with q˜1 and q˜2 as generators for
I. As a result, we must also replace ℓ with ℓ̃ = (ℓ1 , ℓ2 , ℓ3 − a1,3 ℓ2 + a2,3 ℓ1 , . . . , ℓg+1 −
a1,g+1 ℓ2 + a2,g+1 ℓ1 ) since
0=
g+1
X
ℓj qj = ℓ1 q˜1 + ℓ2 q˜2 +
g+1
X
(ℓj − a1,j ℓ2 + a2,j ℓ1 )qj .
j=3
j=1
Similarly, h must be replaced with the linear syzygy h̃ defined as above. It is
easily seen that ℓ̃ and h̃ must also be independent linear syzygies. Finally, setting
bi,j = ai,j + a1,j a2,i − a1,i a2,j for 3 ≤ i < j ≤ g + 1, we claim that
X
z ℓ̃ + v h̃ = (q˜2 , −q˜1 , 0, . . . , 0) +
bi,j (qj ei − qi ej ) .
3≤i<j≤g+1
By definition of q˜1 and q˜2 , it suffices to check equality in the p-th coordinate for
p ≥ 3. Using the above equalities, we see that
z ℓ̃p + v h̃p = zℓp + vhp + a1,p q˜1 + a2,p q˜2
=−
X
ai,p qi +
ap,i qi −
a1,p a2,i qi +
X
(ai,p + a1,p a2,i − a2,p a1,i )qi +
X
bi,p qi +
3≤i<p
g+1
X
a2,p a1,i qi
i=3
X
(ap,i − a1,p a2,i + a2,p a1,i )qi
i>p
3≤i<p
=−
g+1
X
i=3
i>p
3≤i<p
=−
X
X
bp,i qi
i>p
as required. Hence, after replacing q1 and q2 as above, we may assume that q1 =
−(zℓ2 + vh2 ), q2 = zℓ1 + vh1 , and a1,j = a2,j = 0 for all j ≥ 3.
If ai,j 6= 0 for some 3 ≤ i < j ≤ g + 1, then after relabeling the qi we may
assume that a3,4 6= 0. Since a1,j = a2,j = 0 for all j ≥ 3, arguing as in the
preceding paragraph shows that we can replace q3 and q4 with −(zℓ4 + vh4 ) and
zℓ3 + vh3 respectively so that I ⊆ (z, v, q5 , . . . , qg+1 ) has height at most g − 1 by
Krull’s Height Theorem, contradicting that I is an almost complete intersection.
KOSZUL ALMOST COMPLETE INTERSECTIONS
11
Therefore, ai,j = 0 for 3 ≤ i < j ≤ g+1, and we see that (q2 , −q1 , 0, . . . , 0) = zℓ+vh
for some linear forms z and v.
Suppose first that z and v are independent linear forms. Then zℓi + vhi = 0 for
i > 2 implies that (ℓi , hi ) = ai (v, −z) for some ai ∈ k so that ℓ = (ℓ1 , ℓ2 , a3 v, . . . , ag+1 v)
and h = (h1 , h2 , −a3 z, . . . , −ag+1 z). If ai = 0 for all i, then we would have two
independent linear syzygies on q1 and q2 , which we have already noted is impossible above. Hence, after relabeling, we may assume that a3 6= 0. Replacing
q3 with a3 q3 + · · · + ag+1 qg+1 , we may assume that ℓ = (ℓ1 , ℓ2 , v, 0, . . . , 0) and
h = (h1 , h2 , −z, 0, . . . , 0). Therefore, we have q1 = −(zℓ2 + vh2 ) and q2 = zℓ1 + vh1 ,
and furthermore, zq3 = h1 q1 + h2 q2 = z(ℓ1 h2 − ℓ2 h1 ) implies q3 = ℓ1 h2 − ℓ2 h1 so
that I = I2 (M) + (q4 , . . . , qg+1 ) where M is the matrix
ℓ1 h1
M = ℓ2 h2 .
(3.1)
v −z
Suppose now that v = cz for some c ∈ k. Then after replacing ℓ with ℓ + ch,
we may assume that (q2 , −q1 , 0, . . . , 0) = zℓ so that q1 = −zℓ2 , q2 = zℓ1 , and
ℓ = (ℓ1 , ℓ2 , 0, . . . , 0). Note that ℓ1 and ℓ2 must be independent linear forms or else q1
P
and q2 would not be independent. On the other hand, we know that g+1
i=1 hi qi = 0
so that (ℓ1 h2 − ℓ2 h1 )z ∈ (q3 , . . . , qg+1 ). We claim that z is a nonzerodivisor modulo
(q3 , . . . , qg+1 ) so that ℓ1 h2 − ℓ2 h1 ∈ (q3 , . . . , qg+1 ).
To see that the claim holds, we first note that ht(q3 , . . . , qg+1 ) = g − 1 so that
(q3 , . . . , qg+1 ) is a complete intersection. Indeed, if this is not the case, then since
I ⊆ (z, q3 , . . . , qg+1 ) we would have ht I ≤ ht(q3 , . . . , qg+1 ) + 1 ≤ g − 1 by Krull’s
Height Theorem and [Ser00, III, Prop. 17], contradicting that I is an almost complete intersection. If z were a zerodivisor modulo (q3 , . . . , qg+1 ), then there would
be an associated prime P of (q3 , . . . , qg+1 ) such that I ⊆ (z, q3 , . . . , qg+1 ) ⊆ P so
that ht I ≤ g − 1 as (q3 , . . . , qg+1 ) is unmixed, again contradicting that I is an
almost complete intersection. And so, we see that z must be a nonzerodivisor
modulo (q3 , . . . , qg+1 ) as claimed. Write ℓ1 h2 − ℓ2 h1 = a3 q3 + · · · + ag+1 qg+1 for
some ai ∈ k. If ai = 0 for all i, then ℓ1 h2 − ℓ2 h1 = 0 so that (h2 , −h1 ) = b(ℓ2 , −ℓ1 )
for some b ∈ k as ℓ1 and ℓ2 are independent linear forms. In that case, we can
replace h with h − bℓ and assume that h = (0, 0, h3 , . . . , hg+1 ) so that q3 is a zerodivisor modulo (q4 , . . . , qg+1 ). However, we claim that this is impossible. Indeed,
by arguing as above, we see that ht(q4 , . . . , qg+1 ) = g − 2 so that (q4 , . . . , qg+1 ) is a
complete intersection, and so, if q3 were a zerodivisor modulo (q4 , . . . , qg+1 ), there
would be an associated prime P of (q4 , . . . , qg+1 ) such that (q3 , . . . , qg+1 ) ⊆ P so
that ht(q3 , . . . , qg+1 ) ≤ g − 2 as (q4 , . . . , qg+1 ) is unmixed, contradicting our earlier
12
MATTHEW MASTROENI
observation. Hence, after relabeling, we may assume that a3 6= 0. Replacing q3
with a3 q3 + · · · + ag+1 qg+1 = ℓ1 h2 − ℓ2 h1 , we see that h = (h1 , h2 , −z, 0, . . . , 0) and
I = I2 (M) + (q4 , . . . , qg+1 ) where M is the matrix of linear forms in (3.1) with
v = 0.
In both of the above cases, it is easily checked that the Koszul syzygies involving
any two of q1 , q2 , q3 are non-minimal. Let W ′ ⊆ U4 denote the k-span of the other
Koszul syzygies. If W ′ ∩ S+ U 6= 0, then there are linear forms u and w such that
uℓ + wh ∈ W ′ is nonzero. Write
uℓ + wh =
g+1
X
j=4
[b1,j (qj e1 − q1 ej ) + b2,j (qj e2 − q2 ej )] +
X
bi,j (qj ei − qi ej )
3≤i<j≤g+1
for some bi,j ∈ k. Since bi,j 6= 0 for some j ≥ 4, reading off the j-th coordinate
of the above equality yields a linear dependence relation on the qi , which is a
contradiction. Hence, we must have W ′ ∩ S+ U = 0 so that all of the Koszul
syzygies involving at least one of q4 , . . . , qg+1 are part of a minimal set of generators
for U. By Proposition 2.8 and Lemma 2.5, it follows that U = SyzS1 (I) is minimally
generated by all the Koszul syzygies involving at least one of q4 , . . . , qg+1 together
with the linear syzygies ℓ and h.
P
If fg+1 ∈ ((q1 , . . . , qg ) : qg+1 ), then we can write fg+1 qg+1 = − gi=1 fi qi for
some fi ∈ S so that (f1 , . . . , fg+1 ) ∈ SyzS1 (I). It follows from the preceding paragraph that fg+1 ∈ (q1 , . . . , qg ) so that qg+1 is regular on R′ = S/(q1 , . . . , qg ). As
regR′ R = 1, it follows from Corollary 2.4 that R′ is also Koszul. Moreover, because
we can obtain the resolution of R over S by taking the mapping cone of multiS
(R′ ) = 2 and
plication by qg+1 on the resolution of R′ over S, it follows that β2,3
that (ℓ1 , ℓ2 , v, 0, . . . , 0) and (h1 , h2 , −z, 0, . . . , 0) are the independent linear syzygies on q1 , . . . , qg . Hence, induction on g implies that q4 , . . . , qg+1 is a regular
sequence on S/I2 (M). In particular, we see that g = ht I = ht I2 (M) + g − 2
so that ht I2 (M) = 2. Conversely, if I = I2 (M) + (q4 , . . . , qg+1 ) for some 3 × 2
matrix of linear forms M with ht I2 (M) = 2 and some regular sequence of quadrics
q4 , . . . , qg+1 on S/I2 (M), then S/I2 (M) has a Hilbert-Burch resolution by [BH93,
1.4.17] so that regS (S/I2 (M)) = 1, and it follows from Corollary 2.4 that S/I2 (M),
and hence also R = S/I, is Koszul.
If F• denotes the minimal free resolution of S/I2 (M) over S, we obtain the minimal resolution of R by repeatedly taking the mapping cone of multiplication by
qi+1 on the resolution of S/(q1 , . . . , qi ). Since taking the mapping cone of multiplication by qi+1 is the same as tensoring with the Koszul complex on qi+1 , we see
that F• ⊗S K• (q4 , . . . , qg+1 ) is the minimal free resolution of R over S, from which
KOSZUL ALMOST COMPLETE INTERSECTIONS
13
the Betti table is easily deduced. In particular, we have
g+1
g
g
g−1
g
g−1
g−1
S
.
=
+
≤
+
=
+
βi (R) = 2
i
i−1
i
i−1
i
i
i−1
Similarly, the multiplicity of S/(q1 , . . . , qi+1 ) is twice the multiplicity of S/(q1 , . . . , qi )
for i ≥ 3, and so, since e(S/I2 (M)) = 3, we see that e(R) = 3 · 2g−2 .
In the next section, we will show that every Koszul almost complete intersection has at most two linear syzygies so that the above results give a complete
classification of Koszul almost complete intersections, and therefore, Question 1.1
has an affirmative answer for Koszul almost complete intersections with any number of generators. Assuming this result for the time being, we have the following
corollary.
Corollary 3.4. Koszul almost complete intersections are LG-quadratic.
Proof. Let R = S/I be a Koszul almost complete intersection, and assume first
S
that β2,3
(R) = 2 so that I = I2 (M)+(q4 , . . . , qg+1 ) for some 3×2 matrix M = (mij )
of linear forms and q4 , . . . , qg+1 a regular sequence of quadrics on S/I2 (M). Set
S̃ = S[X][y4 , . . . , yg+1] where X = (xij ) is a 3 ×2 generic matrix, I˜ = I2 (X) + (y42 +
2
˜ If we choose a lexicographic order on S̃ with
q4 , . . . , yg+1
+ qg+1 ), and A = S̃/I.
x1,2 > x1,1 > x2,2 > x2,1 > x3,2 > x3,1 and yi greater than the variables in S for all
˜ = in> (I2 (X)) + (y 2, . . . , y 2 ) so that
i, it follows from [Eis95, 15.15] that in> (I)
4
g+1
2
˜
the 2-minors of X together with the yi + qi are a Gröbner basis for I by [Stu90].
Hence, A is G-quadratic. Moreover, we also know that the yi2 + qi form a regular
sequence on S̃/I2 (X), and since the latter ring is Cohen-Macaulay, we see that
A is Cohen-Macaulay. If J denotes the ideal of A generated by the linear forms
xij − mij and ys for i = 1, 2, 3, j = 1, 2, and 4 ≤ s ≤ g + 1, then A/J ∼
= R so that
R will be LG-quadratic if the linear forms generating J are a regular sequence.
Since A is Cohen-Macaulay and I˜ is also an almost complete intersection of height
g, this follows from the fact that ht J = dim A − dim R = dim S̃ − dim S = g + 4
is the number of generators of J.
S
Assume now that β2,3
(R) = 1 so that I = (xz, zw, q3 , . . . , qg+1 ) for some linear
forms x, z, and w and q3 , . . . , qg+1 a regular sequence of quadrics on S/(xz, zw).
In this case, we define Si = S[yi+1 , . . . , yg+1 ] and Ai = Si /Ii for 0 ≤ i ≤ g + 1,
where
2
I0 = (y1 z, y2 z, y32 + q3 , . . . , yg+1
+ qg+1 )
2
I1 = (xz, y2 z, y32 + q3 , . . . , yg+1
+ qg+1 )
2
2
Ii = (xz, zw, q3 , . . . , qi , yi+1
+ qi+1 , . . . , yg+1
+ qg+1 )
(i ≥ 2) .
14
MATTHEW MASTROENI
2
As above, we see that in> (I0 ) = (in> (z)y1 , in> (z)y2 , y32 , . . . , yg+1
) for any monomial
order on S0 in which the yi are greater than every monomial in S so that the generators of I0 are a Gröbner basis and A0 is G-quadratic. In addition, we have Ag+1 = R
and Ai /(yi+1 ) ∼
= Ai+1 for all i < g+1. An initial ideal argument as above shows that
2
2
yi+1 + qi+1 , . . . , yg+1
+ qg+1 is a regular sequence on Si /(xz, zw, q3 , . . . , qi ) for i ≥ 2,
2
and similarly, the yj + qj are a regular sequence on S0 /(y1z, y2 z) and S1 /(xz, y1 z).
Si
Consequently, Ai is a Koszul almost complete intersection with β2,3
(Ai ) = 1 and
ht Ii = g for all i. It then follows from Theorem 3.1 that the Ai have the same
Betti table, hence the same h-polynomial h(t), over their respective polynomial
rings Si . Hence, the Hilbert series of Ai is HAi (t) = h(t)/(1 − t)dim Ai . We then
compute that dim Ai = dim Si − g = dim Si+1 − g + 1 = dim Ai+1 + 1 so that
(1 − t)HAi (t) = HAi+1 (t) for all i < g + 1. This implies that the natural sequence
yi+1
0 → Ai (−1) → Ai → Ai+1 → 0 is exact so that yi+1 is Ai -regular. Therefore, we
see that y1 , . . . , yg+1 is an A0 -sequence, and R is LG-quadratic.
4. Linear Syzygies of Quadratic ACI’s
The following proposition is similar in spirit to Theorem 3.3. However, there are
two important distinctions: We do not assume that R is Koszul, so we lose some
information about the syzygies of the defining ideal I, and to make up for this
loss of information, we must assume that the ground field is infinite. But first, we
make a simple observation which will be useful in the proof.
Remark 4.1. If f1 , . . . , fn ∈ S is a regular sequence of homogeneous forms of
the same degree and f = a1 f1 + · · · + an fn for some ai ∈ k with a1 6= 0, then
((f2 , . . . , fn ) : f ) = ((f2 , . . . , fn ) : f1 ) so that f, f2 , . . . , fn is also a regular sequence.
Proposition 4.2. Suppose that k is an infinite field and that R = S/I is an almost
S
complete intersection defined by quadrics with β2,3
(R) ≥ 2. Then there are quadrics
q1 , . . . , qg+1 and a 3 × 2 matrix of linear forms M such that I = (q1 , . . . , qg+1 ),
q2 , . . . , qg+1 is a regular sequence, and I2 (M) = (q1 , q2 , q3 ).
Proof. Set g = ht I. We note that g ≥ 2, since otherwise we would have I =
S
(xz, yz) for some linear forms x, y, z so that β2,3
(R) = 1. First, we can find
quadrics q1 , . . . , qg+1 such that I = (q1 , . . . , qg+1 ) and q2 , . . . , qg+1 is a regular sequence. Indeed, we can take qg+1 to be any quadric in I. Having found quadrics
qi , . . . , qg+1 ∈ I with i > 2 forming a regular sequence, we know that I is not
contained in any associated prime of (qi , . . . , qg+1 ) since the latter ideal is unmixed
of height g − i + 2 < g. Because I is generated in degree two, this implies I2 * P
for each associated prime P of (qi , . . . , qg+1 ). Since k is infinite, I2 is not a union
of the proper subspaces (I ∩ P )2 for P ∈ Ass(S/(qi , . . . , qg+1 )). Hence, we can find
KOSZUL ALMOST COMPLETE INTERSECTIONS
15
a quadric qi−1 ∈ I so that qi−1 , . . . , qg+1 is a regular sequence. So by induction
we have a regular sequence of quadrics q2 , . . . , qg+1 in I, and we can take q1 to be
any other quadric independent from q2 , . . . , qg+1 since I is minimally generated by
g + 1 quadrics.
Let ℓ and h be be two independent linear syzygies on the qi . Then h1 ℓ − ℓ1 h
is a syzygy on q2 , . . . , qg+1 and, therefore, a linear combination of Koszul syzygies.
Write
X
h1 ℓ − ℓ1 h =
ai,j (qj ei − qi ej )
2≤i<j≤g+1
for some ai,j ∈ k, where e1 , . . . , eg+1 denotes the standard basis of S(−2)g+1 . Note
that ℓ1 and h1 must be independent linear forms, otherwise we could find a nontrivial linear syzygy on q2 , . . . , qg+1 since ℓ and h are independent, but that contradicts
that q2 , . . . , qg+1 is a regular sequence. If h1 ℓ − ℓ1 h = 0, then (−hi , ℓi ) = bi (h1 , −ℓ1 )
for some bi ∈ k for all i ≥ 2 so that h = h1 (1, −b2 , . . . , −bg+1 ). But then
(1, −b2 , . . . , −bg+1 ) must be a syzygy on the qi , contradicting that they are independent quadrics. Hence, we see that h1 ℓ − ℓ1 h 6= 0 so that ai,j 6= 0 for some i, j. Relabeling q2 , . . . , qg+1 if necessary, we may assume that a2,3 6= 0. Then by Remark 4.1,
we can replace q3 with q = h1 ℓ2 − ℓ1 h2 = a2,3 q3 + · · · + a2,g+1 qg+1 . In exchanging q3
−1
−1
for q, we must replace ℓ with ℓ̃ = (ℓ1 , ℓ2 , a−1
2,3 ℓ3 , ℓ4 −a2,3 a2,4 ℓ3 , . . . , ℓg+1 −a2,3 a2,g+1 ℓ3 )
as
g+1
g+1
X
X
−1
(ℓi − a−1
ℓi qi = a2,3 ℓ3 q + ℓ1 q1 + ℓ2 q2 +
0=
2,3 a2,i ℓ3 )qi
i=4
i=1
and we also replace h with the syzygy h̃ defined as above. However, ℓ̃ and h̃ are
still independent linear syzygies since their first coordinates are independent linear
forms. After making the above changes, we have a2,3 = 1 and a2,i = 0 for all
i > 3. Then h1 ℓ3 − ℓ1 h3 = −q2 + a3,4 q4 + · · · + a3,g+1 qg+1 , and we can we replace
q2 with −(h1 ℓ3 − ℓ1 h3 ) as above. In that case, we have q2 = −(h1 ℓ3 − ℓ1 h3 ) and
q3 = h1 ℓ2 − ℓ1 h2 so that
0=
g+1
X
i=1
ℓi qi = ℓ1 (q1 + ℓ2 h3 − ℓ3 h2 ) +
4
X
ℓi qi
i=1
implies that q1 + ℓ2 h3 − ℓ3 h2 ∈ ((q4 , . . . , qg+1 ) : ℓ1 ).
We claim that ℓ1 is a nonzerodivisor modulo (q4 , . . . , qg+1 ). If not, then ℓ1 is
contained in an associated prime of (q4 , . . . , qg+1 ) so that ht(ℓ1 , q4 , . . . , qg+1 ) =
g − 2. But then ht(ℓ1 , h1 , q4 , . . . , qg+1 ) ≤ ht(ℓ1 , q4 , . . . , qg+1 ) + 1 = g − 1, and
since (q2 , . . . , qg+1 ) ⊆ (ℓ1 , h1 , q4 , . . . , qg+1 ), this contradicts ht(q2 , . . . , qg+1 ) = g.
Therefore, ℓ1 is a nonzerodivisor modulo (q4 , . . . , qg+1 ) as claimed so that q1 +ℓ2 h3 −
ℓ3 h2 ∈ (q4 , . . . , qg+1 ). We can then write ℓ2 h3 − ℓ3 h2 = −q1 + c4 q4 + · · · + cg+1 qg+1
16
MATTHEW MASTROENI
for some ci ∈ k. Replacing q1 with ℓ2 h3 − ℓ3 h2 and setting
ℓ1 h1
M = ℓ2 h2
ℓ3 h3
yields I2 (M) = (q1 , q2 , q3 ) as wanted.
(4.1)
Theorem 4.3. If R = S/I is a quadratic almost complete intersection, then
S
β2,3
(R) ≤ 2.
S
Proof. Suppose that β2,3
(R) ≥ 3. Let K be an infinite extension field of k, and for
SK
S
each k-algebra A, set AK = A ⊗k K. Then β2,3
(RK ) = β2,3
(R), dim RK = dim R,
∼
∼
and ISK /I(SK )+ = I/IS+ ⊗S SK = I/IS+ ⊗k K by faithfully flat base change so
that ISK is still a quadratic almost complete intersection, and replacing R with
RK , we may assume that the ground field k is infinite.
By the preceding proposition, there are quadrics q1 , . . . , qg+1 for g = ht I ≥ 2
and a 3 × 2 matrix of linear forms M as in (4.1) such that I = (q1 , . . . , qg+1 ),
q2 , . . . , qg+1 is a regular sequence, and I2 (M) = (q1 , q2 , q3 ). We may assume that
q1 , q2 , q3 are the minors of M as in the proof of the proposition. In that case,
ℓ = (ℓ1 , ℓ2 , ℓ3 , 0, . . . , 0) and h = (h1 , h2 , h3 , 0, . . . , 0) are two independent linear
syzygies on the qi . Let u be a linear syzygy independent from ℓ and h. By
arguing as in the proof of the preceding proposition, we see that u1 , ℓ1 , and h1 are
independent linear forms and that u1 ℓ − ℓ1 u 6= 0.
We claim that u1 ℓ − ℓ1 u is linear independent from h1 ℓ − ℓ1 h. If not, then
u1 ℓ−ℓ1 u = c(h1 ℓ−ℓ1 h) for some nonzero c ∈ k. Setting ũ = u−ch and rearranging
the preceding equality, we have that ũ is a linear syzygy independent from ℓ and h
with ũ1 ℓ−ℓ1 ũ = 0, which is impossible as already noted in the previous paragraph.
Hence, u1 ℓ − ℓ1 u is independent from h1 ℓ − ℓ1 h as claimed. In particular, we note
that g ≥ 3 since there cannot be two independent quadric syzygies on the regular
sequence q2 , q3 . Write
X
u 1 ℓ − ℓ1 u =
bi,j (qj ei − qi ej )
2≤i<j≤g+1
for some bi,j ∈ k. Since h1 ℓ − ℓ1 h = (0, q3 , −q2 , 0, . . . , 0) by assumption, we must
have bi,j 6= 0 for some (i, j) 6= (2, 3), otherwise we would have a contradiction to
the claim. In fact, by replacing u with u − b2,3 h, we may assume that b2,3 = 0.
Next, we claim that there is a j ≥ 4 such that b2,j 6= 0 or b3,j 6= 0. If
not, the first three coordinates of u1 ℓ − ℓ1 u must be zero so that u1 ℓ − ℓ1 u =
(0, 0, 0, −ℓ1u4 , . . . , −ℓ1 ug+1). But this implies that (u4 , . . . , ug+1 ) is a linear syzygy
on q4 , . . . , qg+1 , which is impossible since q4 , . . . , qg+1 is a regular sequence. Hence,
KOSZUL ALMOST COMPLETE INTERSECTIONS
17
relabeling if necessary, we may assume that b2,4 6= 0 so that we can replace q4 with
u1 ℓ2 − ℓ1 u2 = b2,4 q4 + · · · + b2,g+1 qg+1 .
Finally, we claim that there is a j ≥ 5 such that b3,j 6= 0. If not, then u1 ℓ3 −
ℓ1 u3 = b3,4 q4 = b3,4 (u1 ℓ2 − ℓ1 u2 ) as b2,3 = 0. But this implies that (b3,4 u2 − u3 , ℓ3 −
b3,4 ℓ2 ) = r(u1, −ℓ1 ) for some r ∈ k as ℓ1 and u1 are independent linear forms. In
particular, we see that ℓ3 ∈ Span{ℓ1 , ℓ2 } so that I ⊆ (ℓ1 , ℓ2 , q5 , . . . , qg+1 ). However,
ht(ℓ1 , ℓ2 , q5 , . . . , qg+1 ) ≤ g − 1, which contradicts ht I = g. Hence, we must have
g ≥ 4, and after relabeling, we may assume that b3,5 6= 0 and replace q5 with
u1 ℓ3 − ℓ1 u3 = b3,4 q4 + b3,5 q5 + · · · + b3,g+1 qg+1 . But then I ⊆ (ℓ1 , ℓ2 , ℓ3 , q6 , . . . , qg+1 )
and ht(ℓ1 , ℓ2 , ℓ3 , q6 , . . . , qg+1 ) ≤ g − 1, contradicting that ht I = g. Therefore, we
S
must have β2,3
(R) ≤ 2.
5. Future Directions
Using our main result, we have reason to believe that Question 1.1 can be
answered affirmatively for Koszul algebras defined by g = 4 quadrics. However,
Example 2.2 points to unexpected difficulties when g = 5.
Our work also has unintended connections to various other conjectures of interest. For example, the Buchsbaum-Eisenbud-Horrocks Conjecture asks whether
βiS (R) ≥ ci for all i if c = ht I. This conjecture is already known to hold for quo
tients by monomial ideals. An even larger lower bound βiS (R) ≥ ci + c−1
was
i−1
given for monomial ideals I of finite colength that are not complete intersections
by Charalambous and Evans in [Cha91][CE91] and for almost complete intersections directly linked to a complete intersection by Dugger in [Dug00, 2.3]. Among
Koszul almost complete intersections R = S/I, those with one linear syzygy cannot
be directly linked to a complete intersection since they are not Cohen-Macaulay.
Nonetheless, they still satisfy this larger bound. For Koszul ACI’s with two linear
syzygies, at least when the ground field is infinite, the results of the previous section show that I is directly linked to the complete intersection (ℓ1 , h1 , q4 , . . . , qg+1 ),
and Theorem 3.3 yields that Dugger’s result is sharp. A total rank version of the
P
Charalambous-Evans bound i βi (R) ≥ 2c + 2c−1 has been studied in [CEM90],
and Boocher and Seiner have recently established this bound for all monomial
ideals which are not complete intersections. Our work affirmatively answers a
couple questions they pose [BS17, 1.2, 1.3] in the Koszul ACI case.
Another consequence of our structure theorem is that the EGH Conjecture holds
for Koszul almost complete intersections in a strong form similar to [CCV14, 2.1].
This conjecture asks whether, given a graded ideal I ⊆ S = k[x1 , . . . , xn ] containing a homogeneous regular sequence f1 , . . . , fr of degrees 2 ≤ d1 ≤ · · · ≤ dr , there
is a monomial ideal J containing xd11 , . . . , xdr r and having the same Hilbert function
18
MATTHEW MASTROENI
as I. The EGH Conjecture is known to hold when I is a quadratic monomial ideal
by the preceding paper, when I is a complete intersection of quadrics, or when I
is generated by products of linear forms [Abe15], which covers quadratic ideals of
height one. To this list, we add that, if I defines a Koszul almost complete intersection, we can simply take J = (x21 , . . . , x2g , xg xg+1 ) or J = (x21 , . . . , x2g , xg−1 xg )
according to whether I has one or two linear syzygies respectively. Note for the
former case that g + 1 = pdS R ≤ n by Hilbert’s Syzygy Theorem.
Given the previous and new evidence for the above conjectures in the Koszul
case, it is natural to ask whether they hold at the very least for Koszul algebras
in general.
Acknowledgments
The author would like to thank his advisor, Hal Schenck, for many helpful
comments and discussions throughout the development of this work. He would
also like to thank Aldo Conca for pointing out a mistake in an earlier draft of
this paper and the anonymous referee for helping to simplify the exposition and
clarify the history of various results cited. Evidence for the results in this paper
was provided by several computations in Macaulay2 [M2].
References
[Abe15]
[Avr98]
[ACI10]
[ACI15]
[Bac86]
[Bac88]
[BDG15]
[BHI16]
[BS17]
[BC03]
A. Abedelfatah. On the Eisenbud-Green-Harris conjecture. Proc. Amer. Math. Soc.
143 (2015), no. 1, 105–115.
L. Avramov. Infinite free resolutions. Six lectures on commutative algebra (Bellaterra,
1996), 1–118, Progr. Math., 166, Birkhuser, Basel, 1998.
L. Avramov, A. Conca, and S. Iyengar. Free resolutions over commutative Koszul
algebras. Math. Res. Lett. 17 (2010), no. 2, 197–210.
L. Avramov, A. Conca, and S. Iyengar. Subadditivity of syzygies of Koszul algebras.
Math. Ann. 361 (2015), no. 1-2, 511–534.
J. Backelin. On the rates of growth of the homologies of Veronese subrings. Algebra,
algebraic topology and their interactions (Stockholm, 1983), 79–100, Lecture Notes in
Math., 1183, Springer, Berlin, 1986.
J. Backelin. Relations between rates of growth of homologies. Research Reports in
Mathematics 25, Department of Mathematics (1988), Stockholm University.
A. Boocher, A. D’Alı̀, E. Grifo, et al. Edge ideals and DG algebra resolutions. Matematiche (Catania) 70 (2015), no. 1, 215–238.
A. Boocher, S. H. Hassanzadeh, and S. Iyengar. Koszul algebras defined by three
relations. arXiv:1612.01558 [math.AC]
A. Boocher and J. Seiner. Lower bounds for Betti numbers of monomial ideals.
arXiv:1706.09866 [math.AC]
W. Bruns and A. Conca. Gröbner bases and determinantal ideals. Commutative algebra, singularities and computer algebra (Sinaia, 2002), 9–66, NATO Sci. Ser. II Math.
Phys. Chem., 115, Kluwer Acad. Publ., Dordrecht, 2003.
KOSZUL ALMOST COMPLETE INTERSECTIONS
[BH93]
19
W. Bruns and J. Herzog. Cohen-Macaulay rings. Cambridge Studies in Advanced
Mathematics, 39. Cambridge University Press, Cambridge, 1993.
[CCV14] G. Caviglia, A. Constantinescu, and M. Varbaro. On a conjecture by Kalai. Israel J.
Math. 204 (2014), no. 1, 469–475.
[Cha91] H. Charalambous. Betti numbers of multigraded modules. J. Algebra 137 (1991), no.
2, 491–500.
[CE91]
H. Charalambous and E. G. Evans. A deformation theory approach to Betti numbers
of finite length modules. J. Algebra 143 (1991), no. 1, 246–251.
[CEM90] H. Charalambous, E. G. Evans, and M. Miller. Betti numbers for modules of finite
length. Proc. Amer. Math. Soc. 109 (1990), no. 1, 63–70.
[Con14] A. Conca. Koszul algebras and their syzygies. Combinatorial algebraic geometry, 1–31,
Lecture Notes in Math., 2108, Fond. CIME/CIME Found. Subser., Springer, Cham,
2014.
[CDR13] A. Conca, E. De Negri, and M.E. Rossi. Koszul algebras and regularity. Commutative
algebra, 285–315, Springer, New York, 2013.
[Dug00] D. Dugger. Betti numbers of almost complete intersections. Illinois J. Math. 44 (2000),
no. 3, 531–541.
[Eis95]
D. Eisenbud. Commutative algebra. With a view toward algebraic geometry. Graduate
Texts in Mathematics, 150. Springer-Verlag, New York, 1995.
[Fro75]
R. Fröberg. Determination of a class of Poincaré series. Math. Scand. 37 (1975), no.
1, 29–39.
[M2]
D. Grayson and M. Stillman. Macaulay2, a software system for research in algebraic
geometry. Available at http://www.math.uiuc.edu/Macaulay2/.
[Kat06]
M. Katzman. Characteristic-independence of Betti numbers of graph ideals. J. Combin. Theory Ser. A 113 (2006), no. 3, 435–454.
[Kem90] G. Kempf. Some wonderful rings in algebraic geometry. J. Algebra 134 (1990), no. 1,
222–224.
[Kem92] G. Kempf. Syzygies for points in projective space. J. Algebra 145 (1992), no. 1, 219–
223.
[MS05]
E. Miller and B. Sturmfels. Combinatorial commutative algebra. Graduate Texts in
Mathematics, 227. Springer-Verlag, New York, 2005.
[PS73]
C. Peskine and L. Szpiro. Dimension projective finie et cohomologie locale. Applications à la démonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck.
Inst. Hautes Études Sci. Publ. Math. No. 42 (1973), 47–119.
[Pol95]
A. Polishchuk. On the Koszul property of the homogeneous coordinate ring of a curve.
J. Algebra 178 (1995), no. 1, 122–135.
[Ser00]
J.-P. Serre. Local algebra. Translated from the French by CheeWhye Chin and revised
by the author. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2000.
[Stu90]
B. Sturmfels. Gröbner bases and Stanley decompositions of determinantal rings. Math.
Z. 205 (1990), no. 1, 137–144.
Department of Mathematics, University of Illinois, Urbana, IL 61801
E-mail address: [email protected]
| 0math.AC
|
CARLA: An Open Urban Driving Simulator
Alexey Dosovitskiy1, German Ros2,3, Felipe Codevilla1,3, Antonio López3, and Vladlen Koltun1
arXiv:1711.03938v1 [cs.LG] 10 Nov 2017
1
Intel Labs
2
Toyota Research Institute
3
Computer Vision Center, Barcelona
Abstract: We introduce CARLA, an open-source simulator for autonomous driving research. CARLA has been developed from the ground up to support development, training, and validation of autonomous urban driving systems. In addition to open-source code and protocols, CARLA provides open digital assets
(urban layouts, buildings, vehicles) that were created for this purpose and can
be used freely. The simulation platform supports flexible specification of sensor
suites and environmental conditions. We use CARLA to study the performance
of three approaches to autonomous driving: a classic modular pipeline, an endto-end model trained via imitation learning, and an end-to-end model trained via
reinforcement learning. The approaches are evaluated in controlled scenarios of
increasing difficulty, and their performance is examined via metrics provided by
CARLA, illustrating the platform’s utility for autonomous driving research.
Keywords: Autonomous driving, sensorimotor control, simulation
1
Introduction
Sensorimotor control in three-dimensional environments remains a major challenge in machine
learning and robotics. The development of autonomous ground vehicles is a long-studied instantiation of this problem [22, 26]. Its most difficult form is navigation in densely populated urban
environments [21]. This setting is particularly challenging due to complex multi-agent dynamics at
traffic intersections; the necessity to track and respond to the motion of tens or hundreds of other
actors that may be in view at any given time; prescriptive traffic rules that necessitate recognizing
street signs, street lights, and road markings and distinguishing between multiple types of other vehicles; the long tail of rare events – road construction, a child running onto the road, an accident
ahead, a rogue driver barreling on the wrong side; and the necessity to rapidly reconcile conflicting
objectives, such as applying appropriate deceleration when an absent-minded pedestrian strays onto
the road ahead but another car is rapidly approaching from behind and may rear-end if one brakes
too hard.
Research in autonomous urban driving is hindered by infrastructure costs and the logistical difficulties of training and testing systems in the physical world. Instrumenting and operating even one
robotic car requires significant funds and manpower. And a single vehicle is far from sufficient for
collecting the requisite data that cover the multitude of corner cases that must be processed for both
training and validation. This is true for classic modular pipelines [21, 8] and even more so for datahungry deep learning techniques. Training and validation of sensorimotor control models for urban
driving in the physical world is beyond the reach of most research groups.
An alternative is to train and validate driving strategies in simulation. Simulation can democratize
research in autonomous urban driving. It is also necessary for system verification, since some scenarios are too dangerous to be staged in the physical world (e.g., a child running onto the road ahead
of the car). Simulation has been used for training driving models since the early days of autonomous
driving research [22]. More recently, racing simulators have been used to evaluate new approaches
to autonomous driving [28, 3]. Custom simulation setups are commonly used to train and benchmark robotic vision systems [2, 9, 10, 11, 20, 25, 27, 29]. And commercial games have been used
to acquire high-fidelity data for training and benchmarking visual perception systems [23, 24].
While ad-hoc use of simulation in autonomous driving research is widespread, existing simulation
platforms are limited. Open-source racing simulators such as TORCS [28] do not present the com1st Conference on Robot Learning (CoRL 2017), Mountain View, United States.
Figure 1: A street in Town 2, shown from a third-person view in four weather conditions. Clockwise from top left: clear day, daytime rain, daytime shortly after rain, and clear sunset. See the
supplementary video for recordings from the simulator.
plexity of urban driving: they lack pedestrians, intersections, cross traffic, traffic rules, and other
complications that distinguish urban driving from track racing. And commercial games that simulate urban environments at high fidelity, such as Grand Theft Auto V [23, 24], do not support detailed
benchmarking of driving policies: they have little customization and control over the environment,
limited scripting and scenario specification, severely limited sensor suite specification, no detailed
feedback upon violation of traffic rules, and other limitations due to their closed-source commercial
nature and fundamentally different objectives during their development.
In this paper, we introduce CARLA (Car Learning to Act) – an open simulator for urban driving.
CARLA has been developed from the ground up to support training, prototyping, and validation of
autonomous driving models, including both perception and control. CARLA is an open platform.
Uniquely, the content of urban environments provided with CARLA is also free. The content was
created from scratch by a dedicated team of digital artists employed for this purpose. It includes
urban layouts, a multitude of vehicle models, buildings, pedestrians, street signs, etc. The simulation
platform supports flexible setup of sensor suites and provides signals that can be used to train driving
strategies, such as GPS coordinates, speed, acceleration, and detailed data on collisions and other
infractions. A wide range of environmental conditions can be specified, including weather and time
of day. A number of such environmental conditions are illustrated in Figure 1.
We use CARLA to study the performance of three approaches to autonomous driving. The first is a
classic modular pipeline that comprises a vision-based perception module, a rule-based planner, and
a maneuver controller. The second is a deep network that maps sensory input to driving commands,
trained end-to-end via imitation learning. The third is also a deep network, trained end-to-end via
reinforcement learning. We use CARLA to stage controlled goal-directed navigation scenarios of
increasing difficulty. We manipulate the complexity of the route that must be traversed, the presence
of traffic, and the environmental conditions. The experimental results shed light on the performance
characteristics of the three approaches.
2
Simulation Engine
CARLA has been built for flexibility and realism in the rendering and physics simulation. It is
implemented as an open-source layer over Unreal Engine 4 (UE4) [7], enabling future extensions by
2
the community. The engine provides state-of-the-art rendering quality, realistic physics, basic NPC
logic, and an ecosystem of interoperable plugins. The engine itself is free for non-commercial use.
CARLA simulates a dynamic world and provides a simple interface between the world and an agent
that interacts with the world. To support this functionality, CARLA is designed as a server-client
system, where the server runs the simulation and renders the scene. The client API is implemented
in Python and is responsible for the interaction between the autonomous agent and the server via
sockets. The client sends commands and meta-commands to the server and receives sensor readings
in return. Commands control the vehicle and include steering, accelerating, and braking. Metacommands control the behavior of the server and are used for resetting the simulation, changing the
properties of the environment, and modifying the sensor suite. Environmental properties include
weather conditions, illumination, and density of cars and pedestrians. When the server is reset, the
agent is re-initialized at a new location specified by the client.
Environment. The environment is composed of 3D models of static objects such as buildings, vegetation, traffic signs, and infrastructure, as well as dynamic objects such as vehicles and pedestrians.
All models are carefully designed to reconcile visual quality and rendering speed: we use low-weight
geometric models and textures, but maintain visual realism by carefully crafting the materials and
making use of variable level of detail. All 3D models share a common scale, and their sizes reflect
those of real objects. At the time of writing, our asset library includes 40 different buildings, 16
animated vehicle models, and 50 animated pedestrian models.
We used these assets to build urban environments via the following steps: (a) laying out roads
and sidewalks; (b) manually placing houses, vegetation, terrain, and traffic infrastructure; and (c)
specifying locations where dynamic objects can appear (spawn). This way we have designed two
towns: Town 1 with a total of 2.9 km of drivable roads, used for training, and Town 2 with 1.4 km
of drivable roads, used for testing. The two towns are shown in the supplement.
One of the challenges in the development of CARLA was the configuration of the behavior of
non-player characters, which is important for realism. We based the non-player vehicles on the
standard UE4 vehicle model (PhysXVehicles). Kinematic parameters were adjusted for realism.
We also implemented a basic controller that governs non-player vehicle behavior: lane following,
respecting traffic lights, speed limits, and decision making at intersections. Vehicles and pedestrians
can detect and avoid each other. More advanced non-player vehicle controllers can be integrated in
the future [1].
Pedestrians navigate the streets according to a town-specific navigation map, which conveys a
location-based cost. This cost is designed to encourage pedestrians to walk along sidewalks and
marked road crossings, but allows them to cross roads at any point. Pedestrians wander around town
in accordance with this map, avoiding each other and trying to avoid vehicles. If a car collides with
a pedestrian, the pedestrian is deleted from the simulation and a new pedestrian is spawned at a
different location after a brief time interval.
To increase visual diversity, we randomize the appearance of non-player characters when they are
added to the simulation. Each pedestrian is clothed in a random outfit sampled from a pre-specified
wardrobe and is optionally equipped with one or more of the following: a smartphone, shopping
bags, a guitar case, a suitcase, a rolling bag, or an umbrella. Each vehicle is painted at random
according to a model-specific set of materials.
We have also implemented a variety of atmospheric conditions and illumination regimes. These
differ in the position and color of the sun, the intensity and color of diffuse sky radiation, as well as
ambient occlusion, atmospheric fog, cloudiness, and precipitation. Currently, the simulator supports
two lighting conditions – midday and sunset – as well as nine weather conditions, differing in cloud
cover, level of precipitation, and the presence of puddles in the streets. This results in a total of 18
illumination-weather combinations. (In what follows we refer to these as weather, for brevity.) Four
of these are illustrated in Figure 1.
Sensors. CARLA allows for flexible configuration of the agent’s sensor suite. At the time of writing,
sensors are limited to RGB cameras and to pseudo-sensors that provide ground-truth depth and
semantic segmentation. These are illustrated in Figure 2. The number of cameras and their type and
position can be specified by the client. Camera parameters include 3D location, 3D orientation with
respect to the car’s coordinate system, field of view, and depth of field. Our semantic segmentation
3
Figure 2: Three of the sensing modalities provided by CARLA. From left to right: normal vision
camera, ground-truth depth, and ground-truth semantic segmentation. Depth and semantic segmentation are pseudo-sensors that support experiments that control for the role of perception. Additional
sensor models can be plugged in via the API.
pseudo-sensor provides 12 semantic classes: road, lane-marking, traffic sign, sidewalk, fence, pole,
wall, building, vegetation, vehicle, pedestrian, and other.
In addition to sensor and pseudo-sensor readings, CARLA provides a range of measurements associated with the state of the agent and compliance with traffic rules. Measurements of the agent’s state
include vehicle location and orientation with respect to the world coordinate system (akin to GPS
and compass), speed, acceleration vector, and accumulated impact from collisions. Measurements
concerning traffic rules include the percentage of the vehicle’s footprint that impinges on wrong-way
lanes or sidewalks, as well as states of the traffic lights and the speed limit at the current location of
the vehicle. Finally, CARLA provides access to exact locations and bounding boxes of all dynamic
objects in the environment. These signals play an important role in training and evaluating driving
policies.
3
Autonomous Driving
CARLA supports development, training, and detailed performance analysis of autonomous driving
systems. We have used CARLA to evaluate three approaches to autonomous driving. The first is
a modular pipeline that relies on dedicated subsystems for visual perception, planning, and control.
This architecture is in line with most existing autonomous driving systems [21, 8]. The second
approach is based on a deep network trained end-to-end via imitation learning [4]. This approach
represents a long line of investigation that has recently attracted renewed interest [22, 16, 4]. The
third approach is based on a deep network trained end-to-end via reinforcement learning [19].
We begin by introducing notation that is common to all methods and then proceed to describe each
in turn. Consider an agent that interacts with the environment over discrete time steps. At each
time step, the agent gets an observation ot and must produce an action at . The action is a threedimensional vector that represents the steering, throttle, and brake. The observation ot is a tuple of
sensory inputs. This can include high-dimensional sensory observations, such as color images and
depth maps, and lower-dimensional measurements, such as speed and GPS readings.
In addition to momentary observations, all approaches also make use of a plan provided by a highlevel topological planner. This planner takes the current position of the agent and the location of the
goal as input, and uses the A∗ algorithm to provide a high-level plan that the agent needs to follow
in order to reach the goal. This plan advises the agent to turn left, turn right, or keep straight at
intersections. The plan does not provide a trajectory and does not contain geometric information. It
is thus a weaker form of the plan that is given by common GPS navigation applications which guide
human drivers and autonomous vehicles in the physical world. We do not use metric maps.
3.1
Modular pipeline
Our first method is a modular pipeline that decomposes the driving task among the following subsystems: (i) perception, (ii) planning, and (iii) continuous control. Since no metric map is provided
as input, visual perception becomes a critical task. Local planning is completely dependent on the
scene layout estimated by the perception module.
The perception stack uses semantic segmentation to estimate lanes, road limits, and dynamic objects
and other hazards. In addition, a classification model is used to determine proximity to intersections.
4
The local planner uses a rule-based state machine that implements simple predefined polices tuned
for urban environments. Continuous control is performed by a PID controller that actuates the
steering, throttle, and brake. We now describe the modules in more detail.
Perception. The perception stack we describe here is built upon a semantic segmentation network
based on RefineNet [17]. The network is trained to classify each pixel in the image into one of the
following semantic categories: C = {road, sidewalk, lane marking, dynamic object,
miscellaneous static}. The network is trained on 2,500 labelled images produced in the
training environment using CARLA. The probability distributions provided by the network are used
to estimate the ego-lane based on the road area and the lane markings. The network output is also
used to compute an obstacle mask that aims to encompass pedestrians, vehicles, and other hazards.
In addition, we estimate the likelihood of being at an intersection by using a binary scene classifier (intersection/no intersection) based on AlexNet [15]. This network is trained on 500 images
balanced between the two classes.
Local planner. The local planner coordinates low-level navigation by generating a set of waypoints:
near-term goal states that represent the desired position and orientation of the car in the near future.
The goal of the planner is to synthesize waypoints that keep the car on the road and prevent collisions. The local planner is based on a state machine with the following states: (i) road-following,
(ii) left-turn, (iii) right-turn, (iv) intersection-forward, and (v) hazard-stop. Transitions between
states are performed based on estimates provided by the perception module and on topological information provided by the global planner. Further details can be found in the supplement. The local
plan in the form of waypoints is delivered to the controller, along with the vehicle’s current pose and
speed.
Continuous controller. We use a proportional-integral-derivative (PID) controller [6] due to its
simplicity, flexibility, and relative robustness to slow response times. Each controller receives the
current pose, speed, and a list of waypoints, and actuates the steering, throttle and brake, respectively.
We target a cruise speed of 20 km/h. Controller parameters were tuned in the training town.
3.2
Imitation learning
Our second method is conditional imitation learning, a form of imitation learning that uses highlevel commands in addition to perceptual input [4]. This method utilizes a dataset of driving traces
recorded by human drivers in the training town. The dataset D = {hoi , ci , ai i} consists of tuples,
each of which contains an observation oi , a command ci , and an action ai . The commands are
provided by drivers during data collection and indicate their intentions, akin to turn signals. We use
a set of four commands: follow the lane (default), drive straight at the next intersection, turn left
at the next intersection, and turn right at the next intersection. The observations are images from a
forward-facing camera. To increase the robustness of the learned policies, we inject noise during
data collection. The dataset is used to train a deep network to predict the expert’s action a given an
observation o and a control command c. Further details are provided by Codevilla et al. [4].
We have collected around 14 hours of driving data for training. The network was trained using the
Adam optimizer [14]. To improve generalization, we performed data augmentation and dropout.
Further details are provided in the supplement.
3.3
Reinforcement learning
Our third method is deep reinforcement learning, which trains a deep network based on a reward
signal provided by the environment, with no human driving traces. We use the asynchronous advantage actor-critic (A3C) algorithm [19]. This algorithm has been shown to perform well in simulated
three-dimensional environments on tasks such as racing [19] and navigation in three-dimensional
mazes [19, 13, 5]. The asynchronous nature of the method enables running multiple simulation
threads in parallel, which is important given the high sample complexity of deep reinforcement
learning.
We train A3C on goal-directed navigation. In each training episode the vehicle has to reach a goal,
guided by high-level commands from the topological planner. The episode is terminated when
the vehicle reaches the goal, when the vehicle collides with an obstacle, or when a time budget is
exhausted. The reward is a weighted sum of five terms: positively weighted speed and distance
5
traveled towards the goal, and negatively weighted collision damage, overlap with the sidewalk, and
overlap with the opposite lane. Further details are provided in the supplement.
The network was trained with 10 parallel actor threads, for a total of 10 million simulation steps. We
limit training to 10 million simulation steps because of computational costs imposed by the realistic
simulation. This correspond to roughly 12 days of non-stop driving at 10 frames per second. This
is considered limited training data by deep reinforcement learning standards, where it is common
to train for hundreds of millions of steps [19], corresponding to months of subjective experience.
To ensure that our setup is fair and that 10 million simulation steps are sufficient for learning to act
in a complex environment, we trained a copy of our A3C agent to navigate in a three-dimensional
maze (task D2 from Dosovitskiy and Koltun [5]). The agent reached a score of 65 out of 100 after
10 million simulation steps – a good result compared to 60 out of 100 reported by Dosovitskiy and
Koltun [5] after 50 million simulation steps for A3C with less optimized hyperparameters.
4
Experiments
We evaluate the three methods – modular pipeline (MP), imitation learning (IL), and reinforcement
learning (RL) – on four increasingly difficult driving tasks, in each of the two available towns, in
six weather conditions. Note that for each of the three approaches we use the same agent on all
four tasks and do not fine-tune separately for each scenario. The tasks are set up as goal-directed
navigation: an agent is initialized somewhere in town and has to reach a destination point. In these
experiments, the agent is allowed to ignore speed limits and traffic lights. We organize the tasks in
order of increasing difficulty as follows:
• Straight: Destination is straight ahead of the starting point, and there are no dynamic objects
in the environment. Average driving distance to the goal is 200 m in Town 1 and 100 m in
Town 2.
• One turn: Destination is one turn away from the starting point; no dynamic objects. Average driving distance to the goal is 400 m in Town 1 and 170 m in Town 2.
• Navigation: No restriction on the location of the destination point relative to the starting
point, no dynamic objects. Average driving distance to the goal is 770 m in Town 1 and
360 m in Town 2.
• Navigation with dynamic obstacles: Same as the previous task, but with dynamic objects
(cars and pedestrians).
Experiments are conducted in two towns. Town 1 is used for training, Town 2 for testing. We
consider six weather conditions for the experiments, organized in two groups. Training Weather Set
was used for training and includes clear day, clear sunset, daytime rain, and daytime after rain. Test
Weather Set was never used during training and includes cloudy daytime and soft rain at sunset.
For each combination of a task, a town, and a weather set, testing is carried out over 25 episodes. In
each episode, the objective is to reach a given goal location. An episode is considered successful if
the agent reaches the goal within a time budget. The time budget is set to the time needed to reach
the goal along the optimal path at a speed of 10 km/h. Infractions, such as driving on the sidewalk
or collisions, do not lead to termination of an episode, but are logged and reported.
5
Results
Table 1 reports the percentage of successfully completed episodes under four different conditions.
The first is the training condition: Town 1, Training Weather Set. Note that start and goal locations
are different from those used during training: only the general environment and ambient conditions
are the same. The other three experimental conditions test more aggressive generalization: to the
previously unseen Town 2 and to previously unencountered weather from the Test Weather Set.
Results presented in Table 1 suggest several general conclusions. Overall, the performance of all
methods is not perfect even on the simplest task of driving in a straight line, and the success rate
further declines for more difficult tasks. Generalization to new weather is easier than generalization
to a new town. The modular pipeline and the agent trained with imitation learning perform on
par on most tasks and conditions. Reinforcement learning underperforms relative to the other two
approaches. We now discuss these four key findings in more detail.
6
Task
Training conditions
MP IL
RL
Straight
One turn
Navigation
Nav. dynamic
98
82
80
77
95
89
86
83
89
34
14
7
New town
MP IL RL
New weather
MP IL RL
92
61
24
24
100
95
94
89
74
12
3
2
97
59
40
38
98
90
84
82
86
16
2
2
New town & weather
MP
IL
RL
50
50
47
44
80
48
44
42
68
20
6
4
Table 1: Quantitative evaluation of three autonomous driving systems on goal-directed navigation
tasks. The table reports the percentage of successfully completed episodes in each condition. Higher
is better. The tested methods are: modular pipeline (MP), imitation learning (IL), and reinforcement
learning (RL).
Performance on the four tasks. Surprisingly, none of the methods performs perfectly even on
the simplest task of driving straight on an empty street in the training conditions. We believe the
fundamental reason for this is variability in the sensory inputs encountered by the agents. Training
conditions include four different weather conditions. The exact trajectories driven during training
are not repeated during testing. Therefore performing perfectly on this task requires robust generalization, which is challenging for existing deep learning methods.
On more advanced tasks the performance of all methods declines. On the most difficult task of
navigation in a populated urban environment, the two best methods – modular pipeline and imitation
learning – are below 90% success in all conditions and are below 45% in the test town. These
results clearly indicate that performance is far from saturated even in the training conditions, and
that generalization to new environments poses a serious challenge.
Generalization. We study two types of generalization: to previously unseen weather conditions and
to a previously unseen environment. Interestingly, the results are dramatically different for these
two. For the modular pipeline and for imitation learning, the performance in the “New weather”
condition is very close to performance in the training condition, and sometimes even better. However, generalization to a new town presents a challenge for all three approaches. On the two most
challenging navigation tasks, the performance of all methods falls by at least a factor of 2 when
switching to the test town. This phenomenon can be explained by the fact that the models have been
trained in multiple weather conditions, but in a single town. Training with diverse weather supports
generalization to previously unseen weather, but not to a new town, which uses different textures and
3D models. The problem can likely be ameliorated by training in diverse environments. Overall,
our results highlight the importance of generalization for learning-based approaches to sensorimotor
control.
Modular pipeline vs end-to-end learning. It is instructive to analyze the relative performance
of the modular pipeline and the imitation learning approach. These systems represent two general approaches to designing intelligent agents, and CARLA enables a direct controlled comparison
between them.
Surprisingly, the performance of both systems is very close under most testing conditions: the performance of the two methods typically differs by less than 10%. There are two notable exceptions
to this general rule. One is that the modular pipeline performs better under the “New weather” condition than under the training conditions. This is due to the specific selection of training and test
weathers: the perception system happens to perform better on the test weathers. Another difference
between the two approaches is that MP underperforms on navigation in the “New town” condition
and on going straight in “New town & weather”. This is because the perception stack fails systematically under complex weather conditions in the context of a new environment. If the perception
stack is not able to reliably find a drivable path, the rules-based planner and the classic controller
are unable to navigate to the destination in a consistent way. The performance is therefore bimodal:
if the perception stack works, the whole system works well; otherwise it fails completely. In this
sense, MP is more fragile than the end-to-end method.
Imitation learning vs reinforcement learning. We now contrast the performance of the two endto-end trained systems: imitation learning and reinforcement learning. On all tasks, the agent trained
with reinforcement learning performs significantly worse than the one trained with imitation learn7
Task
Training conditions
MP
IL
RL
MP
Opposite lane
Sidewalk
Collision-static
Collision-car
Collision-pedestrian
10.2
18.3
10.0
16.4
18.9
0.45
0.46
0.44
0.51
1.40
33.4
12.9
5.38
3.26
6.35
0.18
0.75
0.42
0.58
17.8
New town
IL
RL
1.12
0.76
0.40
0.59
1.88
0.23
0.43
0.23
0.41
2.55
New weather
MP
IL
RL
New town & weather
MP
IL
RL
16.1
24.2
16.1
20.2
20.4
0.40
0.43
0.45
0.47
1.46
57.3
> 57
4.05
1.86
11.2
0.09
0.72
0.24
0.85
20.6
0.78
0.81
0.28
0.44
1.41
0.21
0.48
0.25
0.37
2.99
Table 2: Average distance (in kilometers) traveled between two infractions. Higher is better.
ing. This is despite the fact that RL was trained using a significantly larger amount of data: 12 days
of driving, compared to 14 hours used by imitation learning. Why does RL underperform, despite
strong results on tasks such as Atari games [18, 19] and maze navigation [19, 5]? One reason is that
RL is known to be brittle [12], and it is common to perform extensive task-specific hyperparameter
search, such as 50 trials per environment as reported by Mnih et al. [19]. When using a realistic
simulator, such extensive hyperparameter search becomes infeasible. We selected hyperparameters
based on evidence from the literature and exploratory experiments with maze navigation. Another
explanation is that urban driving is more difficult than most tasks previously addressed with RL.
For instance, compared to maze navigation, in a driving scenario the agent has to deal with vehicle
dynamics and more complex visual perception in a cluttered dynamic environment. Finally, the poor
generalization of reinforcement learning may be explained by the fact that in contrast with imitation
learning, RL has been trained without data augmentation or regularization such as dropout.
Infraction analysis. CARLA supports fine-grained analysis of driving policies. We now examine
the behavior of the three systems on the hardest task: navigation in the presence of dynamic objects.
We characterize the approaches by average distance traveled between infractions of the following
five types: driving on the opposite lane, driving on the sidewalk, colliding with other vehicles,
colliding with pedestrians, and hitting static objects. Details are provided in the supplement.
Table 2 reports the average distance (in kilometers) driven between two infractions. All approaches
perform better in the training town. For all conditions, IL strays onto the opposite lane least frequently, and RL is the worst in this metric. A similar pattern is observed with regards to veering
onto the sidewalk. Surprisingly, RL collides with pedestrians least often, which could be explained
by the large negative reward incurred by such collisions. However, the reinforcement learning agent
is not successful at avoiding collisions with cars and static objects, while the modular pipeline generally performs best according to this measure.
These results highlight the susceptibility of end-to-end approaches to rare events: breaking or swerving to avoid a pedestrian is a rare occurrence during training. While CARLA can be used to increase
the frequency of such events during training to support end-to-end approaches, deeper advances
in learning algorithms and model architectures may be necessary for significant improvements in
robustness [3].
6
Conclusion
We have presented CARLA, an open simulator for autonomous driving. In addition to open-source
code and protocols, CARLA provides digital assets that were created specifically for this purpose
and can be reused freely. We leverage CARLA’s simulation engine and content to test three approaches to autonomous driving: a classic modular pipeline, a deep network trained end-to-end via
imitation learning, and a deep network trained via reinforcement learning. We challenged these
systems to navigate urban environments in the presence of other vehicles and pedestrians. CARLA
provided us with the tools to develop and train the systems and then evaluate them in controlled
scenarios. The feedback provided by the simulator enables detailed analyses that highlight particular failure modes and opportunities for future work. We hope that CARLA will enable a broad
community to actively engage in autonomous driving research. The simulator and accompanying
assets will be released open-source at http://carla.org.
8
Acknowledgements
CARLA would not have been possible without the development team at the CVC in Barcelona. The
authors are particularly grateful to Nestor Subiron, the principal programmer, and Francisco Perez,
the lead digital artist, for their tireless work. We sincerely thank artists Iris Saez and Alberto Abal,
FX programmer Marc Garcia, and traffic behavior programmer Francisco Bosch. We thank artists
Mario Gonzalez, Juan Gonzalez, and Ignazio Acerenza for their contributions, and programmer
Francisco Molero for his support. Antonio M. López and Felipe Codevilla acknowledge the Spanish
MINECO project TRA2014-57088-C2-1-R and the Spanish DGT project SPIP2017-02237, as well
as the Generalitat de Catalunya CERCA Program and its ACCIO agency. Felipe Codevilla was supported in part by FI grant 2017FI-B1-00162. The authors thank Epic Games for support concerning
the use of UE4.
References
[1] A. Best, S. Narang, D. Barber, and D. Manocha. AutonoVi: Autonomous vehicle planning with dynamic
maneuvers and traffic constraints. In Intelligent Robots and Systems (IROS), 2017.
[2] D. H. Biedermann, M. Ochs, and R. Mester. Evaluating visual ADAS components on the COnGRATS
dataset. In IEEE Intelligent Vehicles Symposium, 2016.
[3] C. Chen, A. Seff, A. L. Kornhauser, and J. Xiao. DeepDriving: Learning affordance for direct perception
in autonomous driving. In International Conference on Computer Vision (ICCV), 2015.
[4] F. Codevilla, M. Müller, A. Dosovitskiy, A. López, and V. Koltun. End-to-end driving via conditional
imitation learning. arXiv:1710.02410, 2017.
[5] A. Dosovitskiy and V. Koltun. Learning to act by predicting the future. In International Conference on
Learning Representations (ICLR), 2017.
[6] M. T. Emirler, I. M. C. Uygan, B. A. Güvenç, and L. Güvenç. Robust PID steering control in parameter
space for highly automated driving. International Journal of Vehicular Technology, 2014.
[7] Epic Games. Unreal Engine 4. https://www.unrealengine.com.
[8] U. Franke. Autonomous driving. In Computer Vision in Vehicle Technology. 2017.
[9] A. Gaidon, Q. Wang, Y. Cabon, and E. Vig. Virtual worlds as proxy for multi-object tracking analysis. In
Computer Vision and Pattern Recognition (CVPR), 2016.
[10] V. Haltakov, C. Unger, and S. Ilic. Framework for generation of synthetic ground truth data for driver
assistance applications. In German Conference on Pattern Recognition (GCPR), 2013.
[11] A. Handa, R. A. Newcombe, A. Angeli, and A. J. Davison. Real-time camera tracking: When is high
frame-rate best? In European Conference on Computer Vision (ECCV), 2012.
[12] P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, and D. Meger. Deep reinforcement learning
that matters. arXiv:1709.06560, 2017.
[13] M. Jaderberg, V. Mnih, W. M. Czarnecki, T. Schaul, J. Z. Leibo, D. Silver, and K. Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. In International Conference on Learning Representations (ICLR), 2017.
[14] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference on
Learning Representations (ICLR), 2015.
[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural
networks. In Neural Information Processing Systems (NIPS), 2012.
[16] Y. LeCun, U. Muller, J. Ben, E. Cosatto, and B. Flepp. Off-road obstacle avoidance through end-to-end
learning. In Neural Information Processing Systems (NIPS), 2005.
[17] G. Lin, A. Milan, C. Shen, and I. Reid. RefineNet: Multi-path refinement networks for high-resolution
semantic segmentation. In Computer Vision and Pattern Recognition (CVPR), 2017.
[18] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller,
A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, et al. Human-level control through deep
reinforcement learning. Nature, 518(7540), 2015.
9
[19] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu.
Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning (ICML), 2016.
[20] M. Müller, N. Smith, and B. Ghanem. A benchmark and simulator for UAV tracking. In European
Conference on Computer Vision (ECCV), 2016.
[21] B. Paden, M. Cáp, S. Z. Yong, D. S. Yershov, and E. Frazzoli. A survey of motion planning and control
techniques for self-driving urban vehicles. IEEE Transactions on Intelligent Vehicles, 1(1), 2016.
[22] D. Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In Neural Information Processing Systems (NIPS), 1988.
[23] S. R. Richter, V. Vineet, S. Roth, and V. Koltun. Playing for data: Ground truth from computer games. In
European Conference on Computer Vision (ECCV), 2016.
[24] S. R. Richter, Z. Hayder, and V. Koltun. Playing for benchmarks. In International Conference on Computer Vision (ICCV), 2017.
[25] G. Ros, L. Sellart, J. Materzynska, D. Vázquez, and A. López. The SYNTHIA dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Computer Vision and Pattern
Recognition (CVPR), 2016.
[26] D. Silver, J. A. Bagnell, and A. Stentz. Learning from demonstration for autonomous navigation in
complex unstructured terrain. International Journal of Robotics Research, 29(12), 2010.
[27] J. Skinner, S. Garg, N. Sünderhauf, P. I. Corke, B. Upcroft, and M. Milford. High-fidelity simulation for
evaluating robotic vision performance. In Intelligent Robots and Systems (IROS), 2016.
[28] B. Wymann, E. Espié, C. Guionneau, C. Dimitrakakis, R. Coulom, and A. Sumner. TORCS, The Open
Racing Car Simulator. http://www.torcs.org.
[29] Z. Zhang, H. Rebecq, C. Forster, and D. Scaramuzza. Benefit of large field-of-view cameras for visual
odometry. In International Conference on Robotics and Automation (ICRA), 2016.
10
Supplementary Material
S.1
Simulator Technical Details
S.1.1
Client and Server Information Exchange
CARLA is designed as a client-server system. The server runs and renders the CARLA world. The
client provides an interface for users to interact with the simulator by controlling the agent vehicle
and certain properties of the simulation.
Commands. The agent vehicle is controlled by 5 types of commands sent via the client:
• Steering: The steering wheel angle is represented by a real number between -1 and 1, where
-1 and 1 correspond to full left and full right, respectively.
• Throttle: The pressure on the throttle pedal, represented as a real number between 0 and 1.
• Brake: The pressure on the brake pedal, represented as a real number between 0 and 1.
• Hand Brake: A boolean value indicating whether the hand brake is activated or not.
• Reverse Gear: A boolean value indicating whether the reverse gear is activated or not.
Meta-commands. The client is also able to control the environment and the behavior of the server
with the following meta commands:
• Number of Vehicles: An integer number of non-player vehicles to be spawned in the city.
• Number of Pedestrians: An integer number of pedestrians to be spawned in the city.
• Weather Id: An index of the weather/lighting presets to use. The following are currently
supported: Clear Midday, Clear Sunset, Cloudy Midday, Cloudy Sunset, Soft Rain Midday,
Soft Rain Sunset, Medium Rain Midday, Cloudy After Rain Midday, Cloudy After Rain
Sunset, Medium Rain Sunset, Hard Rain Midday, Hard Rain Sunset, After Rain Noon,
After Rain Sunset.
• Seed Vehicles/Pedestrians: A seed that controls how non-player vehicles and pedestrians
are spawned. It is possible to have the same vehicle/pedestrian behavior by setting the same
seed.
• Set of Cameras: A set of cameras with specific parameters such as position, orientation,
field of view, resolution and camera type. Available camera types include an optical RGB
camera, and pseudo-cameras that provide ground-truth depth and semantic segmentation.
Measurements and sensor readings. The client receives from the server the following information
about the world and the player’s state:
• Player Position: The 3D position of the player with respect to the world coordinate system.
• Player Speed: The player’s linear speed in kilometers per hour.
• Collision: Cumulative impact from collisions with three different types of objects: cars,
pedestrians, or static objects.
• Opposite Lane Intersection: The current fraction of the player car footprint that overlaps
the opposite lane.
• Sidewalk Intersection: The current fraction of the player car footprint that overlaps the
sidewalk.
• Time: The current in-game time.
• Player Acceleration: A 3D vector with the agent’s acceleration with respect to the world
coordinate system.
• Player Orientation: A unit-length vector corresponding to the agent car orientation.
• Sensor readings: The current readings from the set of camera sensors.
• Non-Client-Controlled agents information: The positions, orientations and bounding boxes
for all pedestrians and cars present in the environment.
• Traffic Lights information: The position and state of all traffic lights.
• Speed Limit Signs information: Position and readings from all speed limit signs.
11
S.1.2
Environment
CARLA provides two towns: Town 1 and Town 2. Figure S.1 shows maps of these towns and representative views. A large variety of assets were produced for CARLA, including cars and pedestrians.
Figure S.2 demonstrates this diversity.
S.2
Driving Systems Technical Details
In this section we provide additional technical details for the autonomous driving systems we have
tested.
S.2.1
Modular Pipeline
Perception module. Training of the semantic segmentation network was performed using
Adam [14] with learning rate λ0 = 10−6 for 300 epochs with batch size 2. The back-end ResNet is
pre-trained on ImageNet and frozen during training. No data augmentation is used.
The intersection classifier network is trained on 500 images balanced between the two classes. We
used Adam with learning rate λ0 = 10−3 for 500 epochs with batch size 5. No pre-training or data
augmentation are used.
Local planning for modular pipeline. In the road-following state, the local planner uses
the ego-lane mask computed from the semantic segmentation to select points that maintain a fixed
distance with the right edge of the road. The left-turn at intersections is more complex due
to the temporary absence of lane markings, the longer distance to the target lane, and the limited
field of view of the forward-facing camera. To deal with these challenges, we first compute waypoints towards the center of the intersection with a predefined skew angle; this helps improve the
visibility of the target lane. An auxiliary camera (along with its computed semantic segmentation
map) is used to determine the shape and alignment of the vehicle with respect of the target lane.
In the second step, the waypoints are laid out to yield a smooth trajectory from the center of the
intersection to the target lane. The right-turn state uses a similar strategy. However, as turning
right is easier given the target lane proximity, the number of waypoints needed is lower and only the
forward-facing information is required. The intersection-forward state is handled similarly
to road-following. The hazard-stop mode is activated when the dynamic obstacle map presents
a cumulative probability of obstacle above a predefined threshold. In this case, the system generates
a special waypoint to request an emergency break from the controller.
S.2.2
Imitation Learning
Architecture. Table S.1 details the configuration of the network used in the imitation learning
approach [4]. The network is composed of four modules: a perception module that is focused on
processing image inputs, a measurement module that processes the speed input, a joint input module
that merges the perception and measurement information, and a control module that produces motor
commands from the joint input representation. The control module consists of 4 identical branches:
command-conditional modules for predicting the steering angle, brake, and throttle – one for each
of the four commands. One of the four command-conditional modules is selected based on the input
command. The perception module is implemented by a convolutional network that takes a 200×88
image as input and outputs a 512-dimensional vector. All other modules are implemented by fullyconnected networks. The measurement module takes as input a measurement vector and outputs a
128-dimensional vector.
Training details. We trained all networks with Adam [14]. We used mini-batches of 120 samples.
We balanced the mini-batches, using the same number of samples for each command. Our starting
learning rate was 0.0002 and it was multiplied by 0.5 every 50,000 mini-batch iterations. We trained
for 294,000 iterations in total. Momentum parameters were set to β1 = 0.7 and β2 = 0.85. We used
no weight decay, but performed 50% dropout after hidden fully-connected layers and 20% dropout
on convolutional layers. To further reduce overfitting, we performed extensive data augmentation by
adding Gaussian blur, additive Gaussian noise, pixel dropout, additive and multiplicative brightness
variation, contrast variation, and saturation variation. Before feeding a raw 800 × 600 image to
12
Town 1 Map
Town 2 Map
Figure S.1: The two CARLA towns. Left: views and a map of CARLA Town 1. Right: views and
a map of CARLA Town 2.
13
Figure S.2: Diversity of cars and pedestrians currently available in CARLA.
14
module
Perception
Measurement
Joint input
Control
input dimension
channels
kernel
stride
200 × 88 × 3
98 × 48 × 32
96 × 46 × 32
47 × 22 × 64
45 × 20 × 64
22 × 9 × 128
20 × 7 × 128
9 × 3 × 256
7 · 1 · 256
512
1
128
128
512 + 128
512
256
256
32
32
64
64
128
128
256
256
512
512
128
128
128
512
256
256
1
5
3
3
3
3
3
3
3
−
−
−
−
−
−
−
−
−
2
1
2
1
2
1
2
1
−
−
−
−
−
−
−
−
−
Table S.1: Exact configurations of all network modules for the imitation learning approach.
the network, we cropped 171 pixels at the top and 45 at the bottom, and then resized the resulting
800 × 384 image to a resolution of 200×88.
Training data. The expert training data was collected from two sources: an automated agent and
human driver data. The automated agent has access to privileged information such as locations of
dynamic objects, ego-lane, states of traffic lights. 80% of the demonstrations were provided by the
automated agent and 20% by a human driver.
In order to improve the robustness of the learned policy, we injected noise into the expert’s steering
during training data collection. Namely, at random points in time we added a perturbation to the
steering angle provided by the driver. The perturbation is a triangular impulse: it increases linearly,
reaches a maximal value, and then linearly declines. This simulates gradual drift from the desired
trajectory, similar to what might happen with a poorly trained controller. The triangular impulse is
parametrized by its starting time t0 , duration τ ∈ R+ , sign σ ∈ {−1, +1}, and intensity γ ∈ R+ :
2(t − t0 )
sperturb (t) = σγ max 0, 1 −
.
(1)
−1
τ
Every second of driving we started a perturbation with probability pperturb . We used pperturb = 0.1
in our experiments. The sign of each perturbation was sampled at random, the duration was sampled
uniformly from 0.5 to 2 seconds, and intensity was fixed to 0.15.
S.2.3
Reinforcement Learning
We base our A3C agent on the network architecture proposed by Mnih et al. [18]. The input to
the network consists of two most recent images observed by the agent, resized to 84 × 84 pixels,
as well as a vector of measurements. The measurement vector includes the current speed of the
car, distance to goal, damage from collisions, and the current high-level command provided by
the topological planner, in one-hot encoding. The inputs are processed by two separate modules:
images by a convolutional module, measurements by a fully-connected network. The outputs of the
two modules are concatenated and further processed jointly. We trained A3C with 10 parallel actor
threads, for a total of 10 million environment steps. We used 20-step rollouts, following Jaderberg et
al. [13], with initial learning rate 0.0007 and entropy regularization 0.01. Learning rate was linearly
decreased to zero over the course of training.
The reward is a weighted sum of five terms: distance traveled towards the goal d in km, speed v in
km/h, collision damage c, intersection with the sidewalk s (between 0 and 1), and intersection with
the opposite lane o (between 0 and 1).
rt = 1000 (dt−1 − dt ) + 0.05 (vt − vt−1 ) − 0.00002 (ct − ct−1 ) − 2 (st − st−1 ) − 2 (ot − ot−1 ). (2)
15
S.3
Experimental Setup
S.3.1
Types of Infractions
We characterize the approaches by average distance travelled between infractions of the following
five types:
•
•
•
•
•
Opposite lane: More than 30% of the car’s footprint is over wrong-way lanes.
Sidewalk: More than 30% of the car’s footprint is over the sidewalk.
Collision with static object: Car makes contact with a static object, such as pole or building.
Collision with car: Car makes contact with another car.
Collision with pedestrian: Car makes contact with a pedestrian.
The duration of each violation is limited to 2 seconds, so if the car remains on the sidewalk for 10
seconds, this will be counted as 5 violations, not one.
16
| 1cs.CV
|
Noname manuscript No.
(will be inserted by the editor)
Investigating the Parameter Space of Evolutionary
Algorithms
Moshe Sipper · Weixuan Fu · Karuna
Ahuja · Jason H. Moore
arXiv:1706.04119v3 [cs.NE] 10 Oct 2017
Received: October 11, 2017/ Accepted: date
Abstract The practice of evolutionary algorithms involves the tuning of many
parameters. How big should the population be? How many generations should
the algorithm run? What is the (tournament selection) tournament size? What
probabilities should one assign to crossover and mutation? Through an extensive series of experiments over multiple evolutionary algorithm implementations and problems we show that parameter space tends to be rife with viable
parameters, at least for 25 the problems studied herein. We discuss the implications of this finding in practice.
Keywords Evolutionary Algorithms · Genetic Programming · Meta-Genetic
Algorithm · Parameter Tuning · Hyper-Parameter
1 Introduction
One of the crucial tasks of the evolutionary computation (EC) practitioner
is the tuning of parameters. The fitness-select-vary paradigm comes with a
plethora of parameters relating to the population, the generations, and the
operators of selection, crossover, and mutation. It seems natural to ask whether
the myriad parameters can be obtained through some clever methodology
(perhaps even an evolutionary one) rather than by trial and error; indeed,
as we shall see below, such methods have been devised. Our own interest
in the issue of parameters stems partly from a desire to better understand
evolutionary algorithms (EAs) and partly from our recent investigation into
the design and implementation of an accessible artificial intelligence system
[16].
Institute for Biomedical Informatics (IBI), Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104. E-mail: {sipper,weixuanf,karunaa,jhmoore}@upenn.edu ·
M. Sipper is also with the Department of Computer Science, Ben-Gurion University, BeerSheva 8410501, Israel.
2
Moshe Sipper et al.
In this paper we examine key parameters, asking whether we might improve
the parameter-seeking process. EC practitioners often employ commonly used
parameters “selected by conventions, ad hoc choices, and very limited experimental comparisons” [7]. We experimented with a large and variegated
assortment of problems in what is arguably one of the most extensive EC
experiments, concluding that parameter space, in fact, tends to be rife with
viable parameters, at least for the 25 problems studied herein.
We begin in the next section by delineating previous work on finding parameters through intelligent means. In Section 3 we turn to our own experiments by first describing the software and datasets we used. Section 4 presents
our first approach for obtaining “good” parameter sets, which was based on
a meta-genetic algorithm, followed by a random-search approach in Section 5.
We discuss our findings and conclude in Section 6.
2 Previous Work: Seeking Parameters
Before delving into a detailed discussion, let us make the distinction between
parameters, which are part of the model being evolved, and hyper-parameters
(also called meta-parameters), which are not part of the model and need to
be set by the user before running the evolutionary process, either manually or
algorithmically. Examples of parameters are synaptic weights in deep neural
networks and ephemeral random constants (ERCs) in genetic programming
(GP), while examples of hyper-parameters include the number of hidden layers
in deep neural networks and several standard parameters in GP: population
size, generation count, crossover rate, mutation probability, and selection type.
[4] noted that there are two major forms of setting
(hyper-)parameter values: parameter tuning and parameter control. Tuning
refers to the common practice of seeking good values for the parameters before running the algorithm, then running the algorithm using these values,
which remain fixed during the run. Parameter control means that values for
the parameters are changed during the run.
An early work by [19] looked into the problem of VLSI layout, aiming to
minimize the overall connection length between the components (cells) of a
circuit. They devised a genetic algorithm that used three operators: crossover
(order, cycle, or partially mapped—PMX), mutation (pairwise interchange
of genes), and inversion (taking a random segment in a solution string and
flipping it). They used a meta-GA to optimize crossover rate, inversion rate,
and mutation rate. The individuals in the meta-GA population consisted of
three integers in the range [0,20] (the overall search space was thus quite
small, comprising 8000 combinations). An individual’s fitness was defined as
the quality of the best layout found when a GA was run with the parameters it
embodied. The meta-GA was run on four test circuits, had a population size of
20, ran for 100 generations, and used uniform crossover (select parameter from
either parent at random) and mutation (add random noise; no inversion used).
They noted that crossover rate converged to values in the range 20-40%, with
Investigating the Parameter Space of Evolutionary Algorithms
3
little difference in best fitness as long as the rate was within this range. The
mutation rate evolved by the meta-GA was 0.5-1.5%, and the inversion rate
was 0-30%. The authors then adopted a crossover rate of 33%, an inversion rate
of 15%, and a mutation rate of 0.5%. These were used to run the optimizing
GA on circuits of interest (different than those used in the meta-GA phase)
and compare it with other techniques.
Another early work by [8] described meta-evolutionary programming. Their
flavor of evolutionary programming used mutation only to evolve solutions to
two functions in R2 . Mutation took a single parent and generated an offspring by adding Gaussian random noise with zero mean and variance equal
to F (x, y)—the function under investigation; this was applied to all members
of the population. The meta-algorithm attached a perturbation term to each
individual in the population, which was used as variance during mutation.
This term was then evolved along with the solution. They compared both
these algorithms (meta- and non-meta) with each other, and with a standard,
Holland-style genetic algorithm, concluding that both versions of evolutionary
programming outperformed the GA. They also concluded that the standard
evolutionary-programming method attained better results over the two functions studied, but that the meta-version was more robust to changes in the
function.
[26] applied a genetic algorithm to nonlinear constrained mixed discreteinteger optimization problems, using a meta-GA to optimize population size,
crossover probability, mutation probability, and crossover operator. The total
number of parameter combinations was 19,200. The fitness of an individual in
the meta-GA population was taken as the optimum found by a GA run with
the parameters defined by the individual. Their findings showed insensitivity
to crossover rate but high sensitivity to mutation rate. Four-point crossover
outperformed one-, two-, and three-point crossover.
[10] (see also [6]) noted that, “it is natural to expect adaptation to be used
not only for finding solutions to a problem, but also for tuning the algorithm
to the particular problem.” Their paper provided a short survey of adaptation techniques in evolutionary computation. They defined four categories of
adaptation: static—constant throughout run and tuned externally; and dynamic, refined further into deterministic—parameter altered by a deterministic rule, adaptive—some feedback from the evolutionary algorithm determines
the change in the parameter, and self-adaptive—the parameters to be adapted
are encoded into the chromosomes and undergo crossover and mutation. They
also defined four levels of adaptation: environment—such as changes to the
fitness function (e.g., weights), population—parameters that apply to the entire population are adapted, individual—parameters held within an individual
affecting only that individual, and component—parameters specific to a component or gene within an individual (such as self-adaptation of component-level
mutation steps sizes and rotation angles in evolution strategies).
[17] presented meta-Lamarckian learning in the context of memetic algorithms (MA), which incorporate local improvement procedures within traditional GAs. Their paper investigated the adaptive choice of local search (LS)
4
Moshe Sipper et al.
methods to ensure robustness in MA search. In addition to Darwinian evolution they also studied Lamarckian learning, where the genotype reflected the
result of improvement through placing the locally improved individual back
into the population to compete for reproductive opportunities. They studied
two adaptive meta-Lamarckian learning strategies, a heuristic approach based
on subproblem decomposition, and a stochastic approach based on a biased
roulette wheel. They tested their system on continuous parametric benchmark
test problems and on a wing-design problem. They concluded that, “the strategies presented are effective in producing search performances that are close to
the best traditional MA with a LS chosen to suit the problem in hand. Given
that such knowledge is often not available a priori, this ability to tackle new
problems in a robust way is of significant value.”
[18] proposed the utilization of logistic regression for tuning the parameters
of a “transgenetic” algorithm—an evolutionary algorithm that deals, basically,
with a population of chromosomes and a population of transgenetic vectors.
They cited symbiogenesis as their inspiration, a theory of evolution according
to which new cell organelles, new bodies, new organs, and new species arise
from symbiosis, wherein independent organisms merge to form composites.
The chromosomes of a transgenetic algorithm do not share genetic material directly. There are no crossover and mutation operations but rather transgenetic
vectors that obtain and insert information into the chromosomes. They used
logistic regression to set two main parameters of their algorithm (population
size and maximum length of transgenetic vector that met certain constraints),
their problem of interest being Traveling Salesman. They showed that their
algorithm outperformed a standard memetic algorithm.
[4], mentioned above, described an efficient technique for adapting control parameter settings associated with differential evolution (DE). DE uses a
floating-point encoding for global optimization over continuous spaces, creating new candidate solutions by combining the parent individual and several
other individuals of the same population. A candidate replaces the parent only
if it has better fitness. DE has three parameters: amplification factor of the
difference vector, crossover control parameter, and population size. In [4], the
parameter control technique was based on the self-adaptation of the first two
parameters, which were encoded within an individual’s genome. Their testbed
consisted of twenty-one benchmark functions from [27]. They concluded that
self-adaptive DE is better or comparable to the original DE and some other
evolutionary algorithms they examined.
[5]—in his chapter in the book Parameter Setting in Evolutionary Algorithms ([13])—provided a thirty-year perspective of parameter setting in evolutionary computation. He wrote that, “It is not surprising, then, that from
the beginning EA practitioners have wanted to know the answers to questions
like:
– Are there optimal settings for the parameters of an EA in general?
– Are there optimal settings for the parameters of an EA for a particular class of fitness landscapes?
Investigating the Parameter Space of Evolutionary Algorithms
5
– Are there robust settings for the parameters of an EA that produce
good performance over a broad range of fitness landscapes?
– Is it desirable to dynamically change parameter values during an
EA run?
– How do changes in a parameter affect the performance of an EA?
– How do landscape properties affect parameter value choices?”
He went on to review static parameter-setting strategies, where he mentioned a two-level EA, the top level of which evolved the parameters of the
lower-level EA. [5] stated that the, “key insight from such studies is the robustness of EAs with respect to their parameter settings. Getting ‘in the ball park’
is generally sufficient for good EA performance.” Our study herein not only
confirms this observation through numerous experiments, but also presents
the novel finding that the ballpark can be quite large. Of dynamic parametersetting strategies he opined that, “it is difficult to say anything definitive and
general about the performance improvements obtained through dynamic parameter setting strategies.” He added, interestingly, “My own view is that
there is not much to be gained in dynamically adapting EA parameter settings when solving static optimization problems. The real payoff for dynamic
parameter setting strategies is when the fitness landscapes are themselves dynamic ...” [5] also discussed the different aspects of setting various standard
parameters: parent population size, offspring population size, selection, reproductive operators; adapting the representation; and parameterless EAs.
[11] provided a survey of self-adaptive parameter control in evolutionary
computation, where—as noted above—control parameters are added to the
(evolving) genome. He complemented the taxonomy offered by [6], dividing
parameter setting into two main categories: tuning and control. Tuning was
further divided into tuning by hand, tuning by design of experiments, and
tuning by meta-evolution; while control was divided as previously into deterministic, adaptive, and self-adaptive. This paper mainly focused on function
optimization techniques, such as the covariance matrix self-adaptation evolution strategy (CMSA-ES). A mention of meta-evolution noted that they,
“have to be chosen problem-dependent, which is the obvious drawback of the
approach.” He concluded with the observation that most theoretical work on
self-adaptation concentrated on mutation, stating that, “A necessary condition
for the success of self-adaptation is a tight link between strategy parameters
and fitness.”
[7] (see also [20, 21]) presented a conceptual framework for parameter tuning based on a three-tier hierarchy of: a problem, an evolutionary algorithm
(EA), and a tuner. They argued that parameter tuning could be considered
from two different perspectives, that of configuring an evolutionary algorithm
by choosing parameter values that optimize its performance, and that of analyzing an evolutionary algorithm by studying how its performance depends
on its parameter values. Furthermore, they distinguished between analyzing
an evolutionary algorithm by studying how its performance depends on the
problems it is solving, and analyzing an evolutionary algorithm by studying
6
Moshe Sipper et al.
how its performance varies when executing independent repetitions of its run.
They noted the existence of two types of parameters, qualitative (e.g., crossover
type) and quantitative (e.g., crossover rate). They opined that, “using tuning
algorithms is highly rewarding. The efforts are moderate and the gains in performance can be very significant. Second, by using tuning algorithms one does
not only obtain superior parameter values, but also much information about
parameter values and algorithm performance. This information can be used
to obtain a deeper understanding of the algorithm in question.” The paper
discussed a wide range of tuning algorithms, which they classified as sampling
methods, model-based methods, screening methods, and meta-evolutionary algorithms. Of interest in their discussion of meta-evolutionary GAs was [14],
an early (possibly first) though limited work; and the description of multiobjective meta-GAs, which tuned for more than a single objective, e.g., speed
and accuracy. They opined that, “parameter tuning in EC has been a largely
ignored issue for a long time ... In the current EC practice parameter values are
mostly selected by conventions, ad hoc choices, and very limited experimental
comparisons.”[italics added] This latter observation—with which we wholly
concur—forms part of our motivation for the current study.
[1] carried out “the largest empirical analysis so far on parameter tuning
in search-based software engineering.” They performed experiments in the
domain of test generation for object-oriented software using genetic algorithms.
The objective was to derive sets of test cases (suites) for a given class, such
that the test suite maximized a chosen coverage criterion while minimizing the
number of tests and their length. A test case in this domain was a sequence of
method calls that constructed objects and called methods on them. Because
their goal was to study the effects of tuning, they analyzed all the possible
combinations of the selected parameter values. They concluded that, “tuning
can improve performance, but default values coming from the literature can
be already sufficient.”
[3] studied neural networks, showing that random experiments were more
efficient than grid experiments for hyper-parameter optimization in the case
of several learning algorithms on several datasets. They wrote that, “random
experiments are more efficient because not all hyperparameters are equally
important to tune ... Random experiments are also easier to carry out than
grid experiments for practical reasons related to the statistical independence
of every trial.” This paper partly motivated our choice of random search in
Section 5.
[22] is perhaps the most relevant paper to our current research, presenting a
meta-EA called REVAC (Relevance Estimation and Value Calibration), which
they used on a suite of 25 real-valued benchmark functions (real-parameter
optimization functions defined for the CEC 2005 Special Session on RealParameter Optimization, including five unimodal functions and twenty multimodal functions; [23]). They chose to improve G-CMA-ES, which they considered a hard-to-improve evolutionary algorithm, cycling through parent selectionrecombination-mutation-survivor selection-evaluation over a population of G-
Investigating the Parameter Space of Evolutionary Algorithms
7
CMA-ES parameter vectors. They were indeed successful in improving the
algorithm’s performance.
Our aim is to go further, casting our net much wider in terms of problem
domains, seeking to better understand parameter space.
3 Software and Datasets
We chose to work with two very different evolutionary-algorithm packages:
Distributed Evolutionary Algorithms in Python (DEAP) [9]—which uses treebased GP, and M4GP [12]—which is a stack-based evolutionary algorithm. We
ran our experiments on a cluster of 224 cores (Intel® Xeon® E5-2650L), with
2 threads per core.
DEAP, available at github.com/DEAP, comes with five sample problems:
1. Symbolic Regression, with data points generated from the quartic polynomial x4 + x3 + x2 + x.
2. Even-Parity: find the parity, even or odd, of n Boolean inputs (we set n to
8).
3. Multiplexer 3-8: reproduce the behavior of an electronic multiplexer with
3 address bits (inputs) and 8 data lines (outputs).
4. Artificial Ant: evolve simple controllers—“artificial ants”—that are able to
eat all the food located in a given two-dimensional, grid environment.
5. Spambase: return true if an email is spam, false otherwise.
We performed a preliminary investigation of these five problems, essentially
running DEAP with parameters tuned by hand (a common-enough undertaking in the EC community). We found that for the first four problems one can
attain an accuracy level of close to 1, while for Spambase our preliminary investigation set the attainable accuracy level at 0.93. More detailed descriptions
of each problem can be found on the DEAP website.
M4GP is entirely different, based on stack-based GP. This serves to reinforce our conclusions by running our experiments on two very different types
of GP algorithms. M4GP uses a nearest centroid distance metric to make
classifications, with each program producing multiple outputs. We ran M4GP
over problems from PMLB, a new publicly available dataset suite (accessibly
hosted on GitHub) initialized with 165 real-world, simulated, and toy benchmark datasets for evaluating supervised classification methods [15]. Note that
PMLB focuses on classification benchmarks, whereas of the DEAP sample
problems above only Spambase involves classification. Thus, our study includes
different types of problems.
The preliminary investigation in this case consisted of delving into the
data provided by the PMLB authors. “Once the datasets were scaled,” wrote
[15], “we performed a comprehensive grid search of each of the ML method’s
parameters using 10-fold cross-validation to find the best parameters (according to mean cross-validation balanced accuracy) for each ML method on each
dataset. This process resulted in a total of over 5.5 million evaluations of the
13 ML methods over the 165 data sets.”
8
Moshe Sipper et al.
Table 1 PMLB results by [15]. (a) “Easier” problems: 10 datasets for which a balanced
accuracy of 1 was attained most frequently. (b) “Harder” problems: 10 datasets whose
average balanced accuracy was in the range [0.9, 0.95]. Shown for each problem: number of
features, number of classes, and number of samples.
(a) “Easier” problems
problem
mofn-3-7-10
clean2
clean1
mushroom
irish
agaricus-lepiota
corral
xd6
mux6
threeOf9
features
10
168
168
22
5
22
6
9
6
9
classes
2
2
2
2
2
2
2
2
2
2
samples
1324
6598
476
8124
500
8145
160
973
128
512
(b) “Harder” problems
problem
breast-cancer-wisconsin
wdbc
tokyo1
new-thyroid
spambase
vote
soybean
house-votes-84
breast-w
molecular-biology promoters
features
30
30
44
5
57
16
35
16
9
58
classes
2
2
2
3
2
2
18
2
2
2
samples
569
569
959
215
4601
435
675
435
699
106
This available data saved us the need to run initial investigative experiments over PMLB. Each row of the 5.5-million table of results represents a
single run, of a single ML algorithm, using a specific set of parameters; a row
contains six columns: dataset name, classifier (machine learning algorithm),
parameters, accuracy, macro-averaged F1 score, balanced accuracy. Given that
one usually wants first and foremost to solve a problem we focused on accuracy, specifically, balanced accuracy (the rightmost value of each row), which
is a normalized version of accuracy that accounts for class imbalance by calculating accuracy on a per-class basis, then averaging the per-class accuracies
[24, 25].
We composed two suites of ten datasets each: 1) 10 datasets for which
a balanced accuracy of 1 was attained most frequently (Table 1(a)), and 2)
10 datasets whose average balanced accuracy was in the range [0.9, 0.95] (Table 1(b)).
Investigating the Parameter Space of Evolutionary Algorithms
9
Fig. 1 The meta-genetic algorithm. The population consisted of simple linear genomes that
contained five parameters. An individual’s fitness was obtained by launching an entire GP
evolutionary run with the parameters given in the genome. pop: population size, gen: generation count, xo: crossover probability, mu: mutation probability, tour: size of tournament
for tournament selection.
4 Searching for Parameters using a Meta-Genetic Algorithm
As done in a number of previous works discussed in Section 2, we ran a metalevel genetic algorithm over the space of GP parameters, of which we identified
five major ones:
1.
2.
3.
4.
5.
Population size (∈ N, [100, 3000]),
Number of generations (∈ N, [100, 2000]),
Crossover rate (∈ R, [0, 1]),
Mutation rate (∈ R, [0, 1]),
Tournament size (∈ N, [3, 100]).
Figure 1 shows a schematic of the meta-GA’s workings. The meta-GA
population comprised individuals with simple linear genomes that encoded
the above five parameters as either integers or real values, respectively. An individual’s fitness was obtained by launching entire GP evolutionary runs with
the parameters given in the genome. The GP in question was either DEAP or
M4GP. DEAP’s goal was to solve the five problems listed in Section 3 (regression, parity, mux, ant, spam), while M4GP was set lose on PMLB datasets.
Each GP was run n times, where n was the number of problems it was set to
solve (5 for DEAP, 10 for M4GP). Fitness was then a simple average of the n
highest fitness values obtained during the n GP runs.
The meta-GA had a population size ranging from 100-300, ran for 50-160
generations, and used tournament selection with tournament size 3. Variance
operators included: two-point crossover—with probability pxo = 0.5 perform
crossover on a pair of parents by selecting two (of four possible) parameter
boundaries and switching the parameters between them; and value mutation—
10
Moshe Sipper et al.
with probability pm = 0.2 perform mutation on a single parent by selecting
one of the five values at random, and mutating it by generating a random
integer (population size, generation count, size of tournament) or real value
(crossover probability, mutation probability) in the range specified above.
We experimented with the meta-GA for approximately two months, performing tens of thousands of evolutionary runs. We noted that numerous good
parameter sets kept emerging, quite often appearing at random generation
zero. Hence, we designed several putative improvements to the algorithmic
process. First, we tweaked parameters such as crossover and mutation, and
also added elitism (2%). It seemed that easier problems were causing the GP
to move into local minima. To correct for this effect we introduced a weighted
fitness function, where fitness was not computed as a simple average of the n
GP runs but rather a weighted one, with weights learned adaptively: examine
every m ∈ [1, 5] generations the average fitness per problem attained by all
GP runs over that particular problem; then increase weights of below-average
problems and decrease weights of above-average problems.
Other tweaks are not described here for brevity. Suffice it to say that after
numerous runs of the various versions of the meta-GA, we eventually concluded
that there seemed to be numerous successful parameter sets.
5 Searching for Parameters using Random Search
How rife is parameter space with good parameters, i.e., ones whose use by a
GP run results in success (which needs to be carefully defined)? Given the
findings by [3] and the high cost grid search would incur in our case, we opted
for random search.
We generated parameter sets at random, ran the following sets of experiments, and recorded the successful ones:
1. DEAP over the 5 sample problems (regression, parity, mux, ant, spam).
– Generate random parameter sets with parameters in the following ranges:
population size – [100, 1000]; generation count – [100, 1000]; crossover
rate – [0, 1]; mutation rate – [0, 1]; tournament size – [3, 30].
– Total runs (i.e., random parameter sets generated, with 5 GP runs
launched per each, attempting to solve all five problems): 2693; number
of successful parameter sets found: 110.
– Success criterion of a parameter set: accuracy of 0.97 attained for all 5
problems but Spambase, where the accuracy threshold was set to 0.93.
Figure 2 shows our results.
2. M4GP over the 10 problems of Table 1(a).
– Generate random parameter sets with parameters in the following ranges:
population size – [100, 2000]; generation count – [100, 2000]; crossover
rate – [0, 1]; mutation rate – [0, 1]; tournament size – [3, 30].
– Total runs: 2610; number of successful parameter sets found: 207.
– Success criterion of a parameter set: accuracy of 0.97 attained for all
10 problems.
Investigating the Parameter Space of Evolutionary Algorithms
11
(a)
(b)
(c)
(d)
(e)
Fig. 2 DEAP run over 5 problems. Shown are plots for the successful parameter sets found.
Figure 3 shows our results.
3. M4GP over the 10 problems of Table 1(b).
– Generate random parameter sets with parameters in the following ranges:
population size – [100, 2000]; generation count – [100, 2000]; crossover
rate – [0, 1]; mutation rate – [0, 1]; tournament size – [3, 30].
– Total runs: 5432; number of successful parameter sets found: 48.
– Success criterion of a parameter set: accuracy of 0.88 attained for all
10 problems.
Figure 4 shows our results.
The above experiments involved a total of 93,615 GP runs, each with a
population size and generation count that could both be as high as 1000 or
2000.
Perhaps evolution is not providing any added value and simple random
search would suffice? This is a standard question one should ask, hence we
performed the following:
12
Moshe Sipper et al.
(a)
(b)
(c)
(d)
(e)
Fig. 3 M4GP over the 10 problems of Table 1(a). Shown are plots for the successful parameter sets found.
1. DEAP: For each of the 110 “good” parameter sets found, generate pop size×
gen count × [1, 5] random solutions and check how many of them pass the
same 5-problem criterion employed above.
Result: none passed.
The total number of random solutions examined was 47,028,011.
2. M4GP: For each of the 48 “good” Table 1(b) parameter sets found, generate
pop size×gen count×[1, 10] random solutions and check how many of them
pass the same 10-problem criterion employed above.1
Result: none passed.
The total number of random solutions examined was 205,377,967.
Random solutions were generated by using the respective software package
to generate generation zero of size pop size × gen count. This not only saved
programming time but also prevented any bias vis-a-vis the GP experiments.
1 Note: The number of actual solutions generated per test point could be less than 5 or
10, respectively, if the criterion failed early on.
Investigating the Parameter Space of Evolutionary Algorithms
13
(a)
(b)
(c)
(d)
(e)
Fig. 4 M4GP over the 10 problems of Table 1(b). Shown are plots for the successful parameter sets found.
6 Concluding Remarks
We performed what is arguably one of the most extensive EC experiments.
Studying our results in Figures 2, 3, and 4, a clear trend emerges: There is very
little trend. We clearly see that good parameters range over the entire spectrum, somewhat in contrast with common lore. At least for 25 the problems
studied herein we note that:
– Population size need not be maximal.
– Generation count need not be maximal.
– At most, population size and generation count should not both be very
low. We remark that recent evolutionary findings suggest the use of fewer
generations [2].
– Tournament size need not be in the commonly used range of 3-7.
– Crossover rate can be high, low, or intermediate.
– Mutation rate can be high, low, or intermediate.
14
Moshe Sipper et al.
– Crossover-mutation pairs show no tendency to aggregate anywhere. Crossovermutation pairs should perhaps not both be low (which makes sense given
that an evolutionary algorithm requires inter-generational variation).
The yield (percentage) of good random parameter sets may seem relatively
small, but that is not the point, rather, our major observation concerns the
diversity in parameter space. Moreover, this yield will likely be far higher in
most situations encountered by evolutionary-algorithm practitioners, as we
were aiming at a broad spectrum of problems—quite a high and somewhat
unconventional bar—rather than at the more common case of a single problem.
To test this latter hypothesis we ran two additional sets of experiments:
1. DEAP: For each of the 5 problems, generate 100 random parameter sets
and perform a GP run for each, recording the best accuracy attained.
2. M4GP: For each of the 10 problems of Table 1(b), generate 100 random
parameter sets and perform a GP run for each, recording the best accuracy
attained.
Figure 5 shows the results of these 1500 runs.
Our findings suggest that perhaps one need not always spend too much time
and resources on tuning hyper-parameters and that random search is a good
choice for such tuning. Moreover, robustness to hyper-parameter tuning is a
desired quality of an evolutionary algorithm and if one’s algorithm requires
very specific parameters, the chances of finding them are slim; this would
essentially be a needle-in-a-haystack situation in hyper-parameter space.
Acknowledgments
We thank William La Cava, Randal Olson, and Ryan Urbanowicz for helpful
discussions. This research was supported by National Institutes of Health grant
no. AI116794.
References
1. Arcuri, A., Fraser, G.: Parameter tuning or default values? an empirical
investigation in search-based software engineering. Empirical Software
Engineering 18(3), 594–623 (2013)
2. Arnold,
C.:
Evolution
runs
faster
on
short
timescales.
Quanta
Magazine
(2017).
www.quantamagazine.org/
20170314-time-dependent-rate-phenomenon-evolution-viruses
3. Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization. Journal of Machine Learning Research 13(Feb), 281–305 (2012)
4. Brest, J., Greiner, S., Boskovic, B., Mernik, M., Zumer, V.: Self-adapting
control parameters in differential evolution: a comparative study on numerical benchmark problems. IEEE Transactions on Evolutionary Computation 10(6), 646–657 (2006)
Investigating the Parameter Space of Evolutionary Algorithms
15
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
(m)
(n)
(o)
Fig. 5 Results for single-problem test. For each of the above 15 problems, 100 random
parameter sets were generated, and a GP run executed for every one, recording the best
accuracy attained. (a)-(e): DEAP, (f)-(o):M4GP
5. De Jong, K.: Parameter setting in eas: a 30 year perspective. In: Parameter
setting in evolutionary algorithms, pp. 1–18. Springer (2007)
6. Eiben, Á.E., Hinterding, R., Michalewicz, Z.: Parameter control in evolutionary algorithms. IEEE Transactions on Evolutionary Computation
3(2), 124–141 (1999)
7. Eiben, A.E., Smit, S.K.: Parameter tuning for configuring and analyzing
evolutionary algorithms. Swarm and Evolutionary Computation 1(1), 19–
31 (2011)
16
Moshe Sipper et al.
8. Fogel, D.B., Fogel, L.J., Atmar, J.W.: Meta-evolutionary programming.
In: Signals, systems and computers, 1991. 1991 Conference record of the
twenty-fifth asilomar conference on, pp. 540–545. IEEE (1991)
9. Fortin, F.A., De Rainville, F.M., Gardner, M.A., Parizeau, M., Gagné, C.:
DEAP: Evolutionary algorithms made easy. Journal of Machine Learning
Research 13, 2171–2175 (2012)
10. Hinterding, R., Michalewicz, Z., Eiben, A.E.: Adaptation in evolutionary
computation: A survey. In: Evolutionary Computation, 1997., IEEE International Conference on, pp. 65–69. IEEE (1997)
11. Kramer, O.: Evolutionary self-adaptation: a survey of operators and strategy parameters. Evolutionary Intelligence 3(2), 51–65 (2010)
12. La Cava, William, Silva, Sara, Vanneschi, Leonardo, Spector, Lee, Moore,
Jason H.: Genetic programming representations for multi-dimensional feature learning in biomedical classification. In: EvoStar 2017. Spring (2017).
To Appear
13. Lobo, F., Lima, C.F., Michalewicz, Z.: Parameter setting in evolutionary
algorithms, vol. 54. Springer Science & Business Media (2007)
14. Mercer, R.E., Sampson, J.: Adaptive search using a reproductive metaplan. Kybernetes 7(3), 215–228 (1978)
15. Olson, R.S., La Cava, W., Orzechowski, P., Urbanowicz, R.J., Moore, J.H.:
PMLB: A large benchmark suite for machine learning evaluation and comparison. arXiv e-print. https://arxiv.org/abs/1703.00512 (2017)
16. Olson, R.S., Sipper, M., La Cava, W., Tartarone, S., Vitale, S., Fu, W.,
Holmes, J.H., Moore, J.H.: A system for accessible artificial intelligence.
In: Genetic Programming Theory & Practice XV. Springer (2017). URL
https://arxiv.org/abs/1705.00594. (submitted)
17. Ong, Y.S., Keane, A.J.: Meta-lamarckian learning in memetic algorithms.
IEEE Transactions on Evolutionary Computation 8(2), 99–110 (2004)
18. Ramos, I.C., Goldbarg, M.C., Goldbarg, E.G., Neto, A.D.D.: Logistic regression for parameter tuning on an evolutionary algorithm. In: 2005
IEEE Congress on Evolutionary Computation, vol. 2, pp. 1061–1068. IEEE
(2005)
19. Shahookar, K., Mazumder, P.: A genetic approach to standard cell placement using meta-genetic parameter optimization. IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems 9(5), 500–511
(1990)
20. Smit, S.K., Eiben, A.: Parameter tuning of evolutionary algorithms: Generalist vs. specialist. In: European Conference on the Applications of Evolutionary Computation, pp. 542–551. Springer (2010)
21. Smit, S.K., Eiben, A.E.: Comparing parameter tuning methods for evolutionary algorithms. In: 2009 IEEE congress on evolutionary computation,
pp. 399–406. IEEE (2009)
22. Smit, S.K., Eiben, A.E.: Beating the world champion evolutionary algorithm via revac tuning. In: IEEE Congress on Evolutionary Computation,
pp. 1–8. IEEE (2010)
Investigating the Parameter Space of Evolutionary Algorithms
17
23. Suganthan, P.N., Hansen, N., Liang, J.J., Deb, K., Chen, Y.P., Auger,
A., Tiwari, S.: Problem definitions and evaluation criteria for the cec 2005
special session on real-parameter optimization. KanGAL report 2005005,
2005 (2005)
24. Urbanowicz, R.J., Moore, J.H.: Exstracs 2.0: description and evaluation
of a scalable learning classifier system. Evolutionary Intelligence 8(2-3),
89–116 (2015)
25. Velez, D.R., et al.: A balanced accuracy function for epistasis modeling in
imbalanced datasets using multifactor dimensionality reduction. Genetic
Epidemiology 31(4), 306–315 (2007)
26. Wu, S.J., Chow, P.T.: Genetic algorithms for nonlinear mixed discreteinteger optimization problems via meta-genetic parameter optimization.
Engineering Optimization+ A35 24(2), 137–159 (1995)
27. Yao, X., Liu, Y., Lin, G.: Evolutionary programming made faster. IEEE
Transactions on Evolutionary computation 3(2), 82–102 (1999)
| 9cs.NE
|
arXiv:1207.4074v1 [math.PR] 17 Jul 2012
An analytical comparison of coalescent-based
multilocus methods: The three-taxon case
Sebastien Roch∗
April 9, 2018
Abstract
Incomplete lineage sorting (ILS) is a common source of gene tree incongruence in multilocus analyses. A large number of methods have been
developed to infer species trees in the presence of ILS. Here we provide a
mathematical analysis of several coalescent-based methods. Our analysis is
performed on a three-taxon species tree and assumes that the gene trees are
correctly reconstructed along with their branch lengths.
1
Introduction
Incomplete lineage sorting (ILS) is an important confounding factor in phylogenetic analyses based on multiple genes or loci [Mad97, DR09]. ILS is a populationlevel phenomenon that is caused by the failure of two lineages to coalesce in a
population, leading to the possibility that one of the lineages first coalesces with
a lineage from a less closely related population. As a result, it can produce extensive gene tree incongruence that must be accounted for appropriately in multilocus
analyses [DR06].
A large number of methods have been developed to address this source of incongruence [LYK+ 09]. Several such methods rely on a statistical model of ILS
known as the multispecies coalescent. In this model, populations are connected by
a phylogeny. Independent coalescent processes are performed in each population
∗
Department of Mathematics, University of Wisconsin–Madison.
1
and assembled to produce gene trees. Several methods have been shown to be statistically consistent under the multispecies coalescent, that is, they are guaranteed
to return the correct species tree given enough loci.
The performance and accuracy of coalescent-based multilocus methods have
been the subject of numerous simulation studies [LR11, LYPE09, YN10]. In
this paper, we complement such studies with a detailed analytical comparison
in a tractable test case, a three-taxon species tree. We analyze 7 methods: maximum likelihood (ML), GLASS/Maximum Tree (MT), R∗ , STAR, minimizing
deep coalescences (MDC), STEAC, and shallowest coalescences (SC). Under the
assumption that gene trees are reconstructed without estimation error, we derive
the exponential decay rate of the failure probability as the internal branch length
of the species tree varies. The analysis, which relies on large-deviations theory,
reveals that ML and GLASS/MT are more accurate in this setting than the other
methods—especially in the regime where ILS is more common.
2
2.1
Materials and Methods
Multispecies coalescent: Three-taxon case
We first describe the statistical model under which our analysis is performed, the
multispecies coalescent. We only discuss the three-taxon case. For more details,
see [DR09] and references therein.
A weighted rooted tree is called ultrametric if each leaf is exactly at the same
distance from the root. For a three-leaf ultrametric tree G with leaves a, b, and c,
we denote by ab|c the topology where a and b are closer to each other than to c,
and similarly for ac|b, bc|a. The topology of G is denoted by T [G].
Let S be an ultrametric species phylogeny with three taxa. We assume that
all haploid populations in S have population size N . We denote the current populations by A, B and C (which we identify with the leaves of S) and we assume
that S has topology AB|C. The ancestral populations are AB (corresponding
to the immediate ancestor to populations A and B) and ABC (corresponding to
the ancestor of populations A, B and C). The corresponding divergence times
(backwards in time from the present) are denoted by τAB and τABC with the assumption τAB ≤ τABC . All times are given in units of N generations. For a
population X, we let τXP be the divergence time of the parent population of X. Let
X = {A, B, C, AB, ABC} be the set of all populations in S.
We consider L loci ` = 1, . . . , L and, for each locus, we sample one lineage
2
(`)
from each population at time 0. For locus `, we denote by IX the number of
(`)
lineages entering population X and by OX the number of lineages exiting pop(`)
(`)
ulation X (backwards in time), where necessarily IX ≥ OX . Similarly, for
(`)
(`)
k = OX + 1, . . . , IX , the time of the coalescent event bringing the number of
(`,k)
lineages from k to k − 1 in population X and locus ` is TX . We denote by
G1 , . . . , GL the corresponding ultrametric gene trees (including both topology and
branch lengths).
Then, under the multispecies coalescent, assuming the loci are unlinked, the
likelihood of the gene trees is given by
f (G1 , . . . , GL |S) =
L
Y
exp
`=1
X O(`)
(`)
(`,OX +1)
P
X
−
τX − TX
2
X∈X
(`)
IX
X
−
(`)
k
(`,k+1)
(`,k)
TX
+ TX
(1)
2
k=OX +1
(`,I
(`)
+1)
where we let TX X
= τX for convenience [RY03].
The parameter governing the extent of incomplete lineage sorting is the length
of the internal branch of S
t = τABC − τAB .
The probability that the lineages from A and B fail to coalesce in branch AB, an
event we denote by FAIL` for locus ` (and its complement by S UCCESS` ), is
1 − p = e−t .
Note that, in that case, all three gene-tree topologies are equally likely. Of course,
1 − p → 1 as t → 0.
2.2
Multilocus methods
A basic goal of multilocus analyses is to reconstruct a species phylogeny (including possibly estimates of the divergence times) from a collection of gene trees.
Here we assume that the data consists of L gene trees G1 , . . . , GL corresponding
to L unlinked loci generated under the multispecies coalescent. We assume further that the gene trees are ultrametric and that their topology and branch lengths
are estimated without error.
3
We consider several common multilocus methods. In our setting, several of
these methods are in fact equivalent and we therefore group them below. Note
further that we only consider statistically consistent methods, that is, methods that
are guaranteed to converge on the right species phylogeny as the number of loci L
increases to +∞ (at least, in the test case we described above). We briefly describe
these methods. For more details, see e.g. [LYK+ 09] and references therein.
ML/GLASS/MT Under the multispecies coalescent, maximum likelihood (ML)
selects the topology and divergence times that maximizes the likelihood (1).
In the GLASS method [MR10], the species phylogeny is reconstructed from
a distance matrix in which the entries are the minimum gene coalescence times
across loci. The equivalent Maximum Tree (MT) method was introduced and
studied in [LP07, ELP07, LYP10].
A key result in [LYP10] is that, in the constant-population case, the term inside
the exponential in the likelihood (1) is monotonically decreasing in the divergence
times. As a result, because GLASS and MT select the phylogeny with the largest
possible divergence times, maximum likelihood is equivalent to GLASS and MT
in this context. See [LYP10] for details.
R∗ /STAR/MDC In the R∗ consensus method [Bry03, DDBR09], for each threetaxon set (here, we only have one such set), we include the topology that appears in
highest frequency among the loci and we reconstruct the most resolved phylogeny
that is compatible with these three-taxon topologies.
In the STAR method [LYPE09], the species phylogeny is reconstructed from
a distance matrix in which the entries are the average ranks of gene coalescence
times across loci. Here the root has the highest rank and the rank decreases by one
as one goes from the root to the leaves.
The minimizing deep coalescences (MDC) method [Mad97, TN09] selects the
species phylogeny that requires the smallest number of “extra lineages,” that is,
lineages that fail to coalesce in a branch of the species phylogeny.
On a three-taxon phylogeny, there is only three distinct rooted topologies. In
each case, the most recent divergence is assigned rank 1 in STAR and the other
divergence is assigned rank 2. Hence selecting the topology corresponding to the
lowest average rank is equivalent to selecting the most common topology among
all loci—which is what R∗ does. A similar argument shows that MDC also selects
the R∗ consensus tree in our test case.
4
STEAC/SC In the STEAC method [LYPE09], the species phylogeny is reconstructed from a distance matrix in which the entries are the average coalescence
times across loci. The shallowest coalescences (SC) method is similar to STEAC
in that it uses average coalescence times. The difference between the two methods
is in how they deal with multiple alleles per population. Since we only consider
the single-allele case, the two methods are equivalent here.
2.3
Large-deviations approach
As mentioned above, we consider estimation methods that are statistically consistent in the sense that they are guaranteed to converge on the correct species
phylogeny as the number of loci L increases to +∞. To compare different methods, we derive the rate of exponential decay of the probability of failure. Let S be
a species phylogeny with internal branch length t and assume that G1 , . . . , GL are
unlinked gene trees generated under the multispecies coalescent. As L → +∞,
large-deviations theory (see e.g. [Dur96]) gives a characterization of the (exponential) decay rate
αM (t) = − lim
L→+∞
1
ln P[Method M fails given L loci from S].
L
That is, roughly
P[Method M fails given L loci from S] ≈ e−LαM (t) ,
for large L. As the notation indicates, the key parameter that influences the decay
rate is the length of the internal branch t of the species phylogeny. In particular, we
expect that αM (t) is increasing in t as a larger t makes the reconstruction problem
easier.
To derive αM (t), we express the probability of failure as a large deviation event
of the form
" L
#
X
P[Method M fails given L loci from S] = P
Y` > yL ,
`=1
where y is a constant and {Y` }L`=1 are independent identically distributed random
variables. The particular choice of random variables depends on the method, as
we explain below. Let
φ(s) = E[esY` ],
5
be the moment-generating function of Y` (which does not depend on ` by assumption). Then the decay rate is given by
αM (t) = ys∗ − ln φ(s∗ ),
(2)
where s∗ > 0 is the solution (if it exists) to
φ0 (s∗ )
= y,
φ(s∗ )
provided there is an s > 0 such that φ(s) < +∞, y > E[Y` ] and Y` is not a point
mass at E[Y` ]. For more details on large-deviations theory, see e.g. [Dur96].
3
Results
3.1
A domination result
We first argue that, given perfectly reconstructed unlinked gene trees under the
multispecies coalescent, ML/GLASS/MT always has a greater probability of success than R∗ /STAR/MDC and STEAC/SC—or, in fact, any other method. Indeed
note that the probability of success can be divided into two cases:
1. The case where S UCCESS` occurs for at least one locus `, an event of probability (1 − (1 − p)L ). In that case, ML/GLASS/MT necessarily succeeds
whereas the other two methods succeed with probability < 1.
2. The case where FAIL` occurs for all loci `, an event of probability (1 − p)L .
In that case, all methods succeed with probability 1/3 by symmetry. For
instance, for ML/GLASS/MT, any pair of populations is equally likely to
lead to the smallest inter-species distance. A similar argument applies to
the other two methods.
Hence, overall ML/GLASS/MT succeeds with greater probability.
3.2
Decay rates
We derive the decay rates for the methods above. The results are plotted in Figure 1. The asymptotic regimes are highlighted in Figures 2 and 3. All proofs can
be found in the appendix.
6
ML/GLASS/MT In this case, the decay rate can be derived directly without
using (2). Following the derivation in [MR10] (see also [LYP10] for a similar
argument), ML/GLASS/MT succeeds with probability
1
(1 − (1 − p)L ) + (1 − p)L .
3
Then we get the following:
Claim 1 (ML/GLASS/MT) The decay rate of ML/GLASS/MT on S is
αML (t) = t.
(`)
R∗ /STAR/MDC For a locus `, we let ZAB be 1 if FAIL` occurs and T [G` ] =
AB|C, and 0 otherwise. We let
ZAB =
L
X
(`)
ZAB .
`=1
(`)
(`)
Similarly, we define ZAC , ZBC , ZAC and ZBC . Then R∗ /STAR/MDC fails if
ZAB + (L − ZAC − ZBC − ZAB ) < max{ZAC , ZBC }.
It can be shown that
1
ln P[2ZAC + ZBC > L].
L→+∞ L
αR∗ (t) = − lim
Then we get the following:
Claim 2 (R∗ /STAR/MDC) The decay rate of R∗ /STAR/MDC on S is
s
!
1 −t
2 −t
1 −t
αR∗ (t) = − ln 2
e
1− e
+ e
.
3
3
3
As t → 0,
3
αR∗ (t) = t2 + O(t3 ),
4
and, as t → +∞,
αR∗ (t) ≈
t 1 4
− ln .
2 2 3
7
(`)
STEAC/SC For a locus `, we let DAB be the time to the most recent common
ancestor of A and B in G` (in units of N generations). We let
DAB =
L
X
(`)
DAB .
`=1
(`)
(`)
Similarly, we define DAC , DBC , DAC and DBC . Then STEAC/SC fails if
DAB > min{DAC , DBC }.
It can be shown that
αSTEAC (t) =
lim −
L→+∞
1
ln P[DAB − DAC > 0].
L
Then we get the following:
Claim 3 (STEAC/SC) The decay rate of STEAC/SC on S is
−s∗ t
3e
− s2∗ e−t
αSTEAC (t) = − ln
,
3(1 − s2∗ )
where 0 < s∗ < 1 is the unique solution to the fixed-point equation
1
s∗ = [6s∗ − 3t(1 − s2∗ )]e(1−s∗ )t .
2
Further, as t → 0,
3
αSTEAC (t) = t2 + O(t3 ),
8
and, as t → +∞,
αSTEAC (t) ≈ t − ln t − 0.1656.
4
Discussion
As can be seen from Figures 1, 2 and 3 as well as from the asymptotics, ML/GLASS/MT does indeed give a larger decay rate for all t. In fact, the decay rate of
ML/GLASS/MT is significantly higher, especially as t → 0 that is, under high
levels of incomplete lineage sorting. For instance, to be concrete, if L = 500 loci
and t = 0.1 (in units of N generations), the probability of failure is approximately:
8
1.9 × 10−22 for ML/GLASS/MT; 0.038 for R∗ /STAR/MDC; 0.16 for STEAC/SC.
Intuitively, this difference in behavior arises from the fact that ML/GLASS/MT requires only one successful locus, whereas R∗ /STAR/MDC and STEAC/SC rely on
an average over all loci.
Comparing R∗ /STAR/MDC and STEAC/SC, note that αR∗ (t) is higher than
αSTEAC (t) for small t but that the situation is reversed for large t. In fact, in
the limit t → +∞, αSTEAC (t) grows at roughly the same rate as the optimal
αML (t). At large t, STEAC/SC has somewhat of an advantage in that the expectation gap in the failure event increases linearly with t, whereas it saturates under
R∗ /STAR/MDC.
The analysis described here ignores several features that influence the accuracy of species tree reconstruction. Notably we have assumed that gene trees,
including their branch lengths, are reconstructed without error. On real sequence
datasets, the uncertainty arising from gene-tree estimation plays an important role.
For instance, although GLASS/MT achieves the optimal decay rate in our setting,
these methods are in fact sensitive to sequence noise because they rely on the
computation of a minimum over loci—the very feature that leads to their superior
performance here. Extending our analysis to incorporate gene tree estimation error is an important open problem which should help in the design of multilocus
methods. It is important to note that, under appropriate modeling of sequence data,
ML is not in general equivalent to GLASS/MT and is likely to be more robust to
estimation error. In particular our analysis suggest that ML may be significantly
more accurate than other methods in multilocus studies.
Other extensions deserve further study. Often many alleles are sampled from
each population. Note that the benefit of multiple alleles is known to saturate
as the number of alleles increases [Ros02]. This is because the probability of
observing any number of alleles at the top of a branch is uniformly bounded in the
number alleles existing at the bottom.
Further, the molecular clock assumption, although it may be a reasonable first
approximation in the context of recently diverged populations, should not be necessary for our analysis. One should also consider larger numbers of taxa, varying
population sizes, etc.
Simulation studies may provide further insight into these issues. However an
analytical approach, such as the one we have used here, is valuable in that it allows
the study of an entire class of models in one analysis. It can also provide useful,
explicit predictions to guide the design of reconstruction procedures.
9
5
Acknowledgments
This work was supported by NSF grant DMS-1007144 and an Alfred P. Sloan Research Fellowship. Part of this work was performed while the author was visiting
the Institute for Pure and Applied Mathematics (IPAM) at UCLA.
10
3
ML
2.5
R*
STEAC
α
2
1.5
1
0.5
0
0
0.5
1
1.5
t
Figure 1: Decay rates.
11
2
2.5
3
−3
8
x 10
7
R*
STEAC
6
α
5
4
3
2
1
0
0
0.02
0.04
t
0.06
0.08
0.1
Figure 2: Decay rates as t → 0. The dotted lines indicate the respective predicted
asymptotics.
12
100
90
80
ML
R*
STEAC
70
α
60
50
40
30
20
10
0
0
20
40
t
60
Figure 3: Decay rates as t → +∞.
13
80
100
References
[Bry03]
David Bryant. A classification of consensus methods for phylogenetics. In Bioconsensus (Piscataway, NJ, 2000/2001), volume 61 of
DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages 163–183.
Amer. Math. Soc., Providence, RI, 2003.
[DDBR09] James H. Degnan, Michael DeGiorgio, David Bryant, and Noah A.
Rosenberg. Properties of consensus methods for inferring species
trees from gene trees. Systematic Biology, 58(1):35–54, 2009.
[DR06]
J. H. Degnan and N. A. Rosenberg. Discordance of species trees with
their most likely gene trees. PLoS Genetics, 2(5), May 2006.
[DR09]
James H. Degnan and Noah A. Rosenberg. Gene tree discordance,
phylogenetic inference and the multispecies coalescent. Trends in
ecology and evolution, 24(6):332–340, 2009.
[Dur96]
Richard Durrett. Probability: theory and examples. Duxbury Press,
Belmont, CA, second edition, 1996.
[ELP07]
Scott V. Edwards, Liang Liu, and Dennis K. Pearl. High-resolution
species trees without concatenation. Proceedings of the National
Academy of Sciences, 104(14):5936–5941, 2007.
[LP07]
Liang Liu and Dennis K. Pearl. Species trees from gene trees: Reconstructing bayesian posterior distributions of a species phylogeny using estimated gene tree distributions. Systematic Biology, 56(3):504–
514, 2007.
[LR11]
Adam D. Leaché and Bruce Rannala. The accuracy of species tree
estimation under simulation: A comparison of methods. Systematic
Biology, 60(2):126–137, 2011.
[LYK+ 09] Liang Liu, Lili Yu, Laura Kubatko, Dennis K. Pearl, and Scott V. Edwards. Coalescent methods for estimating phylogenetic trees. Molecular Phylogenetics and Evolution, 53(1):320 – 328, 2009.
[LYP10]
Liang Liu, Lili Yu, and Dennis Pearl. Maximum tree: a consistent estimator of the species tree. Journal of Mathematical Biology, 60:95–
106, 2010. 10.1007/s00285-009-0260-0.
14
[LYPE09] Liang Liu, Lili Yu, Dennis K. Pearl, and Scott V. Edwards. Estimating species phylogenies using coalescence times among sequences.
Systematic Biology, 58(5):468–477, 2009.
[Mad97]
Wayne P. Maddison. Gene trees in species trees. Systematic Biology,
46(3):523–536, 1997.
[MR10]
Elchanan Mossel and Sébastien Roch. Incomplete lineage sorting:
Consistent phylogeny estimation from multiple loci. IEEE/ACM
Trans. Comput. Biology Bioinform., 7(1):166–171, 2010.
[Roc12]
Sebastien Roch. An analytical comparison of coalescent-based multilocus methods: The three-taxon case. Preprint, 2012.
[Ros02]
N. A. Rosenberg. The probability of topological concordance of gene
trees and species trees. Theor. Popul. Biol., 61(2):225–247, March
2002.
[RY03]
Bruce Rannala and Ziheng Yang. Bayes estimation of species divergence times and ancestral population sizes using dna sequences from
multiple loci. Genetics, 164(4):1645–1656, 2003.
[TN09]
Cuong Than and Luay Nakhleh. Species tree inference by minimizing
deep coalescences. PLoS Comput Biol, 5(9):e1000501, 09 2009.
[YN10]
Yun Yu and Luay Nakhleh. The performance of methods for inferring
species trees from multi-locus data. Preprint, 2010.
15
A
A.1
Proofs
ML/GLASS/MT
Decay rate Following the derivation in [MR10] (see also [LYP10] for a similar
argument), ML/GLASS/MT succeeds with probability
1
(1 − (1 − p)L ) + (1 − p)L ,
3
where the two terms correspond to the two cases described in Section 3.1. Hence
the decay rate of the failure probability is
1
1
L
L
αML (t) = lim − ln 1 − (1 − (1 − p) ) + (1 − p)
L→+∞
L
3
2
1
(1 − p)L
= lim − ln
L→+∞
L
3
1 2
= lim − ln − ln(1 − p)
L→+∞
L 3
= − ln(1 − p)
= t.
A.2
R∗ /STAR/MDC
(`)
Definitions For a locus `, we let ZAB be 1 if FAIL` occurs and T [G` ] = AB|C,
and 0 otherwise. We let
L
X
(`)
ZAB =
ZAB .
`=1
(`)
(`)
Similarly, we define ZAC , ZBC , ZAC and ZBC . Then R∗ /STAR/MDC fails if
ZAB + (L − ZAC − ZBC − ZAB ) < max{ZAC , ZBC },
an event we denote by E. The second term on the LHS comes form the fact that,
given S UCCESS` , T [G` ] = AB|C. To deal with the term on the RHS, we re-write
E as
2 max{ZAC , ZBC } + min{ZAC , ZBC } > L,
and we use the auxiliary events
E 0 = {2ZAC + ZBC > L},
16
and
E 00 = {2ZBC + ZAC > L},
to bound P[E] as follows
P[E 0 ] ≤ P[E] ≤ P[E 0 ∪ E 00 ] ≤ 2P[E 0 ],
where we used that P[E 0 ] = P[E 00 ] (by symmetry) in a union bound, and the fact
that, on E 0 ,
2 max{ZAC , ZBC } + min{ZAC , ZBC } ≥ 2ZAC + ZBC > L.
Hence
−
1
1
1
1
1
ln P[E 0 ] ≥ − ln P[E] ≥ − ln 2P[E 0 ] = − ln P[E 0 ] − ln 2,
L
L
L
L
L
and, taking a limit as L → +∞,
αR∗ (t) = lim −
L→+∞
1
1
ln P[E] = lim − ln P[E 0 ],
L→+∞
L
L
provided the limit exists.
Moment-generating function
moment-generating function
In order to compute the limit above, we use the
(`)
(`)
φ(s) = E[exp(s[2ZAC + ZBC ])],
(which does not depend on `) as described in Section 2.3. Dividing up the expectation into the four possible cases, we have
1
1
φ(s) = p + (1 − p) + (1 − p)(es + e2s ) < +∞,
3
3
for all s ∈ R. Letting
1
Wp = (1 − p),
3
the derivative of φ(s) is
φ0 (s) = Wp (es + 2e2s ).
17
Decay rate
By large-deviations theory, we are looking for a solution to
1=
φ0 (s)
.
φ(s)
Letting ω = es , we get the quadratic equation
(p + Wp ) + Wp (ω + ω 2 ) = Wp (ω + 2ω 2 ),
or, rearranging,
(p + Wp ) = Wp ω 2 ,
whose solution is
s
ω∗ = es∗ =
p + Wp
.
Wp
Then
1
ln P[E 0 ]
L→+∞
L
1
= lim − ln P[2ZAC + ZBC > L]
L→+∞
L
= s∗ − ln φ(s∗ ).
αR∗ (t) =
lim −
Noting that
s
φ(s∗ ) = p + Wp + Wp
= 2(p + Wp ) +
p + Wp
p + Wp
+ Wp
Wp
Wp
q
(p + Wp )Wp ,
we get
p
s∗ − ln φ(s∗ ) = ln
p + Wp
p
p
2 Wp (p + Wp ) + Wp p + Wp
Rearranging, we have finally
q
αR∗ (t) = − ln 2 Wp (p + Wp ) + Wp
s
!
1 −t
2 −t
1 −t
e
1− e
+ e
.
= − ln 2
3
3
3
18
!
,
Asymptotics By a Taylor expansion, we get as t → 0 that
3
αR∗ (t) = t2 + O(t3 ).
4
On the other hand, as t → +∞,
#!
" s
2
1
1
1 − e−t + e−t/2
αR∗ (t) = − ln e−t/2 2
3
3
3
=
t
− βt
2
where
lim βt =
t→+∞
A.3
1 4
ln .
2 3
STEAC/SC
(`)
Definitions For a locus `, we let DAB be the time to the most recent common
ancestor of A and B in G` (in units of N generations). We let
DAB =
L
X
(`)
DAB .
`=1
(`)
(`)
Similarly, we define DAC , DBC , DAC and DBC . Then STEAC/SC fails if
DAB > min{DAC , DBC },
an event we denote by E. Once again, to deal with the term on the RHS, we
re-write E as
DAB − min{DAC , DBC } > 0,
and we use the auxiliary events
E 0 = {DAB − DAC > 0},
and
E 00 = {DAB − DBC > 0},
to bound P[E] as follows
P[E 0 ] ≤ P[E] ≤ P[E 0 ∪ E 00 ] ≤ 2P[E 0 ],
19
where we used that P[E 0 ] = P[E 00 ] (by symmetry) in a union bound, and the fact
that E 0 implies E. Hence
1
1
1
1
1
− ln P[E 0 ] ≥ − ln P[E] ≥ − ln 2P[E 0 ] = − ln P[E 0 ] − ln 2,
L
L
L
L
L
and, taking a limit as L → +∞,
1
1
αSTEAC (t) = lim − ln P[E] = lim − ln P[E 0 ],
L→+∞
L→+∞
L
L
provided the limit exists.
Moment-generating function In order to compute the limit above, we need the
moment-generating function
(`)
(`)
φ(s) = E[exp(s[DAB − DAC ])],
(which does not depend on `). Dividing up the expectation into the four possible
cases, we have
h
i
φ(s) = pe−st E esẼ0 E e−sE0
1
+ (1 − p)E[e−sE1 ]
3
1
+ (1 − p)E[esE1 ]
3
1
+ (1 − p)
3
where we used:
(`)
1. In the case S UCCESS` , DAB − τAB = Ẽ0 where Ẽ0 is an exponential mean 1
conditioned to be less than t. Independently, using the memoryless property
(`)
of the exponential, DAC − τABC = E0 where E0 is an exponential mean 1.
Hence
(`)
(`)
DAB − DAC = τAB + Ẽ0 − τABC − E0 = −t + Ẽ0 − E0 .
(`)
2. In the case FAIL
` and T [G` ] = AB|C, DAB − τABC = Ẽ1 where Ẽ1 the
3
minimum of 2 independent exponentials mean 1, that is, an exponential
(`)
mean 1/ 32 = 1/3. Moreover, DAC − τABC = Ẽ1 + E1 where E1 is an
exponential mean 1 independent of Ẽ1 . Hence
(`)
(`)
DAB − DAC = τABC + Ẽ1 − τABC − Ẽ1 − E1 = −E1 .
20
(`)
3. In the case FAIL` and T [G` ] = AC|B, DAC − τABC = Ẽ1 where Ẽ1 is an
(`)
exponential mean 1/ 32 = 1/3. Moreover, DAB − τABC = Ẽ1 + E1 where
E1 is an exponential mean 1 independent of Ẽ1 .
(`)
(`)
DAB − DAC = τABC + Ẽ1 + E1 − τABC − Ẽ1 = E1 .
(`)
(`)
4. In the case FAIL` and T [G` ] = BC|A, DAC = DAB .
Note that
E[esE0 ] = E[esE1 ] =
1
,
1−s
for all |s| < 1, and
sẼ0
E[e
1
]=
p
Z
t
esx e−x dx =
0
1 − e−(1−s)t
.
p(1 − s)
Hence
− e−(1−s)t 1
φ(s) = pe
p(1 − s) 1 + s
1
1
1
+
+1
+ (1 − p)
3
1+s 1−s
e−st − e−t 1 −t 3 − s2
=
+ e
1 − s2
3
1 − s2
3e−st − s2 e−t
.
=
3(1 − s2 )
−st 1
The derivative of φ(s) is
[−3te−st − 2se−t ][3(1 − s2 )] − [3e−st − s2 e−t ][−6s]
[3(1 − s2 )]2
[18s − 9t(1 − s2 )]e−st − 6se−t
=
[3(1 − s2 )]2
φ0 (s) =
Decay rate
By large-deviations theory, we are looking for a solution to
0=
φ0 (s)
[6s − 3t(1 − s2 )]e−st − 2se−t
=
.
φ(s)
(1 − s2 )(3e−st − s2 e−t )
21
(3)
Note that the denominator on the RHS is positive on s ∈ (0, 1), and that
φ0 (0)
= −t
φ(0)
and
lim−
s→1
φ0 (s)
= +∞,
φ(s)
so that by [Dur96] there is a solution 0 < s∗ < 1 to (3). The solution s∗ must
satisfy
[6s∗ − 3t(1 − s2∗ )]e−s∗ t − 2s∗ e−t = 0.
(4)
which can be re-written as the fixed-point equation
1
s∗ = [6s∗ − 3t(1 − s2∗ )]e(1−s∗ )t ≡ Ft (s∗ ),
2
0 < s∗ < 1.
(5)
Note that Ft (0) = −3tet ≤ 0 and Ft (1) = 3 > 1. Moreover,
1
t
[6 + 6ts]e(1−s)t − [6s − 3t(1 − s2 )]e(1−s)t
2
2
1 (1−s)t
e
[6 + 3t2 (1 − s2 )] > 1,
=
2
Ft0 (s) =
for 0 < s < 1. Hence Ft (s) − s is strictly increasing and has a unique solution in
(0, 1). Eq. (5) is easily solved numerically.
Then
1
ln P[E 0 ]
L→+∞
L
1
= lim − ln P[DAB − DAC > 0]
L→+∞
L
= − ln φ(s∗ ).
αSTEAC (t) =
lim −
Asymptotics We consider asymptotics when t → 0. Define
3
sε = t + ε−1 t2 .
4
22
Evaluating the LHS in (4) (for ε > 0 small but fixed) as t → 0 gives
[6sε − 3t(1 − s2ε )]e−sε t − 2sε e−t
9
9 2
3 2
−1 2
3
3
= t + 6ε t − 3t 1 − t + O(t )
1 − t + O(t )
2
16
4
2
t
3
− t + 2ε−1 t2 1 − t + + O(t3 )
2
2
3 2
3
−1 2
3
3
= t + 6ε t + O(t ) 1 − t + O(t )
2
4
2
3
t
−1 2
3
− t + 2ε t
1 − t + + O(t )
2
2
3 2
= 4ε−1 +
t + O(t3 ),
2
so that, because
4(−ε)−1 +
3
3
< 0 and 4ε−1 + > 0,
2
2
the solution of (4) satisfies s−ε < s∗ < sε for 0 < ε <
Then
8
3
and t small enough.
3e−sε t − s2ε e−t
3(1 − s2 )
ε
9 2
1
3 2
t2
3
3
3
=
3 1 − t + O(t ) −
t + O(t )
1 − t + + O(t )
3
4
16
2
9
× 1 + t2 + O(t3 )
16
3
3
9
2
= 1+t − −
+
+ O(t3 ).
4 16 16
φ(sε ) =
Since this holds for all ε > 0 small we get
3
αSTEAC (t) = − ln φ(s∗ ) = t2 + O(t3 ).
8
For the t → +∞ asymptotics, let
u=
1
t
and σ = (1 − s)t.
23
Substituting in (5), we get
1
1 − σu = [6(1 − σu) − 3σ(2 − σu)]eσ ,
2
which after rearranging becomes
3eσ − 1 − 3σeσ
3σeσ − σ − 32 σ 2 eσ
1
=
3σeσ
σ(1 + 2(3eσ −1−3σe
σ) )
u =
(6)
≡ F(σ).
We have F(0) = +∞. Moreover, letting σ∗ be the only positive solution to
G(σ∗ ) ≡ 3eσ∗ − 1 − 3σ∗ eσ∗ = 0,
(7)
we have F(σ∗ ) = 0. Note that G 0 (σ) = −3σeσ < 0, G(0) = 2 and limσ→+∞ G(σ) =
−∞, so that σ∗ is well-defined. Noticing that G appears in the denominator of (6)
as well we get that F is strictly decreasing between σ = 0 and σ = σ∗ . Hence the
limit t → +∞ is equivalent to the limit σ → σ∗− . Finally, in that limit, letting σt
be the σ-value giving rise to the value u = 1/t
αSTEAC (t) = − ln φ(s∗ )
3e−s∗ t − s2∗ e−t
= − ln
3(1 − s2∗ )
3eσt − (1 − σtt )2
= t − ln
3(1 − (1 − σtt )2 )
= t − ln
σt2
2σt
−
t
t2
σt2
2σt
3( t − t2 )
3eσt − 1 +
= t − ln t − βt ,
where
3eσ∗ − 1
3σ∗ eσ∗
lim βt = ln
= ln
= σ∗ − ln 2,
t→+∞
6σ∗
6σ∗
where we used (7).
24
| 5cs.CE
|
Backdoors to Tractable Valued CSP
Robert Ganian, M. S. Ramanujan, and Stefan Szeider
arXiv:1612.05733v1 [cs.DS] 17 Dec 2016
Algorithms and Complexity Group, TU Wien, Vienna, Austria
Abstract. We extend the notion of a strong backdoor from the CSP setting to
the Valued CSP setting (VCSP, for short). This provides a means for augmenting
a class of tractable VCSP instances to instances that are outside the class but of
small distance to the class, where the distance is measured in terms of the size of
a smallest backdoor. We establish that VCSP is fixed-parameter tractable when
parameterized by the size of a smallest backdoor into every tractable class of
VCSP instances characterized by a (possibly infinite) tractable valued constraint
language of finite arity and finite domain. We further extend this fixed-parameter
tractability result to so-called “scattered classes” of VCSP instances where each
connected component may belong to a different tractable class.
1 Introduction
Valued CSP (or VCSP for short) is a powerful framework that entails among others the
problems CSP and MAX-CSP as special cases [26]. A VCSP instance consists of a finite set of cost functions over a finite set of variables which range over a domain D, and
the task is to find an instantiation of these variables that minimizes the sum of the cost
functions. The VCSP framework is robust and has been studied in different contexts in
computer science. In its full generality, VCSP considers cost functions that can take as
values the rational numbers and positive infinity. CSP (feasibility) and Max-CSP (optimisation) arise as special cases by limiting the values of cost functions to {0, ∞} and
{0, 1}, respectively. Clearly VCSP is in general intractable. Over the last decades much
research has been devoted into the identification of tractable VCSP subproblems. An
important line of this research (see, e.g., [17,18,25]) is the characterization of tractable
VCSPs in terms of restrictions on the underlying valued constraint language Γ , i.e.,
a set Γ of cost functions that guarantees polynomial-time solvability of all VCSP instances that use only cost functions from Γ . The VCSP restricted to instances with cost
functions from Γ is denoted by VCSP[Γ ].
In this paper we provide algorithmic results which allow us to gradually augment
a tractable VCSP based on the notion of a (strong) backdoor into a tractable class of
instances, called the base class. Backdoors where introduced by Williams et al. [27,28]
for SAT and CSP and generalize in a natural way to VCSP. Let C denote a tractable
class of VCSP instances over a finite domain D. A backdoor of a VCSP instance P
into C is a (small) subset B of the variables of P such that for all partial assignments
α that instantiate B, the restricted instance P|α belongs to the tractable class C. Once
The authors acknowledge support by the Austrian Science Fund (FWF, project P26696).
Robert Ganian is also affiliated with FI MU, Brno, Czech Republic.
we know such a backdoor B of size k we can solve P by solving at most |D|k tractable
instances. In other words, VCSP is then fixed parameter tractable parameterized by
backdoor size. This is highly desirable as it allows us to scale the tractability for C to
instances outside the class, paying for an increased “distance” from C only by a larger
constant factor.
In order to apply this backdoor approach to solving a VCSP instance, we first need
to find a small backdoor. This turns out to be an algorithmically challenging task. The
fixed-parameter tractability of backdoor detection has been subject of intensive research
in the context of SAT (see, e.g., [16]) and CSP (see, e.g., [2]). In this paper we extend
this line of research to VCSP.
First we obtain some basic and fundamental results on backdoor detection when
the base class is defined by a valued constraint language Γ . We obtain fixed-parameter
tractability for the detection of backdoors into VCSP[Γ ] where Γ is a valued constraint
language with cost functions of bounded arity. In fact, we show the stronger result: fixedparameter tractability also holds with respect to heterogeneous base classes of the form
VCSP[Γ1 ] ∪ · · · ∪ VCSP[Γℓ ] where different assignments to the backdoor variables
may result in instances that belong to different base classes VCSP[Γi ]. A similar result
holds for CSP, but the VCSP setting is slightly more complicated as a valued constraint
language of finite arity over a finite domain is not necessarily finite.
Secondly, we extend the basic fixed-parameter tractability result to so-called scattered base classes of the form VCSP[Γ1 ] ⊕ · · · ⊕ VCSP[Γℓ ] which contain VCSP
instances where each connected component belongs to a tractable class VCSP[Γi ] for
some 1 ≤ i ≤ ℓ—again in the heterogeneous sense that for different assignments to the
backdoor variables a single component of the reduced instance may belong to different classes VCSP[Γi ]. Backdoors into a scattered base class can be much smaller than
backdoors into each single class it is composed of, hence the gain is huge if we can handle scattered classes. This boost in scalability does not come for free. Indeed, already
the “crisp” case of CSP, which was the topic of a recent SODA paper [14], requires a
sophisticated algorithm which makes use of advanced techniques from parameterized
algorithm design. This algorithm works under the requirement that the constraint languages contain all unary constraints (i.e., is conservative); this is a reasonable requirement as one needs these unary cost functions to express partial assignments (see also
Section 2 for further discussion). Here we lift the crisp case to general VCSP, and this
also represents our main technical contribution.
To achieve this, we proceed in two phases. First we transform the backdoor detection problem from a general scattered class VCSP(Γ1 )⊕· · ·⊕ VCSP(Γℓ ) to a scattered
class VCSP(Γ1′ ) ⊕ · · · ⊕ VCSP(Γℓ′ ) over finite valued constraint languages Γi′ . In the
subsequent second phase we transform the problem to a backdoor detection problem
into a scattered class VCSP(Γ1′′ ) ⊕ · · · ⊕ VCSP(Γℓ′′ ) where each Γi′′ is a finite crisp
language; i.e., we reduce from the VCSP setting to the CSP setting. We believe that
this sheds light on an interesting link between backdoors in the VCSP and CSP settings.
The latter problem can now be solved using the known algorithm [14].
Related Work
Williams et al. [27,28] introduced backdoors for CSP or SAT as a theoretical tool to capture the overall combinatorics of instances. The purpose was an analysis of the empirical
behaviour of backtrack search algorithms. Nishimura et. al [22] started the investigation
on the parameterized complexity of finding a small SAT backdoor and using it to solve
the instance. This lead to a number of follow-up work (see [16]). Parameterized complexity provides here an appealing framework, as given a CSP instance with n variables,
one can trivially find a backdoor of size ≤ k into a fixed tractable class of instances by
trying all subsets of the variable set containing ≤ k variables; but there are Θ(nk ) such
sets, and therefore the running time of this brute-force algorithm scales very poorly in k.
Fixed-parameter tractability removes k from the exponent providing running times of
the form f (k)nc which yields a significantly better scalability in backdoor size.
Extensions to the basic notion of a backdoor have been proposed, including backdoors with empty clause detection [6], backdoors in the context of learning [7], heterogeneous backdoors where different instantiations of the backdoor variables may
result in instances that belong to different base classes [15], and backdoors into scattered classes where each connected component of an instance may belong to a different
tractable class [14]. Le Bras et al. [20] used backdoors to speed-up the solution of hard
problems in materials discovery, using a crowd sourcing approach to find small backdoors.
The research on the parameterized complexity of backdoor detection was also successfully extended to other problem areas including disjunctive answer set programming [11,10], abstract argumentation [9], and integer linear programming [13]. There
are also several papers that investigate the parameterized complexity of backdoor detection for CSP. Bessière et al. [1], considered “partition backdoors” which are sets of variables whose deletion partitions the CSP instance into two parts, one falls into a tractable
class defined by a conservative polymorphism, and the other part is a collection of independent constraints. They also performed an empirical evaluation of the backdoor
approach which resulted in promising results. Gaspers et al. [15] considered heterogeneous backdoors into tractable CSP classes that are characterized by polymorphisms. A
similar approach was also undertaken by Carbonnel et al. [3] who also considered base
classes that are “h-Helly” for a fixed integer h under the additional assumption that the
domain is a finite subset of the natural numbers and comes with a fixed ordering.
2 Preliminaries
2.1 Valued Constraint Satisfaction
For a tuple t, we shall denote by t[i] its i-th component. We shall denote by Q the set of
all rational numbers, by Q≥ 0 the set of all nonnegative rational numbers, and by Q≥0
the set of all nonnegative rational numbers together with positive infinity, ∞. We define
α + ∞ = ∞ + α = ∞ for all α ∈ Q≥0 , and α · ∞ = ∞ for all α ∈ Q≥0 . The elements
of Q≥0 are called costs.
For every fixed set D and m ≥ 0, a function ϕ from Dm to Q≥0 will be called a
cost function on D of arity m. D is called the domain, and here we will only deal with
finite domains. If the range of ϕ is {0, ∞}, then ϕ is called a crisp cost function.
With every relation R on D, we can associate a crisp cost function ϕR on D which
maps tuples in R to 0 and tuples not in R to ∞. On the other hand, with every mary cost function ϕ we can associate a relation Rϕ defined by (x1 , . . . , xm ) ∈ Rϕ ⇔
ϕ(x1 , . . . , xm ) < ∞. In the view of the close correspondence between crisp cost functions and relations we shall use these terms interchangeably in the rest of the paper.
A VCSP instance consists of a set of variables, a set of possible values, and a
multiset of valued constraints. Each valued constraint has an associated cost function
which assigns a cost to every possible tuple of values for the variables in the scope of
the valued constraint. The goal is to find an assignment of values to all of the variables
that has the minimum total cost. A formal definition is provided below.
Definition 1 (VCSP). An instance P of the VALUED C ONSTRAINT SATISFACTION
P ROBLEM, or VCSP, is a triple (V, D, C) where V is a finite set of variables, which
are to be assigned values from the set D, and C is a multiset of valued constraints. Each
c ∈ C is a pair c = (x, ϕ), where x is a tuple of variables of length m called the scope
of c, and ϕ : Dm → Q≥0 is an m-ary cost function. An assignment for the instance
P is a mapping τ from V to D. We extend τ to a mapping from V k to Dk on tuples
of variables by applying τ componentwise. The cost of an assignment τ is defined as
follows:
CostP (τ ) =
X
ϕ(τ (x)).
(x,ϕ)∈C
The task for VCSP is the computation of an assignment with minimum cost, called
a solution to P.
For a constraint c, we will use var(c) to denote the set of variables which occur in
the scope of c. We will later also deal with the constraint satisfaction problem, or CSP.
Having already defined VCSP, it is advantageous to simply define CSP as the special
case of VCSP where each valued constraint has a crisp cost function.
The following representation of a cost function will sometimes be useful for our
purposes. A cost table for an m-ary cost function ϕ is a table with Dm rows and m + 1
columns with the following property: each row corresponds to a unique tuple a =
(a1 , . . . , am ) ∈ Dm , for each i ∈ [m] the position i of this row contains ai , and position
m + 1 of this row contains ϕ(a1 , . . . , am ).
A partial assignment is a mapping from V ′ ⊆ V to D. Given a partial assignment τ ,
the application of τ on a valued constraint c = (x, ϕ) results in a new valued constraint
c|τ = (x′ , ϕ′ ) defined as follows. Let x′ = x \ V ′ (i.e., x′ is obtained by removing
′
all elements in V ∩ x from x) and m′ = |x′ |. Then for each a′ ∈ Dm , we set
′
′
ϕ (a ) = ϕ(a) where for each i ∈ [m]
τ (x[i]) if x[i] ∈ V ′
a[i] =
a′ [i − j] otherwise, where j = |{ x[p] | p ∈ [i] } ∩ V ′ |.
Intuitively, the tuple a defined above is obtained by taking the original tuple a′ and
enriching it by the values of the assignment τ applied on the “missing” variables from
x. In the special case when x′ is empty, the valued constraint c|τ becomes a nullary
constraint whose cost function ϕ′ will effectively be a constant. The application of τ on
a VCSP instance P then results in a new VCSP instance P|τ = (V \ V ′ , D, C ′ ) where
C ′ = { c|τ | c ∈ C }. It will be useful to observe that applying a partial assignment τ can
be done in time linear in |P| (each valued constraint can be processed independently,
and the processing of each such valued constraint consists of merely pruning the cost
table).
2.2 Valued Constraint Languages
A valued constraint language (or language for short) is a set of cost functions. The arity
of a language Γ is the maximum arity of a cost function in Γ , or ∞ if Γ contains cost
functions of arbitrarily large arities. Each language Γ defines a set VCSP[Γ ] of VCSP
instances which only use cost functions from Γ ; formally, (V, D, C) ∈ VCSP[Γ ] iff
each (x, ϕ) ∈ C satisfies ϕ ∈ Γ . A language is crisp if it contains only crisp cost
functions.
A language Γ is globally tractable if there exists a polynomial-time algorithm
which solves VCSP[Γ ].1 . Similarly, a class H of VCSP instances is called tractable
if there exists a polynomial-time algorithm which solves H. For technical reasons, we
will implicitly assume that every language contains all nullary cost functions (i.e., constants); it is easily seen that adding such cost functions into a language has no impact
on its tractability.
There are a few other properties of languages that will be required to formally state
our results. A language Γ is efficiently recognizable if there exists a polynomial-time
algorithm which takes as input a cost function ϕ and decides whether ϕ ∈ Γ . We note
that every finite language is efficiently recognizable.
A language Γ is closed under partial assignments if for every instance P ∈ VCSP[Γ ]
and every partial assignment τ on P and every valued constraint c = (x, ϕ) in P, the
valued constraint c|τ = (x′ , ϕ′ ) satisfies ϕ′ ∈ Γ . The closure of a language Γ under
partial assignments, is the language Γ ′ ⊇ Γ containing all cost functions that can be
obtained from Γ via partial assignments; formally, Γ ′ contains a cost function ϕ′ if and
only if there exists a cost function ϕ ∈ Γ such that for a constraint c = (x, ϕ) and an
assignment τ : X → D defined on a subset X ⊆ var(c) we have c|τ = (x′ , ϕ′ ).
If a language Γ is closed under partial assignments, then also VCSP[Γ ] is closed under partial assignments, which is a natural property and provides a certain robustness of
the class. This robustness is also useful when considering backdoors into VCSP[Γ ] (see
Section 3), as then every superset of a backdoor remains a backdoor. Incidentally, being
closed under partial assignments is also a property of tractable classes defined in terms
of a polynomial-time subsolver [27,28] where the property is called self-reducibility.
A language is conservative if it contains all unary cost functions [18]. We note
that being closed under partial assignments is closely related to the well-studied property of conservativeness. Crucially, for every conservative globally tractable language
Γ , its closure under partial assignments Γ ′ will also be globally tractable; indeed, one
1
The literature also defines the notion of tractability [17,19], which we do not consider here.
We remark that, to the best of our knowledge, all known tractable constraint languages are also
globally tractable [17,19]
can observe that every instance P ∈ VCSP[Γ ′ ] can be converted, in linear time, to
a solution-equivalent instance P ′ ∈ VCSP[Γ ] by using infinity-valued (or even sufficiently high-valued) unary cost functions to model the effects of partial assignments.
2.3 Parameterized Complexity
We give a brief and rather informal review of the most important concepts of parameterized complexity. For an in-depth treatment of the subject we refer the reader to other
sources [5,8,12,21].
The instances of a parameterized problem can be considered as pairs (I, k) where I
is the main part of the instance and k is the parameter of the instance; the latter is
usually a non-negative integer. A parameterized problem is fixed-parameter tractable
(FPT) if instances (I, k) of size n (with respect to some reasonable encoding) can be
solved in time O(f (k)nc ) where f is a computable function and c is a constant independent of k. The function f is called the parameter dependence, and algorithms with
running time in this form are called fixed-parameter algorithms. Since the parameter
dependence is usually superpolynomial, we will often give the running times of our algorithms in O∗ notation which suppresses polynomial factors. Hence the running time
of an FPT algorithm can be simply stated as O∗ (f (k)).
The exists a completeness theory which allows to obtain strong theoretical evidence
that a parameterized problem is not fixed-parameter tractable. This theory is based on a
hierarchy of parameterized complexity classes W[1] ⊆ W[2] ⊆ . . . where all inclusions
are believed to be proper. If a parameterized problem is shown to be W[i]-hard for some
i ≥ 1, then the problem is unlikely to be fixed-parameter tractable, similarly to an
NP-complete problem being solvable in polynomial time [5,8,12,21].
3 Backdoors into Tractable Languages
This section is devoted to establishing the first general results for finding and exploiting
backdoors for VCSP. We first present the formal definition of backdoors in the context
of VCSP and describe how such backdoors once found, can be used to solve the VCSP
instance. Subsequently, we show how to detect backdoors into a single tractable VCSP
class with certain properties. In fact, our proof shows something stronger. That is, we
show how to detect heterogeneous backdoors into a finite set of VCSP classes which
satisfy these properties. The notion of heterogeneous backdoors is based on that introduced by Gaspers et al. [15]. For now, we proceed with the definition of a backdoor.
Definition 2. Let H be a fixed class of VCSP instances over a domain D and let P =
(V, D, C) be a VCSP instance. A backdoor into H is a subset X ⊆ V such that for each
assignment τ : X → D, the reduced instance P|τ is in H.
We note that this naturally corresponds to the notion of a strong backdoor in the
context of Constraint Satisfaction and Satisfiability [27,28]; here we drop the adjective
“strong” because the other kind of backdoors studied on these structures (so-called weak
backdoors) do not seem to be useful in the general VCSP setting. Namely, in analogy to
the CSP setting, one would define a weak backdoor of a VCSP instance P = (V, D, C)
into H as a subset X ⊆ V such that for some assignment τ : X → D (i) the reduced
instance P|τ is in H and (ii) τ can be extended to an assignment to V of minimum
cost. However, in order to ensure (ii) we need to compare the cost of τ with the costs of
all other assignments τ ′ to V . If X is not a strong backdoor, then some of the reduced
instances P|τ ′ restricted to X will be outside of H, and so in general we have no efficient
way of determining a minimum cost assignment for it.
We begin by showing that small backdoors for globally tractable languages can
always be used to efficiently solve VCSP instances as long as the domain is finite
(assuming such a backdoor is known).
Lemma 1. Let H be a tractable class of VCSP instances over a finite domain D. There
exists an algorithm which takes as input a VCSP instance P along with a backdoor X
of P = (V, D, C) into H, runs in time O∗ (|D||X| ), and solves P.
Proof. Let B be a polynomial-time algorithm which solves every P in H, i.e., outputs
a minimum-cost assignment in P; the existence of B follows by the tractability of H.
Consider the following algorithm A. First, A branches on the at most |D||X| -many partial assignments of X. In each branch, A then applies the selected partial assignment τ
to obtain the instance P|τ in linear time. In this branch, A proceeds by calling B on P|τ ,
and stores the produced assignment along with its cost. After the branching is complete
A reads the table of all of the at most |D||X| assignments and costs outputted by B, and
selects one assignment (say α) with a minimum value (cost) a. Let τ be the particular
partial assignment on X which resulted in the branch leading to α. A then outputs the
assignment α ∪ τ along with the value (cost) a.
⊓
⊔
Already for crisp languages it is known that having a small backdoor does not necessarily allow for efficient (i.e., fixed-parameter) algorithms when the domain is not
bounded. Specifically, the W[1]-hard k-clique problem can be encoded into a CSP
with only k variables [23], which naturally contains a backdoor of size at most k for
every crisp language under the natural assumption that the language contains the empty
constraint. Hence the finiteness of the domain in Lemma 1 is a necessary condition for
the statement to hold.
Next, we show that it is possible to find a small backdoor into VCSP[Γ ] efficiently
(or correctly determine that no such small backdoor exists) as long as Γ has two properties. First, Γ must be efficiently recognizable; it is easily seen that this condition is
a necessary one, since detection of an empty backdoor is equivalent to determining
whether the instance lies in VCSP[Γ ]. Second, the arity of Γ must be bounded. This
condition is also necessary since already in the more restricted CSP setting it was shown
that backdoor detection for a wide range of natural crisp languages (of unbounded arity)
is W[2]-hard [15].
Before we proceed, we introduce the notion of heterogeneous backdoors for VCSP
which represent a generalization of backdoors into classes defined in terms of a single
language. For languages Γ1 , . . . , Γℓ , a heterogeneous backdoor is a backdoor into the
class H = VCSP[Γ1 ] ∪ · · · ∪ VCSP[Γℓ ]; in other words, after each assignment to
the backdoor variables, all cost functions in the resulting instance must belong to a
language from our set. We now show that detecting small heterogenous backdoors is
fixed-parameter tractable parameterized by the size of the backdoor.
Lemma 2. Let Γ1 , . . . , Γℓ be efficiently recognizable languages over a domain D of
size at most d and let q be a bound on the arity of Γi for every i ∈ [ℓ]. There exists an
algorithm which takes as input a VCSP instance P over D and an integer k, runs in
time O∗ ((ℓ · d · (q + 1))k ), and either outputs a backdoor X of P into VCSP[Γ1 ] ∪ · · · ∪
VCSP[Γℓ ] such that |X| ≤ k or correctly concludes that no such backdoor exists.
Proof. The algorithm is a standard branching algorithm (see also [15]). Formally, the
algorithm is called Detectbd, takes as input an instance P = (V, D, C), integer k, a set
of variables B of size at most k and in time O∗ ((ℓ·d·(q+1))k ) either correctly concludes
that P has no backdoor Z ⊇ B of size at most k into VCSP[Γ1 ] ∪ · · · ∪ VCSP[Γℓ ]
or returns a backdoor Z of P into VCSP[Γ1 ] ∪ · · · ∪ VCSP[Γℓ ] of size at most k. The
algorithm is initialized with B = ∅.
In the base case, if |B| = k, and B is a backdoor of P into VCSP[Γ1 ] ∪ · · · ∪
VCSP[Γℓ ] then we return the set B. Otherwise, we return N O. We now move to the
description of the case when |B| < k.
In this case, if for every σ : B → D there is an i ∈ [ℓ] such that P|σ ∈ VCSP[Γi ],
then it sets Z = B and returns it. That is, if B is already found to be a backdoor of
the required kind, then the algorithm returns B. Otherwise, it computes an assignment
σ : B → D and valued constraints c1 , . . . , cℓ in P|σ such that for every i ∈ [ℓ], the cost
function of ci is not in Γi . Observe that for some σ, such a set of constraints must exist.
Furthermore, since every Γi is efficiently recognizable and B has size at most k, the
selection of these valued constraints takes time O∗ (dk ). The algorithm now constructs
a set Y as follows. Initially, Y = ∅. For each i ∈ [ℓ], if the scope of the constraint ci
contains more than q variables then it adds to Y an arbitrary q + 1-sized subset of the
scope of ci . Otherwise, it adds to Y all the variables in the scope of c. This completes
the definition of Y . Observe that any backdoor set for the given instance which contains
B must also intersect Y . Hence the algorithm now branches on the set Y . Formally, for
every x ∈ Y it executes the recursive calls Detectbd(P, k, B ∪{x}). If for some x ∈ Y ,
the invoked call returned a set of variables, then it must be a backdoor set of the given
instance and hence it is returned. Otherwise, the algorithm returns N O.
Since the branching factor of this algorithm is at most ℓ · (q + 1) and the set B,
whose size is upper bounded by k, is enlarged with each recursive call, the number of
nodes in the search tree is bounded by O((ℓ · (q + 1))k ). Since the time spent at each
node is bounded by O∗ (dk ), the running time of the algorithm Detectbd is bounded by
O∗ ((ℓ · (q + 1) · d)k ).
⊓
⊔
Combining Lemmas 1 and 2, we obtain the main result of this section.
Corollary 1. Let Γ1 , . . . , Γℓ be globally tractable and efficiently recognizable languages
each of arity at most q over a domain of size d. There exists an algorithm which solves
2
VCSP in time O∗ ((ℓ · d · (q + 1))k +k ), where k is the size of a minimum backdoor of
the given instance into VCSP[Γ1 ] ∪ · · · ∪ VCSP[Γℓ ].
4 Backdoors into Scattered Classes
Having established Corollary 1 and knowing that both the arity and domain restrictions
of the language are necessary, it is natural to ask whether it is possible to push the
frontiers of tractability for backdoors to more general classes of VCSP instances. In
particular, there is no natural reason why the instances we obtain after each assignment
into the backdoor should necessary always belong to the same language Γ even if Γ itself is one among several globally tractable languages. In fact, it is not difficult to show
that as long as each “connected component” of the instance belongs to some tractable
class after each assignment into the backdoor, then we can use the backdoor in a similar
fashion as in Lemma 1. Such a generalization of backdoors from single languages to
collections of languages has recently been obtained in the CSP setting [14] for conservative constraint languages. We proceed by formally defining these more general
classes of VCSP instances, along with some other required notions.
4.1 Scattered Classes
A VCSP instance (V, D, C) is connected if for each partition of its variable set into
nonempty sets V1 and V2 , there exists at least one constraint c ∈ C such that var(c) ∩
V1 6= ∅ and var(c) ∩ V2 6= ∅. A connected component of (V, D, C) is a maximal connected subinstance (V ′ , D, C ′ ) for V ′ ⊆ V , C ′ ⊆ C. These notions naturally correspond
to the connectedness and connected components of standard graph representations of
VCSP instances.
Let Γ1 , . . . , Γd be languages. Then the scattered class VCSP(Γ1 )⊕· · ·⊕VCSP(Γd )
is the class of all instances (V, D, C) which may be partitioned into pairwise variable
disjoint subinstances (V1 , D, C1 ), . . . , (Vd , D, Cd ) such that (Vi , D, Ci ) ∈ VCSP[Γi ]
for each i ∈ [d]. Equivalently, an instance P is in VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γd ) iff
each connected component in P belongs to some VCSP[Γi ], i ∈ [d].
Lemma 3. Let Γ1 , . . . , Γd be globally tractable languages. Then there exists a polynomialtime algorithm solving VCSP for all instances P ∈ VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γd ).
It is worth noting that while scattered classes on their own are a somewhat trivial
extension of the tractable classes defined in terms of individual languages, backdoors
into scattered classes can be much smaller than backdoors into each individual globally tractable language (or, more precisely, each individual class defined by a globally
tractable language). That is because a backdoor can not only simplify cost functions to
ensure they belong to a specific language, but it can also disconnect the instance into
several “parts”, each belonging to a different language, and furthermore the specific
language each “part” belongs to can change for different assignments into the backdoor. As a simple example of this behavior, consider the boolean domain, let Γ1 be the
globally tractable crisp language corresponding to Horn constraints [24], and let Γ2 be
a globally tractable language containing only submodular cost functions [4]. It is not
difficult to construct an instance P = (V1 ∪ V2 ∪ {x}, {0, 1}, C) such that (a) every
assignment to x disconnects V1 from V2 , (b) in P|x7→0 , all valued constraints over V1
are crisp Horn constraints and all valued constraints over V2 are submodular, and (c) in
P|x7→1 , all valued constraints over V1 are submodular and all valued constraints over
V2 are crisp Horn constraints. In the hypothetical example above, it is easy to verify
that x is a backdoor into VCSP[Γ1 ] ⊕ VCSP[Γ2 ] but the instance does not have a small
backdoor into neither VCSP[Γ1 ] nor VCSP[Γ2 ].
It is known that backdoors into scattered classes can be used to obtain fixed-parameter
algorithms for CSP, i.e., both finding and using such backdoors is FPT when dealing
with crisp languages of bounded arity and domain size [14]. Crucially, these previous
results relied on the fact that every crisp language of bounded arity and domain size is
finite (which is not true for valued constraint languages in general). We formalize this
below.
Theorem 1 ([14, Lemma 1.1]). Let Γ1 , . . . , Γℓ be globally tractable conservative crisp
languages over a domain D, with each language having arity at most q and containing
at most p relations. There exists a function f and an algorithm solving VCSP in time
O∗ (f (ℓ, |D|, q, k, p)), where k is the size of a minimum backdoor into VCSP[Γ1 ] ⊕
· · · ⊕ VCSP[Γℓ ].
Observe that in the above theorem, when q and |D| are bounded, p is immediately
bounded. However, it is important that we formulate the running time of the algorithm in
this form because in the course of our application, these parameters have to be bounded
separately. Our goal for the remainder of this section is to extend Theorem 1 in the
VCSP setting to also cover infinite globally tractable languages (of bounded arity and
domain size). Before proceeding, it will be useful to observe that if each Γ1 , . . . , Γℓ is
globally tractable, then the class VCSP[Γ1 ] ⊕ · · · ⊕ VCSP[Γℓ ] is also tractable (since
each connected component can be resolved independently of the others).
4.2 Finding Backdoors to Scattered Classes
In this subsection, we prove that finding backdoors for VCSP into scattered classes is
fixed-parameter tractable. This will then allow us to give a proof of our main theorem,
stated below.
Theorem 2. Let ∆1 , . . . , ∆ℓ be conservative, globally tractable and efficiently recognizable languages over a finite domain and having constant arity. Then VCSP is fixedparameter tractable parameterized by the size of a smallest backdoor of the given instance into VCSP(∆1 ) ⊕ · · · ⊕ VCSP(∆ℓ ).
Recall that the closure of a conservative and globally tractable language under partial assignments is also a globally tractable language. Furthermore, every backdoor
of the given instance into VCSP(∆1 ) ⊕ · · · ⊕ VCSP(∆ℓ ) is also a backdoor into
VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ) where Γi is the closure of ∆i under partial assignments.
Due to Lemma 1, it follows that it is sufficient to compute a backdoor of small size into
the scattered class VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ) where each Γi is closed under partial
assignments.
Our strategy for finding backdoors to scattered classes defined in terms of (potentially infinite) globally tractable languages relies on a two-phase transformation of the
input instance. In the first phase (Lemma 4), we show that for every choice of Γ1 , . . . , Γd
(each having bounded domain size and arity), we can construct a set of finite languages
Γ1′ , . . . , Γd′ and a new instance P ′ such that there is a one-to-one correspondence between backdoors of P into Γ1 ⊕ · · · ⊕ Γd and backdoors of P ′ into Γ1′ ⊕ · · · ⊕ Γd′ . This
allows us to restrict ourselves to only the case of finite (but not necessarily crisp) languages as far as backdoor detection is concerned. In the second phase (Lemma 5), we
transform the instance and languages one more time to obtain another instance P ′′ along
with finite crisp languages Γ1′′ , . . . , Γd′′ such that there is a one-to-one correspondence
between the backdoors of P ′′ and backdoors of P ′ . We crucially note that the newly
constructed instances are equivalent only with respect to backdoor detection; there is no
correspondence between the solutions of these instances.
Before proceeding, we introduce a natural notion of replacement of valued constraints which is used in our proofs.
Definition 3. Let P = (V, D, C) be a VCSP instance and let c = (x, ϕ) ∈ C. Let ϕ′
be a cost function over D with the same arity as ϕ. Then the operation of replacing ϕ
in c with ϕ′ results in a new instance P ′ = (V, D, (C \ {c}) ∪ {(x, ϕ′ )}).
Lemma 4. Let Γ1 , . . . , Γℓ be efficiently recognizable languages closed under partial
assignments, each of arity at most q over a domain D of size d. There exists an algorithm
which takes as input a VCSP instance P = (V, D, C) and an integer k, runs in time
O∗ (f (ℓ, d, k, q)) for some function f and either correctly concludes that P has no
backdoor into VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ) of size at most k or outputs a VCSP
instance P ′ = (V, D′ , C ′ ) and languages Γ1′ , . . . , Γℓ′ with the following properties.
1. For each i ∈ [ℓ], the arity of Γi′ is at most q
2. For each i ∈ [ℓ], Γi′ is over D′ and D′ ⊆ D
3. Each of the languages Γ1′ , . . . , Γℓ′ is closed under partial assignments and contains
at most g(ℓ, d, k, q) cost functions for some function g.
4. For each X ⊆ V , X is a minimal backdoor of P into VCSP(Γ1 )⊕· · ·⊕ VCSP(Γℓ )
of size at most k if and only if X is a minimal backdoor of P ′ into VCSP(Γ1′ ) ⊕
· · · ⊕ VCSP(Γℓ′ ) of size at most k.
Proof. We will first define a function mapping the valued constraints in C to a finite set
whose size depends only on ℓ, d, k and q. Subsequently, we will show that every pair of
constraints in C which are mapped to the same element of this set are, for our purposes
(locating a backdoor), interchangeable. We will then use this observation to define the
new instance P ′ and the languages Γ1′ , . . . , Γℓ′′ . To begin with, observe that if the arity
of a valued constraint in P is at least q + k + 1, then P has no backdoor of size at most k
into VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ). Hence, we may assume without loss of generality
that the arity of every valued constraint in P is at most q + k.
Let F be the set of all functions from [q + k] × 2[q+k] × D[q+k] → 2[ℓ] ∪ {⊥},
where ⊥ is used a special symbol expressing that F is “out of bounds.” Observe that
(q+k)+log(q+k)
|F | ≤ η(ℓ, d, k, q) = (2ℓ + 1)(2d)
. We will now define a function Type :
C → F as follows. We assume without loss of generality that the variables in the scope
of each constraint in C are numbered from 1 to |var(c)| based on their occurrence in
the tuple x where c = (x, ϕ). Furthermore, recall that |var(c)| ≤ q + k. For c ∈ C, we
define Type(c) = δ ∈ F where δ is defined as follows. Let r ≤ q + k, Q ⊆ [q + k] and
γ : [q +k] → D. Let γ[Q∩[r]] denote the restriction of γ to the set Q∩[r]. Furthermore,
recall that c|γ[Q∩[r]] denotes the valued constraint resulting from applying the partial
assignment γ on the variables of c corresponding to all those indices in Q ∩ [r].
Then, δ(r, Q, γ) = ⊥ if r 6= |var(c)|. Otherwise, δ(r, Q, γ) = L ⊆ [ℓ] where
i ∈ [ℓ] is in L if and only if c|γ[Q∩[r]] ∈ VCSP(Γi ). This completes the description
of the function Type; observe that Type(c) can be computed in time which is upperbounded by a function of ℓ, d, k, q.
For every δ ∈ F , if there is a valued constraint c ∈ C such that Type(c) = δ, we
pick and fix one arbitrary such valued constraint c∗δ = (x∗δ , ϕ∗δ ). We now proceed to the
definition of the instance P ′ and the languages Γ1′ , . . . , Γℓ′′ .
Observe that for 2 constraints c = (x1 , ϕ), c′ = (x′ , ϕ′ ) ∈ C, if Type(c) =
Type(c′ ) then |var(c)| = |var(c′ )|. Hence, the notion of replacing ϕ in c with ϕ′ is
well-defined (see Definition 3). We define the instance P ′ as the instance obtained from
P by replacing each c = (x, ϕ) ∈ C with the constraint (x, ϕ∗δ ) where δ = Type(c).
For each i ∈ [ℓ] and cost function ϕ ∈ Γi , we add ϕ to the language Γi′ if and only if
for some δ ∈ F and some set Q ⊆ var(c∗δ ) and assignment γ : Q → D, the constraint
c|γ[Q] = (x \ Q, ϕ). Clearly, for every i ∈ [ℓ], |Γi′ | ≤ dq · |F | ≤ dq · η(ℓ, d, k, q).
Finally, for each Γi′ , we compute the closure of Γi′ under partial assignments and add
each relation from this closure into Γi′ . Since the size of each Γi′ is bounded initially
in terms of ℓ, d, k, q, computing this closure can be done in time O∗ (λ(ℓ, d, k, q)) for
some function λ. Since each cost function has arity q and domain D, the size of the
final language Γi′ obtained after this operation is blown up by a factor of at most dq ,
implying that in the end, |Γi′ | ≤ d2q · |F | ≤ d2q · η(ℓ, d, k, q).
Now, observe that the first two statements of the lemma follow from the definition
of the languages {Γi′ }i∈[ℓ] . Furthermore, the number of cost functions in each Γi′ is
bounded by dq · η(ℓ, d, k, q), and so the third statement holds as well. Therefore, it
only remains to prove the final statement of the lemma. Before we do so, we state a
straightforward consequence of the definition of P ′ .
Observation 1 For every Y ⊆ V , γ : Y → D and connected component H′ of P ′ |γ ,
there is a connected component H of P|γ and a bijection ψ : H → H′ such that for
every c ∈ H, Type(c) = Type(ψ(c)). Furthermore, for every c = (x, ϕ) ∈ H, the
constraint ψ(c) is obtained by replacing ϕ in c with ϕ∗Type(c) .
We now return to the proof of Lemma 4. Consider the forward direction and let X
be a backdoor of size at most k for P into VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ) and suppose
that X is not a backdoor for P ′ into VCSP(Γ1′ ) ⊕ · · · ⊕ VCSP(Γℓ′ ). Then, there is an
assignment γ : X → D such that for some connected component H of P ′ |γ , there is
no i ∈ ℓ such that all constraints in H′ lie in VCSP(Γi′ ). By Observation 1 above, there
is a connected component H in P|γ and a bijection ψ : H → H′ such that for every
c ∈ H, Type(c) = Type(ψ(c)). Since X is a backdoor for P, there is a j ∈ ℓ such
that all constraints in H lie in VCSP(Γj ). Pick an arbitrary constraint c = (x, ϕ) ∈ H.
Let c′ = (x, ϕ∗Type(c) ) be the constraint ψ(c). By definition of ϕ∗Type(c) it follows that
c′ |γ ∈ VCSP(Γj ). The fact that this holds for an arbitrary constraint in H along with
the fact that ψ is a bijection implies that every constraint in H′ is in fact in VCSP(Γj′ ),
a contradiction. The argument in the converse direction is symmetric. This completes
the proof of the final statement of the lemma.
The time taken to compute P ′ and the languages Γ1′ , . . . , Γℓ′ is dominated by the
time required to compute the function Type. Since the languages Γ1 , . . . , Γℓ are ef-
ficiently recognizable, this time is bounded by O∗ (|F |), completing the proof of the
lemma.
⊓
⊔
Lemma 5. Let Γ1 , . . . , Γℓ be efficiently recognizable languages closed under partial
assignments, each of arity at most q over a domain D of size d. Let P ′ = (V, D′ , C ′ )
be the VCSP instance and let Γ1′ , . . . , Γℓ′ be languages returned by the algorithm of
Lemma 4 on input P and k. There exists an algorithm which takes as input P ′ , these
languages and k, runs in time O∗ (f (ℓ, d, k, q)) for some function f and outputs a CSP
instance P ′′ = (V ′′ ⊇ V, D′′ , C ′′ ) and crisp languages Γ1′′ , . . . , Γℓ′′ with the following
properties.
1. For each i ∈ [ℓ], the arity of Γi′′ is at most q + 1
2. D′′ ⊇ D and |D′′ | ≤ β(q, d, k) for some function β.
3. The number of relations in each of the languages Γ1′′ , . . . , Γℓ′′ is at most α(q, d, k)
for some function α.
4. if X is a minimal backdoor of arity at most k of P ′′ into CSP(Γ1′′ )⊕· · ·⊕CSP(Γℓ′′ ),
then X ⊆ V .
5. For each X ⊆ V , X is a minimal backdoor of P ′ into VCSP(Γ1′ )⊕· · ·⊕VCSP(Γℓ′ )
if and only if X is a minimal backdoor of P ′′ into CSP(Γ1′′ ) ⊕ · · · ⊕ CSP(Γℓ′′ ).
Proof. We propose a fixed-parameter algorithm A, and show that it has the claimed
properties. It will be useful to recall that we do not distinguish between crisp cost functions and relations. We also formally assume that D′ does not intersect the set of rationals Q; if this is not the case, then we simply rename elements of D′ to make sure that
this holds. Within the proof, we will use a ◦ b to denote the concatenation of vector a
by element b.
First, let Ti be the set of all values which are returned
by at least one cost function
S
from Γi′ , i ∈ [ℓ], for at least one input. Let T = i∈[ℓ] Ti . Observe that |T | is upperbounded by the size, domain and arity of our languages. Let us now set D′′ = D′ ∪T ∪ǫ.
Intuitively, our goal will be to represent the cost function in each valued constraint in
P ′ by a crisp cost function with one additional variable which ranges over T , where T
corresponds to a specific value which occurs in one of our base languages. Note that this
satisfies Condition 2. of the lemma, and that T can be computed in linear time from the
cost tables of Γ1′ , . . . , Γℓ′ . We will later construct k + 1 such representations (each with
its own additional variable) to ensure that the additional variables are never selected by
minimal backdoors.
Next, for each language Γi′ , i ∈ [ℓ], we compute a new crisp language Γi′′ as follows.
For each ϕ ∈ Γi′ of arity t, we add a new relation ψ of arity t + 1 into Γi′′ , and for each
tuple (x1 , . . . , xt ) of elements from D′ we add the tuple (x1 , . . . , xt , ϕ(x1 , . . . , xt ))
into ψ; observe that this relation exactly corresponds to the cost table of ϕ. We then
compute the closure of Γi′′ under partial assignments and add each relation from this
closure into Γi′′ . Observe that the number of relations in Γi′′ is bounded by a function
of |T | and |Γi′ |, and furthermore the number of tuples in each relation is upper-bounded
′
by q |D | , and so Conditions 1. and 3. of the lemma hold. The construction of each Γi′′
from Γi′ can also be done in linear time from the cost tables of Γ1′ , . . . , Γℓ′ .
Finally, we construct a new instance P ′′ = (V ′′ , D′′ , C ′′ ) from P ′ = (V, D′ , C ′ ) as
follows. At the beginning, we set V ′′ := V . For each c′ = (x′ , ϕ′ ) ∈ C ′ , we add k + 1
unique new variables vc1′ , . . . , vck+1
into V ′′ and add k + 1 constraints c′′1 , . . . , c′′k+1
′
′′
′′i
into C . For i ∈ [k + 1], each c = (x′ ◦ vci ′ , ψ ′′ ) where ψ ′′ is a relation that is constructed similarly as the relations in our new languages Γi′′ above. Specifically, for each
tuple (x1 , . . . , xt ) of elements from D′ we add the tuple (x1 , . . . , xt , ϕ′ (x1 , . . . , xt ))
into ψ ′′ , modulo the following exception. If ϕ′ (x1 , . . . , xt ) 6∈ D′′ , then we instead add
the tuple (x1 , . . . , xt , ǫ) into ψ ′′ . Clearly, the construction of our new instance P ′′ takes
′
time at most O(|C ′ | + q |D | ). This concludes the description of A.
It remains to argue that Conditions 4. and 5. of the lemma hold. First, consider a
minimal backdoor X of size at most k of P ′′ into CSP[Γ1′′ ] ⊕ · · · ⊕ CSP[Γℓ′′ ], and
assume for a contradiction that there exists some c′ = (x′ , ϕ′ ) ∈ C ′ and i ∈ [k + 1]
such that vci ′ ∈ X. First, observe that this cannot happen if the whole scope of c′′i lies
in X. By the size bound on X, there exists j ∈ [k + 1] such that vcj′ 6∈ X. Then for each
partial assignment τ of X, the relation ϕ′′ in c′′j belongs to the same globally tractable
language as the rest of the connected component of P ′′ containing the scope of c′′ (after
applying τ ). Since the relation ϕ′′ in c′′j is precisely the same as in c′′i and the scope of
c′′i must lie in the same connected component as that of c′′j , it follows that X \ {vxi ′ }
is also a backdoor of P ′′ into CSP(Γ1′′ ) ⊕ · · · ⊕ CSP(Γℓ′′ ). However, this contradicts
the minimality of X.
Finally, for Condition 5., consider an arbitrary backdoor X of P ′ into VCSP(Γ1′ ) ⊕
· · · ⊕ VCSP(Γℓ′ ), and let us consider an arbitrary assignment from X to D′′ . It will
be useful to note that while the contents of relations and/or cost functions in individual
(valued) constraints depend on the particular choice of the assignment to X, which
variables actually occur in individual components depends only on the choice of X and
remains the same for arbitrary assignments.
Now observe that each connected component P CSP of P ′′ after the application of
the (arbitrarily chosen) assignment will fall into one of the following two cases. P CSP
could contain a single variable vc′ with a single constraint whose relation lies in every
language Γi′′ , i ∈ [ℓ]; this occurs precisely when the whole scope of a valued constraint
c′ ∈ C ′ lies in X, and the relation will either contain a singleton element from T or
be the empty relation. In this case, we immediately conclude that P CSP ∈ CSP(Γi ) for
each i ∈ [ℓ].
Alternatively, P CSP contains at least one variable v ∈ V . Let P VCSP be the unique
connected component of P ′ obtained after the application of an arbitrary assignment
from X to D′ which contains v. Observe that the variable sets of P CSP and P VCSP
only differ in the fact that P CSP may contain some of the newly added variables vc′ for
various constraints c′ . Now let us consider a concrete assignment τ from X to D′ along
with an i ∈ [ℓ] such that after the application of τ , the resulting connected component
P VCSP belongs to VCSP(Γi′ ). It follows by our construction that applying the same
assignment τ in P ′′ will result in a connected component P CSP corresponding to P VCSP
such that P VCSP ∈ CSP(Γi′′ ); indeed, whenever Γi′ contains an arbitrary cost function
ϕ(x) = β, the language Γi′′ will contain the relation (x ◦ β).
By the above, the application of an assignment from X to D′ in P ′′ will indeed
result in an instance in CSP(Γ1′′ ⊕ · · · ⊕ Γℓ′′ ). But recall that the domain of P ′′ is D′′ ,
which is a superset of D′ ; we need to argue that the above also holds for assignments
τ from X to D′′ . To this end, consider an arbitrary such τ and let τ0 be an arbitrary
assignment from X to D′ which matches τ on all mappings into D′ . Let us compare
the instances Pτ′′0 and Pτ′′ . By our construction of P ′ , whenever τ maps at least one
variable from the scope of some constraint c′′ to D′′ \ D′ , the resulting relation will be
the empty relation. It follows that each constraint in Pτ′′ will either be the same as in
Pτ′′ , or will contain the empty relation. But since the empty relation is included in every
language Γ1′′ , . . . , Γℓ′′ , we conclude that each connected component of Pτ′′ must belong
to at least one language Γi′′ , i ∈ [ℓ]. This shows that X must also be a backdoor of P ′′
into CSP[Γ1′′ ] ⊕ · · · ⊕ CSP[Γℓ′′ ].
For the converse direction, consider a minimal backdoor X of P ′′ into CSP[Γ1′′ ] ⊕
· · · ⊕ CSP[Γℓ′′ ]. Since we already know that Condition 4. holds, X must be a subset
of V . The argument from the previous case can then simply be reversed to see that X
will also be a backdoor of P ′ into VCSP[Γ1′ ] ⊕ · · · ⊕ VCSP[Γℓ′ ]; in fact, the situation
in this case is much easier since only assignments into D′ need to be considered.
Summarizing, we gave a fixed-parameter algorithm and then showed that it satisfies
each of the required conditions, and so the proof is complete.
⊓
⊔
We are now ready to prove Theorem 2, which we restate for the sake of convenience.
Theorem 2. Let ∆1 , . . . , ∆ℓ be conservative, globally tractable and efficiently recognizable languages over a finite domain and having constant arity. Then VCSP is fixedparameter tractable parameterized by the size of a smallest backdoor of the given instance into VCSP(∆1 ) ⊕ · · · ⊕ VCSP(∆ℓ ).
Proof. For each i ∈ [ℓ], let Γi denote the closure of ∆i under partial assignments.
Observe that every backdoor of the given instance into VCSP(∆1 ) ⊕ · · · ⊕ VCSP(∆ℓ )
is also a backdoor into VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ). Furthermore, each VCSP(Γi )
is tractable since VCSP(∆1 ) is tractable and conservative. Hence, it is sufficient to
compute and use a backdoor of size at most k into VCSP(Γ1 ) ⊕ · · · ⊕ VCSP(Γℓ ).
The claimed algorithm has two parts. The first one is finding a backdoor into VCSP(Γ1 )⊕
· · · ⊕ VCSP(Γℓ ) and the second one is using the computed backdoor to solve VCSP.
Given an instance P and k, we first execute the algorithm of Lemma 4 to compute the
instance P ′ , and the languages Γ1′ , . . . , Γℓ′ with the properties stated in the lemma. We
then execute the algorithm of Lemma 5 with input P ′ , k, and Γ1′ , . . . , Γℓ′ to compute the
CSP instance P ′′ and crisp languages Γ1′′ , . . . , Γℓ′′ with the stated properties. Following
this, we execute the algorithm of Theorem 1 with input P ′′ , k. If this algorithm returns
N O then we return N O as well. Otherwise we return the set returned by this algorithm
as a backdoor of size at most k for the given instance P. Finally, we use the algorithm
of Lemma 1 with H set to be the class VCSP(Γ1 ) ⊕ · · ·⊕ VCSP(Γℓ ), to solve the given
instance.
The correctness as well as running time bounds follow from those of Lemmas 4
and 5, Theorem 1, and Lemma 1. This completes the proof of the theorem.
⊓
⊔
5 Concluding Remarks
We have introduced the notion of backdoors to the VCSP setting as a means for augmenting a class of globally tractable VCSP instances to instances that are outside the
class but of small distance to the class. We have presented fixed-parameter tractability
results for solving VCSP instances parameterized by the size of a smallest backdoor
into a (possibly scattered and heterogeneous) tractable class satisfying certain natural
properties.
Our work opens up several avenues for future research. Since our main objective
was to establish the fixed-parameter tractability of this problem, we have not attempted
to optimize the runtime bounds for finding backdoors to scattered classes. As a result,
it is quite likely that a more focussed study of scattered classes arising from specific
constraint languages will yield a significantly better runtime. A second interesting direction would be studying the parameterized complexity of detection of backdoors into
tractable VCSP classes that are characterized by specific fractional polymorphisms.
References
1. C. Bessiere, C. Carbonnel, E. Hebrard, G. Katsirelos, and T. Walsh. Detecting and exploiting subproblem tractability. In F. Rossi, editor, IJCAI 2013, Proceedings of the 23rd
International Joint Conference on Artificial Intelligence, Beijing, China, August 3-9, 2013.
IJCAI/AAAI, 2013.
2. C. Carbonnel and M. C. Cooper. Tractability in constraint satisfaction problems: a survey.
Constraints, 21(2):115–144, 2016.
3. C. Carbonnel, M. C. Cooper, and E. Hebrard. On backdoors to tractable constraint languages.
In B. O’Sullivan, editor, Principles and Practice of Constraint Programming - 20th International Conference, CP 2014, Lyon, France, September 8-12, 2014. Proceedings, volume
8656 of Lecture Notes in Computer Science, pages 224–239. Springer Verlag, 2014.
4. D. A. Cohen, P. G. Jeavons, and S. Zivny. The expressive power of valued constraints:
Hierarchies and collapses. Theoretical Computer Science, 409(1):137–153, 2008.
5. M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk,
and S. Saurabh. Parameterized Algorithms. Springer, 2015.
6. B. N. Dilkina, C. P. Gomes, and A. Sabharwal. Tradeoffs in the complexity of backdoor detection. In Principles and Practice of Constraint Programming - CP 2007, 13th International
Conference, CP 2007, Providence, RI, USA, September 23-27, 2007, Proceedings, volume
4741 of Lecture Notes in Computer Science, pages 256–270. Springer Verlag, 2007.
7. B. N. Dilkina, C. P. Gomes, and A. Sabharwal. Backdoors in the context of learning. In
O. Kullmann, editor, Theory and Applications of Satisfiability Testing - SAT 2009, 12th International Conference, SAT 2009, Swansea, UK, June 30 - July 3, 2009. Proceedings, volume
5584 of Lecture Notes in Computer Science, pages 73–79. Springer Verlag, 2009.
8. R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Texts in
Computer Science. Springer Verlag, 2013.
9. W. Dvorák, S. Ordyniak, and S. Szeider. Augmenting tractable fragments of abstract argumentation. Artificial Intelligence, 186:157–173, 2012.
10. J. K. Fichte and S. Szeider. Backdoors to normality for disjunctive logic programs. ACM
Trans. Comput. Log., 17(1), 2015.
11. J. K. Fichte and S. Szeider. Backdoors to tractable answer set programming. Artificial
Intelligence, 220:64–103, Mar. 2015.
12. J. Flum and M. Grohe. Parameterized Complexity Theory, volume XIV of Texts in Theoretical Computer Science. An EATCS Series. Springer Verlag, Berlin, 2006.
13. R. Ganian and S. Ordyniak. The complexity landscape of decompositional parameters for
ILP. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence. AAAI Press,
2016. To appear.
14. R. Ganian, M. S. Ramanujan, and S. Szeider. Discovering archipelagos of tractability for
constraint satisfaction and counting. In Proceedings of the Twenty-Seventh Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January 10-12,
2016, pages 1670–1681, 2016.
15. S. Gaspers, N. Misra, S. Ordyniak, S. Szeider, and S. Živný. Backdoors into heterogeneous
classes of SAT and CSP. In C. E. Brodley and P. Stone, editors, Proceedings of the TwentyEighth AAAI Conference on Artificial Intelligence, July 27 -31, 2014, Québec City, Québec,
Canada, pages 2652–2658. AAAI Press, 2014.
16. S. Gaspers and S. Szeider. Backdoors to satisfaction. In H. L. Bodlaender, R. Downey, F. V.
Fomin, and D. Marx, editors, The Multivariate Algorithmic Revolution and Beyond - Essays
Dedicated to Michael R. Fellows on the Occasion of His 60th Birthday, volume 7370 of
Lecture Notes in Computer Science, pages 287–317. Springer Verlag, 2012.
17. P. Jeavons, A. A. Krokhin, and S. Živný. The complexity of valued constraint satisfaction.
Bulletin of the European Association for Theoretical Computer Science, 113, 2014.
18. V. Kolmogorov and S. Živný. The complexity of conservative valued CSPs. J. of the ACM,
60(2):Art. 10, 38, 2013.
19. A. Krokhin, A. Bulatov, and P. Jeavons. The complexity of constraint satisfaction: An algebraic approach. In Structural Theory of Automata, Semigroups and Universal Algebra
(Montreal, 2003), number 207 in NATO Science Series II: Mathematics, Physics, and Chemistry, pages 181–213, 2005.
20. R. LeBras, R. Bernstein, C. P. Gomes, B. Selman, and R. B. van Dover. Crowdsourcing
backdoor identification for combinatorial optimization. In F. Rossi, editor, IJCAI 2013, Proceedings of the 23rd International Joint Conference on Artificial Intelligence, Beijing, China,
August 3-9, 2013, 2013.
21. R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford Lecture Series in Mathematics and its Applications. Oxford University Press, Oxford, 2006.
22. N. Nishimura, P. Ragde, and S. Szeider. Detecting backdoor sets with respect to Horn and
binary clauses. In Proceedings of SAT 2004 (Seventh International Conference on Theory
and Applications of Satisfiability Testing, 10–13 May, 2004, Vancouver, BC, Canada), pages
96–103, 2004.
23. C. H. Papadimitriou and M. Yannakakis. On the complexity of database queries. J. of
Computer and System Sciences, 58(3):407–427, 1999.
24. T. J. Schaefer. The complexity of satisfiability problems. In Conference Record of the Tenth
Annual ACM Symposium on Theory of Computing (San Diego, Calif., 1978), pages 216–226.
ACM, 1978.
25. J. Thapper and S. Živný. Necessary conditions for tractability of valued CSPs. SIAM J.
Discrete Math., 29(4):2361–2384, 2015.
26. S. Živný. The Complexity of Valued Constraint Satisfaction Problems. Cognitive Technologies. Springer, 2012.
27. R. Williams, C. Gomes, and B. Selman. Backdoors to typical case complexity. In G. Gottlob and T. Walsh, editors, Proceedings of the Eighteenth International Joint Conference on
Artificial Intelligence, IJCAI 2003, pages 1173–1178. Morgan Kaufmann, 2003.
28. R. Williams, C. Gomes, and B. Selman. On the connections between backdoors, restarts, and
heavy-tailedness in combinatorial search. In Informal Proc. of the Sixth International Conference on Theory and Applications of Satisfiability Testing, S. Margherita Ligure - Portofino,
Italy, May 5-8, 2003 (SAT 2003), pages 222–230, 2003.
| 8cs.DS
|
Correlation Properties of Signal at Mobile Receiver
for Different Propagation Environments
Cezary Ziółkowski, Jan M. Kelner, Leszek Nowosielski
Institute of Telecommunications, Faculty of Electronics
Military University of Technology
Warsaw, Poland
{cezary.ziolkowski, jan.kelner, leszek.nowosielski}@wat.edu.pl
Abstract—An issue of the parameter selection in various
branches of a multi-antenna receiver system determines its
effectiveness. A significant effect on these parameters are
correlation properties of received signals. In this paper, the
assessment of the signal correlation properties for different
environmental conditions is presented. The obtained results
showed that depending on the receiver speed, the adaptive
selection of the delays in the different RAKE receiver branches
provide minimization of the correlation between the signals.
Particularly low levels of the signal correlation could be obtained
in complex propagation environments such as urban and bad
urban.
Keywords—propagation; channel model; autocorrelation
function; correlational properties; angle distribution; modified
Laplacian; delay spread; mobile receiver; RAKE receiver; type of
environment; modeling; simulation
I. INTRODUCTION
The impact of the environment on the autocorrelation
function (ACF) properties is important for the development of
multi-antenna systems and processing methods of the received
signals. For individual branches of the system, the evaluation
of ACF properties allows to determine the parameters that
provide minimize the correlation of the received signals. In
RAKE receiver [1][2][3][4], knowing ACF properties is the
basis for choosing the values of the coefficients and delays of
each branch of the tapped delay line. Superposition of all
signals from line ensures minimization of frequency selective
fading. This mean that ACF properties play a large role in
minimizing intersymbol interference and, as a consequence,
they significantly affect the capacity of the entire system.
The angular spread of power is one of the main factors that
determine ACF properties of the received signal. Hence, the
assessment of these properties is determined by selection of the
appropriate angular spread model that provides mapping the
effects of actual propagation phenomena. The basic function
that defines a statistical spread model is the probability density
function (PDF) of angle of arrival (AOA). In this set of models,
we can distinguish analytical (geometrical) and empirical
models. The analytical models are based on the geometrical
structures that described the spatial location of the scattering
elements. The empirical models whereas use standard PDFs. In
this case, the criterion for selection of model parameters is to
minimize the approximation error with respect to measurement
data. The analytical models characterized by considerable
complexity of the analytical description compared to empirical
models. In addition, in [5], the comparison of the results of
approximation shows that the empirical models provide a
smaller approximation error of measurement data with respect
to the analytical models.
Influence of the propagation environment on the ACF
properties is the main purpose of this paper. The assessment of
this impact is based on the selected PDF of AOA empirical
model that minimizes the approximation error of data from
several measurement campaigns. The modified Laplacian
distribution is one of the fundamental empirical models that is
used to describe PDF of AOA. The accuracy of measurement
data mapping for this model is presented in [6][7], and a
comparative assessment with other models is contained in [8].
The presented results are the reason for choosing this model to
assess ACF properties of the received signal by mobile receiver
in different propagation conditions. Propagation environment
type is defined on the basis of the rms delay spread (DS), στ.
Adaptation of the modified Laplacian uses the relationship
between the parameters of this model and DS. Dispersion of
AOA versus DS allows the assessment of the impact of
propagation environment on AFC properties that among other
things, determine the efficacy of using multi-antenna systems.
A method of determining the empirical model parameters as a
function of DS is shown based on measurement data in [9].
The remainder of the paper is organized as follows. In
Section II, we present the theoretical basis for the
determination of ACF analytical form of received signal by a
moving receiver. This section also shows the PDF of AOA and
an adapting method its parameters to the type of propagation
environment. The calculation results of averaging procedure
for the different environmental parameters are presented in
Section III. From the obtained results, final conclusions are
contained in Section IV.
II. NORMALIZED CORRELATION FUNCTION
AND ANGLE DISTRIBUTION
The fundamental model that defines ACF of the received
signal by the mobile receiver is a two-dimensional isotropic
Cezary Ziółkowski, Jan M. Kelner, and Leszek Nowosielski,
“Correlation properties of signal at mobile receiver for different propagation environments,”
in 2018 14th International Conference on Advanced Trends in Radioelectronics,
Telecommunications and Computer Engineering (TCSET),
Lviv - Slavske, Ukraine, 20-24.02.2018., pp. 1–4.
Clark model. For this model, the analytic form of the low-pas
received signal, x(t), for the unmodulated carrier is [10]
N
x t an exp 2πi f Dm t cos n i n
For the Clark model, PDF of AOA, f (φ), is described by the
uniformly distributed. In this case, we have
r J 0 2π
n 1
where N is the number of propagation paths, an represents the
signal amplitude of the nth path, fDm is the maximum frequency
of the Doppler shift, φn means angle between the nth path and
velocity vector of the receiver, γn is the random phase of signal
component from the nth path.
In practice, all parameters of x(t) are statistically
independent and γn values are equally likely. It is the premise
that PDF of each γn (n = 1,2,…,N) can be assumed to be
uniformly distributed. By using these assumptions, the ACF,
R(τ), for time interval that conditions stationary properties of
x(t), is
R E x t x* t
ak exp 2πi f Dm t cos k i k
k 1
E N
a exp 2πi f t cos i
l
Dm
l
l
l 1
where J0(⸱) is the zero-order Bessel function of the first kind
and τ’ = fDmτ is a normalized time.
However, a number of measurement results executed in real
environment significantly different from the uniform PDF. This
fact is shown in many publications, for example [6]. In this
paper, our attention focuses on the empirical model f (φ) that
provides a simple analytical description and minimizes
approximation errors of measurement data. In [8], the
comparative analysis of empirical models shows that the
modified Laplacian gives the smallest approximation error for
different propagation scenarios. Therefore, in the paper, this
PDF is used to analyze the signal ACF. Modified Laplacian has
a form
N
N
E an2 exp 2πi f Dm cos n
n 1
N
n 1
where E{⸱} is an averaging operator after sets of an and φn
values.
The statistical properties of each AOA are the same for all
propagation paths. Thus, we have
R E exp 2πif Dm cos E an2
N
n 1
E a
2
where C 1 exp π
1
n 1
2
n
such that
where
rI irQ
rI E cos 2π f Dm cos
9.44 0.40 1
Considering (4) and (6), the analytical forms of the in-phase
and quadrature components of ACF, we can express as
2
π
cos 2π cos exp d
π
and
rQ E sin 2π f Dm cos are the in-phase and quadrature
components of r(τ).
is a normalized factor
f d 1 .
rI , C
Hence, the normalized ACF, r(τ) has a form
for π, π
Adaptation of f (φ) to the type of propagation environment
consists in appropriately adjusting the PDF parameters. In this
procedure, a statistical relationship between DS and the rms
angle spread (AS), σφ, is used [6]. Details of adaptation method
are shown in [9]. Considering the results of several
measurement campaigns, the relationship between λ and στ is
approximated by [9]
is the average power of received signal.
r R R 0 E exp 2πi f Dm cos
π
N
where
exp
π
E an2 E exp 2πi f Dm cos n
f C
rQ , C
2
π
sin 2π cos exp d
π
Cezary Ziółkowski, Jan M. Kelner, and Leszek Nowosielski,
“Correlation properties of signal at mobile receiver for different propagation environments,”
in 2018 14th International Conference on Advanced Trends in Radioelectronics,
Telecommunications and Computer Engineering (TCSET),
Lviv - Slavske, Ukraine, 20-24.02.2018., pp. 1–4.
Formulas (8) and (9) are the basis to determine the module
of ACF, r(τ’, στ), which enables the ACF assessment of the
received signal in different environmental conditions.
III. NUMERICAL CALCULATION
In the paper, we use numerical calculation for analysis of
signal ACF. Such an approach to analyzed problem results
from the complexity of analytic expressions (8) and (9).
Numerical calculation of integrals are made for three different
values of στ (0.1 μs, 0.2 μs, 1.0 μs) that define the propagation
environments of rural area and typical urban types. The
obtained results of rI(τ’, στ), rQ(τ’, στ), and | r(τ’, στ) | are
presented in Figs. 1-3.
Fig. 1. In-phase component of ACF for different environment types.
As shown in graphs, the strong correlation of the signal
occurs for rural area, and | r(τ’, στ) | graph is converges to the
Clark model with the increased complexity of the propagation
environment. This means that the application of the multiantenna system in rural area gives a significantly smaller
improvement in the quality and channel capacity compared to
an environment that has a large
value. The graph of
| r(τ’, στ) | shows that in individual branches, the selection of the
delays minimizes correlate the received signals in the multiantenna system. The minimum value of ACF that can be
obtained for the first local minimum (τ’ = τ’min) are shown in
Fig. 4 for different propagation environments (versus στ). In
addition, the maximum values of | r(τ’, στ) | (the first local
maximum τ’ = τ’min) are presented in this figure.
Fig. 4. First local minimum and maximum of | r(τ’, στ) | for different
environment types.
The figure indicates that for the higher the complexity of
propagation environment, the ACF value can be a smaller.
Comparing graphs shows that the use of procedure for delaying
the signals in individual branches can provide the efficiency
increase of signal reception.
A measure
r r m ax, r m in ,
Fig. 2. Quadrature component of ACF for different environment types.
that gives you the opportunity to evaluate the use effectiveness
of the first local minimum of the module of ACF in the
branches receiver, is shown in Fig. 5.
Fig. 3. Module of ACF for different environment types.
Fig. 5. Difference between first local maximum and minimum of ACF for
different environment types.
Cezary Ziółkowski, Jan M. Kelner, and Leszek Nowosielski,
“Correlation properties of signal at mobile receiver for different propagation environments,”
in 2018 14th International Conference on Advanced Trends in Radioelectronics,
Telecommunications and Computer Engineering (TCSET),
Lviv - Slavske, Ukraine, 20-24.02.2018., pp. 1–4.
The graph of Δr(στ) shows that the selection of the signal
delays in various branches of the receiver could provide several
times reduction of the signal correlation for complex
propagation environments such as urban and bad urban areas.
IV. CONCLUSION
The issue of the parameter selection in the various branches
of the multi-antenna system and processing method determines
their effectiveness. A significant effect on these parameters are
the correlation properties of the received signals. In this paper,
the assessment of signal correlation properties for different
environmental conditions is presented. The obtained results
showed that depending on the receiver speed, the adaptive
selection of the delays in the different RAKE receiver branches
provide minimization of the correlation between the signals.
This mean that ACF properties play a large role in minimizing
intersymbol interference and, as a consequence, they
significantly affect the capacity of the entire system.
Particularly low levels of the signal correlation could be
obtained in complex propagation environments such as urban
and bad urban.
REFERENCES
[1]
[2]
I. Syed, Channel shortening equalizers for UWB RAKE receivers:
Design and performance analysis. Saarbrücken: VDM Verlag, 2009.
[3] C. Buchacher, A hybrid equalizer/RAKE receiver: Multipath
interference mitigation for HSDPA in large delay spread channels.
Saarbrücken, Germany: VDM Verlag, 2008.
[4] M. Wang, Y. Han, and W. Sheng, “A Bayesian approach to adaptive
RAKE receiver,” IEEE Access, pp. 1–7, 2017.
[5] C. Ziółkowski and J. M. Kelner, “Estimation of the reception angle
distribution based on the power delay spectrum or profile,” Int. J.
Antennas Propag., vol. 2015, p. e936406, Dec. 2015.
[6] K. I. Pedersen, P. E. Mogensen, and B. H. Fleury, “A stochastic model
of the temporal and azimuthal dispersion seen at the base station in
outdoor propagation environments,” IEEE Trans. Veh. Technol., vol. 49,
no. 2, pp. 437–447, Mar. 2000.
[7] K. I. Pedersen, P. E. Mogensen, and B. H. Fleury, “Power azimuth
spectrum in outdoor environments,” Electron. Lett., vol. 33, no. 18, pp.
1583–1584, Aug. 1997.
[8] C. Ziółkowski and J. M. Kelner, “Empirical models of the azimuthal
reception angle—Part I: Comparative analysis of empirical models for
different propagation environments,” Wirel. Pers. Commun., vol. 91, no.
2, pp. 771–791, Jul. 2016.
[9] C. Ziółkowski and J. M. Kelner, “Empirical models of the azimuthal
reception angle—Part II: Adaptation of the empirical models in
analytical and simulation studies,” Wirel. Pers. Commun., vol. 91, no. 3,
pp. 1285–1296, Dec. 2016.
[10] G. L. Stüber, Principles of mobile communication, 3rd ed. New York,
USA: Springer, 2011.
M. Singh, Study of DS-CDMA RAKE receiver in realistic mobile radio
environment: Case study. S.l.: VDM Verlag Dr. Müller, 2011.
Cezary Ziółkowski, Jan M. Kelner, and Leszek Nowosielski,
“Correlation properties of signal at mobile receiver for different propagation environments,”
in 2018 14th International Conference on Advanced Trends in Radioelectronics,
Telecommunications and Computer Engineering (TCSET),
Lviv - Slavske, Ukraine, 20-24.02.2018., pp. 1–4.
| 7cs.IT
|
Nominal C-Unification
Mauricio Ayala-Rincón1⋆ , Washington de Carvalho-Segundo1, Maribel
Fernández2 and Daniele Nantes-Sobrinho1
arXiv:1709.05384v1 [cs.PL] 15 Sep 2017
1
Depts. de Matemática e Ciência da Computação, Universidade de Brası́lia, Brazil
[email protected], [email protected], [email protected],
[email protected]
2
Department of Informatics, King’s College London, UK
Abstract. Nominal unification is an extension of first-order unification
that takes into account the α-equivalence relation generated by binding
operators, following the nominal approach. We propose a sound and complete procedure for nominal unification with commutative operators, or
nominal C-unification for short, which has been formalised in Coq. The
procedure transforms nominal C-unification problems into simpler (finite families) of fixpoint problems, whose solutions can be generated by
algebraic techniques on combinatorics of permutations.
1
Introduction
Unification, where the goal is to solve equations between first-order terms, is
a key notion in logic programming systems, type inference algorithms, protocol analysis tools, theorem provers, etc. Solutions to unification problems are
represented by substitutions that map variables (X, Y, . . . ) to terms.
When terms include binding operators, a more general notion of unification
is needed: unification modulo α-equivalence. In this paper, we follow the nominal
approach to the specification of binding operators [20,30,26], where the syntax of
terms includes, in addition to variables, also atoms (a, b, . . . ), which can be abstracted, and α-equivalence is axiomatised by means of a freshness relation a#t
and name-swappings (a b). For example, the first-order logic formula ∀a.a ≥ 0
can be written as a nominal term ∀([a]geq(a, 0)), using function symbols ∀ and
geq and an abstracted atom a. Nominal unification [30] is the problem of solving
equations between nominal terms modulo α-equivalence; it is a decidable problem and efficient nominal unification algorithms are available [11,9,24], that compute solutions consisting of freshness contexts (containing freshness constraints
of the form a#X) and substitutions.
In many applications, operators obey equational axioms. Nominal reasoning
and unification have been extended to deal with equational theories presented
by rewrite rules (see, e.g., [18,17,5]) or defined by equational axioms (see, e.g.,
[14,19]). The case of associative and commutative nominal theories was considered in [3], where a parametric {α, AC}-equivalence relation was formalised in
⋆
Work supported by the Brazilian agencies FAPDF (DE 193.001.369/2016), CAPES
(Proc. 88881.132034/2016-01, 2nd author) and CNPq (PQ 307009/2013, 1st author).
2
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
Coq. However, only equational deduction was considered (not unification). In
this paper, we study nominal C-unification.
Contributions: We present a nominal C-unification algorithm, based on a set
of simplification rules, which transforms a given nominal C-unification problem
h∆, Qi, where ∆ is a freshness context and Q a set of freshness constraints and
equations, respectively of the form a#? s and s ≈? t, into a finite set of triples of
the form h∇, σ, P i, consisting of a freshness context ∇, a substitution σ and a
set of fixpoint equations P , of the form π.X ≈? X. The simplifications are based
on the deduction rules for freshness and α-C-equivalence (denoted as ≈{α,C} ).
The role of fixpoint equations in nominal C-unification is tricky: while in standard nominal unification [30], solving a fixpoint equation of the form (a b).X ≈?
X reduces to checking whether the constraints a#X, b#X (a and b fresh in X)
are satisfied, and in this case the solution is the identity substitution, in nominal
C-unification, for ∗ and + commutative operators, one can have additional combinatory solutions of the form {X/a+b}, {X/(a+b)∗. . .∗(a+b)}, {X/f (a)+f (b)},
etc. We show that in general there is no finitary representation of solutions using
only freshness contexts and substitutions, hence a nominal C-unification problem
may have a potentially infinite set of independent most general unifiers (unlike
standard C-unification, which is well-known to be finitary).
We adapt the proof of NP-completeness of syntactic C-unification to show
that nominal C-unification is NP-complete as well.
The simplification rules were formalised to be sound and complete in Coq.
The formalisation and a full version of the present paper with all details of the
proofs are available at http://ayala.mat.unb.br/publications.html.
Related work: To generate the set of combinatorial solutions for fixpoint equations we can use an enumeration procedure given in [4], which is based on the
combinatorics of permutations. By combining the simplification and enumeration methods, we obtain a nominal C-unification procedure in two phases: a
simplification phase, described in this paper, which outputs a finite set of most
general solutions that may include fixpoint constraints, and a generation phase,
which eliminates the fixpoint constraints according to [4].
Several extensions of the nominal unification algorithm have been defined, in
addition to the equational extensions already mentioned.
An algorithm for nominal unification of higher-order expressions with recursive let was proposed in [23]; as for nominal C-unification, fixpoint equations are
obtained in the process. Using the techniques in [4], it is possible to proceed
further and generate the combinatorial solutions of fixpoint equations.
Recently, Aoto and Kikuchi [1] proposed a rule-based procedure for nominal
equivariant unication [13], an extension of nominal unification that is useful in
confluence analysis of nominal rewriting systems [2,16].
Furthermore, several formalisations and implementations of the nominal unification algorithm are available. For example, formalisations of its soundness and
completeness were developed by Urban et al [30,29], Ayala-Rincón et al [6], and
Kumar and Norrish [22] using, respectively, the proof assistants Isabelle/HOL,
PVS and HOL4. An implementation in Maude using term graphs [10] is also
Nominal C-Unification
3
available. Urban and Cheney used a nominal unification algorithm to develop
a Prolog-like language called α-Prolog [12]. Our formalisation of nominal Cunification is based on the formalisation of equivalence modulo {α, AC} presented in [3]. The representations of permutations and terms are similar, but
here we deal also with substitutions and unification rules, and prove soundness
and completeness of the unification algorithm.
Non nominal reasoning modulo equational theories has been subject of formalisations. For instance, in [25], Nipkow presented a set of Isabelle/HOL tactics
for reasoning modulo A, C and AC; Braibant and Pous [8] designed a plugin for
Coq, with an underlying AC-matching algorithm, that extends the system tactic
rewrite to deal with AC function symbols; also, Contejean [15] formalised in
Coq the correction of an AC-matching algorithm implemented in CiME.
Syntactic unification with commutative operators is an NP-complete problem
and its solutions can be finitely generated [21,28]. Since C-unification problems
are a particular case of nominal C-unification problems, our simplification algorithm, checked in Coq, is also a formalisation of the C-unification algorithm.
Organisation: Section 2 presents basic concepts and notations. Section 3 introduces the formalised equational and freshness inference rules for nominal
C-unification, and also discusses NP-completeness; Section 4 shows how single
fixpoint equations are solved giving rise to infinite independent solutions and
also we briefly explain how fixpoint solutions are combined in [4] in order to
produce solutions. Section 5 concludes and describes future work.
2
Background
Consider countable disjoint sets of variables X := {X, Y, Z, · · · } and atoms A :=
{a, b, c, · · · }. A permutation π is a bijection on A with a finite domain, where
the domain (i.e., the support ) of π is the set dom(π) := {a ∈ A | π · a 6= a}.
The inverse of π is denoted by π −1 . Permutations can be represented by lists of
swappings, which are pairs of different atoms (a b); hence a permutation π is a
finite list of the form (a1 b1 ) :: . . . :: (an bn ) :: nil, where nil denotes the identity
permutation; concatenation is denoted by ⊕ and, when no confusion may arise,
:: and nil are omitted. We follow Gabbay’s permutative convention: Atoms differ
on their names, so for atoms a and b the expression a 6= b is redundant. Also,
(a b) and (b a) represent the same swapping.
We will assume as in [3] countable sets of function symbols with different
equational properties such as associativity, commutativity, idempotence, etc.
Function symbols have superscripts that indicate their equational properties;
thus, fkC will denote the k th function symbol that is commutative and fj∅ the
j th function symbol without any equational property.
Nominal terms are generated by the following grammar:
s, t := hi | ā | [a]t | hs, ti | fkE t | π.X
hi denotes the unit (that is the empty tuple), ā denotes an atom term, [a]t
denotes an abstraction of the atom a over the term t, hs, ti denotes a pair,
fkE t the application of fkE to t and, π.X a moderated variable or suspension.
Suspensions of the form id.X will be represented just by X.
4
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
The set of variables occurring in a term t will be denoted as S
V ar(t). This
notation extends to a set S of terms in the natural way: V ar(S) = t∈S V ar(t).
As usual, | | will be used to denote the cardinality of sets as well as to denote
the size or number of symbols occurring in a given term.
Definition 1 (Permutation action). The action of a permutation on atoms
is defined as: nil · a := a; (b c) :: π · a := π · a; and, (b c) :: π · b := π · c. The action
of a permutation on terms is defined recursively as:
π · hi := hi
π · hu, vi := hπ · u, π · vi
π · fkE t
:= fkE (π · t)
′
π · a := π · a
π · ([a]t) := [π · a](π · t)
π · (π . X) := (π ′ ⊕ π) . X
Notice that according to the definition of the action of a permutation over
atoms, the composition of permutations π and π ′ , usually denoted as π ◦ π ′ ,
corresponds to the append π ′ ⊕ π. Also notice that π ′ ⊕ π · t = π · (π ′ · t). The
difference set between two permutations π and π ′ is the set of atoms where the
action of π and π ′ differs: ds(π, π ′ ) := {a ∈ A | π · a 6= π ′ · a}.
A substitution σ is a mapping from variables to terms such that its domain, dom(σ) := {X | X 6= Xσ}, is finite. For X ∈ dom(σ), Xσ is called
the image of X. Define the image of σ as im(σ) := {Xσ | X ∈ dom(σ)}. Let
dom(σ) = {X1 , · · · , Xn }, then σ can be represented as a set of bindings in the
form {X1 /t1 , · · · , Xn /tn }, where Xi σ = ti , for 1 ≤ i ≤ n.
Definition 2 (Substitution action). The action of a substitution σ on a term
t, denoted tσ, is defined recursively as follows:
aσ
:= a
(fkE t)σ := fkE tσ
hiσ
:= hi
hs, tiσ := hsσ, tσi
([a]t)σ := [a]tσ
(π.X)σ := π · Xσ
The following result can be proved by induction on the structure of terms.
Lemma 1 (Substitutions and Permutations Commute). (π ·t)σ = π ·(tσ)
The inference rules defining freshness and α-equivalence are given in Fig. 1
and 2. The symbols ∇ and ∆ are used to denote freshness contexts that are sets of
constraints of the form a#X, meaning that the atom a is fresh in X. The domain
of a freshness context dom(∆) is the set of atoms appearing in it; ∆|X denotes
the restriction of ∆ to the freshness constraints on X: {a#X | a#X ∈ ∆}. The
rules in Fig. 1 are used to check if an atom a is fresh in a nominal term t under
a freshness context ∇, also denoted as ∇ ⊢ a#t. The rules in Fig. 2 are used to
check if two nominal terms s and t are α-equivalent under some freshness context
∇, written as ∇ ⊢ s ≈α t. These rules use the inference system for freshness
constraints: specifically freshness constraints are used in rule (≈α [ab]).
Example 1. Let σ = {X/[a]a}. Verify that h(a b).X, f (e)iσ ≈α hX, f (e)iσ.
By dom(π)#X and ds(π, π ′ )#X we abbreviate the sets {a#X | a ∈ dom(π)}
and {a#X | a ∈ ds(π, π ′ )}, respectively.
Key properties of the nominal freshness and α-equivalence relations have
been extensively explored in previous works [3,6,29,30].
Nominal C-Unification
∇ ⊢ a # hi
(#hi)
∇ ⊢ a#t
∇ ⊢ a # [b]t
∇ ⊢ a#b
(# a[b])
(# atom)
∇ ⊢ a#t
∇ ⊢ a # fkE t
(π −1 · a#X) ∈ ∇
∇ ⊢ a # π.X
(# var)
5
(# app) ∇ ⊢ a # [a]t (# a[a])
∇ ⊢ a#s ∇ ⊢ a#t
∇ ⊢ a # hs, ti
(# pair)
Fig. 1. Rules for the freshness relation
∇ ⊢ hi ≈α hi
(≈α hi)
∇ ⊢ s ≈α t
∇ ⊢ [a]s ≈α [a]t
∇ ⊢ a ≈α a
(≈α [aa])
ds(π, π ′ )#X ⊆ ∇
′
∇ ⊢ π.X ≈α π .X
(≈α atom)
∇ ⊢ s ≈α t
∇ ⊢ fkE s ≈α fkE t
∇ ⊢ s ≈α (a b) · t ∇ ⊢ a # t
(≈α var)
∇ ⊢ [a]s ≈α [b]t
∇ ⊢ s0 ≈α t0 ∇ ⊢ s1 ≈α t1
∇ ⊢ hs0 , s1 i ≈α ht0 , t1 i
(≈α app)
(≈α [ab])
(≈α pair)
Fig. 2. Rules for the relation ≈α
2.1
The relation ≈{α,C} as an extension of ≈α
In [3], the relation ≈α was extended to deal with associative and commutative
theories. Here we will consider α-equivalence modulo commutativity, denoted as
≈{α,C} . This means that some function symbols in our syntax are commutative,
and therefore the rule for function application (≈α app) in Fig. 2 should be
replaced by the rules in Fig. 3.
∇ ⊢ s ≈{α,C} t
∇ ⊢ fkE s ≈{α,C} fkE t
, E 6= C or both s and t are not pairs (≈{α,C} app)
∇ ⊢ s0 ≈{α,C} ti , ∇ ⊢ s1 ≈{α,C} t(i+1) mod 2
∇ ⊢ fkC hs0 , s1 i ≈{α,C} fkC ht0 , t1 i
, i = 0, 1 (≈{α,C} C)
Fig. 3. Additional rules for {α, C}-equivalence
The following properties for ≈{α,C} were formalised as simple adaptations of
the formalisations given in [3] for ≈α .
Lemma 2 (Inversion). The inference rules of ≈{α,C} are invertible.
This means, for instance, that for rules (≈α [ab]) one has ∇ ⊢ [a]s ≈{α,C}
[b]t implies ∇ ⊢ s ≈{α,C} (a b) · t and ∇ ⊢ a # t; and for (≈{α,C} app), ∇ ⊢
fkC hs0 , s1 i ≈{α,C} fkC ht0 , t1 i implies ∇ ⊢ s0 ≈{α,C} t0 and ∇ ⊢ s1 ≈{α,C} t1 , or
∇ ⊢ s0 ≈{α,C} t1 and ∇ ⊢ s1 ≈{α,C} t0 .
Lemma 3 (Freshness preservation). If ∇ ⊢ a # s and ∇ ⊢ s ≈{α,C} t then
∇ ⊢ a # t.
Lemma 4 (Intermediate transitivity for ≈{α,C} with ≈α ). If ∇ ⊢ s ≈{α,C}
t and ∇ ⊢ t ≈α u then ∇ ⊢ s ≈{α,C} u.
Lemma 5 (Equivariance). ∇ ⊢ π · s ≈{α,C} π · t whenever ∇ ⊢ s ≈{α,C} t.
6
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
Lemma 6 (Equivalence).
⊢ ≈{α,C}
is an equivalence relation.
Remark 1. According to the grammar for nominal terms, function symbols have
no fixed arity: any function symbol can apply to any term. Despite this, in the
syntax of our Coq formalisation commutative symbols apply only to tuples.
3
A nominal C-unification algorithm
Inference rules are given that transform a nominal C-unification problem into
a finite family of problems that consist exclusively of fixpoint equations of the
form π.X ≈? X, together with a substitution and a set of freshness constraints.
Definition 3 (Unification problem). A unification problem is a pair h∇, P i,
where ∇ is a freshness context and P is a finite set of equations and freshness
constraints of the form s ≈? t and a#? s, respectively, where ≈? is symmetric, s
and t are terms and a is an atom. Nominal terms in the equations preserve the
syntactic restriction that commutative symbols are only applied to tuples.
Equations of the form π.X ≈? X are called fixpoint equations. Given h∇, P i,
by P≈ , P# , Pfp≈ and Pnfp≈ we will resp. denote the sets of equations, freshness
constraints, fixpoint equations and non fixpoint equations in the set P .
Example 2. Given the nominal unification problem P = h∅, {[a][b]X ≈? [b][a]X}i,
the standard unification algorithm [30] reduces it to h∅, {X ≈? (a b).X}i, which
gives the solution h{a#X, b#X}, idi. However, we will see that infinite independent solutions are feasible when there is at least a commutative operator.
We design a nominal C-unification algorithm using one set of transformation rules to deal with equations (Fig. 4) and another set of rules to deal with
freshness constraints and contexts (Fig. 5). These rules act over triples of the
form h∇, σ, P i, where σ is a substitution. The triple that will be associated by
default with a unification problem h∇, P i is h∇, id, P i. We will use calligraphic
uppercase letters (e.g., P, Q, R, etc) to denote triples.
Remark 2. Let ∇ and ∇′ be freshness contexts and σ and σ ′ be substitutions.
– ∇′ ⊢ ∇σ denotes that ∇′ ⊢ a # Xσ holds for each (a#X) ∈ ∇, and
– ∇ ⊢ σ ≈ σ ′ that ∇ ⊢ Xσ ≈{α,C} Xσ ′ for all X (in dom(σ) ∪ dom(σ ′ )).
Definition 4 (Solution for a triple or problem). A solution for a triple
P = h∆, δ, P i is a pair h∇, σi, where the following conditions are satisfied:
1. ∇ ⊢ ∆σ;
3. ∇ ⊢ sσ ≈{α,C} tσ if s ≈? t ∈ P ;
2. ∇ ⊢ a # tσ if a#? t ∈ P ; 4. there is a substitution λ such that ∇ ⊢ δλ ≈ σ.
A solution for a unification problem h∆, P i is a solution for the associated
triple h∆, id, P i. The solution set for a problem or triple P is denoted by UC (P).
Definition 5 (More general solution and complete set of solutions). For
h∇, σi and h∇′ , σ ′ i in UC (P), we say that h∇, σi is more general than h∇′ , σ ′ i,
denoted h∇, σi 4 h∇′ , σ ′ i, if there exists a substitution λ satisfying ∇′ ⊢ σλ ≈ σ ′
and ∇′ ⊢ ∇λ. A subset V of UC (P) is said to be a complete set of solutions of P
if for all h∇′ , σ ′ i ∈ UC (P), there exists h∇, σi in V such that h∇, σi 4 h∇′ , σ ′ i.
Nominal C-Unification
7
We will denote the set of variables occurring in the set P of a problem h∇, P i
or triple P = h∇, σ, P i as V ar(P ). We also will write V ar(P) to denote this set.
h∇, σ, P ⊎ {s ≈? s}i
h∇, σ, P i
(≈? refl)
h∇, σ, P ⊎ {hs1 , t1 i ≈? hs2 , t2 i}i
h∇, σ, P ∪ {s1 ≈? s2 , t1 ≈? t2 }i
(≈? pair)
h∇, σ, P ⊎ {fkE s ≈? fkE t}i
, if E 6= C (≈? app)
h∇, σ, P ∪ {s ≈? t}i
h∇, σ, P ⊎ {fkC s ≈? fkC t}i
where s = hs0 , s1 i and t = ht0 , t1 i
(≈? C)
,
v = hti , t(i+1) mod 2 i, i = 0, 1
h∇, σ, P ∪ {s ≈? v}i
h∇, σ, P ⊎ {[a]s ≈? [a]t}i
h∇, σ, P ∪ {s ≈? t}i
(≈? [aa])
h∇, σ, P ⊎ {[a]s ≈? [b]t}i
h∇, σ, P ∪ {s ≈? (a b) t, a#? t}i
(≈? [ab])
h∇, σ, P ⊎ {π.X ≈? t}i let σ ′ := σ{X/π −1 · t}
/ V ar(t) (≈? inst)
*
+ , if X ∈
S
{a#? Y σ ′ }
′
−1
∇, σ , P {X/π · t} ∪ Y ∈dom(σ ′ ),
a#Y ∈∇
h∇, σ, P ⊎ {π.X ≈? π ′ .X}i
h∇, σ, P ∪ {π ⊕ (π ′ )−1 .X ≈? X}i
, if π ′ 6= id (≈? inv)
Fig. 4. Reduction rules for equational problems
The unification algorithm proceeds by simplification. Derivation with rules
of Figs. 4 and 5 is respectively denoted by ⇒≈ and ⇒# . Thus, h∇, σ, P i ⇒≈
h∇, σ ′ , P ′ i means that the second triple is obtained from the first one by application of one rule. We will use the standard rewriting nomenclature, e.g., we will
say that P is a normal form or irreducible by ⇒≈ , denoted by ⇒≈ -nf, whenever
there is no Q such that P ⇒≈ Q; ⇒∗≈ and ⇒+
≈ denote respectively derivations
in zero or more and one or more applications of the rules in Fig. 4.
The only rule that can generate branches is (≈? C), which is an abbreviation
for two rules providing the different forms in which one can relate the arguments
s and t in an equation fkC s ≈? fkC t for a commutative function symbol (s, t are
tuples, by the syntactic restriction in Definition 3): either hs0 , s1 i ≈? ht0 , t1 i or
hs0 , s1 i ≈? ht1 , t0 i.
The syntactic restriction on arguments of commutative symbols being only
tuples, is not crucial since any equation of the form fkC π.X ≈? t can be translated into an equation of form fkC hπ.X1 , π.X2 i ≈? t, where X1 and X2 are new
variables and ∇ is extended to ∇′ in such a way that both X1 and X2 inherit
all freshness constraints of X in ∇: ∇′ = ∇ ∪ {a#Xi | i = 1, 2, and a#X ∈ ∇}.
In the rule (≈? inst) the inclusion of new constraints in the problem, given in
[
{a#? Y σ′ } is necessary to guarantee that the new substitution σ ′ is
′
Y ∈dom(σ ),
compatible with the freshness context ∇.
a#Y ∈∇
Example 3. Let ∗3 be a commutative function symbol. Below, we show how the
problem P = h∅, {[e](a b).X ∗ Y ≈? [f ](a c)(c d).X ∗ Y }i reduces (via rules in
3
Infix notation is adopted for commutative symbols: s ∗ t abbreviates ∗hs, ti.
8
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
h∇, σ, P ⊎ {a#? hi}i
h∇, σ, P i
h∇, σ, P ⊎ {a#? f t}i
h∇, σ, P ∪ {a#? t}i
h∇, σ, P ⊎ {a#? [b]t}i
h∇, σ, P ∪ {a#? t}i
(#? hi)
(#? app)
(#? a[b])
h∇, σ, P ⊎ {a#? b̄}i
h∇, σ, P i
(#? ab̄)
h∇, σ, P ⊎ {a#? [a]t}i
h∇, σ, P i
(#? a[a])
h∇, σ, P ⊎ {a#? π.X}i
h{(π −1 · a)#X} ∪ ∇, σ, P i
h∇, σ, P ⊎ {a#? hs, ti}i
h∇, σ, P ∪ {a#? s, a#? t}i
(#? var)
(#? pair)
Fig. 5. Reduction rules for freshness problems
Figs. 4 and 5). Application of rule (≈? C) gives two branches which reduce
into two fixpoint problems: Q1 and Q2 . Highlighted terms show where the rules
are applied. For brevity, let π1 = (a c)(c d)(e f ), π2 = (a b)(e f )(c d)(a c), π3 =
(a c)(c d)(e f )(a b) and, σ = {X/(e f )(a b).Y }.
h∅, id, { [e](a b).X ∗ Y ≈? [f ](a c)(c d).X ∗ Y }i
⇒(≈? [ab])
h∅, id, { (a b).X ∗ Y ≈? π1 .X ∗ (e f ).Y , e#? (a c)(c d).X ∗ Y }i ⇒(≈? C)
branch 1:
h∅, id, { (a b).X ≈? π1 .X , Y ≈? (e f ).Y , e#? (a c)(c d).X ∗ Y }i
⇒(≈? inv)(2×) h∅, id, {(a b)[π1 ]−1 .X ≈? X, [(e f )]−1 .Y ≈? Y, e#? (a c)(c d).X ∗ Y }i
⇒ (#? app),
h∅, id, {π2 .X ≈? X, (e f ).Y ≈? Y, e#? (a c)(c d).X , e#? Y }i
(#? pair)
⇒(#? var)(2×) h{e#X, e#Y }, id, {π2 .X ≈? X, (e f ).Y ≈? Y }i = Q1
branch 2:
h∅, id, { (a b).X ≈? (e f ).Y , Y ≈? π1 .X, e#? (a c)(c d).X ∗ Y }i
⇒(≈? inst)
h∅, σ, { Y ≈? (a c)(c d)(e f )(e f )[(a b)]−1 .Y , e#? π1 [(a b)]−1 .Y ∗ Y }i
⇒(≈? inv)
h∅, σ, {[(a c)(c d)(a b)]−1 .Y ≈? Y, e#? π3 .Y ∗ Y }i
⇒ (#? app),
h∅, σ, {(a b)(c d)(a c).Y ≈? Y, e#? π3 .Y , e#? Y }i
(#? pair)
⇒(#? var)(2×) h{e#Y, f #Y }, σ, {(a b)(c d)(a c).Y ≈? Y }i = Q2
Definition 6 (Set of ⇒≈ and ⇒# -normal forms). We denote by P⇒≈ (resp.
P⇒# ) the set of normal forms of P with respect to ⇒≈ (resp. ⇒# ).
Definition 7 (Fail and success for ⇒≈ ). Let P be a triple, such that the
rules in Fig. 4 give rise to a normal form h∇, σ, P i. The rules in Fig. 4 are
said to fail if P contains non fixpoint equations. Otherwise h∇, σ, P i is called
a successful triple regarding ⇒≈ (i.e., in a successful triple, P consists only of
fixpoint equations and, possibly, freshness constraints).
The rules in Fig. 5 will only be applied to successful triples regarding ⇒≈ .
Definition 8 (Fail and success for ⇒# ). Let Q = h∇, σ, Qi be a successful
triple regarding ⇒≈ , and Q′ = h∇′ , σ, Q′ i its normal form via rules in Fig. 5,
that is Q ⇒∗# Q′ and Q′ is in Q⇒# . If Q′ contains freshness constraints it is
said that ⇒# fails for Q; otherwise, Q′ will be called a successful triple for ⇒# .
Nominal C-Unification
9
Remark 3. Since in a successful triple regarding ⇒≈ , Q, one has only fixpoint
equations and ⇒# acts only over freshness constraints, Q′ in the definition above
contains only fixpoint equations and freshness constraints. Also, by a simple case
analysis on t one can check that any triple with freshness constraints a#? t is
reducible by ⇒# , except when t ≡ ā. Hence the freshness constraints in Q′ would
be only of the form a#? ā.
The relation ⇒≈ , starts from a triple with the identity substitution and
always maintains a triple h∇, σ ′ , P ′ i in which the substitution σ ′ does not affect
the current problem P ′ . The same happens for ⇒# since the substitution does
not change with this relation. This motivates the next definition and lemma.
Definition 9 (Valid triple). P = h∇, σ, P i is valid if im(σ) ∩ dom(σ) = ∅ and
dom(σ) ∩ V ar(P ) = ∅.
Remark 4. A substitution σ in a valid triple P is idempotent, that is, σσ = σ.
Lemma 7 is proved by case analysis on the rules used by ⇒≈ and ⇒# .
Lemma 7 (Preservation of valid triples). If P = h∇, σ, P i is valid and
P ⇒≈ ∪ ⇒# P ′ = h∇′ , σ ′ , P ′ i, then P ′ is also valid.
From now on, we consider only valid triples.
Lemma 8 (Termination of ⇒≈ and ⇒# ). There is no infinite chain of reductions ⇒≈ (or ⇒# ) starting from an arbitrary triple P = h∇, σ, P i.
Proof. – The proof for ⇒≈ is by well-founded induction on P using the measure
kPk
k, |Pnfp≈ |i with a lexicographic ordering, where kP k =
P = h|V ar(P≈ )|, kPP
s≈? t ∈ P≈ |s| + |t| +
a#? u∈P# |u|. Note that this measure decreases after
each step h∇, σ, P i ⇒≈ h∇, σ ′ , P ′ i: for (≈? inst), |V ar(P≈ )| > |V ar(P≈′ )|; for
(≈? refl), (≈? pair), (≈? app), (≈? [aa]), (≈? [ab]) and (≈? C), |V ar(P≈ )| ≥
|V ar(P≈′ )|, but kP k > kP ′ k; and, for (≈? inv), both |V ar(P≈ )| = |V ar(P ′ ≈)|
′
|.
and kP k = kP ′ k, but |Pnfp≈ | > |Pnfp
≈
– The proof for ⇒# is by induction on P using as measure kP# k. It can be
checked that this measure decreases after each step: h∇, σ, P i ⇒# h∇, σ ′ , P ′ i.
To solve a unification problem, h∇, P i, one builds the derivation tree for
⇒≈ , labelling the root node with h∇, id, P i. This tree has leaves labelled with
⇒≈ -nf’s that are either failing or successful triples. Then, the tree is extended
by building ⇒# -derivations starting from all successful leaves. The extended
tree will include failing leaves and successful leaves. The successful leaves will be
labelled by triples P ′ in which the problem P ′ consists only of fixpoint equations.
Since ⇒≈ and ⇒# are both terminating (Lemma 8), the process described above
must be also terminating.
Definition 10 (Derivation tree for h∇, P i). A derivation tree for the unification problem h∇, P i, denoted as Th∇,P i , is a tree with root label P = h∇, id, P i
built in two stages:
10
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
– Initially, a tree is built, whose branches end in leaf nodes labelled with the
triples in P⇒≈ . The labels in each path from the root to a leaf correspond to
a ⇒≈ -derivation.
– Further, for each leaf labelled with a successful triple Q in P⇒≈ , the tree is
extended with a path to a new leaf that is labelled with a Q̄ ∈ Q⇒# . The labels
in the extended path from the node with label Q to the new leaf correspond to
a ⇒# -derivation.
Remark 5. For h∇, P i, all labels in the nodes of Th∇,P i are valid by Lemma 7.
The next lemma is proved by case analysis on elements of P⇒≈ and P⇒# .
Lemma 9 (Characterisation of leaves of Th∇,P i ). Let h∇, P i be a unification problem. If P ′ = h∇′ , σ ′ , P ′ i is the label of a leaf in Th∇,P i , then P ′ can be
partitioned as follows: P ′ = P ′′ ∪P⊥ , where P ′′ is the set of all fixpoint equations
in P ′ and P⊥ = P ′ − P ′′ . If P⊥ 6= ∅ then UC (P ′ ) = ∅.
The next definition is motivated by the previous characterisation of the labels
of leaves in derivation trees.
Definition 11 (Successful leaves). Let h∇, P i be a unification problem. A
leaf in Th∇,P i that is labelled with a triple of the form Q = h∇′ , σ ′ , Qi, where Q
consists only of fixpoint equations, is called a successful leaf of Th∇,P i . In this
case Q is called a successful triple of Th∇,P i . The sets of successful leaves and
triples of Th∇,P i are denoted respectively by SL(Th∇,P i ) and ST (Th∇,P i ).
The soundness theorem states that successful leaves of Th∇,P i produce correct
solutions. The proof is by induction on the number of steps of ⇒≈ and ⇒# and
uses Lemma 9 and auxiliary results on the preservation of solutions by ⇒≈ and
⇒# . Proving preservation of solutions for rules (≈? [ab]) and (≈? inst) is not
straightforward and uses Lemmas 1 2, 3 and 5 to check that the four conditions
of Def. 4 are valid before, if one supposes their validity after the rule application.
Theorem 1 (Soundness of Th∇,P i ). Th∇,P i is correct, i.e., if P ′ = h∇′ , σ, P ′ i
is the label of a leaf in Th∇,P i , then 1. UC (P ′ ) ⊆ UC (h∇, id, P i), and 2. if P ′
contains non fixpoint equations or freshness constraints then UC (P ′ ) = ∅.
The completeness theorem guarantees that the set of successful triples provides a complete set of solutions. Its proof uses case analysis on the rules of the
relations ⇒≈ and ⇒# by an argumentation similar to the one used for Theorem
1. For ⇒# one has indeed equivalence: P ⇒# P ′ , implies UC (P) = UC (P ′ ). The
same is true for all rules of the relation ⇒≈ except the branching rule (≈? C),
for which it is necessary to prove that all solutions of a triple reduced by (≈? C)
must belong to the set of solutions of one of its sibling triples.
Theorem 2 (Completeness of Th∇,P i ). Let h∇, P i and
S Th∇,P i be a unification
problem and its derivation tree. Then UC (h∇, id, P i) = Q∈ST (Th∇,P i ) UC (Q).
Nominal C-Unification
11
Corollary 1 (Generality of successful triples). Let P = h∇, P i be a unification problem and h∆′ , σ ′ i ∈ UC (P). Then there exists a successful triple
Q ∈ ST (Th∇,P i ) where Q = h∆, σ, Qi such that h∆′ , σ ′ i ∈ UC (Q), and hence,
∆′ ⊢ ∆σ ′ and there exists λ such that ∆′ ⊢ σλ ≈ σ ′ .
S
Proof. By Theorem 2, UC (P) = P ′ ∈ST (Th∇,P i ) UC (P ′ ). Then there exists Q ∈
ST (Th∇,P i ) such that h∆′ , σ ′ i ∈ UC (Q). Suppose Q = h∆, σ, Qi. Then by the
first and fourth conditions of the definition of solution (Def. 4) we have that
∆′ ⊢ ∆σ ′ and there exists λ such that ∆′ ⊢ σλ ≈ σ ′ .
Remark 6. The nominal C-unification problem is to decide, for a given P, if
UC (P) is non empty; that is, whether P has nominal C-unifiers. To prove that
this problem is in NP, a non-deterministic procedure using the reduction rules in
the same order as in Definition 10 is designed. In this procedure, whenever rule
(≈? C) applies, only one of the two possible branches is guessed. In this manner,
if the derivation tree has a successful leaf, this procedure will guess a path to
the successful leaf, answering positively to the decision problem. According to
the measures used in the proof of termination Lemma 8, reduction with both
the relations ⇒≈ and ⇒# is polynomially bound, which implies that this nondeterministic procedure is polynomially bound.
To prove NP-completeness, one can polynomially reduce the well-known NPcomplete positive 1-in-3-SAT problem into nominal C-unification, as done in [7]
for the C-unification problem. An instance of the positive 1-in-3-SAT problem
consists of a set of clauses C = {Ci |1 ≤ i ≤ n}, where each Ci is a disjunction of
three propositional variables, say Ci = pi ∨ qi ∨ ri . A solution of C is a valuation
whit exactly one variable true in each clause. The proposed reduction of C into
a nominal C-unification problem would require just a commutative function
symbol, say ⊕, two atoms, say a and b, a variable for each clause Ci , say Yi ,
and a variable for each propositional variable p in C, say Xp . Instantiating Xp
as a or b, would be interpreted as evaluating p as true or false, respectively.
Each clause Ci = pi ∨ qi ∨ ri in C is translated into an equation Ei of the form
((Xpi ⊕ Xqi ) ⊕ Xri ) ⊕ Yi ≈? ((b ⊕ b) ⊕ a) ⊕ ((b ⊕ a) ⊕ b). The nominal C-unification
problem for C is given by PC = h∅, {Ei |1 ≤ i ≤ n}i. Simplifying PC would not
introduce freshness constraints since the problem does not include abstractions.
Thus, to conclude it is only necessary to check that h∅, σi is a solution for PC if
and only if σ instantiates exactly one of the variables Xpi , Xqi and Xri in each
equation with a and the other two with b, which means that C has a solution.
4
Generation of solutions for successful leaves of Th∇,P i
To build solutions for a successful leaf P = h∇, σ, P i in the derivation tree
of a given unification problem, we will select and combine solutions generated
for fixpoint equations π.X ≈? X, for each X ∈ V ar(P ). We introduce the
notion of pseudo-cycle of a permutation, in order to provide precise conditions
to build terms t by combining the atoms in dom(π), such that π · t ≈{α,C} t. For
convenience, we use the algebraic cycle representation of permutations. Thus,
12
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
instead of sequences of swappings, permutations in nominal terms will be read
as products of disjoint cycles [27].
Example 4. (Continuing Example 3) The permutations (a b)::(e f )::(c d)::(a c)::
nil and (a b) :: (c d) :: (a c) :: nil are respectively represented as the product of
permutation cycles (a b c d)(e f ) and (a b c d)(e)(f ).
Permutation cycles of length one are omitted. In general the cyclic representation of a permutation consists of the product of all its cycles.
Let π be a permutation with dom(π) = n. Given a ∈ dom(π) the elements of
the sequence a, π(a), π 2 (a), . . . cannot be all distinct. Taking the first k ≤ n, such
that π k (a) = a, we have the k-cycle (a π(a) . . . π k−1 (a)), where π j+1 (a) is the
successor of π j (a). For the 4-cycle in the permutation (a b c d) (e f ), the 4-cycles
generated by a, b, c and d are the same: (a b c d) = (b c d a) = (c d a b) = (d a b c).
Def. 12 establishes the notion of a pseudo-cycle w.r.t. a k-cycle κ. Intuitively,
given a k-cycle κ and a commutative function symbol ∗, a pseudo-cycle w.r.t κ,
(A0 . . . Al ), is a cycle whose elements are either atom terms built from the atoms
in κ or terms of the form A′i ∗ A′j , for A′i , A′j elements of a pseudo-cycle w.r.t κ.
Definition 12 (Pseudo-cycle). Let κ = (a0 a1 . . . ak−1 ) be a k-cycle of a
permutation π. A pseudo-cycle w.r.t. κ is inductively defined as follows:
1. κ = (a0 · · · ak−1 ) is a pseudo-cycle w.r.t. κ, called trivial pseudo-cycle of κ.
2. κ′ = (A0 ... Ak′ −1 ) is a pseudo-cycle w.r.t. κ, if the following conditions are
simultaneously satisfied:
(a) each element of κ′ is of the form Bi ∗ Bj , where ∗ is a commutative
function symbol in the signature, and Bi , Bj are different elements of
κ′′ , a pseudo-cycle w.r.t. κ. κ′ will be called a first-instance pseudo-cycle
of κ′′ w.r.t. κ.
(b) Ai 6≈α,C Aj for i 6= j, 0 ≤ i, j ≤ k ′ − 1;
(c) for each 0 ≤ i < k ′ − 1, κ · Ai ≈{α,C} A(i+1)mod k′ .
The length of the pseudo-cycle κ, denoted by |κ|, consists of the number of
elements in κ. A pseudo-cycle of length one will be called unitary.
Example 5. A (Continuing Example 2) Considering ∗ a commutative function
symbol in the signature, the unitary pseudo-cycles of κ = (a b) generate infinite independent solutions for the leaf h∅, id, {X ≈? (a b).X}i. Examples of
these solutions are: h∅, {X/a∗b}i, h∅, {X/(a∗a)∗(b∗b)}i, h∅, {X/(a∗b)∗(a∗b)}i,
h∅, {X/((a ∗ a) ∗ a) ∗ ((b ∗ b) ∗ b)}i, h∅, {X/(a ∗ (a ∗ a)) ∗ (b ∗ (b ∗ b))}i, etc.
B (Continuing Examples 3 and 4) In Q1 and Q2 we have the occurrences of
the 4-cycle κ = (a b c d). Suppose ∗, ⊕, + are commutative operators in the
signature. The following are pseudo-cycles w.r.t. κ: κ = (a b c d); κ1 =
((a ∗ b) (b ∗ c) (c ∗ d) (d ∗ a)); κ2 = ((a ⊕ c) (b ⊕ d)); κ11 = (((a ∗ b) + (b ∗ c)) ((b ∗
c)+(c∗d))((c∗d)+(d∗a)) ((d∗a)+(a∗b))); κ12 = (((a∗b)∗(c∗d)) ((b∗c)∗(d∗a)));
κ21 = (((a ⊕ c) ∗ (b ⊕ d))); κ121 = (((a ∗ b) ∗ (c ∗ d)) ∗ ((b ∗ c) ∗ (d ∗ a))). κ1 and
κ2 are first-instance pseudo-cycles of κ, and κ11 and κ12 of κ1 and κ21 of κ2 .
Nominal C-Unification
13
Notice that, |κ| = |κ1 | = |κ11 | = 4, |κ12 | = 2, and |κ21 | = |κ121 | = 1. Also, κ1
corresponds to ((a ∗ d) (b ∗ a) (c ∗ b) (d ∗ c)), a first-instance pseudo-cycle of κ.
Finally, observe that for the elements of the unitary pseudo-cycles κ21 and
κ121 , say s = (a ⊕ c) ∗ (b ⊕ d) and t = ((a ∗ b) ∗ (c ∗ d)) ∗ ((b ∗ c) ∗ (d ∗ a)), {X/s}
and {X/t} (resp. {Y /s} and {Y /t}) are solutions of the fixpoint equation
(a b c d)(e f ).X ≈? X (resp. (a b c d).Y ≈? Y ).
Let κ be a pseudo-cycle. Notice that only item 2 of Def. 12 may build a
first-instance pseudo-cycle κ′ w.r.t. κ with fewer elements. If |κ′ | < |κ| then,
due algebraic properties of cycles and commutativity of the operator applied
(∗), one must have that |κ′ | = |κ|/2. Thus, unitary pseudo-cycles can only be
generated from cycles of length a power of two. This is the intuition behind the
next theorem, proved by induction on the size of the cycle κ.
Theorem 3. A pseudo-cycle κ generates unitary pseudo-cycles iff |κ| is a power
of two.
Notice that, according item 2.c of Def. 12, if κ′ = (A0 . . . Ak′ −1 ) is a pseudocycle w.r.t. π then π · Ak′ −1 ≈{α,C} A0 ; particularly, if k ′ = 1 then π · A0 ≈{α,C}
A0 . Below, given P = h∅, {π.X ≈? X}i a fixpoint equational problem, we call
a combinatory solution of P, a substitution {X/t}, such that π · t ≈C t, and t
contains only atoms from π and commutative function symbols, built as unary
pseudo-cycles w.r.t. κ a cycle in π.
The next theorem is proved by contradiction, supposing that κ has an odd
factor and using Theorem 3.
Theorem 4. Let P = h∅, {π.X ≈? X}i be a fixpoint problem. P has a combinatory solution iff there exists a unitary pseudo-cycle κ w.r.t. π.
Remark 7. Since one can generate infinitely many unitary pseudo-cycles from a
given 2n -cycle κ in π, n ∈ N, there exist infinite independent solutions for the
fixpoint problem h∅, {π.X ≈? X}i.
General solutions for fixpoint problems. To compute the set of solutions for a
fixpoint equation, we use a method described in [4], which is based on the computation of unitary extended pseudo-cycles (epc). We refer to [4] for the definition
of extended pseudo-cycles and an algorithm to enumerate all the solutions of a
successful leaf in the derivation tree.
Pseudo-cycles are built just from atom terms in dom(π) and commutative
function symbols, while epc’s consider all nominal syntactic elements including
new variables, and also non commutative function symbols. The soundness and
completeness of the generator of solutions described in [4] relies on the properties
of pseudo-cycles described above, in particular the fact that only unitary pseudocycles generate solutions.
Example 6. A (Continuing Example 5 A) General solutions for the fixpoint
problem h∅, {(a b).X ≈? X}i include not only those generated by pseudocycles, but also any other generated by unitary epc’s from (a b), which include other syntactic elements to the pseudo-cycle construction; for instance:
14
M. Ayala-Rincón, W. de Carvalho, M. Fernández, D. Nantes-Sobrinho
h∅, {X/h a∗h b}i, h{a, b#Y, {X/(ha, Y i∗ha, Y i)⊕(hb, Y i∗hb, Y i)}i, h∅, {X/([i]a∗
[i]b) ⊕ ([i]a∗ [i]b)}i, h{a, b#Z}, {X/(hZ, ai ∗ (a ∗ a)) ⊕ (hZ, bi ∗ (b ∗ b))}i and
h∅, {X/(([b]a ∗ [b]a) ∗ b) ⊕ (([a]b ∗ [a]b) ∗ a)}i.
B (Continuing Example 5 B) Notice that constraints e#X and e#Y in Q1 =
h{e#X, e#Y }, id, {(a b c d)(e f ).X ≈? X, (e f ).Y ≈? Y }i avoid any combinatory solution for atoms in (e f ), but there are unitary epc’s w.r.t (a b c d)
that give solutions; for instance, epc’s s1 = (a ⊕ c) ∗ (b ⊕ d), s2 = (h a ⊕
h c) ∗ (h b ⊕ h d) and s3 = ([i]a ⊕ [i]c) ∗ ([i]b ⊕ [i]d), generate solutions of the
form h{e, f #X, e, f #Y }, {X/si}i, i = 1, 2, 3; the epc v = (h ha, Zi⊕h hc, Zi)∗
(h hb, Zi ⊕ h hd, Zi) the solution h{e, f #X, e, f #Y, a, b, c, d, e, f #Z}, {X/v}i,
etc.
5
Conclusions and future work
A Coq formalisation of a sound and complete nominal C-unification algorithm
was obtained by combining ⇒≈ - and ⇒# -reduction. The algorithm builds finite
derivation trees such that the leaves provide a complete set of most general
unifiers consisting of freshness contexts, substitutions and fixpoint equations. We
have also introduced the notion of pseudo-cycle to eliminate fixpoint equations
by generating their possibly infinite set of solutions. The generator is based
on the analysis of permutation cycles, followed by a brute-force enumeration
procedure. An implementantion of the algorithm, as well as its extension to deal
with different equational theories is left as future work.
References
1. T. Aoto and K. Kikuchi. A Rule-Based Procedure for Equivariant Nominal Unification. In Pre-proc. of Higher-Order Rewriting (HOR), pages 1–5, 2016.
2. T. Aoto and K. Kikuchi. Nominal Confluence Tool. In Proc. of the 8th Int.
Joint Conf.: Automated Reasoning (IJCAR), volume 9706 of LNCS, pages 173–
182. Springer, 2016.
3. M. Ayala-Rincón, W. Carvalho-Segundo, M. Fernández, and D. Nantes-Sobrinho.
A Formalisation of Nominal Equivalence with Associative-Commutative Function
Symbols. ENTCS, 332:21–38, 2017.
4. M. Ayala-Rincón, W. Carvalho-Segundo, M. Fernández, and D. Nantes-Sobrinho.
On Solving Nominal Fixpoint Equations. In Proc. of the 11th Int. Symp. on Frontiers of Combining Systems (FroCoS), volume 10483 of LNCS, pages 209–226.
Springer, 2017.
5. M. Ayala-Rincón, M. Fernández, and D. Nantes-Sobrinho. Nominal Narrowing. In
Proc. of the 1st Int. Conf. on Formal Structures for Computation and Deduction
(FSCD), volume 52 of LIPIcs, pages 11:1–11:17, 2016.
6. M. Ayala-Rincón, M. Fernández, and A. C. Rocha-oliveira. Completeness in PVS
of a Nominal Unification Algorithm. ENTCS, 323:57–74, 2016.
7. F. Baader and T. Nipkow. Term Rewriting and All That. Cambridge UP, 1998.
8. T. Braibant and D. Pous. Tactics for Reasoning Modulo AC in Coq. In In Proc.
of the 1st. Int. Conf. on Certified Programs and Proofs (CPP), volume 7086 of
LNCS, pages 167–182. Springer, 2011.
9. C. F. Calvès. Complexity and implementation of nominal algorithms. PhD Thesis,
King’s College London, 2010.
Nominal C-Unification
15
10. C. F. Calvès and M. Fernández. Implementing Nominal Unification. ENTCS,
176(1):25–37, 2007.
11. C. F. Calvès and M. Fernández. The First-order Nominal Link. In Proc. of the
20th Int. Symp. Logic-based Program Synthesis and Transformation (LOPSTR),
volume 6564 of LNCS, pages 234–248. Springer, 2011.
12. J. Cheney. αProlog Users Guide & Language Reference Version 0.3 DRAFT, 2003.
13. J. Cheney. Equivariant unification. J. of Autom. Reasoning, 45(3):267–300, 2010.
14. R. A. Clouston and A. M. Pitts. Nominal Equational Logic. ENTCS, 172:223–257,
2007.
15. E. Contejean. A Certified AC Matching Algorithm. In Proc. of the 15th Int. Conf.
on Rewriting Techniques and Applications, (RTA), volume 3091 of LNCS, pages
70–84. Springer, 2004.
16. M. Fernández and M. J. Gabbay. Nominal Rewriting. Information and Computation, 205(6):917–965, 2007.
17. M. Fernández and M. J. Gabbay. Closed nominal rewriting and efficiently computable nominal algebra equality. In Proc. of the 5th Int. Work. on Logical Frameworks and Meta-languages: Theory and Practice (LFMTP), volume 34 of EPTCS,
pages 37–51, 2010.
18. M. Fernández, M. J. Gabbay, and I. Mackie. Nominal Rewriting Systems. In
Proc. of the 6th Int. Conf. on Principles and Practice of Declarative Programming
(PPDP), pages 108–119. ACM Press, 2004.
19. M. J. Gabbay and A. Mathijssen. Nominal (Universal) Algebra: Equational Logic
with Names and Binding. J. of Logic and Computation, 19(6):1455–1508, 2009.
20. M. J. Gabbay and A. M. Pitts. A New Approach to Abstract Syntax with Variable
Binding. Formal Aspects of Computing, 13(3-5):341–363, 2002.
21. D. Kapur and P. Narendran. Matching, Unification and Complexity. SIGSAM
Bulletin, 21(4):6–9, 1987.
22. R. Kumar and M. Norrish. (Nominal) Unification by Recursive Descent with
Triangular Substitutions. In Proc. of Interactive Theorem Proving, 1st Int. Conf.
(ITP), volume 6172 of LNCS, pages 51–66. Springer, 2010.
23. T. Kutsia, J. Levy, M. Schmidt-Schauß, and M. Villaret. Nominal Unification of
Higher Order Expressions with Recursive Let. In Proc. of the 26th Int. Sym. on
Logic-Based Program Synthesis and Transformation (LOPSTR), volume 10184 of
LNCS, pages 328–344. Springer, 2016.
24. J. Levy and M. Villaret. An Efficient Nominal Unification Algorithm. In Proc. of
the 21st Int. Conf. on Rewriting Techniques and Applications (RTA), volume 6 of
LIPIcs, pages 209–226, 2010.
25. T. Nipkow. Equational Reasoning in Isabelle. Science of Computer Programming,
12(2):123–149, 1989.
26. A. M. Pitts. Nominal Sets: Names and Symmetry in Computer Science. Cambridge
UP, 2013.
27. Sagan, B. E. The Symmetric Group: Representations, Combinatorial Algorithms,
and Symmetric Functions. Springer, 2nd edition, 2001.
28. J. H. Siekmann. Unification of Commutative Terms. In Proc. of the Int. Symposium on Symbolic and Algebraic Manipulation, volume 72 of LNCS, pages 22–29.
Springer, 1979.
29. C. Urban. Nominal Unification Revisited. In Proc. of the 24th Int. Work. on
Unification (UNIF), volume 42 of EPTCS, pages 1–11, 2010.
30. C. Urban, A. M. Pitts, and M. J. Gabbay. Nominal Unification. Theoretical
Computer Science, 323(1-3):473–497, 2004.
| 6cs.PL
|
www.nature.com/scientificreports
OPEN
Three representations of the Ising
model
Joost Kruis1 & Gunter Maris1,2
received: 21 June 2016
accepted: 07 September 2016
Published: 04 October 2016
Statistical models that analyse (pairwise) relations between variables encompass assumptions about
the underlying mechanism that generated the associations in the observed data. In the present paper
we demonstrate that three Ising model representations exist that, although each proposes a distinct
theoretical explanation for the observed associations, are mathematically equivalent. This equivalence
allows the researcher to interpret the results of one model in three different ways. We illustrate the
ramifications of this by discussing concepts that are conceived as problematic in their traditional
explanation, yet when interpreted in the context of another explanation make immediate sense.
Scientific advances can be achieved by two types of theories: those that simply seek to identify correlations
between observable events without regard to linking mechanisms; and those that specify the mechanisms governing the relations between observable events (Bandura, p. 21).1.
Examining the structure of observed associations between measured variables is an integral part in many
branches of science. At face value, associations (or their quantification in the form of correlations) inform about
a possible relation between two variables, yet contain no information about the nature and directions of these
relations. Making causal inferences from associations requires the specification of a mechanism that explains the
emergence of the associations2,3. By constructing an explanatory model to account for associations in the data,
that has testable consequences at the level of the joint distribution of variables, it is possible to test the adequacy of
the model against the data. When the model is deemed sufficiently adequate with respect to the data, this is often
perceived as justification for the proposed causal interpretation4.
We can discern (at least) three general frameworks, each representing a different mechanism to explain the
emergence of associations between variables, with their own collection of corresponding explanatory models.
These frameworks, and their corresponding statistical models, all originate from different disciplines and have
received considerable attention in diverse fields such as, physics, mathematics, statistical mechanics, causality,
biology, epidemiology, and social sciences. Although both authors originate from psychology, and primarily
illustrate their findings with examples from this field, the frameworks and models discussed in this paper clearly
transcend the domain of psychology, and as such have a multidisciplinary relevance. In the current paper we
refer to these frameworks as, respectively, the common cause-, reciprocal affect-, and common effect framework.
The common cause framework explains the observed associations through a latent (unobserved) variable acting as a common cause with respect to the manifest (observed) variables5. Causal models that propose a common cause mechanism as generating the associations between manifest variables, are also known as reflective
models6,7; the manifest variables are indicators of the latent variable and reflect its current state. In the statistical
literature, models in this framework are therefore often referred to as latent variable models. Latent variable
models have proven to be extremely successful at fitting observed multivariate distributions, at the same time
their theoretical and philosophical underpinning remains problematic. Latent variables are both a powerful and
controversial concept, in psychology for example, the idea of psychological constructs as intelligence8,9 and personality10 being latent variables has been the subject of many intense debate. In particular about the question
whether one should take a realist interpretation of latent variables, that is the latent variable signifying a real but
hidden entity, to justify the use of latent variable analysis11,12. An important reason for this is that a latent cause is
never observed, and similar to physics around the turn of the 20th century, there was need for…
Popper, p. 21113. …an epistemological programme: to rid the theory of ‘unobservables’, that is, of magnitudes
inaccessible to experimental observation; to rid it, one might say, of metaphysical elements.
1
University of Amsterdam, Department of Psychology, Amsterdam, The Netherlands. 2Cito, Psychometric Research
Center, Arnhem, The Netherlands. Correspondence and requests for materials should be addressed to J.K. (email:
[email protected])
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
1
www.nature.com/scientificreports/
a
c
b
Χ1
Χ2
Χ1
Χ2
Χ1
Θ
Χ4
Χ2
Ε
Χ3
Χ4
Χ3
Χ4
Χ3
Figure 1. Three frameworks for explaining observed associations in the parametrisation of their
prototypical statistical models. (a) The Rasch model from the common cause framework as a DAG. (b) The
Curie-Weiss model from the reciprocal affect framework as an undirected graph. (c) The collider selection bias
model from the common effect framework as a DAG.
It is in the reciprocal affect framework that we find such a programme without ‘unobservables’. In this framework the relations between variables are represented as is, in that, observed associations between manifest variables are explained as a consequence of mutualistic relations between these variables14. The idea that variables
are associated as a consequence of reciprocal affect has been formalised in the field of network analysis15, and
has been studied extensively in diverse field of science such as mathematics, physics, and biology16–19. The Ising
model is a suitable example of a statistical model in this framework, as it captures all main effects and pairwise
interactions between variables20,21. While originally a model of ferromagnetism from statistical physics, in the
last decade the Ising model has been adopted within the social sciences, where the network perspective has been
gaining much popularity22–25. As an alternative for the latent variable perspective, the network approach has lead
to valuable new insights about, for example, psychopathology and the aforementioned concept of intelligence26.
In the, final, common effect framework, observed associations between manifest variables are explained as
arising from (unknowingly) conditioning on a common effect of these variables; the manifest variables are marginally independent with respect to each other, yet their collective state leads to the occurrence (or absence) of the
effect27,28. Variables can act as the collective cause towards an effect in (at least) two ways, either as the separate
indicators of an artificial compound score (e.g. Socio-Economic-Status, SAT)29, or as determinants of a naturally
occurring phenomenon, such as agents culminating into the outbreak of an epidemic. In the statistical literature,
the term collider variable models is used for this framework, as the collective state of the independent variables
collides into the effect4,30. Because of the independence, one would naturally expect not to find any associations
between these variables. However, from the literature on causality it is known that conditioning on a collider variable introduces (spurious) correlations among the variables functioning as the collective cause. This phenomenon
is known as endogenous selection bias and will results in the observation of associations between the manifest
variables30–34.
It is clear that each of these frameworks proposes a radically different explanation for the emergence of associations between a set of manifest variables. In this paper we argue that these differences only exist with respect
to the theoretical interpretation of these frameworks. Specifically, we demonstrate that the prototypical statistical
models for binary data in each framework are mathematically equivalent, and that this equivalence extends to
more realistic models that capture all main effects and pairwise interactions between the observed variables.
Through this we obtain three, statistically equivalent, representations of the Ising model that each explain the
occurrence of associations between binary variables by a theoretically very distinct mechanism.
Results
Prototypical models. To enhance the readability of this section we start by introducing the variables that
return in all discussed models, and clarify the mathematical notation used in the text and equations for the
distribution functions. We will denote random variables with capital letters and possible realisations of random
variables with lower case letters. We represent vectors with bold-faced letters, and use boldfaced capital letters to
indicate matrices for parameters. Manifest variables are denoted with Roman letters, whereas we use Greek letters
to indicate latent variables and parameters that need be estimated.
In the context of the paper we are primarily interested in the vector X = [X1, X2, …, XN], consisting of N binary
random variables that can take +1 and −1 as values, as we look to the mechanism by which the three frameworks
explain the observed associations between the realisations of this vector denoted by x = [x1, x2, … , xn].
Furthermore, each of the models we discuss includes a vector containing the main effect parameters δ = [δ1,
δ2, …, δn], consisting of N numbers in . Except for equation (1) which we write out in full, we use ∑i and ∏i to
denote respectively ∑ ni =1 and ∏ni =1 for the remainder of the equations. Additionally, we use p(x) to denote
p(X = x), which extends to all variables in both conditional, and joint probability distributions, such that we can
read it as the probability of observing some realisation of the random vector X, optionally, conditional on, or
together with, the realisation of some other variable.
We consider the Rasch model35, an Item Response Theory (IRT) model from the field of psychometrics, as the
prototypical model for binary data in the common cause framework. Historically, the Rasch model has been
developed for modelling the responses of persons to binary scored items on a test. The model is graphically represented in Fig. 1(a) as a Directed Acyclic Graph (DAG) 36, where the latent random variable Θ acts as the
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
2
www.nature.com/scientificreports/
common cause of the manifest random variables X. The Rasch model is characterised by the following distribution for the manifest variables (X), conditional on the latent variable (Θ ∈ ):
exp(x [θ + δ ])
∏ exp( + [θ + δ ])i + exp(i − [θ + δ ])
p (X = x Θ = θ ) =
i
i
i
(1)
The marginal probabilities for X can be obtained by endowing Θwith a distribution as shown in equation (2).
While this gives us an expression for the manifest probabilities of the Rasch model, for almost all choices for the
distribution of f(θ), this expression becomes computationally intractable.
p ( x) =
x i [θ + δ i ])
f (θ) dθ
∫ ∏ exp( + [θ +exp(
δ ]) + exp( − [θ + δ ])
i
i
i
(2)
In the traditional interpretation of the Rasch model, xi indicates whether the response to item i is correct
(xi = +1) or incorrect (xi = −1). In this context, the continuous random variable (Θ) represents the latent ability
being examined by the set of items (X). The vector δ contains the item main effects, where δi represents the easiness of item i, such that −δi represents the difficulty of item i with respect to the measured ability. The response
of an individual on item i is a trade-off between the ability of the person (θ) and the item difficulty (−δi). When
the ability of the person is greater than the difficulty of the item, the probability for a correct response will be
higher than for an incorrect response (θ > −δi ⇒ p(xi = 1) > p(xi = −1)), if the ability of the person is lower than
the item difficulty the reverse holds (θ < −δi ⇒ p(xi = 1) < p(xi = −1)). As such, persons with a greater ability will
always have a higher probability of giving a correct response, and persons always has a higher probability for a
correct response on an easy item than on a more difficult item. A key property of the Rasch model is that of local
independence, which entails that only variation in Θdetermines the probability for a response on an item. That
is, conditionally on the state of the latent variable all manifest variables are independent, such that marginally
(with respect to the latent variable) they are dependent. Consequently, any observed associations between the
manifest variables can be traced back to the influence of the latent variable. It is the latent variable that causes the
associations on the manifest variables, which is why the Rasch model falls within the common cause framework.
For the reciprocal affect framework we examine the Curie-Weiss model from statistical physics37–39, originally
used to model the state of a set of magnetic moments, for which the thermodynamical properties correspond to
that of the classical Curie-Weiss theory of magnetism, and where the pairwise interactions between the magnetic
moments are replaced by the mean magnetisation. Graphically, the Curie-Weiss model can be represented as an
undirected graph wherein the manifest variables (X), representing the set of magnetic moments, are fully connected with each other, and all connection are of equal strength, as illustrated in Fig. 1(b). The distribution of the
manifest variables (X) in the Curie-Weiss model is given by:
p ( x) =
2
1
1
exp ∑x i δ i + ∑x i
i
Z
2 i
(3)
In the conventional interpretation of the Curie-Weiss model, xi indicates that the magnetic spin of moment i
is upward (xi = +1) or downward (xi = −1), whereas the main effect for each moment (δi) indicates the natural
preference of moment i to be in an upward (δi > 0) or downward (δi < 0) spin position, due to the external magnetic field not present in X. In equation (3) we use Z to represents the normalising constant, in thermodynamical
systems often referred to as the partition function, that makes the distribution sum to one. In the Curie-Weiss
model, this partition function sums over all 2N possible configurations of the vector X, which we denote in this
paper as ∑ x , and is given by the following expression:
Z=
2
1
∑ exp ∑x iδi + 2 ∑x i
x
i
i
(4)
As the pairwise interactions between the magnetic moments are replaced by the mean magnetisation, all interactions are captured by the squared sum of the set of moments in the exponential of the Curie-Weiss distribution.
By averaging over individual interactions between magnetic moments, the Curie-Weiss model is the simplest
non-trivial model that exhibits a phase transition in statistical mechanics. However, because the model violates
fundamental principles of statistical physics, and its predictions are only partially verified by experiments, it is
considered as being mainly of theoretical interest37. Nonetheless, due to its simplicity, the Curie-Weiss model has
been useful in understanding the dynamics of equivalent phenomena in more realistic systems, such as the Ising
model39. Still, it is clear that as the observed associations between magnetic moments are presumed to emerge due
to the magnetic interaction between these moments themselves, the Curie-Weiss model falls within the reciprocal
affect framework.
Whereas in the Rasch model, from the common cause framework, associations between manifest variables
are explained by the latent variable Θ, in the Curie-Weiss model, from the reciprocal affect framework, these
associations are captured in the squared sum of the set of moments in the exponential of the distribution. The key
ingredient for establishing the connection between these two models has been known for a long time, and has
also been rediscovered quite a few times in quite diverse fields of science37,40–45. It was in his Brandeis lecture that
Mark Kac37 established the relation between the Curie-Weiss model and the Rasch model through an ingenuous
use of the following Gaussian integral:
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
3
www.nature.com/scientificreports/
exp(a2) =
2
∫ exp(2aθπ − θ ) dθ
(5)
What Kac realized is that whenever you see the exponential of a square, you can replace it with the right hand
side integral from equation (5). In the Methods section we demonstrate that applying this Gaussian identity to
the Curie-Weiss distribution from equation (3) linearises the squared sum in the exponential, and introduces a
random variable Θ, such that we obtain a latent variable representation of the Curie-Weiss model. We then show
that because the square in the exponential is gone, we can rewrite the expression for the latent variable representation of the Curie-Weiss model such that, both the marginal distribution of the manifest variables, and that of the
manifest variables conditional on θ, are identical to that of the Rasch model from equation (1) and equation (2).
Having established the relation between the prototypical models from the common cause, and reciprocal affect
framework we turn to our third framework.
For the common effect framework we consider X as a set of independent random variables, which we will
collectively call the cause, together with a single dependent binary random variable (E), which we will call the
effect. Their joint distribution, given in equation (6), is a collider structure and can be graphically represented in
a DAG as illustrated in Fig. 1(c).
e
exp( 1 [∑i x i ]2 )
exp(x i δ i )
2
p (x , e ) = ∑
1
sup exp( 2 [∑i x i ]2 )
i exp( + δ i ) + exp( − δ i )
x
1 −e
1
exp( 2 [∑i x i ]2 )
× 1 −
1
sup exp( 2 [∑i x i ]2 )
x
(6)
In this collider distribution, xi indicates if cause i is active (xi = +1) or inactive (xi = −1), whereas (e) indicates
whether the effect was present (e = 1) or absent (e = 0) at that time. The main effect for each cause (δi) denotes
the natural predisposition for cause i to be active (δi > 0) or inactive (δi < 0) at any given time. As mentioned, and
shown by equation (9) in the Methods section, are the individual causes independent of each other in the marginal distribution of X. As a consequence when we marginalise X with respect to E, the causes will not show any
associations among each other. From the literature on causality it is however known that selection with respect to
a common effect variable will introduce (spurious) correlations among the causes. That is, by using only observations of x where the common effect is present (e = 1), this set of observations will show a pattern of associations
among the causes. This is known as endogenous selection bias with respect to a collider variable, and can be
mathematically represented in the distribution of the causes conditionally on the effect. In the Methods section
we demonstrate that when we apply this selection bias mechanism to the collider structure from equation (6),
the distribution of the collective cause (X) conditionally on the effect, exactly gives the Curie-Weiss model from
statistical physics, and hence, the Rasch model.
Realistic models. Having studied the three statistical explanations in their simplest non-trivial form, we
conclude that, although their theoretical interpretation is radically different, the three models are mathematically indistinguishable. Still, the simplicity of the prototypical models for each framework also makes them often
unrealistic with respect to the observed reality. Specifically, the Rasch model and simple collider model only allow
for main effects between the observations, and do not consider possible pairwise interactions. The Curie-Weiss
model does allow some crude form of interaction, however, as the individual interactions between nearest neighbours are replaced by the mean interaction, one must make the (often) unrealistic assumption that all observations are interconnected with the same strength. Fortunately, we can swiftly generalise all three prototypical
models to more realistic forms.
We start with the Ising model, of which the Curie-Weiss model is the simplest form, from the reciprocal affect
framework. Like the Curie-Weiss model, the Ising model was originally introduced in statistical physics as a
model for magnetism, with the same possible values and interpretation for X and its possible realisations.
However, instead of only considering the mean magnetisation, the Ising model captures all pairwise interactions
between the set of manifest variables (X). The distribution of the Ising model, where ∑ i , j is the sum over all
distinct pairs of magnetic moments, is commonly written as follows:
p ( x) =
1
exp ∑x i δ i +
i
Z
∑x ix jσ ij
i ,j
(7)
The pairwise interactions are represented in the Ising model distribution by the symmetric N × N connectivity
matrix ∑ in . In this connectivity matrix, σij modulates the reciprocal affect relation between xi and xj, indicating if
moments i and j prefer to be in identical (σij > 0), or opposing (σij < 0) spin positions, wherein the higher the absolute
value of σij, the stronger this preference. Under the condition that all off-diagonal entries of Σ are equal, the Ising
model reduces to the prototypical Curie-Weiss model. Because the diagonal values of the connectivity matrix in the
Ising model are arbitrary, i.e., the probability of X is independent of these values, we can choose the values for the
diagonal in such a way that the connectivity matrix becomes positive (semi) definite. As a consequence the eigenvalue
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
4
www.nature.com/scientificreports/
decomposition of the matrix will also be non-negative. As clarified in the Methods section, by applying this transformation to the Ising model distribution from equation (7) we obtain an eigenvalue representation of the Ising model:
p ( x) =
1
exp ∑x i δ i +
i
Z
1
∑ 2 λr ∑qir x i
r
i
2
(8)
where λr is the r non-negative eigenvalue of the vector Λ = [λ1, λ2, …, λN], and qir the value of the i row and rth
column of the N × N eigenvector matrix Q. In the equation above ∑r should be read as ∑ nr =1, we continue this
practice in the notation of the applicable equations in the Methods section. In the Methods section we demonstrate how this eigenvalue representation allows us to connect the Ising model from the reciprocal affect framework to the more realistic models in both the other frameworks. First, by applying the Gaussian identity from Kac
to the squared sum in the exponent for each eigenvalue in equation (8), we obtain a latent variable representation
of the Ising model46, with as many latent dimensions as there are non-zero eigenvalues. This latent variable representation of the Ising model is then shown to be the multidimensional IRT model47 from the common cause
framework, of which the Rasch model is the simplest instance, but allows for more than one latent variable to
explain the observed associations between the manifest variables. Similarly, we can introduce (independent)
effect variables for each eigenvalue in equation (8), such that we obtain a collider representation of the Ising
model where endogenous selection bias has taken place. We then show that this distribution is a version of the
common effect model as seen in equation (6), that is extended such that the collective cause can collide into more
then one common effect.
th
th
Discussion
We have shown that the mathematical equivalence of the simple prototypical models from the common cause,
reciprocal affect, and common effect framework, extends to the more realistic counterparts of these models.
That is, there exist three, statistically indistinguishable, representations of the Ising model that explain observed
associations either through marginalisation with respect to latent variables, through reciprocal affect between
variables, or through conditioning on common effect variables. We therefore argue that these are not three different models, but just one model for which three distinct theoretical interpretations have been developed in
different fields of science. Consequently, any set of associations between variables that is sufficiently described by
a model in one framework, can be explained as emerging from the mechanism represented by any of the three
theoretical frameworks. We illustrate the implications of this by considering one of the most controversial topics
in the common cause framework, differential item functioning (DIF)48, and discuss it in the context of the three
possible interpretations.
In it’s traditional (common cause) framework DIF indicates that, conditional on the level of the latent variable,
the probability for some response is dependent on group membership. For items that exhibit DIF it is not only
variation in Θthat determines the probability for a response on an item. From a common cause perspective the
occurrence of DIF is a violation of local independence, and as such measurement invariance. In the context of
ability testing DIF is often perceived as indication of item bias49. As a fictitious example, consider the situation
where on certain items from the Revised NEO Personality Inventory (NEO PI-R)50, that intents to measure the
Big Five personality traits51, we find that for a group of subjects with the same latent trait score, those that listed
their occupation as being a manager always have a higher probability of giving a correct response on these items,
compared to subjects that have no occupation as a manager.
Needless to say, in this context items that exhibit DIF are seen as bad because they pose a problem for both the
reliability and validity of a test. From a reciprocal affect perspective, identification of DIF would exhibit itself in
the form of differences in the estimated pairwise associations between items depending on group membership.
As such the appearance of DIF in the NEO PI-R example would also be viewed as troublesome. However, in contrast to the common cause framework, the appearance of DIF in a network model is at least informative in that
our model might be incomplete, i.e., the network is missing a node. The interpretation of DIF in a common effect
framework is best understood, in the context of the current example, by considering the answer to the question:
What causes people to obtain a managerial position as occupation? It is safe to say that in most cases a persons personality is an important factor in this process. In other words, people that are selected to become manager, get this
position because they posses a certain set of personality traits associated with being a successful manager52. As
such, the items in the NEO PI-R that show DIF in this case measure those personality traits that are most sought
after in managers. More broadly in the context of the common effect framework, the occurrence of DIF indicates
how well an item predicts differences in the effect. In contrast to the disruptive interpretation of DIF in both the
common cause and reciprocal affect frameworks, the occurrence of DIF within the context of the common effect
framework is actually both sensible and informative.
The previous example clearly demonstrates how fundamental concepts, that are firmly established in their
traditional framework as being problematic, can be perceived as neutral and informative or even desirable in
another context. Having multiple possible interpretations for the same model allows for more plausible explanations when it comes to the theoretical concepts and the causal inferences we obtain from the measurement model
applied to our data. For example in the context of psychopathology, depression has been habitually being treated
as a common cause variable for which its symptoms are the interchangeable indicators. Measures of these symptoms with the popular Beck Depression Inventory53 have shown to fit a latent variable model with one underlying
general depression factor and three highly inter-correlated sub-factors, or a two-factor solution well54. However,
in a common cause framework depression symptoms, such as sleep problems, loss of energy, and trouble concentrating, are assumed independent of one another, as they are purely caused by the latent variable interpreted as
depression. More recently it has been shown that a network model can also give an accurate description of data
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
5
www.nature.com/scientificreports/
on depression symptoms26. The reciprocal affect representation of depression, where symptoms can directly influence each other, is as an explanation more in line with our perceived reality. Furthermore, the historical success
of theoretically very implausible models, such as the latent variable model, can thus in retrospect, arguably be
explained by the equivalence of these three models.
Being able to interpret the outcome of an applied measurement model from theoretically very distinct perspectives, instead of only the perspective as traditionally assumed by the model, is great progress, as it allows for
novel explanations that might be a better reflection of our perceived reality. Furthermore, in their different fields
of application different aspects of these models have been studied and different methodology has been developed.
Through their connection much of these developments become available to all fields of application.
Methods
In this section we clarify the mathematics involved in connecting the simple prototypical models, as well as the
more realistic Ising model representations, for the three different frameworks. In the first proof we demonstrate
the equivalence between the simple collider, Curie-Weiss and Rasch models, the prototypical (yet unrealistic)
models for respectively the common effect, reciprocal affect, and common cause explanation for observed associations between a set of binary variables.
Proof for the equivalence of the simple prototypical models. Collider to Curie-Weiss. The simple
collider model from the common effect framework is characterised by the following joint probability distribution
p(x, e):
p (x , e ) =
∏
i
exp 1 [∑i x i ]2
exp(x i δ i )
2
exp( + δ i) + exp( − δ i ) sup exp 1 [∑i x i ]2
2
x
(
1
exp 2 [∑i x i ]2
×1 −
1
sup exp 2 [∑i x i ]2
x
(
)
(
)
(
e
)
1 −e
)
(9)
In order to connect this collider model to the Curie-Weiss model we introduce endogenous selection bias
on the set of manifest variables forming the collective cause, by conditioning on the effect being present. This is
mathematically presented as the conditional distribution p(x|e = 1), proportional to the product of the marginal
distribution for the cause p(x), and the probability of observing the effect given the cause p(e = 1|x), defined by:
p (x|e = 1) ∝ p (x) p (e = 1|x) =
∏
i
(
)
1
exp 2 [∑i x i ]2
exp(x i δ i )
exp( + δ i ) + exp( − δ i ) sup exp 1 [∑i x i ]2
x
(
2
)
(10)
We can simplify the expression for p(x|e = 1) by recognising that the product of exponentials in the numerator can be rewritten as a sum within the exponential. Furthermore, the denominator of the expression is only
dependent on the sum of X, and thus independent of the specific pattern that the realisation of X takes. As a consequence p(x|e = 1) is only proportional to the numerator of equation (10), such that we can write:
2
1
p (x|e = 1) ∝ exp ∑x i δ i + ∑x i
i
2 i
(11)
In order to obtain a valid probability density function we have to add the appropriate normalising constant
that makes the probabilities sum to one again. In this case this translates to dividing the expression in equation (11) for a certain realisation of X by the sum of this expression for all possible configurations of X:
p (x|e = 1) =
(
1
exp ∑ i x i δ i + 2 [∑i x i ]2
(
∑ x exp ∑ i x i δ i +
1
[ ∑i
2
)
x i ]2
)
(12)
It can quickly be verified that the resulting expression in equation (12) is identical to the distribution for the
Curie-Weiss model introduced in equation (3), with the same normalising constant as given in equation (4). Thus
proofing that, conditional on the effect being present, the distribution of the collective cause in the collider model
is equivalent to the distribution of a set of directly interacting magnetic moments in the Curie-Weiss model.
Curie-Weiss to Rasch. Next, we will connect the Curie-Weiss model from statistical physics to the Rasch model
from psychometrics. We start from the distribution function of the Curie-Weiss model, where we use Z to denote
the appropriate normalising constant:
p ( x) =
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
(
1
exp ∑ i x i δ i + 2 [∑i x i ]2
Z
)
(13)
6
www.nature.com/scientificreports/
Next we use Kac’s Gaussian identity from equation (5) to linearise the quadratic sum in the exponential of the
Curie-Weiss distribution, to that end let a2 = 1 [∑i x i ]2, so we can rewrite it in the following way:
2
1
exp [∑i x i ]2 =
2
(
1
exp 2 2 [∑i x i ] θ − θ 2
∫
π
) dθ
(14)
By incorporating this transformation we obtain a latent variable representation of the Curie-Weiss model
p ( x) =
∫ exp(∑i
x i δ i + [∑i x i ] θ − θ 2 )
dθ
Z π
(15)
Which we can simplify further by merging the two sums in the exponential:
p ( x) =
2
∫ exp(∑i xi [Zθ +π δi] − θ ) dθ
(16)
Where we will use Z ⁎ = Z π to denote the appropriate normalising constant. For the next step towards our goal
we multiply both the numerator and denominator of equation (16) by ∏i [exp( + [θ + δ i ]) + exp( − [θ + δ i ])],
such that we obtain the equivalent expression:
p ( x) =
2
[
exp( [
i ])
i ])]
dθ
∫ exp(∑i xi [θZ+⁎ δi] − θ ) ∏∏i [exp(
[exp( + [θ + δ ]) + exp( − [θ + δ ])]
+ θ+δ
+
− θ+δ
i
i
i
(17)
Next we rearrange the expression in equation (17) by switching the denominators of both factors, taking the
sum in the first numerator out of the exponential so it becomes a product, and transferring exp (−θ2) out the
numerator of the first factor, and into the numerator of the second factor:
p ( x) =
x i [θ + δ i ])
∫ ∏ exp( + [θ +exp(
δ ]) + exp( − [θ + δ ])
i
i
i
2
∏ [exp( + [θ + δ i ]) + exp( − [θ + δ i ])]exp( − θ )
× i
dθ
⁎
Z
(18)
The resulting expression can be recognised as p (x) = ∫ p (x θ) f (θ) dθ, the marginal probability for some
realisation of X where the latent variable Θis integrated out. Let us denote the second factor of the expression in
equation (18) as the distribution of the latent variable (f(θ)), which gives us:
p ( x) =
x i [θ + δ i ])
f (θ)dθ
∫ ∏ exp( + [θ +exp(
δ ]) + exp( − [θ + δ ])
i
i
i
(19)
Such that the distribution of the set of binary random variables (X), conditionally on the latent variable (Θ) , is:
p ( x| θ ) =
exp(x [θ + δ ])
∏ exp( + [θ + δ ])i + exp(i − [θ + δ ])
i
i
i
(20)
Again, it is readily seen that the resulting latent variable expression of the Curie-Weiss model in equation (20)
is identical to the distribution of the Rasch model from equation (1). This completes our first proof in which we
demonstrated that by conditioning on the effect in the collider model from the common effect framework, the
distribution of set of binary random variables (X) is equivalent to that of the Curie-Weiss model from the reciprocal affect framework. Furthermore, when we linearise the quadratic sum in the exponential of the Curie-Weiss
model, we obtain a latent variable representation of this model where the distribution of the manifest random
variables (X) given the latent variable (Θ) is equivalent to that of the Rasch model from the common cause
framework. Consequently, given the equivalence of the collider model and the Curie-Weiss model, and that of
the Curie-Weiss model and the Rasch model, we can conclude that the collider model and the Rasch model are
also equivalent.
In the next proof we demonstrate that this equivalence relation between the three frameworks extends to the
more realistic models of these frameworks, as those allow pairwise interactions between the random variables in
the set X. We start with the conventional representation of the full Ising model from the reciprocal affect framework and rewrite this into an equivalent eigenvalue representation. Subsequently we connect this to both a latent
variable representation equivalent to the multidimensional IRT model from the common cause framework, and
a collider representation from the common effect framework.
Three representations of the Ising model. Conventional to Eigenvalue representation. The distribution
of the Ising model is commonly written as follows:
p ( x) =
1
exp ∑x i δ i +
i
Z
∑x ix jσ ij
i ,j
(21)
Where the partition function Z, that makes the distribution sum to one, is given by:
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
7
www.nature.com/scientificreports/
Z=
∑ exp ∑x iδi + ∑x ix jσ ij
x
i
i ,j
(22)
In order to connect the Ising model from the reciprocal affect framework to the models from both other
frameworks, we first have to rewrite it into matrix notation such that we can obtain the eigenvalue representation
of the Ising model. To that end we first rewrite the sum over all distinct i, j pairs in the exponent, as a function of
the sum over i and the sum over j:
p ( x) =
1
1 n n
exp ∑x i δ i + ∑∑ x i x j σ ij
i
Z
2 i=1 j=1
(23)
Such that we may rewrite the Ising model in matrix notation:
p ( x) =
1
1
exp xTδ + xTΣx
2
Z
(24)
All parameters, except for entries on the diagonal of the connectivity matrix, are identifiable from the data.
However, as xi xj = 1 when i = j any diagonal entry for the connectivity matrix will be cancelled out by the partition function. With the observation that the diagonal values of Σ are thus arbitrary (i.e., do not change the
probabilities), we can shift them in such a way that the connectivity matrix Σbecomes positive (semi) definite,
and hence its eigenvalue decomposition non-negative. This allows for the transformation Σ + cI = QΛQT, where
c contains the chosen values for the diagonal of the connectivity matrix, that when implemented gives:
p ( x) =
1
1
exp xTδ + xTQΛQT x
2
Z
(25)
By taking the expression out of its matrix notation we obtain the eigenvalue representation of the Ising model:
p ( x) =
1
exp ∑x i δ i +
i
Z
1
∑ 2 λr ∑qir x i
r
i
2
(26)
Eigenvalue to Latent Variable representation. We obtain a latent variable representation of the Ising model by
applying Kac’s Gaussian identity to the squared sum in the exponent of equation (26). To that end let a2 be
1
λ [∑i qir x i ]2 for each of the N eigenvalues in the Ising model, and replace this with the right hand side integral
2 r
from equation (5):
2
1
exp λr ∑qir x i =
2 i
∫
1
1
exp 2 λr ∑qir x i θr − θr2 dθr
2
π
i
(27)
N
#
Incorporating this transformation into the Ising model, and letting Z = Z π , we get he latent variable
representation of the Ising model, where the number of non-zero eigenvalues represents the number of latent
dimensions in the model:
p ( x) =
∫ Z1# exp ∑xiδi + ∑λr ∑qir xi θr − θr2 dθ
i
r
i
(28)
In order to connect this latent variable representation to the multidimensional IRT model we the multiply we
multiply both the numerator and denominator of equation (28) by ∑ x exp(∑i x i δ i + ∑r λr [∑i qir x i ] θr ), such
that we obtain the equivalent expression:
p ( x) =
exp(∑ i x i δ i + ∑ r λ r [∑ i qir x i ] θr − θr2 )
Z#
∑ x exp(∑i x i δ i + ∑r λr [∑i qir x i ] θr )
×
dθ
∑ x exp(∑i x i δ i + ∑r λr [∑i qir x i ] θr )
∫
(29)
Next we rearange and simplify the expression in equation (29). To that end let us merge the sums over xi in the
exponential, and denote λrqir as αir, where αir is the value of the ith row and rth column of the N × N matrix A in .
We can then rewrite the sum over r as a product of the vector αiT , containing the ith row of the matrix A, and the
vector θ. Furthermore, we switch the denominators of both factors and transfer exp(∑r θr2 ) out the numerator of
the first factor, to the numerator of the second factor:
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
8
www.nature.com/scientificreports/
p ( x) =
T
T
2
∑i x i [δ i + αi θ]) ∑ x exp(∑i x i [δ i + αi θ] − ∑r θr )
dθ
∫ ∑exp(
Z#
exp(∑ x [δ + α T θ])
x
i
i
i
i
(30)
In the resulting expression we can again recognise a function for the marginal probability of X where all latent
dimensions (Θ) are integrated out. Finally we take the sum over i out of the exponential such that we obtain:
p ( x) =
T
x i [δ i + αi θ])
f (θ)dθ
∫ ∏ ∑exp(
exp(x [δ + α T θ])
x
i
i
i
(31)
i
We can recognise this particular latent variable representation of the Ising model as a multidimensional IRT
model47 from the common cause framework, where the vector Θ represent the set of latent abilities measured by
the items in X. In addition to this vector, we also find the matrix A in the model, where the ith row contains the
discrimination parameters for all latent variables on item i. In the traditional interpretation of the IRT framework,
the discrimination parameter quantifies how well the item measures the corresponding latent variable, or in model
terms, the degree to which the probability for item responses varies with respect to each latent variable in Θ. We
obtain the following expression for the conditional probabilities of X given the vector of latent variables (Θ):
p ( x| θ ) =
∏
i
exp(x i [δ i + αiT θ])
∑ x exp(x i [δ i + αiT θ])
(32)
Note that, as the number of non-zero eigenvalues represents the number of latent dimensions in the model,
under the condition that only the first eigenvalue is non-zero, and the discrimination parameters with respect to
the single resulting latent variable are 1 for all the items, the model reduces to the Rasch model from equation (1).
Eigenvalue to Collider representation. To acquire an collider representation of the Ising model we start again
from the eigenvalue representation of the Ising model:
p ( x) =
1
exp ∑x i δ i +
i
Z
2
1
∑ 2 λr ∑qir x i
r
i
(33)
By taking the partition function out of the expression we obtain the following proportionality relation:
p (x) ∝ exp ∑x i δ i +
i
2
1
∑ 2 λr ∑qir x i
r
i
(34)
Next we can introduce a set of (independent) effect variables (E = [E1, E2, …, Em]) for each eigenvalue, such
that we obtain a collider representation of the Ising model where endogenous selection bias has taken place for
multiple effect variables. To that end, we recognise p(x) from equation (34) as p(x|e = 1), the conditional probability of the collective cause (X) given that all effects are present (E = 1), proportional to the product of the marginal distribution for the collective cause [p (x) ∝ ∑i exp(x i δ i )], and the probability of observing the effects given
the collective cause p (e = 1|x) ∝ exp ∑r 1 λr [∑i qir x i ]2 . Taking the sum over i in p(x), and the sum over r in
2
p(e = 1|x) out of their respective exponential, and adding the appropriate normalising constant to make the probabilities sum to one we obtain the following expression:
(
p (x|e = 1) =
∏
i
)
exp(x i δ i )
exp( + δ i ) + exp( − δ i )
∏
r
(
)
1
λ [∑i qir x i ]2
2 r
1
sup exp 2 λ r [∑i qir x i ]2
x
exp
(
)
(35)
Such that we can write the joint distribution of causes and effect variables as a common effect representation
of the Ising model:
p (x , e ) =
∏
i
exp 1 λ r [∑i q x i ]2
exp(x i δ i )
ir
2
∏
exp( + δ i) + exp( − δ i ) r sup exp 1 λr [∑i q x i ]2
ir
2
x
(
1
exp 2 λr [∑i qir x i ]2
× 1 −
1
sup exp 2 λr [∑i qir x i ]2
x
(
(
)
(
)
e
)
r
1 −e r
)
(36)
We can quickly recognise a collider model in this distribution that is extended for as much common effect
variables as there are non-negative eigenvalues. With this we have completed our second, and final, set of proofs,
showing that three, statistically equivalent, representations of the Ising model exist that explain observed associations between binary variables as arising either through marginalisation with respect to latent variables, through
reciprocal affect between variables, or through conditioning on common effect variables.
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
9
www.nature.com/scientificreports/
References
1. Bandura, A. Social cognitive theory: An agentic perspective. Asian J. Soc. Psychol. 2, 21–41 (1999).
2. Costner, H. L. Theory, deduction, and rules of correspondence. Am. J. Sociol. 75, 245–263 (1969).
3. Edwards, J. R. & Bagozzi, R. P. On the nature and direction of relationships between constructs and measures. Psychol. Methods 5,
155–174 (2000).
4. Pearl, J. Causality: models, reasoning and inference. Economet. Theor. 19, 675–685 (2003).
5. Reichenbach, H. The Direction of Time (University of California Press, Berkeley, 1956).
6. Fornell, C. & Bookstein, F. L. Two structural equation models: Lisrel and PLS applied to consumer exit-voice theory. J. Mark. Res. 19,
440–452 (1982).
7. Markus, K. A. & Borsboom, D. Reflective measurement models, behavior domains, and common causes. New Ideas Psychol. 31,
54–64 (2013).
8. Spearman, C. ‘General intelligence’ objectively determined and measured. Am. J. Psychol. 15, 201–293 (1904).
9. Jensen, A. R. The g factor: The science of mental ability (Praeger, Westport, 1998).
10. McCrae, R. R. & Costa, P. T. Empirical and theoretical status of the five-factor model of personality traits in The SAGE Handbook of
Personality Theory and Assessment: Volume 1 – Personality Theories and Models 273–294 (SAGE, London, 2008).
11. Borsboom, D., Mellenbergh, G. J. & van Heerden, J. The concept of validity. Psychol. Rev. 111, 1061–1071 (2004).
12. Borsboom, D. Psychometric perspectives on diagnostic systems. J. Clin. Psychol. 64, 1089–1108 (2008).
13. Popper, K. R. The Logic of Scientific Discovery (Hutchinson, London, 1959).
14. Van Der Maas, H. L. et al. A dynamical model of general intelligence: the positive manifold of intelligence by mutualism. Psychol.
Rev. 113, 842–861 (2006).
15. Epskamp, S., Maris, G., Waldorp, L. J. & Borsboom, D. Network Psychometrics in Handbook of Psychometrics. (Wiley, New York,
2015).
16. Lee, T.-D. & Yang, C.-N. Statistical theory of equations of state and phase transitions II. Lattice gas and Ising model. Phys. Rev. 87,
410–419 (1952).
17. Besag, J. Spatial interaction and the statistical analysis of lattice systems. J. R. Stat. Soc. Series B Stat. Methodol. (Methodological) 36,
192–236 (1974).
18. Besag, J. Statistical analysis of non-lattice data. Statistician 24, 179–195 (1975).
19. Barzel, B. & Barabási, A. L. Universality in network dynamics. Nat. Phys. 9, 673–681 (2013).
20. Ising, E. Beitrag zur theorie des ferromagnetismus. Zeit. Phys. 31, 253–258 (1925).
21. Jaynes, E. T. Information theory and statistical mechanics. Phys. Rev. 106, 620–630 (1957).
22. Barabási, A. L. The network takeover. Nat. Phys. 8, 14–16 (2012).
23. Cramer, A. O., Waldorp, L. J., van der Maas, H. L. & Borsboom, D. Comorbidity: A network perspective. Behav. Brain Sci. 33,
137–150 (2010).
24. Borsboom, D. & Cramer, A. O. Network analysis: an integrative approach to the structure of psychopathology. Annu. Rev. Clin.
Psychol. 9, 91–121 (2013).
25. Schmittmann, V. D. et al. Deconstructing the construct: A network perspective on psychological phenomena. New Ideas Psychol. 31,
43–53 (2013).
26. van Borkulo, C. D. et al. A new method for constructing networks from binary data. Sci. Rep. 4, 5918 (2014).
27. Blalock, H. M. Causal models involving unmeasured variables in stimulus-response situations in Causal Models in the Social Sciences
335–347 (Aldine-Atherton, Chicago, 1971).
28. Bollen, K. & Lennox, R. Conventional wisdom on measurement: A structural equation perspective. Psychol. Bull. 110, 305–314 (1991).
29. Hauser, R. M. Disaggregating a social-psychological model of educational attainment. Soc. Sci. Res. 1, 159–188 (1972).
30. Greenland, S., Pearl, J. & Robins, J. M. Causal diagrams for epidemiologic research. Epidemiology 10, 37–48 (1999).
31. Heckman, J. J. Sample selection bias as a specification error. Econometrica 47, 153–161 (1979).
32. Greenland, S. Quantifying biases in causal models: classical confounding vs collider-stratification bias. Epidemiology 14, 300–306
(2003).
33. Hernán, M. A., Hernández-Daz, S. & Robins, J. M. A structural approach to selection bias. Epidemiology 15, 615–625 (2004).
34. Elwert, F. & Winship, C. Endogenous selection bias: The problem of conditioning on a collider variable. Annu. Rev. Sociol. 40, 31–53 (2014).
35. Rasch, G. Probabilistic Models for some Intelligence and Attainment Tests (The Danish Institute of Educational Research, Copenhagen,
1960).
36. Christofides, N. Graph Theory: An Algorithmic Approach (New York: Academic Press Inc, 1975).
37. Kac, M. Statistical physics: Phase transitions and superfluidity in Brandeis University Summer Institute in Theoretical Physics, Vol. 1
(eds Chretien, M., Gross, E. & Deser, S.) 241–305 (Gordon and Breach Science Publishers, New York, 1968).
38. Stanley, H. E. Introduction to Phase Transitions and Critical Phenomena. (Oxford University Press, New York, 1971).
39. Ellis, R. S. & Newman, C. M. The statistics of curie-weiss models. J. Stat. Phys. 19, 149–161 (1978).
40. Olkin, I. & Tate, R. F. Multivariate correlation models with mixed discrete and continuous variables. Ann. Math. Stat. 32, 448–465
(1961).
41. Emch, G. G. & Knops, H. J. Pure thermodynamical phases as extremal kms states. J. Math. Phys. 11, 3008–3018 (1970).
42. Cox, D. R. & Wermuth, N. A note on the quadratic exponential binary distribution. Biometrika 81, 403–408 (1994).
43. McCullagh, P. Exponential mixtures and quadratic exponential families. Biometrika 81, 721–729 (1994).
44. Lauritzen, S. L. Graphical models (Oxford University Press, USA, 1996).
45. Anderson, C. J. & Yu, H.-T. Log-multiplicative association models as item response models. Psychometrika 72, 5–23 (2007).
46. Marsman, M., Maris, G., Bechger, T. & Glas, C. Bayesian inference for low-rank ising networks. Sci. Rep. 5, 9050 (2015).
47. Reckase, M. Multidimensional Item Response Theory (Springer, 2009).
48. Mellenbergh, G. J. Item bias and item response theory. Int. J. Educ. Res. 13, 127–143 (1989).
49. Meredith, W. Measurement invariance, factor analysis and factorial invariance. Psychometrika 58, 525–543 (1993).
50. Costa, P. T. & MacCrae, R. R. Revised NEO personality inventory (NEO PI-R) and NEO five-factor inventory (NEO FFI): Professional
manual (Psychological Assessment Resources, 1992).
51. Goldberg, L. R. An alternative “description of personality”: the big-five factor structure. J. Pers. Soc. Psychol. 59, 1216–1229 (1990).
52. Barrick, M. R. & Mount, M. K. The big five personality dimensions and job performance: a meta-analysis. Pers. Psychol. 44, 1–26 (1991).
53. Beck, A. T., Ward, C. H., Mendelson, M., Mock, J. & Erbaugh, J. An inventory for measuring depression. Arch. Gen. Psychiatry 562,
53–63 (1961).
54. Beck, A. T., Steer, R. A. & Carbin, M. G. Psychometric properties of the beck depression inventory: Twenty-five years of evaluation.
Clin. Psychol. Rev. 8, 77–100 (1988).
Acknowledgements
We thank Dr. Maria A. Bolsinova and Dr. Maarten Marsman for proof reading the manuscript. This work was
supported by by NWO (The Netherlands Organisation for Scientific Research), No. 022.005.0 (JK), and No.
CI1-12-S037 (GM).
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
10
www.nature.com/scientificreports/
Author Contributions
J.K. and G.M. wrote the main manuscript. All authors reviewed the manuscript.
Additional Information
Competing financial interests: The authors declare no competing financial interests.
How to cite this article: Kruis, J. and Maris, G. Three representations of the Ising model. Sci. Rep. 6, 34175;
doi: 10.1038/srep34175 (2016).
This work is licensed under a Creative Commons Attribution 4.0 International License. The images
or other third party material in this article are included in the article’s Creative Commons license,
unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license,
users will need to obtain permission from the license holder to reproduce the material. To view a copy of this
license, visit http://creativecommons.org/licenses/by/4.0/
© The Author(s) 2016
Scientific Reports | 6:34175 | DOI: 10.1038/srep34175
11
| 10math.ST
|
Sensitivity to the Neutrino Mass Hierarchy
Emilio Ciuffoli, Jarah Evslin and Xinmin Zhang
arXiv:1305.5150v4 [hep-ph] 20 Apr 2016
Theoretical Physics Division, IHEP, CAS, YuQuanLu 19B, Beijing 100049, China
In the next decade, a number of experiments will attempt to determine the neutrino mass hierarchy.
Feasibility studies for such experiments generally determine the statistic ∆χ2 . As the hierarchy is
a discrete choice, ∆χ2 does not obey a one degree of freedom χ2 distribution and so the number
of σ’s of confidence of the hierarchy determination is not the square root of ∆χ2 . We present a
simple Bayesian formula for the sensitivity to the hierarchy that can be expected from the median
experiment as a function of ∆χ2 .
In the next two decades a number of reactor,
accelerator and atmospheric neutrino experiments
will attempt to determine the neutrino mass hierarchy, which is the sign of the mass difference
2
∆M31
= M32 − M12 where Mi is the ith eigenvalue
of the neutrino mass matrix. If the sign is positive (negative), one says that the hierarchy is normal (inverted). Most of these experiments are still
in the planning stages, where the key role is played
by studies of the sensitivity of a given design to the
hierarchy.
Such studies determine, either analytically or via
Monte Carlo simulations,
∆χ2 = χ2I − χ2N
(1)
where χ2N (χ2I ) is the χ2 statistic equal to a weighted
sum of the squares of the differences between the
data and predictions given the normal (inverted) hierarchy, choosing all of the nuisance parameters so
as to minimize χ2N (χ2I ). The goal of these experiments is not to determine whether each of the hierarchies is consistent with the data, as would be usual
in a frequentist approach. Rather, as it is already
well accepted that precisely one of the hierarchies is
manifested in nature, the goal of these experiments
is to determine which of the hierarchies provides the
best fit to the data. In this paper we will use the
test statistic ∆χ2 to answer this question as follows.
We will define the best fit hierarchy to be that which
yields the lowest value of χ2 , and so the hierarchy
determined by the experiment simply corresponds to
the sign of ∆χ2 .
The critical question is then, given ∆χ2 , what is
the sensitivity of a typical experiment to the hierarchy? In Ref. [1] the authors showed that the most
naive answer, the p value that would be obtained
if ∆χ2 satisfied a one degree of freedom χ2 distri-
bution, gives the incorrect answer. Indeed ∆χ2 is
not necessarily positive and so such a prescription
would not even always be defined. In this note we
will provide an analytic answer (13) to this question
and will compare our answer to the results of simulations of Daya Bay II and disappearance data at
NOνA.
Nested hypotheses
To begin, we will describe just why the p value is
not the answer to the question stated above. Consider N data points {yi } generated by an experiment
trying to determine an unknown quantity x. We will
use the approximation in which these data points yi
(0)
follow a Gaussian distribution peaked at yi (x) with
(0)
variance σi2 (x). Both yi (x) and σi (x) are known
functions of x. An experimenter is interested in two
hypotheses. Hypothesis (A) states that x is a real
number. Hypothesis (B) states that x = x0 , for a
particular real number x0 . Clearly hypothesis (B) is
a special case of hypothesis (A), so these hypotheses
are said to be nested. In particular, (B) is obtained
from (A) by fixing one, otherwise unconstrained, real
number, the number x.
For any given value of x, the experimenter can define a statistic χ2 (x) by simulating the experiment
with that value of x and calculating the weighted
sum of the squares of differences between his measured and simulated results
χ2 (x) =
X (yi − y (0) (x))2
i
i
σi (x)2
.
(2)
The experimenter then determines a best fit x, for
which χ2 (x) is minimized. He then asks how compatible his results are with the hypothesis (B). To
determine this, he calculates
δχ2 = χ2 (x0 ) − χ2 (x).
(3)
2
Unlike ∆χ2 defined in Eq. (1), δχ2 is manifestly nonnegative, because x is defined so as to give the lowest
value of χ2 .
Just what value of δχ2 should the experimenter
expect? 75 years ago Wilks proved [2] that if hypothesis (B) is true then δχ2 will obey a χ2 distribution
with a single degree of freedom. The experimenter
can then determine a conditional probability that
given (B), the experiment would have gone as badly
as it did
!!
r
1
δχ2
2
pW (δχ ) =
1 − erf
.
(4)
2
2
For example, if he found δχ2 = 9, then pW would
only be about 0.13%, and so a frequentist experimenter might conclude that he has ruled out (B)
with 3σ of confidence.
Non-nested Hypotheses
As described in Ref. [1], the determination of the
hierarchy is qualitatively different. The two hypotheses are the normal hierarchy (NH) and the
inverted hierarchy (IH). These hypotheses are not
nested, and they correspond to a discrete choice, not
the fixing of a degree of freedom. So the conditions
for Wilks’ theorem are strongly violated. As was observed in general in Ref. [3] and in this context in
Ref. [1], this means that the statistic ∆χ2 defined in
Eq. (1) does not follow a χ2 distribution.
Just what distribution does ∆χ2 follow? Let us
begin with the simple case in which there are no nuisance parameters, which was applied to a toy model
of the hierarchy determination in Ref. [1].
An experiment will produce a set of numbers {yi },
which we assemble into a vector y. The normal and
inverted hierarchies yield two theoretical estimates
of this vector which we will denote yN and yI respectively. Again let us assume that the measured numbers yi are normally distributed about their mean
with a variance σi2 , which for simplicity we take to
be independent of the hierarchy. Without loss of
generality, let us assume for the moment that the
true hierarchy is normal. Then the measured numbers will be
yi = yiN + σi gi
(5)
where gi is a standard Gaussian random variable.
The statistic ∆χ2 is then easily determined to be
∆χ2 = χ2I − χ2N
X (yi − y I )2 X (yi − y N )2
i
i
−
=
2
2
σ
σ
i
i
i
i
=
X (y N + σi gi − y I )2 − (y N + σi gi − y N )2
i
i
i
i
σi2
i
=
(6)
X (y N − y I )2
i
i
σi2
i
+
X 2(y N − y I )
i
i
i
σi
gi .
This identifies ∆χ2 as a Gaussian distributed random variable with mean given by the first term on
the right hand side
∆χ2 =
X (y N − y I )2
i
i
i
σi2
(7)
and standard deviation given by the second term [1]
s
q
X 4(y N − y I )2
i
i
σ∆χ2 =
(8)
= 2 ∆χ2 .
2
σ
i
i
Note that ∆χ2 is the ∆χ2 statistic without statistical fluctuations, for example it may be given by
the theoretical spectra of ν e observed at a reactor
experiment, of νµ and ν µ at an iron calorimeter atmospheric neutrino experiment, or of νe (ν e ) appearance at an accelerator experiment running in the
neutrino (antineutrino) mode. In an atmospheric
neutrino experiment one may use the spectra as a
function of energy, zenith angle and even the inelasticity of the events [4]. ∆χ2 is the statistic most
often reported in the literature. We will now use
Eq. (8) to relate ∆χ2 to three quantities of interest.
What is the probability that the hierarchy which
yields the lowest χ2 is indeed the true hierarchy?
Let us first consider the case in which the normal hierarchy is manifested in nature. The correct
hierarchy will be determined by the experiment if
∆χ2 > 0. The statistic ∆χ2 is centered on theqposi-
tive value ∆χ2 with a standard deviation of 2 ∆χ2
q
and so the closest negative value is ∆χ2 /2 σ’s from
the mean, on one side of the distribution. For example, if ∆χ2 = 9 then a negative value will be excluded at 1.5σ’s on one side, yielding a probability of
successfully determining the hierarchy of 93.3%, considerably less than the 99.7% that one may naively
suspect just by taking the square root of ∆χ2 . More
3
generally, the probability of correctly determining
the hierarchy is
pc (∆χ2 ) =
s
1
1 + erf
2
∆χ2
.
8
(9)
In a more standard terminology, pc is the sensitivity to the hierarchy of the binary classification test
whose classification function is the sign of ∆χ2 . We
will refer to it simply as the “probability of success”
in what follows. In Ref. [5] the authors obtained a
similar result which differs as a result of their formula (5.11) for the probability of success for a given
∆χ2 .
If instead the inverted hierarchy is correct, the
calculation proceeds identically. As we have approximated σi to be hierarchy independent, the probability of success is identical for both hierarchies. This
is the quantity quoted in a number of studies such
as Refs. [6–8].
Second, what is the sensitivity of a typical experiment to the hierarchy?
A “typical experiment” is one in which |∆χ2 | obtains its average value |∆χ2 |. As the probability
of successfully determining the hierarchy is a monotonic function of ∆χ2 , the average value of ∆χ2
corresponds to the median value of the probability of success and so we will refer to such experiments as median experiments. The sensitivity of
the sign(∆χ2 ) test to the hierarchy is the probability that a fit to the correct hierarchy yields a lower
value of χ2 than one to the wrong hierarchy. Since
|∆χ2 | is fixed, this is simply the probability that
∆χ2 has the correct sign.
Again the calculation will proceed identically for
both hierarchies, so we may restrict our attention to
the case in which the normal hierarchy is correct.
Therefore the question is, given that ∆χ2 is positive
and ∆χ2 is equal to either ∆χ2 or −∆χ2 , what is
the probability pv that ∆χ2 = ∆χ2 .
Let L± be the likelihood, given the normal hierarchy, that ∆χ2 = ±∆χ2 , which is easily found
using the fact that ∆χ2 obeys a normal distribution
p
centered at ∆χ2 with standard deviation 2 ∆χ2 .
Using the fact that the distribution of ∆χ2 is odd
with respect to a change in the hierarchy, the Bayes
factor for the normal hierarchy is
L+
2
= e∆χ /2 .
L−
(10)
In particular, symmetric Bayesian priors assigning a
50% chance to each hierarchy yield a probability of
success of
pv =
1
L+
=
L+ + L−
1 + e−∆χ2 /2
(11)
for median experiments, those in which |∆χ2 | =
|∆χ2 |. For example, if ∆χ2 = 9 then the probability
that a median experiment correctly determines the
hierarchy will be 98.9%. While this is better than
the mean probability of success 93.3%, it still falls
noticeably short of the 99.7% which one might expect from Wilks’ theorem. In Ref. [9] it was noted
that the sensitivity (11) is equal to the posterior
probability of determining the correct hierarchy.
Given ∆χ2 determined either from Monte Carlo
simulations or from Asimov data, one may express
the sensitivity to the hierarchy expected at a median
experiment in terms of a number s of standard deviations σ. We will convert probabilities into standard
deviations using the one-sided Gaussian distribution
s
1
.
(12)
1 + erf √
pv (∆χ2 ) =
2
2
While the double-sided Gaussian is also often used
in the literature, we have checked that this choice of
convention has a small effect on our results.
Using Eq. (11) one now finds that the number of
σ’s of sensitivity is
!
2
√
1 − e−∆χ /2
−1
2
s(∆χ ) = 2 erf
.
(13)
1 + e−∆χ2 /2
This function is plotted in Fig. 1. For example, if
∆χ2 = 9 then a median experiment determines the
hierarchy with a sensitivity of 2.3σ instead of the 3σ
which might be expected. Had we insteaded opted
for the double-sided Gaussian convention for s, we
would have instead found 2.5σ.
A general Bayesian prior of b and 1 − b for the
normal and inverted hierarchies leads to a sensitivity
−∆χ2 /2 !
1
√
1
+
1
−
−1
b e
. (14)
s(∆χ2 ) = 2 erf
1
1 + b − 1 e−∆χ2 /2
4
s
7
6
5
4
3
2
1
10
20
30
40
50
DΧ2
FIG. 1: For a given ∆χ2 statistic determined from theoretical spectra, the black curve is the number s of σ’s
of sensitivity of the determination of the mass hierarchy
by a median experiment. A median experiment is one in
which |∆χ2 | obtains its median value. For comparison,
the dashed curve uses the two-sided definition of s and
the red curve is the square root of ∆χ2 .
Third, what is the probability p(s) that the hierarchy
will be determined with a sensitivity of at least sσ?
Note first that for a general experimental outcome
∆χ2 , the probability of success
pv =
=
L+
L+ + L−
2 −∆χ2 )2 /8∆χ2
e−(∆χ
e−(∆χ2 −∆χ2 )2 /8∆χ2 + e−(∆χ2 +∆χ2 )2 /8∆χ2
1
=
(15)
1 + e−∆χ2 /2
FIG. 2: The black, red, blue, purple and green curves
are the probability of a hierarchy determination with 1σ,
2σ, 3σ, 4σ and 5σ of sensitivity as a function of ∆χ2 .
The dashed line represents a median experiment, and its
intersections with the curves yield the same information
as Fig. 1.
responds not to a point but to a K-dimensional subset of the N -dimensional vector space in which y
is valued. The nuisance parameters xi are coordinates on these subsets. If the standard deviations σi
vary sufficiently slowly, then the inverse covariance
matrix defines a metric on this space. Recall that,
in the case of the normal (inverted) hierarchy, the
nuisance parameters x are chosen to minimize χ2N
(χ2I ). Geometrically, this minimization corresponds
to choosing the point in each subset which is closest to y, the coordinates of the point are the nuisance parameters which minimize the corresponding
χ2 statistic.
is independent of ∆χ2 . Using this fact, an argument
similar to those above leads to
In this framework, it is easy to combine data from
multiple
experiments. They can simply be added to
s
2
∆χ − 4arctanh erf √2
1
y
as
new
components. For example, one can combine
q
1 + erf
.
p(s) =
2
2
a
forecast
spectrum of Daya Bay II with a value of
8∆χ
the nuisance parameter θ13 determined at Daya Bay
(16)
and RENO by letting the first N − 2 components of
y correspond to the νe spectrum at Daya Bay II and
This function is plotted in Fig. 2.
the next two to the relative survival probabilities obParallel nuisance parameters
served at Daya Bay and RENO. The single nuisance
In reality there is no single experimental result
parameter θ13 yields a curve in the N -dimensional
yN or yI which is predicted by a given hierarchy.
space of observations for each hierarchy. The curve
The results also depend on a number of nuisance
is parameterized by θ13 . The last two coordinates of
parameters, such as the neutrino mass matrix pathis curve are simply the relative survival probabilirameters and the flux normalization of the source.
ties expected at Daya Bay and RENO as a function
We will assemble these nuisance parameters into a
of the parameter θ13 . The χ2 to be minimized is the
vector x = {xi }.
distance in the full N dimensional space, so it automatically combines determinations of θ13 at RENO,
If the final data consists of N numbers, such as
Daya Bay and Daya Bay II without the need for any
the number of events in N energy bins, and if there
penalty terms.
are K nuisance parameters, then each hierarchy cor-
5
Now let us make two approximations. First, we
approximate yN and yI to be linear (or affine) functions of the nuisance parameters x, so that the subspaces corresponding to theoretical predictions are
hyperplanes. The resulting models are called linear regression models. Model selection in one dimensional non-nested linear regression models was
first studied in Ref. [10]. Ref. [3] presented a statistic, generalizing ∆χ2 , which is Gaussian distributed
and distinguishes the models. The properties of this
statistic, in the case of linear regression models, were
determined in Ref. [11].
One may object that the spectra are not indeed
linear functions of the neutrino mass matrix. However the essential point is that they be approximately
linear in a regime whose size is the precision to which
an experiment can determine the nuisance parameters. This is a much easier criterion. Later we
will compare our analytical results to simulations in
which no such approximation is made, and we will
see that the resulting error is small.
For now we will make the further approximation
that one obtains the same value of ∆χ2 for any value
of the nuisance parameters chosen for the normal hierarchy if the nuisance parameters for the inverse
hierarchy are chosen so as to minimize ∆χ2I . In
other words, ∆χ2 , is independent of the choice of
the nuisance parameters so long as each χ2 is properly minimized. Geometrically this means that the
hyperplanes corresponding to the theoretical values
yN and yI are parallel.
Again assume that the normal hierarchy is correct. If xT is the true value of the nuisance parameters, then the theoretical values of the observables
yN will be linear functions yiN of xT . χ2N (χ2I ) is just
the minimum distance squared from the observations
yi = yiN + σi gi to the hyperplane corresponding to
the normal (inverted) hierarchy. The statistical fluctuation vector g = σi gi can be decomposed into a
two vectors, g⊥ and gk such that g⊥ is perpendicular to the hyperplanes and gk is parallel.
To determine χ2N or χ2I , one must choose the nuisance parameters x at which it is minimized. χ2 will
be minimized for the choice of nuisance parameters
xT + gk . In other words, the parallel part of g yields
the statistical error in the determination of the nuisance parameters. We have assumed that this error
is the same for both hierarchies. For this choice of
FIG. 3: In this figure the hierarchy is normal and ∆χ2 is
independent of the nuisance parameters. The two parallel lines are the expected measurements corresponding to
various values of the nuisance parameters for the two hierarchies. As a result of statistical fluctuations yiN + σi gi
is measured instead of the theoretical value yiN . The parallel part of g determines the effect of this fluctuation on
the best fit nuisance parameters and the perpendicular
part its effect on ∆χ2 .
nuisance parameters, the theoretical predictions for
k
k
yi are yiN + σi gi and yiI + σi gi in the cases of the two
hierarchies.
Now we are ready to calculate
∆χ2 = χ2I − χ2N
(17)
k
X (yi − y I − σi g )2 X (yi − y N − σi g k )2
i
i
i
i
=
−
2
2
σ
σ
i
i
i
i
=
X (y N + σi g ⊥ − y I )2 − (y N + σi g ⊥ − y N )2
i
i
i
i
=
X (y N − y I )2
i
i
σi2
i
=
i
i
i
σi2
X (y N − y I )2
i
i
σi2
i
+
X 2(y N − y I )
i
i
+
i
σi
X 2(y N − y I )
i
i
i
σi
gi⊥ .
gi .
In the last step we used the identity
X 2(y N − y I )
i
i
i
σi
k
gi = 0
(18)
which follows from the fact that, using the metric
1/σi2 , the vector (yiN − yiI ) is perpendicular to the
k
hyperplanes and so to σi gi .
Just as in Eq. (6), Eq. (17) describes a normal
2
distribution
q centered at ∆χ and with standard deviation 2 ∆χ2 . As a result, Eqs. (9) and (13) for
6
the probability of success and number of σ’s of sensitivity in the median experiment remain correct.
General nuisance parameters
Of course, ∆χ2 does depend on the nuisance parameters, and so the hyperplanes corresponding to
the theoretical data are not parallel and the above
results are only approximate. This fact was first
noted in Ref. [3], where it was concluded that as a
result ∆χ2 is not normally distributed. Its distribution leptokurtic.
This observation can be intuitively understood as
follows. Imagine that ∆χ2 depends so strongly upon
the nuisance parameters that a 1σ change in the nuisance parameters can reduce the sensitivity to the
hierarchy by several σ’s. As a result, most of the
experiments in which the hierarchy determination is
incorrect will be those in which the nuisance parameter is such that ∆χ2 is much smaller. Thus the
tails of the distribution of ∆χ2 will grow as a result
of those simulations in which the nuisance parameters take a nonstandard value. Clearly, this effect
is only present in simulations in which the nuisance
parameters are allowed to vary, and so simulations
that fix the nuisance parameters will yield values of
∆χ2 which, upon using Eq. (13), overestimate the
sensitivity to the hierarchy.
In Ref. [3] the author proposed a new statistic
which does follow a Gaussian distribution even in
this more general setting. However, in the case of
the hierarchy determinations planned in the near future, the angle between the hypersurfaces is actually
quite small. This is reflected in the observation [13]
that even a 1σ variation in θ13 only leads to about a
one third of a σ variation in the confidence. Therefore the approximate treatment of the ∆χ2 statistic
above is quite precise.
To illustrate this point, in Fig. 4 we present the
distribution of the ∆χ2 statistic in simulations which
combine the ν e spectrum measured at Daya Bay II
with MINOS’ 4% determination of the atmospheric
mass difference [12] and also with an optimistic 1%
forecast determination at an upgraded NOνA . All of
2
the nuisance parameters are fixed except for |∆M32
|,
2
2
which is chosen to minimize χI and χN as described
above. Following [14] we have considered 6 years
of exposure at a 20 kton detector for Daya Bay II
which detects ν e via inverse β decay on the 10% of
FIG. 4: The distribution of ∆χ2 in 50,000 experiments
with each hierarchy is shown, combining the data with
MINOS’ 4% determination of the atmospheric mass
splitting (red curve) and with an optimistic 1% determination at NOνA (black curve). The dashed curves
are the corresponding
q Gaussian distributions centered
at ∆χ2 with width 2
∆χ2 .
its mass consisting of free protons. The baselines
and reactor fluxes are identical to Ref. [14]. The
leptonic CP-violating angle δ is set to π/2.
We find that the distribution of ∆χ2 is indeed well
approximated by a Gaussian distribution
centered
q
at ∆χ2 with standard deviation 2 ∆χ2 . ∆χ2 ∼
11 (20) for Daya Bay II with MINOS (NOνA) yielding 2.6σ (3.9σ) of sensitivity at the median experiment, with a rate of successfully determining the
hierarchy of 94.6% (98.5%) in good agreement with
Eq. (9).
In Fig. 5 we present the distribution of ∆χ2 in simulations in which δ = 0 and π, although we always
fit to a δ = π/2 theoretical mode as the appearance mode at T2K and NOνA cannot distinguish 0
and π [13, 15]. At δ = 0 (π) we find ∆χ2 = 17
(22) yielding 3.5σ (4.2σ) of sensitivity, confirming
the expectations of Ref. [16]. Despite the fact that
the model used for fitting differs from that used to
generate the data, the distribution of ∆χ2 described
in this paper approximates the simulated data well.
Frequentist confidence
A frequentist notion of confidence can be made
well defined even in this context [17, 18]. Imagine
that an experiment measures ∆χ2 . This differs from
the expected ∆χ2 for the normal (inverted) hierarchy by |∆χ2 ∓ ∆χ2 | which corresponds to a frequentist incompatibility of
|∆χ2 ∓ ∆χ2 |
q
2 ∆χ2
(19)
7
FIG. 5: As in Fig. 4, but using only a 1% determination of the atmospheric mass splitting. The simulations
reported in the red and black curves use δ = 0 and π
respectively, although the fitting is always performed assuming δ = π/2. As can be seen, if δ = π, the hierarchy
determination will be more reliable [13, 14].
σ’s.
In particular, in the case of the median experiment
with the true hierarchy, ∆χ2 = ∆χ2 . Therefore
thepinverted hierarchy is excluded at a confidence
of ∆χ2 σ’s. In this sense it might be tempting
to ignore the results of this paper
p and to identify
the frequentist incompatibility ∆χ2 with the confidence in the hierarchy determination expected in a
median experiment.
While such a definition of confidence is welldefined, it has a very unattractive feature. Consider an experiment with an expected ∆χ2 = 16.
The general arguments in this note imply that if
the hierarchy is normal (inverted) then ∆χ2 will follow a Gaussian distribution centered on 16 (−16)
with a width of σ = 8. In the frequentist sense,
the median experiment will yield |∆χ2 | = 16 and so
is incompatible with the false hierarchy with 4σ of
confidence while the 98th percentile experiment will
yield ∆χ2 = 0 and so is incompatible with the false
hierarchy with 2σ of confidence. An identification of
the sensitivity to the hierarchy with the frequentist
incompatibility would therefore imply that even the
98th percentile of experimental outcomes will yield
a 2σ sensitivity to the hierarchy.
Now consider the somewhat unlikely case in which
due to statistical fluctuations, the results of this experiment are indeed in the 98th percentile, so that
∆χ2 = 0. Now the experimentalist will be asked
to provide the hierarchy with 2σ of confidence. Of
course he cannot, the experiment has not yielded
any preference for either hierarchy, even at the 2σ
level that was promised for a 98th percentile experi-
ment when the funding was requested. In this sense,
the identification of the frequentist incompatibility
with the confidence in the hierarchy determination
is misleading: the confidence can be nonzero even
when no information is obtained.
The basic problem with the application of the frequentist notion of confidence in this example is that
both hierarchies have been ruled out with equal confidence. Ruling out both hierarchies can be useful
when searching for new physics, testing assumptions
regarding backgrounds, etc. Although in that case
one would use a χ2 test and not a ∆χ2 test, as the
latter is insensitive to effects that affect both hierarchies similarly. However, for the purpose of determining which hierarchy is manifested in nature it is
reasonable to assume that one of the hierarchies is
indeed correct. In this case one is led to the Bayesian
constructions described in this note.
Acknowledgement
We have benefited beyond measure from correspondence with Xin Qian, Yoshitaro Takaesu and Tyce
DeYoung. We thank Amir Nawaz Khan for identifying two typos in an earlier version of this paper.
JE is supported by the Chinese Academy of Sciences
Fellowship for Young International Scientists grant
number 2010Y2JA01. EC and XZ are supported in
part by the NSF of China.
[1] X. Qian, A. Tan, W. Wang, J. J. Ling, R. D. McKeown and C. Zhang, “Statistical Evaluation of
Experimental Determinations of Neutrino Mass
Hierarchy,” Phys. Rev. D 86 (2012) 113011
[arXiv:1210.3651 [hep-ph]].
[2] S. S. Wilks, “The Large-Sample Distribution of the
Likelihood Ratio for Testing Composite Hypotheses” Ann. Math. Stat. 9, No. 1 (1938), 60-62.
[3] D. R. Cox, “Tests of Separate Families of Hypotheses,” Proc. Fourth Berkeley Symp. on Math. Statist.
and Prob. 1 (1961), 105-123.
[4] M. Ribordy and A. Y. Smirnov, “Improving the
neutrino mass hierarchy identification with inelasticity measurement in PINGU and ORCA,”
arXiv:1303.0758 [hep-ph].
[5] S. -F. Ge, K. Hagiwara, N. Okamura and
Y. Takaesu, “Determination of mass hierarchy with
8
[6]
[7]
[8]
[9]
[10]
[11]
[12]
medium baseline reactor neutrino experiments,”
arXiv:1210.8141 [hep-ph].
L. Zhan, Y. Wang, J. Cao and L. Wen, “Experimental Requirements to Determine the Neutrino Mass
Hierarchy Using Reactor Neutrinos,” Phys. Rev. D
79 (2009) 073007 [arXiv:0901.2976 [hep-ex]].
X. Qian, D. A. Dwyer, R. D. McKeown, P. Vogel, W. Wang and C. Zhang, “Mass Hierarchy
Resolution in Reactor Anti-neutrino Experiments:
Parameter Degeneracies and Detector Energy Response,” PRD, 87, 033005 (2013) [arXiv:1208.1551
[physics.ins-det]].
E. Ciuffoli, J. Evslin and X. Zhang, “Mass Hierarchy Determination Using Neutrinos from Multiple
Reactors,” JHEP 1212 (2012) 004 [arXiv:1209.2227
[hep-ph]].
M. Blennow, “On the Bayesian approach to neutrino
mass ordering,” arXiv:1311.3183 [hep-ph].
H. Hotelling, “The Selection of Variates for Use
in Prediction with Some Comments on the General Problem of Nuisance Parameters,” Ann. Math.
Statist. Volume 11, 3 (1940) 271-283
M. H. Pesaran, “On the General Problem of Model
Selection,” Rev. of Econ. Stud. 41 (1974) 153-171.
H. Nunokawa, S. J. Parke and R. Zukanovich Funchal, “Another possible way to determine the neutrino mass hierarchy,” Phys. Rev. D 72 (2005)
013009 [hep-ph/0503283].
[13] E. Ciuffoli, J. Evslin and X. Zhang, “Optimizing
Medium Baseline Reactor Neutrino Experiments,”
arXiv:1302.0624 [hep-ph].
[14] Y. -F. Li, J. Cao, Y. Wang and L. Zhan, “Unambiguous Determination of the Neutrino Mass Hierarchy Using Reactor Neutrinos,” arXiv:1303.6733
[hep-ex].
[15] H. Minakata and S. J. Parke, “Correlated, Precision Measurements of θ23 and ∆ using only
the Electron Neutrino Appearance Experiments,”
arXiv:1303.6178 [hep-ph].
[16] H. Minakata, H. Nunokawa, S. J. Parke and
R. Zukanovich Funchal, “Determining neutrino
mass hierarchy by precision measurements in
electron and muon neutrino disappearance experiments,” Phys. Rev. D 74 (2006) 053008
[hep-ph/0607284].
[17] Y. Takaesu, “Mass Hierarchy Using Medium Baseline Detector,” presented at the International Workshop on RENO-50.
[18] A. U. Abeysekara, R. Alfaro, C. Alvarez,
J. D. lvarez, R. Arceo, J. C. Arteaga-Velzquez,
H. A. A. Solares and A. S. Barber et al., “Sensitivity of the High Altitude Water Cherenkov Detector to Sources of Multi-TeV Gamma Rays,”
arXiv:1306.5800 [astro-ph.HE].
| 10math.ST
|
T. S. Rappaport, G. R. MacCartney, Jr., S. Sun, H. Yan, S. Deng, “Small-Scale, Local Area, and Transitional Millimeter Wave Propagation
for 5G Communications,” in IEEE Transactions on Antennas and Propagation, Special Issue on 5G, Nov. 2017.
Small-Scale, Local Area, and Transitional Millimeter Wave
Propagation for 5G Communications
Theodore S. Rappaport, George R. MacCartney Jr., Shu Sun, Hangsong Yan, and Sijia Deng
arXiv:1707.07816v2 [cs.IT] 15 Aug 2017
Abstract—This paper studies radio propagation mechanisms that
impact handoffs, air interface design, beam steering, and MIMO for
5G mobile communication systems. Knife edge diffraction (KED)
and a creeping wave linear model are shown to predict diffraction
loss around typical building objects from 10 to 26 GHz, and human
blockage measurements at 73 GHz are shown to fit a double knifeedge diffraction (DKED) model which incorporates antenna gains.
Small-scale spatial fading of millimeter wave received signal voltage
amplitude is generally Ricean-distributed for both omnidirectional
and directional receive antenna patterns under both line-of-sight
(LOS) and non-line-of-sight (NLOS) conditions in most cases, although the log-normal distribution fits measured data better for the
omnidirectional receive antenna pattern in the NLOS environment.
Small-scale spatial autocorrelations of received voltage amplitudes
are shown to fit sinusoidal exponential and exponential functions for
LOS and NLOS environments, respectively, with small decorrelation
distances of 0.27 cm to 13.6 cm (smaller than the size of a handset)
that are favorable for spatial multiplexing. Local area measurements
using cluster and route scenarios show how the received signal
changes as the mobile moves and transitions from LOS to NLOS
locations, with reasonably stationary signal levels within clusters.
Wideband mmWave power levels are shown to fade from 0.4 dB/ms
to 40 dB/s, depending on travel speed and surroundings.
Index Terms—Millimeter wave, diffraction, human blockage,
small-scale fading, spatial autocorrelation, propagation, channel transition, mobile propagation, MIMO, spatial consistency
I. I NTRODUCTION
Driven mainly by the pervasive usage of smartphones and
the emergence of the Internet of Things (IoT), future 5G mobile
networks will become as pervasive as electrical wiring [1] and will
offer unprecedented data rates and ultra-low latency [2]–[4]. For
the first time in the history of radio, millimeter-wave (mmWave)
frequencies will be used extensively for mobile and fixed access,
thus requiring accurate propagation models that predict how the
channel varies as people move about. Remarkable progress has
been made in modeling large-scale propagation path loss at
mmWave frequencies [4]–[14], and it is well understood that for
an assumption of unity gain antennas across all frequencies, Friis
equation predicts that path loss is greater at mmWave compared to
today’s UHF/microwave cellular systems [4]–[6], [8]–[10], [15].
Also, rain and atmospheric attenuation are well understood, and
reflection and scattering are more dominant than diffraction at
mmWave bands [4], [16]–[19].
Broadband statistical spatial channel models (SSCMs) and
simulators that faithfully predict the statistics of signal strength,
and the number and direction of arrival and departure of multipath
components, have been developed by a consortium of companies
and universities [20] and from measurements in New York City
This material is based upon work supported by the NYU WIRELESS Industrial
Affiliates program, NSF research grants 1320472, 1302336, and 1555332, and
funding from Nokia. The authors acknowledge Amitava Ghosh of Nokia for his
support of this work. The authors thank Yunchou Xing, Jeton Koka, Ruichen Wang,
and Dian Yu for their help in conducting the measurements. The authors also thank
Prof. Henry Bertoni for his insights and advisement in diffraction theory.
T. S. Rappaport (email: [email protected]), G. R. MacCartney, Jr., S. Sun, H. Yan,
and S. Deng are with NYU WIRELESS Research Center, NYU Tandon School of
Engineering, 9th Floor, 2 MetroTech Center, Brooklyn, NY 11201.
[21]. These models are being used to develop air-interfaces for
5G systems [22] [23]. Elsewhere in this issue, [4] summarizes
standard activities for large-scale mmWave channel modeling.
Little is known, however, about the small-scale behavior of
wideband mmWave signals as a mobile user moves about a local
area. Such information is vital for the design of handoff mechanisms and beam steering needed to rescue the communication
link from deep fades. In this paper, propagation measurements
investigate diffraction, human blocking effects, small-scale spatial
fading and autocorrelation, local area channel transitions, and
stationarity of signal power in local area clusters at frequencies
ranging from 10 to 73 GHz. Diffraction measurements for indoor
and outdoor materials at 10, 20, and 26 GHz are presented
in Section II, and two diffraction models, i.e., the knife edge
diffraction (KED) model and a creeping wave linear model, are
used to fit the measured results. We predict the rapid signal
decay as a mobile moves around a diffracting corner. In Section
III, measurements at 73 GHz are presented and a double knifeedge diffraction (DKED) antenna gain model that uses directional
antenna patterns is shown to describe minimum and maximum
fade depths caused by human blockage. Small-scale fading and
correlation studies at 73 GHz are presented in Section IV, where
small-scale fading distributions and spatial autocorrelations of
received voltage amplitudes in LOS and NLOS environments
with omnidirectional and directional antennas are provided and
analyzed. In Section V, route and cluster scenarios are used to
study local area channel transitions and stationarity, where analysis
for channel transition from a NLOS to a LOS region and local
area path loss variations are provided. Conclusions are given in
Section VI. Channel models given here may be implemented
for small-scale propagation modeling and real-time site-specific
mobile channel prediction and network control [24].
II. D IFFRACTION M EASUREMENTS AND M ODELS
A. Introduction of Diffraction Measurements and Models
Accurate characterization of diffraction at cmWave and
mmWave frequencies is important for understanding the rate of
change of signal strength for mobile communications since future
5G mmWave systems will have to rely less on diffraction as a
dominant propagation mechanism [6], [8]. Published indoor and
outdoor diffraction measurements show that diffraction has little
contribution to the received signal power using various materials
and geometries (edges, wedges, and circular cylinders) at 60 GHz
and 300 GHz [25]–[27]. It was shown that the KED model agreed
well with diffraction measurements for cuboids at 300 GHz [28],
vegetation obstacles at 2.4, 5, 28, and 60 GHz [29], and for
human blocking at 60 GHz [13]. Apart from the KED model,
uniform theory of diffraction (UTD) models are also used. An
overview of the Geometrical Theory of Diffraction (GTD) and
the UTD are provided in [30], as well as their utility to solve
practical problems. Besides the KED and UTD models, Mavridis
et al. presented a creeping wave linear model [31] to estimate
the diffraction loss by a perfectly conducting or lossy circular
cylinder for both transverse-magnetic (TM) and transverse-electric
(TE) polarizations at 60 GHz. In the following sub-sections, we
describe diffraction measurements conducted in 2015 around the
engineering campus of New York University [17], where the
frequency dependency of diffraction at 10, 20, and 26 GHz in
realistic indoor and outdoor scenarios was investigated to yield
simple yet accurate diffraction models for wireless planning.
B. Diffraction Measurement System
Diffraction measurements were performed by transmitting a
continuous wave (CW) signal generated by an Agilent E8257D
PSG analog signal generator through a pyramidal horn antenna at
the transmitter (TX). An identical horn antenna was used at the
receiver (RX) to receive signal energy around a corner test material
(e.g. a stone pillar). The RX antenna was fed to an E4407B
ESA-E spectrum analyzer that measured received power which
was subsequently recorded on a laptop with LabVIEW software.
During the measurements, the TX antenna was set one meter from
the knife edge, sufficiently in the far field, and was fixed to a tripod
and aimed at the knife edge, whereas the RX antenna was set 2
meters from the knife edge (also in the far field) and was fixed
on a rotatable gimbal attached to a translatable linear track that
was made to from an approximate arc around the knife edge (see
Fig. 1). Diffraction loss was measured at 10, 20, and 26 GHz
using identical pairs of antennas at the TX and RX to measure
each frequency, separately. For each frequency, Table I lists the
flange type, antenna gain, and half-power beamwidth (HPBW) of
the antenna pairs used. The TX and RX were stationed at a wide
range of angles, both in the lit and shadowed region, and the horns
always had their boresights focused on the corner knife-edge of
the indoor and outdoor materials that were studied. More details
regarding the measurement system are given in [17], [32].
C. Diffraction Measurement Description
The indoor measurements were performed at 90◦ (right-angle)
wall corners made of drywall, wood, and semi-transparent plastic
board with 2 cm thickness. Outdoor measurements studied one
rounded stone pillar corner and one right-angle marble building
corner. During the measurements, the TX and RX were placed
on either side of the corner (knife edge) of the test material. A
diagram of the corner diffraction geometry is shown in Fig. 1
where d1 is the distance between the TX and the corner knifeedge, and d2 is the distance between the corner knife-edge and the
RX. Both d1 (1 m) and d2 (2 m) remained constant throughout
the diffraction measurement campaign. The β and α values are the
incident and diffraction angles, respectively, where two (outdoor)
or three (indoor) fixed values between 10◦ and 39◦ were chosen
for β. The RX antenna was mounted on a motorized linear track
(see Fig. 1) that translated in step increments of 0.875 cm, which
corresponds to approximately a 0.5◦ increment in diffraction angle
(α) for each step increment. At each step increment, the RX
antenna was adjusted to point perfectly towards the knife-edge
TABLE I: Antenna parameters used diffraction measurements.
Measured
Frequency
10 GHz
20 GHz
26 GHz
Flange
Type
WR-75
WR-51
WR-28
Antenna
Gain
20 dBi
20 dBi
24.5 dBi
HPBW
(Az./El.)
17◦ /17◦
17◦ /17◦
10.9◦ /8.6◦
Far Field
Distance
0.47 m
0.46 m
0.83 m
corner. The length of the track was 35.5 cm and was used to
measure a 20◦ swath of diffraction angles over the entire length
of the track. At each measurement location, five consecutive linear
tracks (see Fig. 1) were used to provide a 100◦ diffraction angle
arc around the corner which covered a broad range of the shadow
region where the TX antenna is shadowed by the corner object
with respect to RX antenna. Additionally, a smaller range of
diffraction angles was measured in the lit region where the TX
and RX antennas were in view of each other but were not pointed
at each other since they were always aimed at the corner.
At each location, prior to the diffraction measurements, a free
space calibration in an open area with both antennas pointed at
each other on boresight was conducted with a 3 m (d1 + d2 = 3
m) transmitter-receiver (T-R) separation distance to provide a free
space power reference for each frequency. The diffraction loss was
then obtained by calculating the difference between the measured
received signal power at each step increment of the RX antenna
during the diffraction measurements and the free space calibration
received power (the TX and RX antenna gains were deducted from
all power measurements).
D. Theoretical Diffraction Models
1) KED Model: The KED model is suitable for applications
with sharp knife edges and has a simple form yet high prediction
accuracy [33]. In general, diffraction loss over complex and
irregular obstructions can be difficult to calculate, but typical
obstructions for 5G wireless will involve common building partitions which are generally simple in nature and with dimensions
that appear infinite at such small wavelengths, such as a wall or
building corner, thus justifying simple diffraction models.
The diffraction loss (as compared to free space) is obtained by
calculating the electric field strength Ed [V/m] at the RX based on
the specific Fresnel diffraction parameter ν [34]. The ratio of Ed
and the free space field strength E0 can be computed by summing
all the secondary Huygens’ sources in the knife edge plane and
is given by [17], [34]:
1+j
Ed
= F (ν) =
E0
2
∞
Z
2
e−j(π/2)t dt
(1)
ν
where F (ν) is the complex Fresnel integral and ν is the Fresnel
diffraction parameter is defined as [34]:
s
ν=h
2(d1 + d2 )
=α
λd1 d2
s
2d1 d2
λ(d1 + d2 )
(2)
RX
Track
TX
β
h
d2
d1
RX Locations
d2
d1
TX Location
α
RX Linear Track
RX
Track
Fig. 1: Top view of the corner diffraction geometry [17].
E(α, d2 , k) ∼ Ei e−jkαRh √
kd2
∞
X
Dp Rh · exp(−ψp α)
(5)
The expression for the diffraction power loss in dB based on the
creeping wave linear model is given by [17], [32]:
(6)
In order to facilitate the computation of G(α), a simple linear
model (7) based on minimum mean squared error (MMSE) estimation between the model and measured data was proposed in [36]
to estimate the diffraction loss caused by a curved surface at a
single frequency, based on the creeping wave linear model [17],
[32]:
P (α) = n · α + c
(7)
where n is the linear slope of diffraction loss calculated by MMSE
for each specific frequency and object radius, and c is the anchor
point set to 6.03 dB, the diffraction loss estimated by KED with
diffraction angles α = β = 0◦ .
E. Indoor Diffraction Results and Analysis
The indoor diffraction loss measurements for the drywall corner,
wooden corner, and plastic board are plotted with the KED model
(3) at 10, 20, and 26 GHz as a function of the diffraction angle in
Fig. 2. Three different TX incident angles were used to measure
diffraction loss for each frequency. Since the measurements in [17]
were conducted with the TX and RX along a constant radius
Shadow Region
Lit Region
30
20
0
-40
-30
-20
-10
0
10
20
30
40
50
60
70
Diffraction Angle α (°)
70
Relative Loss (dB)
60
50
40
Measurement Data at 10, 20, & 26 GHz for Wooden Corner
w/ V-V Polarization for All Incident Angles
Relative Loss at 10 GHz
Relative Loss at 20 GHz
Relative Loss at 26 GHz
KED at 10 GHz
KED at 20 GHz
KED at 26 GHz
Shadow Region
Lit Region
30
20
10
0
-40
-30
-20
-10
0
10
20
30
40
50
60
Diffraction Angle α (°)
70
60
where Ei is the incident field from the TX that impinges upon the
obstruction, Rh is the radius of cylinder for the diffraction corner,
k is the wave number of the carrier frequency, α is the diffraction
angle, d2 is the distance between the launch point at the rounded
corner edge and the RX, Dp is the excitation coefficient and ψp
is the attenuation constant. Due to the computational complexity
of (4), a reasonable approximation for E on a flat surface can be
obtained by considering only the p = 1 term in (4), which is given
by [17], [31], [32], [35], [36]:
G(α)[dB] = −P (α)[dB] = 20 log10 E
40
Relative Loss at 10 GHz
Relative Loss at 20 GHz
Relative Loss at 26 GHz
KED at 10 GHz
KED at 20 GHz
KED at 26 GHz
10
(4)
p=1
E ∼ Ei Dp Rh · exp(−ψp α)
50
Measurement Data at 10, 20, & 26 GHz for Drywall Corner
w/ V-V Polarization for All Incident Angles
(3)
where P (ν) is the power loss of the diffracted signal for the value
of ν, compared to free space case for the same distance.
2) Convex Surfaces based Diffraction Model: Although the
KED model has broad applications for various geometries, it
requires the diffraction corner to be in the shape of a sharp
knife edge and does not account for the radius of curvature of an
obstacle. When a diffraction corner is rounded in shape, such as
a stone pillar corner (it resembles a circular cylinder), a creeping
wave linear model can better predict diffraction loss [31], [35]. A
creeping ray field at the RX antenna behind the circular object for
an incident plane wave is given by [35]:
e−jkd2
60
Relative Loss (dB)
G(ν)[dB] = −P (ν)[dB] = 20 log10 |F (ν)|
70
Relative Loss (dB)
where λ is the wavelength, α is the diffraction angle, h is the
effective height (or width) of the obstructing screen with an infinite
width (or height) placed between the TX and RX at the distances
d1 and d2 , respectively, under the conditions that d1 , d2 h, and
d1 , d2 λ. These conditions were met for 10, 20 and 26 GHz
measurements in both the indoor and outdoor environments, as
shown in Fig. 1.
Based on (1) and (2), the diffraction power gain G(ν) in dB
produced in a knife edge by the KED model is expressed as [17],
[34]:
50
40
Measurement Data at 10, 20, & 26 GHz for Plastic Board
w/ V-V Polarization for All Incident Angles
Relative Loss at 10 GHz
Relative Loss at 20 GHz
Relative Loss at 26 GHz
KED at 10 GHz
KED at 20 GHz
KED at 26 GHz
Shadow Region
Lit Region
30
20
10
0
-40
-30
-20
-10
0
10
20
30
40
50
60
70
Diffraction Angle α (°)
Fig. 2: Diffraction measurements for the indoor drywall corner (top),
wooden corner (middle), and plastic board (bottom) compared to the KED
model at 10, 20 and 26 GHz [17].
(d1 and d2 , respectively) from the corner of each test material,
diffraction loss can be represented as a function of the diffraction
angle α in (2) for each frequency (or wavelength λ), without the
need for the TX incident angle.. Fig. 2 shows diffraction loss
increases to approximately 30 dB as compared to free space as
the RX antenna moves from the edge of the lit region (0◦ ) into the
shadow region (20◦ ) for the drywall corner and wooden corner,
respectively. This demonstrates the rapid signal degradation that
occurs when diffraction is the primary propagation mechanism in
mobile systems. Note the good fit between the drywall diffraction
measurements and the simple KED model in the early and deep
shadow regions for all three frequencies, while the KED model
overestimates diffraction loss by 5-10 dB compared with the
wooden corner diffraction measurements at 10, 20, and 26 GHz.
As for the plastic board material, the KED model overestimates
the measured diffraction loss at small diffraction angles near the
lit/shadow region boundary (from 0◦ to 30◦ ), and underestimates
diffraction loss in portions of the deep shadow region (diffraction
angles greater than 30◦ ), most noticeably at 20 GHz and 26 GHz.
Fig. 2 indicates slightly less loss occurs at lower frequencies,
implying frequency dependence, where divergence from pure
diffraction theory can be attributed to reflections and scattering
in the indoor environment and potential transmissions through the
test materials. In general, the observations match the KED model
trend. Oscillation patterns of the measured data in the shadow
region observed in Fig. 2 indicate that the measured diffraction
signal includes corner diffraction, penetration through the material,
and partial scattering in the measurement environment. We note
that the diffraction loss observed in the lit region is due to
the measurement procedure where the TX and RX antennas
were never aligned on boresight (except at 0◦ ) since they were
constantly pointed directly at the knife-edge corner.
Penetration loss was measured for typical building materials
in the same indoor environment at 73 GHz and showed that copolarized penetration loss ranged from 0.8 dB/cm (lowest loss
material – drywall) to 9.9 dB/cm (highest loss material – steel
door), with standard deviation σ about the average loss ranging
from 0.3 dB/cm (lowest σ – drywall) to 2.3 dB/cm (highest σ –
clear glass). Additional details can be found in Table II of [37].
F. Outdoor Diffraction Results and Analysis
The outdoor marble corner and stone pillar measurement results
are shown in Fig. 3 with the MMSE creeping wave linear models
(7) at 10, 20, and 26 GHz. It can be seen from Fig. 3 that
(7) predicts a linearly increasing diffraction loss into the deeply
shadowed region, as opposed to the leveling off seen in Fig. 2 from
(3). Fig. 3 also shows the outdoor stone pillar measurement results
70
Outdoor Measurement Data at 10, 20, & 26 GHz for Pillar
w/ V-V Polarization for All Incident Angles
Relative Loss (dB)
60
50
40
Lit Region
Shadow Region
Relative Loss at 10 GHz
Relative Loss at 20 GHz
Relative Loss at 26 GHz
n10 GHz = 0.75
30
20
10
n20 GHz = 0.88
n26 GHz = 0.96
0
-30
-20
-10
0
10
20
30
40
50
60
70
Diffraction Angle α (°)
Outdoor Measurement Data at 10, 20, & 26 GHz for Marble Corner
w/ V-V Polarization for All Incident Angles
Relative Loss (dB)
60
50
40
Lit Region
Shadow Region
30
Relative Loss at 10 GHz
Relative Loss at 20 GHz
Relative Loss at 26 GHz
n10 GHz = 0.62
20
10
n20 GHz = 0.77
n26 GHz = 0.96
0
-30
-20
-10
0
10
20
30
40
50
60
Diffraction Angle α (°)
Relative Loss vs Diffraction Angle at 10 GHz
for Pillar w/ V-V Polarization
Relative Loss (dB)
60
50
40
30
Diffraction Loss for 21° Incident Angle
Diffraction Loss for 35° Incident Angle
Knife Edge Diffraction Model at 10 GHz
MMSE Linear Fit with σ = 2.75 dB
Lit Region
at 10 GHz compared to the KED model and creeping wave linear
model, where the creeping wave linear model provides a better fit
to the measured relative diffraction loss than the KED model, in
the shadow region. Diffraction loss for each frequency is plotted
as a function of diffraction angle and includes the measured loss at
two TX incident angles. The measured data matches well with the
creeping wave linear model derived via MMSE for each frequency,
in the shadow region. The creeping wave linear model slopes are
0.75, 0.88, and 0.96 for the stone pillar measurements and 0.62,
0.77, and 0.96 for the marble corner measurements at 10, 20, and
26 GHz, respectively. Fig. 3 shows outdoor obstructions cause an
even greater loss when a mobile moves into a deeply shadowed
region, showing as much as 50 dB of loss when solely based on
diffraction. Based on the increase in slope values with frequency,
it is easily seen that diffraction loss increases with frequency
in the outdoor environment. The simple creeping wave linear
model (7) fits well with the measured data and has a much lower
standard deviation compared with the KED model [17], indicating
a good overall match between the creeping wave linear model and
measured data. Similar to the indoor environment diffraction loss
measurements, the high diffraction loss in the lit region is caused
by the TX and RX pointing off boresight towards the corner of
the test material.
When comparing the slope values of the two outdoor materials,
the rougher surface (stone) with a slightly rounded edge has a
greater slope (greater attenuation) for identical frequencies as
compared to the smoother surface (marble) straight edge. We note
that using the MMSE method to derive the typical slope values
instead of calculating the theoretical slope value provides system
engineers a useful parameter while reducing the computational
complexity of the diffraction model. The simple slope values are
useful for mobile handoff design since a mobile at 20 GHz would
see approximately 21 dB of fading when moving around a marble
corner from a diffraction angle of 0◦ to 20◦ . For a person moving
at a speed of 1 m/s, this results in about a 21 dB/s initial fade rate
from LOS to NLOS. Fig. 3 shows more oscillation patterns in the
marble corner measurements in the deeply shadowed region than
in the stone pillar measurements, which indicates more prevalent
scattering when measuring the marble corner. We note that the
diffraction loss for outdoor building corners at mmWaves using
directional antennas can be better predicted with a simple linear
model (creeping wave), whereas the KED model agrees well with
indoor diffraction loss measurements. Both the creeping wave and
KED models can be used in network simulations and ray-tracers
with short computation time and good accuracy while considering
approximately 5-6 dB standard deviation (see [17] for mean error
and standard deviation values between the measured data and
models derived from the data). For cross-polarized diffraction
measurement results, see [32].
III. H UMAN B LOCKAGE M EASUREMENTS AND M ODELS
Shadow Region
A. Introduction of mmWave Human Blockage
20
10
Diffraction Loss = 0.75 * α + 6.03 dB
0
-30
-20
-10
0
10
20
30
40
50
60
70
Diffraction Angle α (°)
Fig. 3: Measured diffraction loss for the outdoor stone pillar (top) and
marble corner (middle) with creeping wave linear models at 10, 20, and
26 GHz, and stone pillar measurement results compared to the KED and
the creeping wave linear models at 10 GHz (bottom) [17].
In mmWave communications, attenuation caused by human
blockage (when a human body blocks the LOS path between
a transmitter and receiver) will greatly impact cellphone link
performance, and phased array antennas will need to adapt to
find other propagation paths when blocked by a human [15], [38].
This is in sharp contrast to omnidirectional antennas used at sub-6
GHz frequencies. Understanding this severe blockage effect and
employing appropriate models for mobile system simulation are
important for properly designing future mmWave antennas and
beam steering algorithms [22], [39], [40]. One of the earliest
human blockage measurement studies [41] was conducted at 60
GHz for indoor wireless local area networks (WLAN) with T-R
separation distances of 10 m or less in a typical office environment.
Results showed that the signal level decreased by as much as 20
dB when a person blocked the direct path between omnidirectional
TX and RX antennas, with deep fades reaching 30 dB using
directive antennas [42], [43]. In addition to human blockage
measurements at 60 GHz, [44] provided a human induced cluster
blockage model based on ray tracing [45], a random walk model,
and a diffraction model from contributions to 802.11ad [46], [47].
The probability distributions for four parameters (duration, decay
time, rise time, and mean attenuation) were generated by the
human-induced cluster blockage model and were validated with
the Kolmogorov-Smirnov test [44]. The mobile and wireless communications enablers for the twenty-twenty information society
(METIS) and 3rd Generation Partnership Project (3GPP) also
proposed their own human blockage models [20], [48], [49] based
on KED models for one or multiple edges. In the following subsections, 73 GHz human blockage measurements and an improved
human blockage model are described.
B. Human Blockage Measurement System
A real-time spread spectrum correlator channel sounder system
described in [38] was used for the human blockage measurements.
A pseudorandom noise (PN) sequence of length 2047 was generated at baseband with a field programmable gate array (FPGA)
and high-speed digital-to-analog converter (DAC). This wideband
sequence was modulated to an intermediate-frequency (IF) of
5.625 GHz which was then upconverted to a center frequency of
73.5 GHz (1 GHz null-to-null RF bandwidth). The transmit power
at the TX was -5.8 dBm, and identical TX and RX horn antennas
with 15◦ azimuth and elevation (Az./El.) HPBW and 20 dBi of
gain were used. The received signal was downconverted to IF, and
then demodulated to its in-phase (I) and quadrature (Q) baseband
signals that were sampled at 1.5 Giga-Samples (GS/s) via a highspeed analog-to-digital converter (ADC). The digital signals were
then correlated in software via a Fast Fourier Transform (FFT)
matched filter to create the I and Q channel impulse response
(CIR), and subsequent power delay profiles (PDPs) (I 2 +Q2 ). The
system had a multipath resolution of 2 ns and an instantaneous
dynamic range of 40 dB and could capture PDPs with a minimum
consecutive snapshot interval of 32.752µs to measure rapid fading.
Power was computed as the area under the PDP, and the voltage
was found as the square root of power [34].
C. Blockage Measurement Environment Description
Human blockage measurements were conducted in an open
laboratory using a 5 m T-R separation distance. High gain narrowbeam horn antennas were used at the TX and RX with both
antenna heights set to 1.4 m relative to the ground and aligned
on boresight. Nine measurements were recorded with a human
blocker walking at a perpendicular orientation through the LOS
path between the TX and RX at an approximate 1 m/s speed. This
perpendicular walk was performed at 0.5 m increments between
the TX and RX starting at 0.5 m from the TX for measurement
one, and 4.5 m from the TX for measurement nine, as depicted
in Fig. 4. We note that 0.5 m is the typical distance for a
5m
1
TX
2
3
4
5
6
7
8
9
RX
0.5 m 1.0 m 1.5 m2.0 m 2.5 m 3.0 m 3.5 m 4.0 m 4.5 m
Fig. 4: Depiction of nine measurement locations where at each indicator
separated by 0.5 m, the human blocker walked at a perpendicular
orientation between the TX and RX antennas [38].
h1
Screen Blocker
w1
A
B
w2
h2
(a) 3D screen projection.
w1
A
αw1
D2w1
w
D1w1
TS
S
SR
D2w2
w2
αw2
B
D1w2
(b) Top view of screen projection.
Fig. 5: (a) 3D and (b) top-down projection of screen blocker.
person to view the screen of a smartphone. For each of the nine
perpendicular walks, 500 PDPs were recorded per second in a fivesecond window, resulting in 2500 PDPs for each measurement.
The dimensions of the human blocker were: bbreadth = 0.47 m;
bdepth = 0.28 m; bheight = 1.80 m. Detailed information about the
experiment is given in [38], [50].
D. KED Blockage Model
Knife-edge diffraction is commonly used to model human
blockage by modeling a thin rectangular screen as the blocker [20],
[48], [49], [51]. In DKED modeling, the rectangular screen is
considered infinitely high, such that diffraction loss only occurs
from the two side edges of the body. A typical screen blocker
with height h and width w is displayed in Fig. 5 from both a 3D
and top view screen projection. The dimensions for the top view
of the screen are defined as follows: w is the width of the screen
def
def
def
def
from w1 to w2; r = AB; w = w1w2; h = h1h2; T S = AS;
def
def
def
def
SR = SB; D2w1 = Aw1; D1w1 = w1B; D2w2 = Aw2;
def
D1w2 = w2B; αw1 and αw2 are the diffraction angles for the w1
and w2 edges of the screen, respectively [38]. From Fig. 5b, the
two side edges of the screen are denoted w1 and w2, where the
distance between the edges is the body depth (bdepth ) or width of
the screen w, since the blocker walks through the LOS path at a
perpendicular orientation. As the screen moves between the TX
and RX antennas and blocks the LOS path, the screen is always
considered perpendicular to the solid line drawn between the two,
to reduce computational complexity.
Diffraction loss is calculated via numerical approximation by
Total Rec. Power
Fresnel integration and the diffraction parameter as follows [52]:
v
Z
sin
S(v) =
0
πv 2
2
dv
(9a)
dv
(9b)
2 · AS · SB
The diffraction parameter v is calculated based on the distance
from the TX to the screen, from the screen to the RX, the
diffraction angle α (see Fig. 5b), and the carrier wavelength λ.
The ± sign in (10) is applied as + to both edges for NLOS
conditions. When calculating the diffraction parameter (10) under
unobstructed (LOS) conditions for the screen edge (w1 or w2)
closest to the straight line drawn between the TX and RX, the ±
is treated as “−”, whereas the ± is treated as “+” for the screen
edge farthest from the straight line drawn between the TX and
RX [52].
The individual received signal caused by knife-edge diffraction
from the w1 and w2 edges is Fw1 and Fw2 , respectively. The
complex signals corresponding to the edges can be added in order
to determine the combined diffraction loss observed at the RX. The
total diffraction loss power in log-scale is determined by taking
the magnitude squared of the summed signals as follows [52]:
Lscreen [dB] = 20 log10
Fw1 + Fw2
(11)
The DKED model (11) has been adopted by METIS and others [20], [48], but it does not consider the antenna radiation
pattern (it assumes an omnidirectional antenna [53]) and has
been shown to underestimate diffraction loss in the deepest fades
when using directional antennas, which are sure to be employed
by mmWave mobile devices [38]. This physical phenomenon
occurs when the blocker obstructs the LOS path between antennas
(i.e. deep fading), leaving only the off-boresight antenna gains
to contribute to the received signal strength, which is slightly
less than the directive gain. To account for the impact of nonuniform gain directional antennas on human blockage, antenna
gain is considered in (11) (see [38]). The following azimuth farfield power radiation pattern of a horn antenna (general for any
directional antenna) for a given half-power beamwdith (HPBW)
is approximated by [38], [39]:
G(θ) = sinc2 (a · sin(θ)) · cos2 (θ)
where:
sinc
2
a · sin
HPBWAZ
2
!
· cos
2
HPBWAZ
2
=
1
2
The DKED model in (11) can be extended to include TX and RX
antenna gains for the projected angles θ between the TX and the
screen, and the screen and RX as follows [38]:
LScreen A.G. [dB] = 20 log10
Fw1 ·
p
p
GD2w1 · GD1w1
!
+Fw2 ·
p
p
GD2w2 · GD1w2
Rel. Pow. (dB)
Rel. Pow. (dB)
-30
2.1
2.2
-20
-30
3
(12)
-20
-30
-40
2.8
3
-30
2.6
2
0
-20
-30
3.2
2.4
2.6
0
-10
-20
-30
-40
2.6
2.8
3
2.6
2.8
3
Time (sec.)
Meas 9: 4.5 m
0
-10
-20
-30
-40
2.6
2.2
Time (sec.)
Meas 6: 3.0 m
Time (sec.)
Meas 8: 4.0 m
0
Time (sec.)
2.4
-10
-40
2.4
3.2
-10
2.6
-20
Time (sec.)
Meas 5: 2.5 m
0
2.8
0
-10
-40
2.2
Time (sec.)
Meas 7: 3.5 m
(10)
λ(AS + SB)
2
-10
2.6
Rel. Pow. (dB)
vw1|w2 = ±αw1|w2
-20
-40
1.9
-40
and where the diffraction parameter v is derived by [34]:
s
0
Rel. Pow. (dB)
0
πv 2
2
-30
Rel. Pow. (dB)
cos
-20
3GPP/METIS
Meas 3: 1.5 m
-10
Time (sec.)
Meas 4: 2.0 m
Rel. Pow. (dB)
v
Z
0
-10
-40
1.8
DKED-AG. Worst Case (out-of-phase)
Meas 2: 1.0 m
Rel. Pow. (dB)
where the numerical approximations of Fresnel integration for
C(v) and S(v) are:
C(v) =
DKED-AG. Best Case (in-phase)
Rel. Pow. (dB)
(8)
Rel. Pow. (dB)
Fw1|w2
(1−j) 1+j − (C(v) + j · S(v)) , if v > 0
2
2
= (1−j) 1+j
+ (C(−v) + j · S(−v)) , otherwise
2
2
DKED-AG.
Meas 1: 0.5 m
0
-10
-20
-30
-40
2.8
3
3.2
Time (sec.)
2.7
2.8
2.9
3
3.1
Time (sec.)
Fig. 6: Comparison of measured received power of human blockage at
73 GHz and the DKED-AG model in (12) [38].
where GD2w1 , GD1w1 , GD2w2 , and GD1w2 are the linear power
gains (normalized to the directive gain such that G0◦ = 1) of the
antennas based on the point-source projections Aw1; w1B; Aw2;
w2B; A to w1, w1 to B, A to w2, and w2 to B (see Fig. 5b).
When the screen does not obstruct the LOS path between the TX
and RX, the normalized gains are set to G(θ) = 1, since the slight
variations of antenna patterns have little effect on diffraction loss
in the unobstructed case.
E. Human Blockage Results and Analysis
For each of the nine measurement paths, the area under the
curve of each of the 2500 PDPs was integrated to calculate the
received power in 2 ms increments. In Fig. 6 the received power
(red) is compared to the DKED antenna gain (DKED-AG) model
(green) (12), in addition to showing the constructive (signals inphase) and destructive (signals out of phase) sum of received
signals of the upper (blue) and lower (black) bound of the fade
envelope, respectively. Fig. 6 represents loss as compared to a free
space reference with no blockage between the TX and RX. From
Fig. 6, we observe gain in the received signal as the human enters
the TX/RX LOS path, and then deep attenuations as the human
blocks the LOS path. Due to the fact that identical antennas were
used at the TX and RX, the envelopes of the received signal power
were similar in the two different symmetrical cases (i.e., Meas. 1
and Meas. 9). The best case scenario (minimum diffraction loss)
is found by summing the magnitudes of received field components
from the w1 and w2 edges of the blocker, represented by the blue
dashed line in Fig.
minimum loss, (12)
as:
p 6. Forp
p is reformulated
p
20 log10 (|Fw1 · GD2w1 · GD1w1 |+|Fw2 · GD2w2 · GD1w2 |).
The worst case scenario (maximum diffraction loss) is found by
taking the difference of the magnitudes of received signals from
the w1 and w2 edges, and is represented by the
line
black dotted
p
in Fig 6, where (12) is computed as: 20 log10 |Fw1 · GD2w1 ·
p
p
p
GD1w1 | − |Fw2 · GD2w2 · GD1w2 | .
It was previously demonstrated that diffraction loss models that
do not account for antenna gain pattern can severely underestimate
the diffraction loss when the blocker is close to either antenna
(a critical issue for mobile phone use) [38]. The DKED-AG
model (12) accurately predicts what is measured, and predicts the
deepest attenuation caused by a human blocker in excess of 40 dB.
To model multiple blockers, the screen model can be replicated
multiple times. These results show that adaptive antenna array
and beamforming techniques will be employed to find suitable
reflectors and scatterers in the signal transmission to overcome
severe blockage attenuation in future 5G communication systems.
The DKED-AG model in (12) may be extended [51] to consider
the top and bottom screen edges, phase corrections, and nonperpendicular screen orientations, although the simple model (12)
matches the human blocking measurements with confidence. It
can be seen in Fig. 6 (Meas. 1 and 9) that the signal strength
drops off at a rate of 0.4 dB/ms as the blocker moves at 1
m/s and begins to shadow the TX (RX). Mobile handoffs and
beam steering schemes will be needed to rescue the mobile from
severe fades by the use of electrically scanning beams at the submillisecond level, a feat easily accomplished with sub millisecond
packets in an air interface standard. An additional technique for
mitigating the effects of rapid fading could include rapid re-routing
around obstacles via handoff to another access point (AP) in a
network cluster [54]. Note that just prior to the deep shadowing
events in Fig. 6 there is a slight increase/scintillation of signal
strength of ∼ 2 dB peak-to-peak amplitudes (noticed by others
in [55]), which could be used to detect the imminent presence of
an obstruction such that the RX adapts its beam in anticipation of
the pending deep fade. The 3GPP/METIS blockage model [20],
[48] shown in Fig. 6 underestimates the deep fades of shadowing
events [38], especially when the blocker is close to the TX or
RX antenna, since the full directive gain of the TX and RX
antennas is not available across the diffraction obstacle, and thus is
unable to contribute to the received signal strength from diffraction
around the blocker during the shadowing event [38]. We note that
the 3GPP/METIS model only offers reasonable agreement to the
measured loss when the blocker is far (several meters) from the
TX and RX antenna.
IV. S MALL -S CALE S PATIAL S TATISTICS
A. Introduction of Small-Scale Spatial Statistics
Small-scale fading and small-scale autocorrelation characteristics are crucial for the design of future mmWave communication
systems, especially in multiple-input multiple-output (MIMO)
channel modeling. Previous studies on small-scale fading characteristics focused on sub-6 GHz frequencies, yet investigations
at mmWave are scarce. Wang et al. [56] showed that smallscale fading of received power in indoor corridor scenarios with
omnidirectional antennas at both TX and RX could be well
described by Ricean distributions with K-factors ranging from 5
dB to 10 dB based on their indoor corridor measurements at 15
GHz with a bandwidth of 1 GHz, and ray tracing results using
a ray-optical based channel model validated by measurements.
Henderson et al. compared Rayleigh, Ricean, and the Two-WaveDiffuse-Power (TWDP) distributions to find the proper small-scale
fading distribution of received voltage magnitudes for a measured
2.4 GHz indoor channel [57] where the Ricean distribution had
highest modeling accuracy in most indoor cases [57]. The authors in [58] demonstrated the use of the TWDP fading model
for mmWave communications. It was reported that log-normal
distribution had a good fit to measured received signal envelopes
in some indoor mobile radio channels [59]. Important work on
wideband directional small-scale fading also appears in [60]–
[64]. In the following subsections, small-scale fading distributions
TABLE II: Hardware Specifications of Small-Scale Fading and Local
Area Channel Transition Measurements.
Campaign
Broadcast Sequence
TX and RX Antenna Type
TX/RX Chip Rate
Slide Factor γ
RF Null-to-Null Bandwidth
PDP Threshold
TX/RX Intermediate Freq.
TX/RX Local Oscillator
Carrier Frequency
TX Antenna Gain
RX Antenna Gain
Max TX Power / EIRP
TX Az. and El. HPBW
TX/RX Heights
RX Az. and El. HPBW
TX-RX Antenna Pol.
Max Measurable Path Loss
73 GHz Small-Scale
Fading and
Correlation
Measurements
73 GHz Local Area
Channel Transition
Measurements
11th order PN Code (L = 211 − 1 = 2047)
Rotatable pyramidal horn antenna
500 Mcps / 499.9375 Mcps
8000
1 GHz
20 dB down from max peak
5.625 GHz
67.875 GHz (22.625 GHz × 3)
73.5 GHz
27 dBi
9.1 dBi
14.2 dBm / 41.2 dBm
20 dBi
14.3 dBm / 41.3 dBm
7◦
4.0 m / 1.4 m
4.0 m / 1.5 m
60◦
15◦
V-V (vertical-to-vertical)
168 dB
180 dB
of total power and autocorrelation characteristics of received
voltage amplitudes at 73 GHz in urban microcell environments are
investigated based on a measurement campaign conducted during
the summer of 2016 around the engineering campus of New York
University in downtown Brooklyn.
B. Measurement System for Small-Scale Spatial Statistics
The TX system for small-scale fading and autocorrelation
measurements at 73 GHz was identical to the TX system used
for the human blocking measurements with the difference only
for TX antennas and transmit powers as identified in Table II.
The RX side of the system captured the RF signal via steerable
horn antennas and downconverted the signal to an IF of 5.625
GHz, which was then demodulated into its baseband in-phase
(I) and quadrature-phase (Q) signals which were correlated via a
common sliding correlation architecture [5], [6], [8], [34] where
the time-dilated I and Q channel voltages were sampled by an
oscilloscope and then squared and added together in software to
generate a PDP. Antennas with 27 dBi gain (7◦ Az./El. HPBW)
and 9.1 dBi gain (60◦ Az./El.HPBW) were used at the TX and
RX sides, respectively [62].
C. Small-Scale Measurement Environment and Procedure
In the summer of 2016, a set of small-scale linear track
measurements were conducted at 73 GHz on the campus of NYU
Tandon School of Engineering in downtown Brooklyn, New York,
representative of an urban microcell (UMi) environment [62].
The measurement environment, and the TX and RX locations are
depicted in Fig. 7. One TX location with an antenna height of
4.0 m above the ground and two RX locations with an antenna
height of 1.4 m were selected to perform the measurements, where
one RX was LOS to the TX while the other was NLOS. The
TX was placed near the southwest corner of the Dibner library
building (top center in Fig. 7), the LOS RX was located 79.9
m away from the TX, and the NLOS RX was shadowed by the
southeast corner of a building (Rogers Hall on the map) with a
T-R separation distance of 75.0 m [62]. Other specifications about
the measurement hardware are detailed in Table II.
Fig. 7: 2D map of the 73 GHz small-scale measurement environment and
the locations of TX and RX. Pointing to the top of the map is 0◦ .
A fixed 35.31-cm spatial linear track (about 87 wavelengths
at 73.5 GHz) was used at each RX location in the small-scale
fading measurements, over which the RX antenna was moved in
increments of half-wavelength (2.04 mm) over 175 track positions
[62]. At each RX, six sets of small-scale fading measurements
were performed, where the elevation angle of the RX antenna
remained fixed at 0◦ (parallel to the horizon) and a different
azimuth angle was chosen for each set of the measurements with
the adjacent azimuth angles separated by 60◦ , such that the RX
antenna swept over the entire azimuth plane after rotating through
the six pointing angles. The RX antenna was pointing at a fixed
angle while moving along the linear track for each set of the
measurements and a PDP was acquired at each track position for
each pointing angle. The TX antenna elevation angle was always
fixed at 0◦ (parallel to horizon). Under the LOS condition, the TX
antenna was pointed at 90◦ in the azimuth plane, directly towards
the RX location; for NLOS, the TX antenna azimuth pointing
angle was 200◦ , roughly towards the southeast corner of Rogers
Hall in Fig. 7 [62]. Due to space limitations, we show here only
one track orientation at each RX (along the direction of the street
beside the RX), but more results and observations are detailed
in [62] which show the fading depths are a function of antenna
orientations and environment.
As a comparison, the 28 GHz small-scale measurements presented in [65] investigated the small-scale fading and autocorrelation of individual resolvable multipath voltage amplitudes using
a 30◦ Az./El. HPBW RX antenna, whereas this paper studies
73 GHz fading and autocorrelation using a wider HPBW (60◦ )
RX antenna, and focuses on received signal voltage amplitude by
integrating the area under the entire PDP curve and then taking
the square root of the total power, instead of individual multipath
voltage amplitude at each location along a track.
Fig. 8: Measured 73 GHz small-scale directional PDPs over 175 track
positions in LOS. The RX horn antenna (60◦ HPBW) was pointing on
boresight to the TX, and track orientation orthogonal to the T-R line.
Fig. 9: Measured 73 GHz small-scale directional PDPs over 175 track
positions in NLOS. The RX horn antenna (60◦ HPBW) was pointing to
the TX but was obstructed by a building corner, and the track orientation
was along the direction of the street.
Typical measured small-scale directional PDPs over 175
track positions on the 35.31-cm (about 87 wavelengths at 73.5
GHz) linear track in the NLOS environment are depicted in
Fig. 9, where the track orientation was along the direction of
the street, and the RX antenna was pointing to the TX but
was obstructed by a building corner [62]. Fig. 9 shows there is
very moderate power variation (4.1 dB) over different local-area
track positions, albeit with rich and varying multipath components.
E. Small-Scale Spatial Statistics Results and Analysis
D. Small-Scale Measurement Results
Fig. 8 illustrates typical measured small-scale directional PDPs
over 175 track positions on the 35.31-cm (about 87 wavelengths
at 73.5 GHz) linear track in the LOS environment, where the RX
horn antenna with 60◦ HPBW was pointing on boresight to the
TX, and the track orientation was orthogonal to the T-R line. The
total power in Figs. 8 and 9 is computed as the area under the PDP
at a particular track position over the 1 GHz RF bandwidth. Fig. 8
shows there is 11 dB power variation over different track positions,
but the power variation is only 3.7 dB when the track orientation
was in the direction of the T-R line (not shown), indicating little
small-scale spatial fading [62].
1) Omnidirectional Small-Scale Spatial Statistics: As described
above, a rotatable directive horn antenna was used at the RX
side to capture directional PDPs in the small-scale fading and
correlation measurements. In channel modeling, however, omnidirectional statistics are often preferred, since arbitrary antenna
patterns can be implemented according to one’s own needs if
accurate temporal and spatial statistics are known [23]. Therefore,
we synthesized the approximated omnidirectional received power
at every track interval by taking the area under the curve of
each directional PDP and summing powers using the approach
presented in [39] and on Page 3040 from [6], thereby computing
omnidirectional received power. Although the RX antenna did not
80
100
Measured LOS
Log-normal, = 0.91 dB
Ricean, K = 10 dB
Rayleigh
60
10 1
40
10 0
10 -1
10 -2
20
0
-4
10 -3
10 -4
-4
-3
-2
-1
0
-3
1
-2
2
-1
3
0
4
Signal Level (dB relative to mean)
Probability (%) (<= abscissa)
Probability (%) (<= abscissa)
100
80
Measured NLOS
Log-normal, = 0.65 dB
Ricean, K = 19 dB
Rayleigh
60
10 1
40
10 0
10 -1
10 -2
20
0
-1.5
10 -3
10 -4
-1.5
-1
-0.5
0
-1
0.5
-0.5
1
0
1.5
Signal Level (dB relative to mean)
Fig. 10: CDF of the measured small-scale spatial fading distribution of
the received voltage amplitude for the omnidirectional RX antenna pattern
in the LOS environment at 73 GHz with a 1 GHz RF bandwidth.
sweep the entire 4π Steradian sphere, the azimuth plane spanned
±30◦ with respect to the horizon, ensuring that a large majority
of the arriving energy was captured, as verified in [39].
Fig. 10 illustrates the cumulative distribution function (CDF)
of the measured small-scale received voltage amplitude at 73
GHz with a 1 GHz RF bandwidth over the 35.31-cm length
track with 175 track positions in increments of half-wavelength
(2.04 mm) for the omnidirectional RX antenna pattern in the
LOS environment [62]. Superimposed with the measured curve
are the CDFs of the Rayleigh distribution, the zero-mean lognormal distribution with a standard deviation of 0.91 dB (obtained
from the measured data), and the Ricean distribution with a Kfactor of 10 dB obtained from the measured data by dividing the
total received power contained in the LOS path by the power
contributed from all the other reflected or scattered paths. As
shown in Fig. 10, the measured 73 GHz small-scale spatial fading
in the LOS environment can be approximated by the Ricean
distribution with a K-factor of 10 dB, indicating that there is
a dominant path (i.e., the LOS path) contributing to the total
received power, and that the received signal voltage amplitude
varies little over the 35.31-cm (about 87 wavelengths) length
track. The log-normal distribution does not fit the measured data
well in the regions of -3 to -2.5 dB and +1.2 to +1.5 dB
about the mean. The maximum fluctuation of the received voltage
amplitude is merely 3 dB relative to the mean value, whereas
the fades are much deeper for the Rayleigh distribution. The
physical reason for this is the presence of a dominant LOS path.
The small-scale spatial fading in the NLOS environment for the
omnidirectional RX antenna pattern is illustrated in Fig. 11, and
the zero-mean log-normal distribution with a standard deviation
of 0.65 dB (obtained from the measured data) is selected to fit
the measured result, and Ricean and Rayleigh distributions are
also given as a reference [62]. As evident from Fig. 11, the
measured NLOS small-scale spatial fading distribution matches
the log-normal fitted curve almost perfectly. In contrast, the Ricean
distribution with K = 19 dB does not fit the measured data as
well as the log-normal distribution in the tail region around -0.6
dB to -0.8 dB of the relative mean signal level (as shown by
the inset in Fig. 11), since the Ricean K = 19 dB distribution
predicts more occurrences of deeper fading events, whereas the
log-normal distribution with a 0.65 dB standard deviation predicts
a more compressed fading range of -0.8 dB to +0.8 dB about the
mean, which was observed for the wideband NLOS signals. The
fact that the local fading of received voltage amplitudes in the
NLOS environment is log-normal instead of Rayleigh is similar
to models in [66] for urban mobile radio channels. For a NLOS
environment, there may not be a dominant path, yet the transmitted
Fig. 11: CDF of the measured small-scale spatial fading distribution of
the received voltage amplitude for the omnidirectional RX antenna pattern
in the NLOS environment at 73 GHz with a 1 GHz RF bandwidth.
broadband signal experiences frequency-selective fading (when
the signal bandwidth is larger than the coherence bandwidth of
the channel [34]). Different frequency components of the signal
experience uncorrelated fading, thus it is highly unlikely that all
parts of the signal will simultaneously experience a deep fade,
and the fades over frequency tend to be very sharp, taking up a
small portion of the total power received over the entire signal
bandwidth [16]. Consequently, the total received power changes
very little over a small-scale local area. This is a distinguishing
feature of wideband mobile signals as compared to narrowband
signals.
Apart from small-scale spatial fading, small-scale spatial autocorrelation is also important for wireless modem design. Spatial
autocorrelation characterizes how the received voltage amplitudes
correlate at different linear track positions within a local area [65].
Spatial autocorrelation coefficient functions can be calculated
using Eq. (13), where Xk denotes the k th linear track position,
E[ ] is the expectation operator where the average of voltage
amplitudes is taken over all the positions Xk , and ∆X represents
the spacing between different antenna positions on the track [62].
The measured 73 GHz spatial autocorrelation of the received
voltage amplitudes in LOS and NLOS environments with a 1 GHz
RF bandwidth are depicted in Fig. 12 and Fig. 13, respectively.
Note that a total of 175 linear track positions over the 35.31-cm
length track were measured during the measurements, yielding a
maximum spatial separation of 174 half-wavelengths on a single
track. Only up to 60 half-wavelengths, however, are shown herein
because little change is found thereafter and it provides 100
autocorrelation data points for all spatial separations on a single
track, thus improving the reliability of the statistics. According
to Fig. 12, the received omnidirectional signal voltage amplitude
first becomes uncorrelated at a spatial separation of about 3.5λ,
then becomes slightly anticorrelated for separations of 3.5λ to
10λ, and becomes slightly correlated for separations between 10λ
and 18λ, and decays towards 0 sinusoidally after 18λ. Therefore,
the spatial correlation can be modeled by a “damped oscillation”
function of (14) [62] [67]:
f (∆X) = cos(a∆X)e−b∆X
(14)
where ∆X denotes the space between antenna positions, a is an
oscillation distance with units of radians/λ (wavelength), T =
2π/a can be defined as the spatial oscillation period with units of λ
or cm, and b is a constant with units of λ−1 whose inverse d = 1/b
is the spatial decay constant with units of λ. a and b are obtained
using the minimum mean square error (MMSE) method to find
the best fit between the empirical spatial autocorrelation curve
ρ= q
E
E
Ak (Xk ) − Ak (Xk )
Ak (Xk ) − Ak (Xk )
Ak (Xk + ∆X) − Ak (Xk + ∆X)
2
E
(13)
2
Ak (Xk + ∆X) − Ak (Xk + ∆X)
TABLE III: Spatial correlation model parameters in (14) for 73 GHz, 1
GHz RF bandwidth (λ=0.41 cm).
a (rad/λ)
T = 2π/a
b (λ−1 )
d = 1/b
LOS Omnidirectional
0.45
14.0λ (5.71 cm)
0.10
10.0λ (4.08 cm)
NLOS Omnidirectional
0
Not used
0.26
3.85λ (1.57 cm)
LOS
Directional
0.33 to 0.50
12.6λ to 19.0λ (5.14
cm to 7.76 cm)
0.03 to 0.15
6.67λ to 33.3λ (2.72
cm to 13.6 cm)
NLOS
Directional
0
Not used
0.04 to 1.49
0.67λ to 25.0λ (0.27
cm to 10.2 cm)
100
Probability (%) (<= abscissa)
Fig. 12: Measured 73 GHz broadband spatial autocorrelation coefficients
of the received voltage amplitude in the LOS environment, and the
corresponding fitting model. The T-R separation distance is 79.9 m.
Condition
60
40
20
0
-4
Fig. 13: Measured 73 GHz broadband spatial autocorrelation coefficients
of the received voltage amplitude in the NLOS environment, and the
corresponding fitting model. The T-R separation distance is 75.0 m.
and theoretical exponential model given by (14). The “damped
oscillation” pattern can be explained by superposition of multipath
components with different phases at different linear track positions. As the separation distance of linear track positions increases,
the phase differences among individual multipath components will
oscillate as the separation distance of track positions increases
due to alternating constructive and destructive combining of the
multipath phases. This “damped oscillation” pattern is obvious
in LOS environment where phase difference among individual
multipath component is not affected by shadowing effects that
occurred in NLOS environments. The form of (14) also guarantees
that the spatial autocorrelation coefficient is always 1 for ∆X = 0,
and converges to 0 when ∆X approximates infinity. The spatial
autocorrelation curve for NLOS environment in Fig. 13 exhibits
a different trend from that in Fig. 12, which is more akin to an
exponential distribution without damping, but can still be fitted
using Eq. (14) with a set to 0 [62]. The constants a, b, are
provided in Table III, where T is the oscillation period, and d
represents the spatial decay constant. From Fig. 13 and Table III
it is clear that after 1.57 cm (3.85 wavelengths at 73.5 GHz) in
the NLOS environment, the received voltage amplitudes become
uncorrelated (the correlation coefficient decreases to 1/e [22]).
We note that Samimi [65] found individual multipath voltage
amplitudes received using a 30◦ Az./El. HPBW antenna became
uncorrelated at physical distances of 0.52 cm (0.48 wavelengths
at 28 GHz) and 0.67 cm (0.62 wavelengths at 28 GHz) in LOS
and NLOS environments, respectively – smaller decorrelation
distances compared to the present 73 GHz results measured using
a 60◦ Az./El. HPBW antenna.
2) Directional Small-Scale Spatial Statistics: Since mobile
devices will use directional antennas, directional statistics are
270° (b)
330°
30 °
90 °
150°
210°
Ricean
K=7dB
Ricean
K=17dB
80
-3
-2
-1
0
1
2
Signal Level (dB relative to mean)
3
4
Fig. 14: Measured 73 GHz LOS small-scale spatial fading distributions of
the directional received voltage amplitude, and the corresponding Ricean
fitting curves with different K factors. The angles in the legend denote
the receiver antenna azimuth angle, and ”b” denotes boresight to the TX.
also of interest. In this subsection, we will investigate smallscale spatial fading and autocorrelation of the received voltage
amplitudes associated with directional antennas at the RX.
Small-scale fading of received voltage amplitudes along the linear track using the 7◦ Az./El. HPBW TX antenna and 60◦ Az./El.
HPBW RX antenna in LOS and NLOS environments are shown
in Fig. 14 and Fig. 15, respectively, where the TX and RX were
placed as shown in Fig. 7, and each measured curve corresponds
to a unique RX antenna azimuth pointing angle relative to true
north as specified in the legend. There was no signal for the RX
azimuth pointing angle of 270◦ in the NLOS environment, thus the
corresponding results are absent in Fig. 15. The strongest pointing
directions are 270◦ and 150◦ in Figs. 14 and 15, respectively
[62]. As shown in Figs. 14 and 15, the measured directional
spatial autocorrelation coefficients resemble Ricean distributions
in both LOS and NLOS environments. Possible reason for such
distributions is that only one dominant path (accompanied with
several weaker paths) is captured by the horn antenna due to
its directionality, given the fact that mmWave propagation is
directional and the channel is sparse [22]. The Ricean K-factor
for received voltage amplitudes for various RX pointing directions
ranges from 7 dB to 17 dB for the LOS environment, and 9 dB
to 21 dB for the NLOS case, as shown in Figs. 14 and 15, as the
RX is moved over a 35.31 cm (86.5 wavelengths at 73.5 GHz)
track.
Figs. 16 and 17 illustrate the spatial autocorrelation coefficients
of the received voltage amplitudes for individual antenna pointing
angles in LOS and NLOS environments in downtown Brooklyn
(see Fig. 7), respectively. As shown by Fig. 16, all of the six
spatial autocorrelation curves in the LOS environment exhibit
sinusoidally exponential decaying trends, albeit with different
oscillation patterns and decay rates. The TX-RX boresight-toboresight pointing angle (270◦ RX pointing angle, corresponding
Probability (%) (<= abscissa)
the southeast corner of Rogers Hall in Fig. 7. It is clear from Fig.
4 in [4] that correlation distances vary among typical mmWave
measurements [18] due to the site-specific nature of propagation.
100
90 ° (s, to TX)
150°
210°
330°
30 °
Ricean
K=9dB
Ricean
K=21dB
80
60
40
20
0
-4
-3
-2
V. L OCAL A REA C HANNEL T RANSITION
A. Introduction of Local Area Channel Transition
-1
0
1
2
3
4
Signal Level (dB relative to mean)
Spatial Autocorrelation Coefficient
Fig. 15: Measured 73 GHz NLOS small-scale spatial fading distributions
of the directional received voltage amplitude, and the corresponding
Ricean fitting curves with different K factors. The angles in the legend
denote the receiver antenna azimuth angle, and ”s, to TX” denotes along
the direction of the street and pointing to the TX side.
1
270° (b)
330°
90 °
150°
210°
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
0
5
10
15
20
25
Separation x (Number of Wavelengths)
30
Fig. 16: Measured 73 GHz LOS spatial autocorrelation coefficients of the
directional received voltage amplitudes. The angles in the legend denote
the receiver antenna azimuth angle, and ”b” denotes boresight to the TX.
to the strongest received power) yields the smallest oscillation
since there is only a single LOS component in the PDP, while
the other pointing directions contain two or more multipath
components with varying phases that result in larger oscillation
[62]. The spatial decay constants for all the curves in Fig. 16
are given in Table III. Compared with the omnidirectional case
displayed in Fig. 12, it is clear that for the LOS environment,
the spatial autocorrelation of both omnidirectional and directional
received voltage amplitude obeys similar distribution, namely,
the sinusoidal-exponential function, with similar decorrelation
distances. On the other hand, most of the spatial autocorrelation
curves for the directional received voltage amplitude shown in
Fig. 17 are also in line with that given by Fig. 13. One exception
for the 73 GHz spatial correlation is found at 30◦ pointing angle
that is the second strongest pointing direction (see Fig. 17), where
decorrelation was much more gradual and decreases to 1/e at 25.0λ
(10.2 cm), probably due to the presence of a dominant path with a
relatively constant signal level, likely caused by the diffraction of
Fig. 17: Measured 73 GHz NLOS spatial autocorrelation coefficients of
the directional received voltage amplitudes. The angles in the legend
denote the receiver antenna azimuth angle, and “s, to TX” denotes along
the direction of the street and pointing to the TX side.
Large-scale channel characteristics, such as the autocorrelation
of shadow fading and delay spread over distance at the mobile,
inter-site correlation of shadow fading at the mobile for two base
stations or a base station for one mobile, and properties of local
area channel transition, play an important role in constructing
channel models for wireless communication systems [20], [68]–
[70]. While sufficient studies on large-scale channel characteristics
have been conducted at sub-6 GHz frequencies [67], [68], [71],
[72], similar studies have been rarely conducted at mmWave
frequencies. Guan et al. investigated spatial autocorrelation of
shadow fading at 920 MHz, 2400 MHz, and 5705 MHz in curved
subway tunnels [71]. Results showed that the 802.16J model was
a better fit to the measured data than an exponential model,
and the mean decorrelation distances were found to be several
meters. Another measurement campaign carried out in urban
macro-cell (UMa) environments at 2.35 GHz showed that a double
exponential model fit well with the autocorrelation coefficients
of shadow fading samples extracted from all measured routes,
while for individual routes, an exponentially decaying sinusoid
model had better fitting performance [67]. On the other hand,
an exponential function was adopted in the 3GPP channel model
(Releases 9 and 11) to describe the normalized autocorrelation of
shadow fading versus distances [69], [70]. Moreover, Kolmonen et
al. [72] investigated interlink correlation of eigenvectors of MIMO
correlation matrices based on a multi-site measurement campaign
at 5.3 GHz using a bandwidth of 100 MHz. Results showed
that the first eigenvectors for both x- and y-oriented arrays were
highly correlated when two RX locations were largely separated.
In the following subsections, local area and channel transition
measurements at 73 GHz are described and analyzed.
B. Measurement System, Environment, and Procedure for Local
Area Channel Transition
The measurement system used for local area channel transition
measurements was identical to the one used in the previous
small-scale fading and autocorrelation measurements described
in Section IV, except from the antennas [50] [73]. A 27 dBi
gain (7◦ azimuth and elevation Az./El. HPBW) and 20.0 dBi
gain (15◦ Az./El. HPBW) antenna were used at the TX and RX,
respectively. Detailed specifications regarding the measurement
system are provided in Table II.
The local area channel transition measurements were conducted
at 73 GHz in the MetroTech Commons courtyard next to 2 and 3
MetroTech Center in downtown Brooklyn. During measurements,
the TX and RX antennas were set to 4.0 m and 1.5 m above
ground level, respectively. For each set of cluster or route scenario RX locations, the TX antenna remained fixed and pointed
towards a manually selected azimuth and elevation pointing angle
that resulted in the strongest received power at the starting RX
position (RX81 for route measurements, and RX51 and RX61
for the LOS and NLOS cluster measurements, respectively). For
each specific TX-RX combination, five consecutive and identical
azimuth sweeps (∼3 minutes per sweep and ∼2 minutes between
C. Local Area Channel Transition Results and Stationarity
The route measurements mimicked a person moving along
an urban canyon from a NLOS to a LOS region, in order to
understand the evolution of the channel during the transition.
Fig. 20 displays the omnidirectional path loss for each of the
RX locations (RX81 to RX96) where the received power from
the individual directional measurements at the RX was summed
up to determine the entire omnidirectional received power at each
measurement location (out of the 5 sweeps, the maximum power
at each angle was used, although variation was less than a dB
between sweeps) [39], [74].
The transition from LOS to NLOS in Fig. 20 is quite abrupt,
where path loss increases by ∼8 dB from RX92 to RX91, similar
Fig. 18: 2D map of TX and RX locations for route NLOS to LOS transition measurements. The yellow star is the TX location, blue dots represent
LOS RX locations, and red squares indicate NLOS RX locations. N = 0◦ .
Fig. 19: 2D map of TX and RX locations for cluster measurements with
LOS and NLOS RX clusters. The yellow star is the TX location, blue
dots represent LOS RX locations, and red squares indicate NLOS RX
locations. Pointing to the top is 0◦ .
160
140
Path Loss (dB)
sweeps) were conducted at the RX in HPBW step increments
(15◦ ) where a PDP was recorded at each RX azimuth pointing
angle and resulted in at most 120 PDPs ( 360
15 × 5 = 120) per
combination (some angles did not have detectable signal above
the noise). The best RX pointing angle in the azimuth plane was
selected as the starting point for the RX azimuth sweeps (elevation
remained fixed for all RX’s), at each RX location measured.
For the route measurements, 16 RX locations were measured
for a fixed TX location (L8) with the RX locations positioned in 5
m adjacent increments of each other forming a simulated route in
the shape of an “L” around a building corner from a LOS to NLOS
region, as provided in Fig. 18. The LOS location (five: RX92 to
RX96) and NLOS location (11: RX81 to RX91) T-R separation
distances (Euclidean distance between TX and RX) varied from
29.6 m to 49.1 m and 50.8 m to 81.5 m, respectively. The TX
antenna at L8 kept the same azimuth and elevation pointing angle
of 100◦ and 0◦ , respectively, during each experiment (see Fig. 18.
Therefore, the LOS measurements have the TX and RX antennas
roughly on boresight in the LOS situation, but they are not exactly
on boresight throughout the entire experiment over all measured
locations. The general layout of measurements consisted of the
RX location starting at RX81, approximately 54 m along an urban
canyon (Bridge Street: 18 m width), with the TX antenna pointed
towards the opening of the urban canyon (see Fig. 18). The LOS
locations were in clear view of the RX, but with some nearby
minor foliage and lamppost obstructions.
For the cluster measurements, 10 RX locations were measured
for a fixed TX location (L11), with two sets of RX clusters, one
in LOS (RX61 to RX65) and the other in NLOS (RX51 to RX
55). For each cluster of RX’s, the adjacent distance between each
RX location was 5 m, however, the path of adjacent RX locations
took the shape of a semi-circle as displayed in Fig. 19. The LOS
cluster T-R separation distances (Euclidean distance between TX
and RX) varied between 57.8 m and 70.6 m with a fixed TX
antenna azimuth and elevation departure angle of 350◦ and -2◦ ,
respectively, and fixed RX elevation angles of +3◦ , to ensure
rough elevation and azimuth alignment for all RX locations. For
the NLOS cluster, the T-R separation distances were between 61.7
m and 73.7 m with a fixed TX antenna azimuth and elevation
departure angle of 5◦ and -2◦ , respectively, and fixed RX elevation
angles of +3◦ . The LOS cluster of RX’s was located near the
opening of an urban canyon near some light foliage, while the
TX location was ∼57 m along an urban canyon (Lawrence Street:
18 m width). The NLOS cluster of RX locations was around
the corner of the urban canyon opening in a courtyard area (see
Fig. 19), also with nearby moderate foliage and lampposts.
Path Loss for LOS to NLOS Transition at 73 GHz
Free Space Path Loss
LOS PL Data
NLOS PL Data
CI LOS: n = 2.53; σ = 2.5 dB
CI NLOS: n = 3.61; σ = 5.6 dB
120
NLOS->LOS Transition: 24.8 dB
100
135
130
125
120
115
110
45
80
60
100
101
50 55 60 65
102
T-R Separation (m)
Fig. 20: Omnidirectional path loss for route measurements for an RX
transitioning from a NLOS to a LOS region.
to the abrupt diffraction loss noticed in Section II. Path loss then
increases 9 dB further from RX91 to RX90, 1 dB from RX90
to RX89, 6 dB from RX 89 to RX88, and 1 dB from RX88 to
RX 87. This observation shows a large initial drop in 8 dB at
the LOS to NLOS transition region, but an overall 25 dB drop
in signal power when moving from LOS conditions to deeply
shadowed NLOS conditions approximately 25 meters farther along
a perpendicular urban canyon (∼10 m increase in Euclidean T-R
separation distance), when using an omnidirectional RX antenna.
The 25 dB drop in signal strength over a 25 m path around
a corner (1 dB/m) is important for handoff considerations. The
signal fading rate is 35 dB/s for vehicle speeds of 35 m/s, or 1
73 GHz Polar Plot at RX 87 for TX: L8
Environment: NLOS
0°
330°
30°
TR Separation: 60.6 m
TX Height: 4 m
300°
60°
RX Height: 1.5 m
Measurement: 3
270°
90°
-110 -100 -90 -80
dBm
240°
120°
TXAZ/EL: 100°/ 0°
TX HPBWAZ/EL: 7°/ 7°
150°
180°
(a) RX87: NLOS
73 GHz Polar Plot at RX 92 for TX: L8
Environment: LOS
0°
330°
30°
TR Separation: 49.1 m
TX Height: 4 m
300°
60°
RX Height: 1.5 m
Measurement: 5
270°
-94
-79
240°
-64
90°
dBm
120°
TXAZ/EL: 100°/ 0°
RXEL: 0°
TX HPBWAZ/EL: 7°/ 7°
RX HPBWAZ/EL: 15°/ 15°
210°
Measurement Set
Omnidirectional Received Power σ [dB]
Route - LOS: RX92 to RX96
1.2
Route - NLOS: RX81 to RX91
7.9
Cluster - LOS: RX61 to RX65
4.3
Cluster - NLOS: RX51 to RX55
2.2
RXEL: 0°
RX HPBWAZ/EL: 15°/ 15°
210°
TABLE IV: Omnidirectional received power standard deviation for the
large-scale route and cluster scenario measurements.
150°
180°
(b) RX92: LOS
Fig. 21: Route scenario polar plots of RX azimuth spectra for RX 87
(NLOS location) and RX 92 (LOS location) that show the evolution of
AOA energy around a corner.
dB/s for walking speeds of 1 m/s. This motivates the use of beam
scanning and phased array technologies in the handset for urban
mobile mmWave communications that will search for and find
the strongest signal paths [40], and future work will study the best
antenna pointing angles at each location from these measurements.
Azimuth power spectra are useful to study how the arriving
energy changes as an RX moves from NLOS to LOS. Work
in [22] showed that energy arrives in directional lobes in mmWave
channels. Fig. 21 displays an RX polar plot from RX87 (NLOS
location) and RX92 (LOS location), where the TX was pointed
in the 100◦ direction towards the street opening (see Fig. 18).
Fig. 21a shows the power azimuth spectra at RX87 (∼25 meters down the urban street canyon) where energy from the TX
waveguides and reflects down Bridge Street such that there is one
main broad lobe at the RX oriented in the 0◦ direction and a
small narrow lobe in the 180◦ direction from weak reflectors and
scattering. The large azimuth spread in the main lobe demonstrates
the surprisingly reflective nature of the channel at the 73 GHz
mmWave band [22], [75].
Fig. 21b displays the power azimuth spectra at RX92 in LOS
with a strong central lobe coming from the direction of the TX
(285◦ ). A relatively strong secondary lobe (100◦ ) is also apparent
in Fig. 21b with energy contributions from reflections off of
the building to the east of RX92 and additional reflectors and
scatterers from nearby lampposts and signs. Table IV displays the
standard deviations of the omnidirectional received power values
measured along the LOS and NLOS routes shown in Fig. 18.
The omnidirectional received power standard deviation of 1.2 dB
is generally small for the LOS locations but is much larger in
NLOS (7.9 dB) due to substantial scattering along the route of
RX locations along the urban canyon where path loss tends to
increase non-linearly over log-distance.
The cluster measurements for the TX at L11 were designed
to understand the stationarity of received power in a local area
(larger than small-scale distances) on the order of many hundreds
to thousands of wavelengths (5 to 10 meters) at mmWave. In
LOS, the cluster of five RX locations with a fixed directional TX
antenna resulted in an omnidirectional received power standard
deviation of 4.3 dB over local area of 5 m x 10 m, a relatively
small variation, indicating a reasonably stationary average received
power for a local set of RX locations in LOS at 73 GHz. The
NLOS cluster resulted in an even lower omnidirectional received
power standard deviation of 2.2 dB, over a 5 m x 10 m local
area. The small fluctuation in received power over the local area
of the LOS and NLOS clusters implies that received power does
not significantly vary over RX locations separated by even a few
to several meters in a dense urban environment at mmWave. As an
aside, the directional CI model path loss exponent (PLE) using a 1
m free space reference in the route measurements was 2.53 in LOS
(a bit higher than free space due to elevation mismatch over the
route) and 3.61 in NLOS (for a single TX beam) [50]. Recent work
in [76] at 28 GHz studied the stationarity of wideband mmWave
channels, and reported smaller stationary regions than at 2 GHz.
VI. C ONCLUSION AND D ISCUSSION
Measurements and analysis were presented on diffraction, human blockage, small-scale fading and local area channel transition
at mmWave frequencies. We showed the KED model is suitable
for modeling diffraction at cmWave and mmWave bands in indoor
scenarios, while a creeping wave linear model is applicable to outdoor environments. Indoor and outdoor diffraction measurements
confirm theoretical simulations [17] that mmWave diffraction will
not be a dominant propagation mechanism. Measurements show
significant attenuation of 30 dB in indoor channels and 40 dB or
more in outdoor channels in the deep shadow region. Therefore,
systems will need to be designed to overcome extremely large
fades when a signal path is blocked and when diffraction is
no longer viable. MmWave antenna systems at the TX and RX
will need to cooperatively search for secondary paths at different
pointing directions. The large fades due to diffraction loss have
implications on the design of physical layer protocols and frame
structures to maintain a link, while finding other spatial paths.
Human blockage measurements showed a person can induce
more than 40 dB of loss when standing 0.5 m from the TX or
RX antenna, with a signal decay rate of 0.4 dB/ms at walking
speeds. The DKED-AG model given here incorporates directional
antenna patterns to accurately predict the upper and lower envelopes of measured received power during a blockage which
better agrees with real-world measurements when compared to the
3GPP/METIS blockage model (that was shown to underestimate
human blockage, most severely when close to the TX or RX).
Scintillation with ∼2 dB peak-to-peak amplitudes can be seen
in the measurements and model just before the blocker enters
the field of view of the RX, suggesting that deep fades may be
predictable just before they occur. This information may be used
in the design of beam steering or handoff algorithms.
Small-scale spatial fading statistics of received signal voltage
amplitudes of 1 GHz bandwidth signals at 73 GHz were measured
over a 35.31-cm (∼ 87 wavelengths) linear track and show only
small power variations when such a wide bandwidth is used.
Fading in LOS locations for omnidirectional RX antennas obeyed
the Ricean distribution with a K-factor of 10 dB, while fading in
NLOS locations can be described by the log-normal distribution
with a standard deviation of 0.65 dB. The fading depth ranges from
-3 dB to +1.5 dB relative to the mean for LOS, and -0.8 dB to +0.8
dB for NLOS. For a 60◦ directional RX antenna, fading in both
LOS and NLOS environments follows the Ricean distribution,
where the K-factor ranges from 7 dB to 17 dB for LOS, and
9 dB to 21 dB for NLOS, depending on the RX orientation in
relation to the environment and the TX [62], while the fading
depth varies between -4 dB to +2 dB relative to the mean for
both LOS and NLOS environments.
Spatial autocorrelation modeling of instantaneous total received
signal voltage amplitudes showed that the sinusoidal-exponential
distribution fits measurements in the LOS environment for both
omnidirectional and directional RX antennas. In the NLOS environment, the spatial autocorrelation can be modeled by the
exponential distribution for both omnidirectional and directional
RX antennas. Table III shows the oscillation distance/period and
spatial decay constant to represent the autocorrelation, where rapid
decorrelation of received voltage amplitudes occurred over 0.67
to 33.3 wavelengths (0.27 cm to 13.6 cm), depending on the
RX orientation in relation to the environment and the TX [62].
The short correlation distance, in general, is favorable for spatial
multiplexing in MIMO since it allows for uncorrelated spatial data
streams to be transmitted from closely-spaced (a fraction to several
wavelengths) antennas [40]. Furthermore, the small fading depth
indicates that the signal quality of an established link between a
TX and an RX will not vary much when the RX moves within a
local area on the order of a few tens of wavelengths (within 13.6
cm), as shown in Figs. 16 and 17.
Local area channel transition measurements show an initial 8
dB power loss and an overall 25 dB power loss when an RX
moves from LOS to NLOS around a building corner and along a
street canyon in an urban microcell environment. However, local
area path loss measurements (cluster measurements) suggest omnidirectional received power has a relatively stationary mean over
a relatively large area (5 m x 10 m) in LOS and NLOS scenarios,
respectively, as indicated in Table IV. Wideband mmWave signals
in typical indoor and outdoor environments were shown to fade
at rates from 0.4 dB/ms to 40 dB/s, depending on the speed and
environment. The results presented here will aid the 5G wireless
community as it develops models for small-scale fading and spatial
consistency for handoff algorithms and beam scanning techniques.
R EFERENCES
[1] T. S. Rappaport, “The wireless revolution,” IEEE Communications Magazine,
vol. 29, no. 11, pp. 52–71, Nov. 1991.
[2] F. Boccardi et al., “Five disruptive technology directions for 5G,” IEEE
Communications Magazine, vol. 52, no. 2, pp. 74–80, Feb. 2014.
[3] A. Ghosh et al., “Millimeter-wave enhanced local area systems: A high-datarate approach for future wireless networks,” IEEE Journal on Selected Areas
in Communications, vol. 32, no. 6, pp. 1152–1163, June 2014.
[4] T. S. Rappaport et al., “Overview of millimeter wave communications for
fifth-generation (5G) wireless networks-with a focus on propagation models,”
IEEE Trans. Ant. Prop., Special Issue on 5G, to appear Nov. 2017.
[5] ——, “Millimeter Wave Mobile Communications for 5G Cellular: It Will
Work!” IEEE Access, vol. 1, pp. 335–349, May 2013.
[6] ——, “Wideband millimeter-wave propagation measurements and channel
models for future wireless communication system design,” IEEE Transactions
on Communications, vol. 63, no. 9, pp. 3029–3056, Sept. 2015.
[7] ——, “Broadband millimeter-wave propagation measurements and models
using adaptive-beam antennas for outdoor urban cellular communications,”
in IEEE Trans. Ant. Prop., vol. 61, no. 4, Apr. 2013, pp. 1850–1859.
[8] G. R. MacCartney, Jr. et al., “Indoor office wideband millimeter-wave
propagation measurements and models at 28 GHz and 73 GHz for ultra-dense
5G wireless networks,” IEEE Access, vol. 3, pp. 2388–2424, Oct. 2015.
[9] S. Sun et al., “Investigation of prediction accuracy, sensitivity, and parameter
stability of large-scale propagation path loss models for 5G wireless communications (Invited Paper),” IEEE Transactions on Vehicular Technology,
vol. 65, no. 5, pp. 2843–2860, May 2016.
[10] T. A. Thomas et al., “A prediction study of path loss models from 2-73.5 GHz
in an urban-macro environment,” in 2016 IEEE 83rd Vehicular Technology
Conference (VTC2016-Spring), May 2016, pp. 1–5.
[11] O. H. Koymen et al., “Indoor mm-Wave channel measurements: Comparative
study of 2.9 GHz and 29 GHz,” in 2015 IEEE Global Telecommunications
Conference Workshops (Globecom Workshops), Dec. 2015, pp. 1–6.
[12] Y. Yoon et al., “Site specific analysis on mm-Wave radio propagation in commercial small urban,” in 2015 Int. Conf. on Information and Communication
Technology Convergence (ICTC), Oct. 2015, pp. 1312–1314.
[13] A. Maltsev et al., “Impact of polarization characteristics on 60-GHz indoor
radio communication systems,” IEEE Antennas and Wireless Propagation
Letters, vol. 9, pp. 413–416, May 2010.
[14] G. R. MacCartney, Jr. and T. S. Rappaport, “Study on 3GPP rural macrocell
path loss models for millimeter wave wireless communications,” in 2017
IEEE International Conference on Communications (ICC), May 2017.
[15] H. T. Friis, “A note on a simple transmission formula,” Proceedings of the
IRE, vol. 34, no. 5, pp. 254–256, May 1946.
[16] T. S. Rappaport, R. W. Heath, Jr., R. C. Daniels, and J. N. Murdock,
Millimeter Wave Wireless Communications. Pearson/Prentice Hall, 2015.
[17] S. Deng, G. R. MacCartney, Jr., and T. S. Rappaport, “Indoor and outdoor 5G
diffraction measurements and models at 10, 20, and 26 GHz,” in 2016 IEEE
Global Telecommunications Conference (GLOBECOM), Dec. 2016, pp. 1–7.
[18] M. Rumney, “Testing 5G: Time to throw away the cables,” Microwave
Journal, Nov. 2016.
[19] D. Solomitckii et al., “Characterizing the impact of diffuse scattering in urban millimeter-wave deployments,” IEEE Wireless Communications Letters,
vol. 5, no. 4, pp. 432–435, Aug. 2016.
[20] 3GPP, “Technical specification group radio access network; channel model
for frequency spectrum above 6 GHz (Release 14),” 3rd Generation Partnership Project (3GPP), TR 38.900 V14.2.0, Dec. 2016.
[21] S. Sun, G. R. MacCartney, Jr., and T. S. Rappaport, “A novel millimeter-wave
channel simulator and applications for 5G wireless communications,” in 2017
IEEE International Conference on Communications (ICC), May 2017.
[22] M. K. Samimi and T. S. Rappaport, “3-D millimeter-wave statistical channel
model for 5G wireless system design,” IEEE Transactions on Microwave
Theory and Techniques, vol. 64, no. 7, pp. 2207–2225, July 2016.
[23] T. S. Rappaport et al., “Investigation and comparison of 3GPP and NYUSIM
channel models for 5G wireless communications,” in 2017 IEEE 86th
Vehicular Technology Conference (VTC Fall), Sep. 2017.
[24] H. Wang and T. S. Rappaport, “A parametric formulation of the UTD
diffraction coefficient for real-time propagation prediction modeling,” IEEE
Antennas and Wireless Propagation Letters, vol. 4, pp. 253–257, 2005.
[25] A. Maltsev et al., “Characteristics of indoor millimeter-wave channel at 60
GHz in application to perspective WLAN system,” in Proceedings of the
Fourth European Conference on Antennas and Propagation, Apr. 2010.
[26] J. S. Lu et al., “Measurement and characterization of various outdoor 60
GHz diffracted and scattered paths,” in 2013 IEEE Military Communications
Conference (MILCOM), Nov. 2013, pp. 1238–1243.
[27] M. Jacob et al., “Diffraction in mm and Sub-mm Wave indoor propagation
channels,” IEEE Transactions on Microwave Theory and Techniques, vol. 60,
no. 3, pp. 833–844, Mar. 2012.
[28] T. Kleine-Ostmann et al., “Diffraction measurements at 60 GHz and 300 GHz
for modeling of future THz communication systems,” in 2012 37th International Conference on Infrared, Millimeter, and Terahertz Waves (IRMMWTHz), Sept. 2012, pp. 1–2.
[29] Y. Corre et al., “Analysis of outdoor propagation and multi-cell coverage
from ray-based simulations in sub-6 GHz and mmwave bands,” in 2016 10th
European Conference on Antennas and Propagation (EuCAP), Apr. 2016.
[30] P. H. Pathak, G. Carluccio, and M. Albani, “The uniform geometrical theory
of diffraction and some of its applications,” IEEE Antennas and Propagation
Magazine, vol. 55, no. 4, pp. 41–69, Aug. 2013.
[31] T. Mavridis et al., “Creeping wave model of diffraction of an obliquely
incident plane wave by a circular cylinder at 60 GHz,” IEEE Transactions
on Antennas and Propagation, vol. 62, no. 3, pp. 1372–1377, Mar. 2014.
[32] S. Deng, “Indoor and outdoor 5G diffraction measurements and models at
10, 20, and 26 GHz,” Master’s thesis, TR 2016-001, New York University,
New York, May 2016.
[33] T. A. Russell, C. W. Bostian, and T. S. Rappaport, “A deterministic approach
to predicting microwave diffraction by buildings for microcellular systems,”
IEEE Transactions on Antennas and Propagation, vol. 41, no. 12, pp. 1640–
1649, Dec. 1993.
[34] T. S. Rappaport, Wireless Communications: Principles and Practice, 2nd ed.
Upper Saddle River, NJ: Prentice Hall, 2002.
[35] L. Piazzi and H. L. Bertoni, “Effect of terrain on path loss in urban
environments for wireless applications,” IEEE Transactions on Antennas and
Propagation, vol. 46, no. 8, pp. 1138–1147, Aug. 1998.
[36] N. Tervo et al., “Diffraction measurements around a building corner at
10 GHz,” in 2014 1st International Conference on 5G for Ubiquitous
Connectivity (5GU), Nov. 2014, pp. 187–191.
[37] J. Ryan, G. R. MacCartney, Jr., and T. S. Rappaport, “Indoor office wideband
penetration loss measurements at 73 ghz,” in 2017 IEEE International
Conference on Communications Workshop (ICCW), May 2017, pp. 1–6.
[Online]. Available: https://arxiv.org/pdf/1703.08030.pdf
[38] G. R. MacCartney, Jr. et al., “Millimeter-wave human blockage at 73
GHz with a simple double knife-edge diffraction model and extension for
directional antennas,” in 2016 IEEE 84th Vehicular Technology Conference
(VTC2016-Fall), Sept. 2016, pp. 1–6.
[39] S. Sun et al., “Synthesizing omnidirectional antenna patterns, received power
and path loss from directional antennas for 5G millimeter-wave communications,” in 2015 IEEE Global Communications Conference (GLOBECOM),
Dec. 2015, pp. 3948–3953.
[40] ——, “MIMO for millimeter-wave wireless communications: beamforming,
spatial multiplexing, or both?” IEEE Communications Magazine, vol. 52,
no. 12, pp. 110–121, Dec. 2014.
[41] K. Sato and T. Manabe, “Estimation of propagation-path visibility for indoor
wireless lan systems under shadowing condition by human bodies,” in 48th
IEEE Vehicular Technology Conference (VTC1998), vol. 3, May 1998, pp.
2109–2113.
[42] S. Collonge, G. Zaharia, and G. El Zein, “Influence of the human activity
on wide-band characteristics of the 60 GHz indoor radio channel,” IEEE
Transactions on Wireless Communications, vol. 3, no. 6, pp. 2396–2406,
Nov. 2004.
[43] ——, “Wideband and dynamic characterization of the 60 GHz indoor radio propagation – future homeWLAN architectures,” Annales Des
Télécommunications, vol. 58, no. 3, pp. 417–447, 2003.
[44] A. Maltsev et al., “Channel Models for 60 GHz WLAN Systems,” doc.: IEEE
802.11-09/0334r8, May 2010.
[45] S. Y. Seidel and T. S. Rappaport, “Site-specific propagation prediction for
wireless in-building personal communication system design,” IEEE Transactions on Vehicular Technology, vol. 43, no. 4, pp. 879–891, Nov. 1994.
[46] M. Jacob, C. Mbianke, and T. Kürner, “Human body blockage - guidelines
for TGad MAC development,” doc.: IEEE 802.11-09/1169r0, Nov. 2009.
[47] M. Jacob, S. Priebe, and T. Kürner, “Modeling the human induced 60 GHz
channel dynamics,” doc.: IEEE 802.11-11/1170r0, Nov. 2009.
[48] METIS, “METIS Channel Model,” METIS2020, Deliverable D1.4 v3, July
2015.
[49] J. Medbo and F. Harrysson, “Channel modeling for the stationary UE
scenario,” in 2013 7th European Conference on Antennas and Propagation
(EuCAP), Apr. 2013, pp. 2811–2815.
[50] G. R. MacCartney and T. S. Rappaport, “A flexible millimeter-wave channel
sounder with absolute timing,” IEEE Journal on Selected Areas in Communications, vol. 35, no. 6, pp. 1402–1418, June 2017.
[51] J. Kunisch and J. Pamp, “Ultra-wideband double vertical knife-edge model
for obstruction of a ray by a person,” in 2008 IEEE International Conference
on Ultra-Wideband, vol. 2, Sept. 2008, pp. 17–20.
[52] Z. Wang et al., “A diffraction measurement model and particle filter tracking
method for RSS-based DFL,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 11, pp. 2391–2403, Nov. 2015.
[53] J. Medbo, J. E. Berg, and F. Harrysson, “Temporal radio channel variations
with stationary terminal,” in IEEE 60th Vehicular Technology Conference
(VTC2004-Fall), vol. 1, Sept. 2004, pp. 91–95.
[54] A. Ghosh, “The 5G mmWave radio revolution,” Microwave Journal, vol. 59,
no. 9, pp. 22–36, Sept. 2016.
[55] M. Jacob, C. Mbianke, and T. Kürner, “A dynamic 60 GHz radio channel
model for system level simulations with MAC protocols for IEEE 802.11ad,”
in IEEE International Symposium on Consumer Electronics (ISCE 2010),
June 2010, pp. 1–5.
[56] Q. Wang et al., “Ray-based analysis of small-scale fading for indoor corridor
scenarios at 15 GHz,” in 2015 Asia-Pacific Symposium on Electromagnetic
Compatibility (APEMC), May 2015, pp. 181–184.
[57] A. H. Henderson, C. J. Durkin, and G. D. Durgin, “Finding the right smallscale fading distribution for a measured indoor 2.4 GHz channel,” in 2008
IEEE Antennas and Propagation Society International Symposium, July 2008,
pp. 1–4.
[58] J. M. Romero-Jerez et al., “The fluctuating two-ray fading model for
mmwave communications,” in 2016 IEEE Globecom Workshops (GC Wkshps), Dec. 2016, pp. 1–6.
[59] S. L. Cotton and W. G. Scanlon, “Higher order statistics for lognormal
small-scale fading in mobile radio channels,” IEEE Antennas and Wireless
Propagation Letters, vol. 6, pp. 540–543, Nov. 2007.
[60] J. M. Holtzman and L. M. A. Jalloul, “Rayleigh fading effect reduction
with wideband DS/CDMA signals,” IEEE Transactions on Communications,
vol. 42, no. 234, pp. 1012–1016, Feb. 1994.
[61] G. D. Durgin, V. Kukshya, and T. S. Rappaport, “Wideband measurements of
angle and delay dispersion for outdoor and indoor peer-to-peer radio channels
at 1920 MHz,” IEEE Transactions on Antennas and Propagation, vol. 51,
no. 5, pp. 936–944, May 2003.
[62] S. Sun et al., “Millimeter wave small-scale spatial statistics in an urban microcell scenario,” in 2017 IEEE International Conference on Communications
(ICC), May 2017, pp. 1–7.
[63] M. K. Samimi and T. S. Rappaport, “Local multipath model parameters
for generating 5G millimeter-wave 3GPP-like channel impulse response,”
in 2016 10th European Conference on Antennas and Propagation (EuCAP),
Apr. 2016, pp. 1–5.
[64] D. Dupleich et al., “Investigations on fading scaling with bandwidth and
directivity at 60 GHz,” in 2017 11th European Conference on Antennas and
Propagation (EUCAP), Mar. 2017, pp. 3375–3379.
[65] M. K. Samimi et al., “28 GHz millimeter-wave ultrawideband small-scale
fading models in wireless channels,” in 2016 IEEE 83rd Vehicular Technology
Conference (VTC2016-Spring), May 2016, pp. 1–6.
[66] T. S. Rappaport, S. Y. Seidel, and K. Takamizawa, “Statistical channel impulse response models for factory and open plan building radio communicate
system design,” IEEE Transactions on Communications, vol. 39, no. 5, pp.
794–807, May 1991.
[67] Y. Zhang et al., “A Novel Spatial Autocorrelation Model of Shadow Fading
in Urban Macro Environments,” in 2008 IEEE Global Telecommunications
Conference, Nov. 2008, pp. 1–5.
[68] N. Jalden et al., “Inter- and intrasite correlations of large-scale parameters
from macrocellular measurements at 1800 MHz,” EURASIP Journal on
Wireless Communications and Networking, no. 10, pp. 1–12, July 2007.
[69] 3GPP, “Further advancements for E-UTRA physical layer aspects (Release
9),” 3rd Generation Partnership Project (3GPP), TR 36.814, Mar. 2010.
[Online]. Available: http://www.3gpp.org/DynaReport/36814.htm
[70] ——, “Technical specification group radio access network; evolved universal
terrestrial radio access (E-UTRA); mobility enhancements in heterogeneous
networks (Release 11),” 3rd Generation Partnership Project (3GPP), TR
36.839 V11.1.0, Dec. 2012.
[71] K. Guan et al., “Measurements and analysis of large-scale fading characteristics in curved subway tunnels at 920 MHz, 2400 MHz, and 5705 MHz,”
IEEE Transactions on Intelligent Transportation Systems, vol. 16, no. 5, pp.
2393–2405, Oct. 2015.
[72] V. M. Kolmonen et al., “Measurement-based evaluation of interlink correlation for indoor multiuser MIMO channels,” IEEE Antennas and Wireless
Propagation Letters, vol. 9, pp. 311–314, Apr. 2010.
[73] G. R. MacCartney, Jr. et al., “A flexible wideband millimeter-wave channel
sounder with local area and NLOS to LOS transition measurements,” in 2017
IEEE International Conference on Communications (ICC), May 2017.
[74] G. R. MacCartney, Jr., M. K. Samimi, and T. S. Rappaport, “Omnidirectional
path loss models in New York City at 28 GHz and 73 GHz,” in IEEE 25th
International Symposium on Personal Indoor and Mobile Radio Communications (PIMRC), Sept. 2014, pp. 227–331.
[75] M. K. Samimi et al., “28 GHz angle of arrival and angle of departure
analysis for outdoor cellular communications using steerable beam antennas
in New York City,” in 2013 IEEE 77th Vehicular Technology Conference
(VTC-Spring), June 2013, pp. 1–6.
[76] R. Wang et al., “Stationarity region of mm-wave channel based on outdoor
microcellular measurements at 28 GHz,” in IEEE MILCOM 2017, Oct. 2017.
| 7cs.IT
|
Emergence of Structured Behaviors from Curiosity-Based Intrinsic Motivation
Nick Haber, Damian Mrowca, Li Fei-Fei, Daniel L. K. Yamins
( [nhaber, mrowca, feifeili, yamins]@stanford.edu )
Department of Psychology and Computer Science, Stanford University, Stanford, CA 94305, USA
arXiv:1802.07461v1 [cs.LG] 21 Feb 2018
Abstract
Infants are experts at playing, with an amazing ability to generate novel structured behaviors in unstructured environments
that lack clear extrinsic reward signals. We seek to replicate
some of these abilities with a neural network that implements
curiosity-driven intrinsic motivation. Using a simple but ecologically naturalistic simulated environment in which the agent can
move and interact with objects it sees, the agent learns a world
model predicting the dynamic consequences of its actions. Simultaneously, the agent learns to take actions that adversarially
challenge the developing world model, pushing the agent to
explore novel and informative interactions with its environment.
We demonstrate that this policy leads to the self-supervised
emergence of a spectrum of complex behaviors, including ego
motion prediction, object attention, and object gathering. Moreover, the world model that the agent learns supports improved
performance on object dynamics prediction and localization
tasks. Our results are a proof-of-principle that computational
models of intrinsic motivation might account for key features
of developmental visuomotor learning in infants. Keywords:
Development learning, Curiosity, Neural Network Models
Introduction
Within the first year of life, humans exhibit a wide range of
interesting, apparently spontaneous, visuomotor behaviors —
including navigating their environment, seeking out and attending to objects, and engaging physically with these objects
in novel and surprising ways (Fantz, 1964; Twomey & Westermann, 2017; Hurley et al., 2010; Hurley & Oakes, 2015;
Goupil et al., 2016; Begus et al., 2014; Gopnik et al., 2009). In
short, young children are excellent at playing, and their ability to make sense of and (re)structure their environments sets
them apart from even the most advanced autonomous robots.
Play capacity in this period likely interacts with infants’ powerful abilities to understand and model their environment. By
six months of age or younger, infants can orient themselves
in a complex environment, account for the presence, number
and visual properties of objects they interact with, and have a
sense of how these objects behave dynamically Spelke (1985);
Stahl & Feigenson (2015); Baillargeon (2007).
But how exactly do such young children know how to
play? And how do such behaviors relate to their world-model
building abilities? One natural idea is that infants’ worldmodeling capacities are the result of built-in core systems,
including those for e.g. object attention and permanence, selflocalization, number sense, and intuitive physics (Spelke &
Kinzler, 2007). Once operational, such systems would naturally give the infant a basis on which to make judgments about
which sequences of actions would be interesting to perform.
A related but alternative idea is that the intrinsic motivation of curiosity can itself drive the development of worldmodel making (Schmidhuber, 2010). This idea relies on a
virtuous cycle in which, by seeking out novel but replicable
Figure 1: Problem setting. How do agents learn from object
interactions while moving around in the physical world?
interactions, the child pushes the boundaries of what its worldmodel-prediction systems can achieve, giving itself useful data
on which to improve and develop these systems. As worldmodeling capacity improves, what used to be novel becomes
old hat, and the cycle starts again. Related to the conception of
the “scientist in the crib” (Gopnik et al., 2009), in this account,
aside from being fun, play behaviors may be extremely useful
and highly structured, driving the self-supervised learning of
a variety of representations underlying sensory judgments and
motor planning capacities (Mitash et al., 2017).
Building on recent work in artificial intelligence (Schmidhuber, 2010; Pathak et al., 2017; Kulkarni et al., 2016; Jaderberg
et al., 2016) we make a computational model of an agent
driven by curiosity-based intrinsic motivation. We present a
simple simulated interactive environment in which an agent
can move around and physically act on objects it sees (Fig. 1).
In this world, interesting interactions are sparse unless actively
sought after. We then describe a neural network architecture
through which the agent learns a world model that seeks to
predict the consequences of its own actions. In addition, as
the agent optimizes the accuracy of its world model, a separate neural network simultaneously learns an agent action
policy that seeks to take actions that adversarially challenge
the current state of its world model. We demonstrate that
this architecture stably engages in the virtuous reinforcement
learning cycle described above, spontaneously learning to
understand self-generated ego motion and to selectively pay
attention to, localize, and interact with objects, without having
to have any of these concepts built in. This learning occurs
through emergent self-curricularization in which new capacities arise at distinct “developmental milestones.” This work is
computational proof-of-principle of how intrinsic motivation
could drive aspects of visuomotor learning and play in infants
and young children, and of how more flexible intrinsically
motivated autonomous robots might be constructed.
S t a te
St at e
World
Model
Λψ
2
Ac t i o
argmax
n
L t L t+1 L t+2 ... L t+T
st
St a t e
t+1
t+1
t+1
αego
αobj
αobj
1
S t a te
ωφ
A c t i o n Lo s s
Environment
1
A c t i o n Lo s s
Loss
Model
n
Actio
t
t
t
αego
αobj
αobj
2
s t-1
t
t
t
αego
αobj
αobj
st
1
2
Lt
s t+1
t-1
t-1
t-1
αego
αobj
αobj
1
2
Figure 2: Self-supervised curiosity model. We train a dynamical world model (blue), while simultaneously learning a loss model
(red) that predicts the world model’s loss to choose actions that lead to novel and surprising events in the environment (black).
Agent Architecture and Environment
We place an agent in a physically realistic simulated environment built in the Unity 3D simulation framework. The agent
consists of a world model and a loss model. The world model
is tasked to learn dynamics from visuomotor inputs. The loss
model tries to estimate the world model’s losses several time
steps into the future to choose actions that antagonize the
world model’s learning. Our self-supervised curiosity system
is depicted in Figure 2. We emphasize that we do not initialize our model with pretrained weights so as to explore what
world representation and behaviors emerge from this simple
antagonistic setup in a physically embodied environment.
Interaction environment
Our environment consists of a simple square room in which
an agent and several objects are initially placed randomly. The
agent is modeled as an invisible sphere that can move around
and, in discrete time steps, receives an RGB image from its
forward-facing camera. In order to model interaction with
objects requiring some attention and proximity, the agent can
apply forces and torques in all three dimensions to objects that
are both in view and within a fixed distance δ. We refer to a
state in which the agent can act on the object as a play state,
and the object as in play. Although the remaining environment
is static, the agent and objects can collide with every part.
We define a state in the state space st ∈ S to consist of the
images captured at times t and t −1 by the agent. In state st , the
agent specifies an action at ∈ A, which leads to the next state
st+1 . The action space A ∈ R2+6N is continuous. The first 2
dimensions specify ego motion, restricting agent movement to
forward/backward motion v f wd and horizontal planar rotation
vθ . The remaining 6N dimensions specify the forces fx , fy , fz
and torques τx , τy τz applied to N objects sorted from the lowerleftmost to the upper-rightmost object relative to the agent’s
field of view. This representation is unambiguous as objects
can only be acted on when in view, and if k < N objects are
in play, the environment accepts all 2 + 6N-tuples but only
applies the leftmost k force-torque pairs. All coordinates are
bounded by constants and normalized to 1.
World model
Given a slice of history ht = (st−kb , at−kb . . . st+k f , at+k f ) ∈ H,
we can describe a generalized dynamical problem by an input map ξ : H → X and a true-value map η : H → Y and
require the world model (blue in Figure 2) map ξ(h) to η(h)
— regardless of whether this is well-defined. Let ω denote
this world model, so that ω(ξ(h)) ∈ Y . For each prediction,
a loss Lwm (ω(ξ(h)), η(h)) is incurred. In theory, future prediction makes an attractive dynamical problem, with H =
{h = (st−k , at−k . . . st , at , st+1 )}, ξ(h) = (st−k , at−k . . . st , at ),
and η(h) = st+1 . In practice, we find inverse dynamical
prediction useful — filling in a missing action, instead of
predicting the future — and concretely in our case, H =
{(at−2 , st−1 , at−1 , st , at , st+1 , at+1 , st+2 )} with ξ(h) excluding
at and η(h) = ζ(at ). Here ζ zeros out the force and torque
components not corresponding to objects in play, as they have
no observable effect, and bins the action into classes, with Lwm
the softmax cross-entropy loss. We bin each dimension by
x < −0.1, −0.1 ≤ x ≤ 0.1, and x > 0.1.
We train a convolutional neural network ωφ for this task
from scratch with stochastic gradient descent, with randomly
initialized parameters φ. We use twelve convolutional layers,
with two-stride max pools every other layer and one hidden
layer to encode all states into a lower-dimensional latent space
with shared weights. The latent states {λt+i } concatenated
with the given actions {at+ j } are then input to a two-layer
MLP to predict at .
Loss model
The agent’s goal is to antagonize the world model, so if
it could predict the loss incurred at future time steps as a
function of its options, a policy could be made. In practice,
we do this explicitly, except predicting only a discretization
of the loss for ease of training. Given st and a proposed
next action a, the loss model Λ (red in Figure 2) predicts
a probability distribution over Cl discrete (via thresholding)
classes of world model loss for a set number T of future
time steps. It is penalized with a softmax cross-entropy loss
Llm (Λ(st ), (Lwm (η( ht+s ), ωφ (ξ(ht ))) | s ∈ 1 . . . T )). We use a
separate convolutional neural network Λψ with parameters ψ,
with twelve convolutional layers with two-stride max pools every other layer and one hidden layer to encode the state which
is then concatenated with the proposed at . This representation
is then used as input to a two-hidden-layer MLP to infer the
prediction. Note that all future losses aside from the first one,
depend not only on the state of the world model, but also on
future actions taken, and the loss model hence needs to predict
in expectation over future policy. The loss predictions are
usefully interpreted as loss prediction maps Λst (a) on action
space given a current state st as depicted in Figure 4.
Action policy
Given the loss prediction model, the agent can use a simple
mechanism to choose its actions. The loss model provides,
given st and a proposed next action a, T probability distributions p(k | st , a) = (p1lm (k | st ) . . . pT (k | st , at )), k ∈ {1 . . .Cl }.
Given a real-valued function σ of these T probability distributions, we can define our policy π(a|st ) as a distribution
π(a|st ) ∼ exp(βσ(p(k | st , a)),
(1)
with hyperparameter β. For our purposes, taking expectation
in loss class k and sum over all time steps is sufficient. In
practice, we execute our policy by evaluating σ for K uniform
random samples in A. We then sample from a K-way discrete
distribution with probabilities proportional to equation (1). In
choosing this policy mechanism, we opt to start with a naı̈ve
approach over using more sophisticated reinforcement learning standards in an effort to focus on studying the structure of
the proposed self-supervision signal. By explicitly predicting
loss separately for several time steps into the future, our results
admit easy visualization and interpretation.
Experiments
We randomly situate the agent and up to two objects in a square
10x10 unit room with play distance δ = 2. The agent trains
on 16 blue objects with different shapes, i.e. cones, cylinders,
cuboids, pyramids, and spheres of varied aspect ratios. We
reinitialize the play scene every 8, 000 to 30, 000 steps.
We first place the agent with one object in the room and
show that it learns to predict ego motion, and attend, localize
and navigate towards objects by evaluating the world model’s
training loss, the agent’s play state frequency, and the inverse
dynamics prediction performance on a fixed validation set.
In a second experiment we increase the number of objects
to two, and demonstrate that the agent learns to gather both
objects and further prefers to play with two objects over one
object by looking at the frequency of 1 and 2 object play states
and object-agent distances. Overarching these results is the
observation that these behaviors emerge in a specific order
akin to developmental milestones.
We compare our learned world model with curious policy
with T = 40 (LW/CP-40) against a baseline with a world
model with fixed random weights following a random policy
(RW/RP) and a baseline where the world model weights are
learned with a random policy (LW/RP).
Ego motion learning
Figure 3 (a) shows the training loss curves of LW/CP-40 and
the baselines. RW/RP does not learn well since most of its
weights are fixed to be random. LW/RP quickly converges to
a low value as it learns from a constant random distribution
without an antagonistic policy. The LW/CP-40 loss dips before increasing as the loss model first needs to learn which
actions lead to higher loss before being able to antagonize the
world model effectively. This first dip in loss corresponds to
the world model learning ego motion. The ego motion error
reported in Table 1 is close to the error reached at this point.
Emergence of object attention
As the LW/CP-40 loss increases after the initial ego motion
dip the agent starts to attend to objects which is reflected in an
increase of object interactions as shown in Figure 3 (b). At the
final stage, the agent interacts with the object about 60% of the
time which indicates that it learns to localize and attend to the
object (Table 1). At the same time, the increasing world model
loss shows that these object interactions are much harder to
predict than ego motion. The baselines almost never interact
with the object and thus experience lower ego motion losses.
Improved inverse dynamics prediction
We evaluate the inverse dynamics prediction performance on
a held out validation set gathered from the environment while
following a random action policy. To measure the models’
performance on predicting ego motion and object actions separately, we divide the validation set into two sets. The first
set contains all examples in which the agent is in a play state.
The second set consists of all remaining examples. As we can
see in Figure 3 (c) and in Table 1, LW/CP-40 and LW/RP perform well on predicting ego motion as no antagonistic policy
is necessary to encounter ego motion. However, our policy
outperforms the baselines on predicting object interactions
by a significant margin showing that focusing on object interactions does indeed improve inverse dynamics prediction
performance as seen in Figure 3 (d) and Table 1.
Improved object detection and localization
To quantify the world model’s object presence and localization
performance, we train a linear regression/logistic regression
with elastic net regularization on features from various world
model layers on an offline dataset generated by gathering
data online while following a random action policy. Half of
the object presence training data contains an object. For the
localization experiment, the second image is guaranteed to
contain the object. Both training datasets consists of 16,000
Ego motion learning
*
Emergence of object attention
object interaction learning
°
(a) World model training loss
*
°
°
*
(b) 1 object play frequencies
(c) Ego motion validation loss
(d) 1 object interaction validation loss
Figure 3: 1 object experiments. Learned world model with curious policy (LW/CP-40) is compared against learning a world
model following a random policy (LW/RP) and a world model with random weights with a random policy (RW/RP). (a) World
model cross-entropy loss during training. (b) Object play state frequency in %. (c) Ego motion prediction cross-entropy loss on
held out validation set. (d) 1 object interaction prediction cross-entropy loss on held out validation set.
Table 1: Performance comparison. Learned world model with
a curious policy (LW/CP-40) is compared against learned
world model with random policy (LW/RP) and random world
model with random policy (RW/RP). Ego motion (v f wd , vθ )
and interaction ( f , τ) accuracy in % is compared for play and
non-play states. Object frequency and presence are measured
in % and localization in mean pixel error.
Task
RW/RP
LW/RP
LW/CP-40
Play v f wd accuracy
Play vθ accuracy
Play f , τ accuracy
Non-play v f wd accuracy
Non-play vθ accuracy
Object frequency
61.8
78.8
20.8
59.2
75.0
0.2
84.5
95.2
39.8
94.2
97.9
0.4
93.6
97.9
44.4
92.3
97.9
59.0
Presence error
Localization error [px]
3.9
15.14
1.2
5.99
0.6
4.80
image pairs labeled with the object’s presence or pixel-wise 2d
position of its centroid respectively. The respective validation
and test sets comprise 8,000 image pairs each. As can be see
in Table 1 our model outperforms the baselines on the object
presence and localization task, indicating that it learns better
visual features.
Navigation and planning
In addition to object localization, the agent also exhibits navigation and planning abilities. In Figure 4 we give visualizations of loss maps projected onto the agent’s position at the
respective time. The loss prediction maps are generated by
uniformly sampling 1000 actions a from the action space A,
evaluating Λst (a) and applying a postprocessing smoothing
algorithm. We truncate the figure at five out of the 40 time
steps our loss model predicts. The loss maps show the agent
predicting a higher loss (red) for actions moving it towards the
object to reach a play state. Consequently, our curious policy
will take actions that navigate the agent closer to the object.
Emergence of multi-object interactions
At the beginning of the training of the 2 object experiment we
observe similar stages as for the 1 object experiment (Figure 5
(a)). The loss dips as the agent learns to predict its ego motion
and rises when its attention shifts towards objects which it then
interacts with. This stage is followed by a another loss increase
which corresponds to the agent gathering and playing with
both objects simultaneously. This is reflected in an increase in
2 object play time (Figure 5 (c)) over 1 object play time (Figure
5 (b)). Consequently, the average distance between the agent
and the objects decreases over time as seen in Figure 5 (c). It
drops to about 2 units which equals the maximum interaction
distance. The LW/RP baseline quickly drops and flattens out.
LW/CP-20 with T = 20 instead of T = 40 learns to interact
with one object but not with two objects simultaneously.
Discussion and Future Work
We observe that a simple and general intrinsic motivation
mechanism based on adversarially antagonizing the loss of a
dynamically constructed model of the world allows an agent
to stably generate a spectrum of emergent naturalistic behaviors. Through self-curricularization in an active learning Set-
t
t+1
t+2
t+3
t+4
Figure 4: Navigation and planning behavior. The loss model predicts higher loss if the agent turns towards the object. Red colors
correspond to high and blue colors to low loss predictions. The center of the heat map corresponds to the agents position.
Ego motion learning
2 object loss
Emergence of object attention
1 object learning
°
*
1 object loss
2 object learning
x
(a) World model training loss
*
°
x
(b) 1 object play frequencies
x
(c) 2 object play frequencies
x
(d) Object-agent distances
Figure 5: 2 object experiments. Learned world model with curious policy (LW/CP-40) is compared against the same setup but
with T = 20 (LW/CP-20) and a learned world model following a random policy (LW/RP). (a) World model cross-entropy loss
during training. (b) Object play state frequency for 1 object in %. (c) Object play state frequency for 2 objects in %. (d) Average
distance between agent and objects in Unity units.
tles (2011) process the agent achieves several “developmental
milestones” of suitably increasing complexity as it learns to
“play”. Starting with random actions, it quickly learns the
dynamics of its own ego motion. Then, without being given an
explicit supervision signal as to the presence or location of an
object, it discards ego motion prediction as boring and begins
to focus its attention on objects, which are more interesting.
Lastly, when multiple objects are available, it gathers the objects so as to bring them into interaction range of each other.
Throughout, the agent finds its way towards a more challenging data distribution that is at each moment just hard enough
to expose the agent to new situations, but still understandable
and exploitable by the agent. This intrinsically motived policy
leads to performance gains in its understanding both of the
object dynamics, as well as other tasks which the system was
not explicitly learning.
This occurs without any pretrained visual backbone — the
visual system world model was intentionally not initialized
with filter weights pretrained on (e.g.) ImageNet classification.
This result constitutes partial progress in replacing the training
of a visual backbone through a task such as large-scale image
classification with an interactive self-supervised task and is
a proof-of-concept that more complex milestones can be potentially reached while developing an understanding of object
categories and physical relations.
From a machine learning perspective, this combination of
spontaneous behavior leading to an improved world model is
well suited to designing agents that must act effectively in the
many real-world reinforcement learning scenarios in which rewards are sparse or potentially unknown. Here, we ultimately
seek to develop algorithms that will control autonomous robots
that learn to operate in complex unpredictable environments.
From a cognitive science perspective, these results suggest a
route toward using intrinsically motivated learning systems to
model emergence of spontaneous behavior in young children.
In this domain, we seek to make computational models that
describe key aspects of real infant learning.
However, to truly achieve either of these goals, a variety of
limitations of the current work will need to be overcome in
future work. First, to make the connection to cognitive science
more realistic, our environment and agent themselves need
to be more realistic. On the one hand, better graphics and
physics, with more varied and interesting visual objects, will
be important to allow better transfer the learned behavior to
real-world visuomotor interactions. It will also be important to
create a properly embodied agent with visible arms and tactile
feedback, allowing for more realistic interactions. In our current work, which we stress is more a proof-of-principle than
a full cognitive model, we fail to address the fact that infants
have severely limited mobility and motor control. To be able
to make realistic predictions for actual infant developmental
milestones, it will be especially critical to more realistically
model the known developmental trajectory of the motor system. In addition, including other animate agents is another
way for more complex interactions, but potentially also for
better learning through imitation.
Second, it will be important to improve the reinforcement
learning techniques used to better handle more complex interactions beyond those demonstrated here. For interactions that
are part of a larger experiment, e.g. placing an object on a table or a ramp and then watching it fall, more sophisticated RL
policies than those used here are likely to be necessary, with
better ability to handle temporally extended reward schedules.
It will also likely be necessary to use recurrent networks to
meet working memory demands in such scenarios.
Third, our world model needs to use better representations
to improve at predicting such complex interactions. Our current approach especially suffers from degenerate cases in the
inverse dynamical prediction problem — the problem does
not correspond to a well-defined map. For example, when
given a sequence of images in which an object rests on the
ground, the action sequence is under-determined: the agent
could have been pushing down on the object, or not. Though
the latent-space approach of (Pathak et al., 2017) is meant in
part to ameliorate this issue, we have not yet found an entirely
effective solution in this context. Choices for how to blend
components of a generalized inverse dynamical prediction task
and the interaction of loss terms seem to be key in resolving
this. Taking these next steps will not only help us to understand how infants learn, but also to develop AI systems that
learn without human supervision.
Acknowledgements
This work was supported by an Understanding Human Cognition award from the James S. McDonnell foundation (DLKY),
a Simons Collaboration on the Global Brain grant (DLKY), a
Berry Foundation postdoctoral fellowship (NH), and by hardware donation from the NVIDIA Corporation.
References
Baillargeon, R. (2007). The acquisition of physical knowledge in infancy: A summary in eight lessons. In Blackwell
handbook of childhood cognitive development (pp. 47–83).
Blackwell Publishers Ltd.
Begus, K., Gliga, T., & Southgate, V. (2014, 10). Infants learn
what they want to learn: Responding to infant pointing leads
to superior learning. PLOS ONE, 9(10), 1-4.
Fantz, R. L. (1964). Visual experience in infants: Decreased
attention to familiar patterns relative to novel ones. Science,
146(3644), 668-670.
Gopnik, A., Meltzoff, A., & Kuhl, P. (2009). The scientist in
the crib: Minds, brains, and how children learn. HarperCollins.
Goupil, L., Romand-Monnier, M., & Kouider, S. (2016).
Infants ask for help when they know they dont know. Proceedings of the National Academy of Sciences, 113(13).
Hurley, K. B., Kovack-Lesh, K. A., & Oakes, L. M. (2010,
Dec). The influence of pets on infants’ processing of cat
and dog images. Infant Behav Dev, 33(4), 619–628.
Hurley, K. B., & Oakes, L. M. (2015, Jan). Experience and
distribution of attention: Pet exposure and infants’ scanning
of animal images. J Cogn Dev, 16(1), 11–30.
Jaderberg, M., Mnih, V., Czarnecki, W. M., Schaul, T., Leibo,
J. Z., Silver, D., & Kavukcuoglu, K. (2016). Reinforcement
learning with unsupervised auxiliary tasks. arXiv preprint
arXiv:1611.05397.
Kulkarni, T. D., Narasimhan, K., Saeedi, A., & Tenenbaum,
J. B. (2016). Hierarchical deep reinforcement learning:
Integrating temporal abstraction and intrinsic motivation.
CoRR, abs/1604.06057.
Mitash, C., E. Bekris, K., & Boularias, A. (2017, 03). A
self-supervised learning system for object detection using
physics simulation and multi-view pose estimation.
Pathak, D., Agrawal, P., Efros, A. A., & Darrell, T. (2017).
Curiosity-driven exploration by self-supervised prediction.
arXiv preprint arXiv:1705.05363.
Schmidhuber, J. (2010, Sept). Formal theory of creativity, fun,
and intrinsic motivation (1990 – 2010). IEEE Transactions
on Autonomous Mental Development, 2(3), 230-247.
Settles, B. (2011). Active learning (Vol. 18). Morgan &
Claypool Publishers.
Spelke, E. S. (1985). Object permanence in five-month-old
infants. Cognition, 20(3), 191–208.
Spelke, E. S., & Kinzler, K. D. (2007). Core knowledge.
Developmental science, 10(1), 89–96.
Stahl, A. E., & Feigenson, L. (2015, Apr). Cognitive development. Observing the unexpected enhances infants’ learning
and exploration. Science, 348(6230), 91–94.
Twomey, K. E., & Westermann, G. (2017, Oct). Curiositybased learning in infants: a neurocomputational approach.
Dev Sci.
| 1cs.CV
|
arXiv:1607.02482v1 [math.AC] 8 Jul 2016
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL
RINGS
MARK W. ROGERS AND CAMERON WICKHAM
Abstract. For a Noetherian local ring pR, mq having a finite residue field of cardinality q,
we study the connections between the ideal Z pRq of Rrxs, which is the set of polynomials
that vanish on R, and the ideal Z pmq, the polynomials
ś that vanish on m, using what we
call π-polynomials: polynomials of the form πpxq “ qi“1 px ´ ci q, where c1 , . . . , cq is a set
of representatives of the residue classes of m. When R is Henselian we prove that πpRq “ m
and show that a generating set for Z pRq may be obtained from a generating set for Z pmq
by composing with πpxq. When m is principal and has index of nilpotency e, we prove that
if e ď q then Z pmq “ px, mqe , and if Ş
e “ q ` 1 then Z pmq “ px, mqe ` pxq ´ mq´1 xq. When
R is finite, we prove that Z pRq “ qi“1 Z pci ` mq is a minimal primary decomposition.
We determine when Z pRq is nonzero, regular, or principal, respectively, and do the same
for Z pmq. We prove that when R is complete, repeated application of πpxq ` x to elements
of R will produce a sequence converging to the roots of πpxq. We show that Z pRq is the
intersection of the principal ideals generated by the π-polynomials.
1. Introduction
One of the surprising facts about finite rings is that a polynomial can be nonzero and
yet induce the zero function. An interesting first example is provided by Fermat’s Little
Theorem: If p is prime, then the nonzero polynomial xp ´ x induces the zero function on Zp .
Using this we can build obvious examples, such as ppxp ´ xq and pxp ´ xq2 on Zp2 , and more
surprising examples, such as pxp ´ xqp ´ pp´1 pxp ´ xq on Zpp`1 ; see Corollary 4.5 for more
information on these examples. For a ring R, it’s easy to see that the set of polynomials in
Rrxs that induce the zero function on R is an ideal of Rrxs; we call this the zero-function
ideal of R, denoted Z pRq. The zero-function ideal has been studied, often with particular
focus on the rings R “ Zpn , for its connection with integer-valued polynomials and functions
induced by polynomials [F, J, P, Z] and coding theory [LRS].
For most of our paper, pR, mq is a Noetherian local ring; we will see that Z pRq is only
nonzero when theˇ residue
field R “ R{m is finite, so we focus most of our attention on this
ˇ
case and let q “ ˇRˇ. Many authors have studied the problem of finding a generating set for
Z pRq, most often in the case R “ Zpn [D, B, F, L, NW, P, W]. In this paper, we argue that
in many cases, focus should be shifted from Z pRq to the simpler ideal Z pmq, which is the
set of polynomials that induce the zero function on m. The connection between the two is
the ideal Z pR, mq, the set of polynomials that take elements of R into m; it’s easy to show
that Z pR, mq “ pxq ´ x, mq. Certainly polynomials in Z pR, mq can be composed with those
in Z pmq to obtain polynomials in Z pRq; visually, Z pmq ˝ Z pR, mq Ď Z pRq. One of the
main themes of this paper is to show that in some sense the opposite is true: A generating
Date: July 11, 2016.
2010 Mathematics Subject Classification. Primary 13M10, 13B25; Secondary 11C08, 13J15, 13E10,
13F20.
Key words and phrases. finite ring, polynomial functions, vanishing polynomials, Artinian rings.
1
2
MARK W. ROGERS AND CAMERON WICKHAM
set for Z pRq can be obtained by composing generators of Z pR, mq with generators of Z pmq
(see Theorem 4.2).
Example 1.1. Let R “ Z9 , so that m “ p3q and R – Z3 . By direct computation or by
Theorem 4.4, Z pmq “ px, mq2 . In Lemma 2.5 we easily find that with πpxq “ x3 ´ x,
Z pR, mq “ pπpxq, mq “ px3 ´ x, 3q; from this, according to Theorem 4.2, we deduce
Z pRq “ Z pmq ˝ Z pR, mq “ px3 ´ x, 3q2 .
This makes it clear that the simpler ideal Z pmq controls the structure of the generating set
of the more complicated ideal Z pRq.
The polynomial xq ´ x has played an important role in the `research
on zero-functions,
˘
q
due primarily to the fact that the image of x ´ x generates Z R and, to a lesser extent,
the fact that when R is finite, xq ´x maps R surjectively onto m. See A. Bandini’s paper [B]
for applications of surjectivity; we generalize the surjectivity result in Corollary 2.11 and
apply it in Theorem 3.2 and Proposition 4.1. A secondary theme of this paper is that there
is actually a class of polynomials with these properties that can play the
of xq ´ x; we
śrole
q
call these π-polynomials, defined to be polynomials of the form πpxq “ i“1 px ´ ci q where
c1 , . . . , cq is any set of representatives of the residue classes of m. As we will see, xq ´ x
is a π-polynomial when R is Henselian (which holds true in the common case where R is
finite). If R is complete, we also provide a computational way of obtaining the factorization
of any π-polynomial, such as xq ´ x itself. In the case of xq ´ x, this method is as simple
as choosing any element of R and repeatedly taking the qth power; the results converge to
a root of xq ´ x. (See Theorems 2.10, 5.2.)
The ideal Z pRq for R “ Zpn was studied as early as 1929 by L. E. Dickson [D, Theorem 27]; in that work, the polynomials in Z pRq were referred to as residual polynomials. Dickson found a generating set for Z pZpn q when n ď p. In our notation, he found
Z pRq “ pπpxq, mqn , where πpxq “ xp ´ x and m “ pR. We generalize and recover this work
as another application of our Theorem 4.2: We show in Theorem 4.4 and its corollary that
if pR, mq is an Artinian local ring with a principal maximal ideal having index of nilpotency
e ď q, then Z pmq “ px, mqe , and thus Z pRq “ pπpxq, mqe for any π-polynomial. When
e ą q, the situation is more complicated, but we take care of the case e “ q ` 1; the result
is related to results on Z pRq for specific rings R ([B, Theorem 2.1] and [L, Theorem II]).
As further indication of the importance of π-polynomials and Z pmq, we provide two
additional results. Under suitable conditions, we prove in Proposition 2.7 that Z pRq is
the intersection of the principal ideals generated by the π-polynomials, and in Proposition
2.9 we provide a minimal primary decomposition of Z pRq as the intersection of the ideals
Z pci ` mq, where c1 , . . . , cq is a set of represetatives of the residue classes of m. Since
generators for Z pci ` mq may be obtained from generators for Z pmq by composition with
x ´ ci , this shows that a primary decomposition for Z pRq may be obtained from knowing
only a generating set for Z pmq. This result on primary decomposition is a generalization
of results from the paper [P] of G. Peruginelli, which was concerned with the ring R “ Zpn .
The remaining theme of our paper is provided in Theorem 3.2, Theorem 3.3, and Corollary
3.4, where we identify conditions under which Z pRq is nonzero, principal, and regular, and
the same for Z pmq; these results explain why we often focus our attention on finite rings.
The results generalize, have some overlap with, and were inspired by R. Gilmer’s paper [G].
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL RINGS
3
2. Zero-Function Ideals and π-polynomials
We begin with a precise definition of the zero-function ideal of a ring, and we define a
class of polynomials that plays an important role the study of zero-function ideals.
Definition 2.1. Let R be a commutative ring with identity, let S be a subset of R, and let
J be an ideal of R. The set Z pS, Jq of polynomials in Rrxs which map S into J is easily
seen to be an ideal of Rrxs. When the ideal J is omitted, it is assumed to be zero. The
focus of this paper is on Z pRq, which we call the zero-function ideal of R.
Definition 2.2. Suppose the local ring pR, mq has a finite residue field. If c1 , . . . ,ś
cq is any set
of representatives of the residue classes of m, then we call the polynomial πpxq “ qi“1 px´ci q
a π-polynomial for R.
Example 2.3. We mentioned in the introduction that for R “ Z9 , Z pmq “ px, 3q2 .
However, for R “ Z8 , Z pmq Ľ px, 2q3 . In fact, according to Theorem 4.4 and a few
brief calculations, Z pmq “ px, 2q3 ` px2 ´ 2xq “ px2 ´ 2x, 4xq. We may then use Theorem 4.2 to compose with the π-polynomial πpxq “ x2 ´ x and conclude that Z pRq “
ppx2 ´ xq2 ´ 2px2 ´ xq, 4px2 ´ xqq.
The following basic result is a generalized factor theorem that will be useful in a few
proofs. The part related to units appears in Gilmer’s proof of Theorem 4 in [G].
Lemma 2.4. Let R be a commutative ring with identity. If f pxq P Rrxs is a polynomial
with roots c1 , c2 , . . . , cn such that each difference ci ´ cj (i ‰ j) is either a regular element
or a unit, then px ´ c1 qpx ´ c2 q . . . px ´ cn q divides f pxq in Rrxs.
Proof. By the Factor Theorem, f pxq “ px ´ c1 qf1 pxq for some f1 pxq P Rrxs. Substituting
c2 in for x, we obtain 0 “ f pc2 q “ pc2 ´ c1 qf1 pc2 q. Since c2 ´ c1 is either a regular element
or a unit, c2 is a root of f1 pxq; we obtain some f2 pxq P Rrxs with f1 pxq “ px ´ c2 qf2 pxq. So
f pxq “ px ´ c1 qpx ´ c2 qf2 pxq. By continuing to evaluate at the c’s in this manner, we arrive
at f pxq “ px´ c1 qpx´ c2 q ¨ ¨ ¨ px´ cn qfn pxq for some polynomial fn pxq P Rrxs, as desired.
Convention. Throughout this paper, let R be a Noetherian local ring with maximal ideal
m. Unless otherwise specified, the residue field R{m will be denoted by R. The image in R
of an element r P R will be denoted by r. If the residue field is finite, c1 , . . . , cq will denote
a set ofś
representatives of the residue classes of m and πpxq will denote the π-polynomial
πpxq “ qi“1 px ´ ci q.
The following lemma may be viewed as a generalization of Fermat’s Little Theorem; it
is well-known, at least in special cases, as remarked by D. J. Lewis in [L]. A simple but
important consequence of this lemma is that πpRq Ď m. Later in Corollary 2.11 we will
show that if R is Henselian, then πpRq “ m.
Lemma 2.5. If pR, mq is a Noetherian local ring with finite residue field of cardinality q,
then Z pR, mq “ pπpxq, mq for any π-polynomial πpxq.
` ˘
Proof. Let f pxq P Z pR, mq; then f pxq P Z R . By Lemma 2.4, f pxq is in the ideal generated
by πpxq in Rrxs; pull this back to Rrxs to get f pxq P pπpxq, mq.
For the opposite containment,
certainly the constant polynomials in m are in Z pR, mq.
śq
Now suppose πpxq “ i“1 px ´ ci q. Since any element in R is congruent modulo m to one
of the ci , the polynomial πpxq is in Z pR, mq, as desired.
4
MARK W. ROGERS AND CAMERON WICKHAM
Example 2.6. Let R “ Zp5q (Z localized at the prime ideal p5q), so that m “ p5q, R “ Z5 ,
and q “ 5. The polynomial xq ´ x is not a π-polynomial since it doesn’t factor completely
over R Ď Q: x5 ´ x “ xpx ´ 1qpx ` 1qpx2 ` 1q. However, by Lagrange’s Theorem applied
to the group of units of R, it is still true that x5 ´ x P Z pR, mq. According to the lemma,
we expect x5 ´ x P pπpxq, mq for any π-polynomial. In fact, if for example we let πpxq “
px ´ 2qpx ` 1qxpx ´ 1qpx ` 2q, then x5 ´ x “ πpxq ` 5px3 ´ 1q P pπpxq, mq.
In the next result we show that the zero-functions are precisely those polynomials that
are multiples of each π-polynomial.
Proposition 2.7. Let pR, mq be a Noetherian ring with finite residue field R of cardinality
q. The zero-function ideal
Ş of R is the intersection of the principal ideals generated by the
π-polynomials: Z pRq “ pπpxqq.
Proof. The fact that any zero-function f pxq P Z pRq is a multiple
of any π-polynomial πpxq
Ş
follows immediately
from
Lemma
2.4.
This
shows
Z
pRq
Ď
pπpxqq.
Ş
Let f pxq P pπpxqq and let r P R; we show f prq “ 0. We may extend r to a set r, c2 , . . . , cq
of representatives of the residue classes of m, and thus πpxq “ px ´ rqpx ´ c2 q ¨ ¨ ¨ px ´ cq q
is a π-polynomial having r as a root. Since f pxq is a multiple of πpxq, f prq “ 0. Thus
f pxq P Z pRq.
In the following result we give a minimal primary decomposition of Z pRq if R is finite.
First, we give a simple example:
Example 2.8. For R “ Z9 , as mentioned in the introduction, we have Z pmq “ px, 3q2 “
px2 , 3xq, so the proposition below gives the following minimal primary decomposition of
Z pRq:
ppx3 ´ xq2 , 3px3 ´ xqq “ px2 , 3xq X ppx ´ 1q2 , 3px ´ 1qq X ppx ´ 2q2 , 3px ´ 2qq,
where the ideals on the right are primary for the maximal ideals px, 3q, px ´ 1, 3q, and
px ´ 2, 3q, respectively.
Proposition 2.9. Let pR, mq be a finite local ring with residue field R of cardinality
q. Let c1 , . . . , cq be a set of representatives of the residue classes of m. Then Z pRq “
Ş
q
i“1 Z pci ` mq is a minimal primary decomposition of Z pRq. For each i, the associated
prime of Z pci ` mq is the maximal ideal px ´ ci , mq.
Proof. Certainly any zero-function
on R is also zero on each ci ` m. For the other inclusion,
Ť
use the fact that R “ qi“1 pci ` mq. For the statement about associated primes, let e
be the index of nilpotency of m, so that me “ 0 but me´1 ‰ 0, and use the fact that
px ´ ci , mqe Ď Z pci ` mq.
As for the minimality
of the decomposition,
Ş
ś let j be an integer between
Ş 1 and q; we show
that Z pRq Ĺ i‰j Z pci ` mq. Let hpxq “ i‰j px ´ ci qe ; then hpxq P i‰j Z pci ` mq since
ś
me “ 0. To see that hpxq is not a zero-function, note that hpcj q “ i‰j pcj ´ ci qe is a
product of units, and is thus nonzero.
Next we provide an equivalent way to view π-polynomials, provided the ring is Henselian;
of course, this holds for the finite local rings in which we are mainly interested. For an
example where the two conditions below are not equivalent, see Example 2.6. This theorem
is also needed in our proof that π-polynomials map R surjectively onto m.
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL RINGS
5
Theorem 2.10. Let R be a Henselian local ring with finite residue field R of cardinality q.
For any polynomial πpxq P Rrxs, the following statements are equivalent:
(i) The polynomial πpxq is a π-polynomial.
(ii) The polynomial πpxq is monic and maps to xq ´ x in Rrxs.
Proof. Let πpxq be any π-polynomial. Since R is a field with q elements, by Lagrange’s
theorem on the group of units of R, xq ´ x is a zero-function on R. By Lemma 2.4, πpxq
divides xq ´ x in Rrxs. Since these are monic polynomials of the same degree, they are
equal.
For the converse, suppose πpxq is any monic polynomial withśπpxq “ xq ´ x. Let R “
td1 , . . . , dq u; as discussed in the previous paragraph, xq ´ x “ qi“1 px ´ di q in Rrxs. By
Hensel’s Lemma, this factorization of πpxq can be pulled
śqback to a factorization in Rrxs:
There exist ci in R with ci “ di such that πpxq “
i“1 px ´ ci q. Thus πpxq is a πpolynomial.
In the following corollary, we improve upon part of Lemma 2.5 by showing that the
induced function π : R Ñ m is actually surjective when R is Henselian. This generalizes
Lemma 1.3 of [B], where A. Bandini proved that, for any prime p, πpRq “ m in case R “ Zpn
and πpxq “ xp ´ x. We use this corollary in our Theorem 3.2, where we characterize finite
rings with principal zero-function ideals, expanding upon Gilmer [G]. It is used again in
Proposition 4.1, which is fundamental for our Theorem 4.2, which shows that generators
for Z pRq may be obtained by composing generators for Z pmq with a π-polynomial.
Corollary 2.11. If pR, mq is a Henselian local ring with finite residue field R of cardinality
q, then πpRq “ πpcq “ m for any π-polynomial πpxq and any coset c of m.
Proof. We show that πpcq Ď πpRq Ď m Ď πpcq. The first containment is clear since c Ď R,
and we saw the second containment in Lemma 2.5. For the final containment, let m P m.
By Theorem 2.10, the polynomial πpxq ´ m is still a π-polynomial, and thus it factors over
R: πpxq ´ m “ px ´ c1 qpx ´ c2 q ¨ ¨ ¨ px ´ cq q. This shows that for each i, πpci q “ m. Since
c1 , c2 , . . . , cq is a set of representatives of the residue classes of m, one of them, say cj , is in
c. Thus m “ πpcj q P πpcq, as desired.
3. When Z pRq and Z pmq are Nonzero, Regular, or Principal
In the upcoming Theorems 3.2 and 3.3 we will use the following result from B. R. McDonald. McDonald states and proves the theorem for any finite local ring pR, mq, but the
theorem and proof still hold when R is just Artinian. The notation McDonald uses is different from ours but the part we will use is that over an Artinian local ring, any regular
polynomial is an associate of a monic polynomial. McDonald writes µf where we would
write f , the image of f in Rrxs.
Theorem 3.1. [M, Theorem XIII.6, p. 259] Let f be a regular polynomial in Rrxs. Then
there is a monic polynomial f ˚ with µf “ µf ˚ and, for an element a in R, f paq “ 0 if and
only if f ˚ paq “ 0. Furthermore, there is a unit v in Rrxs with vf “ f ˚ .
One of the motivations for the current paper is Theorem 4 from [G], which states that
if pR, mq is a zero-dimensional local ring, then Z pRq is principal if and only if either R is
infinite (when Z pRq “ 0) or R is a finite field (when Z pRq is generated by xq ´ x.) The
following result recovers half of Gilmer’s result, using some of the same ideas but a few
different ones as well. For example, Gilmer used a result of E. Snapper; instead we use the
6
MARK W. ROGERS AND CAMERON WICKHAM
result of McDonald mentioned above. Also, we make the connection with π-polynomials
and Z pmq. The other half of Gilmer’s result is recovered in Theorem 3.3.
Theorem 3.2. Let pR, mq be a finite local ring and let πpxq be any π-polynomial for R.
The following statements are equivalent:
(1) R is a field.
(2) Z pRq is principal.
(3) Z pRq “ pπpxqq.
(4) Z pmq is principal.
(5) Z pmq “ pxq.
Proof. If R is a field then Rrxs is a principal ideal domain, so Z pRq is principal.
Assume Z pRq is principal. Since πpRq Ď m and me “ 0 for some e ě 1, Z pRq contains
regular polynomials, such as πpxqe ; thus, the generator of Z pRq must be regular. According
to Theorem 3.1, we may assume the generator is monic: Z pRq “ pf pxqq for some monic
polynomial in Rrxs. Let r be a nonzero element in the annihilator of m, so that, by
Lemma 2.5, rπpxq is a polynomial of degree q in Z pRq “ pf pxqq. This forces f pxq to have
degree at most q, and since we know that all polynomials in Z pRq are multiples of πpxq
(Proposition 2.7), the degree of f pxq is exactly q. Since f pxq is a monic multiple of πpxq
with the same degree, f pxq “ πpxq.
If Z pRq “ pπpxqq, then according to Corollary 2.11, 0 “ πpRq “ m. From this we easily
conclude that Z pmq “ pxq, and thus Z pmq is principal.
If Z pmq is principal, we use an argument similar to the part where we assumed Z pRq
is principal. Since Z pmq contains xe for some e ě 1, Z pmq contains regular elements, so
the generator of Z pmq is regular, and we may assume it is monic: Z pmq “ pf pxqq for some
monic f pxq P Rrxs. Let r be a nonzero element in the annihilator of m, so that rx P Z pmq.
This forces f pxq to have degree at most 1. Since f pxq is monic and f p0q “ 0, f pxq “ x, as
desired.
If Z pmq “ x, then m “ 0, so R is a field.
In the next proposition it becomes clear how the conditions of being Artinian or having
a finite residue field affect Z pRq and Z pmq. We will use the concept of the embedding
dimension of R, denoted edim R; this is the minimal number of generators of the maximal
ideal. Recall that depth R ď dim R ď edim R.
Theorem 3.3. Let pR, mq be a Noetherian local ring.
(1) Z pmq contains nonzero polynomials if and only if depth R “ 0.
(2) Z pRq contains nonzero polynomials if and only if depth R “ 0 and R is finite.
(3) Z pmq contains regular polynomials if and only if dim R “ 0.
(4) Z pRq contains regular polynomials if and only if dim R “ 0 and R is finite.
(5) Z pmq is generated by a regular polynomial if and only if edim R “ 0.
(6) Z pRq is generated by a regular polynomial if and only if edim R “ 0 and R is finite.
In order to make the similarity with the other parts more clear, parts (4), (5), and (6)
of the previous proposition were not stated as concisely as possible. Before we present the
proof, we state and prove a corollary to clarify those three parts. In the development of this
paper, we were particularly interested in the monic polynomials in Z pRq, so this corollary
explains why we were mainly focused on finite rings.
Corollary 3.4. Let pR, mq be a Noetherian local ring.
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL RINGS
(3)
(4)
(5)
(6)
Z pmq
Z pRq
Z pmq
Z pRq
7
contains regular polynomials if and only if R is Artinian.
contains regular polynomials if and only if R is a finite ring.
is generated by a regular polynomial if and only if R is a field.
is generated by a regular polynomial if and only if R is a finite field.
Proof of Corollary 3.4. Part (3) is clear since a Noetherian local ring is Artinian if and
only if it has dimension 0. Part (4) follows from the proposition since a zero dimensional
Noetherian local ring has a finite residue field if and only if the ring is finite. Parts (5) and
(6) follow since edim R “ 0 if and only if R is a field.
Proof of Theorem 3.3. (1): If R has depth 0 then since R is Noetherian, there is a nonzero
element m that annihilates m; thus mx P Z pmq.
If R does not have depth 0, then R contains a regular element t. Let gpxq “ g0 ` g1 x `
¨ ¨ ¨ ` gn xn P Z pmq. Since gptq “ gpt2 q “ gpt3 q “ ¨ ¨ ¨ “ gptn`1 q “ 0, there is a matrix
equation
fi» fi » fi
»
1 t1
t2
¨¨¨
tn
g0
0
4
2n ffi — ffi
—1 t2
—0ffi
t
¨
¨
¨
t
g
1
ffi— ffi — ffi
—
— ..
..
..
.. ffi — .. ffi “ — .. ffi
..
–.
.
.
.
. fl – . fl –.fl
1 tn`1 tpn`1q2 ¨ ¨ ¨
tpn`1qn
The determinant of this Vandermonde matrix is
ź
ź
ptj ´ ti q “
1ďiăjďn`1
gn
0
ti ptj´i ´ 1q.
1ďiăjďn`1
Since each tj´i ´ 1 is a unit, this determinant is an associate of a power of t; namely tk
where k “ npn ` 1qpn ` 2q{6. After multiplying both sides of the matrix equation by the
adjugate and dividing by the unit, we find that tk gi “ 0 for each i. Since tk is regular, the
polynomial gpxq is zero, as desired.
(2): Assume R “ tc1 , . . . , cq u and R has depth 0. Since R is Noetherian with depth 0,
there
śq is a nonzero element m that annihilates m. By Lemma 2.5, the polynomial πpxq “
i“1 px ´ ci q is in Z pR, mq, so the polynomial mπpxq is a nonzero element of Z pRq.
For the converse, if R is not finite, then there is an infinite sequence of elements tcn uně1
such that no two come from the same residue class of m; thus each difference cj ´ ci (j ‰ i)
is a unit. On the other hand, if R does not have depth 0, then there is a regular element
t P m. Consider the sequence ttn uně1 ; for i ą j we have ti ´ tj “ tj pti´j ´ 1q. Since tj is
regular and ti´j ´ 1 is a unit, each difference ti ´ tj (i ‰ j) is regular.
In either case, we may apply Lemma 2.4 to conclude that any nonzero polynomial in the
zero-function ideal has arbitrarily high degree. This is a contradiction, so the zero-function
ideal contains only the zero polynomial, as desired.
(3): If R has dimension 0 then me “ 0 for some e ě 1. Thus xe P Z pmq.
Conversely, if R has positive dimension, then for any minimal prime ideal p of R, R{p is
an integral domain of positive dimension; in particular, it has positive depth. According to
Part (1), Z pm{pq is the zero ideal of of R{p. The image of any f pxq P Z pmq in pR{pqrxs is
in Z pm{pq, and thus f pxq P prxs. Since the coefficients of f pxq are in every minimal prime,
they are all nilpotent, so f pxq is not regular.
0. Let e be such that me “ 0 and let R “ tc1 , . . . , cq u.
(4): Suppose R is finite and dim R “ ś
By Lemma 2.5, the polynomial πpxq “ qi“1 px´ci q is in Z pR, mq, so the regular polynomial
πpxqe is in Z pRq.
8
MARK W. ROGERS AND CAMERON WICKHAM
Conversely, suppose either R is infinite or dim R ě 1. If R is infinite, then Z pRq does
not contain regular polynomials since it is the zero ideal, by Part (2). If dim R ě 1, then
for every minimal prime ideal p of R, R{p is an integral domain of positive dimension; in
particular, it has positive depth. According to Part (2), Z pR{pq is the zero ideal of of R{p.
The image of any f pxq P Z pRq in pR{pqrxs is in Z pR{pq, and thus f pxq P prxs. This shows
that f pxq cannot be regular.
(5): If edim R “ 0 then R is a field, so m “ 0 and Z pmq “ pxq.
Conversely, assume Z pmq “ pf pxqq for some regular polynomial f pxq P Rrxs. By Part
(3), dim R “ 0, so me “ 0 for some e ě 1. Due to this and the result from McDonald
(our Theorem 3.1), we may assume f pxq is monic. By Part (1), depth R “ 0, so there is an
element m P m with mx P Z pmq. This shows that the monic polynomial f pxq must have
degree 1; the only choice is f pxq “ x. From this we see that m “ 0 so R is a field.
(6): If edim R “ 0 and R is finite, then R is a finite field, so Z pRq “ pπpxqq for any
π-polynomial, by Lemma 2.5.
Conversely, assume Z pRq “ pf pxqq for some regular polynomial f pxq. From Part (4) we
know that dim R “ 0 and R is finite; this implies that R is finite. By Theorem 3.2, R is a
finite field, so edim R “ 0.
In the following example we illustrate the use of this theorem and contrast the behavior
of the zero-function ideals over rings with finite and infinite residue fields.
Example 3.5. The ring R “ Z2 JS, T K{pS 2 , ST q is a complete Noetherian local ring with
depth zero, dimension one, and a finite residue field with q “ 2 elements; let s and t be
the images of S and T in R. According to Theorem 3.3, Z pmq is nonzero but does not
contain regular polynomials, and the same goes for Z pRq. We argue that Z pmq “ psxq
and conclude that Z pRq “ pspx2 ´ xqq by applying Theorem 4.2. Certainly Z pmq Ě psxq,
since the annihilator of the maximal ideal of R is sR. For the opposite containment, let
R̃ “ R{sR – Z2 JT K, a local Noetherian ring of positive depth. If f pxq P Z pmq, then
f˜ P Z pm̃q, where m̃ and f˜ are the images of m and f in R̃ and R̃rxs. By Theorem 3.3 (1),
Z pm̃q “ 0, so we conclude that f P psq. Since f p0q “ 0, f P psq X pxq “ psxq, as desired.
If we switch to an infinite coefficient ring and residue field, say, Q instead of Z2 , we
still have Z pmq “ psxq (nonzero but containing no regular polynomials). However, Theorem 4.2 does not apply (for one thing, π-polynomials don’t exist). In fact, Theorem 3.3 (2)
guarantees Z pRq “ 0 instead of Z pRq “ pspx2 ´ xqq.
4. Obtaining Z pRq from Z pmq; Applications
The following proposition is the key to our main result, Theorem 4.2.
Proposition 4.1. Let pR, mq be a Henselian local ring with finite residue field R and let
πpxq be an arbitrary π-polynomial. Any f pxq P Z pRq may be written in the form
f pxq “ p0 pπpxqq ` xp1 pπpxqq ` x2 p2 pπpxqq ` ¨ ¨ ¨ ` xq´1 pq´1 pπpxqq
with each pi pxq P Z pmq.
Proof. Let f pxq P Z pRq. In the polynomial ring Rrx, ys “ Rrysrxs, f may be divided
by the monic polynomial πpxq ´ y, so that f pxq “ Qpx, yqpπpxq ´ yq ` Rpx, yq for some
Qpx, yq, Rpx, yq P Rrx, ys with Rpx, yq “ 0 or the degree of Rpx, yq with respect to x is
less than q. Now set y “ ppxq to obtain f pxq “ p0 pπpxqq ` xp1 pπpxqq ` x2 p2 pπpxqq ` ¨ ¨ ¨ `
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL RINGS
9
xq´1 pq´1 pπpxqq where the polynomials pi pyq P Rrys are the coefficients of the powers of x
in Rpx, yq.
It remains to see that each pi pxq P Z pmq. Let m P m. Since (according to Corollary 2.11)
π maps each coset of m onto m, there exists a set c1 , c2 , . . . , cq of representatives of the
residue classes of m, with πpci q “ m for each ci ; each difference ci ´ cj (i ‰ j) is a unit.
Since f pxq P Z pRq, we may evaluate f pxq at ci for each i from 1 to q to obtain
0 “ p0 pmq ` ci p1 pmq ` c2i p2 pmq ` ¨ ¨ ¨ ` ciq´1 pq´1 pmq.
In matrix form, this system becomes
» fi »
1 c1 c21 ¨ ¨ ¨
0
—0ffi —1 c2 c2 ¨ ¨ ¨
2
— ffi —
— .. ffi “ — .. ..
.. . .
–.fl –. .
.
.
2
0
1 cq cq ¨ ¨ ¨
fi
fi»
c1q´1
p0 pmq
ffi
—
c2q´1 ffi
ffi — p1 pmq ffi
ffi
..
.. ffi —
fl
.
. fl–
q´1
pq´1 pmq
cq
ś
The matrix is a Vandermonde matrix, and its determinant is 1ďiăjďq pcj ´ ci q, which is
a unit since it is a product of units. Thus, the matrix is invertible, so each pi pmq “ 0, as
desired.
An application of the previous result, we come to our main theorem, which states that generators for Z pRq can be obtained by composing generators for Z pmq with any π-polynomial,
which we could roughly describe by writing Z pRq “ Z pmq ˝ Z pR, mq, if one keeps in mind
Lemma 2.5 which states that Z pR, mq “ pπpxq, mq. An obvious consequence of the next
theorem is that the minimal number of generators of Z pRq is less than or equal to the
minimal number of generators of Z pmq.
Theorem 4.2. Suppose pR, mq is a Henselian local ring with finite residue field R of cardinality q and let πpxq be an arbitrary π-polynomial. If Z pmq “ pF1 pxq, . . . , Fn pxqq then
Z pRq “ pF1 pπpxqq, . . . , Fn pπpxqqq.
Proof. Since πpRq Ď m, certainly Z pRq Ě pF1 pπpxqq, . . . , Fn pπpxqqq. Now let f pxq P
Z pRq. Use Proposition 4.1 to write f pxq “ p0 pπpxqq ` xp1 pπpxqq ` x2 p2 pπpxqq ` ¨ ¨ ¨ `
xq´1 pq´1 pπpxqq with each pi pxq P Z pmq. Since each pi pxq is an Rrxs-linear combination of
F1 pxq, . . . , Fn pxq, each pi pπpxqq is an Rrxs-linear (actually Rrπpxqs-linear) combination of
F1 pπpxqq, . . . , Fn pπpxqq. Since f pxq is an Rrxs-linear combination of the pi pπpxqq, the proof
is complete.
Remark 4.3. The equality Z pRq “ Z pmq ˝ Z pR, mq should not be taken too literally.
Certainly polynomials in Z pmq composed with polynomials in Z pR, mq are in Z pRq, but
it’s not true that every polynomial in Z pRq can be obtained in that way. For example,
xpx2 ´ xq P Z pZ2 q, but since it’s degree is not even, it does not equal f px2 ´ xq for any
polynomial f pxq.
As mentioned in the introduction, the theorem below is a version of results of Dickson [D,
p. 22, Theorem 27], Bandini [B, Theorem 2.1], and Lewis [L, Theorem II], adapted for Z pmq
rather than Z pRq, and for finite local rings rather than specific rings. We then recover the
results for Z pRq in Corollary 4.5 as an application of our main theorem, Theorem 4.2.
Theorem 4.4. Let pR, mq be a finite local ring with principal maximal ideal m “ pmq; set
q “ |R{m|. Suppose e is the index of nilpotency of m. If e ď q then Z pmq “ px, mqe ; if
e “ q ` 1, then Z pmq “ px, mqe ` pxq ´ mq´1 xq.
10
MARK W. ROGERS AND CAMERON WICKHAM
Proof. We prove the first result using induction on e. The base case e “ 1 is clear, since then
R is a field, m “ 0, and Z pmq “ pxq. Assume the result is true for rings whose maximal
ideal has index of nilpotency e ´ 1 ď q; we prove the result for a ring whose maximal
ideal has index `of nilpotency
e ď q. The containment Ě is clear. Let f pxq P Z pmq;
˘
then f pxq P Z R{me´1 . By induction, f pxq P px, mqe´1 , and thus f pxq P px, mqe´1 .
ře´1 k e´1´k
We have f pxq “ k“0
x m
fk pxq for some fk pxq P Rrxs; it remains to see that each
fk pxq P px, mq, i.e. that fk p0q P m.
For each r P R,
0 “ f prmq “
e´1
ÿ
prmqk me´1´k fk prmq “ me´1
k“0
e´1
ÿ
r k fk prmq.
k“0
ře´1
ře´1 k
Since the annihilator of me´1 is m, k“0 r k fk prmq P m, and thus k“0
r fk p0q P m. This
` ˘
` ˘
ře´1
k
q
shows that k“0 fk p0qx P Z R . Since by Lemma 2.5 Z R “ px ´ xq, this polynomial
with degree less than q must be the zero polynomial. Therefore each fk p0q P m, as desired.
For the second result, the only part of the containment Z pmq Ě px, mqe ` pxq ´ mq´1 xq
that is not clear is xq ´ mq´1 x P Z pmq; for this, take any rm P m and compute prmqq ´
mq´1 prmq “ mq pr q ´rq P mq`1 “ 0. For the opposite containment, assume f pxq P Z pmq and
reduce module me´1
to obtain a similar expression for f pxq, and again deduce that
` as
˘ above
ře´1
ře´1
k
q
k
q
k“0 fk p0qx “ upx ´ xq
k“0 fk p0qx P Z R “ px ´ xq. Since e ´ 1 “ q, we must have
for some unit u P R; thus each fk p0q “ 0 except for fq p0q “ u and f1 p0q “ ´u . Define
polynomials gk pxq identical to fk pxq except for gq pxq “ fq pxq ´ u and g1 pxq “ f1 pxq ` u.
Now the constant term of each gk pxq is in m and we have
f pxq “
e´1
ÿ
xk me´1´k fk pxq “ upxq ´ mq´1 xq `
k“0
e´1
ÿ
xk me´1´k gk pxq
k“0
so that f pxq P px, mqe ` pxq ´ mq´1 xq, as desired.
The following corollary follows immediately from the theorem and Theorem 4.2.
Corollary 4.5. Let pR, mq be a finite local ring with principal maximal ideal m “ pmq; set
q “ |R{m|. Suppose e is the index of nilpotency of m, and let πpxq be any π-polynomial. If
e ď q then Z pRq “ pπpxq, mqe ; if e “ q ` 1 then Z pRq “ pπpxq, mqe ` pπpxqq ´ mq´1 πpxqq.
5. Factoring π-Polynomials
The following lemma is the heart of a more constructive approach (Theorem 5.2) to
the converse part of the proof of Theorem 2.10, which gave two equivalent conditions for
π-polynomials. Note that according to this lemma, if me “ 0, then for any r P R, the
sequence tpn prqu stabilizes at n “ e ´ 1.
Lemma 5.1. Let pR, mq be a Noetherian local ring with finite residue field R of cardinality
q, and let πpxq be any polynomial mapping to xq ´ x in Rrxs. Let p0 pxq “ x and pn pxq “
πppn´1 pxqq ` pn´1 pxq, so that pn pxq denotes the function obtained by successively applying
the function πpxq ` x, n times. For every n ě 1,
pn pxq ´ pn´1 pxq P Z pR, mn q .
POLYNOMIALS INDUCING THE ZERO FUNCTION ON LOCAL RINGS
11
Proof. We use induction on
` n. For
˘ the base case (n “ 1) just note that p1 pxq ´ p0 pxq “
pπpxq ` xq ´ x “ πpxq P Z R, m1 by Lemma 2.5.
n
Now assume the induction hypothesis:
`
˘For some n ě 1, pn pxq ´ pn´1 pxq P Z pR, m q.q We
n`1
show that pn`1 pxq ´ pn pxq P Z R, m
. Since πpxq is a polynomial mapping to x ´ x
modulo mrxs, there is some mpxq P mrxs such that πpxq “ xq ´ x ´ mpxq; thus pn pxq may
also be viewed as applying xq ´ mpxq, n times. For simplicity of notation, set c “ pn pxq,
a “ pn´1 pxq, and b “ c ´ a, so that c “ πpaq ` a “ aq ´ mpaq. We have
pn`1 pxq ´ pn pxq “ πppn pxqq ` pn pxq ´ pn pxq
“ πpcq
“ cq ´ c ´ mpcq
“ pa ` bqq ´ c ´ mpcq
ˆ ˙
q q´2 2
q
q´1
“ a ` qa b `
a b ` ¨ ¨ ¨ ` bq ´ c ´ mpcq
2
ˆ ˙
q q´2 2
q´1
“ qa b `
a b ` ¨ ¨ ¨ ` bq ` mpaq ´ mpcq
2
since aq ´ c `“ mpaq. ˘By induction, b P Z pR, mn q, and since `q P m, we˘ see that the first
term is in Z R, mn`1 . Since b P Z pR, mn q and n ě 1, b2 P Z R, mn`1 , which takes care
of all but the last two terms: mpaq ´ mpcq.
řdeg mpxq
Now mpaq ´ mpcq “ i“1
mi pai ´ ci q, where mi is the coefficient of xi in mpxq, and is
thus in m. Since ´b “ a´c is a factor
of a˘i ´ci for all positive integers i and ´b P Z pR, mn q,
`
it follows that mpaq ´ mpcq P Z R, mn`1 , as desired.
The following theorem provides, in particular, a more constructive approach to the proof
of the result in Theorem 2.10 which states that any monic polynomial πpxq mapping to
xq ´ x is actually a π-polynomial. In a ring with me “ 0, it allows discovery of the roots by
successively applying the function πpxq ` x (e ´ 1 times) to representatives of the residue
classes of m. When πpxq “ xq ´ x, this amounts to successively taking qth powers. In the
case of a finite ring, the resulting roots of xq ´ x are called Teichmüller elements in Jian
Jun Jiang’s paper [J].
Theorem 5.2. Let pR, mq be a complete Noetherian local ring with finite residue field R “
tc1 , . . . , cq u. Let πpxq be any polynomial mapping to xq ´ x in Rrxs and let pn pxq be the
function obtained by applying πpxq`x successively, n times. The limit limnÑ8 pn pci q exists.
Set di “ limnÑ8 pn pci q; then di is a root of πpxq and di “ ci . If πpxq is monic, then there
is a factorization
πpxq “ px ´ d1 qpx ´ d2 q ¨ ¨ ¨ px ´ dq q,
and thus πpxq is a π-polynomial.
Proof. The limit exists since, by Lemma 5.1, the sequence tpn pci quně1 is a Cauchy sequence.
For any r P R, πprq ` r and r are in the same coset of m, since πprq “ r q ´ r ´ mprq P m. We
may apply this fact successively, beginning with r “ ci , to see that each pn pci q is congruent
to ci modulo m. Since m is closed under the m-adic topology, we conclude that di “ ci .
To see that πpdi q “ 0, use the Cauchy sequence mentioned above and the fact that
polynomials are continuous under the m-adic topology:
πpdi q “ πplim pn pci qq “ lim πppn pci qq “ limppn`1 pci q ´ pn pci qq “ 0.
12
MARK W. ROGERS AND CAMERON WICKHAM
If πpxq is monic then it must have degree q; an application of Lemma 2.4 completes the
proof.
Example 5.3. With R “ Z125 , we have q “ 5 and e “ 3. We can choose elements 0, 1, 2, 3, 4
to be representatives of the elements of R{m “ Z5 . We factor the polynomial
πpxq “ x5 ` 5x4 ` 40x3 ` 85x2 ` 24x ` 50 “ x5 ´ x ´ mpxq
where mpxq “ ´p5x4 ` 40x3 ` 85x2 ` 25x ` 50q. Applying p2 pxq to 0, 1, 2, 3, 4 yields
50, 31, 72, 18, 74. According to the theorem, πpxq factors in Rrxs as
πpxq “ px ´ 50qpx ´ 31qpx ´ 72qpx ´ 18qpx ´ 74q.
Acknowledgment: The authors thank Richard Belshoff, Jung-Chen Liu and Les Reid for
helpful discussions and suggestions.
References
Bandini, A., Functions f : Z{p Z ÝÑ Z{pn Z induced by polynomials of Zrxs, Annali di Matematica
181 (2002), 95–104. doi:10.1007/s102310100031
[]D
Dickson, L. E., Introduction to the Theory of Numbers, The University of Chicago Press, 1929.
[]F
Frisch, S., Polynomial Functions on Finite Commutative Rings, Lecture Notes in Pure and Appl.
Mathematics 205, Dekker 1999, 323–336.
[]G
Gilmer, R., The Ideal of Polynomials Vanishing on a Commutative Ring, Proc. Amer. Math. Soc. 127
(1999), no. 5, 1265–1267. doi:10.1090/S0002-9939-99-04634-1
[]J
Jiang, Jian Jun, On the Number Counting of Polynomial Functions, J. Math. Research & Exp. 30
(2010), no. 2, 241–248. doi:10.3770/j.issn:1000-341X.2010.02.006
[]L
Lewis, D. J., Ideals and Polynomial Functions, Amer. J. Math. 78 (1956), no. 1, 71–77.
doi:10.2307/2372485
[]LRS Lahtonen, J., Ryu, J., and Suvitie, E., On the Degree of the Inverse of Quadratic Permutation Polynomial Interleavers, IEEE Trans. Inf. Theory 58 (2012), no. 6, 3925–3932. doi:10.1109/TIT.2012.2190492
[]M McDonald, B. R., Finite Rings with Identity, Pure and Applied Mathematics, New York, Marcel
Dekker Inc, 1974.
[]NW Niven, I. and Warren, L. J., A Generalization of Fermat’s Theorem, Proc. Amer. Math. Soc. 8 (1957),
no. 2, 306–313. doi:10.1090/S0002-9939-1957-0086804-0
[]P
Peruginelli, G., Primary decomposition of the ideal of polynomials whose fixed divisor is divisible by a
prime power, J. Alg. 398 (2014), 227–242. doi:10.1016/j.jalgebra.2013.09.016
[]W Werner, N. J., Polynomials that kill each element of a finite ring, J. Alg. and Its Appl. 13 (2014), no.
3, 1350111-1–12. doi:10.1142/S0219498813501119
[]Z
Zhang, Q., Polynomial functions and permutation polynomials over some finite commutative rings,
J. Number Theory 105 (2004), 192–202. doi:10.1016/j.jnt.2003.09.009
[]B
n
Department of Mathematics, Missouri State University, Springfield, MO 65897, USA
E-mail address: [email protected]
Department of Mathematics, Missouri State University, Springfield, MO 65897, USA
E-mail address: [email protected]
| 0math.AC
|
Calcul matriciel généralisé sur les
domaines de Prüfer
arXiv:1508.00345v1 [math.AC] 3 Aug 2015
Gema M. Díaz-Toca, Henri Lombardi, 27 juillet 2015
Résumé
Dans cet article nous présentons un algorithme pour calculer la
forme de Hermite d’une pseudo-matrice sur un domaine de Prüfer. Ceci
nous donne des preuves constructives des principaux résultats théoriques
concernant les modules de présentation finie et projectifs de type fini
sur le domaines de Prüfer, et nous permet de discuter la résolution
des systèmes linéaires sur les domaines de Prüfer. Nous généralisons
ainsi la méthodologie développée par Henri Cohen pour les domaines
de Dedekind. Nous présentons également des résultats concernant la
réduction de Smith sur les domaines de Prüfer de dimension 6 1.
Abstract
In this paper, we first present an algorithm for computing the
Hermite normal form of pseudo-matrices over Prüfer domains. This
algorithm allows us to provide constructive proofs of the main theoretical results on finitely presented modules over Prüfer domains and to
discuss the resolution of linear systems. In some sense, we generalize the
methodology developed by Henri Cohen for Dedekind domains. Finally,
we present some results over Prüfer domains of dimension one about
the Smith normal form.
Introduction
La résolution algorithmique des systèmes linéaires sur les corps ou sur
les anneaux principaux est classique, et elle est équivalente au traitement
des matrices par des manipulations élémentaires (et des manipulations de
Bezout dans le cas des anneaux principaux) de façon à les ramener à une
forme réduite convenable.
On vise ici la généralisation de ce type de procédé à des domaines de Prüfer
arbitraires.
À cet effet, on adapte pour un domaine de Prüfer arbitraire le calcul matriciel
généralisé d’Henri Cohen [Cohen, Chapitre 1], qu’il a introduit dans le cadre
du traitement algorithmique des anneaux d’entiers de corps de nombres.
Notre but est de montrer qu’avec un calcul matriciel généralisé on peut
obtenir pour un domaine de Prüfer l’analogue des réductions classiques pour
un domaine de Bezout.
2
Calcul matriciel généralisé
Ceci donne un traitement algorithmique systématique et hh agréable ii, qui
permet d’obtenir, comme conséquences du calcul matriciel considéré, les
théorèmes généraux sur les domaines de Prüfer.
En particulier cela donne un moyen pratique pour discuter les systèmes
linéaires à coefficients et inconnues dans le domaine de Prüfer considéré.
Du point de vue du Calcul Formel, notre approche permet de traiter des
situations plus générales que les approches usuelles. Par exemple, pour les
anneaux d’entiers de corps de nombres, comme le véritable ingrédient de nos
algorithmes se limite à l’inversion des idéaux de type fini, notre approche
permet de traiter le cas où la factorisation du discriminant est impossible
en pratique, et plus généralement le cas où le calcul d’une base d’entiers
sur Z est trop coûteux.
Notons que pour les domaines de Dedekind, et plus généralement pour les
domaines de Prüfer de dimension 6 1, on obtient des résultats plus précis,
du type de la réduction de Smith des matrices sur les anneaux principaux.
Cela sera discuté dans la section 4. Les méthodes ici relèvent des calculs
modulaires et sont les seules vraiment efficaces pour éviter l’explosion de la
taille des objets dans les calculs intermédiaires.
1. Quelques rappels
Le premier paragraphe de cette section donne des définitions classiques
que nous donnons si nécessaire sous forme constructive. Le deuxième est
consacré à des rappels concernant les calculs avec les idéaux de type fini
sur les domaines de Prüfer. Le troisième rappelle les principaux résultats
concernant les calculs matriciels sur les domaines de Bezout.
Notre référence pour les résultats classiques exposés dans cette section est
le livre [Modules], dans lequel tous les résultats sont démontrés de manière
algorithmique.
Quelques définitions
Un anneau A est dit zéro-dimensionnel si l’axiome suivant est vérifié
∀a ∈ A, ∃n ∈ N ∃x ∈ A, xn (1 − ax) = 0.
Un anneau intègre A est dit de dimension 6 1 lorsque pour tout b =
6 0
dans A, l’anneau quotient A/hbi est zéro-dimensionnel.
Sur un anneau A arbitraire, un idéal de type fini a = ha1 , . . . , an i est dit
P
localement principal s’il existe s1 , . . ., sn ∈ A tels que i∈J1..nK si = 1
et si a ⊆ hai i pour chaque si .
1. Quelques rappels
3
Un anneau A arbitraire est dit arithmétique si tout idéal de type fini est
localement principal.
Un idéal de type fini a = ha1 , . . . , an i est dit inversible s’il existe un élément régulier c et un idéal de type fini b tel que ab = hci. Il revient au
même de dire que a est localement principal et contient un élément régulier.
Un domaine de Prüfer est un anneau arithmétique intègre. De manière
équivalente, c’est un anneau intègre dans lequel tout idéal de type fini non
nul est inversible. Dans ce cas, pour tout a 6= 0 dans l’idéal de type fini a,
le transporteur (hai : a) est de type fini, avec l’égalité a (hai : a) = hai.
Une matrice A = (aij ) sur un anneau arbitraire est dite en forme de
Smith lorsque :
– tous les coefficients aij sont nuls sauf éventuellement des hh coefficients
diagonnaux ii ai = aii ,
– et ai divise ai+1 pour i > 1.
Lorsque la matrice M admet une réduction de Smith, le module Coker M est
isomorphe à la somme directe des A/hai i et les idéaux hai i sont uniquement
déterminés en vertu du fait 1.1.
Pour un anneau intègre, cela signifie que les ai sont eux-mêmes bien déterminés hh à une unité multiplicative près ii.
On appelle anneau de Smith un anneau A sur lequel toute matrice M
est équivalente à une matrice en forme de Smith.
L’idéal déterminantiel d’ordre k d’une matrice M est l’idéal de type
fini Dk (M ) engendré par les mineurs d’ordre k de M .
L’idéal de Fitting d’ordre k d’un module de présentation finie P , conoyau
d’une matrice M ∈ Mn,m (A), est défini par l’égalité Fk (P ) := Dn−k (M ). Il
ne dépend pas de la présentation choisie pour P .
On rappelle les résultats suivants.
1.1. Fait. ([Modules, théorème V-9.1], [MRR, théorème V-2.4])
Soient a1 ⊆ · · · ⊆ an et b1 ⊆ · · · ⊆ bm des idéaux de A avec n 6 m. Si un
module P est isomorphe à la fois à A/a1 ⊕· · ·⊕A/an et A/b1 ⊕· · ·⊕A/bm ,
alors on a
1. bk = A pour n < k 6 m ;
2. et bk = ak pour 1 6 k 6 n.
1.2. Fait. ([ACMC, exercice IV-16])
Soit P ' A/a1 ⊕ · · · ⊕ A/an avec des idéaux de type fini a1 ⊇ · · · ⊇ an .
Alors pour k ∈ J1..nK, on a Fn−k (P ) = a1 · · · ak .
4
Calcul matriciel généralisé
1.3. Fait. ([Modules, exercice XVI-9])
Sur un anneau arithmétique zéro-dimensionnel, toute matrice se ramène
par manipulations élémentaires de lignes et de colonnes à une matrice en
forme de Smith.
Idéaux fractionnaires de type fini
Il est souvent confortable de considérer, sur un pied d’égalité avec les idéaux
de type fini, les idéaux fractionnaires de type fini.
Pour un anneau intègre A de corps de fractions K, nous noterons Iff(A)
l’ensemble des idéaux fractionnaires de type fini de A, c’est-à-dire l’ensemble
des sous-A-modules de type fini de K.
Un idéal fractionnaire de type fini b peut toujours s’écrire
b=
1
a
a=
1
a
ha1 , . . . , an i avec a1 , . . . , an ∈ A, et a ∈ A∗ .
On dit qu’il est localement principal lorsque a est localement principal.
Il admet alors la même matrice de localisation principale (à coefficients
dans A) que a, avec les mêmes propriétés.
Le produit de deux idéaux fractionnaires est défini de façon naturelle et
l’on a xa yb = xy ab pour x, y ∈ K.
Concernant les transporteurs, il faut distinguer entre l’idéal fractionnaire
(a : b)K = { x ∈ K | xb ⊆ a }
et l’idéal (au sens usuel) (a : b)A = { x ∈ A | xb ⊆ a }.
Un idéal fractionnaire localement principal non nul b admet un inverse b−1
au sens du produit défini dans Iff(A) : c’est l’idéal fractionnaire (A : b)K .
Notons que l’application bilinéaire θ : b × b−1 → A, (x, y) 7→ xy est une
dualité, c’est-à-dire qu’elle permet d’identifier b−1 au A-module dual de b.
Un idéal fractionnaire de type fini contenu dans A est un idéal de type fini
usuel et il est appelé un idéal entier si l’on est dans le contexte des idéaux
fractionnaires.
Calculs sur les idéaux de type fini dans les domaines
de Prüfer
Matrice de localisation principale
Un anneau intègre Z est un domaine de Prüfer d’un point de vue algorithmique lorsque l’on dispose d’un algorithme général qui permet d’inverser
un idéal de type fini arbitraire.
Comme l’a remarqué Dedekind lorsqu’il insiste sur ce qu’il estime être la
propriété la plus décisive dans les anneaux d’entiers de corps de nombres,
on peut formuler l’inversibilité de l’idéal a = ha1 , . . . , an i ⊆ Z (les ai 6= 0)
de la manière suivante.
1. Quelques rappels
5
1. Il existe γ1 , . . ., γn dans K tels que
dans Z.
P
i
γi ai = 1 et chacun des γi aj est
Deux formulations équivalentes sont les suivantes.
2. L’idéal a est localement principal, autrement dit il existe s1 , . . ., sn ∈ A
de somme 1 tels que l’on ait si a ⊆ hai i pour chaque si non nul.
3. Il existe une matrice de localisation principale pour (a1 , . . . , an )
i.e., une matrice C = (cij ) ∈ Mn (Z) satisfaisant
– Tr(C) = 1,
– chaque ligne de C est proportionnelle à [ a1 . . . an ], i.e.
ci,` ci,j
= 0 pour tous i, j, `.
a` aj
Pour passer de 1 à 2 on pose si = γi ai , et dans l’autre sens γi =
Pour passer de 2 à 3 on pose cij =
si
ai .
si aj
ai .
La matrice de localisation principale permet de nombreux calculs, comme
indiqués dans [Modules, théorème IX-2.3, lemme IX-2.5, et exercice XVI-8].
En particulier pour un anneau intègre, on a C 2 = C (matrice de projection
de rang 1) et In − C est une matrice de présentation de l’idéal a pour le
système générateur (a1 , . . . , an ). En tant que Z-module, a est isomorphe à
l’image de C dans Zn .
Notons que, lorsque la divisibilité est explicite dans Z, une matrice de
localisation principale pour (a1 , . . . , an ) est connue à partir de sa diagonale.
Nous ferons cette hypothèse de divisibilité explicite dans la suite .
Le système générateur (a1 , . . . , an ) d’un idéal de type fini sera toujours
accompagné d’une liste d’éléments si de somme 1 qui satisfont les inclusions
si a ⊆ hai i.
Sans l’hypothèse de divisibilité explicite, il faudrait travailler avec les matrices de localisation principale elles-mêmes, qui sont des objets plus lourds à
manipuler.
Pour qu’un anneau intègre Z soit un domaine de Prüfer, il suffit que les
idéaux à deux générateurs soient inversibles, ce qui s’exprime comme suit.
Pour tous a, b ∈ Z, il existe u, v, s, t ∈ Z satisfaisant :
s + t = 1, sa = vb, tb = wa.
Autrement dit, on a pour (a, b) la matrice de localisation principale
t
v
(1)
w
.
s
Supposons a 6= 0. Alors ts = vw, ht, wi = at ha, bi et ht, wi ht, vi = hti. Donc,
−1
−1
si t, a =
6 0, ht, wi = 1t ht, vi et ha, bi = a1 ht, vi (notation des idéaux
fractionnaires).
6
Calcul matriciel généralisé
Opérations élémentaires sur les idéaux de type fini
Nous supposons travailler avec un domaine de Prüfer explicite. Autrement
dit une boite noire 1 donne pour touteP
suite finie non nulle (a1 , . . . , an ) = (a)
une suite finie (s1 , . . . , sn ) = (s) avec i si = 1 et si aj ∈ hai i pour tous i, j.
Nous supposons aussi que le domaine de Prüfer est à divisibilité explicite.
Tout ce qui est écrit par la suite fonctionne aussi bien avec des idéaux
fractionnaires de type fini qu’avec des idéaux de type fini entiers. Nous
parlerons donc simplement d’idéaux de type fini.
Signification de sk = 0. Si sk = 0, en additionnant
les égalités si ak =
P
ci,k ai pour i 6= k, on obtient l’égalité ak = i6=k ci,k ai , autrement dit le
générateur ak de l’idéal a est superflu. En supprimant la ligne et la colonne
correspondante, on obtient une matrice de localisation principale pour les
ai restants.
Inversement, si l’on rajoute un générateur combinaison linéaire de ceux
déjà donnés, on peut obtenir une matrice de localisation principale pour le
nouveau système générateur en rajoutant une ligne nulle pour le nouveau
générateur, et la colonne correspondant au nouveau générateur qui exprime
la combinaison linéaire comme indiqué ci-avant.
Pour a et b de type fini, on a 1 ∈ (a : b)Z + (b : a)Z (voir [Modules,
exercice XVI-8]). On écrit a = ha1 , . . . , an i et b = hb1 , . . . , bp i. On a des si
de somme 1 tels que si a ⊆ hai i, des tj de somme 1 tels que tj b ⊆ hbj i, et
pour chaque couple (i, j) deux éléments uij et vij de somme 1 tels que
uij hai , bj i ⊆ hai i et vij hai , bj i ⊆ hbj i .
On obtient donc
uij si tj (a + b) ⊆ hai i ⊆ a et vij si tj (a + b) ⊆ hbj i ⊆ b.
P
P
On prend s = ij uij si tj et t = ij vij si tj . On a alors s + t = 1 et
s(a + b) ⊆ a (i.e. sb ⊆ a) et t(a + b) ⊆ b (i.e. ta ⊆ b).
Notons que lorsque p = 1 la matrice de localisation principale pour (b1 ) est
simplement I1 .
Somme (pgcd). Avec les mêmes données qu’au point précédent, on voit
que si l’on pose
P
P
ui = j uij si tj et
vj = i vij si tj ,
P
P
on a ui (a + b) ⊆ hai i, vj (a + b) ⊆ hbj i et i ui + j vj = 1.
Ainsi la liste (u1 , . . . , un , v1 , . . . , vp ) convient pour le système générateur
(a1 , . . . , an , b1 , . . . , bp ) de l’idéal a + b.
1. Dans les cas que l’on traite sur machine, cette boite noire cache un algorithme.
1. Quelques rappels
7
NB : cette procédure explique pourquoi on peut se ramener au cas où la
boite noire fonctionne uniquement pour les idéaux à deux générateurs.
Intersection (ppcm). Avec les mêmes données qu’aux points précédents,
on a sb + ta = a ∩ b . En effet, on a évidemment sb ⊆ b ∩ a et ta ⊆
b ∩ a. Et pour l’autre inclusion sb + ta ⊇ a ∩ b, si m ∈ a ∩ b, alors
m = (s + t)m = sm + tm ∈ sb + ta.
Inverse ([Modules, lemme IX-2.5]).
alors on a
b1
..
hai hbi = hai avec .
bn
Soit a =
P
i
ai xi = [ a1 · · · an ]X,
x1
.
= C .. = CX
xn
(où C est une matrice de localisation principale pour (a)). En effet
X
ai bi = [ a1 · · · an ] C X = [ a1 · · · an ] X = a,
i
donc a ∈ ba. En outre
X
X
X
bi aj =
cik xk aj =
cij xk ak = cij
xk ak = cij a,
k
k
k
donc ba ⊆ hai. Ainsi ba = hai.
Test d’appartenance. Si a = ha1 , . . . , an i et si s1 , . . . , sn ∈ Z vérifient
P
si a ⊆ hai i et i si = 1, alors x ∈ a si, et seulement si, chaque si x ∈ hai i.
Comme on suppose que Z est à divisibilité explicite on a un test d’appartenance aux idéaux fractionnaires, et donc aussi un test pour l’inclusion d’un
idéal fractionnaire dans un autre.
NB : un calcul plus général mais plus compliqué, sans l’hypothèse que Z
est intègre, est donné dans l’implication 1 ⇒ 3 du théorème XVI-2.2 du
livre modules)
Formes réduites de matrices sur un domaine de Bezout
On considère un domaine de Bezout B (cas particulier de domaine de Prüfer).
1.4. Définition. Une matrice de Bezout de taille n est une matrice
carrée, égale à la matrice identité, à l’exception de quatre coefficients en
aii aij
positions (i, i), (j, j), (i, j) et (j, i) avec 1 6 i < j 6 n et
= 1.
aji ajj
Nous pouvons noter cette matrice Bz(n, i, j; aii , aij , aji , ajj ). Par exemple
8
Calcul matriciel généralisé
(avec des cases vides en guise de zéros)
1
u
v
1
B = Bz(6, 2, 4; u, v, s, t) =
s
t
1
avec
ut − sv = 1.
1
Effectuer une manipulation de Bezout sur une matrice M c’est la multiplier à gauche (manipulation de lignes) ou à droite (manipulation de
colonnes) par une matrice de Bezout.
u v
Si l’on a dans un anneau une égalité matricielle [ a b ]
= [ g 0 ], alors
s t
pour une matrice A ∈ M6,n (A) ayant des coefficients a et b dans la ligne
k en positions (k, 2) et (k, 4), effectuer le produit A B revient à faire les
manipulations de colonnes simultanées suivantes sur la matrice A :
C2 ← uC2 + sC4
C4 ← vC2 + tC4
Le résultat sera (entre autres) que les coefficients en position (k, 2) et (k, 4)
seront remplacés par g et 0.
De la même manière, effectuer un produit B 0 A par une matrice de Bezout
du style B 0 = tB permettra de remplacer un couple de coefficients (a, b)
situés sur une même colonne par un couple (g, 0).
Réduction de Hermite (manipulations de colonnes, échanges de
lignes)
Soit une matrice F ∈ Mm,n (B). Il existe un entier k ∈ J0.. inf(m, n)K, une
matrice C ∈ GLn (B) et une matrice de permutation P ∈ GLm (B) telles
que l’on ait
0
T
PFC=
0
G
avec T ∈ Mk (B) triangulaire inférieure de déterminant 6= 0.
Forme réduite triangulaire (manipulations de colonnes et de lignes)
Il existe un entier k ∈ J0.. inf(m, n)K, une matrice C 0 ∈ GLn (B) et une
matrice L ∈ GLm (B) telles que l’on ait
L F C0 =
T0
0
0
0
avec T 0 ∈ Mk (B) triangulaire inférieure de déterminant 6= 0.
2. Pseudo-bases et pseudo-matrices
9
Réduction de Smith en dimension 6 1, (manipulations de lignes
et de colonnes)
1.5. Théorème. Un domaine de Bezout de dimension 6 1 est un anneau
de Smith.
Démonstration. D’après le paragraphe précédent, il suffit de savoir traiter
une matrice carrée (triangulaire) M de déterminant d =
6 0. Comme l’anneau quotient B/hdi est zéro-dimensionnel et arithmétique, toute matrice
sur B/hdi se ramène par manipulations élémentaires de lignes et de colonnes
à une matrice en forme de Smith (fait 1.3).
On peut donc ramener, modulo d, la matrice M à une forme diagonale de
Smith, disons Diag(a1 , . . . , an ). Notons dk = Dk (M ).
Puisque B est un domaine de Bezout, on a di = hdi i pour des éléments di
de B, et 1 = d0 | d1 | d2 | . . . | dn = d. On a donc des éléments c1 , . . ., cn
tels que dk = c1 · · · ck pour k ∈ J1..nK.
Quand on fait l’extension des scalaires B → B0 = B/hdi, on obtient
ha1 . . . ak i = Dk (M ) = dk , c’est-à-dire ha1 · · · ak , di = hc1 · · · ck i.
En fait, on a aussi hak , di = hck i. Pour le voir, on considère le B-module
conoyau Coker(M ) = K. On a dK = 0, donc
K = K/dK ' B0 /ha1 i ⊕ · · · ⊕ B0 /han i ' B/ha1 , di ⊕ · · · ⊕ B/han , di .
En notant a0i = pgcd(ai , d), comme hai i ⊆ hai−1 i, on a a0i−1 | a0i . Cela
montre 2 que dk = ha01 · · · a0k i. Comme dk = hc1 · · · ck i, puisque B est intègre
et les ci non nuls, on obtient de proche en proche ha0k i = hck i.
Les manipulations élémentaires modulo d se relèvent en des manipulations
élémentaires sur B, d’où une égalité LM C = M 0 , où M 0 = Diag(a1 , . . . , an ).
On voit que la k-ème ligne de M 0 est multiple de ck = pgcd(ak , d). On
peut donc écrire M 0 = Diag(c1 , . . . , cn )C1 avec C1 ∈ Mn (B). Et comme
det(M 0 ) = d, on obtient det(C1 ) = 1. D’où enfin LM C2 = Diag(c1 , . . . , cn )
avec C2 = CC1−1 .
2. Pseudo-bases et pseudo-matrices
Dans cette section, A est un anneau intègre et l’on s’intéresse
aux applications linéaires entre modules qui sont sommes
directes de modules projectif de rang 1. Nous notons K le
corps de fractions
2. Ici on utilise le fait que deux matrices de même format qui ont des conoyaux
isomorphes ont les mêmes idéaux déterminantiels : ce sont les idéaux de Fitting du
conoyau.
10
Calcul matriciel généralisé
Dans cette section on développe un hh calcul matriciel généralisé ii pour
les applications linéaires entre modules projectifs de type fini qui sont
isomorphes à des sommes directes d’idéaux inversibles.
Pseudo-bases
Soit E un A-module projectif de type fini isomorphe à une somme directe
finie d’idéaux inversibles, E = E1 ⊕ · · · ⊕ Er avec Ei ' ei ∈ Iff(A) (inversibles). On peut regarder E comme un sous-A-module du K-espace vectoriel
ES obtenu en inversant les éléments du monoïde S = A∗ . Comme (Ei )S
est un K-module projectif de rang 1, il est isomorphe à K et l’on obtient
E = e1 e1 ⊕ · · · ⊕ er er , où les ei ∈ ES forment une base du K-espace vectoriel ES .
Notons que si be1 = ah1 pour deux idéaux de type fini e1 et h1 on a l’égalité
E1 = e1 e1 = h1 ( ab e1 ) = h1 e01 (e1 et h1 sont deux A-modules isomorphes).
Ainsi la décomposition E = E1 ⊕ · · · ⊕ Er ne donne pas une écriture unique
sous forme ei ei pour les Ei .
On appelle pseudo-base de E, un r-uplet
(e1 , e1 ), . . . , (er , er ) tel que E = e1 e1 ⊕ · · · ⊕ er er ,
où les ei sont des éléments de ES et les ei des idéaux fractionnaires
Pinversibles de A. Un x ∈ E s’écrit alors de manière unique sous forme i xi ei
où les xi ∈ ei (et donc xi ei ∈ E).
2.1. Fait. Donner une pseudo-base E = ((e1 , e1 ), . . . , (er , er )) pour le
module E, revient exactement à donner un isomorphisme
P
ψE : e1 × · · · × er −→ E, (x1 , . . . , xr ) 7−→ i∈J1..rK xi ei ,
dans lequel les images inverses des ei (après l’extension de ψE à ES ) forment
la base canonique de Kr (e1 × · · · × er ⊆ Kr ).
Remarque. Dans la suite, on pourrait imaginer que tous les modules projectifs qui interviennent sont des sous-A-modules de K-espaces vectoriels Kn ,
où n est supérieur ou égal au rang r du module.
Dans un tel contexte les ei sont soumis à la seule contrainte d’être des
vecteurs linéairement indépendants du K-espace vectoriel Kn considéré,
et ES est simplement le sous-K-espace vectoriel de Kn dont une base est
donnée par les vecteurs ei . Les ei sont donnés par une matrice M ∈ Mn,r (K).
On pourrait dans ce cadre aussi bien dire que la pseudo-base E est donnée
par une matrice M ∈ Mn,r (K) et une liste de r idéaux fractionnaires inversibles de A.
Il nous semble cependant préférable de garder pour nos explications un
cadre plus général.
2. Pseudo-bases et pseudo-matrices
11
Pseudo-matrices
Matrice d’une application linéaire sur des pseudo-bases
Rappelons que toute cette section s’applique à un anneau intègre arbitraire A
pour des modules qui admettent des pseudo-bases, c’est-à-dire qui sont
isomorphes à des sommes directes finies d’idéaux inversibles.
Soit ϕ : E → H une application linéaire entre modules projectifs de rangs m
et n. Si E = (e1 , e1 ), . . . , (em , em ) et H = (h1 , h1 ), . . . , (hn , hn ) sont
des pseudo-bases de E et H, l’application ϕS : ES → HS admet une matrice
A ∈ Kn×m sur les K-bases (e1 , . . . , em ) et (h1 , . . . , hn ) de ES et HS .
Les coefficients aij de cette matrice ne sont pas arbitraires dans K .
P
En effet, vue l’égalité ϕS (ej ) = i∈J1..mK aij hi , l’inclusion ϕS (E) ⊆ H est
satisfaite si, et seulement si, pour tout j ∈ J1..mK, tout xj ∈ ej et tout
i ∈ J1..nK, on a aij xj ∈ hi .
Autrement dit, pour tous i, j, on doit avoir
aij ej ⊆ bi , c’est-à-dire encore
aij ∈ hi e−1
.
j
2.2. Définition et notation. (Avec les notations ci-dessus)
1. On appelle matrice de ϕ sur les pseudo-bases E et H la donnée
A = (h1 , . . . , hn ; e1 , . . . , em ; A) = (h; e; A), où A = (aij )ij ∈ Mn,m (K).
Cette donnée satisfait les inclusions aij ej ⊆ hi . On note A = ME,H (ϕ) .
2. On appelle pseudo-matrice toute hh matrice ii (h; e; A) de ce type, i.e.
satisfaisant les inclusions aij ej ⊆ hi .
3. Pour des listes e et h fixées, l’ensemble formé par ces pseudo-matrices
est noté Mh;e (A) .
4. Cet ensemble est muni d’une structure naturelle de A-module : pour
additionner A et B, on additionne A et B.
5. On définit également le produit de deux pseudo-matrices de formats
convenables :
Si B = (c; h; B) et A = (h; e; A), alors B · A := (c; e; B · A)
Dans le cas envisagé au point 1 ci-dessus, les modules Mh;e (A) et LA (E, H)
sont isomorphes.
Les pseudo-matrices ont donc pour hh indices de colonnes ii des couples (j, ej ),
pour hh indices de lignes ii des couples (i, hi ), et les coefficients aij satisfont
les inclusions aij ej ⊆ hi . On pourra visualiser les choses comme suit, par
exemple pour une pseudo-matrice A de format 3 × 4 :
12
Calcul matriciel généralisé
e1
h1 a11
A = h2 a21
h3 a31
e2
a12
a22
a32
e3
a13
a23
a33
e4
a14
a24 aij ej ⊆ hi .
a34
Notez que le module Mn,m (A) des matrices usuelles correspond au cas
particulier où les idéaux ei et hj sont tous égaux à h1i.
En l’absence de modules de référence E et H, si (e1 , . . . , em ) et (h1 , . . . , hn )
désignent les bases canoniques de Km et Kn , la pseudo-matrice A est vue
comme la matrice d’une application A-linéaire ϕ : E → F , où
E = e1 × · · · × em = e1 e1 + · · · + em em ⊆ Km et
F
=
h1 × · · · × hn
=
h1 h1 + · · · + hn hn
⊆
Kn .
Calcul matriciel généralisé, premiers résultats.
2.3. Théorème et définition. (Déterminant d’une application linéaire
sur des pseudo-bases) Avec les notations précédentes, on suppose que E
et H ont même rang n et l’on considère la matrice A = ME,H (ϕ) d’une
application A-linéaire ϕ : E → H. On a donc A ∈ Mh;e (A).
Q
Q
Notons h = i∈J1..nK hi et e = i∈J1..nK ei . Le déterminant de la matrice A,
qui est un élément de K, est noté detE,H (ϕ) et on l’appelle le déterminant de ϕ sur les pseudo-bases E et H. Ci-après ∆ := detE,H (ϕ) =
det(A).
1. On a ∆e ⊆ h, i.e. ∆ ∈ he−1 . En particulier si e = h on a ∆ ∈ A.
2. Les propriétés suivantes sont équivalentes.
a. L’application linéaire ϕ est un isomorphisme.
b. On a l’égalité ∆e = h.
c. L’application linéaire ϕ est surjective.
Dans ce cas, la matrice A est une pseudo-matrice inversible.
Son inverse A−1 = (e; h; A−1 ) est un élément de Me;h (A) : c’est la
matrice de l’isomorphisme ϕ−1 sur les pseudo-bases H et E.
3. L’application linéaire ϕ est injective si, et seulement si, ∆ 6= 0.
4. Lorsque E = H et E = H, ∆ est un élément de A, il ne dépend que
de ϕ, et c’est le déterminant de ϕ au sens hh usuel ii, noté det(ϕ).
P
Démonstration. 1. Puisque aij ej ⊆ hi et ∆ = σ∈Pn sign(σ)(a1j1 · · · anjn ),
on obtient ∆e ⊆ h.
2a ⇔ 2c. Évidemment si ϕ est un isomorphisme, alors ϕ est un surjective.
Si ϕ est surjective, alors ϕ est scindable ([Modules, théorème XIII-2.1]),
2. Pseudo-bases et pseudo-matrices
13
donc H ' E ⊕ Ker ϕ, et puisque les modules E et H ont le même rang, ϕ
est un isomorphisme.
2a ⇒ 2b. Si ψ est l’isomorphisme réciproque, il est représenté par une
matrice B ∈ Me;h (A) sur les pseudo-bases H et E, et l’on a AB = (h; h; In ).
1
1
Si on applique le point 1 avec det(B) = ∆
, on obtient ∆
h ⊆ e et donc
h ⊆ ∆e.
e définit
2b ⇒ 2a. Si ∆e = h, nous allons montrer que la matrice A−1 = 1 A
∆
une pseudo-matrice de Me;h (A). Ainsi cette pseudo-matrice définit un application linéaire ψ : H → E sur les pseudo-bases H et E, et comme ϕ ◦ ψ
et ψ ◦ ϕ sont représentées par la matrice identité (sur les pseudo-bases
convenables), ψ et ϕ sont deux isomorphismes réciproques.
Il faut montrer
e
aij
1
∆
e
aij hj ⊆ ei . On note que de manière générale
Y
r6=i
er ⊆
Y
s6=j
hs , ou encore e
aij hj e ⊆ ei h .
(2)
D’où le résultat puisque ∆e = h.
3. Les modules E et H sont sans torsion et peuvent être vus comme des sousA-modules de Kn . Donc ϕ est injective si, et seulement si, ϕS : Kn → Kn
est injective. Et ϕS est injective si, et seulement si, ∆ 6= 0.
4. Le déterminant d’un endomorphisme se comporte bien par extension
des scalaires. En particulier lorsque l’on passe du A-module E au K-espace vectoriel ES le déterminant ne change pas. Or la matrice de ϕ sur la
pseudo-base ((e1 , e1 ), . . . , (en , en )) n’est autre que la matrice de ϕS sur la
base (e1 , . . . , en ).
Commentaire. On notera que les deux choses suivantes sont totalement
indépendantes des hh vecteurs ii ej et hi qui interviennent dans les pseudobases E et H.
– Le fait que la matrice est bien celle d’une application linéaire de E dans H :
ceci est contrôlé par les inclusions aij ej ⊆ hi .
– Le fait que l’application linéaire est un isomorphisme : ceci est contrôlé
par l’égalité ∆e = h.
Matrice de changement de pseudo-bases
Considérons deux pseudo-bases
différentes, E = (e1 , e1 ), . . . , (er , er )
et E 0 = (e01 , e01 ), . . . , (e0r , e0r ) , pour un même module E, avec les ei et
les e0i dans le K-espace vectoriel ES .
La pseudo-matrice A = ME 0 ,E (IdE ) s’appelle la matrice de passage de E
à E 0 : la matrice A ∈ Mr (K) a pour colonnes les e0i exprimés sur les ei . C’est
la matrice de passage de la base (e1 , . . . , er ) à la base (e01 , . . . , e0r ) dans ES .
14
Calcul matriciel généralisé
L’important, pour que cela fonctionne au niveau du A-module E, est que
les coefficients pij de la matrice de passage P vérifient pij e0j ∈ ei et que le
déterminant ∆ vérifie ∆e0 = e.
En fait, chaque fois que l’on a une pseudo-matrice inversible P pour deux
familles (e1 , . . . , en ) et (e01 , . . . , e0n ), si l’on considère un A-module
E qui
admet une pseudo-base du type E = (e1 , e1 ), . . . , (en , en ) , la matrice P
peut être vue comme une matrice de changement de pseudo-base pour E. Il
suffit pour cela de considérer les vecteurs e0j qui sont donnés par les colonnes
P
de la matrice P (i.e., e0j = i pij ei ).
L
En effet, considérons le module E 0 = j e0j e0j , et E 0 = (e01 , e01 ), . . . , (e0n , e0n ) ,
la matrice P est celle d’une application linéaire ϕ : E 0 → E sur les pseudobases E 0 et E. Mais, vu que les e0i sont donnés par les colonnes de P ,
on a ϕS = IdES . Ceci montre que ϕ est une application linéaire d’inclusion
de E 0 dans E. Enfin, puisque cette inclusion est un isomorphisme, elle est
surjective et l’on obtient E = E 0 et ϕ = IdE .
Idéaux déterminantiels
On a défini le déterminant d’une application linéaire sur des pseudo-bases.
On généralise les choses comme suit, en application du point 1 du théorème 2.3.
2.4. Définition. (Idéaux déterminantiels d’une pseudo-matrice)
1. Pour une pseudo-matrice carrée A = (h; e; A) on définit son idéal
déterminant, noté det(A) : c’est l’idéal
Y
Y
det(A) = det(A) e h−1 , où e =
ej et h =
hi .
j
i
2. Soient β = [β1 , . . . , βr ] ⊆ J1..nK et α = [α1 , . . . , αr ] ⊆ J1..mK des
listes extraites en ordre croissant. Nous notons Aβ,α la pseudo-matrice
extraite sur les lignes (d’indices dans) β et les colonnes α. Précisément
Aβ,α = (hβ1 , . . . , hβr ; eα1 , . . . , eαr , Aβ,α ).
On dit que l’idéal
Qr
Qr
mβ,α (A) := det(Aβ,α ) = det(Aβ,α )( i=0 eαi )( j=0 hβj )−1
est l’idéal mineur d’ordre r de la matrice A extrait sur les
lignes β et les colonnes α.
3. Pour une pseudo-matrice arbitraire, et r 6 inf(m, n) on définit l’idéal
déterminantiel d’ordre r de A, noté Dr (A), comme la somme des
idéaux mineurs d’ordre r de A.
Pour r > inf(m, n) on définit Dr (A) = 0.
Pour r 6 0 on définit Dr (A) = h1i.
2. Pseudo-bases et pseudo-matrices
15
2.5. Fait. Si A et B sont deux pseudo-matrices carrées telles que le
produit AB est défini, on a det(AB) = det(A) det(B).
Une pseudo-matrice carrée est inversible si, et seulement si, son idéal déterminant est égal à h1i.
Démonstration. On laisse le calcul du premier point au lecteur.
Le deuxième point est donné dans le théorème 2.3.
2.6. Théorème et définition. (Idéaux déterminantiels d’une application
linéaire) On continue avec la notation A = (h1 , . . . , hn ; e1 , . . . , em ; A).
1. Lorsque m = n, l’idéal déterminant de la matrice A = ME,H (ϕ) ne
dépend pas des pseudo-bases choisies pour E et H. On l’appelle l’idéal
déterminant de l’application linéaire ϕ, noté det(ϕ). On a les
équivalences
– det(ϕ) = h1i si, et seulement si, ϕ est un isomorphisme,
– det(ϕ) 6= h0i si, et seulement si, ϕ est injective.
2. Les idéaux déterminantiels de la matrice A ne dépendent pas des pseudobases choisies pour E et H. On les appelle les idéaux déterminantiels
de l’application linéaire ϕ, on les note Dr (ϕ).
3. L’application linéaire est surjective si, et seulement si, Dn (ϕ) = h1i
si, et seulement si, elle admet un inverse à droite. Elle est injective
si, et seulement si, Dm (ϕ) 6= 0. Elle admet un inverse à gauche si, et
seulement si, Dm (ϕ) = h1i.
4. Soit s ∈ A∗ tel que les modules E[1/s] et H[1/s] sont libres sur A[1/s].
Notons ϕs : E[1/s] → H[1/s] l’extension de ϕ par A → A[1/s]. Alors
pour chaque r on a : Dr (ϕ)A[1/s] = Dr (ϕs ) (idéaux déterminantiels
usuels définis pour une application linéaire entre modules libres).
5. On a les inclusions
{0} = D1+min(m,n) (ϕ) ⊆ · · · ⊆ D1 (ϕ) ⊆ D0 (ϕ) = h1i = A
(3)
Plus précisément pour tous k, r ∈ N on a une inclusion
Dk+r (ϕ) ⊆ Dk (ϕ) Dr (ϕ)
(4)
Démonstration. Le point 1 résulte du fait 2.5 et du théorème 2.3. Les autres
points (comme d’ailleurs le point 1 ) sont conséquences du point 4 (qui est
facile) et des résultats analogues concernant les applications linéaires entre
modules libres. En effet, un idéal est caractérisé par ses localisations en
des éléments comaximaux 3 . Et les propriétés d’injectivité et de surjectivité
pour des applications linéaires également.
3. Puisque l’anneau est intègre, l’idéal est même l’intersection de ses localisés, vus
comme sous-A-modules de K.
16
Calcul matriciel généralisé
Un cas important est celui des idéaux mineurs d’ordre 1,
mij (A) = aij ej h−1
,
i
qui interviendront dans les manipulations de matrices conduisant à des
formes réduites agréables.
Pseudo-matrices par blocs
Les calculs par blocs se font pour les pseudo-matrices aussi bien que pour
les matrices usuelles.
Considérons une pseudo-matrice M = (h1 , . . . , hn ; e1 , . . . , em ; M ) et des
entiers p ∈ J2..n − 1K, q ∈ J2..m − 1K. Décomposons M en quatre blocs
A ∈ Kp×q , B ∈ Kp×(m−q) , C ∈ K(n−p)×q et D ∈ K(n−p)×(m−q) ,
M=
A
B
C
D
ce qui donne quatre pseudo-matrices A = (h1 , . . . , hp ; e1 , . . . , eq ; A), etc.
Alors on note
M=
A
B
C
D
.
Pivot de Gauss pour les pseudo-matrices
2.7. Définition et notation.
– Une pseudo-matrice est dite strictement carrée lorsqu’elle est de la
forme A = (e; e; A). On note l’algèbre des pseudo-matrices correspondante par
Me1 ,...,en (A) = Me (A) .
– On appelle pseudo-matrice élémentaire une pseudo-matrice strictement carrée A telle que A est une matrice élémentaire usuelle (sur K).
Si aij ∈ K est le coefficient non nul en dehors de la diagonale, on doit
donc avoir aij ej ⊆ ei .
– On appelle manipulation élémentaire de colonnes (resp. de
lignes) sur une pseudo-matrice la postmultiplication (resp. la prémultiplication) par une pseudo-matrice élémentaire.
Exemples.
a b
Manipulation de colonnes. Considérons une pseudo-matrice L = c a b .
2. Pseudo-bases et pseudo-matrices
17
Si m11 (L) ⊇ m12 (L), c’est-à-dire si aa ⊇ bb, on a une pseudo-matrice éléa b
a 1 − ab
mentaire P =
, de sorte que l’on obtient, comme pour le pivot
b 0 1
de Gauss classique
a b
LP = c a 0
Ainsi, dans le processus hh pivot de Gauss ii ce qui tient lieu, pour une
pseudo-matrice, de hh a11 divise a12 ii dans le cas des matrices usuelles, c’est
ii
maintenant hh m11 ⊇ m12 ii, ou encore hh m12 m−1
11 ⊆ A , ce qui se dit aussi
hh m
i
i
11 divise m12 .
a
c1 a
.
Manipulation de lignes. Considérons une pseudo-matrice L =
c2 b
Si m11 (L) ⊇ m21 (L), c’est-à-dire si ac2 ⊇ bc1 , on a une pseudo-matrice
c1 c2
a
c1
1
0
c
a
élémentaire P =
, de sorte que l’on obtient P L = 1
.
c2 − ab 1
c2 0
Ainsi, dans le processus hh pivot de Gauss ii ce qui tient lieu, pour une
pseudo-matrice, de hh a11 divise a21 ii dans le cas des matrices usuelles, c’est
maintenant hh m11 ⊇ m21 ii, ce qui se dit aussi hh m11 divise m21 ii.
Voici une autre généralisation naturelle d’un résultat classique pour les
matrices usuelles.
2.8. Lemme du mineur inversible pour les pseudo-matrices.
(Pivot généralisé) On considère une pseudo-matrice
A = (h1 , . . . , hn ; e1 , . . . , em ; A).
On suppose que l’idéal mineur mJ1..kK,J1..kK (A) est égal à h1i.
1. (Pivot sur les colonnes) Il existe une pseudo-matrice inversible
C = (e1 , . . . , em ; h1 , . . . , hk , ek+1 , . . . , em , C)
C1
C2
décrite par blocs comme C =
0
Im−k
avec C1 = AJ1..kK,J1..kK −1
telle que la pseudo-matrice AC soit de la forme
2. (Pivots sur les lignes et les colonnes)
Ik
0
B
D
.
18
Calcul matriciel généralisé
La pseudo-matrice A est équivalente à une pseudo-matrice
A0 = (h1 , . . . , hn ; h1 , . . . , hk , ek+1 , . . . , em ; A0 ) =
Ik
0
0
A00
et l’on a Dr (A00 ) = Dk+r (A) pour tout r ∈ Z.
En particulier si Dk+1 (A) = h0i, on a une forme réduite
LAC =
Ik
0
0
0
.
NB : Si A possède un mineur d’ordre k > 0 égal à h1i, on obtient des
résultats analogues après permutation de lignes et/ou de colonnes.
Démonstration. On recopie la démonstration usuelle pour les matrices
usuelles.
A1 A2
1. Notons A =
où A1 ∈ Mk (K).
A3 A4
C1 C2
La matrice C est la matrice par blocs C =
où C1 = A1 −1
0 Im−k
et C2 = −C1 A2 . On obtient la matrice A C donnée dans l’énoncé.
2. On continue en multipliant à gauche par la matrice strictement carrée
définie par blocs L =
Ik
0
−B I`
∈ Mh1 ,...,hn (Z) avec ` = n − k. On obtient
pour L A C la pseudo-matrice donnée dans l’énoncé.
Enfin Dr (A00 ) = Dk+r (A0 ) = Dk+r (A) pour tout r ∈ Z.
Remarque. On a obtenu A0 = L A C avec C inversible et L strictement
carrée inversible. Si en outre on a ei = hi pour i ∈ J1..kK, la matrice C est
également strictement carrée.
3. Domaines de Prüfer, systèmes linéaires,
réduction de Hermite
Dans cette section, Z est un domaine de Prüfer à divisibilité
explicite, avec K pour corps de fractions.
3. Réduction de Hermite
19
Systèmes linéaires, généralités
On considère un système linéaire AX = B de n équations à m inconnues
sur Z. Dans la situation usuelle, tous les idéaux indexant les lignes et les
colonnes des matrices A, B (données) et X (inconnue) sont égaux à h1i,
les inconnues (coordonnées de X) sont dans Z et A ∈ Mn,m (Z). Plus généralement, on considère des pseudo-matrices A = (h1 , . . . , hn ; e1 , . . . , em ; A)
et B = (h; b; B), et la matrice inconnue X est dans Me;b (Z).
Comme on le voit sur la démonstration, le premier résultat vaut sur un
anneau intègre arbitraire.
3.1. Proposition. (Principe local-global) Soient s1 , . . ., s` comaximaux
dans Z. Le système linéaire AX = B admet une solution X ∈ Me;b (Z) si,
et seulement si, il admet une solution dans chaque Me;b (Zsi ).
Démonstration. La condition est évidemment nécessaire. Soit Xi =
1
Yi
sk
i
une solution dans Me;b (Zsi ) avec Yi ∈ Me;b (Z), de sorte que AYi = ski B. On
P`
P`
écrit i=1 ui ski = 1 avec les ui ∈ Z. Alors X = i=1 ui Yi est une solution
dans Me;b (Z).
Le point 1 du théorème suivant, bien que cas très particulier, est intéressant
car le calcul de la solution se fait dans K. Le point 2 reprend dans le langage
des pseudo-matrices le point 3 du théorème XII-3.2 dans [ACMC].
3.2. Théorème. (Systèmes linéaires sur un domaine de Prüfer, I)
1. Cas particulier : on donne une pseudo-matrice carrée avec det(A) 6= h0i
et l’on suppose que les idéaux mineurs d’ordre 1 du vecteur colonne B
sont tous contenus dans det(A). Alors la solution X calculée dans K
fournit une pseudomatrice X ∈ Me,b (Z), autrement dit xj b ∈ ej pour
j ∈ J1..nK.
2. Cas général : le système linéaire admet une solution si, et seulement
si, les idéaux déterminantiels de [ A | B ] sont égaux à ceux de A.
Démonstration. Le point 1 peut se démontrer par calcul direct en utilisant
la règle de Cramer. Il est clair aussi qu’il résulte du point 2.
2. Considérons des éléments comaximaux si de Z tels que, après localisation
en chacun des si , chacun des idéaux hj , ei et b est principal, c’est-à-dire libre
de rang 1. Ceci nous ramène au cas des matrices usuelles, pour lesquelles
l’équivalence est établie dans [ACMC, XII-3.2].
20
Calcul matriciel généralisé
Manipulations du type Bezout
Les manipulations de colonnes (resp. de lignes) de type Bezout que nous
examinons correspondent à des changements de pseudo-base pour le module
de départ (resp. d’arrivée).
On désire faire apparaître un coefficient nul de la même façon que ce que
l’on fait dans le cas des domaines de Bezout avec des matrices usuelles.
On commence par un cas particulier.
On traite la pseudo-matrice particulière suivante de format 1 × 2
a b
L= c a a .
Plutôt que de donner tout de suite la solution, expliquons d’où elle vient.
Cela nous prend une quinzaine de lignes, mais cela éclaire la chose. On a
une suite exacte scindable
t
[ −1 1 ]
[1 1]
0 → a ∩ b −−−−→ a ⊕ b −−−→ a + b → 0
ψ
qui donne un isomorphisme (a + b) ⊕ (a ∩ b) −→ a ⊕ b.
Plus précisément, regardons a ⊕ b comme
le sous-Z-module E de K2 muni
de la pseudo-base E = (e1 , a), (e2 , b) ((e1 , e2 ), base canonique de K2 ), i.e.
E = e1 a ⊕ e2 b = E1 ⊕ E2 .
Il nous faut maintenant trouver une pseudo-base H = (f1 , a+b), (f2 , a∩b)
de E, c’est-à-dire une matrice B ∈ M2 (K) convenable, dont les colonnes
exprimeront f1 et f2 sur la base (e1 , e2 ).
Pour cela, on utilise l’égalité (a : b)Z + (b : a)Z = h1i. On a donc
s ∈ (b : a)Z et t ∈ (a : b)Z avec s + t = 1,
d’où s(a + b) ⊆ b, t(a + b) ⊆ a et sa + tb = a ∩ b. D’où la matrice de passage
de E à H avec f1 = te1 + se2 et f2 = −e1 + e2 .
B = MH,E (IdM ) =
a
b
a+b
t
s
a∩b
−1
,
1
avec pour inverse
a+b
ME,H (IdM ) =
a∩b
a b
1 1
.
−s t
Notons que l’égalité a b = (a + b)(a ∩ b) nous assurait que la première
matrice était celle d’un isomorphisme, donc il n’y a aucune surprise à voir
que la matrice inverse est bien celle d’une application Z-linéaire.
En résumé B est une pseudo-matrice inversible et l’on obtient le résultat
3. Réduction de Hermite
21
souhaité :
LB = c
a+b
a
a∩b
0
Le cas général.
On veut maintenant traiter la pseudo-matrice générale de format 1 × 2
a b
L= c a b .
On se demande d’abord si le pivot de Gauss fonctionne (c’est-à-dire est-ce
que aa ⊇ bb ou aa ⊆ bb ?), auquel cas on fait appel au pivot de Gauss.
Si ce n’est pas le cas, il suffit de se ramener au cas particulier précédent,
en considérant les idéaux a0 = b−1 a et b0 = a−1 b. 4 On a alors une pseudomatrice inversible
a0 b0
a0 b0
a b 0
P =
avec LP = c ab ab .
b 0 a
En fin de compte ce sont donc les idéaux a0 et b0 qui seront traités selon la
méthode donnée dans le cas particulier étudié en premier. En particulier,
on prendra s et t ∈ A tels que s + t = 1, sa0 ⊆ b0 et tb0 ⊆ a0 (i.e. saa ⊆ bb
et tbb ⊆ aa).
Réduction de Hermite : images et noyaux d’applications
linéaires
Dans cette sous-section, on montre comment la réduction de Hermite des
pseudo-matrices permet de démontrer sous forme très concrète les théorèmes
de structure généraux concernant aussi bien les systèmes linéaires que les
modules de présentation finie sur les domaines de Prüfer.
3.3. Théorème. (Réduction de Hermite)
Soit Z un domaine de Prüfer de corps de fractions K pour lequel on a un
test de divisibilité et un algorithme d’inversion des idéaux de type fini.
Soit A = ((hi )i∈J1..nK ), (ej )j∈J1..mK ), A) avec A = (aij ) ∈ Mn,m (K) une
pseudo-matrice sur Z (i.e., aij ej ⊆ hi pour tous i, j).
On peut calculer une pseudo-matrice inversible
C = ((ej )j∈J1..mK , (e0j )j∈J1..mK , C) avec C = (cj,k ) ∈ Mm (K)
(on regarde C comme opérant un changement de pseudo-base au départ), et
une matrice de permutation L ∈ Mn (Z) (la pseudo-matrice L correspondante
4. Dans le cas où l’on mène tous les calculs avec des idéaux entiers, on considère
un élément x ∈ K tels que les idéaux fractionnaires a0 = xb−1 a et b0 = xa−1 b soient
entiers, i.e. x ∈ ba−1 ∩ ab−1 . Simplement, on peut prendre pour x un multiple commun
des numérateurs de a et b, ce qui donne xb−1 et xa−1 ∈ A, et a fortiori a0 et b0 entiers.
22
Calcul matriciel généralisé
opère un changement de pseudo-base à l’arrivée), telles que la matrice L A C
T
soit en forme réduite de Hermite : H =
G
inférieure de déterminant non nul.
0
, T carrée triangulaire
0
Démonstration. Cela résulte des manipulations de colonnes de type pivot
de Gauss et de type Bezout décrites précédemment. Tout se passe comme
pour une matrice sur un anneau de Bezout intègre, à ceci près que l’on doit
remplacer les matrices usuelles par des pseudo-matrices.
Notons que le test de divisibilité est utilisé uniquement pour vérifier que
dans certains cas, des manipulations élémentaires du type pivot de Gauss
sont admissibles.
On pourrait donc se passer de cette hypothèse en faisant fonctionner l’algorithme de réduction de Hermite uniquement avec des manipulations de
type Bezout.
Cependant, comme déjà indiqué, nous faisons toujours l’hypothèse que
l’anneau Z est un domaine de Prüfer à divisibilité explicite (au sens constructif).
Les théorèmes qui suivent sont tous des corollaires faciles du théorème 3.3
concernant les pseudo-matrices.
3.4. Théorème. (Image d’une pseudo-matrice) L’image de toute pseudomatrice est isomorphe à une somme directe d’idéaux inversibles.
Démonstration. On effectue une réduction de Hermite de la matrice. La
multiplication à droite par une pseudo-matrice inversible ne change pas le
module image. Il suffit de traiter l’image d’une pseudo-matrice (h; e0 ; H)
où H =
T
0
est en forme réduite de Hermite. C’est aussi l’image
0
G
T
de la pseudo-matrice (h; e01 , . . . , e0r ; H1 ) où H1 =
. Comme det(T ) 6= 0,
G
l’application linéaire
e01 × · · · × e0r −→ h1 × · · · × hn
définie par H1 est injective. Donc Im A = Im H = Im H1 ' e01 × · · · × e0r .
Précisément, supposons que l’on traite la pseudo-matrice A d’une application
Z-linéaire ϕ : E → H sur des pseudo-bases E = (e1 , e1 ), . . . , (em , em )
et H = (h1 , h1 ), . . . , (hn , hn ) de E et H. Alors l’image de ϕ est le module
somme directe e01 h01 ⊕ · · · ⊕ e0r h0r , où les h0i sont les vecteurs exprimés par les
colonnes de la matrice H1 sur les hi (qui sont des éléments de K ⊗Z H).
3.5. Théorème. (Images)
1. Tout Z-module projectif de type fini est isomorphe à une somme directe
d’idéaux inversibles, autrement dit, il possède une pseudo-base.
3. Réduction de Hermite
23
2. L’image d’une application linéaire entre Z-modules projectifs de type fini est un module projectif de type fini dont on peut calculer une
pseudo-base.
3. – Tout sous-Z-module de type fini de Kn est projectif de type fini.
– Si g1 , . . ., gm ∈ Kn et si e1 , . . ., em sont des idéaux de type
Pfini non
nuls, on peut calculer une pseudo-base du Z-module E = j gj ej .
Démonstration. 1. Puisqu’un module projectif de type fini est isomorphe à
l’image d’une matrice de projection usuelle, cela résulte du théorème 3.4.
2. Il s’agit pour l’essentiel d’une reformulation du théorème 3.4. En effet, vu
le point 1, tout module projectif de type fini possède une pseudo-base. Alors
une application linéaire entre modules projectifs de type fini est représentée
par une pseudo-matrice sur les pseudo-bases des modules.
3. Les deux points sont clairement équivalents. Le second résulte du théorème 3.4 car E est l’image d’une pseudo-matrice (hdi , . . . , hdi ; e1 , . . . , em ; G),
où G a pour colonnes les gj exprimés sur la base canonique de Kn , et où
l’élément d ∈ K est choisi de façon que gij ej ⊆ dZ pour tous les (i, j).
3.6. Théorème. (Noyaux)
Soit ϕ : E → H une application linéaire entre modules projectifs de type
fini sur un domaine de Prüfer. Le noyau Ker ϕ est facteur direct dans E (a
fortiori projectif de type fini). Plus précisément, on peut calculer une pseudobase de E dont les derniers termes forment une pseudo-base de Ker ϕ.
Démonstration. Comme dans le théorème précédent, il suffit de faire la
démonstration pour le noyau d’une pseudo-matrice, et l’on prend les notations du théorème 3.3. Le changement de pseudo-base donné
par la matrice C
0
0
fournit une nouvelle pseudo-base (u1 , e1 ), . . . , (um , em ) (les uj sont exprimés par les colonnes de la matrice C), pour laquelle l’application linéaire
est exprimée par la nouvelle pseudo-matrice H. Pour cette matrice il est
clair que le noyau est fourni par la somme directe des sous-modules uj e0j
pour les numéros j des colonnes nulles de H.
Pseudo-matrices surjectives
La proposition suivante reprend le point 3 du théorème 2.6, qui généralise
un théorème bien connu pour les matrices usuelles.
Dans le point 3 du théorème 2.6, l’application linéaire ϕ, ici représentée
par A, admet un inverse à droite ψ du fait qu’elle est surjective et que l’image
est un module projectif. Le module source de ϕ est égal à Im ψ ⊕ Ker ϕ.
Lorsque Ker ϕ est une somme directe de modules projectifs de rang 1,
la matrice A peut être complétée en la pseudo-matrice inversible B qui
représente l’application linéaire ϕ|Im ψ ⊕ IdKer ϕ , laquelle a pour source
Im ψ ⊕ Ker ϕ et pour image Im ϕ ⊕ Ker ϕ.
24
Calcul matriciel généralisé
Une fois que l’existence de la pseudo-matrice B est assurée par un argument
général, il reste le problème algorithmique d’un calcul aussi simple que
possible de cette pseudo-matrice.
Nous sommes intéressés ici par une démonstration matricielle algorithmique
du résultat dans le contexte des domaines de Prüfer, pour lesquels les calculs
sur les pseudo-matrices sont simplifiés.
3.7. Proposition. (Pseudo-matrices surjectives)
Soit Z un domaine de Prüfer et soit n ∈ J1..mK.
Soit A = (h1 , . . . , hn ; e1 , . . . , em ; A) ∈ Mh,e (Z). Cette pseudo-matrice représente une application linéaire surjective si, et seulement si, Dn (A) = h1i.
Dans ce cas, A peut être complétée en une pseudo-matrice inversible
B = (h1 , . . . , hm ; e1 , . . . , em ; B).
La matrice AB
−1
est alors en forme de Smith :
AB −1 =
h1
h1
1
h2
0
h2
..
.
0
..
.
1
..
hn
0
.
···
···
···
..
.
..
0
.
hn
0
..
.
0
1
hn+1
0
..
.
..
.
0
···
···
···
hm
0
..
.
.. .
.
0
Et la pseudo-matrice extraite sur les n premières colonnes de B −1 est un
inverse à droite de A.
Démonstration. On démontre l’implication difficile.
D’après le théorème 3.3, on peut calculer une pseudo-matrice inversible
C = ((ej )j∈J1..mK ; e01 , . . . , e0n , hn+1 , . . . , hm ; C) avec C = (cj,k ) ∈ Mm (K)
telle que la matrice A C soit en forme réduite de Hermite : H = T
0
,
où T est carrée triangulaire inférieure. On a Dn (T ) = Dn (A) = h1i, donc T
est inversible. En fait les idéaux mineurs d’ordre 1 diagonaux mii (T ) ont
leur produit égal à h1i, donc sont tous égaux à h1i, ce qui autorise la
méthode du pivot par manipulations élémentaires de colonnes. On termine
en inversant la matrice diagonale obtenue. On a ainsi ramené T à la forme
(h1 , . . . , hn ; h1 , . . . , hn ; In ). Ceci donne une pseudo-matrice inversible C 0 telle
que A C C 0 = In
et l’on pose B := (C C 0 )−1 .
0
B1
En regardant B comme une pseudo-matrice par blocs B =
B1
obtient In
= A et B1 = A.
0
B2
, on
B2
3. Réduction de Hermite
25
Double réduction de Hermite, conoyaux
3.8. Théorème. (Double réduction de Hermite)
Mêmes hypothèses qu’au théorème 3.3. On peut calculer une pseudo-matrice
inversible C opérant un changement de pseudo-base au départ, et une pseudomatrice inversible L opérant un changement de pseudo-base à l’arrivée, telles
que L A C =
non nul.
T
0
0
0
, où T est triangulaire inférieure de déterminant
Démonstration. On fait une réduction de Hermite sur les colonnes, suivie
d’une réduction de Hermite sur les lignes.
Le théorème suivant est une forme précise, hh pseudo-matricielle ii, du théorème XIV-2.5 de [Modules].
3.9. Théorème. (Conoyaux)
Soit ϕ : E → H une application linéaire entre modules projectifs de type
fini sur un domaine de Prüfer.
1. Le saturé de Im ϕ dans H est en facteur direct dans H. Plus précisément, on peut calculer une pseudo-base de H dont les premiers termes
forment une pseudo-base du saturé de Im ϕ.
2. Le module conoyau Coker ϕ est un module de présentation finie, somme
directe de son sous-module de torsion et d’un module projectif de type
fini.
3. Le sous-module de torsion de Coker ϕ est le quotient d’un module projectif par un sous-module projectif de même rang. Il est isomorphe au
conoyau d’une pseudo-matrice carrée injective.
En particulier tout module de présentation finie sans torsion est projectif.
Démonstration. Le point 1 résulte du théorème 3.8 et du fait que lorsque
l’on sature l’image d’une pseudo-matrice carrée injective, on obtient le
module d’arrivée en entier. Le reste suit facilement.
Remarque. Dans le théorème 3.9, la structure du sous-module de torsion et
celle du sous-module projectif de type fini ne sont pas décryptées de manière
complètement satisfaisante. En particulier, ce théorème, bien que constructif, ne donne pas de solution algorithmique générale aux deux problèmes
suivants :
– déterminer si deux modules de présentation finie de torsion sont ou ne
sont pas isomorphes ;
– déterminer si deux modules projectifs de type fini sont ou ne sont pas
isomorphes.
26
Calcul matriciel généralisé
On verra dans la section 4 que lorsque le domaine de Prüfer Z est de
dimension 1, on sait répondre à la première question, mais pas nécessairement à la seconde, qui se ramènera à la question de savoir si un idéal de
type fini e est principal. Ce problème n’admet pas de solution algorithmique
générale, même pour les domaines de Dedekind.
Systèmes linéaires, via la réduction de Hermite
Le théorème 3.3 donne le corollaire suivant pour le traitement des systèmes
linéaires.
On notera que les tests proposés ici sont explicites parce que le divisibilité
dans Z est explicite.
3.10. Théorème. (Systèmes linéaires sur un domaine de Prüfer, II)
On considère un système linéaire AX = B de n équations à m inconnues
sur Z. Dans la situation usuelle, tous les idéaux indexant les lignes et les
colonnes des matrices A, B (données) et X (inconnue) sont égaux à h1i,
les inconnues (coordonnées de X) sont dans Z et A ∈ Mn,m (Z).
Les résultats s’appliquent aussi bien avec des pseudo-matrices A = (h; e; A)
et B = (h; b; B), en adaptant le format de la colonne X aux données : les
inconnues sont des éléments xj ∈ K soumis aux conditions xj b ∈ ej .
On considère une double réduction de Hermite L A C =
T
0
, avec la
0
0
matrice T ∈ Mr (K) triangulaire inférieure injective. Les pseudo-matrices
C = (e, e0 , C) et L = (h0 , h, L) sont inversibles, et l’on introduit les nouvelles
inconnues Y = C −1 X.
1. La solution générale du système homogène sans second membre AX = 0
est donnée par y1 = · · · = yr = 0. Ceci donne pour Y le sous-Z-module
K = {0} × · · · × {0} × e0r+1 × · · · × e0m
et pour X, le Z-module C(K) paramétré par
(yr+1 , . . . , ym ) ∈ e0r+1 × · · · × e0m ,
c’est aussi l’image de la pseudo-matrice extraite de C sur les m − r
dernières colonnes.
2. Le système linéaire avec second membre AX = B admet une solution
si, et seulement si, le sous-Z-module de Kn codé par LB est contenu
dans l’image de la pseudo-matrice
T
h01 , . . . , h0n ; e01 , . . . , e0r ;
.
0
– Les contraintes données par l’annulation des n − r dernières coordonnées de L B signifient exactement que le système A X = B a une
solution dans Km .
4. Réduction de Smith en dimension 1
27
– Les contraintes sur les r premières coordonnées de L B peuvent être
testées de proche en proche, vue la forme triangulaire de T .
– On peut donc tester l’existence d’une solution et calculer une solution
particulière lorsqu’il en existe une.
Démonstration. Il faut juste préciser la question du test pour une solution
particulière : toute solution donne par soustraction une solution avec les
hh inconnues auxiliaires ii nulles, donc le test peut porter uniquement sur
l’équation T Z = B 0 avec Z et B 0 donnés par troncatures de Y et LB aux r
premières coordonnées. Comme T est injective, la solution, si elle existe, est
unique.
Commentaire. L’énoncé du corollaire sur les systèmes linéaires est un peu
plus simple dans le cas d’un système linéaire usuel, mais il n’est pas
hh nettement plus simple ii, car les idéaux e0 et h0 introduits par la double
j
i
réduction de Hermite, ainsi que les coefficients de T dans K et non dans Z,
semblent des ingrédients inévitables.
4. Calculs modulaires et réduction de Smith
en dimension 1
Dans cette section, Z est un domaine de Prüfer à divisibilité
explicite, de dimension 6 1.
4.1. Définition. Soit A = (h1 , . . . , hn ; e1 , . . . , em ; A) ∈ Mh,e (Z) et p =
inf(m, n). La pseudo-matrice A est dite en forme de Smith si la matrice
A = (aij ) a tous ses coefficients nuls hormis éventuellement des coefficients
aii et si les idéaux mineurs d’ordre 1 ci := mii (A) = aii ei h−1
satisfont les
i
relations attendues :
c1 ⊇ c2 ⊇ · · · ⊇ cp .
4.2. Théorème. Soit Z un domaine de Dedekind explicite, i.e., un domaine de Prüfer noethérien à divisibilité explicite. Toute pseudo-matrice
sur Z est équivalente à une pseudo-matrice en forme de Smith.
Démonstration. L’algorithme est le même que pour la réduction de Smith
d’une matrice sur un anneau principal, en remplaçant les matrices usuelles
par des pseudo-matrices. On peut se reporter à [Modules, Algorithme IV2.2].
28
Calcul matriciel généralisé
Sur l’unicité d’une réduite de Smith usuelle
Lorsqu’une matrice M ∈ Mn,m (A) admet une réduction de Smith dont la
partie carrée est égale à Diag(a1 , . . . , an ), le module Coker M est isomorphe
à la somme directe des A/hai i :
Coker M ' A/ha1 i ⊕ · · · ⊕ A/han i avec des idéaux ha1 i ⊇ · · · ⊇ han i .
Les idéaux ei = hai i sont uniquement déterminés (fait 1.1).
Plus précisément, si aucun des ai n’est inversible, la liste (e1 , . . . , en ) ne
dépend que du module Coker M .
Cela ne signifie pas nécessairement que pour deux matrices équivalentes en
forme de Smith A = (aij ) et A0 = (a0ij ) les a0ii soient associés aux aii .
Néanmoins, c’est vrai lorsque A est intègre, et le lemme suivant montre que
c’est aussi le cas pour un anneau zéro-dimensionnel.
4.3. Lemme. Soient a et b ∈ A zéro-dimensionnel. Si hai = hbi, alors a
et b sont associés : a = wb pour un élément w ∈ A× .
Démonstration. On a a = ub et b = va, donc a(uv − 1) = 0. Comme A est
zéro-dimensionnel on a un idempotent e tel que uv −1 est nilpotent modulo e
et inversible modulo 1 − e. Dans la composante A/hei, on a uv = 1 + x avec
x nilpotent, donc u est inversible. Dans la composante A/h1 − ei, a = 0
donc b = 0 et l’on a a = 1.b. En bref a = wb avec w = u.(1 − e) + 1.e
inversible.
On montre maintenant un lemme qui complète en quelque sorte le lemme
précédent.
4.4. Lemme. Soit Diag(a1 , . . . , an )Qune matrice en forme de Smith sur un
anneau A, et des ui ∈ A× tels que i ui = 1. Définissons a0i = ai ui , alors
les deux matrices Diag(a1 , . . . , an ) et Diag(a01 , . . . , a0n ) sont élémentairement
équivalentes.
0
Démonstration. Il suffit
le cas suivant,
de traiter
avec b = ca, a = au,
0
a 0
a 0
uv = 1, b0 = bv, A =
et B =
.
0 b
0 b0
a 0 1 a a0 − a 2 a0 a0 − a 3 a0 a0 − a 4 a0 0
−→
−→
−→
−→
0 b
0 b
b
b
0 b0
0 b0
1
2
3
4
:
:
:
:
C2
C1
L2
C2
←
←
←
←
En bref
C2 + (u − 1)C1 .
C1 + C2 .
L2 − cvL1 , avec b − cv(a0 − a) = b − ca + cva = cva = vb = b0 .
C2 + (v −
1)C1 .
1 0
1 u−1
1 0
1 v−1
A
= B.
−cv 1
0 1
1 1
0 1
4. Réduction de Smith en dimension 1
29
La structure des modules de présentation finie de torsion, et quelques calculs matriciels
Rappelons la démonstration de la proposition suivante [Modules, XVI-4.6].
4.5. Proposition. Un module de présentation finie de torsion P sur un
domaine de Prüfer Z de dimension 1 est isomorphe à une somme directe
Z/a1 ⊕ · · · ⊕ Z/an , avec des idéaux inversibles a1 ⊇ · · · ⊇ an .
Si les ai sont 6= h1i cette écriture est unique 5 . Plus généralement, deux
écritures de ce type de même longueur sont identiques (fait 1.1).
Démonstration. Le module P est annulé par un élément δ =
6 0, parce qu’il
est de torsion et de type fini (il suffit que δ annule chacun des générateurs).
L’anneau quotient Z0 = Z/hδi est zéro-dimensionnel et arithmétique, donc
est un anneau de Smith ([Modules, XVI-4.2]).
Ainsi, on a un isomorphisme de Z-modules (ce sont aussi des Z0 -modules)
du type suivant
P = P/δP ' Z0 /ha1 i ⊕ · · · ⊕ Z0 /han i ' Z/hδ, a1 i ⊕ · · · ⊕ Z/hδ, an i ,
avec les inclusions hδ, a1 i ⊇ · · · ⊇ hδ, an i .
En pratique, cela se passe précisément comme suit. Le module P est présenté
par une matrice M ∈ Mn,m (Z) avec dn := Dn (M ) = F0 (P ) 6= h0i.
Pour un δ 6= 0 dans dn , on calcule sur Z0 une réduite de Smith par manipulations élémentaires L M C =
D
0
avec D = Diag(a1 , . . . , an )
et ha1 i ⊇ · · · ⊇ han i. On en déduit la structure de P comme ci-dessus, de
sorte que Dk (M ) = Fn−k (P ) = ha1 · · · ak , δi = a1 · · · ak et ak = hak , δi pour
chaque k ∈ J0..nK.
On en déduit le corollaire suivant.
4.6. Corollaire. Sur un domaine de Prüfer de dimension 6 1, la structure
d’un module de présentation finie de torsion P est caractérisée par ses
idéaux de Fitting Fk (P ).
Démonstration. En effet, dans la proposition 4.5, notons di = Fn−i (P ). On
a alors d1 = a1 , et plus généralement dk = a1 · · · ak pour k ∈ J2..nK (fait
1.2) ; donc ak = (dk : dk−1 ) car ak 6= h0i.
Le corollaire suivant est plus surprenant.
5. Dans la mesure où nous supposons que Z est à divisibilité explicite, on pourra se
débarrasser des idéaux égaux à h1i.
30
Calcul matriciel généralisé
4.7. Corollaire. Soit M ∈ Mn,m (Z) avec Dn (M ) 6= 0. Soit δ 6= 0 dans
Dn (M ). On considère l’anneau arithmétique zéro-dimensionnel Z0 = Z/hδi.
On note x pour x mod δ. On calcule par manipulations élémentaires une
réduite de Smith de M , disons
M0 ≡ D
0
mod δ avec D = Diag(a1 , . . . , an ) et ha1 i ⊇ · · · ⊇ han i .
Alors :
1. ha1 · · · ai , δi = ha1 , δi · · · hai , δi pour i ∈ J2..nK.
2. Dn (M ) = ha1 , δi · · · han , δi.
Q
Q
Q
=
3.
i∈J ai , δ dans les cas suivants :
i∈I ai , δ
i∈I∪J ai , δ :
I = J1..kK et J = Jk + 1..`K ⊆ J1..nK, ou vice versa.
Q
Q
Q
4.
i∈I∪J ai :
i∈I ai =
i∈J ai dans les cas suivants :
I = J1..kK et J = Jk + 1..`K ⊆ J1..nK, ou vice versa.
Démonstration. 1 et 2. Cas particulier de ce qui a été expliqué dans la
démonstration de la proposition 4.5.
Le point 3 résulte des égalités di = a1 · · · ai , car Z est un domaine de Prüfer.
Le point 4 en résulte parce que lorsqu’on a trois idéaux I ⊆ J ⊆ K dans
un anneau A, on a l’égalité
(J : K)A mod I = (J/I : K/I)A/I .
Et maintenant, oh surprise !, on obtient la hh réduction de Smith avec pseudobases ii d’une matrice carrée de déterminant non nul sans effort.
4.8. Proposition. Soit M ∈ Mn (Z) une matrice avec d = det(M ) 6= 0 (Z
domaine de Prüfer de dimension 1). Cette matrice est équivalente à une
pseudo-matrice de la forme
a1 a2 · · · an
h1i 1
0 ··· 0
.
..
. ..
h1i
0
1
M1 = . .
.. .. . . . . . . 0
h1i 0 · · · 0
1
Q
avec a1 ⊇ a2 ⊇ · · · ⊇ an , i ai = hdi et chaque ai de la forme hai , di.
Démonstration. On reprend ce qui a été expliqué dans la démonstration
de la proposition 4.5 (éventuellement avec δ = d). En examinant ce qui se
passe dans le calcul de la réduction de Smith modulaire par manipulations
4. Réduction de Smith en dimension 1
31
élémentaires, on obtient sur Z une égalité et une congruence
LM C = M 0 ≡ Diag(a1 , . . . , an ) mod δ .
h1i
a1
Considérons la pseudo-matrice 6 E := ...
an
···
h1i
.
M0
Puisque M 0 = Diag(a1 , . . . , an ) mod δ, elle est de la forme
a1
h1i
a1 + δx11
h1i
δx12
a2
E = .
..
δx21
..
.
a2 + δx22
δxn1
···
an
···
···
..
h1i
δx1n
..
.
..
.
.
.
an + δxnn
Oh merveille ! c’est une pseudo-matrice de changement de pseudo-base. En
Qn
effet, i=1 ai = hdi avec d = det(M ) = det(LM C) = det(M 0 ) = det(E),
donc det(E) = h1i. Son inverse est la pseudo-matrice
a1
E −1
h1i
.
= ..
h1i
···
an
M0
−1
Posons C1 = CE −1 . On obtient LM C1 = M1 de la forme annoncée, où L
et C1 sont des pseudo-matrices de changements de base (L et M sont des
matrices usuelles vues comme des pseudo-matrices ayant les idéaux h1i en
indices de lignes et en colonnes).
Voici maintenant le résultat analogue pour une matrice hh usuelle ii plus large
que haute. Il nous faut faire ici appel à la proposition 3.7 concernant les
pseudo-matrices surjectives en espérant qu’elle soit efficace sur un domaine
de Prüfer de dimension 1.
4.9. Théorème. Soit M ∈ Mn,m (Z) une matrice avec Dn (M ) 6= 0. Soit
un élément δ non nul de Dn (M ). Cette matrice est équivalente à une
6. On vérifie immédiatement que c’est bien une pseudo-matrice.
32
Calcul matriciel généralisé
pseudo-matrice de la forme
h1i
a1
1
a2
0
h1i
0
M1 = .
.
.. ..
h1i 0
Qn
avec a1 ⊇ a2 ⊇ · · · ⊇ an , i=1 ai
Lm
et k=1 ak ' Zm .
1
..
.
···
···
···
..
.
..
.
0
an
0
..
.
···
···
0
1
···
···
am
0
..
.
..
.
0
= Dn (M ), chaque ai de la forme hai , δi
Démonstration. On reprend ce qui a été expliqué dans la démonstration
de la proposition 4.5. En examinant ce qui se passe dans le calcul de la
réduction de Smith modulaire par manipulations élémentaires, on obtient
sur Z une égalité et une congruence
LM C = M 0 ≡ D
mod δ
0
avec D = Diag(a1 , . . . , an ) et ha1 i ⊇ · · · ⊇ han i.
h1i
···
···
···
a1
7
Considérons la pseudo-matrice E := ...
an
Elle est de la forme
h1i
h1i
···
a1 a1 + δx11
δx12
···
a2 δx21
a2 + δx22
E = .
.
..
..
..
.
an
δxn1
···
h1i
.
M0
h1i
δx1n
···
···
an + δxnn
···
h1i
δx1m
..
.
.
..
.
δxnm
Oh merveille ! c’est une pseudo-matrice surjective. En effet,
Qn
0
i=1 ai = Dn (M ) avec Dn (M ) = Dn (LM C) = Dn (M ) = Dn (E),
donc Dn (E) = h1i. D’après la proposition 3.7, on peut la compléter en
une pseudo-matrice inversible B = (a1 , . . . , am ; h1i , . . . , h1i ; B) telle que la
pseudo-matrice M 0 B −1 soit la matrice M1 de l’énoncé. Avec C1 = CB −1 , on
obtient LM C1 = M1 , où L et C1 sont des pseudo-matrices de changements
de base (L et M sont des matrices usuelles vues comme des pseudo-matrices
ayant les idéaux h1i en indices de lignes et en colonnes).
7. On vérifie immédiatement que c’est bien une pseudo-matrice.
4. Réduction de Smith en dimension 1
33
Références
[ACMC] Lombardi H. & Quitté C. Algèbre Commutative. Méthodes
constructives. Calvage&Mounet (2011).
[CACM] English version of [ACMC]. Springer (2015).
[Cohen] Cohen H. Advanced topics in computational number theory. Graduate texts in mathematics 193. Springer-Verlag (1999).
[Modules] Díaz-Toca G.-M., Lombardi H. & Quitté C. Modules sur
les anneaux commutatifs. Calvage&Mounet (2014).
[MRR] Mines R., Richman F. & Ruitenburg W. A Course in Constructive Algebra. Universitext. Springer-Verlag, (1988).
Table des matières
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 Quelques rappels . . . . . . . . . . . . . . . . . . . . . . . . .
Quelques définitions . . . . . . . . . . . . . . . . . . . . .
Idéaux fractionnaires de type fini . . . . . . . . . . .
Calculs sur les idéaux de type fini dans les domaines de Prüfer
Matrice de localisation principale . . . . . . . . . . .
Opérations élémentaires sur les idéaux de type fini .
Formes réduites de matrices sur un domaine de Bezout .
Réduction de Hermite (manipulations de colonnes,
échanges de lignes) . . . . . . . . . . . . . . . . . .
Forme réduite triangulaire (manipulations de colonnes
et de lignes) . . . . . . . . . . . . . . . . . . . . . .
Réduction de Smith en dimension 6 1, (manipulations
de lignes et de colonnes) . . . . . . . . . . . . . . .
2 Pseudo-bases et pseudo-matrices . . . . . . . . . . . . . . . . .
Pseudo-bases . . . . . . . . . . . . . . . . . . . . . . . . .
Pseudo-matrices . . . . . . . . . . . . . . . . . . . . . . .
Matrice d’une application linéaire sur des pseudo-bases
Calcul matriciel généralisé, premiers résultats. . . . .
Matrice de changement de pseudo-bases . . . . . . . .
Idéaux déterminantiels . . . . . . . . . . . . . . . . . . .
Pseudo-matrices par blocs . . . . . . . . . . . . . . . . . .
Pivot de Gauss pour les pseudo-matrices . . . . . . . . . .
3 Réduction de Hermite . . . . . . . . . . . . . . . . . . . . . .
Systèmes linéaires, généralités . . . . . . . . . . . . . . . .
Manipulations du type Bezout . . . . . . . . . . . . . . .
Réduction de Hermite : images et noyaux d’applications
linéaires . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pseudo-matrices surjectives . . . . . . . . . . . . . . . . .
Double réduction de Hermite, conoyaux . . . . . . . . . .
Systèmes linéaires, via la réduction de Hermite . . . . . .
4 Réduction de Smith en dimension 1 . . . . . . . . . . . . . . .
– 35 –
1
2
2
4
4
4
6
7
8
8
9
9
10
11
11
12
13
14
16
16
18
19
20
21
23
25
26
27
36
Table des matières
Sur l’unicité d’une réduite de Smith usuelle . . . . . . . .
La structure des modules de présentation finie de torsion,
et quelques calculs matriciels . . . . . . . . . . . . . . .
28
29
| 0math.AC
|
arXiv:1802.00509v1 [cs.CV] 1 Feb 2018
Learning Semantic Segmentation with Diverse Supervision
Linwei Ye
University of Manitoba
Zhi Liu
Shanghai University
Yang Wang
University of Manitoba
[email protected]
[email protected]
[email protected]
Abstract
Models based on deep convolutional neural networks
(CNN) have significantly improved the performance of semantic segmentation. However, learning these models requires a large amount of training images with pixel-level
labels, which are very costly and time-consuming to collect. In this paper, we propose a method for learning CNNbased semantic segmentation models from images with several types of annotations that are available for various computer vision tasks, including image-level labels for classification, box-level labels for object detection and pixel-level
labels for semantic segmentation. The proposed method is
flexible and can be used together with any existing CNNbased semantic segmentation networks. Experimental evaluation on the challenging PASCAL VOC 2012 and SIFTflow benchmarks demonstrate that the proposed method can
effectively make use of diverse training data to improve the
performance of the learned models.
1. Introduction
Semantic segmentation is one of the key problems in
computer vision. The goal is to assign a label to each pixel
in the image to indicate the object class of this pixel. Like
many areas in visual recognition, the performance of semantic segmentation has improved significantly in the past
few years thanks to the adoption of deep convolutional neural network (CNN) [1, 2, 3, 17, 28].
A limitation of current CNN-based semantic segmentation methods is that they require a large amount of training
images with pixel-level annotations which are onerous and
time-consuming to acquire. This severely restricts the applicability and scalability of existing semantic segmentation
methods to real-world scenarios. On the other hand, images
with weaker supervision labels (e.g. image-level tags, object bounding boxes) are available at a much larger scale for
other computer vision tasks, e.g. image classification [12]
and object detection [7]. In addition, these weaker labels
are much easier and cheaper to collect than pixel-level labels. For example, collecting bounding boxes for each class
in an image is about 15 times faster/cheaper than pixel-level
labeling [15]. Unsurprisingly, the image-level label is the
easiest to collect, since it only requires annotating whether
an object class is present without specifying the exact location in the image.
Several weakly supervised semantic segmentation methods have been proposed to leverage image-level labels [11,
20, 22, 24] or box-level labels [4, 10, 19] for learning semantic segmentation models. These image-level label based
models generally employ smooth or constraint terms to infer potential segmentation masks from small and sparse discriminative activation regions which lack dense localization
or shape information. On the other hand, box-level label
based methods tend to use proposals [23] or unsupervised
segmentation approaches [25] to obtain the candidate segmentaion mask from each bounding box. A few related
works [8, 13, 27] explore to use both image-level and boxlevel labels as this paper. Unlike these existing methods, we
propose a flexible CNN-based framework to learn semantic segmentation models from diverse data in an end-to-end
fashion.
In this paper, we introduce an approach for learning semantic segmentation from diverse data with different types
of annotations. In particular, we consider three types of annotations (see Fig. 1) commonly used in computer vision:
image-level labels for image classification, box-level labels
for object localization/detection, and pixel-level labels for
semantic segmentation. Our proposed network consists of
three losses to train the network from training images with
different types of labels. Available training data with one
of these three types of labels can be easily added to our
learning scheme and boost the segmentation performance
accordingly. In addition, our model is flexible and can be
used in combination with any existing CNN-based architectures in semantic segmentation. By exploiting training
data with diverse supervision, our method outperforms approaches that only use images with pixel-level supervision.
2. Related Work
The performance of semantic segmentation has improved significantly in the past few years due to the success
Horse
Person
Image
Image-level label
Box-level label
Pixel-level label
Figure 1. The proposed approach can learn semantic segmentation models by making use of training images annotated for different visual
recognition tasks: image-level labels for image classification, box-level labels for object detection, and pixel-level labels for semantic
object segmentation.
of CNN. Long et al. [17] propose to convert fully connected
layers into convolutional layers and use in-network upsampling to produce feature map with the desired spatial size,
and then make dense labeling prediction of an image based
on the feature map. Meanwhile, coarse and fine features
are fused between different layers with skip connections to
generate fine details for the segmentation results. Chen et
al. [3] develop a dilated convolution layer to enlarge the
field of view (FOV) which is gradually reduced due to the
convolution operations. Thus the network can incorporate
larger context information without increasing the number of
parameters. In this paper, we consider these two segmentation frameworks as our base network.
In order to learn semantic segmentation, we usually need
images with pixel-level labels which are very expensive to
acquire. To alleviate this limitation, there has been work on
weakly supervised semantic segmentation [20, 21, 22, 24].
The goal is to learning semantic segmentation models from
labels that are less expensive and easier to obtain, e.g.
image-level labels. Some of these approaches [21, 22] are
based on multiple instance learning (MIL) framework that
considers each image as a bag of pixels and learns semantic segmentation from bag labels (e.g. image-level labels).
Pathak et al. [20] propose weakly supervised learning with
a series of linear constraints. Different linear constraints
are incorporated into a loss function on the latent distribution to optimize CNN by standard stochastic gradient descent (SGD). Qi et al. [24] introduce an object localization
branch to select and aggregate object proposals for suppressing error accumulation in training segmentation network. A new loss that combines seeding, expansion and
constrain-to-boundary is proposed in [11]. These three cues
are integrated into a segmentation network to exploit imagelevel labels.
Another line of work on weakly supervised semantic
segmentation aims to use box-level labels as the supervision. Dai et al. [4] first generate candidate segmentation
masks based on object proposal methods, and then iteratively feed these masks as supervision and update masks
from the trained network in a recursive training fashion. Papandreou et al. [19] develop an EM algorithm with a bias to
use weak label images to estimate object labels and iteratively generate latent pixel labels to optimize the CNN pa-
rameters. Khoreva et al. [10] view the problem as an issue
of input label noise and explore training as denoising. Various computer vision processing techniques (including cropping, grabcut segmentation [25] and MCG [23]) are used
to obtain effective supervision from box-level labels. Their
method can be applied in both semantic segmentation and
instance segmentation. Our approach explores a boundary
based method to obtain per-pixel labeling of each mask as
a soft segmentation and uses the proposed loss function to
address ambiguity when a pixel belongs to more than one
masks.
In order to exploit both image-level and box-level labels,
Kumar et al. [13] propose a latent structural support vector machine (LSVM) formulation. They use the latent variables model missing information in the annotation for learning the parameters of a specific-class segmentation model.
Guillaumin et al. [8] explore the object-background segmentation in the ImageNet [5]. The segmentation process
can be propagated at the image and class level by transferring appearance models from the previous stage. Xu et
al. [27] build a unified model to incorporate different combinations of weak supervision (image level tags, bounding
boxes, partial labels) to produce a pixel-wise labeling.
3. Our Approach
Given a set of training images with different forms of annotations (image-level labels, box-level labels, pixel-level
labels), our goal is to learn a semantic segmentation model
from such diverse data. The proposed approach builds on
any off-the-shelf semantic segmentation network. Figure 2
illustrates the overall architecture of our approach. The input image of arbitrary size is fed into the fully convolutional
network to generate a feature map which has the same size
as the input image. Then a separate branch is designed to
train the network for each of the three different annotations
that the image could have. Note that all the three branches
in Fig. 2 operate on the same feature map produced by the
same fully convolutional network. So the supervision signals in any of these branches will propagate to the fully convolutional network.
We use FCN [17] as an example segmentation network
to present our approach in this section. FCN uses replace
Global average
pooling
Horse
Person
Feature Vector
(1 (C+1))
Fully
Convolutional
Network
Input image
(h w 3)
Soft Seg.
Feature map
(h w (C+1))
Annotation-Specific
Loss Module
Figure 2. An overview of our approach. Given an image of arbitrary size, a per-class feature map is generated by a fully convolutional
network (e.g., FCN [17]). The feature map has the same spatial size as the input image. The number of channels of the feature map
corresponds to the number of classes (including the background class). Our model then feeds this per-class feature map to one of three
different branches. According to the available label of this input image (e.g. image-level, box-level, or pixel-level), our annotation-specific
loss module will use one of the three branches accordingly. In each branch of the annotation-specific loss module, a different loss function
is designed according to the form of the label.
the fully connected layers in standard CNN architecture
(i.e. VGG [26]) with convolutional layers for semantic segmentation. In our approach, we replace the last softmax
layer in FCN with the proposed annotation-specific loss
module shown in Fig. 2, so that we can learn the network
from diverse data with different annotations. For each input image, the network produces a per-class feature map
f ∈ Rh×w×(C+1) , where h and w are the height and width
of the input image, and C is the total number of object categories (excluding the background class). This feature map
can be used to generate the final semantic segmentation, e.g.
by assigning the class with the highest value for each pixel.
If we have ground-truth per-pixel labels on a training image, we can define a loss function to measure the difference
between f and the ground-truth per-pixel labels. However,
the challenge is that some of our training images are not
annotated at the pixel level, but we still want to utilize all
these available data with weaker supervision labels (imagelevel labels and box-level labels). To this end, we need to
somehow convert the feature map f to a different form of
prediction label, so that an appropriate loss function can be
defined. In our model, the feature map is then fed to one of
the three different branches with an annotation-specific loss
according to the available label of this input image. In the
following, we explain the details of each branch.
3.1. Image-level label
This branch (see Fig. 3) is used when the training image
only has the image-level label l = [l1 , l2 , ..., lC ], where lc is
a binary label indicating the presence/absence of the object
class c in this image. If lc = 1, the object class c is present
Horse
Person
•
•
•
non-train
non-tv
Figure 3. Image-level label training branch. We use the FCN [17]
segmentation network and replace the last softmax loss layer with
the specific loss module. The length of output vector is the same
as the number of classes.
somewhere in the image. If lc = 0, the object class c is
not in the image. Since image-level labels do not contain
any localization or shape information of the objects, they
can not be used directly to provide supervision for semantic
segmentation models which require per-pixel labeling. Instead, we use the feature map f to predict an image-level
classification label and define a loss function in terms of the
image-level classification. We attach a global average pooling layer [14] on the feature map f to extract a label vector
v ∈ R1×(C+1) in which each element of v corresponds to a
specific category, i.e.:
P
fi,c
vc = i∈Ω
∀c ∈ {0, 1, 2, ..., C}
(1)
|Ω|
where Ω denotes the set of pixels in the image and fi,c is
the value of the i-th pixel in the c-th channel in the feature
map f .
Then we apply a sigmoid activation layer to v and use
a multi-category binary cross-entropy loss to measure the
difference between v and the ground-truth label l as follows:
C
e−vc
1 X
1
)+(1−lc )·log(
)
lc ·log(
−v
c
C c=1
1+e
1 + e−vc
(2)
Here, we ignore the loss of the background class which exists in every image so as to push the network to correctly
predict object categories.
Limage = −
3.2. Box-level label
If an image has the box-level label (i.e. object bounding box), it is sent to the second branch in our annotationspecific loss module. A box-level label provides the rough
localization and size of an object in the image, but it does
not provide detailed shape information of the object. In this
branch, we convert the ground-truth box-level labeling of
the training images to a per-pixel labeling in each bounding
box. Then we use the per-pixel labeling to define the loss
function used in this branch.
(a)
(b)
sofa
uncertain
regions
(c)
(d)
(e)
Figure 4. Example of a segmentation mask obtained from the boxlevel label of an object: (a) input image; (b) UCM; (c) box-level
label (i.e. bounding box); (d) multi-scale segmentation; (e) final
object mask for this object. This figure is best viewed in color.
Darker pixels in Fig. 4 (b) correspond to locations that are
more likely to be true object boundaries. Within each annotated box, we first normalize boundary strength to be within
[0, 1] and then use three strength thresholds of { 41 , 24 , 43 } to
generate segmentation masks of different scales (Fig. 4 (d)).
From these three coarse-to-fine segmentation masks, we fill
regions starting from the inner center region to the maximum occluded boundary and choose a certain mask of a
scale until it takes up at least α% of the region in this box as
a confident mask. Finally, we assign the corresponding category label to this confident mask and mark the remaining
regions segmented in the coarsest segmentation as uncertain
regions. Other regions that do not correspond to foreground
in any of these three segmentation masks in the box are considered as background. The generated mask is shown in
(Fig. 4 (e)). Our method for generating mask labeling is
motivated by the observations in [10]: the area of an object mask should be a reasonable size within an annotated
box, and the inner box region should have higher chances of
overlapping with the corresponding object. We use α = 30
in our experiments.
We also need to consider the ambiguity when a pixel belongs to more than one mask (e.g. in the case of overlapping
segmented masks due to overlapping bounding boxes). In
order to deal with this case, we define a binary value si,c
to indicate whether the i-th pixel belongs to the c-th object
class (including uncertain regions). If a pixel belongs to
two masks with different object classes c and c0 , we assign
si,c = 1 and si,c0 = 1 as a soft segmentation pixel. Then
PC
we denote si = c=0 si,c and define the loss function for
the box-level label as follows:
Lbox = −
(b)
(a)
sofa
cat
tv
soft
ignore
(c)
(d)
(e)
Figure 5. Example of soft segmentation annotation obtained from
box-level labels: (a) input image; (b) pixel-level ground truth; (c)
box-level labels; (d) object masks corresponding to each of the
bounding box; (e) final segmentation mask.
Figure 4 illustrates the process of generating a mask labeling from a box-level label. We leverage ultrametric contour map (UCM) [18] to obtain an object mask of each annotated box. As shown in Fig. 4 (b), UCM assigns different boundary strength values based on region differences.
C
efi,c
1 X 1 X
1(si,c = 1) log( PC
)
fi,j
C
si c=0
j=0 e
(3)
i∈Ω
The intuition of Eq. 3 is as follows. If a pixel belongs to a
single mask, we consider this pixel to be a positive instance
of the corresponding object class. But if a pixel belongs to
multiple masks of K different object classes, we essentially
consider this pixel to be a positive instance of each of the K
object classes with a weight of 1/K. Similar to [10], for the
remaining pixels belonging to uncertain regions (i.e. white
regions in Fig. 5), we ignore the loss on these pixels during
training.
3.3. Pixel-level label
If a training image has pixel-level labels, we use the standard fully supervised semantic segmentation learning approach. Let pi ∈ {0, 1, ..., C} be the label of pixel i in the
image. The loss for the pixel-level label image can be calculated by the summation of softmax losses from all pixels
as:
Lpixel = −
C
1 XX
efi,c
)
1(pi = c) log( PC
fi,j
C c=0
j=0 e
(4)
i∈Ω
3.4. Learning
Let Simage , Sbox , Spixel denote the subset of images
with image-level, box-level, and pixel-level labels. Our
final loss is the summarization of the losses defined in
Sec. 3.1 3.2 3.3 over images in each of the corresponding
subset:
X
X (n)
X
(n)
(n)
Limage +
Lfinal =
Lbox +
Lpixel (5)
n∈Simage
(n)
n∈Sbox
(n)
P
nii / i ti
P
• mean accuracy (mAcc): (1/C) i nii /ti
P
P
• mean IU (mIU): (1/C) i nii /(ti + j nji − nii )
• pixel accuracy (pAcc):
n∈Spixel
(n)
Here we use Limage , Lbox , Lpixel to denote the corresponding
loss defined on the n-th training image in Simage , Sbox and
Spixel , respectively.
To optimize the loss function, we use stochastic gradient descent with momentum of 0.9, weight decay of 0.0005
and learning rate of 0.0001. The batch size is set to 1 in
order to take an image of arbitrary size as the input and
the total iteration is based on the original setting of adopted
segmentation framework and adjusted proportionally with
the amount of available images. In each mini-batch, we
pick images from one of the subsets Simage , Sbox , Spixel .
In other words, in each mini-batch, only one branch of the
annotation-specific loss module is activated.
4. Experiments
This section describes the experimental evaluation of our
method. We first describe the dataset and the experiment
setup in Sec. 4.1. Then we present the experimental results
in Sec. 4.2 and comparison with other methods in Sec. 4.3.
We perform additional ablation analysis in Sec. 4.4.
4.1. Dataset and Setup
Dataset: We evaluate the proposed approach on the PASCAL VOC 2012 benchmark dataset [6]. The dataset
consists of images of 20 object categories and the background class, and also has instance-level label augmented
by [9] to produce box-level annotation. Following the prior
work [19, 24], we use the augmented train dataset [9] which
contains 10582 images with pixel-level labels. From the
original PASCAL VOC 2012 train dataset, we create three
subsets (for image-level, box-level, pixel-level). We consider three different ratios of |Spixel |, |Sbox |, |Simage | shown
in Table 1.
Evaluation metrics: We use the following four evaluation
metrics defined in [17] to measure the performance of the
semantic segmentation results. Specifically, let nij be the
number
of pixels of class i predicted to be class j and ti =
P
n
be
the total number of pixels of class i. These four
j ij
metrics are defined as follows:
P
i
• frequency
IU
P
P weighted P
( k tk )−1 i ti nii /(ti + j nji − nii )
(fwIU):
Methods for comparison: We consider FCN [17] and
DeepLab [2] as our base networks and reimplement them
in pytorch 1 . For each base network, we compare several
different variants of our model. Assuming FCN is used as
the base network, these variants are defined as follows:
• FCN-p: this method ignores the training images with
box-level or image-level annotations (i.e. Simage and
Sbox ). It only learns the model parameters from the
subset Spixel with pixel-level annotations.
• FCN-p+i: this method uses Spixel and Simage , but it ignores Sbox .
• FCN-p+b: this method uses Spixel and Sbox , but it ignores
Simage .
• FCN-p+b+i (i.e. our method): this is our method that
makes use of all three types of data.
4.2. Experimental Results
Table 2 shows the results of learning the FCN model [17]
from diverse data on the PASCAL VOC 2012 validation
dataset. From the results, we can see that the proposed approach can effectively improve the semantic segmentation
performance by exploiting diverse supervision. In particular, using additional weak supervision data leads to better
performance. Our approach using all three subsets outperforms other alternatives that only use part of the data. In
addition, the relative sizes of the three subsets affect the final performance. If we have more images with pixel-level
annotations, the performance is better. This is intuitive since
pixel-level annotations provide the most informative supervision for the learning algorithm. When we have more training images with image-level labels and less with pixel-level
labels (e.g. when |Spixel | : |Sbox | : |Simage | = 1 : 5 : 10), the
performance boost of our approach is even more significant.
Our proposed method is also fairly general and can be
used with any existing CNN-based semantic segmentation
models. We show the generality by replacing FCN [17] with
DeepLab [2] (without CRFs). The results are shown in Table 3. Our approach using all available training data outperforms other alternatives that only use part of the training
data.
In order to compare the proposed method with fully supervised model, we hereby calculate upper bounds of corresponding variants of our model. The upper bound uses fully
pixel-level labels with the same amount of training images
1 https://github.com/pytorch/pytorch
method
FCN-p
FCN-p+i
FCN-p+b
FCN-p+b+i
annotation
pixel
pixel+image
pixel+box
pixel+box+image
p:b:i=1:1:1
3527
3527+3528
3527+3527
3527+3527+3528
p:b:i=1:2:3
1763
1763+5293
1763+3526
1763+3526+5293
p:b:i=1:5:10
662
662+6610
662+3310
662+3310+6610
Table 1. Three different settings of the training data considered in the experiments. Here we use “p:b:i” to denote the ratio of the images in
the three subsets Spixel , Sbox , Simage . In each setting, we list the number of training images in each subset.
Method (p:b:i=1:5:10)
FCN-p
FCN-p+i
FCN-p+b
FCN-p+b+i
Method (p:b:i=1:2:3)
FCN-p
FCN-p+i
FCN-p+b
FCN-p+b+i
Method (p:b:i=1:1:1)
FCN-p
FCN-p+i
FCN-p+b
FCN-p+b+i
pAcc
84.79
85.48
87.18
87.47
pAcc
86.73
87.35
88.00
88.16
pAcc
88.24
88.52
88.94
88.96
mAcc
67.95
68.43
72.25
72.65
mAcc
72.41
72.30
74.49
75.40
mAcc
72.96
74.20
75.99
76.35
mIU
51.08
53.91
57.07
57.51
mIU
55.58
57.04
58.94
58.97
mIU
58.13
59.73
60.20
60.88
fwIU
77.42
77.67
80.20
80.80
fwIU
80.13
80.59
81.70
82.02
fwIU
81.93
82.26
82.96
83.13
ratio of |Spixel | : |Sbox | : |Simage | = 1 : 1 : 1, our model is
only about 2% less than the upper bound in terms of mean
IU.
Figure 7 shows some qualitative examples of various
FCN models. From the qualitative results, we can see that
using the images in Simage (in addition to Spixel ) helps to
predict exact categories in the images (see examples in row
2 and 5 in Fig. 7). The additional images in Sbox help to
obtain a better localization and shape for the objects (see
examples in the top and bottom rows in Fig. 7). By making
full use of all available data with different forms of labels,
the proposed approach generally produces more precise
semantic segmentation results than using part of them.
Table 2. Comparison of different variants of FCN models on the
PASCAL VOC 2012 validation dataset.
Method (p:b:i=1:5:10)
DeepLab-p
DeepLab-p+i
DeepLab-p+b
DeepLab-p+b+i
Method (p:b:i=1:2:3)
DeepLab-p
DeepLab-p+i
DeepLab-p+b
DeepLab-p+b+i
Method (p:b:i=1:1:1)
DeepLab-p
DeepLab-p+i
DeepLab-p+b
DeepLab-p+b+i
pAcc
87.64
87.99
88.31
88.65
pAcc
88.55
88.84
89.53
89.60
pAcc
89.29
89.96
90.05
90.11
mAcc
75.50
76.17
77.56
78.22
mAcc
77.57
78.41
79.46
79.70
mAcc
78.21
80.10
80.15
80.39
mIU
49.56
50.08
53.44
53.86
mIU
53.85
53.98
55.08
56.43
mIU
56.20
56.53
57.41
58.05
fwIU
81.80
82.68
82.82
83.41
fwIU
83.31
83.69
84.65
84.74
fwIU
84.20
85.13
85.41
85.49
Figure 6. Comparison of different variants of FCN models and
their corresponding upper bound on the PASCAL VOC 2012 validation dataset under different ratios of training data splits. ∗ ub is
the upper bound which uses fully pixel-level labels with the same
amount of training images to learn the FCN model. The lighter
colors show the results of different variants of the proposed model
and the corresponding darker colors show upper bound results.
Table 3. Comparison of different variants of DeepLab models on
the PASCAL VOC 2012 validation dataset.
4.3. Comparison with Other Methods
to produce the segmentation performance of our model. In
other words, this is the best performance we can get if all
the training data have pixel-level labels. As shown in Fig. 6,
the performance of different variants of our model is comparable to the upper bound performance obtained by learning FCN on the training data using pixel-level labels under
different ratios of training data splits. In particular, with the
In Table 4, we compare our model with [27] which is also
developed for semantic segmentation under various forms
of supervision. Since [27] uses the SIFT-flow dataset [16]
in the experiment, we apply our approach on the SIFT-flow
in order to compare with [27]. SIFT-flow contains 2,688
images with pixel-level labels for 33 semantic categories,
including 2,488 for training and 200 for testing. Following the setting in [27], we use a sampling ratio of 0.5 for
Image
Ground Truth
FCN-p
FCN-p+i
FCN-p+b
FCN-p+b+i
Figure 7. Some qualitative examples of our approach on the PASCAL VOC 2012 validation dataset. In each row, from left to right: Image,
Ground Truth and segmentation results generated by FCN-p, FCN-p+i, FCN-p+b and FCN-p+b+i.
image-level labels and pixel-level labels respectively. Our
model outperforms [27] by a large margin in terms of pixel
accuracy and mean accuracy.
We also compare our results with the state-of-the-art
weakly supervised semantic segmentation methods [4, 10,
19] that use box-level labels. Similarly, we follow their settings with the same numbers of pixel-level and box-level labels in training on PASCAL VOC 2012. Table 5 shows the
semantic segmentation results. Note that all three compared
methods are originally proposed for weakly supervised semantic segmentation rather than diverse supervision. Since
the source codes of these methods are not available, we directly take the performance numbers in those papers. However, these methods in Table 5 use different backbone networks and some of them apply CRF post-processing to refine the results. So it is difficult to have a fair comparison
directly. This can be seen from the fact that the fully supervised versions of these methods give slightly different
Method
Xu et al. [27]
Ours
pAcc
72.8
82.08
mAcc
36.5
58.76
Table 4. Comparison to [27] on the SIFT-flow dataset under the
same setting (p:i=1:1).
performance numbers (see Table 5). Although our model
is not specifically designed for weakly supervised semantic
segmentation, it can still achieve comparable performance
in this setting.
4.4. Ablation Analysis
In this section, we perform additional ablation analysis of our approach. We conduct experiments on the FCN
model under the ratio of |Spixel | : |Sbox | : |Simage | = 1 : 5 :
10, which has the least amount of pixel-level labels and the
method has to rely more on images with the box-level or
Method
WSSL [19]
BoxSup [4]
SDI [10]
Ours
val. semi / full
65.1 / 67.6 (96.30)
63.5 / 63.8 (99.53)
65.8 / 69.1 (95.22)
60.14 / 63.14 (95.25)
test. semi / full
66.6 / 70.3 (94.74)
66.2 / - (-)
66.9 / 70.5 (94.89)
61.42 / 62.2 (98.75)
Table 5. Comparison of different weakly-supervised segmentation
methods on the PASCAL VOC 2012 validation and test datasets
under the same setting as [10] in terms of mIU. The numbers of
pixel-level and box-level annotations are 1.4k and 9k, respectively.
For each method, we show the mIU in both semi-supervised and
fully-supervised settings. Since different methods use different
design choices (such as backbone network, CRF postprocessing,
etc), it is difficult to compare their performance numbers directly.
This can be seen from the fact that these different methods have
different mIU in the fully supervised setting. The numbers in (·)
indicate the ratio of the mIU between the semi-supervised and
fully-supervised settings. In the ideal case, this number will be
close to 100 indicating that the semi-supervised setting achieves
the upper-bound (full-supervised) performance.
image-level labels.
First, we explore and compare other alternative ways of
exploiting the box-level supervision. We consider four different ways of using box-level supervision as follows:
Hard segmentation: this method uses the same approach to obtain the mask labeling from box-level labels
as Sec. 3.2, but it does not keep all overlapping segmented
masks of different object classes and weight them in the
loss. Instead, every pixel has to belong to a certain class.
We randomly assign a class to the overlapping region of
multiple segmented masks and then uses it as the pseudo
pixel-level label.
GrabCut: this method takes each box-level label and applies GrabCut [25] to obtain the binary mask of the object
inside the bounding box. It then considers the binary mask
as the pseudo pixel-level labels.
MCG: this method uses the MCG algorithm [23] to generate a set of candidate object masks for an input image. Inside each box-level label, an object mask that has the maximum overlap with the corresponding bounding box is selected as the object mask for the object inside this bounding
box. Similarly, this object mask is used as the pseudo pixellevel labels. This strategy has been used in [10].
Raw box: this method assigns the class label to each pixel
within a bounding box and considers the pixels outside of
any bounding boxes in the image as background. Therefore,
Eq. 3 can also be adopted to address ambiguity caused by
overlapping bounding boxes.
Table 6 shows the results of different strategies of using
box-level labels in the proposed approach, respectively. The
segmentation performance with our proposed box-level supervision outperforms all other alternatives. In addition, the
proposed box-level supervision loss in Eq. 3 shows its ef-
FCN
Pseudo
pixel-level
Our loss
Grabcut
MCG
Hard seg.
Raw box
Ours
pAcc
86.68
86.71
87.30
87.04
87.47
mAcc
71.91
70.87
71.80
71.07
72.65
mIU
55.60
56.19
57.35
56.91
57.51
fwIU
79.78
79.51
80.33
79.87
80.80
Table 6. Comparison of different strategies to use box-level labels
in FCN on the PASCAL VOC 2012 validation dataset.
fectiveness to address the ambiguity caused by overlapping
regions with multiple uncertain labels. Even using raw box
method in the box-level supervision, the semantic segmentation performance is comparable to their pseudo pixel-level
methods which require additional pre-processing with segmentation or proposal.
Figure 8. Effect of using three strength thresholds of scales to generate segmentation masks.
We also analyze the effects of some of the hyperparameters in our method. Figure 8 shows the performance of using one strengh threshold ( 12 , 24 , or 34 ) versus using all three
strength thresholds (our method) when generating the segmentation masks. We can see that using all three strength
thresholds outperforms just using one threshold.
5. Conclusion
We have proposed a method for learning CNN-based semantic segmentation models from training images with diverse annotations, including image-level labels, box-level
labels and pixel-level labels. The proposed model consists
of three specific losses for each of three annotations and
can be easily used together with any existing CNN-based
semantic segmentation networks. The experimental results
show that our method can effectively make use of diverse
training data with different levels of supervision.
Acknowledgement: This work was supported by NSERC,
the National Natural Science Foundation of China under
Grant No. 61771301, and the University of Manitoba GETS
funding program. We also thank NVIDIA for the GPU donations.
References
[1] V. Badrinarayanan, A. Kendall, and R. Cipolla. Segnet: A
deep convolutional encoder-decoder architecture for image
segmentation. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 2017.
[2] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International
Conference on Learning Representations, 2015.
[3] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. Deeplab: Semantic image segmentation with
deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915, 2016.
[4] J. Dai, K. He, and J. Sun. Boxsup: Exploiting bounding
boxes to supervise convolutional networks for semantic segmentation. In IEEE International Conference on Computer
Vision, 2015.
[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database. In
IEEE Conference on Computer Vision and Pattern Recognition, 2009.
[6] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and
A. Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 2010.
[7] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. In IEEE Conference on Computer Vision and
Pattern Recognition, 2014.
[8] M. Guillaumin, D. Küttel, and V. Ferrari. Imagenet autoannotation with segmentation propagation. International
Journal of Computer Vision, 2014.
[9] B. Hariharan, P. Arbeláez, L. Bourdev, S. Maji, and J. Malik.
Semantic contours from inverse detectors. In International
Conference on Computer Vision, 2011.
[10] A. Khoreva, R. Benenson, J. Hosang, M. Hein, and
B. Schiele. Simple does it: Weakly supervised instance and
semantic segmentation. In IEEE Conference on Computer
Vision and Pattern Recognition, 2017.
[11] A. Kolesnikov and C. H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In European Conference on Computer Vision,
2016.
[12] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing Systems, 2012.
[13] M. P. Kumar, H. Turki, D. Preston, and D. Koller. Learning specific-class segmentation from diverse data. In IEEE
International Conference on Computer Vision, 2011.
[14] M. Lin, Q. Chen, and S. Yan. Network in network. In International Conference on Learning Representations, 2014.
[15] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick,
J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Dollár.
Microsoft COCO: Common objects in context. In European
Conference on Computer Vision, 2014.
[16] C. Liu, J. Yuen, and A. Torralba. Nonparametric scene parsing via label transfer. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 2011.
[17] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In IEEE Conference on
Computer Vision and Pattern Recognition, 2015.
[18] K.-K. Maninis, J. Pont-Tuset, P. Arbelaez, and L. Van Gool.
Convolutional oriented boundaries: From image segmentation to high-level tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
[19] G. Papandreou, L.-C. Chen, K. Murphy, and A. L. Yuille.
Weakly-and semi-supervised learning of a dcnn for semantic
image segmentation. In IEEE International Conference on
Computer Vision, 2015.
[20] D. Pathak, P. Krahenbuhl, and T. Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In IEEE International Conference on Computer Vision,
2015.
[21] D. Pathak, E. Shelhamer, J. Long, and T. Darrell. Fully convolutional multi-class multiple instance learning. In International Conference on Learning Representations, 2015.
[22] P. O. Pinheiro and R. Collobert. From image-level to pixellevel labeling with convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2015.
[23] J. Pont-Tuset, P. Arbelaez, J. T. Barron, F. Marques,
and J. Malik. Multiscale combinatorial grouping for image segmentation and object proposal generation. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
39(1):128–140, 2017.
[24] X. Qi, Z. Liu, J. Shi, H. Zhao, and J. Jia. Augmented feedback in semantic segmentation under image level supervision. In European Conference on Computer Vision. Springer,
2016.
[25] C. Rother, V. Kolmogorov, and A. Blake. Grabcut: Interactive foreground extraction using iterated graph cuts. ACM
Transactions on Graphics, 2004.
[26] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. International
Conference on Learning Representations, 2015.
[27] J. Xu, A. G. Schwing, and R. Urtasun. Learning to segment
under various forms of weak supervision. In IEEE Conference on Computer Vision and Pattern Recognition, 2015.
[28] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet,
Z. Su, D. Du, C. Huang, and P. H. Torr. Conditional random
fields as recurrent neural networks. In IEEE International
Conference on Computer Vision, 2015.
| 1cs.CV
|
arXiv:cs/0610120v2 [cs.MS] 22 Oct 2008
Classdesc and Graphcode: support for scientific
programming in C++
Russell K. Standish
School of Mathematics
University of New South Wales
[email protected]
http://parallel.hpc.unsw.edu.au/rks
Duraid Madina
Department of Systems Studies
University of Tokyo
[email protected]
April 3, 2018
Abstract
Object-oriented programming languages such as Java and Objective
C have become popular for implementing agent-based and other objectbased simulations since objects in those languages can reflect (i.e. make
runtime queries of an object’s structure). This allows, for example, a
fairly trivial serialisation routine (conversion of an object into a binary
representation that can be stored or passed over a network) to be written.
However C++ does not offer this ability, as type information is thrown
away at compile time. Yet C++ is often a preferred development environment, whether for performance reasons or for its expressive features such
as operator overloading.
In scientific coding, changes to a model’s code takes place constantly,
as the model is refined, and different phenomena are studied. Yet traditionally, facilities such as checkpointing, routines for initialising model
parameters and analysis of model output depend on the underlying model
remaining static, otherwise each time a model is modified, a whole slew of
supporting routines needs to be changed to reflect the new data structures.
Reflection offers the advantage of the simulation framework adapting to
the underlying model without programmer intervention, reducing the effort of modifying the model.
In this paper, we present the Classdesc system which brings many of
the benefits of object reflection to C++, ClassdescMP which dramatically
simplifies coding of MPI based parallel programs and Graphcode a general
purpose data parallel programming environment.
1
1
Introduction
This paper describes Classdesc, ClassdescMP and Graphcode, techniques for
building high performance scientific codes in C++.
Classdesc is a technique for providing automated reflection capabilities in
C++, including serialisation support. ClassdescMP builds on Classdesc’s serialisation capability to provide a simple interface to using the MPI message
passing library with objects. Graphcode implements distributed objects on a
graph, where the objects represent computation, and the links between objects represent communication patterns. It is a higher level of abstraction than
the message passing paradigm of ClassdescMP, yet more general and powerful
than traditional data parallel programming paradigms such as High Performance
Fortran[2] or POOMA[12].
This paper is organised into three sections, describing the three technologies
in more detail. The final section concludes with a description of the current
status of the code, and where it can be obtained from.
2
2.1
Classdesc
Reflection and Serialisation
Object reflection allows straightforward implementation of serialisation (i.e. the
creation of binary data representing objects that can be stored and later reconstructed), binding of scripting languages or GUI objects to ‘worker’ objects and
remote method invocation. Serialisation, for example, requires knowledge of
the detailed structure of the object. The member objects may be able to be
serialised (e.g. a dynamic array structure), but be implemented in terms of a
pointer to a heap object. Also, one may be interested in serialising the object in
a machine independent way, which requires knowledge of whether a particular
bitfield is an integer or floating point variable.
Languages such as Objective C give objects reflection by creating class objects and implicitly including an isa pointer in objects of that class pointing to
the class object. Java does much the same thing, providing all objects with the
native (i.e. non-Java) method getClass() which returns the object’s class at
runtime, as maintained by the virtual machine.
When using C++, on the other hand, at compile time most of the information about what exactly objects are is discarded. Standard C++ does provide
a run-time type information mechanism (RTTI), however this is only required
to return a unique signature for each type used in the program. Not only is this
signature compiler dependent, it could be implemented by the compiler enumerating all types used in a particular compilation, and so the signature for a given
type would differ from program to program! Importantly, standard RTTI does
not provide any information on the internal structure of a type, nor methods
implemented.
The solution to this problem lies (as it must) outside the C++ language per
2
se, in the form of a separate program which parses the interface header files. A
number of C++ reflection systems do this: SWIG[4], being perhaps the oldest,
parses a somewhat simplified C++ syntax with markup, provides exposure to
scripting language of selected top level objects. Reflex[13], a more recent system
than Classdesc[9], is interesting in that it use the GCC-XML parser (based on
the C++ front end of GCC) to parse the input file to build a dictionary of type
properties. Classdesc differs from these other attempts by traversing the data
structures recursively at runtime, providing a genuine solution to serialisation, as
well as allowing “drill down” of simulation objects in an interactive exploration
of a running simulation.
These are generically termed object descriptors. The object descriptor generator only needs to handle class, struct and union definitions. Anonymous structs
used in typedefs are parsed as well. What is emitted in the object descriptor is
a sequence of function calls for each base class and member, similar in nature to
compiler generated constructors and destructors. Function overloading ensures
that the correct sequence of actions is generated at compile time.
For instance, assume that your program had the following class definition:
class jellyfish: public animal
{
double position[D], velocity[D], radius;
int colour;
};
and you wished to generate a serialisation operator called pack. Then this
program will emit the following function declaration for jellyfish:
#include "pack_base.h"
void pack(pack_t *p, string nm, jellyfish& v)
{
pack(p,nm,(animal&)v);
pack(p,nm+".position",v.position,is_array(),1,D);
pack(p,nm+".velocity",v.velocity,is_array(),1,D);
pack(p,nm+".radius",v.radius);
pack(p,nm+".colour",v.colour);
}
The use of auxiliary types like is_array() improves resolution of overloaded
functions, without polluting global namespace further. This function is overloaded for arbitrary types, but is more than a template, so deserves a distinct
name. We call these functions class descriptors (hence the name Classdesc), or
simply an action for short.
Thus, calling pack(p,"",var) where var is of type test1, will recursively
descend the compound structure of the class type, until it reaches primitive
data types which can be handled by the following generic template defined for
primitive data types:
3
template <class T>
void pack_basic(pack_t *p,string desc, T& arg)
{p->append((char*)&arg,sizeof(arg));}
given a utility routine pack t::append that adds a chunk of data to a repository
of type pack t.
This can even be given an easier interface by defining the member template:
template <class T>
pack_t& pack_t::operator<<(T& x)
{::pack(this,"",x);}
so constructions like buf << foo << bla; will pack the objects foo and bla
into the object buf.
Classdesc is released as public domain software, and is available from the
Classdesc website1 . The pack and unpack operations work more or less as described. The type xdr_pack, derived from pack_t uses the standard unix XDR
library to pack the buffer in a machine independent way. This allows checkpoint
files to be transported between machines of different architectures, or to run the
simulation in a client-server mode, with the client downloading a copy of the
simulation whilst the simulation is in progress.
2.2
Object Exposure
Another application of reflection is exposing object internals to an external environment, such as a scripting language, or another object oriented programming
language. For computational science models, adding a scripting language has
many advantages[11]. Initialisation of the model is simply achieved by setting
a few variables within a script. Data collected can be customised without code
recompilation by simple script changes. GUI widgets can allow the real time
monitoring of the model’s variables in a graphical form during debugging and
model development. A drill down facility can be readily provided in the scripting language that allows the model to be stopped, and values of the model’s
variables queried. Being a scripted environment, the same executable can be
used for exploration in a GUI mode, or for production in a batch mode, simply
by using a different script.
Automated techniques for exposing objects to a scripting environment, or to
a different OO environment already exist. Examples include VTK[16], CORBA/IDL
and SWIG[4]. However these either straight-jacket the programmer into using
a particular programming style, or require the class definitions to be coded in a
different language, (often termed an IDL). SWIG at least has the advantage of
being able to parse any ISO standard C/C++ code. Its strong advantage is that
it already has bindings for many popular languages, including TCL, Python,
Perl and Java. Where Classdesc and SWIG might work well together is scripting Fortran applications. An experimental Fortran version of Classdesc (called
1 http://ecolab.sourceforge.net/classdesc.html
4
FClassdesc) was developed under a grant from the Australian Partnership for
Advance Computing, with serialisation of Fortran modules and a descriptor to
produce C-syntax code for use as input to SWIG.
The EcoLab agent based modelling system[20]2 uses Classdesc to expose C++
objects to the TCL scripting language[10]. If the authors had been aware of
SWIG at the time of EcoLab’s development, SWIG would probably have been
used instead, however, Classdesc’s recursive approach to analysing data structures is more useful for interactive exploration of scientific models in a simulation framework than SWIG’s approach of requiring explicit exposure of top
level objects only.
Virtually any model that is implemented as a C++ object can be dropped
into EcoLab, and one instantly has a scriptable simulation system, with GUI
plotting and drill down tools and checkpointing functionality. The main programming constraint is the DCAS requirement (§2.4), although departures from
DCAS tend to result in degraded capability rather than catastrophic failure
(such as uncompilable code).
The exposure of objects into TCL is handled by a descriptor TCL_obj.
Simple data members generate a TCL command which returns the value of
that member, and set the member if an argument is supplied (if corresponding
ostream::operator<< and istream::operator>> are defined).
Member functions whose arguments match a limited range of signatures are
also callable from TCL.
2.3
Resource Aquisition Is Initialisation (RAII)
The RAII principle [21, §14.4.1] uses stack resident objects to control the lifetime
or states of objects elsewhere in the system, such as heap resident objects. One
of the simplest and most obvious application of RAII is prevention of memory
leaks that occur through forgetting to destroy objects once they are no longer
needed. By making the raw pointer a private member of a class, placing the calls
to new and delete within member functions of the class and arranging for the
destructor to call a final delete to dispose of the object, we can then use this
class to declare a stack variable that controls the lifetime of a heap object. Figure
1 shows a simple implementation of the ISO C99 variable length automatic
arrays that is not part of the C++ standard. Like the C version, data is allocated
when control passes through the statement declaring the array variable, and is
deallocated automatically when control leaves the scope containing the array
variable, relieving the programmer of having to remember to delete the object.
Unlike the C version, however, the data is actually allocated on the heap (via
the new statement called in the constructor, rather than the stack. This is often
advantageous as many modern operating systems restrict the stack size to a few
megabytes which cannot support large arrays.
RAII is useful for many other tasks, such as ensuring files are closed and
flushed, network connections are terminated properly, software licenses released
2 http://parallel.hpc.unsw.edu.au/rks/ecolab
5
template <class T> class Array
{
T* data;
public:
Array(size_t n): data(new T[n]) {}
T& operator[](size_t i) {return data[i];}
const T& operator[](size_t i) const {return data[i];}
~Array() {delete [] data;}
};
Figure 1: Implementation of C99 variable length automatic array feature in
C++.
and very importantly, ensuring partly constructed objects are correctly cleaned
up in the event of an exception occurring[21, 14.4.1].
By way of contrast, languages such as Java and C# do not allow the RAII
technique to be deployed, as complex objects cannot reside on the stack. Instead,
garbage collection is relied upon to release objects on the heap that are no longer
needed. Since this occurs at rather indeterminate times (if ever), it cannot be
relied upon for anything other than controlling memory leaks.
One mistake seasoned Java or C# programmers make when writing C++
is to assume that the C++ new operator should be used in the same way as it
is used in the other languages. This leads to code that is hard to debug and
maintain, and has given C++ a reputation for being difficult to avoid memory
problems.
2.4
The DCAS principle
The C++ compiler automatically provides a default constructor, a copy constructor and an assignment operator, if none are explicitly provided by the
programmer, which recursively call the default constructor, copy constructor or
assignment operator repectively of the base classes and members. The use of
Classdesc is analogous — Classdesc recursively applies its descriptor on base
classes and member functions. Since serialisation is the most important Classdesc application, I call this the DCAS principle (Default constructor, Copy
constructor, Assignment and Serialisation). Classes whose members and base
classes are DCAS are also DCAS automatically, alleviating a lot of programmer
effort.
To create a DCAS object from a non-DCAS object requires wrapping. Primitive types (ints, floats, etc) are DCAS, although their default constructors do
not initialise them to any particular value, so some care must be taken with
default constructors of classes taking such types. Pointers, on the other hand
are not DCAS at all. The default constructor for a pointer does not initialise the
pointer to a valid value. The copy constructor and assignment operator merely
copies the pointer, which ends up with two references to the same object.
6
To get around this problem, various solutions have been developed. The
C++ standard defines auto_ptr<T>[21, §14.4.2], which is DCAS (to a degree).
The default constructor set the reference to NULL, and copy and assignment
operators pass control of the target object to the target of the copy or assignment
operation, leaving the original object set to NULL, which breaks some notions of
“copy”. It supports the notion of “resource aquisition is initialisation” or RAII,
so that the pointer is released when the auto_ptr object is destroyed. Its main
use is to provide a means for returning an object by reference from a function,
avoiding any performance penalties of a copy constructor or the possibility of
an exception being thrown during the copy constructor.
The Boost library[5] defined the shared_ptr<T> and intrusive_ptr<T>
concepts, which allow for multiple references to a single object, whilst still supporting RAII, which are DCA. The shared_ptr<T> concept is so useful that it
has been included into TR1[1], which is scheduled to be standardized as part of
the next C++ standard.
However, none of these concepts can be serialised, as these objects are initialised to the address of an object created by an earlier new statement. The
actual type of the object is unknown at serialisation time, only the base class
T declared in the template argument is known. This arrangement allows the
handling of polymorphic objects, which we will return to in §2.6.
The Classdesc package includes the ref<T> concept, which implements a
reference counted dynamic reference class similar to Boost’s intrusive_ptr<T>
concept, which is serialisable (hence DCAS). Instead of creating the target object outside the ref class with a new statement, as done in Boost’s smart pointer
concepts, the target object is created on first dereference. This has the advantage that the reference counter can be stored alongside the object of type T
on the heap like intrusive_ptr<T> does, without the need for T to support
any reference counting API, however T is required to be DCAS. Polymorphism
(§2.6), which requires special treatment, is not supported by ref at all, however.
The use of reference counting (whether the Classdesc ref or the Boost versions) allows heap allocated objects to be used as simply as they would with
classic garbage collection. However copying reference counted references is about
twice as expensive as simple pointer assignment, so under some circumstances,
the use of such classes may be a performance issue. By judicious use of standard
C++ references, and function inlining, this performance impact can be ameliorated, and if necessary, bare pointers can be used within the innermost scope of
pointer chasing algorithms as a performance optimisation.
Reference counted references prove effective in implementing acyclic graph
structues — deleting the reference to the head node is sufficient to ensure that all
nodes are deleted. However, cycles of references will cause objects to remain in
existence, even when no references remain to the graph structure. One possible
means of dealing with this is to perform a graph walk at graph destruction
time, deleting links to objects that have already been traversed in the walk,
thus deleting any cycles. Then deleting the head node reference is sufficient to
delete the entire structure. This operation is most conveniently handled in the
destructor of some Graph class.
7
2.5
Pointers
Pointers create difficulties for Classdesc, since pointers may point to a single
object, an array of objects, functions, members or even nothing at all. When
array sizes are known at compile time, Classdesc issues an object descriptor
that loops over the elements, however arrays allocated dynamically on the heap
through the use of new cannot be handled, even in principle.
Because pointers cause problems with the DCAS and RAII paradigms, it
is worth discussing the uses that pointers are put to in C++, and alternatives
that are available. Many of the uses have been inherited from C, where pointer
usage is almost unavoidable in practical codes. Pointers are used in C++ for
the following purposes:
Passing by reference. This use is inherited from C, but superseded by C++’s
reference types.
Dynamic arrays. C++’s std::vector<T> container can be used for most dynamic array purposes without any extra overhead. It is DCAS, provided
the element type T is also DCAS, and satisfies the RAII technique. If a
standard container is not suitable, then a purpose-built container such as
shown that in figure 1 can be provided.
Strings. std::string provides a safe and DCAS-ready alternative to char*
variables.
Graph structures. Classdesc provides the DCAS-ready ref<T> which is suitable for graphs and trees.
Polymorphic objects. Classdesc provides the poly<T> for handling polymorphic object heirarchies in a DCAS fashion (§2.6).
Opaque handles. Opaque handles are used to improve compile times by hiding the actual implementation details, including instance variables, in a
separate compilation unit. This is not a major problem, but specific methods must be provided by the programmer for construction, destruction,
copying and serialisation of the object referenced by the opaque handle.
These may call automatically generated versions of the methods in the
separate compilation unit to reduce programmer burden.
Libraries. C language API libraries will often use pointers to data structures.
Where the details of these data structures are provided as part of the
interface file, it is possible to use the automatically generated (whether
compiler or Classdesc generated) methods to implement a DCAS wrapper
around these objects. Where opaque handles are used, however, one’s
choices are limited depending on whether the appropriate methods for
implementing copying and serialisation have been provided (some means of
construction and disposing of objects will always be provided), or whether
source code is available.
8
It turns out that one can distinguish between member pointers and normal
pointers quite easily through overloading of object descriptors. Member pointers are relevant for exposing an object’s methods to a scripting interface, for
example, and are also serialised and passed between processes in ClassdescMP
to implement a form of remote procedure calling. Classdesc does not distinguish between pointers to functions and pointers to objects as simple function
overloading is not sufficient to distinguish them. However, the Boost library provides a template metaprogramming technique for distinguishing between function pointers and objects pointers in its types_trait package, so providing
overloading for function pointers is planned for the future.
By default, an attempt to serialise a pointer will issue a runtime warning.
However, if pointer members are genuinely necessary, it is possible for the programmer to specify that pointers either point to a single object of the specified
DCAS type or are NULL if invalid. We call this the graphnode protocol. This
situation is most likely to occur when using a “legacy” library that deals with
pointers, and wrapping the data with something like ref is prohibitive. The
gSOAP package[22] is an example.
Within the Classdesc system it is possible to specify that all pointers of a
given pointer type satisfies the graphnode protocol, or that all pointers within
a given graph structure satisfy the graphnode protocol. The pack descriptor
than walks the graph structure keeping a track of nodes visited so that cycles
are handled, and recursion cut off to avoid stack limits being breached.
2.6
Polymorphism
C++ has two notions of polymorphism, compile-time and runtime. Compiletime polymorphism (aka generic programming) is implemented in terms of templates, and allows the provision of code that can work on many different types
of objects. On the other hand, runtime polymorphism involves the use of virtual
member functions. Whereever generic programming can solve a task, it is to be
preferred over runtime polymorphism, as virtual member functions introduce
procedure call overhead, and inhibit optimisation. Furthermore, the use of a
DCAS class like poly introduces additional overheads.
Nevertheless, there are situations that cannot be solve with compile-time
polymorphism, for example a container containing objects of varying types.
For this purpose, Classdesc’s poly type is useful. To use poly, your object
heirarchy must implement the following interface (provided as an abstract base
class object).
struct object
{
typedef int TypeID;
virtual TypeID type() const=0;
virtual object* clone() const=0;
virtual void pack(pack_t *b) const=0;
virtual void unpack(pack_t *b)=0;
9
virtual ~object() {}
};
The type() method implements a simple runtime type identifier system. In
the case of object, it uses simple integer tags, which are assumed to be allocated
more or less consequitively to types in the type heirarchy. However, any type
may be used provided it is exported as the typedef TypeID, and an appropriate
customised type table class is defined (see below). One possibility, although by
no means the most efficient, is to use the object’s type_info object returned
by C++’s inbuilt run time type identification system[21, §15.4.4].
It is not actually necessary to use this abstract base class to use poly. The
base class (which must be default constructible, hence not abstract) is passed to
the poly template. Classdesc provides an empty concrete class Eobject which
can be used for this purpose.
To assist in deriving classes from object, the Object template is provided.
template <class This, int Type, class Base=object> struct Object;
The first template argument This is the class you’re currently defining, the
second (Type) is the integer value of its type tag and Base is the base class you
are deriving from. Eobject is defined as
class Eobject: public Object<Eobject,0> {};
and a new class (eg foo) with type ID 1 can be defined
class foo: public Object<foo,1,Eobject> {...
This saves having to explicitly provide versions of the virtual functions type(),
clone(), pack() and unpack(), as these are provided by Object. It also provides a utility method cloneT() which executes clone(), but instead of returning a bare object pointer, returns a pointer to an object of the same type as
the calling object (if legally convertible via dynamic_cast).
The synopsis of poly is:
template <class T=Eobject, class TT=SimpleTypeTable<T> >
class poly
{
public:
TT TypeTable;
poly();
poly(const polyref& x);
poly(const T& x);
poly& operator=(const poly& x);
poly& operator=(const T& x);
template <class U> void addObject();
template <class U, class A> void addObject(A);
template <class U, class A1, class A2> void addObject(A1, A2);
10
T* operator->();
T& operator*();
const T* operator->() const;
const T& operator*() const;
template <class U> U& cast();
template <class U> const U& cast();
void swap(poly& x);
};
Most of this is fairly straightforward. However the addObject() and cast()
methods need a little more explanation. To make the poly object an object of
type (say foobar), use the following calls:
poly.addObject<foobar>(); //calls foobar()
poly.addObject(1);
//calls foobar(1)
poly.addObject(1,"hello"); //calls foobar(1,"hello");
poly.addObject(foobar(x,y,z)); //more than 2 arguments
The cast method provides a convenient method casting the poly object to
a specific type. It is equivalent to calling dynamic_cast, but a little easier to
use, ie
poly.cast<foobar>().grunge() <=> dynamic_cast<foobar&>(*poly).grunge();
The return type was chosen to be a reference, not a pointer, as this is the more
convenient form. It can easily be converted to a pointer with the & operator.
The TypeTable member of poly must implement the following interface
class typetable
{
Base& operator[](TypeID);
void register_type(const Base&);
};
where Base is the base type of the poly class, and is basically a database of
reference objects, from which new objects can be constructed using clone(),
given a type identifier. This is used for implementing serialisation. Classdesc
provides simple implementation of this as SimpleTypeTable<Base>, where the
TypeIDs are integers that are reasonably close to each other.
2.7
Member Privacy
Serialisation descriptors need access to all members of an object, including private and protected ones. Since in C++ class namespaces are closed by design
(no new members can be added, except by inheritance), descriptors need to be
placed in a global or an open namespace. This means that friend declarations
need to be added to all class definitions with private or protected areas. The
11
convention adopted by Classdesc is to define two macros that expand to a list
of friend declarations for the descriptors, similar to the following:
#define CLASSDESC_ACCESS(type)\
friend void pack(pack_t *,eco_string,type&);\
friend void unpack(unpack_t *,eco_string,type&);
#define CLASSDESC_ACCESS_TEMPLATE(type)\
friend void pack<>(pack_t *,eco_string,type&);\
friend void unpack<>(unpack_t *,eco_string,type&);
Then placing a CLASSDESC_ACCESS statement in the class definition allows
the descriptor access to the private members of the class:
class foo
{
int bar;
CLASSDESC_ACCESS(foo);
public:
float bar2;
};
An auxiliary program insert-friend is provided as part of the Classdesc
package to automatically insert these macros into class definitions.
For object exposure, only public members need to be processed by the descriptor. Classdesc provides a -respect_private flag to indicate that private
and protected members should be ignored by the descriptor.
3
3.1
ClassdescMP: easy MPI programming in C++
MPIbuf
MPI[17] is an industry standard API for constructing distributed memory parallel applications using the message passing metaphor. Originally designed for
use with Fortran77 and C, it primarily deals with passing arrays of simple types
such as characters, integers or floating point numbers. In a later incarnation,
C++ bindings to the library were provided as part of the MPI-2 standard. It
primarily added support for the MPI namespace, communicators as objects and
support for C++ exception handling. However, messages are fundamentally
composed of arrays of simple types.
Classdesc’s general serialisation operation solves the problem of passing messages of complex objects as the pack descriptor turns a sequence of complex
objects into an array of bytes. In ClassdescMP, the MPIbuf type is derived from
pack_t, so messages can be constructed in a streaming fashion, eg:
buf << a << b << send(1);
12
which sends a and b to process 1.
Streaming MPI messages is not new — it is used in PARA++[6], and in
OOMPI[3], for example. However, in these packages, programmers are required
to provided explicit serialisation routines for complex types.
To receive a message, use the MPIbuf::get():
buf.get() >> a >> b;
Optional arguments to get allow selective reception of messages by source and
tag.
By setting the preprocessor macro HETERO, MPIbuf is derived from xdr_pack
instead of pack_t. This allows ClassdescMP programs to be run on heterogeneous clusters, where numerical representation may differ from processor to
processor.
In MPI-2 C++ bindings, the basic object handling messages is a communicator. In ClassdescMP, an MPIbuf has a communicator. It also has a buffer,
and assorted other housekeeping members. Some of these are used for managing
asynchronous communication patterns:
{
MPIbuf buf;
buf << a << isend(1);
while (something_to_do && !buf.sent()) do_something;
buf.wait();
buf << b << isend(2);
...
}
When buf goes out of scope, an implicit MPI_Wait is called to ensure that the
message has been correctly sent.
Often, one needs to perform all-to-all exchange of data. To do this, we use
an MPIbuf_array:
{
...
tag++;
MPIbuf_array sendbuf(nprocs());
for (unsigned proc=0; proc<nprocs(); proc++)
{
if (proc==myid()) continue;
sendbuf[proc] << requests[proc] << isend(proc,tag);
}
for (int i=0; i<nprocs()-1; i++)
{
MPIbuf b;
b.get(MPI_ANY_SOURCE,tag);
b >> rec_req[b.proc];
}
}
13
This piece of code is copied verbatim from the Graphcode library (§4). Note the
use of a tag variable to ensure that unrelated groups of communication do not
get mixed up. Also, when sendbuf goes out of scope, an implicit MPI_Waitall
called, which ensures that all messages in the group have been sent.
3.2
MPISPMD
Whilst MPIbuf and MPIbuf_array are the heart of ClassdescMP, there is also
some application framework support. Two programming models are supported:
an SPMD mode, which simply wraps up the MPI setup and teardown into
an object, and a master-slave mode in which the master thread controls slave
thread objects via remote method invocation.
The SPMD mode is rather similar to that of PARA++[6]. By instantiating
an object of type MPISPMD, the MPI environment is initialised. One key feature of Classdesc’s implementation is that MPI_Finalize() is called from the
MPISPMD object’s destructor — not only does this save the programmer from
having to remember to do this, but it is also called during stack unwinding if
an exception is thrown. This alleviates the problem with some MPI implementations (eg MPICH) which leave active threads running and consuming CPU
time if MPI_Finalize() is not called.
3.3
MPIslave
The master-slave mode is a more powerful feature of ClassdescMP. Setting up
the structure of a master-slave program is very tedious and error prone. The
MPIslave class is designed to make master-slave algorithms simple to program.
When a MPIslave object is instantiated, a slave “interpreter” object is instantiated on each process to receive messages from the master. As MPIslave
needs to know the type of object to be instantiated on the slave processes, it is
implemented as a template, with the type of slave object passed as the template
parameter.
A message sent to the slave process starts with a method pointer of type:
void (S::*)(MPIbuf&) where S is the slave object type, followed by the arguments to be passed. That method of the slave object is then called, with the
arguments passed through MPIbuf argument, and any return values also passed
through MPIbuf argument:
struct S
{
void foo(MPIbuf& args)
{
int x,y,r;
args >> x >> y;
...
args.reset() << r;
}
14
};
main(int argc, char** argv)
{
MPIslave<S> C;
MPIbuf buf;
int x=1, y=2;
buf << &S::foo << x << y << send(1);
}
When the MPIslave object is destroyed on the master process, it arranges
for all the slave objects to be MPI_Finalized() and destroyed also.
MPIslave also has features for managing a pool of idle slaves:
MPIslave<S> C(argc,argv);
vector<job> joblist;
for (int p=1; p<C.nprocs && p<joblist.size(); p++)
C.exec(C << &S::do_job << joblist[p]);
while (p<joblist.size())
{
process_return(C.get_returnv());
C.exec(C << &S::do_job << joblist[p++]);
}
while (!C.all_idle())
process_return(C.get_returnv());
3.4
Access to underlying MPI functions
The philosophy of ClassdescMP is not to hide the underlying MPI transport
layer. It is possible to mix MPI calls with ClassdescMP calls, which may be
done to provide a more efficient implementation of a particular operation, or
to provide functionality not provided in ClassdescMP (reductions for example).
This allows ClassdescMP to concentrate on providing new functionality, rather
than simply wrapping existing MPI functionality in a new syntax.
In terms of performance, the only overhead ClassdescMP adds is copying
data into the MPIbuf variable. In the case of sending a large array of a simple
type, it may well be more efficient to call the appropriate MPI call directly. On
the other hand, if one is sending a lot of different small variables, it is more
efficient to marshal the data into a single array, before sending it as a single
message, for which task ClassdescMP is extremely effective.
This philosophy of coexisting with the underlying MPI library is in sharp
contrast with PARA++[6], which was designed to allow the transport layer to
be swapped completely for another one (eg PVM). However, MPI is now so
ubiquitous that swapping the transport layer no longer seems to have much of
an advantage.
Currently, ClassdescMP is implemented completely in terms of MPI-1 functionality. As MPI-2 implementations become available, increased performance,
15
and or functionality dependent on MPI-2 functionality may be added. The most
obvious MPI-2 feature to impact ClassdescMP is one-sided messaging, which
would allow the implementation of the global pointer concept[7]. Unfortunately,
one-sided messaging appears to be the one area of MPI-2 left out of existing
implementations, or implemented badly. One could implement one-sided messaging using a standard threads API, such as Posix threads, however in real
applications encountered to date, separating the communication and computation steps (see §4.6) has proved effective, so we haven’t needed to explore
one-sided communication.
4
Graphcode
Whilst MPISPMD and MPIslave provide rather simple application frameworks
for message passing codes, Graphcode provides a far richer framework within
which programming is closer to data parallel programming than the lower level
message passing environment on which it is based. The underlying paradigm
of Graphcode is objects distributed on a graph. Computation takes place within
the objects (vertices of the graph), and communication takes place along the
edges of the graph.
Graphcode calls the PARMETIS parallel graph partitioner[15, 14] to partition
the graph across the available processors, given a suitable weighting of computational and communication costs (which defaults to a uniform weighting). Since
the solution found by PARMETIS is a Pareto non-dominated solution (no other
partitioning exists that has better load balancing and less communication), the
costs do not need to be provided in any normalised fashion — only the leading
order of computational or communication complexity need be provided.
Since traditional data parallel programs can be expressed as a graph (put
aligned data elements on the same node, express communication patterns as
graph links, eg shifts as nearest neighbour connections), it could be argued
that Graphcode embraces and extends the data parallel programming model.
However the data layout within a compute node differs. For instance, if one
considers a 5-point stencil of some hypothetical 2 component field:
u′i,j
=
′
vi,j
=
1
vi,j − (vi−1,j + vi+1,j + vi,j−1 + vi,j+1 )
4
1
ui,j − (ui−1,j + ui+1,j + ui,j−1 + ui,j+1 )
4
(1)
then Graphcode will store ui,j next to vi,j , whereas an HPF implementation
will store ui,j next to ui+1,j . It remains to be seen what impact this has on
performance in typical situations.
16
4.1
Graphcode objects
A Graphcode graph is represented by the Graph class. Nodes of the graph
are polymorphic objects, derived from the object abstract base class. Being
polymorphic allows more complex topologies, such as hypergraphs, where nodes
may belong to more than groupings. For example consider an object class
representing a human being, and also another object representing the families
that human being might belong to, for instance the family e was born into, and
the family e married into:
class human: public object
{
...
};
class family: public object
{
...
};
The relationship belongs to is represented by a link connecting a human object
with a family object. The reverse link represents the relationship contains.
Graphcode objects may be located on any processor, and may need to
be migrated to achieve dynamic load balancing. Objects are accessed through
proxy variables, of type objref. An objref contains the object’s identifier, its
location (processor ID), and may be dereferenced to obtain access to the object
(if the object is located in the current process’s address space), or a copy of the
object (if it exists in the current process’s address space):
class objref
{
public:
GraphID_t ID;
unsigned int proc;
object& operator*();
object* operator->();
const object* operator->() const;
bool nullref() const;
inline void nullify();
void addref(object* o, bool mflag=false);
};
The members nullref() allow one to test whether the objref points to
a copy of the object in the current address space, and nullify() allows one
to remove the copy of the object. addref(&obj,mflag) points the objref at
object obj, setting mflag (managed flag) to true allows the objref destructor
to destroy the object.
17
Several virtual members need to be provided for any object, including virtual serialisation members pack and unpack as described in §2.6, a “virtual
constructor”, a “virtual copy constructor” and a virtual type identifier. “Virtual constructors” are described in [21, §15.6.2], and the exact procedure used
in Graphcode is detailed below.
To migrate objects between processors, Graphcode will arrange the following
sequence of operations:
Code on source processor
objref a;
MPIbuf b;
b << a.ID << a->type() << *a << send(dest);
Code on destination processor
MPIbuf b;
GraphID_t ID;
int type;
b.get() >> ID >> type;
objref& a=objects[ID];
if (a.nullref())
a.addref(archetype[type]->lnew(),true);
b >> *a;
Note the use of the “virtual constructor” lnew(). We use the type information to index into a database of object archetypes, and call lnew() to obtain a
new object of that type. A programmer defining an object class foo defines the
virtual members as follows:
class foo:
{
public:
virtual
virtual
virtual
virtual
virtual
}
public object
int type() {return vtype(*this);}
object* lnew() {return vnew(this);}
object* lcopy() {return vcopy(this);}
void lpack(pack_t *b);
void lunpack(pack_t *);
The template function vnew() returns a pointer to a new object of the same
type as its argument, and vcopy() returns a copy of the object pointed to by
its argument. Both of these functions use the C++ new operator, so can be
disposed of using delete at a later stage.
Graphcode implements its own runtime type identification — the standard
C++ RTTI typeid() call returns a complex object of type type_info. Not only
is it inefficient to transfer the whole type_info object via MPI, and inefficient
to use a complex object to index into the archetype database, we also have the
potential scenario of the object codes on different processors being generated by
18
different compilers (a heterogeneous computer), and hence having incompatible
type_info objects.
Graphcode’s RTTI system is very simple. An object’s virtual type member
makes a call to the template function vtype, which places a version of itself into
the archetype database:
template <class T> int vtype(const T& x)
{
static int t=-1;
if (t==-1)
{
t=archetype.size();
objref *o=new objref; o->addref(x.lnew());
archetype.push_back(o);
}
return t;
}
Having discussed the virtual function interface of object, we are now ready
to present to full definition of object:
class object: public Ptrlist
{
public:
/* serialisation methods */
virtual void lpack(pack_t *buf)=0;
virtual void lunpack(pack_t *buf)=0;
/* virtual "constructors" */
virtual object* lnew() const=0;
virtual object* lcopy() const=0;
virtual int type() const=0;
virtual idxtype weight() const {return 1;}
virtual idxtype edgeweight(const objref& x) const {return 1;}
};
As well as the virtual members we have described, there are two weight functions
used by the ParMETIS partitioner, used to described the computational cost
(weight()) represented by the object, and the communication cost (edgeweight(x))
in transferring a copy of a remote object x into local address space. As can be
seen, these default to 1, but may be overridden by the programmer of the derived object. Finally, object is derived from Ptrlist, which is syntactically
equivalent to a vector of objref’s, and represents the objects linked to this
object.
As can be seen, there is a lot of similarity between the Graphcode object
type and the object type used with the poly<T> class. Graphcode was the first
real application of Classdesc to a polymorphic data structure, and so its design
stongly influenced that of the poly<T>, which was developed later. At a later
19
stage, we hope to migrate the Graphcode API to use the ref<T> and poly<T>
interfaces to more closely couple Graphcode with Classdesc.
4.2
omap
Each object is identified by a unique identifier, so each process maintains a map
object Graph::objects that can be used to locate the objref corresponding to
a particular identifier. Graphcode supplies two possible map objects — a vmap,
using a std::vector which is optimised for contiguous, or nearly contiguous
ranges of object identifiers, and hmap, a hash map implementation suitable for
non-contiguous identifiers. You select the version of omap you wish to use by
using the namespace graphcode_vmap or graphcode_hmap as appropriate.
It might seem puzzling why the Graph type is not a template, with the
omap type as a template argument. The problem is that internally, objects
need to keep track of the map to which they belong in order to regenerate the
neighbourhood linklist after migration. Therfore, the map type will need to be
a template argument to the objref, but the map itself takes objref as a template
argument, unfortunately leading to a circular template definition of objref:
template <class map> class objref;
template <class map>
class omap: public map
{
};
template <class map>
class objref
{
omap<map> Map;
};
...
typedef std::map<int,objref<Map> > Map;
omap<Map> foo;
In practice, compilers cannot cope with this code.
4.3
Standard library syntax
Wherever possible, the syntax of Graphcode’s containers follows that of the
standard library, so should be familiar to C++ programmers. So Ptrlist and
omap have iterators, and an operator[]. One slight departure from the standard library, is that omap::iterator::operator*() returns an objref, not
pair<GraphID_t, objref>, as one might expect if one followed the std::map
model. The reason for this is that objref objects already contain the object’s
20
identifier, and the pair construct is redundant and wasteful. It also leads to
clearer code.
4.4
Graph
Having introduced objects, objrefs and omaps, we are now in a position to
present a skeleton of Graphcode’s Graph class.
class Graph: public Ptrlist
{
public:
omap objects;
void rebuild_local_list();
void clear_non_local();
template <class T> objref& AddObject(const T& type, GraphID_t id);
void gather();
void Prepare_Neighbours();
void Partition_Objects();
inline void Distribute_Objects();
};
Graph contains two main data members — the objects database mentioned
previously, and a list of object references that refers those objects hosted in the
current address space. This list is a base class of Graph, allowing a simple loop
of the form:
for (Ptrlist::iterator i=begin(); i!=end(); i++)
to be, in effect, a data parallel operation.
The member rebuild_local_list() refreshes this list after a migration of
objects, and the member clear_non_local() nullifies those objrefs that are
not hosted locally, reclaiming memory.
Creating a graph involves calls to AddObject to add an object of type T
(which must be derived from object), and adding the links to each object to
form the graph. For example, the code for a 2D 5-point stencil might look like:
for (i=0; i<nx; i++)
for (j=0; j<ny; j++)
AddObject(foo(),mapid(i,j));
for (i=0; i<nx; i++)
for (j=0; j<ny; j++)
{
objref& o=objects[mapid(i,j)];
o->push_back(objects[mapid(i-1,j)]);
o->push_back(objects[mapid(i+1,j)]);
o->push_back(objects[mapid(i,j-1)]);
o->push_back(objects[mapid(i,j+1)]);
}
21
where the user supplied function mapid(,) converts a coordinate into a pin identifier. Boundary conditions can be handled by returning the special identifier
bad_ID when no link is applicable. The Graph::AddObject and Ptrlist::push_back()
members refuse to add on object having a bad_ID identifier.
4.5
Distribution of data over multiple processors
To distribute objects from the master thread to slave threads, according to some
specified distribution, assign the desired destination of the objects to the proc
member, then call Graph::Distribute_Objects(), which broadcasts the entire
graph to all nodes. There is an inverse Graph::gather() function that gathers
data from all the nodes into the master thread copy.
To partition the objects using PARMETIS, you must first distribute the graph
according to some distribution (no matter how naı̈ve and non-optimal), and
then call Graph::Partition_Objects() to redistribute the Graph more optimally by calling the PARMETIS library. Partition_Objects() can be then
called periodically to rebalance the load, if the graph contains mobile agents for
instance.
Whilst is conceptually the easiest to construct the entire computation on the
master process, and distrbute the data using Graph::Distribute_Objects(),
it is possible to for each process to construct just its part of the computation,
and for Graph::Partition_Objects() to rebalance the load without all the
data needing to pass through a single process’s address space.
4.6
Communication and computation steps
In typical Graphcode applications deployed to date, an update involves performing a computation on each object using the values of the neighbouring objects,
storing the results into a backing buffer graph, and then swapping the backing buffer with the original graph, typical of a synchronous updating scheme.
Asynchronous schemes could be employed as well with due care. The only
communication required is to ensure a copy of all neighbours residing on remote processes is transferred to the processor hosting the object being updated.
Whilst this could be done as needed via one-sided messages, it is more efficient
to batch up all the objects that need to be transferred so that only one message is sent between each pair of processes. Since the communication pattern
is already described by the graph’s links, all a programmer needs to do is make
a call to Graph::Prepare_Neighbours() before starting the computation step.
Returning to our 5 point stencil example (Eq (1)), the update code would be
written as:
graph->Prepare_Neighbours(); /* communication step */
for (Ptrlist::iterator p=graph->begin(); p!=graph->end(); p++)
{
foo* b=fooptr(back->objects[p->ID]);
b->u = fooptr(p)->v;
22
b->v = fooptr(p)->u;
for (Ptrlist::iterator n=p->begin(); n!=p->end(); n++)
{
b->u -= 0.25 * fooptr(n)->v;
b->v -= 0.25 * fooptr(n)->u;
}
}
swap(graph,back);
Here graph and back are the graph and backing buffer for the calculation. We
are also assuming a utility function fooptr() written by the programmer to
return a foo* pointer to the object. A typical implementation of this might be:
foo *fooptr(objref& x) {return dynamic_cast<foo*>(&*x);}
foo *fooptr(Ptrlist::iterator x) {return dynamic_cast<foo*>(&**x);}
It is important to use the new dynamic_cast feature of C++ to catch errors such
x not referring to a foo object, or an incorrect combination of dereferencing and
address-of operators. dynamic_cast will return a NULL pointer in case of error,
which typically causes an immediate NULL dereference error. Old fashioned C
style casts (of the type (foo*)) will simply return an invalid pointer in case of
error, which can be very hard to debug.
It should be noted that a Graph object appears as a list of those objects local
to the executing processor. So this code will execute correctly in parallel. Each
time Prepare_Neighbours is called, the message tag is incremented, preventing
subsequent calls from interfering with the delivery of the previous batches of
messages.
4.7
Deployed applications and performance
Within the EcoLab system[18], Graphcode is deployed with two of the example
models provided with the EcoLab software. These models are working scientific
models, not toy examples. The first model is the spatial Ecolab model[19],
where the panmictic Ecolab model (Lotka-Volterra ecology equations, coupled
with mutation) is replicated over a 2D Cartesian grid, and migration is allowed
between neighbouring grid cells. The performance of this model has not been
studied much yet.
The second model is one of jellyfish in assorted lakes on the islands of Palau.
Each jellyfish is represented as a separate C++ object, commonly called agent
based modelling. The jellyfish move around within a continuous space representing the lake, and from time to time bump into each other. In order to determine
if a collision happens in the next timestep, each jellyfish must examine all the
other jellyfish to see if its path intersects that of the other. This is clearly an
O(N 2 ) serial operation, which severely limits scalability of the model.
To improve scalability, the lake is subdivided into a Cartesian grid, and the
jellyfish is allocated to the cell describing its position. If the cells are sufficiently
large that the jellyfish will only ever pass from one cell to its neighbouring cell
23
35
linear
9am
noon
3pm
6pm
9pm
30
Speedup
25
20
15
10
5
0
0
5
10
15
20
No. processors
25
30
Figure 2: Speedup of the Jellyfish application with 1 million jellyfish in the OTM
lake. Speedup is reported as accumulated wall time for a single processor divided
by the accumulated wall time for n processors for that point in the simulation.
Note that the speedup from 16 processors to 32 is more than double.
in a given timestep, then only the jellyfish within the cell, plus those within
the nearest neighbours need to be examined. This reduces the complexity of
the algorithm to dramatically less than O(N 2 ), and also allows the algorithm
to be executed in parallel. In the field of molecular dynamics simulations, this
method is often called a particle in cell method. PARMETIS allows nodes and
edges to be weighted, so in this case we weight each cell by wi = n2i , and each
edge by vij = nj . In figure 2, the speedup (relative performance of the code
running on n processors versus 1 processor) is plotted for different stages of the
simulation. The simulation starts at 7am with the jellyfish uniformly distributed
throughout the lake. As the sun rises in the east, the jellyfish track the sun, and
become concentrated along the shadow lines. PARMETIS is called repeatedly to
rebalance the calculation. As the sun sets at around 5pm, the jellyfish disperse
randomly throughout the lake. In figure 3, the speedup is plotted as function
of simulation time, so the effect of load unbalancing can be seen. It can be seen
that Graphcode delivers scalable performance in this application.
Graphcode has also been deployed in a 3D artificial chemistry model[8] exhibiting superlinear speedup over 64 processors due to the effectively enlarged
24
35
16
2 processors
4 processors
8 processors
16 processors
32 processors
14
12
Speedup
10
8
6
4
2
0
8
10
12
14
16
Time of day (24 hour clock)
18
20
Figure 3: Speedup of the Jellyfish application as a function of simulation time.
This is differential speedup, calculated from the wall time needed to simulate a
6 minute period, so does not include partitioning time.
25
22
memory cache.
5
Current Status
Classdesc and Graphcode are open source packages written in ISO standard
C++. They have been tested on a range of platforms, and compilers, including
Linux, Mac OSX, Cygwin (Windows), Irix, Tru64; gcc and Intel’s icc for Linux,
as well as native C++ compilers for Irix and Tru64.
The source code is distributed through a SourceForge project, available from
http://ecolab.sourceforge.net. The code is managed by the Aegis source code
management system, which is browsable through a web interface. Version numbers of the form x.Dy are considered “production ready” — they have been
tested on a range of platforms, and are more likely to be reliable. These codes
are also available through the SourceForge file release system. The versions
x.y.Dz are under active development, and have only undergone minimal testing
(ie they should compile, but may still have significant bugs). Developers interested in contributing to the code base can register as a developer of the system
by emailing one of the authors.
Classdesc and Graphcode are also included as part of the EcoLab simulation
system, which is available from the same source code repository.
Acknowledgments
This work was funded by a grant from the Australian Partnership for Advanced
Computing (APAC) under the auspices of its Computational Tools and Techniques programme. Computer time was obtained through the Australian Centre
for Advanced Computation and Communication (ac3).
References
[1] Draft technical report on C++ library extensions. Technical Report DTR
19768, International Standards Organization, 2005.
[2] Charles H. Koelbel amd David B. Loveman, Robert S. Schreiber, Guy L.
Steele, Jr., and Mary E. Zosel. The High Performance Fortran Handbook.
MIT Press, Cambridge, Mass., 1994.
[3] J. M. Squyres an B. C. McCandless and A. Lumsdaine. Object oriented
MPI: A class library for the message passing interface. In Proceedings of
the POOMA conference, 1996.
[4] David M. Beazley.
SWIG : An easy to use tool for integrating scripting languages with C and C++.
In Proceedings of 4th Annual USENIX Tcl/Tk Workshop. USENIX, 1996.
http://www.usenix.org/publications/library/proceedings/tcl96/beazley.html.
26
[5] Greg Colvin, Beman Dawes, and Darin Adler.
http://www.boost.org/.
Boost Smart Pointers.
[6] O. Coulaud and E. Dillon. PARA++: C++ bindings for message passing
libraries. users guide. Technical report, INRIA, 1995.
[7] Dennis Gannon, Shridar Diwan, and Elizabeth Johnson. HPC++ and the
Europa call reification model. ACM SIGAPP Applied Computing Review,
4:11–14, 1996.
[8] Duraid Madina, Naoki Ono, and Takashi Ikegami. Cellular evolution in
a 3d lattice artificial chemistry. In Banzhaf et al., editors, Advances in
Artificial Life, volume 2801 of Lecture Notes in Computer Science, pages
59–68, Berlin, 2003. Springer.
[9] Duraid Madina and Russell K. Standish. A system for reflection in C++.
In Proceedings of AUUG2001: Always on and Everywhere, page 207. Australian Unix Users Group, 2001.
[10] J. K. Ousterhout. TCL and the Tk Toolkit. Addison-Wesley, 1994.
[11] John K. Ousterhout. Scripting: Higher-level programming for the 21st
century. IEEE Computer, 31(3):23–30, 1998.
[12] J.V.W. Reynders, P.J. Hinker, J.C. Cummings, S.R. Atlas, S. Banerjee,
W.F. Humphrey, S.R. Karmesin, K. Keahey, M. Srikant, M.D. Tholburn,
et al. POOMA:aframework for scientific simulations of paralllel architectures. In Parallel Programming in C++, pages 547–588. MIT Press, Cambridge, MA, 1996.
[13] S. Roiser and P. Mato. The SEAL C++ reflection system. Interlaken,
Switzerland, 2004. http://chep2004.web.cern.ch/chep2004/.
[14] Kirk Schloegel, George Karypis, and Vipin Kumar. Parallel multilevel algorithms for multi-constraint graph partitioning. In A. Bode et al., editors,
Euro-Par 2000, volume 1900 of Lecture Notes in Computer Science, page
296, Berlin, 2000. Springer.
[15] Kirk Schloegel, George Karypis, and Vipin Kumar. A unified algorithm
for load-balancing adaptive scientific simulations. In Supercomputing 2000,
2000. http://www.sc2000.org/proceedings/techpapr.
[16] Will Schroeder, Ken Martin, and Bill Lorensen. The visualization toolkit :
an object-oriented approach to 3-D graphics. Prentice Hall, Upper Saddle
River, N.J., 1996.
[17] Marc Snir et al. MPI: the complete reference. MIT Press, Cambridge, MA,
1996.
[18] Russell K. Standish.
Ecolab
http://ecolab.sourceforge.net.
27
documentation.
Available
at
[19] Russell K. Standish. Cellular Ecolab. Complexity International, 6, 1999.
[20] Russell K. Standish and Richard Leow. EcoLab: Agent based modeling for C++ programmers. In Proceedings SwarmFest 2003, 2003.
arXiv:cs.MA/0401026.
[21] Bjarne Stroustrup. The C++ Programming Language. Addison-Wesley,
Reading, Mass., 3rd edition, 1997.
[22] Robert A. van Engelen and Kyle Gallivan. The gSOAP toolkit for web
services and peer-to-peer computing networks. In Proceedings of the 2nd
IEEE International Symposium on Cluster Computing and the Grid (CCGrid2002),, pages 28–135, Berlin, Germany, May 2002.
28
| 5cs.CE
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.